/**
 * \file zstd.c
 * Single-file Zstandard library.
 *
 * Generate using:
 * \code
 *	python combine.py -r ../../lib -x legacy/zstd_legacy.h -o zstd.c zstd-in.c
 * \endcode
 */
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */
/*
 * Settings to bake for the single library file.
 *
 * Note: It's important that none of these affects 'zstd.h' (only the
 * implementation files we're amalgamating).
 *
 * Note: MEM_MODULE stops xxhash redefining BYTE, U16, etc., which are also
 * defined in mem.h (breaking C99 compatibility).
 *
 * Note: the undefs for xxHash allow Zstd's implementation to coincide with
 * standalone xxHash usage (with global defines).
 *
 * Note: if you enable ZSTD_LEGACY_SUPPORT the combine.py script will need
 * re-running without the "-x legacy/zstd_legacy.h" option (it excludes the
 * legacy support at the source level).
 *
 * Note: multithreading is enabled for all platforms apart from Emscripten.
 */
#define DEBUGLEVEL 0
#define MEM_MODULE
#undef  XXH_NAMESPACE
#define XXH_NAMESPACE ZSTD_
#undef  XXH_PRIVATE_API
#define XXH_PRIVATE_API
#undef  XXH_INLINE_ALL
#define XXH_INLINE_ALL
#define ZSTD_LEGACY_SUPPORT 0
#ifndef __EMSCRIPTEN__
#define ZSTD_MULTITHREAD
#endif
#define ZSTD_TRACE 0
/* TODO: Can't amalgamate ASM function */
#define ZSTD_DISABLE_ASM 1

/* Include zstd_deps.h first with all the options we need enabled. */
#define ZSTD_DEPS_NEED_MALLOC
#define ZSTD_DEPS_NEED_MATH64
/**** start inlining common/zstd_deps.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* This file provides common libc dependencies that zstd requires.
 * The purpose is to allow replacing this file with a custom implementation
 * to compile zstd without libc support.
 */

/* Need:
 * NULL
 * INT_MAX
 * UINT_MAX
 * ZSTD_memcpy()
 * ZSTD_memset()
 * ZSTD_memmove()
 */
#ifndef ZSTD_DEPS_COMMON
#define ZSTD_DEPS_COMMON

/* Even though we use qsort_r only for the dictionary builder, the macro
 * _GNU_SOURCE has to be declared *before* the inclusion of any standard
 * header and the script 'combine.sh' combines the whole zstd source code
 * in a single file.
 */
#if defined(__linux) || defined(__linux__) || defined(linux) || defined(__gnu_linux__) || \
    defined(__CYGWIN__) || defined(__MSYS__)
#if !defined(_GNU_SOURCE) && !defined(__ANDROID__) /* NDK doesn't ship qsort_r(). */
#define _GNU_SOURCE
#endif
#endif

#include <limits.h>
#include <stddef.h>
#include <string.h>

#if defined(__GNUC__) && __GNUC__ >= 4
# define ZSTD_memcpy(d,s,l) __builtin_memcpy((d),(s),(l))
# define ZSTD_memmove(d,s,l) __builtin_memmove((d),(s),(l))
# define ZSTD_memset(p,v,l) __builtin_memset((p),(v),(l))
#else
# define ZSTD_memcpy(d,s,l) memcpy((d),(s),(l))
# define ZSTD_memmove(d,s,l) memmove((d),(s),(l))
# define ZSTD_memset(p,v,l) memset((p),(v),(l))
#endif

#endif /* ZSTD_DEPS_COMMON */

/* Need:
 * ZSTD_malloc()
 * ZSTD_free()
 * ZSTD_calloc()
 */
#ifdef ZSTD_DEPS_NEED_MALLOC
#ifndef ZSTD_DEPS_MALLOC
#define ZSTD_DEPS_MALLOC

#include <stdlib.h>

#define ZSTD_malloc(s) malloc(s)
#define ZSTD_calloc(n,s) calloc((n), (s))
#define ZSTD_free(p) free((p))

#endif /* ZSTD_DEPS_MALLOC */
#endif /* ZSTD_DEPS_NEED_MALLOC */

/*
 * Provides 64-bit math support.
 * Need:
 * U64 ZSTD_div64(U64 dividend, U32 divisor)
 */
#ifdef ZSTD_DEPS_NEED_MATH64
#ifndef ZSTD_DEPS_MATH64
#define ZSTD_DEPS_MATH64

#define ZSTD_div64(dividend, divisor) ((dividend) / (divisor))

#endif /* ZSTD_DEPS_MATH64 */
#endif /* ZSTD_DEPS_NEED_MATH64 */

/* Need:
 * assert()
 */
#ifdef ZSTD_DEPS_NEED_ASSERT
#ifndef ZSTD_DEPS_ASSERT
#define ZSTD_DEPS_ASSERT

#include <assert.h>

#endif /* ZSTD_DEPS_ASSERT */
#endif /* ZSTD_DEPS_NEED_ASSERT */

/* Need:
 * ZSTD_DEBUG_PRINT()
 */
#ifdef ZSTD_DEPS_NEED_IO
#ifndef ZSTD_DEPS_IO
#define ZSTD_DEPS_IO

#include <stdio.h>
#define ZSTD_DEBUG_PRINT(...) fprintf(stderr, __VA_ARGS__)

#endif /* ZSTD_DEPS_IO */
#endif /* ZSTD_DEPS_NEED_IO */

/* Only requested when <stdint.h> is known to be present.
 * Need:
 * intptr_t
 */
#ifdef ZSTD_DEPS_NEED_STDINT
#ifndef ZSTD_DEPS_STDINT
#define ZSTD_DEPS_STDINT

#include <stdint.h>

#endif /* ZSTD_DEPS_STDINT */
#endif /* ZSTD_DEPS_NEED_STDINT */
/**** ended inlining common/zstd_deps.h ****/

/**** start inlining common/debug.c ****/
/* ******************************************************************
 * debug
 * Part of FSE library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 * You can contact the author at :
 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */


/*
 * This module only hosts one global variable
 * which can be used to dynamically influence the verbosity of traces,
 * such as DEBUGLOG and RAWLOG
 */

/**** start inlining debug.h ****/
/* ******************************************************************
 * debug
 * Part of FSE library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 * You can contact the author at :
 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */


/*
 * The purpose of this header is to enable debug functions.
 * They regroup assert(), DEBUGLOG() and RAWLOG() for run-time,
 * and DEBUG_STATIC_ASSERT() for compile-time.
 *
 * By default, DEBUGLEVEL==0, which means run-time debug is disabled.
 *
 * Level 1 enables assert() only.
 * Starting level 2, traces can be generated and pushed to stderr.
 * The higher the level, the more verbose the traces.
 *
 * It's possible to dynamically adjust level using variable g_debug_level,
 * which is only declared if DEBUGLEVEL>=2,
 * and is a global variable, not multi-thread protected (use with care)
 */

#ifndef DEBUG_H_12987983217
#define DEBUG_H_12987983217


/* static assert is triggered at compile time, leaving no runtime artefact.
 * static assert only works with compile-time constants.
 * Also, this variant can only be used inside a function. */
#define DEBUG_STATIC_ASSERT(c) (void)sizeof(char[(c) ? 1 : -1])


/* DEBUGLEVEL is expected to be defined externally,
 * typically through compiler command line.
 * Value must be a number. */
#ifndef DEBUGLEVEL
#  define DEBUGLEVEL 0
#endif


/* recommended values for DEBUGLEVEL :
 * 0 : release mode, no debug, all run-time checks disabled
 * 1 : enables assert() only, no display
 * 2 : reserved, for currently active debug path
 * 3 : events once per object lifetime (CCtx, CDict, etc.)
 * 4 : events once per frame
 * 5 : events once per block
 * 6 : events once per sequence (verbose)
 * 7+: events at every position (*very* verbose)
 *
 * It's generally inconvenient to output traces > 5.
 * In which case, it's possible to selectively trigger high verbosity levels
 * by modifying g_debug_level.
 */

#if (DEBUGLEVEL>=1)
#  define ZSTD_DEPS_NEED_ASSERT
/**** skipping file: zstd_deps.h ****/
#else
#  ifndef assert   /* assert may be already defined, due to prior #include <assert.h> */
#    define assert(condition) ((void)0)   /* disable assert (default) */
#  endif
#endif

#if (DEBUGLEVEL>=2)
#  define ZSTD_DEPS_NEED_IO
/**** skipping file: zstd_deps.h ****/
extern int g_debuglevel; /* the variable is only declared,
                            it actually lives in debug.c,
                            and is shared by the whole process.
                            It's not thread-safe.
                            It's useful when enabling very verbose levels
                            on selective conditions (such as position in src) */

#  define RAWLOG(l, ...)                   \
    do {                                   \
        if (l<=g_debuglevel) {             \
            ZSTD_DEBUG_PRINT(__VA_ARGS__); \
        }                                  \
    } while (0)

#define STRINGIFY(x) #x
#define TOSTRING(x) STRINGIFY(x)
#define LINE_AS_STRING TOSTRING(__LINE__)

#  define DEBUGLOG(l, ...)                               \
    do {                                                 \
        if (l<=g_debuglevel) {                           \
            ZSTD_DEBUG_PRINT(__FILE__ ":" LINE_AS_STRING ": " __VA_ARGS__); \
            ZSTD_DEBUG_PRINT(" \n");                     \
        }                                                \
    } while (0)
#else
#  define RAWLOG(l, ...)   do { } while (0)    /* disabled */
#  define DEBUGLOG(l, ...) do { } while (0)    /* disabled */
#endif

#endif /* DEBUG_H_12987983217 */
/**** ended inlining debug.h ****/

#if !defined(ZSTD_LINUX_KERNEL) || (DEBUGLEVEL>=2)
/* We only use this when DEBUGLEVEL>=2, but we get -Werror=pedantic errors if a
 * translation unit is empty. So remove this from Linux kernel builds, but
 * otherwise just leave it in.
 */
int g_debuglevel = DEBUGLEVEL;
#endif
/**** ended inlining common/debug.c ****/
/**** start inlining common/entropy_common.c ****/
/* ******************************************************************
 * Common functions of New Generation Entropy library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* *************************************
*  Dependencies
***************************************/
/**** start inlining mem.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef MEM_H_MODULE
#define MEM_H_MODULE

/*-****************************************
*  Dependencies
******************************************/
#include <stddef.h>  /* size_t, ptrdiff_t */
/**** start inlining compiler.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_COMPILER_H
#define ZSTD_COMPILER_H

#include <stddef.h>

/**** start inlining portability_macros.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_PORTABILITY_MACROS_H
#define ZSTD_PORTABILITY_MACROS_H

/**
 * This header file contains macro definitions to support portability.
 * This header is shared between C and ASM code, so it MUST only
 * contain macro definitions. It MUST not contain any C code.
 *
 * This header ONLY defines macros to detect platforms/feature support.
 *
 */


/* compat. with non-clang compilers */
#ifndef __has_attribute
  #define __has_attribute(x) 0
#endif

/* compat. with non-clang compilers */
#ifndef __has_builtin
#  define __has_builtin(x) 0
#endif

/* compat. with non-clang compilers */
#ifndef __has_feature
#  define __has_feature(x) 0
#endif

/* detects whether we are being compiled under msan */
#ifndef ZSTD_MEMORY_SANITIZER
#  if __has_feature(memory_sanitizer)
#    define ZSTD_MEMORY_SANITIZER 1
#  else
#    define ZSTD_MEMORY_SANITIZER 0
#  endif
#endif

/* detects whether we are being compiled under asan */
#ifndef ZSTD_ADDRESS_SANITIZER
#  if __has_feature(address_sanitizer)
#    define ZSTD_ADDRESS_SANITIZER 1
#  elif defined(__SANITIZE_ADDRESS__)
#    define ZSTD_ADDRESS_SANITIZER 1
#  else
#    define ZSTD_ADDRESS_SANITIZER 0
#  endif
#endif

/* detects whether we are being compiled under dfsan */
#ifndef ZSTD_DATAFLOW_SANITIZER
# if __has_feature(dataflow_sanitizer)
#  define ZSTD_DATAFLOW_SANITIZER 1
# else
#  define ZSTD_DATAFLOW_SANITIZER 0
# endif
#endif

/* Mark the internal assembly functions as hidden  */
#ifdef __ELF__
# define ZSTD_HIDE_ASM_FUNCTION(func) .hidden func
#elif defined(__APPLE__)
# define ZSTD_HIDE_ASM_FUNCTION(func) .private_extern func
#else
# define ZSTD_HIDE_ASM_FUNCTION(func)
#endif

/* Compile time determination of BMI2 support */
#ifndef STATIC_BMI2
#  if defined(__BMI2__)
#    define STATIC_BMI2 1
#  elif defined(_MSC_VER) && defined(__AVX2__)
#    define STATIC_BMI2 1 /* MSVC does not have a BMI2 specific flag, but every CPU that supports AVX2 also supports BMI2 */
#  endif
#endif

#ifndef STATIC_BMI2
#  define STATIC_BMI2 0
#endif

/* Enable runtime BMI2 dispatch based on the CPU.
 * Enabled for clang & gcc >=4.8 on x86 when BMI2 isn't enabled by default.
 */
#ifndef DYNAMIC_BMI2
#  if ((defined(__clang__) && __has_attribute(__target__)) \
      || (defined(__GNUC__) \
          && (__GNUC__ >= 5 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8)))) \
      && (defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)) \
      && !defined(__BMI2__)
#    define DYNAMIC_BMI2 1
#  else
#    define DYNAMIC_BMI2 0
#  endif
#endif

/**
 * Only enable assembly for GNU C compatible compilers,
 * because other platforms may not support GAS assembly syntax.
 *
 * Only enable assembly for Linux / MacOS / Win32, other platforms may
 * work, but they haven't been tested. This could likely be
 * extended to BSD systems.
 *
 * Disable assembly when MSAN is enabled, because MSAN requires
 * 100% of code to be instrumented to work.
 */
#if defined(__GNUC__)
#  if defined(__linux__) || defined(__linux) || defined(__APPLE__) || defined(_WIN32)
#    if ZSTD_MEMORY_SANITIZER
#      define ZSTD_ASM_SUPPORTED 0
#    elif ZSTD_DATAFLOW_SANITIZER
#      define ZSTD_ASM_SUPPORTED 0
#    else
#      define ZSTD_ASM_SUPPORTED 1
#    endif
#  else
#    define ZSTD_ASM_SUPPORTED 0
#  endif
#else
#  define ZSTD_ASM_SUPPORTED 0
#endif

/**
 * Determines whether we should enable assembly for x86-64
 * with BMI2.
 *
 * Enable if all of the following conditions hold:
 * - ASM hasn't been explicitly disabled by defining ZSTD_DISABLE_ASM
 * - Assembly is supported
 * - We are compiling for x86-64 and either:
 *   - DYNAMIC_BMI2 is enabled
 *   - BMI2 is supported at compile time
 */
#if !defined(ZSTD_DISABLE_ASM) &&                                 \
    ZSTD_ASM_SUPPORTED &&                                         \
    defined(__x86_64__) &&                                        \
    (DYNAMIC_BMI2 || defined(__BMI2__))
# define ZSTD_ENABLE_ASM_X86_64_BMI2 1
#else
# define ZSTD_ENABLE_ASM_X86_64_BMI2 0
#endif

/*
 * For x86 ELF targets, add .note.gnu.property section for Intel CET in
 * assembly sources when CET is enabled.
 *
 * Additionally, any function that may be called indirectly must begin
 * with ZSTD_CET_ENDBRANCH.
 */
#if defined(__ELF__) && (defined(__x86_64__) || defined(__i386__)) \
    && defined(__has_include)
# if __has_include(<cet.h>)
#  include <cet.h>
#  define ZSTD_CET_ENDBRANCH _CET_ENDBR
# endif
#endif

#ifndef ZSTD_CET_ENDBRANCH
# define ZSTD_CET_ENDBRANCH
#endif

#endif /* ZSTD_PORTABILITY_MACROS_H */
/**** ended inlining portability_macros.h ****/

/*-*******************************************************
*  Compiler specifics
*********************************************************/
/* force inlining */

#if !defined(ZSTD_NO_INLINE)
#if (defined(__GNUC__) && !defined(__STRICT_ANSI__)) || defined(__cplusplus) || defined(__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* C99 */
#  define INLINE_KEYWORD inline
#else
#  define INLINE_KEYWORD
#endif

#if defined(__GNUC__) || defined(__IAR_SYSTEMS_ICC__)
#  define FORCE_INLINE_ATTR __attribute__((always_inline))
#elif defined(_MSC_VER)
#  define FORCE_INLINE_ATTR __forceinline
#else
#  define FORCE_INLINE_ATTR
#endif

#else

#define INLINE_KEYWORD
#define FORCE_INLINE_ATTR

#endif

/**
  On MSVC qsort requires that functions passed into it use the __cdecl calling conversion(CC).
  This explicitly marks such functions as __cdecl so that the code will still compile
  if a CC other than __cdecl has been made the default.
*/
#if  defined(_MSC_VER)
#  define WIN_CDECL __cdecl
#else
#  define WIN_CDECL
#endif

/* UNUSED_ATTR tells the compiler it is okay if the function is unused. */
#if defined(__GNUC__) || defined(__IAR_SYSTEMS_ICC__)
#  define UNUSED_ATTR __attribute__((unused))
#else
#  define UNUSED_ATTR
#endif

/**
 * FORCE_INLINE_TEMPLATE is used to define C "templates", which take constant
 * parameters. They must be inlined for the compiler to eliminate the constant
 * branches.
 */
#define FORCE_INLINE_TEMPLATE static INLINE_KEYWORD FORCE_INLINE_ATTR UNUSED_ATTR
/**
 * HINT_INLINE is used to help the compiler generate better code. It is *not*
 * used for "templates", so it can be tweaked based on the compilers
 * performance.
 *
 * gcc-4.8 and gcc-4.9 have been shown to benefit from leaving off the
 * always_inline attribute.
 *
 * clang up to 5.0.0 (trunk) benefit tremendously from the always_inline
 * attribute.
 */
#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ >= 4 && __GNUC_MINOR__ >= 8 && __GNUC__ < 5
#  define HINT_INLINE static INLINE_KEYWORD
#else
#  define HINT_INLINE FORCE_INLINE_TEMPLATE
#endif

/* "soft" inline :
 * The compiler is free to select if it's a good idea to inline or not.
 * The main objective is to silence compiler warnings
 * when a defined function in included but not used.
 *
 * Note : this macro is prefixed `MEM_` because it used to be provided by `mem.h` unit.
 * Updating the prefix is probably preferable, but requires a fairly large codemod,
 * since this name is used everywhere.
 */
#ifndef MEM_STATIC  /* already defined in Linux Kernel mem.h */
#if defined(__GNUC__)
#  define MEM_STATIC static __inline UNUSED_ATTR
#elif defined(__IAR_SYSTEMS_ICC__)
#  define MEM_STATIC static inline UNUSED_ATTR
#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
#  define MEM_STATIC static inline
#elif defined(_MSC_VER)
#  define MEM_STATIC static __inline
#else
#  define MEM_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
#endif
#endif

/* force no inlining */
#ifdef _MSC_VER
#  define FORCE_NOINLINE static __declspec(noinline)
#else
#  if defined(__GNUC__) || defined(__IAR_SYSTEMS_ICC__)
#    define FORCE_NOINLINE static __attribute__((__noinline__))
#  else
#    define FORCE_NOINLINE static
#  endif
#endif


/* target attribute */
#if defined(__GNUC__) || defined(__IAR_SYSTEMS_ICC__)
#  define TARGET_ATTRIBUTE(target) __attribute__((__target__(target)))
#else
#  define TARGET_ATTRIBUTE(target)
#endif

/* Target attribute for BMI2 dynamic dispatch.
 * Enable lzcnt, bmi, and bmi2.
 * We test for bmi1 & bmi2. lzcnt is included in bmi1.
 */
#define BMI2_TARGET_ATTRIBUTE TARGET_ATTRIBUTE("lzcnt,bmi,bmi2")

/* prefetch
 * can be disabled, by declaring NO_PREFETCH build macro */
#if defined(NO_PREFETCH)
#  define PREFETCH_L1(ptr)  do { (void)(ptr); } while (0)  /* disabled */
#  define PREFETCH_L2(ptr)  do { (void)(ptr); } while (0)  /* disabled */
#else
#  if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_I86)) && !defined(_M_ARM64EC)  /* _mm_prefetch() is not defined outside of x86/x64 */
#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
#    define PREFETCH_L1(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
#    define PREFETCH_L2(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T1)
#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
#    define PREFETCH_L1(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
#    define PREFETCH_L2(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 2 /* locality */)
#  elif defined(__aarch64__)
#    define PREFETCH_L1(ptr)  do { __asm__ __volatile__("prfm pldl1keep, %0" ::"Q"(*(ptr))); } while (0)
#    define PREFETCH_L2(ptr)  do { __asm__ __volatile__("prfm pldl2keep, %0" ::"Q"(*(ptr))); } while (0)
#  else
#    define PREFETCH_L1(ptr) do { (void)(ptr); } while (0)  /* disabled */
#    define PREFETCH_L2(ptr) do { (void)(ptr); } while (0)  /* disabled */
#  endif
#endif  /* NO_PREFETCH */

#define CACHELINE_SIZE 64

#define PREFETCH_AREA(p, s)                              \
    do {                                                 \
        const char* const _ptr = (const char*)(p);       \
        size_t const _size = (size_t)(s);                \
        size_t _pos;                                     \
        for (_pos=0; _pos<_size; _pos+=CACHELINE_SIZE) { \
            PREFETCH_L2(_ptr + _pos);                    \
        }                                                \
    } while (0)

/* vectorization
 * older GCC (pre gcc-4.3 picked as the cutoff) uses a different syntax,
 * and some compilers, like Intel ICC and MCST LCC, do not support it at all. */
#if !defined(__INTEL_COMPILER) && !defined(__clang__) && defined(__GNUC__) && !defined(__LCC__)
#  if (__GNUC__ == 4 && __GNUC_MINOR__ > 3) || (__GNUC__ >= 5)
#    define DONT_VECTORIZE __attribute__((optimize("no-tree-vectorize")))
#  else
#    define DONT_VECTORIZE _Pragma("GCC optimize(\"no-tree-vectorize\")")
#  endif
#else
#  define DONT_VECTORIZE
#endif

/* Tell the compiler that a branch is likely or unlikely.
 * Only use these macros if it causes the compiler to generate better code.
 * If you can remove a LIKELY/UNLIKELY annotation without speed changes in gcc
 * and clang, please do.
 */
#if defined(__GNUC__)
#define LIKELY(x) (__builtin_expect((x), 1))
#define UNLIKELY(x) (__builtin_expect((x), 0))
#else
#define LIKELY(x) (x)
#define UNLIKELY(x) (x)
#endif

#if __has_builtin(__builtin_unreachable) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 5)))
#  define ZSTD_UNREACHABLE do { assert(0), __builtin_unreachable(); } while (0)
#else
#  define ZSTD_UNREACHABLE do { assert(0); } while (0)
#endif

/* disable warnings */
#ifdef _MSC_VER    /* Visual Studio */
#  include <intrin.h>                    /* For Visual 2005 */
#  pragma warning(disable : 4100)        /* disable: C4100: unreferenced formal parameter */
#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
#  pragma warning(disable : 4204)        /* disable: C4204: non-constant aggregate initializer */
#  pragma warning(disable : 4214)        /* disable: C4214: non-int bitfields */
#  pragma warning(disable : 4324)        /* disable: C4324: padded structure */
#endif

/* compile time determination of SIMD support */
#if !defined(ZSTD_NO_INTRINSICS)
#  if defined(__AVX2__)
#    define ZSTD_ARCH_X86_AVX2
#  endif
#  if defined(__SSE2__) || defined(_M_X64) || (defined (_M_IX86) && defined(_M_IX86_FP) && (_M_IX86_FP >= 2))
#    define ZSTD_ARCH_X86_SSE2
#  endif
#  if defined(__ARM_NEON) || defined(_M_ARM64)
#    define ZSTD_ARCH_ARM_NEON
#  endif
#
#  if defined(ZSTD_ARCH_X86_AVX2)
#    include <immintrin.h>
#  endif
#  if defined(ZSTD_ARCH_X86_SSE2)
#    include <emmintrin.h>
#  elif defined(ZSTD_ARCH_ARM_NEON)
#    include <arm_neon.h>
#  endif
#endif

/* C-language Attributes are added in C23. */
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ > 201710L) && defined(__has_c_attribute)
# define ZSTD_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
#else
# define ZSTD_HAS_C_ATTRIBUTE(x) 0
#endif

/* Only use C++ attributes in C++. Some compilers report support for C++
 * attributes when compiling with C.
 */
#if defined(__cplusplus) && defined(__has_cpp_attribute)
# define ZSTD_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
#else
# define ZSTD_HAS_CPP_ATTRIBUTE(x) 0
#endif

/* Define ZSTD_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute.
 * - C23: https://en.cppreference.com/w/c/language/attributes/fallthrough
 * - CPP17: https://en.cppreference.com/w/cpp/language/attributes/fallthrough
 * - Else: __attribute__((__fallthrough__))
 */
#ifndef ZSTD_FALLTHROUGH
# if ZSTD_HAS_C_ATTRIBUTE(fallthrough)
#  define ZSTD_FALLTHROUGH [[fallthrough]]
# elif ZSTD_HAS_CPP_ATTRIBUTE(fallthrough)
#  define ZSTD_FALLTHROUGH [[fallthrough]]
# elif __has_attribute(__fallthrough__)
/* Leading semicolon is to satisfy gcc-11 with -pedantic. Without the semicolon
 * gcc complains about: a label can only be part of a statement and a declaration is not a statement.
 */
#  define ZSTD_FALLTHROUGH ; __attribute__((__fallthrough__))
# else
#  define ZSTD_FALLTHROUGH
# endif
#endif

/*-**************************************************************
*  Alignment
*****************************************************************/

/* @return 1 if @u is a 2^n value, 0 otherwise
 * useful to check a value is valid for alignment restrictions */
MEM_STATIC int ZSTD_isPower2(size_t u) {
    return (u & (u-1)) == 0;
}

/* this test was initially positioned in mem.h,
 * but this file is removed (or replaced) for linux kernel
 * so it's now hosted in compiler.h,
 * which remains valid for both user & kernel spaces.
 */

#ifndef ZSTD_ALIGNOF
# if defined(__GNUC__) || defined(_MSC_VER)
/* covers gcc, clang & MSVC */
/* note : this section must come first, before C11,
 * due to a limitation in the kernel source generator */
#  define ZSTD_ALIGNOF(T) __alignof(T)

# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)
/* C11 support */
#  include <stdalign.h>
#  define ZSTD_ALIGNOF(T) alignof(T)

# else
/* No known support for alignof() - imperfect backup */
#  define ZSTD_ALIGNOF(T) (sizeof(void*) < sizeof(T) ? sizeof(void*) : sizeof(T))

# endif
#endif /* ZSTD_ALIGNOF */

#ifndef ZSTD_ALIGNED
/* C90-compatible alignment macro (GCC/Clang). Adjust for other compilers if needed. */
# if defined(__GNUC__) || defined(__clang__)
#  define ZSTD_ALIGNED(a) __attribute__((aligned(a)))
# elif defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* C11 */
#  define ZSTD_ALIGNED(a) _Alignas(a)
#elif defined(_MSC_VER)
#  define ZSTD_ALIGNED(n) __declspec(align(n))
# else
   /* this compiler will require its own alignment instruction */
#  define ZSTD_ALIGNED(...)
# endif
#endif /* ZSTD_ALIGNED */


/*-**************************************************************
*  Sanitizer
*****************************************************************/

/**
 * Zstd relies on pointer overflow in its decompressor.
 * We add this attribute to functions that rely on pointer overflow.
 */
#ifndef ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
#  if __has_attribute(no_sanitize)
#    if !defined(__clang__) && defined(__GNUC__) && __GNUC__ < 8
       /* gcc < 8 only has signed-integer-overlow which triggers on pointer overflow */
#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("signed-integer-overflow")))
#    else
       /* older versions of clang [3.7, 5.0) will warn that pointer-overflow is ignored. */
#      define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR __attribute__((no_sanitize("pointer-overflow")))
#    endif
#  else
#    define ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
#  endif
#endif

/**
 * Helper function to perform a wrapped pointer difference without triggering
 * UBSAN.
 *
 * @returns lhs - rhs with wrapping
 */
MEM_STATIC
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
ptrdiff_t ZSTD_wrappedPtrDiff(unsigned char const* lhs, unsigned char const* rhs)
{
    return lhs - rhs;
}

/**
 * Helper function to perform a wrapped pointer add without triggering UBSAN.
 *
 * @return ptr + add with wrapping
 */
MEM_STATIC
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
unsigned char const* ZSTD_wrappedPtrAdd(unsigned char const* ptr, ptrdiff_t add)
{
    return ptr + add;
}

/**
 * Helper function to perform a wrapped pointer subtraction without triggering
 * UBSAN.
 *
 * @return ptr - sub with wrapping
 */
MEM_STATIC
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
unsigned char const* ZSTD_wrappedPtrSub(unsigned char const* ptr, ptrdiff_t sub)
{
    return ptr - sub;
}

/**
 * Helper function to add to a pointer that works around C's undefined behavior
 * of adding 0 to NULL.
 *
 * @returns `ptr + add` except it defines `NULL + 0 == NULL`.
 */
MEM_STATIC
unsigned char* ZSTD_maybeNullPtrAdd(unsigned char* ptr, ptrdiff_t add)
{
    return add > 0 ? ptr + add : ptr;
}

/* Issue #3240 reports an ASAN failure on an llvm-mingw build. Out of an
 * abundance of caution, disable our custom poisoning on mingw. */
#ifdef __MINGW32__
#ifndef ZSTD_ASAN_DONT_POISON_WORKSPACE
#define ZSTD_ASAN_DONT_POISON_WORKSPACE 1
#endif
#ifndef ZSTD_MSAN_DONT_POISON_WORKSPACE
#define ZSTD_MSAN_DONT_POISON_WORKSPACE 1
#endif
#endif

#if ZSTD_MEMORY_SANITIZER && !defined(ZSTD_MSAN_DONT_POISON_WORKSPACE)
/* Not all platforms that support msan provide sanitizers/msan_interface.h.
 * We therefore declare the functions we need ourselves, rather than trying to
 * include the header file... */
#include <stddef.h>  /* size_t */
#define ZSTD_DEPS_NEED_STDINT
/**** skipping file: zstd_deps.h ****/

/* Make memory region fully initialized (without changing its contents). */
void __msan_unpoison(const volatile void *a, size_t size);

/* Make memory region fully uninitialized (without changing its contents).
   This is a legacy interface that does not update origin information. Use
   __msan_allocated_memory() instead. */
void __msan_poison(const volatile void *a, size_t size);

/* Returns the offset of the first (at least partially) poisoned byte in the
   memory range, or -1 if the whole range is good. */
intptr_t __msan_test_shadow(const volatile void *x, size_t size);

/* Print shadow and origin for the memory range to stderr in a human-readable
   format. */
void __msan_print_shadow(const volatile void *x, size_t size);
#endif

#if ZSTD_ADDRESS_SANITIZER && !defined(ZSTD_ASAN_DONT_POISON_WORKSPACE)
/* Not all platforms that support asan provide sanitizers/asan_interface.h.
 * We therefore declare the functions we need ourselves, rather than trying to
 * include the header file... */
#include <stddef.h>  /* size_t */

/**
 * Marks a memory region (<c>[addr, addr+size)</c>) as unaddressable.
 *
 * This memory must be previously allocated by your program. Instrumented
 * code is forbidden from accessing addresses in this region until it is
 * unpoisoned. This function is not guaranteed to poison the entire region -
 * it could poison only a subregion of <c>[addr, addr+size)</c> due to ASan
 * alignment restrictions.
 *
 * \note This function is not thread-safe because no two threads can poison or
 * unpoison memory in the same memory region simultaneously.
 *
 * \param addr Start of memory region.
 * \param size Size of memory region. */
void __asan_poison_memory_region(void const volatile *addr, size_t size);

/**
 * Marks a memory region (<c>[addr, addr+size)</c>) as addressable.
 *
 * This memory must be previously allocated by your program. Accessing
 * addresses in this region is allowed until this region is poisoned again.
 * This function could unpoison a super-region of <c>[addr, addr+size)</c> due
 * to ASan alignment restrictions.
 *
 * \note This function is not thread-safe because no two threads can
 * poison or unpoison memory in the same memory region simultaneously.
 *
 * \param addr Start of memory region.
 * \param size Size of memory region. */
void __asan_unpoison_memory_region(void const volatile *addr, size_t size);
#endif

#endif /* ZSTD_COMPILER_H */
/**** ended inlining compiler.h ****/
/**** skipping file: debug.h ****/
/**** skipping file: zstd_deps.h ****/


/*-****************************************
*  Compiler specifics
******************************************/
#if defined(_MSC_VER)   /* Visual Studio */
#   include <stdlib.h>  /* _byteswap_ulong */
#   include <intrin.h>  /* _byteswap_* */
#elif defined(__ICCARM__)
#   include <intrinsics.h>
#endif

/*-**************************************************************
*  Basic Types
*****************************************************************/
#if  !defined (__VMS) && (defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
#  if defined(_AIX)
#    include <inttypes.h>
#  else
#    include <stdint.h> /* intptr_t */
#  endif
  typedef   uint8_t BYTE;
  typedef   uint8_t U8;
  typedef    int8_t S8;
  typedef  uint16_t U16;
  typedef   int16_t S16;
  typedef  uint32_t U32;
  typedef   int32_t S32;
  typedef  uint64_t U64;
  typedef   int64_t S64;
#else
# include <limits.h>
#if CHAR_BIT != 8
#  error "this implementation requires char to be exactly 8-bit type"
#endif
  typedef unsigned char      BYTE;
  typedef unsigned char      U8;
  typedef   signed char      S8;
#if USHRT_MAX != 65535
#  error "this implementation requires short to be exactly 16-bit type"
#endif
  typedef unsigned short      U16;
  typedef   signed short      S16;
#if UINT_MAX != 4294967295
#  error "this implementation requires int to be exactly 32-bit type"
#endif
  typedef unsigned int        U32;
  typedef   signed int        S32;
/* note : there are no limits defined for long long type in C90.
 * limits exist in C99, however, in such case, <stdint.h> is preferred */
  typedef unsigned long long  U64;
  typedef   signed long long  S64;
#endif

/*-**************************************************************
*  Memory I/O API
*****************************************************************/
/*=== Static platform detection ===*/
MEM_STATIC unsigned MEM_32bits(void);
MEM_STATIC unsigned MEM_64bits(void);
MEM_STATIC unsigned MEM_isLittleEndian(void);

/*=== Native unaligned read/write ===*/
MEM_STATIC U16 MEM_read16(const void* memPtr);
MEM_STATIC U32 MEM_read32(const void* memPtr);
MEM_STATIC U64 MEM_read64(const void* memPtr);
MEM_STATIC size_t MEM_readST(const void* memPtr);

MEM_STATIC void MEM_write16(void* memPtr, U16 value);
MEM_STATIC void MEM_write32(void* memPtr, U32 value);
MEM_STATIC void MEM_write64(void* memPtr, U64 value);

/*=== Little endian unaligned read/write ===*/
MEM_STATIC U16 MEM_readLE16(const void* memPtr);
MEM_STATIC U32 MEM_readLE24(const void* memPtr);
MEM_STATIC U32 MEM_readLE32(const void* memPtr);
MEM_STATIC U64 MEM_readLE64(const void* memPtr);
MEM_STATIC size_t MEM_readLEST(const void* memPtr);

MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val);
MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val);
MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32);
MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64);
MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val);

/*=== Big endian unaligned read/write ===*/
MEM_STATIC U32 MEM_readBE32(const void* memPtr);
MEM_STATIC U64 MEM_readBE64(const void* memPtr);
MEM_STATIC size_t MEM_readBEST(const void* memPtr);

MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32);
MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64);
MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val);

/*=== Byteswap ===*/
MEM_STATIC U32 MEM_swap32(U32 in);
MEM_STATIC U64 MEM_swap64(U64 in);
MEM_STATIC size_t MEM_swapST(size_t in);


/*-**************************************************************
*  Memory I/O Implementation
*****************************************************************/
/* MEM_FORCE_MEMORY_ACCESS : For accessing unaligned memory:
 * Method 0 : always use `memcpy()`. Safe and portable.
 * Method 1 : Use compiler extension to set unaligned access.
 * Method 2 : direct access. This method is portable but violate C standard.
 *            It can generate buggy code on targets depending on alignment.
 * Default  : method 1 if supported, else method 0
 */
#ifndef MEM_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
#  ifdef __GNUC__
#    define MEM_FORCE_MEMORY_ACCESS 1
#  endif
#endif

MEM_STATIC unsigned MEM_32bits(void) { return sizeof(size_t)==4; }
MEM_STATIC unsigned MEM_64bits(void) { return sizeof(size_t)==8; }

MEM_STATIC unsigned MEM_isLittleEndian(void)
{
#if defined(__BYTE_ORDER__) && defined(__ORDER_LITTLE_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
    return 1;
#elif defined(__BYTE_ORDER__) && defined(__ORDER_BIG_ENDIAN__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
    return 0;
#elif defined(__clang__) && __LITTLE_ENDIAN__
    return 1;
#elif defined(__clang__) && __BIG_ENDIAN__
    return 0;
#elif defined(_MSC_VER) && (_M_X64 || _M_IX86)
    return 1;
#elif defined(__DMC__) && defined(_M_IX86)
    return 1;
#elif defined(__IAR_SYSTEMS_ICC__) && __LITTLE_ENDIAN__
    return 1;
#else
    const union { U32 u; BYTE c[4]; } one = { 1 };   /* don't use static : performance detrimental  */
    return one.c[0];
#endif
}

#if defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==2)

/* violates C standard, by lying on structure alignment.
Only use if no other choice to achieve best performance on target platform */
MEM_STATIC U16 MEM_read16(const void* memPtr) { return *(const U16*) memPtr; }
MEM_STATIC U32 MEM_read32(const void* memPtr) { return *(const U32*) memPtr; }
MEM_STATIC U64 MEM_read64(const void* memPtr) { return *(const U64*) memPtr; }
MEM_STATIC size_t MEM_readST(const void* memPtr) { return *(const size_t*) memPtr; }

MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(U16*)memPtr = value; }
MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(U32*)memPtr = value; }
MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(U64*)memPtr = value; }

#elif defined(MEM_FORCE_MEMORY_ACCESS) && (MEM_FORCE_MEMORY_ACCESS==1)

typedef __attribute__((aligned(1))) U16 unalign16;
typedef __attribute__((aligned(1))) U32 unalign32;
typedef __attribute__((aligned(1))) U64 unalign64;
typedef __attribute__((aligned(1))) size_t unalignArch;

MEM_STATIC U16 MEM_read16(const void* ptr) { return *(const unalign16*)ptr; }
MEM_STATIC U32 MEM_read32(const void* ptr) { return *(const unalign32*)ptr; }
MEM_STATIC U64 MEM_read64(const void* ptr) { return *(const unalign64*)ptr; }
MEM_STATIC size_t MEM_readST(const void* ptr) { return *(const unalignArch*)ptr; }

MEM_STATIC void MEM_write16(void* memPtr, U16 value) { *(unalign16*)memPtr = value; }
MEM_STATIC void MEM_write32(void* memPtr, U32 value) { *(unalign32*)memPtr = value; }
MEM_STATIC void MEM_write64(void* memPtr, U64 value) { *(unalign64*)memPtr = value; }

#else

/* default method, safe and standard.
   can sometimes prove slower */

MEM_STATIC U16 MEM_read16(const void* memPtr)
{
    U16 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
}

MEM_STATIC U32 MEM_read32(const void* memPtr)
{
    U32 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
}

MEM_STATIC U64 MEM_read64(const void* memPtr)
{
    U64 val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
}

MEM_STATIC size_t MEM_readST(const void* memPtr)
{
    size_t val; ZSTD_memcpy(&val, memPtr, sizeof(val)); return val;
}

MEM_STATIC void MEM_write16(void* memPtr, U16 value)
{
    ZSTD_memcpy(memPtr, &value, sizeof(value));
}

MEM_STATIC void MEM_write32(void* memPtr, U32 value)
{
    ZSTD_memcpy(memPtr, &value, sizeof(value));
}

MEM_STATIC void MEM_write64(void* memPtr, U64 value)
{
    ZSTD_memcpy(memPtr, &value, sizeof(value));
}

#endif /* MEM_FORCE_MEMORY_ACCESS */

MEM_STATIC U32 MEM_swap32_fallback(U32 in)
{
    return  ((in << 24) & 0xff000000 ) |
            ((in <<  8) & 0x00ff0000 ) |
            ((in >>  8) & 0x0000ff00 ) |
            ((in >> 24) & 0x000000ff );
}

MEM_STATIC U32 MEM_swap32(U32 in)
{
#if defined(_MSC_VER)     /* Visual Studio */
    return _byteswap_ulong(in);
#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
  || (defined(__clang__) && __has_builtin(__builtin_bswap32))
    return __builtin_bswap32(in);
#elif defined(__ICCARM__)
    return __REV(in);
#else
    return MEM_swap32_fallback(in);
#endif
}

MEM_STATIC U64 MEM_swap64_fallback(U64 in)
{
     return  ((in << 56) & 0xff00000000000000ULL) |
            ((in << 40) & 0x00ff000000000000ULL) |
            ((in << 24) & 0x0000ff0000000000ULL) |
            ((in << 8)  & 0x000000ff00000000ULL) |
            ((in >> 8)  & 0x00000000ff000000ULL) |
            ((in >> 24) & 0x0000000000ff0000ULL) |
            ((in >> 40) & 0x000000000000ff00ULL) |
            ((in >> 56) & 0x00000000000000ffULL);
}

MEM_STATIC U64 MEM_swap64(U64 in)
{
#if defined(_MSC_VER)     /* Visual Studio */
    return _byteswap_uint64(in);
#elif (defined (__GNUC__) && (__GNUC__ * 100 + __GNUC_MINOR__ >= 403)) \
  || (defined(__clang__) && __has_builtin(__builtin_bswap64))
    return __builtin_bswap64(in);
#else
    return MEM_swap64_fallback(in);
#endif
}

MEM_STATIC size_t MEM_swapST(size_t in)
{
    if (MEM_32bits())
        return (size_t)MEM_swap32((U32)in);
    else
        return (size_t)MEM_swap64((U64)in);
}

/*=== Little endian r/w ===*/

MEM_STATIC U16 MEM_readLE16(const void* memPtr)
{
    if (MEM_isLittleEndian())
        return MEM_read16(memPtr);
    else {
        const BYTE* p = (const BYTE*)memPtr;
        return (U16)(p[0] + (p[1]<<8));
    }
}

MEM_STATIC void MEM_writeLE16(void* memPtr, U16 val)
{
    if (MEM_isLittleEndian()) {
        MEM_write16(memPtr, val);
    } else {
        BYTE* p = (BYTE*)memPtr;
        p[0] = (BYTE)val;
        p[1] = (BYTE)(val>>8);
    }
}

MEM_STATIC U32 MEM_readLE24(const void* memPtr)
{
    return (U32)MEM_readLE16(memPtr) + ((U32)(((const BYTE*)memPtr)[2]) << 16);
}

MEM_STATIC void MEM_writeLE24(void* memPtr, U32 val)
{
    MEM_writeLE16(memPtr, (U16)val);
    ((BYTE*)memPtr)[2] = (BYTE)(val>>16);
}

MEM_STATIC U32 MEM_readLE32(const void* memPtr)
{
    if (MEM_isLittleEndian())
        return MEM_read32(memPtr);
    else
        return MEM_swap32(MEM_read32(memPtr));
}

MEM_STATIC void MEM_writeLE32(void* memPtr, U32 val32)
{
    if (MEM_isLittleEndian())
        MEM_write32(memPtr, val32);
    else
        MEM_write32(memPtr, MEM_swap32(val32));
}

MEM_STATIC U64 MEM_readLE64(const void* memPtr)
{
    if (MEM_isLittleEndian())
        return MEM_read64(memPtr);
    else
        return MEM_swap64(MEM_read64(memPtr));
}

MEM_STATIC void MEM_writeLE64(void* memPtr, U64 val64)
{
    if (MEM_isLittleEndian())
        MEM_write64(memPtr, val64);
    else
        MEM_write64(memPtr, MEM_swap64(val64));
}

MEM_STATIC size_t MEM_readLEST(const void* memPtr)
{
    if (MEM_32bits())
        return (size_t)MEM_readLE32(memPtr);
    else
        return (size_t)MEM_readLE64(memPtr);
}

MEM_STATIC void MEM_writeLEST(void* memPtr, size_t val)
{
    if (MEM_32bits())
        MEM_writeLE32(memPtr, (U32)val);
    else
        MEM_writeLE64(memPtr, (U64)val);
}

/*=== Big endian r/w ===*/

MEM_STATIC U32 MEM_readBE32(const void* memPtr)
{
    if (MEM_isLittleEndian())
        return MEM_swap32(MEM_read32(memPtr));
    else
        return MEM_read32(memPtr);
}

MEM_STATIC void MEM_writeBE32(void* memPtr, U32 val32)
{
    if (MEM_isLittleEndian())
        MEM_write32(memPtr, MEM_swap32(val32));
    else
        MEM_write32(memPtr, val32);
}

MEM_STATIC U64 MEM_readBE64(const void* memPtr)
{
    if (MEM_isLittleEndian())
        return MEM_swap64(MEM_read64(memPtr));
    else
        return MEM_read64(memPtr);
}

MEM_STATIC void MEM_writeBE64(void* memPtr, U64 val64)
{
    if (MEM_isLittleEndian())
        MEM_write64(memPtr, MEM_swap64(val64));
    else
        MEM_write64(memPtr, val64);
}

MEM_STATIC size_t MEM_readBEST(const void* memPtr)
{
    if (MEM_32bits())
        return (size_t)MEM_readBE32(memPtr);
    else
        return (size_t)MEM_readBE64(memPtr);
}

MEM_STATIC void MEM_writeBEST(void* memPtr, size_t val)
{
    if (MEM_32bits())
        MEM_writeBE32(memPtr, (U32)val);
    else
        MEM_writeBE64(memPtr, (U64)val);
}

/* code only tested on 32 and 64 bits systems */
MEM_STATIC void MEM_check(void) { DEBUG_STATIC_ASSERT((sizeof(size_t)==4) || (sizeof(size_t)==8)); }

#endif /* MEM_H_MODULE */
/**** ended inlining mem.h ****/
/**** start inlining error_private.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* Note : this module is expected to remain private, do not expose it */

#ifndef ERROR_H_MODULE
#define ERROR_H_MODULE

/* ****************************************
*  Dependencies
******************************************/
/**** start inlining ../zstd_errors.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_ERRORS_H_398273423
#define ZSTD_ERRORS_H_398273423

#if defined (__cplusplus)
extern "C" {
#endif

/* =====   ZSTDERRORLIB_API : control library symbols visibility   ===== */
#ifndef ZSTDERRORLIB_VISIBLE
   /* Backwards compatibility with old macro name */
#  ifdef ZSTDERRORLIB_VISIBILITY
#    define ZSTDERRORLIB_VISIBLE ZSTDERRORLIB_VISIBILITY
#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
#    define ZSTDERRORLIB_VISIBLE __attribute__ ((visibility ("default")))
#  else
#    define ZSTDERRORLIB_VISIBLE
#  endif
#endif

#ifndef ZSTDERRORLIB_HIDDEN
#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
#    define ZSTDERRORLIB_HIDDEN __attribute__ ((visibility ("hidden")))
#  else
#    define ZSTDERRORLIB_HIDDEN
#  endif
#endif

#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
#  define ZSTDERRORLIB_API __declspec(dllexport) ZSTDERRORLIB_VISIBLE
#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
#  define ZSTDERRORLIB_API __declspec(dllimport) ZSTDERRORLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
#else
#  define ZSTDERRORLIB_API ZSTDERRORLIB_VISIBLE
#endif

/*-*********************************************
 *  Error codes list
 *-*********************************************
 *  Error codes _values_ are pinned down since v1.3.1 only.
 *  Therefore, don't rely on values if you may link to any version < v1.3.1.
 *
 *  Only values < 100 are considered stable.
 *
 *  note 1 : this API shall be used with static linking only.
 *           dynamic linking is not yet officially supported.
 *  note 2 : Prefer relying on the enum than on its value whenever possible
 *           This is the only supported way to use the error list < v1.3.1
 *  note 3 : ZSTD_isError() is always correct, whatever the library version.
 **********************************************/
typedef enum {
  ZSTD_error_no_error = 0,
  ZSTD_error_GENERIC  = 1,
  ZSTD_error_prefix_unknown                = 10,
  ZSTD_error_version_unsupported           = 12,
  ZSTD_error_frameParameter_unsupported    = 14,
  ZSTD_error_frameParameter_windowTooLarge = 16,
  ZSTD_error_corruption_detected = 20,
  ZSTD_error_checksum_wrong      = 22,
  ZSTD_error_literals_headerWrong = 24,
  ZSTD_error_dictionary_corrupted      = 30,
  ZSTD_error_dictionary_wrong          = 32,
  ZSTD_error_dictionaryCreation_failed = 34,
  ZSTD_error_parameter_unsupported   = 40,
  ZSTD_error_parameter_combination_unsupported = 41,
  ZSTD_error_parameter_outOfBound    = 42,
  ZSTD_error_tableLog_tooLarge       = 44,
  ZSTD_error_maxSymbolValue_tooLarge = 46,
  ZSTD_error_maxSymbolValue_tooSmall = 48,
  ZSTD_error_cannotProduce_uncompressedBlock = 49,
  ZSTD_error_stabilityCondition_notRespected = 50,
  ZSTD_error_stage_wrong       = 60,
  ZSTD_error_init_missing      = 62,
  ZSTD_error_memory_allocation = 64,
  ZSTD_error_workSpace_tooSmall= 66,
  ZSTD_error_dstSize_tooSmall = 70,
  ZSTD_error_srcSize_wrong    = 72,
  ZSTD_error_dstBuffer_null   = 74,
  ZSTD_error_noForwardProgress_destFull = 80,
  ZSTD_error_noForwardProgress_inputEmpty = 82,
  /* following error codes are __NOT STABLE__, they can be removed or changed in future versions */
  ZSTD_error_frameIndex_tooLarge = 100,
  ZSTD_error_seekableIO          = 102,
  ZSTD_error_dstBuffer_wrong     = 104,
  ZSTD_error_srcBuffer_wrong     = 105,
  ZSTD_error_sequenceProducer_failed = 106,
  ZSTD_error_externalSequences_invalid = 107,
  ZSTD_error_maxCode = 120  /* never EVER use this value directly, it can change in future versions! Use ZSTD_isError() instead */
} ZSTD_ErrorCode;

ZSTDERRORLIB_API const char* ZSTD_getErrorString(ZSTD_ErrorCode code);   /**< Same as ZSTD_getErrorName, but using a `ZSTD_ErrorCode` enum argument */


#if defined (__cplusplus)
}
#endif

#endif /* ZSTD_ERRORS_H_398273423 */
/**** ended inlining ../zstd_errors.h ****/
/**** skipping file: compiler.h ****/
/**** skipping file: debug.h ****/
/**** skipping file: zstd_deps.h ****/

/* ****************************************
*  Compiler-specific
******************************************/
#if defined(__GNUC__)
#  define ERR_STATIC static __attribute__((unused))
#elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
#  define ERR_STATIC static inline
#elif defined(_MSC_VER)
#  define ERR_STATIC static __inline
#else
#  define ERR_STATIC static  /* this version may generate warnings for unused static functions; disable the relevant warning */
#endif


/*-****************************************
*  Customization (error_public.h)
******************************************/
typedef ZSTD_ErrorCode ERR_enum;
#define PREFIX(name) ZSTD_error_##name


/*-****************************************
*  Error codes handling
******************************************/
#undef ERROR   /* already defined on Visual Studio */
#define ERROR(name) ZSTD_ERROR(name)
#define ZSTD_ERROR(name) ((size_t)-PREFIX(name))

ERR_STATIC unsigned ERR_isError(size_t code) { return (code > ERROR(maxCode)); }

ERR_STATIC ERR_enum ERR_getErrorCode(size_t code) { if (!ERR_isError(code)) return (ERR_enum)0; return (ERR_enum) (0-code); }

/* check and forward error code */
#define CHECK_V_F(e, f)     \
    size_t const e = f;     \
    do {                    \
        if (ERR_isError(e)) \
            return e;       \
    } while (0)
#define CHECK_F(f)   do { CHECK_V_F(_var_err__, f); } while (0)


/*-****************************************
*  Error Strings
******************************************/

const char* ERR_getErrorString(ERR_enum code);   /* error_private.c */

ERR_STATIC const char* ERR_getErrorName(size_t code)
{
    return ERR_getErrorString(ERR_getErrorCode(code));
}

/**
 * Ignore: this is an internal helper.
 *
 * This is a helper function to help force C99-correctness during compilation.
 * Under strict compilation modes, variadic macro arguments can't be empty.
 * However, variadic function arguments can be. Using a function therefore lets
 * us statically check that at least one (string) argument was passed,
 * independent of the compilation flags.
 */
static INLINE_KEYWORD UNUSED_ATTR
void _force_has_format_string(const char *format, ...) {
  (void)format;
}

/**
 * Ignore: this is an internal helper.
 *
 * We want to force this function invocation to be syntactically correct, but
 * we don't want to force runtime evaluation of its arguments.
 */
#define _FORCE_HAS_FORMAT_STRING(...)              \
    do {                                           \
        if (0) {                                   \
            _force_has_format_string(__VA_ARGS__); \
        }                                          \
    } while (0)

#define ERR_QUOTE(str) #str

/**
 * Return the specified error if the condition evaluates to true.
 *
 * In debug modes, prints additional information.
 * In order to do that (particularly, printing the conditional that failed),
 * this can't just wrap RETURN_ERROR().
 */
#define RETURN_ERROR_IF(cond, err, ...)                                        \
    do {                                                                       \
        if (cond) {                                                            \
            RAWLOG(3, "%s:%d: ERROR!: check %s failed, returning %s",          \
                  __FILE__, __LINE__, ERR_QUOTE(cond), ERR_QUOTE(ERROR(err))); \
            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                             \
            RAWLOG(3, ": " __VA_ARGS__);                                       \
            RAWLOG(3, "\n");                                                   \
            return ERROR(err);                                                 \
        }                                                                      \
    } while (0)

/**
 * Unconditionally return the specified error.
 *
 * In debug modes, prints additional information.
 */
#define RETURN_ERROR(err, ...)                                               \
    do {                                                                     \
        RAWLOG(3, "%s:%d: ERROR!: unconditional check failed, returning %s", \
              __FILE__, __LINE__, ERR_QUOTE(ERROR(err)));                    \
        _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                               \
        RAWLOG(3, ": " __VA_ARGS__);                                         \
        RAWLOG(3, "\n");                                                     \
        return ERROR(err);                                                   \
    } while(0)

/**
 * If the provided expression evaluates to an error code, returns that error code.
 *
 * In debug modes, prints additional information.
 */
#define FORWARD_IF_ERROR(err, ...)                                                 \
    do {                                                                           \
        size_t const err_code = (err);                                             \
        if (ERR_isError(err_code)) {                                               \
            RAWLOG(3, "%s:%d: ERROR!: forwarding error in %s: %s",                 \
                  __FILE__, __LINE__, ERR_QUOTE(err), ERR_getErrorName(err_code)); \
            _FORCE_HAS_FORMAT_STRING(__VA_ARGS__);                                 \
            RAWLOG(3, ": " __VA_ARGS__);                                           \
            RAWLOG(3, "\n");                                                       \
            return err_code;                                                       \
        }                                                                          \
    } while(0)

#endif /* ERROR_H_MODULE */
/**** ended inlining error_private.h ****/
#define FSE_STATIC_LINKING_ONLY  /* FSE_MIN_TABLELOG */
/**** start inlining fse.h ****/
/* ******************************************************************
 * FSE : Finite State Entropy codec
 * Public Prototypes declaration
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 * You can contact the author at :
 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */
#ifndef FSE_H
#define FSE_H


/*-*****************************************
*  Dependencies
******************************************/
/**** skipping file: zstd_deps.h ****/

/*-*****************************************
*  FSE_PUBLIC_API : control library symbols visibility
******************************************/
#if defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1) && defined(__GNUC__) && (__GNUC__ >= 4)
#  define FSE_PUBLIC_API __attribute__ ((visibility ("default")))
#elif defined(FSE_DLL_EXPORT) && (FSE_DLL_EXPORT==1)   /* Visual expected */
#  define FSE_PUBLIC_API __declspec(dllexport)
#elif defined(FSE_DLL_IMPORT) && (FSE_DLL_IMPORT==1)
#  define FSE_PUBLIC_API __declspec(dllimport) /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
#else
#  define FSE_PUBLIC_API
#endif

/*------   Version   ------*/
#define FSE_VERSION_MAJOR    0
#define FSE_VERSION_MINOR    9
#define FSE_VERSION_RELEASE  0

#define FSE_LIB_VERSION FSE_VERSION_MAJOR.FSE_VERSION_MINOR.FSE_VERSION_RELEASE
#define FSE_QUOTE(str) #str
#define FSE_EXPAND_AND_QUOTE(str) FSE_QUOTE(str)
#define FSE_VERSION_STRING FSE_EXPAND_AND_QUOTE(FSE_LIB_VERSION)

#define FSE_VERSION_NUMBER  (FSE_VERSION_MAJOR *100*100 + FSE_VERSION_MINOR *100 + FSE_VERSION_RELEASE)
FSE_PUBLIC_API unsigned FSE_versionNumber(void);   /**< library version number; to be used when checking dll version */


/*-*****************************************
*  Tool functions
******************************************/
FSE_PUBLIC_API size_t FSE_compressBound(size_t size);       /* maximum compressed size */

/* Error Management */
FSE_PUBLIC_API unsigned    FSE_isError(size_t code);        /* tells if a return value is an error code */
FSE_PUBLIC_API const char* FSE_getErrorName(size_t code);   /* provides error code string (useful for debugging) */


/*-*****************************************
*  FSE detailed API
******************************************/
/*!
FSE_compress() does the following:
1. count symbol occurrence from source[] into table count[] (see hist.h)
2. normalize counters so that sum(count[]) == Power_of_2 (2^tableLog)
3. save normalized counters to memory buffer using writeNCount()
4. build encoding table 'CTable' from normalized counters
5. encode the data stream using encoding table 'CTable'

FSE_decompress() does the following:
1. read normalized counters with readNCount()
2. build decoding table 'DTable' from normalized counters
3. decode the data stream using decoding table 'DTable'

The following API allows targeting specific sub-functions for advanced tasks.
For example, it's possible to compress several blocks using the same 'CTable',
or to save and provide normalized distribution using external method.
*/

/* *** COMPRESSION *** */

/*! FSE_optimalTableLog():
    dynamically downsize 'tableLog' when conditions are met.
    It saves CPU time, by using smaller tables, while preserving or even improving compression ratio.
    @return : recommended tableLog (necessarily <= 'maxTableLog') */
FSE_PUBLIC_API unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue);

/*! FSE_normalizeCount():
    normalize counts so that sum(count[]) == Power_of_2 (2^tableLog)
    'normalizedCounter' is a table of short, of minimum size (maxSymbolValue+1).
    useLowProbCount is a boolean parameter which trades off compressed size for
    faster header decoding. When it is set to 1, the compressed data will be slightly
    smaller. And when it is set to 0, FSE_readNCount() and FSE_buildDTable() will be
    faster. If you are compressing a small amount of data (< 2 KB) then useLowProbCount=0
    is a good default, since header deserialization makes a big speed difference.
    Otherwise, useLowProbCount=1 is a good default, since the speed difference is small.
    @return : tableLog,
              or an errorCode, which can be tested using FSE_isError() */
FSE_PUBLIC_API size_t FSE_normalizeCount(short* normalizedCounter, unsigned tableLog,
                    const unsigned* count, size_t srcSize, unsigned maxSymbolValue, unsigned useLowProbCount);

/*! FSE_NCountWriteBound():
    Provides the maximum possible size of an FSE normalized table, given 'maxSymbolValue' and 'tableLog'.
    Typically useful for allocation purpose. */
FSE_PUBLIC_API size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog);

/*! FSE_writeNCount():
    Compactly save 'normalizedCounter' into 'buffer'.
    @return : size of the compressed table,
              or an errorCode, which can be tested using FSE_isError(). */
FSE_PUBLIC_API size_t FSE_writeNCount (void* buffer, size_t bufferSize,
                                 const short* normalizedCounter,
                                 unsigned maxSymbolValue, unsigned tableLog);

/*! Constructor and Destructor of FSE_CTable.
    Note that FSE_CTable size depends on 'tableLog' and 'maxSymbolValue' */
typedef unsigned FSE_CTable;   /* don't allocate that. It's only meant to be more restrictive than void* */

/*! FSE_buildCTable():
    Builds `ct`, which must be already allocated, using FSE_createCTable().
    @return : 0, or an errorCode, which can be tested using FSE_isError() */
FSE_PUBLIC_API size_t FSE_buildCTable(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog);

/*! FSE_compress_usingCTable():
    Compress `src` using `ct` into `dst` which must be already allocated.
    @return : size of compressed data (<= `dstCapacity`),
              or 0 if compressed data could not fit into `dst`,
              or an errorCode, which can be tested using FSE_isError() */
FSE_PUBLIC_API size_t FSE_compress_usingCTable (void* dst, size_t dstCapacity, const void* src, size_t srcSize, const FSE_CTable* ct);

/*!
Tutorial :
----------
The first step is to count all symbols. FSE_count() does this job very fast.
Result will be saved into 'count', a table of unsigned int, which must be already allocated, and have 'maxSymbolValuePtr[0]+1' cells.
'src' is a table of bytes of size 'srcSize'. All values within 'src' MUST be <= maxSymbolValuePtr[0]
maxSymbolValuePtr[0] will be updated, with its real value (necessarily <= original value)
FSE_count() will return the number of occurrence of the most frequent symbol.
This can be used to know if there is a single symbol within 'src', and to quickly evaluate its compressibility.
If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).

The next step is to normalize the frequencies.
FSE_normalizeCount() will ensure that sum of frequencies is == 2 ^'tableLog'.
It also guarantees a minimum of 1 to any Symbol with frequency >= 1.
You can use 'tableLog'==0 to mean "use default tableLog value".
If you are unsure of which tableLog value to use, you can ask FSE_optimalTableLog(),
which will provide the optimal valid tableLog given sourceSize, maxSymbolValue, and a user-defined maximum (0 means "default").

The result of FSE_normalizeCount() will be saved into a table,
called 'normalizedCounter', which is a table of signed short.
'normalizedCounter' must be already allocated, and have at least 'maxSymbolValue+1' cells.
The return value is tableLog if everything proceeded as expected.
It is 0 if there is a single symbol within distribution.
If there is an error (ex: invalid tableLog value), the function will return an ErrorCode (which can be tested using FSE_isError()).

'normalizedCounter' can be saved in a compact manner to a memory area using FSE_writeNCount().
'buffer' must be already allocated.
For guaranteed success, buffer size must be at least FSE_headerBound().
The result of the function is the number of bytes written into 'buffer'.
If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError(); ex : buffer size too small).

'normalizedCounter' can then be used to create the compression table 'CTable'.
The space required by 'CTable' must be already allocated, using FSE_createCTable().
You can then use FSE_buildCTable() to fill 'CTable'.
If there is an error, both functions will return an ErrorCode (which can be tested using FSE_isError()).

'CTable' can then be used to compress 'src', with FSE_compress_usingCTable().
Similar to FSE_count(), the convention is that 'src' is assumed to be a table of char of size 'srcSize'
The function returns the size of compressed data (without header), necessarily <= `dstCapacity`.
If it returns '0', compressed data could not fit into 'dst'.
If there is an error, the function will return an ErrorCode (which can be tested using FSE_isError()).
*/


/* *** DECOMPRESSION *** */

/*! FSE_readNCount():
    Read compactly saved 'normalizedCounter' from 'rBuffer'.
    @return : size read from 'rBuffer',
              or an errorCode, which can be tested using FSE_isError().
              maxSymbolValuePtr[0] and tableLogPtr[0] will also be updated with their respective values */
FSE_PUBLIC_API size_t FSE_readNCount (short* normalizedCounter,
                           unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                           const void* rBuffer, size_t rBuffSize);

/*! FSE_readNCount_bmi2():
 * Same as FSE_readNCount() but pass bmi2=1 when your CPU supports BMI2 and 0 otherwise.
 */
FSE_PUBLIC_API size_t FSE_readNCount_bmi2(short* normalizedCounter,
                           unsigned* maxSymbolValuePtr, unsigned* tableLogPtr,
                           const void* rBuffer, size_t rBuffSize, int bmi2);

typedef unsigned FSE_DTable;   /* don't allocate that. It's just a way to be more restrictive than void* */

/*!
Tutorial :
----------
(Note : these functions only decompress FSE-compressed blocks.
 If block is uncompressed, use memcpy() instead
 If block is a single repeated byte, use memset() instead )

The first step is to obtain the normalized frequencies of symbols.
This can be performed by FSE_readNCount() if it was saved using FSE_writeNCount().
'normalizedCounter' must be already allocated, and have at least 'maxSymbolValuePtr[0]+1' cells of signed short.
In practice, that means it's necessary to know 'maxSymbolValue' beforehand,
or size the table to handle worst case situations (typically 256).
FSE_readNCount() will provide 'tableLog' and 'maxSymbolValue'.
The result of FSE_readNCount() is the number of bytes read from 'rBuffer'.
Note that 'rBufferSize' must be at least 4 bytes, even if useful information is less than that.
If there is an error, the function will return an error code, which can be tested using FSE_isError().

The next step is to build the decompression tables 'FSE_DTable' from 'normalizedCounter'.
This is performed by the function FSE_buildDTable().
The space required by 'FSE_DTable' must be already allocated using FSE_createDTable().
If there is an error, the function will return an error code, which can be tested using FSE_isError().

`FSE_DTable` can then be used to decompress `cSrc`, with FSE_decompress_usingDTable().
`cSrcSize` must be strictly correct, otherwise decompression will fail.
FSE_decompress_usingDTable() result will tell how many bytes were regenerated (<=`dstCapacity`).
If there is an error, the function will return an error code, which can be tested using FSE_isError(). (ex: dst buffer too small)
*/

#endif  /* FSE_H */


#if defined(FSE_STATIC_LINKING_ONLY) && !defined(FSE_H_FSE_STATIC_LINKING_ONLY)
#define FSE_H_FSE_STATIC_LINKING_ONLY
/**** start inlining bitstream.h ****/
/* ******************************************************************
 * bitstream
 * Part of FSE library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 * You can contact the author at :
 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */
#ifndef BITSTREAM_H_MODULE
#define BITSTREAM_H_MODULE

/*
*  This API consists of small unitary functions, which must be inlined for best performance.
*  Since link-time-optimization is not available for all compilers,
*  these functions are defined into a .h to be included.
*/

/*-****************************************
*  Dependencies
******************************************/
/**** skipping file: mem.h ****/
/**** skipping file: compiler.h ****/
/**** skipping file: debug.h ****/
/**** skipping file: error_private.h ****/
/**** start inlining bits.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_BITS_H
#define ZSTD_BITS_H

/**** skipping file: mem.h ****/

MEM_STATIC unsigned ZSTD_countTrailingZeros32_fallback(U32 val)
{
    assert(val != 0);
    {
        static const U32 DeBruijnBytePos[32] = {0, 1, 28, 2, 29, 14, 24, 3,
                                                30, 22, 20, 15, 25, 17, 4, 8,
                                                31, 27, 13, 23, 21, 19, 16, 7,
                                                26, 12, 18, 6, 11, 5, 10, 9};
        return DeBruijnBytePos[((U32) ((val & -(S32) val) * 0x077CB531U)) >> 27];
    }
}

MEM_STATIC unsigned ZSTD_countTrailingZeros32(U32 val)
{
    assert(val != 0);
#if defined(_MSC_VER)
#  if STATIC_BMI2
    return (unsigned)_tzcnt_u32(val);
#  else
    if (val != 0) {
        unsigned long r;
        _BitScanForward(&r, val);
        return (unsigned)r;
    } else {
        __assume(0); /* Should not reach this code path */
    }
#  endif
#elif defined(__GNUC__) && (__GNUC__ >= 4)
    return (unsigned)__builtin_ctz(val);
#elif defined(__ICCARM__)
    return (unsigned)__builtin_ctz(val);
#else
    return ZSTD_countTrailingZeros32_fallback(val);
#endif
}

MEM_STATIC unsigned ZSTD_countLeadingZeros32_fallback(U32 val)
{
    assert(val != 0);
    {
        static const U32 DeBruijnClz[32] = {0, 9, 1, 10, 13, 21, 2, 29,
                                            11, 14, 16, 18, 22, 25, 3, 30,
                                            8, 12, 20, 28, 15, 17, 24, 7,
                                            19, 27, 23, 6, 26, 5, 4, 31};
        val |= val >> 1;
        val |= val >> 2;
        val |= val >> 4;
        val |= val >> 8;
        val |= val >> 16;
        return 31 - DeBruijnClz[(val * 0x07C4ACDDU) >> 27];
    }
}

MEM_STATIC unsigned ZSTD_countLeadingZeros32(U32 val)
{
    assert(val != 0);
#if defined(_MSC_VER)
#  if STATIC_BMI2
    return (unsigned)_lzcnt_u32(val);
#  else
    if (val != 0) {
        unsigned long r;
        _BitScanReverse(&r, val);
        return (unsigned)(31 - r);
    } else {
        __assume(0); /* Should not reach this code path */
    }
#  endif
#elif defined(__GNUC__) && (__GNUC__ >= 4)
    return (unsigned)__builtin_clz(val);
#elif defined(__ICCARM__)
    return (unsigned)__builtin_clz(val);
#else
    return ZSTD_countLeadingZeros32_fallback(val);
#endif
}

MEM_STATIC unsigned ZSTD_countTrailingZeros64(U64 val)
{
    assert(val != 0);
#if defined(_MSC_VER) && defined(_WIN64)
#  if STATIC_BMI2
    return (unsigned)_tzcnt_u64(val);
#  else
    if (val != 0) {
        unsigned long r;
        _BitScanForward64(&r, val);
        return (unsigned)r;
    } else {
        __assume(0); /* Should not reach this code path */
    }
#  endif
#elif defined(__GNUC__) && (__GNUC__ >= 4) && defined(__LP64__)
    return (unsigned)__builtin_ctzll(val);
#elif defined(__ICCARM__)
    return (unsigned)__builtin_ctzll(val);
#else
    {
        U32 mostSignificantWord = (U32)(val >> 32);
        U32 leastSignificantWord = (U32)val;
        if (leastSignificantWord == 0) {
            return 32 + ZSTD_countTrailingZeros32(mostSignificantWord);
        } else {
            return ZSTD_countTrailingZeros32(leastSignificantWord);
        }
    }
#endif
}

MEM_STATIC unsigned ZSTD_countLeadingZeros64(U64 val)
{
    assert(val != 0);
#if defined(_MSC_VER) && defined(_WIN64)
#  if STATIC_BMI2
    return (unsigned)_lzcnt_u64(val);
#  else
    if (val != 0) {
        unsigned long r;
        _BitScanReverse64(&r, val);
        return (unsigned)(63 - r);
    } else {
        __assume(0); /* Should not reach this code path */
    }
#  endif
#elif defined(__GNUC__) && (__GNUC__ >= 4)
    return (unsigned)(__builtin_clzll(val));
#elif defined(__ICCARM__)
    return (unsigned)(__builtin_clzll(val));
#else
    {
        U32 mostSignificantWord = (U32)(val >> 32);
        U32 leastSignificantWord = (U32)val;
        if (mostSignificantWord == 0) {
            return 32 + ZSTD_countLeadingZeros32(leastSignificantWord);
        } else {
            return ZSTD_countLeadingZeros32(mostSignificantWord);
        }
    }
#endif
}

MEM_STATIC unsigned ZSTD_NbCommonBytes(size_t val)
{
    if (MEM_isLittleEndian()) {
        if (MEM_64bits()) {
            return ZSTD_countTrailingZeros64((U64)val) >> 3;
        } else {
            return ZSTD_countTrailingZeros32((U32)val) >> 3;
        }
    } else {  /* Big Endian CPU */
        if (MEM_64bits()) {
            return ZSTD_countLeadingZeros64((U64)val) >> 3;
        } else {
            return ZSTD_countLeadingZeros32((U32)val) >> 3;
        }
    }
}

MEM_STATIC unsigned ZSTD_highbit32(U32 val)   /* compress, dictBuilder, decodeCorpus */
{
    assert(val != 0);
    return 31 - ZSTD_countLeadingZeros32(val);
}

/* ZSTD_rotateRight_*():
 * Rotates a bitfield to the right by "count" bits.
 * https://en.wikipedia.org/w/index.php?title=Circular_shift&oldid=991635599#Implementing_circular_shifts
 */
MEM_STATIC
U64 ZSTD_rotateRight_U64(U64 const value, U32 count) {
    assert(count < 64);
    count &= 0x3F; /* for fickle pattern recognition */
    return (value >> count) | (U64)(value << ((0U - count) & 0x3F));
}

MEM_STATIC
U32 ZSTD_rotateRight_U32(U32 const value, U32 count) {
    assert(count < 32);
    count &= 0x1F; /* for fickle pattern recognition */
    return (value >> count) | (U32)(value << ((0U - count) & 0x1F));
}

MEM_STATIC
U16 ZSTD_rotateRight_U16(U16 const value, U32 count) {
    assert(count < 16);
    count &= 0x0F; /* for fickle pattern recognition */
    return (value >> count) | (U16)(value << ((0U - count) & 0x0F));
}

#endif /* ZSTD_BITS_H */
/**** ended inlining bits.h ****/

/*=========================================
*  Target specific
=========================================*/
#ifndef ZSTD_NO_INTRINSICS
#  if (defined(__BMI__) || defined(__BMI2__)) && defined(__GNUC__)
#    include <immintrin.h>   /* support for bextr (experimental)/bzhi */
#  elif defined(__ICCARM__)
#    include <intrinsics.h>
#  endif
#endif

#define STREAM_ACCUMULATOR_MIN_32  25
#define STREAM_ACCUMULATOR_MIN_64  57
#define STREAM_ACCUMULATOR_MIN    ((U32)(MEM_32bits() ? STREAM_ACCUMULATOR_MIN_32 : STREAM_ACCUMULATOR_MIN_64))


/*-******************************************
*  bitStream encoding API (write forward)
********************************************/
typedef size_t BitContainerType;
/* bitStream can mix input from multiple sources.
 * A critical property of these streams is that they encode and decode in **reverse** direction.
 * So the first bit sequence you add will be the last to be read, like a LIFO stack.
 */
typedef struct {
    BitContainerType bitContainer;
    unsigned bitPos;
    char*  startPtr;
    char*  ptr;
    char*  endPtr;
} BIT_CStream_t;

MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC, void* dstBuffer, size_t dstCapacity);
MEM_STATIC void   BIT_addBits(BIT_CStream_t* bitC, BitContainerType value, unsigned nbBits);
MEM_STATIC void   BIT_flushBits(BIT_CStream_t* bitC);
MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC);

/* Start with initCStream, providing the size of buffer to write into.
*  bitStream will never write outside of this buffer.
*  `dstCapacity` must be >= sizeof(bitD->bitContainer), otherwise @return will be an error code.
*
*  bits are first added to a local register.
*  Local register is BitContainerType, 64-bits on 64-bits systems, or 32-bits on 32-bits systems.
*  Writing data into memory is an explicit operation, performed by the flushBits function.
*  Hence keep track how many bits are potentially stored into local register to avoid register overflow.
*  After a flushBits, a maximum of 7 bits might still be stored into local register.
*
*  Avoid storing elements of more than 24 bits if you want compatibility with 32-bits bitstream readers.
*
*  Last operation is to close the bitStream.
*  The function returns the final size of CStream in bytes.
*  If data couldn't fit into `dstBuffer`, it will return a 0 ( == not storable)
*/


/*-********************************************
*  bitStream decoding API (read backward)
**********************************************/
typedef struct {
    BitContainerType bitContainer;
    unsigned bitsConsumed;
    const char* ptr;
    const char* start;
    const char* limitPtr;
} BIT_DStream_t;

typedef enum { BIT_DStream_unfinished = 0,  /* fully refilled */
               BIT_DStream_endOfBuffer = 1, /* still some bits left in bitstream */
               BIT_DStream_completed = 2,   /* bitstream entirely consumed, bit-exact */
               BIT_DStream_overflow = 3     /* user requested more bits than present in bitstream */
    } BIT_DStream_status;  /* result of BIT_reloadDStream() */

MEM_STATIC size_t   BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize);
MEM_STATIC BitContainerType BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits);
MEM_STATIC BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD);
MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* bitD);


/* Start by invoking BIT_initDStream().
*  A chunk of the bitStream is then stored into a local register.
*  Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (BitContainerType).
*  You can then retrieve bitFields stored into the local register, **in reverse order**.
*  Local register is explicitly reloaded from memory by the BIT_reloadDStream() method.
*  A reload guarantee a minimum of ((8*sizeof(bitD->bitContainer))-7) bits when its result is BIT_DStream_unfinished.
*  Otherwise, it can be less than that, so proceed accordingly.
*  Checking if DStream has reached its end can be performed with BIT_endOfDStream().
*/


/*-****************************************
*  unsafe API
******************************************/
MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC, BitContainerType value, unsigned nbBits);
/* faster, but works only if value is "clean", meaning all high bits above nbBits are 0 */

MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC);
/* unsafe version; does not check buffer overflow */

MEM_STATIC size_t BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits);
/* faster, but works only if nbBits >= 1 */

/*=====    Local Constants   =====*/
static const unsigned BIT_mask[] = {
    0,          1,         3,         7,         0xF,       0x1F,
    0x3F,       0x7F,      0xFF,      0x1FF,     0x3FF,     0x7FF,
    0xFFF,      0x1FFF,    0x3FFF,    0x7FFF,    0xFFFF,    0x1FFFF,
    0x3FFFF,    0x7FFFF,   0xFFFFF,   0x1FFFFF,  0x3FFFFF,  0x7FFFFF,
    0xFFFFFF,   0x1FFFFFF, 0x3FFFFFF, 0x7FFFFFF, 0xFFFFFFF, 0x1FFFFFFF,
    0x3FFFFFFF, 0x7FFFFFFF}; /* up to 31 bits */
#define BIT_MASK_SIZE (sizeof(BIT_mask) / sizeof(BIT_mask[0]))

/*-**************************************************************
*  bitStream encoding
****************************************************************/
/*! BIT_initCStream() :
 *  `dstCapacity` must be > sizeof(size_t)
 *  @return : 0 if success,
 *            otherwise an error code (can be tested using ERR_isError()) */
MEM_STATIC size_t BIT_initCStream(BIT_CStream_t* bitC,
                                  void* startPtr, size_t dstCapacity)
{
    bitC->bitContainer = 0;
    bitC->bitPos = 0;
    bitC->startPtr = (char*)startPtr;
    bitC->ptr = bitC->startPtr;
    bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer);
    if (dstCapacity <= sizeof(bitC->bitContainer)) return ERROR(dstSize_tooSmall);
    return 0;
}

FORCE_INLINE_TEMPLATE BitContainerType BIT_getLowerBits(BitContainerType bitContainer, U32 const nbBits)
{
#if STATIC_BMI2 && !defined(ZSTD_NO_INTRINSICS)
#  if (defined(__x86_64__) || defined(_M_X64)) && !defined(__ILP32__)
    return _bzhi_u64(bitContainer, nbBits);
#  else
    DEBUG_STATIC_ASSERT(sizeof(bitContainer) == sizeof(U32));
    return _bzhi_u32(bitContainer, nbBits);
#  endif
#else
    assert(nbBits < BIT_MASK_SIZE);
    return bitContainer & BIT_mask[nbBits];
#endif
}

/*! BIT_addBits() :
 *  can add up to 31 bits into `bitC`.
 *  Note : does not check for register overflow ! */
MEM_STATIC void BIT_addBits(BIT_CStream_t* bitC,
                            BitContainerType value, unsigned nbBits)
{
    DEBUG_STATIC_ASSERT(BIT_MASK_SIZE == 32);
    assert(nbBits < BIT_MASK_SIZE);
    assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
    bitC->bitContainer |= BIT_getLowerBits(value, nbBits) << bitC->bitPos;
    bitC->bitPos += nbBits;
}

/*! BIT_addBitsFast() :
 *  works only if `value` is _clean_,
 *  meaning all high bits above nbBits are 0 */
MEM_STATIC void BIT_addBitsFast(BIT_CStream_t* bitC,
                                BitContainerType value, unsigned nbBits)
{
    assert((value>>nbBits) == 0);
    assert(nbBits + bitC->bitPos < sizeof(bitC->bitContainer) * 8);
    bitC->bitContainer |= value << bitC->bitPos;
    bitC->bitPos += nbBits;
}

/*! BIT_flushBitsFast() :
 *  assumption : bitContainer has not overflowed
 *  unsafe version; does not check buffer overflow */
MEM_STATIC void BIT_flushBitsFast(BIT_CStream_t* bitC)
{
    size_t const nbBytes = bitC->bitPos >> 3;
    assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
    assert(bitC->ptr <= bitC->endPtr);
    MEM_writeLEST(bitC->ptr, bitC->bitContainer);
    bitC->ptr += nbBytes;
    bitC->bitPos &= 7;
    bitC->bitContainer >>= nbBytes*8;
}

/*! BIT_flushBits() :
 *  assumption : bitContainer has not overflowed
 *  safe version; check for buffer overflow, and prevents it.
 *  note : does not signal buffer overflow.
 *  overflow will be revealed later on using BIT_closeCStream() */
MEM_STATIC void BIT_flushBits(BIT_CStream_t* bitC)
{
    size_t const nbBytes = bitC->bitPos >> 3;
    assert(bitC->bitPos < sizeof(bitC->bitContainer) * 8);
    assert(bitC->ptr <= bitC->endPtr);
    MEM_writeLEST(bitC->ptr, bitC->bitContainer);
    bitC->ptr += nbBytes;
    if (bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
    bitC->bitPos &= 7;
    bitC->bitContainer >>= nbBytes*8;
}

/*! BIT_closeCStream() :
 *  @return : size of CStream, in bytes,
 *            or 0 if it could not fit into dstBuffer */
MEM_STATIC size_t BIT_closeCStream(BIT_CStream_t* bitC)
{
    BIT_addBitsFast(bitC, 1, 1);   /* endMark */
    BIT_flushBits(bitC);
    if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
    return (size_t)(bitC->ptr - bitC->startPtr) + (bitC->bitPos > 0);
}


/*-********************************************************
*  bitStream decoding
**********************************************************/
/*! BIT_initDStream() :
 *  Initialize a BIT_DStream_t.
 * `bitD` : a pointer to an already allocated BIT_DStream_t structure.
 * `srcSize` must be the *exact* size of the bitStream, in bytes.
 * @return : size of stream (== srcSize), or an errorCode if a problem is detected
 */
MEM_STATIC size_t BIT_initDStream(BIT_DStream_t* bitD, const void* srcBuffer, size_t srcSize)
{
    if (srcSize < 1) { ZSTD_memset(bitD, 0, sizeof(*bitD)); return ERROR(srcSize_wrong); }

    bitD->start = (const char*)srcBuffer;
    bitD->limitPtr = bitD->start + sizeof(bitD->bitContainer);

    if (srcSize >=  sizeof(bitD->bitContainer)) {  /* normal case */
        bitD->ptr   = (const char*)srcBuffer + srcSize - sizeof(bitD->bitContainer);
        bitD->bitContainer = MEM_readLEST(bitD->ptr);
        { BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
          bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;  /* ensures bitsConsumed is always set */
          if (lastByte == 0) return ERROR(GENERIC); /* endMark not present */ }
    } else {
        bitD->ptr   = bitD->start;
        bitD->bitContainer = *(const BYTE*)(bitD->start);
        switch(srcSize)
        {
        case 7: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[6]) << (sizeof(bitD->bitContainer)*8 - 16);
                ZSTD_FALLTHROUGH;

        case 6: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[5]) << (sizeof(bitD->bitContainer)*8 - 24);
                ZSTD_FALLTHROUGH;

        case 5: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[4]) << (sizeof(bitD->bitContainer)*8 - 32);
                ZSTD_FALLTHROUGH;

        case 4: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[3]) << 24;
                ZSTD_FALLTHROUGH;

        case 3: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[2]) << 16;
                ZSTD_FALLTHROUGH;

        case 2: bitD->bitContainer += (BitContainerType)(((const BYTE*)(srcBuffer))[1]) <<  8;
                ZSTD_FALLTHROUGH;

        default: break;
        }
        {   BYTE const lastByte = ((const BYTE*)srcBuffer)[srcSize-1];
            bitD->bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
            if (lastByte == 0) return ERROR(corruption_detected);  /* endMark not present */
        }
        bitD->bitsConsumed += (U32)(sizeof(bitD->bitContainer) - srcSize)*8;
    }

    return srcSize;
}

FORCE_INLINE_TEMPLATE BitContainerType BIT_getUpperBits(BitContainerType bitContainer, U32 const start)
{
    return bitContainer >> start;
}

FORCE_INLINE_TEMPLATE BitContainerType BIT_getMiddleBits(BitContainerType bitContainer, U32 const start, U32 const nbBits)
{
    U32 const regMask = sizeof(bitContainer)*8 - 1;
    /* if start > regMask, bitstream is corrupted, and result is undefined */
    assert(nbBits < BIT_MASK_SIZE);
    /* x86 transform & ((1 << nbBits) - 1) to bzhi instruction, it is better
     * than accessing memory. When bmi2 instruction is not present, we consider
     * such cpus old (pre-Haswell, 2013) and their performance is not of that
     * importance.
     */
#if defined(__x86_64__) || defined(_M_X64)
    return (bitContainer >> (start & regMask)) & ((((U64)1) << nbBits) - 1);
#else
    return (bitContainer >> (start & regMask)) & BIT_mask[nbBits];
#endif
}

/*! BIT_lookBits() :
 *  Provides next n bits from local register.
 *  local register is not modified.
 *  On 32-bits, maxNbBits==24.
 *  On 64-bits, maxNbBits==56.
 * @return : value extracted */
FORCE_INLINE_TEMPLATE BitContainerType BIT_lookBits(const BIT_DStream_t*  bitD, U32 nbBits)
{
    /* arbitrate between double-shift and shift+mask */
#if 1
    /* if bitD->bitsConsumed + nbBits > sizeof(bitD->bitContainer)*8,
     * bitstream is likely corrupted, and result is undefined */
    return BIT_getMiddleBits(bitD->bitContainer, (sizeof(bitD->bitContainer)*8) - bitD->bitsConsumed - nbBits, nbBits);
#else
    /* this code path is slower on my os-x laptop */
    U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
    return ((bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> 1) >> ((regMask-nbBits) & regMask);
#endif
}

/*! BIT_lookBitsFast() :
 *  unsafe version; only works if nbBits >= 1 */
MEM_STATIC BitContainerType BIT_lookBitsFast(const BIT_DStream_t* bitD, U32 nbBits)
{
    U32 const regMask = sizeof(bitD->bitContainer)*8 - 1;
    assert(nbBits >= 1);
    return (bitD->bitContainer << (bitD->bitsConsumed & regMask)) >> (((regMask+1)-nbBits) & regMask);
}

FORCE_INLINE_TEMPLATE void BIT_skipBits(BIT_DStream_t* bitD, U32 nbBits)
{
    bitD->bitsConsumed += nbBits;
}

/*! BIT_readBits() :
 *  Read (consume) next n bits from local register and update.
 *  Pay attention to not read more than nbBits contained into local register.
 * @return : extracted value. */
FORCE_INLINE_TEMPLATE BitContainerType BIT_readBits(BIT_DStream_t* bitD, unsigned nbBits)
{
    BitContainerType const value = BIT_lookBits(bitD, nbBits);
    BIT_skipBits(bitD, nbBits);
    return value;
}

/*! BIT_readBitsFast() :
 *  unsafe version; only works if nbBits >= 1 */
MEM_STATIC BitContainerType BIT_readBitsFast(BIT_DStream_t* bitD, unsigned nbBits)
{
    BitContainerType const value = BIT_lookBitsFast(bitD, nbBits);
    assert(nbBits >= 1);
    BIT_skipBits(bitD, nbBits);
    return value;
}

/*! BIT_reloadDStream_internal() :
 *  Simple variant of BIT_reloadDStream(), with two conditions:
 *  1. bitstream is valid : bitsConsumed <= sizeof(bitD->bitContainer)*8
 *  2. look window is valid after shifted down : bitD->ptr >= bitD->start
 */
MEM_STATIC BIT_DStream_status BIT_reloadDStream_internal(BIT_DStream_t* bitD)
{
    assert(bitD->bitsConsumed <= sizeof(bitD->bitContainer)*8);
    bitD->ptr -= bitD->bitsConsumed >> 3;
    assert(bitD->ptr >= bitD->start);
    bitD->bitsConsumed &= 7;
    bitD->bitContainer = MEM_readLEST(bitD->ptr);
    return BIT_DStream_unfinished;
}

/*! BIT_reloadDStreamFast() :
 *  Similar to BIT_reloadDStream(), but with two differences:
 *  1. bitsConsumed <= sizeof(bitD->bitContainer)*8 must hold!
 *  2. Returns BIT_DStream_overflow when bitD->ptr < bitD->limitPtr, at this
 *     point you must use BIT_reloadDStream() to reload.
 */
MEM_STATIC BIT_DStream_status BIT_reloadDStreamFast(BIT_DStream_t* bitD)
{
    if (UNLIKELY(bitD->ptr < bitD->limitPtr))
        return BIT_DStream_overflow;
    return BIT_reloadDStream_internal(bitD);
}

/*! BIT_reloadDStream() :
 *  Refill `bitD` from buffer previously set in BIT_initDStream() .
 *  This function is safe, it guarantees it will not never beyond src buffer.
 * @return : status of `BIT_DStream_t` internal register.
 *           when status == BIT_DStream_unfinished, internal register is filled with at least 25 or 57 bits */
FORCE_INLINE_TEMPLATE BIT_DStream_status BIT_reloadDStream(BIT_DStream_t* bitD)
{
    /* note : once in overflow mode, a bitstream remains in this mode until it's reset */
    if (UNLIKELY(bitD->bitsConsumed > (sizeof(bitD->bitContainer)*8))) {
        static const BitContainerType zeroFilled = 0;
        bitD->ptr = (const char*)&zeroFilled; /* aliasing is allowed for char */
        /* overflow detected, erroneous scenario or end of stream: no update */
        return BIT_DStream_overflow;
    }

    assert(bitD->ptr >= bitD->start);

    if (bitD->ptr >= bitD->limitPtr) {
        return BIT_reloadDStream_internal(bitD);
    }
    if (bitD->ptr == bitD->start) {
        /* reached end of bitStream => no update */
        if (bitD->bitsConsumed < sizeof(bitD->bitContainer)*8) return BIT_DStream_endOfBuffer;
        return BIT_DStream_completed;
    }
    /* start < ptr < limitPtr => cautious update */
    {   U32 nbBytes = bitD->bitsConsumed >> 3;
        BIT_DStream_status result = BIT_DStream_unfinished;
        if (bitD->ptr - nbBytes < bitD->start) {
            nbBytes = (U32)(bitD->ptr - bitD->start);  /* ptr > start */
            result = BIT_DStream_endOfBuffer;
        }
        bitD->ptr -= nbBytes;
        bitD->bitsConsumed -= nbBytes*8;
        bitD->bitContainer = MEM_readLEST(bitD->ptr);   /* reminder : srcSize > sizeof(bitD->bitContainer), otherwise bitD->ptr == bitD->start */
        return result;
    }
}

/*! BIT_endOfDStream() :
 * @return : 1 if DStream has _exactly_ reached its end (all bits consumed).
 */
MEM_STATIC unsigned BIT_endOfDStream(const BIT_DStream_t* DStream)
{
    return ((DStream->ptr == DStream->start) && (DStream->bitsConsumed == sizeof(DStream->bitContainer)*8));
}

#endif /* BITSTREAM_H_MODULE */
/**** ended inlining bitstream.h ****/

/* *****************************************
*  Static allocation
*******************************************/
/* FSE buffer bounds */
#define FSE_NCOUNTBOUND 512
#define FSE_BLOCKBOUND(size) ((size) + ((size)>>7) + 4 /* fse states */ + sizeof(size_t) /* bitContainer */)
#define FSE_COMPRESSBOUND(size) (FSE_NCOUNTBOUND + FSE_BLOCKBOUND(size))   /* Macro version, useful for static allocation */

/* It is possible to statically allocate FSE CTable/DTable as a table of FSE_CTable/FSE_DTable using below macros */
#define FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue)   (1 + (1<<((maxTableLog)-1)) + (((maxSymbolValue)+1)*2))
#define FSE_DTABLE_SIZE_U32(maxTableLog)                   (1 + (1<<(maxTableLog)))

/* or use the size to malloc() space directly. Pay attention to alignment restrictions though */
#define FSE_CTABLE_SIZE(maxTableLog, maxSymbolValue)   (FSE_CTABLE_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(FSE_CTable))
#define FSE_DTABLE_SIZE(maxTableLog)                   (FSE_DTABLE_SIZE_U32(maxTableLog) * sizeof(FSE_DTable))


/* *****************************************
 *  FSE advanced API
 ***************************************** */

unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus);
/**< same as FSE_optimalTableLog(), which used `minus==2` */

size_t FSE_buildCTable_rle (FSE_CTable* ct, unsigned char symbolValue);
/**< build a fake FSE_CTable, designed to compress always the same symbolValue */

/* FSE_buildCTable_wksp() :
 * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
 * `wkspSize` must be >= `FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog)` of `unsigned`.
 * See FSE_buildCTable_wksp() for breakdown of workspace usage.
 */
#define FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog) (((maxSymbolValue + 2) + (1ull << (tableLog)))/2 + sizeof(U64)/sizeof(U32) /* additional 8 bytes for potential table overwrite */)
#define FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) (sizeof(unsigned) * FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(maxSymbolValue, tableLog))
size_t FSE_buildCTable_wksp(FSE_CTable* ct, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);

#define FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) (sizeof(short) * (maxSymbolValue + 1) + (1ULL << maxTableLog) + 8)
#define FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) ((FSE_BUILD_DTABLE_WKSP_SIZE(maxTableLog, maxSymbolValue) + sizeof(unsigned) - 1) / sizeof(unsigned))
FSE_PUBLIC_API size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize);
/**< Same as FSE_buildDTable(), using an externally allocated `workspace` produced with `FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxSymbolValue)` */

#define FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) (FSE_DTABLE_SIZE_U32(maxTableLog) + 1 + FSE_BUILD_DTABLE_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) + (FSE_MAX_SYMBOL_VALUE + 1) / 2 + 1)
#define FSE_DECOMPRESS_WKSP_SIZE(maxTableLog, maxSymbolValue) (FSE_DECOMPRESS_WKSP_SIZE_U32(maxTableLog, maxSymbolValue) * sizeof(unsigned))
size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2);
/**< same as FSE_decompress(), using an externally allocated `workSpace` produced with `FSE_DECOMPRESS_WKSP_SIZE_U32(maxLog, maxSymbolValue)`.
 * Set bmi2 to 1 if your CPU supports BMI2 or 0 if it doesn't */

typedef enum {
   FSE_repeat_none,  /**< Cannot use the previous table */
   FSE_repeat_check, /**< Can use the previous table but it must be checked */
   FSE_repeat_valid  /**< Can use the previous table and it is assumed to be valid */
 } FSE_repeat;

/* *****************************************
*  FSE symbol compression API
*******************************************/
/*!
   This API consists of small unitary functions, which highly benefit from being inlined.
   Hence their body are included in next section.
*/
typedef struct {
    ptrdiff_t   value;
    const void* stateTable;
    const void* symbolTT;
    unsigned    stateLog;
} FSE_CState_t;

static void FSE_initCState(FSE_CState_t* CStatePtr, const FSE_CTable* ct);

static void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* CStatePtr, unsigned symbol);

static void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* CStatePtr);

/**<
These functions are inner components of FSE_compress_usingCTable().
They allow the creation of custom streams, mixing multiple tables and bit sources.

A key property to keep in mind is that encoding and decoding are done **in reverse direction**.
So the first symbol you will encode is the last you will decode, like a LIFO stack.

You will need a few variables to track your CStream. They are :

FSE_CTable    ct;         // Provided by FSE_buildCTable()
BIT_CStream_t bitStream;  // bitStream tracking structure
FSE_CState_t  state;      // State tracking structure (can have several)


The first thing to do is to init bitStream and state.
    size_t errorCode = BIT_initCStream(&bitStream, dstBuffer, maxDstSize);
    FSE_initCState(&state, ct);

Note that BIT_initCStream() can produce an error code, so its result should be tested, using FSE_isError();
You can then encode your input data, byte after byte.
FSE_encodeSymbol() outputs a maximum of 'tableLog' bits at a time.
Remember decoding will be done in reverse direction.
    FSE_encodeByte(&bitStream, &state, symbol);

At any time, you can also add any bit sequence.
Note : maximum allowed nbBits is 25, for compatibility with 32-bits decoders
    BIT_addBits(&bitStream, bitField, nbBits);

The above methods don't commit data to memory, they just store it into local register, for speed.
Local register size is 64-bits on 64-bits systems, 32-bits on 32-bits systems (size_t).
Writing data to memory is a manual operation, performed by the flushBits function.
    BIT_flushBits(&bitStream);

Your last FSE encoding operation shall be to flush your last state value(s).
    FSE_flushState(&bitStream, &state);

Finally, you must close the bitStream.
The function returns the size of CStream in bytes.
If data couldn't fit into dstBuffer, it will return a 0 ( == not compressible)
If there is an error, it returns an errorCode (which can be tested using FSE_isError()).
    size_t size = BIT_closeCStream(&bitStream);
*/


/* *****************************************
*  FSE symbol decompression API
*******************************************/
typedef struct {
    size_t      state;
    const void* table;   /* precise table may vary, depending on U16 */
} FSE_DState_t;


static void     FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt);

static unsigned char FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);

static unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr);

/**<
Let's now decompose FSE_decompress_usingDTable() into its unitary components.
You will decode FSE-encoded symbols from the bitStream,
and also any other bitFields you put in, **in reverse order**.

You will need a few variables to track your bitStream. They are :

BIT_DStream_t DStream;    // Stream context
FSE_DState_t  DState;     // State context. Multiple ones are possible
FSE_DTable*   DTablePtr;  // Decoding table, provided by FSE_buildDTable()

The first thing to do is to init the bitStream.
    errorCode = BIT_initDStream(&DStream, srcBuffer, srcSize);

You should then retrieve your initial state(s)
(in reverse flushing order if you have several ones) :
    errorCode = FSE_initDState(&DState, &DStream, DTablePtr);

You can then decode your data, symbol after symbol.
For information the maximum number of bits read by FSE_decodeSymbol() is 'tableLog'.
Keep in mind that symbols are decoded in reverse order, like a LIFO stack (last in, first out).
    unsigned char symbol = FSE_decodeSymbol(&DState, &DStream);

You can retrieve any bitfield you eventually stored into the bitStream (in reverse order)
Note : maximum allowed nbBits is 25, for 32-bits compatibility
    size_t bitField = BIT_readBits(&DStream, nbBits);

All above operations only read from local register (which size depends on size_t).
Refueling the register from memory is manually performed by the reload method.
    endSignal = FSE_reloadDStream(&DStream);

BIT_reloadDStream() result tells if there is still some more data to read from DStream.
BIT_DStream_unfinished : there is still some data left into the DStream.
BIT_DStream_endOfBuffer : Dstream reached end of buffer. Its container may no longer be completely filled.
BIT_DStream_completed : Dstream reached its exact end, corresponding in general to decompression completed.
BIT_DStream_tooFar : Dstream went too far. Decompression result is corrupted.

When reaching end of buffer (BIT_DStream_endOfBuffer), progress slowly, notably if you decode multiple symbols per loop,
to properly detect the exact end of stream.
After each decoded symbol, check if DStream is fully consumed using this simple test :
    BIT_reloadDStream(&DStream) >= BIT_DStream_completed

When it's done, verify decompression is fully completed, by checking both DStream and the relevant states.
Checking if DStream has reached its end is performed by :
    BIT_endOfDStream(&DStream);
Check also the states. There might be some symbols left there, if some high probability ones (>50%) are possible.
    FSE_endOfDState(&DState);
*/


/* *****************************************
*  FSE unsafe API
*******************************************/
static unsigned char FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD);
/* faster, but works only if nbBits is always >= 1 (otherwise, result will be corrupted) */


/* *****************************************
*  Implementation of inlined functions
*******************************************/
typedef struct {
    int deltaFindState;
    U32 deltaNbBits;
} FSE_symbolCompressionTransform; /* total 8 bytes */

MEM_STATIC void FSE_initCState(FSE_CState_t* statePtr, const FSE_CTable* ct)
{
    const void* ptr = ct;
    const U16* u16ptr = (const U16*) ptr;
    const U32 tableLog = MEM_read16(ptr);
    statePtr->value = (ptrdiff_t)1<<tableLog;
    statePtr->stateTable = u16ptr+2;
    statePtr->symbolTT = ct + 1 + (tableLog ? (1<<(tableLog-1)) : 1);
    statePtr->stateLog = tableLog;
}


/*! FSE_initCState2() :
*   Same as FSE_initCState(), but the first symbol to include (which will be the last to be read)
*   uses the smallest state value possible, saving the cost of this symbol */
MEM_STATIC void FSE_initCState2(FSE_CState_t* statePtr, const FSE_CTable* ct, U32 symbol)
{
    FSE_initCState(statePtr, ct);
    {   const FSE_symbolCompressionTransform symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
        const U16* stateTable = (const U16*)(statePtr->stateTable);
        U32 nbBitsOut  = (U32)((symbolTT.deltaNbBits + (1<<15)) >> 16);
        statePtr->value = (nbBitsOut << 16) - symbolTT.deltaNbBits;
        statePtr->value = stateTable[(statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
    }
}

MEM_STATIC void FSE_encodeSymbol(BIT_CStream_t* bitC, FSE_CState_t* statePtr, unsigned symbol)
{
    FSE_symbolCompressionTransform const symbolTT = ((const FSE_symbolCompressionTransform*)(statePtr->symbolTT))[symbol];
    const U16* const stateTable = (const U16*)(statePtr->stateTable);
    U32 const nbBitsOut  = (U32)((statePtr->value + symbolTT.deltaNbBits) >> 16);
    BIT_addBits(bitC, (BitContainerType)statePtr->value, nbBitsOut);
    statePtr->value = stateTable[ (statePtr->value >> nbBitsOut) + symbolTT.deltaFindState];
}

MEM_STATIC void FSE_flushCState(BIT_CStream_t* bitC, const FSE_CState_t* statePtr)
{
    BIT_addBits(bitC, (BitContainerType)statePtr->value, statePtr->stateLog);
    BIT_flushBits(bitC);
}


/* FSE_getMaxNbBits() :
 * Approximate maximum cost of a symbol, in bits.
 * Fractional get rounded up (i.e. a symbol with a normalized frequency of 3 gives the same result as a frequency of 2)
 * note 1 : assume symbolValue is valid (<= maxSymbolValue)
 * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
MEM_STATIC U32 FSE_getMaxNbBits(const void* symbolTTPtr, U32 symbolValue)
{
    const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
    return (symbolTT[symbolValue].deltaNbBits + ((1<<16)-1)) >> 16;
}

/* FSE_bitCost() :
 * Approximate symbol cost, as fractional value, using fixed-point format (accuracyLog fractional bits)
 * note 1 : assume symbolValue is valid (<= maxSymbolValue)
 * note 2 : if freq[symbolValue]==0, @return a fake cost of tableLog+1 bits */
MEM_STATIC U32 FSE_bitCost(const void* symbolTTPtr, U32 tableLog, U32 symbolValue, U32 accuracyLog)
{
    const FSE_symbolCompressionTransform* symbolTT = (const FSE_symbolCompressionTransform*) symbolTTPtr;
    U32 const minNbBits = symbolTT[symbolValue].deltaNbBits >> 16;
    U32 const threshold = (minNbBits+1) << 16;
    assert(tableLog < 16);
    assert(accuracyLog < 31-tableLog);  /* ensure enough room for renormalization double shift */
    {   U32 const tableSize = 1 << tableLog;
        U32 const deltaFromThreshold = threshold - (symbolTT[symbolValue].deltaNbBits + tableSize);
        U32 const normalizedDeltaFromThreshold = (deltaFromThreshold << accuracyLog) >> tableLog;   /* linear interpolation (very approximate) */
        U32 const bitMultiplier = 1 << accuracyLog;
        assert(symbolTT[symbolValue].deltaNbBits + tableSize <= threshold);
        assert(normalizedDeltaFromThreshold <= bitMultiplier);
        return (minNbBits+1)*bitMultiplier - normalizedDeltaFromThreshold;
    }
}


/* ======    Decompression    ====== */

typedef struct {
    U16 tableLog;
    U16 fastMode;
} FSE_DTableHeader;   /* sizeof U32 */

typedef struct
{
    unsigned short newState;
    unsigned char  symbol;
    unsigned char  nbBits;
} FSE_decode_t;   /* size == U32 */

MEM_STATIC void FSE_initDState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD, const FSE_DTable* dt)
{
    const void* ptr = dt;
    const FSE_DTableHeader* const DTableH = (const FSE_DTableHeader*)ptr;
    DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
    BIT_reloadDStream(bitD);
    DStatePtr->table = dt + 1;
}

MEM_STATIC BYTE FSE_peekSymbol(const FSE_DState_t* DStatePtr)
{
    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
    return DInfo.symbol;
}

MEM_STATIC void FSE_updateState(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
{
    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
    U32 const nbBits = DInfo.nbBits;
    size_t const lowBits = BIT_readBits(bitD, nbBits);
    DStatePtr->state = DInfo.newState + lowBits;
}

MEM_STATIC BYTE FSE_decodeSymbol(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
{
    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
    U32 const nbBits = DInfo.nbBits;
    BYTE const symbol = DInfo.symbol;
    size_t const lowBits = BIT_readBits(bitD, nbBits);

    DStatePtr->state = DInfo.newState + lowBits;
    return symbol;
}

/*! FSE_decodeSymbolFast() :
    unsafe, only works if no symbol has a probability > 50% */
MEM_STATIC BYTE FSE_decodeSymbolFast(FSE_DState_t* DStatePtr, BIT_DStream_t* bitD)
{
    FSE_decode_t const DInfo = ((const FSE_decode_t*)(DStatePtr->table))[DStatePtr->state];
    U32 const nbBits = DInfo.nbBits;
    BYTE const symbol = DInfo.symbol;
    size_t const lowBits = BIT_readBitsFast(bitD, nbBits);

    DStatePtr->state = DInfo.newState + lowBits;
    return symbol;
}

MEM_STATIC unsigned FSE_endOfDState(const FSE_DState_t* DStatePtr)
{
    return DStatePtr->state == 0;
}


#ifndef FSE_COMMONDEFS_ONLY

/* **************************************************************
*  Tuning parameters
****************************************************************/
/*!MEMORY_USAGE :
*  Memory usage formula : N->2^N Bytes (examples : 10 -> 1KB; 12 -> 4KB ; 16 -> 64KB; 20 -> 1MB; etc.)
*  Increasing memory usage improves compression ratio
*  Reduced memory usage can improve speed, due to cache effect
*  Recommended max value is 14, for 16KB, which nicely fits into Intel x86 L1 cache */
#ifndef FSE_MAX_MEMORY_USAGE
#  define FSE_MAX_MEMORY_USAGE 14
#endif
#ifndef FSE_DEFAULT_MEMORY_USAGE
#  define FSE_DEFAULT_MEMORY_USAGE 13
#endif
#if (FSE_DEFAULT_MEMORY_USAGE > FSE_MAX_MEMORY_USAGE)
#  error "FSE_DEFAULT_MEMORY_USAGE must be <= FSE_MAX_MEMORY_USAGE"
#endif

/*!FSE_MAX_SYMBOL_VALUE :
*  Maximum symbol value authorized.
*  Required for proper stack allocation */
#ifndef FSE_MAX_SYMBOL_VALUE
#  define FSE_MAX_SYMBOL_VALUE 255
#endif

/* **************************************************************
*  template functions type & suffix
****************************************************************/
#define FSE_FUNCTION_TYPE BYTE
#define FSE_FUNCTION_EXTENSION
#define FSE_DECODE_TYPE FSE_decode_t


#endif   /* !FSE_COMMONDEFS_ONLY */


/* ***************************************************************
*  Constants
*****************************************************************/
#define FSE_MAX_TABLELOG  (FSE_MAX_MEMORY_USAGE-2)
#define FSE_MAX_TABLESIZE (1U<<FSE_MAX_TABLELOG)
#define FSE_MAXTABLESIZE_MASK (FSE_MAX_TABLESIZE-1)
#define FSE_DEFAULT_TABLELOG (FSE_DEFAULT_MEMORY_USAGE-2)
#define FSE_MIN_TABLELOG 5

#define FSE_TABLELOG_ABSOLUTE_MAX 15
#if FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX
#  error "FSE_MAX_TABLELOG > FSE_TABLELOG_ABSOLUTE_MAX is not supported"
#endif

#define FSE_TABLESTEP(tableSize) (((tableSize)>>1) + ((tableSize)>>3) + 3)

#endif /* FSE_STATIC_LINKING_ONLY */
/**** ended inlining fse.h ****/
/**** start inlining huf.h ****/
/* ******************************************************************
 * huff0 huffman codec,
 * part of Finite State Entropy library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 * You can contact the author at :
 * - Source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

#ifndef HUF_H_298734234
#define HUF_H_298734234

/* *** Dependencies *** */
/**** skipping file: zstd_deps.h ****/
/**** skipping file: mem.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: fse.h ****/

/* ***   Tool functions *** */
#define HUF_BLOCKSIZE_MAX (128 * 1024)   /**< maximum input size for a single block compressed with HUF_compress */
size_t HUF_compressBound(size_t size);   /**< maximum compressed size (worst case) */

/* Error Management */
unsigned    HUF_isError(size_t code);       /**< tells if a return value is an error code */
const char* HUF_getErrorName(size_t code);  /**< provides error code string (useful for debugging) */


#define HUF_WORKSPACE_SIZE ((8 << 10) + 512 /* sorting scratch space */)
#define HUF_WORKSPACE_SIZE_U64 (HUF_WORKSPACE_SIZE / sizeof(U64))

/* *** Constants *** */
#define HUF_TABLELOG_MAX      12      /* max runtime value of tableLog (due to static allocation); can be modified up to HUF_TABLELOG_ABSOLUTEMAX */
#define HUF_TABLELOG_DEFAULT  11      /* default tableLog value when none specified */
#define HUF_SYMBOLVALUE_MAX  255

#define HUF_TABLELOG_ABSOLUTEMAX  12  /* absolute limit of HUF_MAX_TABLELOG. Beyond that value, code does not work */
#if (HUF_TABLELOG_MAX > HUF_TABLELOG_ABSOLUTEMAX)
#  error "HUF_TABLELOG_MAX is too large !"
#endif


/* ****************************************
*  Static allocation
******************************************/
/* HUF buffer bounds */
#define HUF_CTABLEBOUND 129
#define HUF_BLOCKBOUND(size) (size + (size>>8) + 8)   /* only true when incompressible is pre-filtered with fast heuristic */
#define HUF_COMPRESSBOUND(size) (HUF_CTABLEBOUND + HUF_BLOCKBOUND(size))   /* Macro version, useful for static allocation */

/* static allocation of HUF's Compression Table */
/* this is a private definition, just exposed for allocation and strict aliasing purpose. never EVER access its members directly */
typedef size_t HUF_CElt;   /* consider it an incomplete type */
#define HUF_CTABLE_SIZE_ST(maxSymbolValue)   ((maxSymbolValue)+2)   /* Use tables of size_t, for proper alignment */
#define HUF_CTABLE_SIZE(maxSymbolValue)       (HUF_CTABLE_SIZE_ST(maxSymbolValue) * sizeof(size_t))
#define HUF_CREATE_STATIC_CTABLE(name, maxSymbolValue) \
    HUF_CElt name[HUF_CTABLE_SIZE_ST(maxSymbolValue)] /* no final ; */

/* static allocation of HUF's DTable */
typedef U32 HUF_DTable;
#define HUF_DTABLE_SIZE(maxTableLog)   (1 + (1<<(maxTableLog)))
#define HUF_CREATE_STATIC_DTABLEX1(DTable, maxTableLog) \
        HUF_DTable DTable[HUF_DTABLE_SIZE((maxTableLog)-1)] = { ((U32)((maxTableLog)-1) * 0x01000001) }
#define HUF_CREATE_STATIC_DTABLEX2(DTable, maxTableLog) \
        HUF_DTable DTable[HUF_DTABLE_SIZE(maxTableLog)] = { ((U32)(maxTableLog) * 0x01000001) }


/* ****************************************
*  Advanced decompression functions
******************************************/

/**
 * Huffman flags bitset.
 * For all flags, 0 is the default value.
 */
typedef enum {
    /**
     * If compiled with DYNAMIC_BMI2: Set flag only if the CPU supports BMI2 at runtime.
     * Otherwise: Ignored.
     */
    HUF_flags_bmi2 = (1 << 0),
    /**
     * If set: Test possible table depths to find the one that produces the smallest header + encoded size.
     * If unset: Use heuristic to find the table depth.
     */
    HUF_flags_optimalDepth = (1 << 1),
    /**
     * If set: If the previous table can encode the input, always reuse the previous table.
     * If unset: If the previous table can encode the input, reuse the previous table if it results in a smaller output.
     */
    HUF_flags_preferRepeat = (1 << 2),
    /**
     * If set: Sample the input and check if the sample is uncompressible, if it is then don't attempt to compress.
     * If unset: Always histogram the entire input.
     */
    HUF_flags_suspectUncompressible = (1 << 3),
    /**
     * If set: Don't use assembly implementations
     * If unset: Allow using assembly implementations
     */
    HUF_flags_disableAsm = (1 << 4),
    /**
     * If set: Don't use the fast decoding loop, always use the fallback decoding loop.
     * If unset: Use the fast decoding loop when possible.
     */
    HUF_flags_disableFast = (1 << 5)
} HUF_flags_e;


/* ****************************************
 *  HUF detailed API
 * ****************************************/
#define HUF_OPTIMAL_DEPTH_THRESHOLD ZSTD_btultra

/*! HUF_compress() does the following:
 *  1. count symbol occurrence from source[] into table count[] using FSE_count() (exposed within "fse.h")
 *  2. (optional) refine tableLog using HUF_optimalTableLog()
 *  3. build Huffman table from count using HUF_buildCTable()
 *  4. save Huffman table to memory buffer using HUF_writeCTable()
 *  5. encode the data stream using HUF_compress4X_usingCTable()
 *
 *  The following API allows targeting specific sub-functions for advanced tasks.
 *  For example, it's possible to compress several blocks using the same 'CTable',
 *  or to save and regenerate 'CTable' using external methods.
 */
unsigned HUF_minTableLog(unsigned symbolCardinality);
unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue);
unsigned HUF_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, void* workSpace,
 size_t wkspSize, HUF_CElt* table, const unsigned* count, int flags); /* table is used as scratch space for building and testing tables, not a return value */
size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize, const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog, void* workspace, size_t workspaceSize);
size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);
int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue);

typedef enum {
   HUF_repeat_none,  /**< Cannot use the previous table */
   HUF_repeat_check, /**< Can use the previous table but it must be checked. Note : The previous table must have been constructed by HUF_compress{1, 4}X_repeat */
   HUF_repeat_valid  /**< Can use the previous table and it is assumed to be valid */
 } HUF_repeat;

/** HUF_compress4X_repeat() :
 *  Same as HUF_compress4X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
 *  If it uses hufTable it does not modify hufTable or repeat.
 *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
 *  If preferRepeat then the old table will always be used if valid.
 *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
size_t HUF_compress4X_repeat(void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned tableLog,
                       void* workSpace, size_t wkspSize,    /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);

/** HUF_buildCTable_wksp() :
 *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
 * `workSpace` must be aligned on 4-bytes boundaries, and its size must be >= HUF_CTABLE_WORKSPACE_SIZE.
 */
#define HUF_CTABLE_WORKSPACE_SIZE_U32 ((4 * (HUF_SYMBOLVALUE_MAX + 1)) + 192)
#define HUF_CTABLE_WORKSPACE_SIZE (HUF_CTABLE_WORKSPACE_SIZE_U32 * sizeof(unsigned))
size_t HUF_buildCTable_wksp (HUF_CElt* tree,
                       const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
                             void* workSpace, size_t wkspSize);

/*! HUF_readStats() :
 *  Read compact Huffman tree, saved by HUF_writeCTable().
 * `huffWeight` is destination buffer.
 * @return : size read from `src` , or an error Code .
 *  Note : Needed by HUF_readCTable() and HUF_readDTableXn() . */
size_t HUF_readStats(BYTE* huffWeight, size_t hwSize,
                     U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
                     const void* src, size_t srcSize);

/*! HUF_readStats_wksp() :
 * Same as HUF_readStats() but takes an external workspace which must be
 * 4-byte aligned and its size must be >= HUF_READ_STATS_WORKSPACE_SIZE.
 * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
 */
#define HUF_READ_STATS_WORKSPACE_SIZE_U32 FSE_DECOMPRESS_WKSP_SIZE_U32(6, HUF_TABLELOG_MAX-1)
#define HUF_READ_STATS_WORKSPACE_SIZE (HUF_READ_STATS_WORKSPACE_SIZE_U32 * sizeof(unsigned))
size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize,
                          U32* rankStats, U32* nbSymbolsPtr, U32* tableLogPtr,
                          const void* src, size_t srcSize,
                          void* workspace, size_t wkspSize,
                          int flags);

/** HUF_readCTable() :
 *  Loading a CTable saved with HUF_writeCTable() */
size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned *hasZeroWeights);

/** HUF_getNbBitsFromCTable() :
 *  Read nbBits from CTable symbolTable, for symbol `symbolValue` presumed <= HUF_SYMBOLVALUE_MAX
 *  Note 1 : If symbolValue > HUF_readCTableHeader(symbolTable).maxSymbolValue, returns 0
 *  Note 2 : is not inlined, as HUF_CElt definition is private
 */
U32 HUF_getNbBitsFromCTable(const HUF_CElt* symbolTable, U32 symbolValue);

typedef struct {
    BYTE tableLog;
    BYTE maxSymbolValue;
    BYTE unused[sizeof(size_t) - 2];
} HUF_CTableHeader;

/** HUF_readCTableHeader() :
 *  @returns The header from the CTable specifying the tableLog and the maxSymbolValue.
 */
HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable);

/*
 * HUF_decompress() does the following:
 * 1. select the decompression algorithm (X1, X2) based on pre-computed heuristics
 * 2. build Huffman table from save, using HUF_readDTableX?()
 * 3. decode 1 or 4 segments in parallel using HUF_decompress?X?_usingDTable()
 */

/** HUF_selectDecoder() :
 *  Tells which decoder is likely to decode faster,
 *  based on a set of pre-computed metrics.
 * @return : 0==HUF_decompress4X1, 1==HUF_decompress4X2 .
 *  Assumption : 0 < dstSize <= 128 KB */
U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize);

/**
 *  The minimum workspace size for the `workSpace` used in
 *  HUF_readDTableX1_wksp() and HUF_readDTableX2_wksp().
 *
 *  The space used depends on HUF_TABLELOG_MAX, ranging from ~1500 bytes when
 *  HUF_TABLE_LOG_MAX=12 to ~1850 bytes when HUF_TABLE_LOG_MAX=15.
 *  Buffer overflow errors may potentially occur if code modifications result in
 *  a required workspace size greater than that specified in the following
 *  macro.
 */
#define HUF_DECOMPRESS_WORKSPACE_SIZE ((2 << 10) + (1 << 9))
#define HUF_DECOMPRESS_WORKSPACE_SIZE_U32 (HUF_DECOMPRESS_WORKSPACE_SIZE / sizeof(U32))


/* ====================== */
/* single stream variants */
/* ====================== */

size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags);
/** HUF_compress1X_repeat() :
 *  Same as HUF_compress1X_wksp(), but considers using hufTable if *repeat != HUF_repeat_none.
 *  If it uses hufTable it does not modify hufTable or repeat.
 *  If it doesn't, it sets *repeat = HUF_repeat_none, and it sets hufTable to the table used.
 *  If preferRepeat then the old table will always be used if valid.
 *  If suspectUncompressible then some sampling checks will be run to potentially skip huffman coding */
size_t HUF_compress1X_repeat(void* dst, size_t dstSize,
                       const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned tableLog,
                       void* workSpace, size_t wkspSize,   /**< `workSpace` must be aligned on 4-bytes boundaries, `wkspSize` must be >= HUF_WORKSPACE_SIZE */
                       HUF_CElt* hufTable, HUF_repeat* repeat, int flags);

size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
#ifndef HUF_FORCE_DECOMPRESS_X1
size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);   /**< double-symbols decoder */
#endif

/* BMI2 variants.
 * If the CPU has BMI2 support, pass bmi2=1, otherwise pass bmi2=0.
 */
size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
#ifndef HUF_FORCE_DECOMPRESS_X2
size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
#endif
size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags);
size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags);
#ifndef HUF_FORCE_DECOMPRESS_X2
size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
#endif
#ifndef HUF_FORCE_DECOMPRESS_X1
size_t HUF_readDTableX2_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags);
#endif

#endif   /* HUF_H_298734234 */
/**** ended inlining huf.h ****/
/**** skipping file: bits.h ****/


/*===   Version   ===*/
unsigned FSE_versionNumber(void) { return FSE_VERSION_NUMBER; }


/*===   Error Management   ===*/
unsigned FSE_isError(size_t code) { return ERR_isError(code); }
const char* FSE_getErrorName(size_t code) { return ERR_getErrorName(code); }

unsigned HUF_isError(size_t code) { return ERR_isError(code); }
const char* HUF_getErrorName(size_t code) { return ERR_getErrorName(code); }


/*-**************************************************************
*  FSE NCount encoding-decoding
****************************************************************/
FORCE_INLINE_TEMPLATE
size_t FSE_readNCount_body(short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
                           const void* headerBuffer, size_t hbSize)
{
    const BYTE* const istart = (const BYTE*) headerBuffer;
    const BYTE* const iend = istart + hbSize;
    const BYTE* ip = istart;
    int nbBits;
    int remaining;
    int threshold;
    U32 bitStream;
    int bitCount;
    unsigned charnum = 0;
    unsigned const maxSV1 = *maxSVPtr + 1;
    int previous0 = 0;

    if (hbSize < 8) {
        /* This function only works when hbSize >= 8 */
        char buffer[8] = {0};
        ZSTD_memcpy(buffer, headerBuffer, hbSize);
        {   size_t const countSize = FSE_readNCount(normalizedCounter, maxSVPtr, tableLogPtr,
                                                    buffer, sizeof(buffer));
            if (FSE_isError(countSize)) return countSize;
            if (countSize > hbSize) return ERROR(corruption_detected);
            return countSize;
    }   }
    assert(hbSize >= 8);

    /* init */
    ZSTD_memset(normalizedCounter, 0, (*maxSVPtr+1) * sizeof(normalizedCounter[0]));   /* all symbols not present in NCount have a frequency of 0 */
    bitStream = MEM_readLE32(ip);
    nbBits = (bitStream & 0xF) + FSE_MIN_TABLELOG;   /* extract tableLog */
    if (nbBits > FSE_TABLELOG_ABSOLUTE_MAX) return ERROR(tableLog_tooLarge);
    bitStream >>= 4;
    bitCount = 4;
    *tableLogPtr = nbBits;
    remaining = (1<<nbBits)+1;
    threshold = 1<<nbBits;
    nbBits++;

    for (;;) {
        if (previous0) {
            /* Count the number of repeats. Each time the
             * 2-bit repeat code is 0b11 there is another
             * repeat.
             * Avoid UB by setting the high bit to 1.
             */
            int repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
            while (repeats >= 12) {
                charnum += 3 * 12;
                if (LIKELY(ip <= iend-7)) {
                    ip += 3;
                } else {
                    bitCount -= (int)(8 * (iend - 7 - ip));
                    bitCount &= 31;
                    ip = iend - 4;
                }
                bitStream = MEM_readLE32(ip) >> bitCount;
                repeats = ZSTD_countTrailingZeros32(~bitStream | 0x80000000) >> 1;
            }
            charnum += 3 * repeats;
            bitStream >>= 2 * repeats;
            bitCount += 2 * repeats;

            /* Add the final repeat which isn't 0b11. */
            assert((bitStream & 3) < 3);
            charnum += bitStream & 3;
            bitCount += 2;

            /* This is an error, but break and return an error
             * at the end, because returning out of a loop makes
             * it harder for the compiler to optimize.
             */
            if (charnum >= maxSV1) break;

            /* We don't need to set the normalized count to 0
             * because we already memset the whole buffer to 0.
             */

            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                assert((bitCount >> 3) <= 3); /* For first condition to work */
                ip += bitCount>>3;
                bitCount &= 7;
            } else {
                bitCount -= (int)(8 * (iend - 4 - ip));
                bitCount &= 31;
                ip = iend - 4;
            }
            bitStream = MEM_readLE32(ip) >> bitCount;
        }
        {
            int const max = (2*threshold-1) - remaining;
            int count;

            if ((bitStream & (threshold-1)) < (U32)max) {
                count = bitStream & (threshold-1);
                bitCount += nbBits-1;
            } else {
                count = bitStream & (2*threshold-1);
                if (count >= threshold) count -= max;
                bitCount += nbBits;
            }

            count--;   /* extra accuracy */
            /* When it matters (small blocks), this is a
             * predictable branch, because we don't use -1.
             */
            if (count >= 0) {
                remaining -= count;
            } else {
                assert(count == -1);
                remaining += count;
            }
            normalizedCounter[charnum++] = (short)count;
            previous0 = !count;

            assert(threshold > 1);
            if (remaining < threshold) {
                /* This branch can be folded into the
                 * threshold update condition because we
                 * know that threshold > 1.
                 */
                if (remaining <= 1) break;
                nbBits = ZSTD_highbit32(remaining) + 1;
                threshold = 1 << (nbBits - 1);
            }
            if (charnum >= maxSV1) break;

            if (LIKELY(ip <= iend-7) || (ip + (bitCount>>3) <= iend-4)) {
                ip += bitCount>>3;
                bitCount &= 7;
            } else {
                bitCount -= (int)(8 * (iend - 4 - ip));
                bitCount &= 31;
                ip = iend - 4;
            }
            bitStream = MEM_readLE32(ip) >> bitCount;
    }   }
    if (remaining != 1) return ERROR(corruption_detected);
    /* Only possible when there are too many zeros. */
    if (charnum > maxSV1) return ERROR(maxSymbolValue_tooSmall);
    if (bitCount > 32) return ERROR(corruption_detected);
    *maxSVPtr = charnum-1;

    ip += (bitCount+7)>>3;
    return ip-istart;
}

/* Avoids the FORCE_INLINE of the _body() function. */
static size_t FSE_readNCount_body_default(
        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
        const void* headerBuffer, size_t hbSize)
{
    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
}

#if DYNAMIC_BMI2
BMI2_TARGET_ATTRIBUTE static size_t FSE_readNCount_body_bmi2(
        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
        const void* headerBuffer, size_t hbSize)
{
    return FSE_readNCount_body(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
}
#endif

size_t FSE_readNCount_bmi2(
        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
        const void* headerBuffer, size_t hbSize, int bmi2)
{
#if DYNAMIC_BMI2
    if (bmi2) {
        return FSE_readNCount_body_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
    }
#endif
    (void)bmi2;
    return FSE_readNCount_body_default(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize);
}

size_t FSE_readNCount(
        short* normalizedCounter, unsigned* maxSVPtr, unsigned* tableLogPtr,
        const void* headerBuffer, size_t hbSize)
{
    return FSE_readNCount_bmi2(normalizedCounter, maxSVPtr, tableLogPtr, headerBuffer, hbSize, /* bmi2 */ 0);
}


/*! HUF_readStats() :
    Read compact Huffman tree, saved by HUF_writeCTable().
    `huffWeight` is destination buffer.
    `rankStats` is assumed to be a table of at least HUF_TABLELOG_MAX U32.
    @return : size read from `src` , or an error Code .
    Note : Needed by HUF_readCTable() and HUF_readDTableX?() .
*/
size_t HUF_readStats(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                     U32* nbSymbolsPtr, U32* tableLogPtr,
                     const void* src, size_t srcSize)
{
    U32 wksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
    return HUF_readStats_wksp(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, wksp, sizeof(wksp), /* flags */ 0);
}

FORCE_INLINE_TEMPLATE size_t
HUF_readStats_body(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                   U32* nbSymbolsPtr, U32* tableLogPtr,
                   const void* src, size_t srcSize,
                   void* workSpace, size_t wkspSize,
                   int bmi2)
{
    U32 weightTotal;
    const BYTE* ip = (const BYTE*) src;
    size_t iSize;
    size_t oSize;

    if (!srcSize) return ERROR(srcSize_wrong);
    iSize = ip[0];
    /* ZSTD_memset(huffWeight, 0, hwSize);   *//* is not necessary, even though some analyzer complain ... */

    if (iSize >= 128) {  /* special header */
        oSize = iSize - 127;
        iSize = ((oSize+1)/2);
        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
        if (oSize >= hwSize) return ERROR(corruption_detected);
        ip += 1;
        {   U32 n;
            for (n=0; n<oSize; n+=2) {
                huffWeight[n]   = ip[n/2] >> 4;
                huffWeight[n+1] = ip[n/2] & 15;
    }   }   }
    else  {   /* header compressed with FSE (normal case) */
        if (iSize+1 > srcSize) return ERROR(srcSize_wrong);
        /* max (hwSize-1) values decoded, as last one is implied */
        oSize = FSE_decompress_wksp_bmi2(huffWeight, hwSize-1, ip+1, iSize, 6, workSpace, wkspSize, bmi2);
        if (FSE_isError(oSize)) return oSize;
    }

    /* collect weight stats */
    ZSTD_memset(rankStats, 0, (HUF_TABLELOG_MAX + 1) * sizeof(U32));
    weightTotal = 0;
    {   U32 n; for (n=0; n<oSize; n++) {
            if (huffWeight[n] > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
            rankStats[huffWeight[n]]++;
            weightTotal += (1 << huffWeight[n]) >> 1;
    }   }
    if (weightTotal == 0) return ERROR(corruption_detected);

    /* get last non-null symbol weight (implied, total must be 2^n) */
    {   U32 const tableLog = ZSTD_highbit32(weightTotal) + 1;
        if (tableLog > HUF_TABLELOG_MAX) return ERROR(corruption_detected);
        *tableLogPtr = tableLog;
        /* determine last weight */
        {   U32 const total = 1 << tableLog;
            U32 const rest = total - weightTotal;
            U32 const verif = 1 << ZSTD_highbit32(rest);
            U32 const lastWeight = ZSTD_highbit32(rest) + 1;
            if (verif != rest) return ERROR(corruption_detected);    /* last value must be a clean power of 2 */
            huffWeight[oSize] = (BYTE)lastWeight;
            rankStats[lastWeight]++;
    }   }

    /* check tree construction validity */
    if ((rankStats[1] < 2) || (rankStats[1] & 1)) return ERROR(corruption_detected);   /* by construction : at least 2 elts of rank 1, must be even */

    /* results */
    *nbSymbolsPtr = (U32)(oSize+1);
    return iSize+1;
}

/* Avoids the FORCE_INLINE of the _body() function. */
static size_t HUF_readStats_body_default(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                     U32* nbSymbolsPtr, U32* tableLogPtr,
                     const void* src, size_t srcSize,
                     void* workSpace, size_t wkspSize)
{
    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 0);
}

#if DYNAMIC_BMI2
static BMI2_TARGET_ATTRIBUTE size_t HUF_readStats_body_bmi2(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                     U32* nbSymbolsPtr, U32* tableLogPtr,
                     const void* src, size_t srcSize,
                     void* workSpace, size_t wkspSize)
{
    return HUF_readStats_body(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize, 1);
}
#endif

size_t HUF_readStats_wksp(BYTE* huffWeight, size_t hwSize, U32* rankStats,
                     U32* nbSymbolsPtr, U32* tableLogPtr,
                     const void* src, size_t srcSize,
                     void* workSpace, size_t wkspSize,
                     int flags)
{
#if DYNAMIC_BMI2
    if (flags & HUF_flags_bmi2) {
        return HUF_readStats_body_bmi2(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
    }
#endif
    (void)flags;
    return HUF_readStats_body_default(huffWeight, hwSize, rankStats, nbSymbolsPtr, tableLogPtr, src, srcSize, workSpace, wkspSize);
}
/**** ended inlining common/entropy_common.c ****/
/**** start inlining common/error_private.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* The purpose of this file is to have a single list of error strings embedded in binary */

/**** skipping file: error_private.h ****/

const char* ERR_getErrorString(ERR_enum code)
{
#ifdef ZSTD_STRIP_ERROR_STRINGS
    (void)code;
    return "Error strings stripped";
#else
    static const char* const notErrorCode = "Unspecified error code";
    switch( code )
    {
    case PREFIX(no_error): return "No error detected";
    case PREFIX(GENERIC):  return "Error (generic)";
    case PREFIX(prefix_unknown): return "Unknown frame descriptor";
    case PREFIX(version_unsupported): return "Version not supported";
    case PREFIX(frameParameter_unsupported): return "Unsupported frame parameter";
    case PREFIX(frameParameter_windowTooLarge): return "Frame requires too much memory for decoding";
    case PREFIX(corruption_detected): return "Data corruption detected";
    case PREFIX(checksum_wrong): return "Restored data doesn't match checksum";
    case PREFIX(literals_headerWrong): return "Header of Literals' block doesn't respect format specification";
    case PREFIX(parameter_unsupported): return "Unsupported parameter";
    case PREFIX(parameter_combination_unsupported): return "Unsupported combination of parameters";
    case PREFIX(parameter_outOfBound): return "Parameter is out of bound";
    case PREFIX(init_missing): return "Context should be init first";
    case PREFIX(memory_allocation): return "Allocation error : not enough memory";
    case PREFIX(workSpace_tooSmall): return "workSpace buffer is not large enough";
    case PREFIX(stage_wrong): return "Operation not authorized at current processing stage";
    case PREFIX(tableLog_tooLarge): return "tableLog requires too much memory : unsupported";
    case PREFIX(maxSymbolValue_tooLarge): return "Unsupported max Symbol Value : too large";
    case PREFIX(maxSymbolValue_tooSmall): return "Specified maxSymbolValue is too small";
    case PREFIX(cannotProduce_uncompressedBlock): return "This mode cannot generate an uncompressed block";
    case PREFIX(stabilityCondition_notRespected): return "pledged buffer stability condition is not respected";
    case PREFIX(dictionary_corrupted): return "Dictionary is corrupted";
    case PREFIX(dictionary_wrong): return "Dictionary mismatch";
    case PREFIX(dictionaryCreation_failed): return "Cannot create Dictionary from provided samples";
    case PREFIX(dstSize_tooSmall): return "Destination buffer is too small";
    case PREFIX(srcSize_wrong): return "Src size is incorrect";
    case PREFIX(dstBuffer_null): return "Operation on NULL destination buffer";
    case PREFIX(noForwardProgress_destFull): return "Operation made no progress over multiple calls, due to output buffer being full";
    case PREFIX(noForwardProgress_inputEmpty): return "Operation made no progress over multiple calls, due to input being empty";
        /* following error codes are not stable and may be removed or changed in a future version */
    case PREFIX(frameIndex_tooLarge): return "Frame index is too large";
    case PREFIX(seekableIO): return "An I/O error occurred when reading/seeking";
    case PREFIX(dstBuffer_wrong): return "Destination buffer is wrong";
    case PREFIX(srcBuffer_wrong): return "Source buffer is wrong";
    case PREFIX(sequenceProducer_failed): return "Block-level external sequence producer returned an error code";
    case PREFIX(externalSequences_invalid): return "External sequences are not valid";
    case PREFIX(maxCode):
    default: return notErrorCode;
    }
#endif
}
/**** ended inlining common/error_private.c ****/
/**** start inlining common/fse_decompress.c ****/
/* ******************************************************************
 * FSE : Finite State Entropy decoder
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */


/* **************************************************************
*  Includes
****************************************************************/
/**** skipping file: debug.h ****/
/**** skipping file: bitstream.h ****/
/**** skipping file: compiler.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: fse.h ****/
/**** skipping file: error_private.h ****/
/**** skipping file: zstd_deps.h ****/
/**** skipping file: bits.h ****/


/* **************************************************************
*  Error Management
****************************************************************/
#define FSE_isError ERR_isError
#define FSE_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */


/* **************************************************************
*  Templates
****************************************************************/
/*
  designed to be included
  for type-specific functions (template emulation in C)
  Objective is to write these functions only once, for improved maintenance
*/

/* safety checks */
#ifndef FSE_FUNCTION_EXTENSION
#  error "FSE_FUNCTION_EXTENSION must be defined"
#endif
#ifndef FSE_FUNCTION_TYPE
#  error "FSE_FUNCTION_TYPE must be defined"
#endif

/* Function names */
#define FSE_CAT(X,Y) X##Y
#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)

static size_t FSE_buildDTable_internal(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
{
    void* const tdPtr = dt+1;   /* because *dt is unsigned, 32-bits aligned on 32-bits */
    FSE_DECODE_TYPE* const tableDecode = (FSE_DECODE_TYPE*) (tdPtr);
    U16* symbolNext = (U16*)workSpace;
    BYTE* spread = (BYTE*)(symbolNext + maxSymbolValue + 1);

    U32 const maxSV1 = maxSymbolValue + 1;
    U32 const tableSize = 1 << tableLog;
    U32 highThreshold = tableSize-1;

    /* Sanity Checks */
    if (FSE_BUILD_DTABLE_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(maxSymbolValue_tooLarge);
    if (maxSymbolValue > FSE_MAX_SYMBOL_VALUE) return ERROR(maxSymbolValue_tooLarge);
    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);

    /* Init, lay down lowprob symbols */
    {   FSE_DTableHeader DTableH;
        DTableH.tableLog = (U16)tableLog;
        DTableH.fastMode = 1;
        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
            U32 s;
            for (s=0; s<maxSV1; s++) {
                if (normalizedCounter[s]==-1) {
                    tableDecode[highThreshold--].symbol = (FSE_FUNCTION_TYPE)s;
                    symbolNext[s] = 1;
                } else {
                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
                    symbolNext[s] = (U16)normalizedCounter[s];
        }   }   }
        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
    }

    /* Spread symbols */
    if (highThreshold == tableSize - 1) {
        size_t const tableMask = tableSize-1;
        size_t const step = FSE_TABLESTEP(tableSize);
        /* First lay down the symbols in order.
         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
         * misses since small blocks generally have small table logs, so nearly
         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
         * our buffer to handle the over-write.
         */
        {   U64 const add = 0x0101010101010101ull;
            size_t pos = 0;
            U64 sv = 0;
            U32 s;
            for (s=0; s<maxSV1; ++s, sv += add) {
                int i;
                int const n = normalizedCounter[s];
                MEM_write64(spread + pos, sv);
                for (i = 8; i < n; i += 8) {
                    MEM_write64(spread + pos + i, sv);
                }
                pos += (size_t)n;
        }   }
        /* Now we spread those positions across the table.
         * The benefit of doing it in two stages is that we avoid the
         * variable size inner loop, which caused lots of branch misses.
         * Now we can run through all the positions without any branch misses.
         * We unroll the loop twice, since that is what empirically worked best.
         */
        {
            size_t position = 0;
            size_t s;
            size_t const unroll = 2;
            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
            for (s = 0; s < (size_t)tableSize; s += unroll) {
                size_t u;
                for (u = 0; u < unroll; ++u) {
                    size_t const uPosition = (position + (u * step)) & tableMask;
                    tableDecode[uPosition].symbol = spread[s + u];
                }
                position = (position + (unroll * step)) & tableMask;
            }
            assert(position == 0);
        }
    } else {
        U32 const tableMask = tableSize-1;
        U32 const step = FSE_TABLESTEP(tableSize);
        U32 s, position = 0;
        for (s=0; s<maxSV1; s++) {
            int i;
            for (i=0; i<normalizedCounter[s]; i++) {
                tableDecode[position].symbol = (FSE_FUNCTION_TYPE)s;
                position = (position + step) & tableMask;
                while (position > highThreshold) position = (position + step) & tableMask;   /* lowprob area */
        }   }
        if (position!=0) return ERROR(GENERIC);   /* position must reach all cells once, otherwise normalizedCounter is incorrect */
    }

    /* Build Decoding table */
    {   U32 u;
        for (u=0; u<tableSize; u++) {
            FSE_FUNCTION_TYPE const symbol = (FSE_FUNCTION_TYPE)(tableDecode[u].symbol);
            U32 const nextState = symbolNext[symbol]++;
            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
            tableDecode[u].newState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
    }   }

    return 0;
}

size_t FSE_buildDTable_wksp(FSE_DTable* dt, const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog, void* workSpace, size_t wkspSize)
{
    return FSE_buildDTable_internal(dt, normalizedCounter, maxSymbolValue, tableLog, workSpace, wkspSize);
}


#ifndef FSE_COMMONDEFS_ONLY

/*-*******************************************************
*  Decompression (Byte symbols)
*********************************************************/

FORCE_INLINE_TEMPLATE size_t FSE_decompress_usingDTable_generic(
          void* dst, size_t maxDstSize,
    const void* cSrc, size_t cSrcSize,
    const FSE_DTable* dt, const unsigned fast)
{
    BYTE* const ostart = (BYTE*) dst;
    BYTE* op = ostart;
    BYTE* const omax = op + maxDstSize;
    BYTE* const olimit = omax-3;

    BIT_DStream_t bitD;
    FSE_DState_t state1;
    FSE_DState_t state2;

    /* Init */
    CHECK_F(BIT_initDStream(&bitD, cSrc, cSrcSize));

    FSE_initDState(&state1, &bitD, dt);
    FSE_initDState(&state2, &bitD, dt);

    RETURN_ERROR_IF(BIT_reloadDStream(&bitD)==BIT_DStream_overflow, corruption_detected, "");

#define FSE_GETSYMBOL(statePtr) fast ? FSE_decodeSymbolFast(statePtr, &bitD) : FSE_decodeSymbol(statePtr, &bitD)

    /* 4 symbols per loop */
    for ( ; (BIT_reloadDStream(&bitD)==BIT_DStream_unfinished) & (op<olimit) ; op+=4) {
        op[0] = FSE_GETSYMBOL(&state1);

        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
            BIT_reloadDStream(&bitD);

        op[1] = FSE_GETSYMBOL(&state2);

        if (FSE_MAX_TABLELOG*4+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
            { if (BIT_reloadDStream(&bitD) > BIT_DStream_unfinished) { op+=2; break; } }

        op[2] = FSE_GETSYMBOL(&state1);

        if (FSE_MAX_TABLELOG*2+7 > sizeof(bitD.bitContainer)*8)    /* This test must be static */
            BIT_reloadDStream(&bitD);

        op[3] = FSE_GETSYMBOL(&state2);
    }

    /* tail */
    /* note : BIT_reloadDStream(&bitD) >= FSE_DStream_partiallyFilled; Ends at exactly BIT_DStream_completed */
    while (1) {
        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
        *op++ = FSE_GETSYMBOL(&state1);
        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
            *op++ = FSE_GETSYMBOL(&state2);
            break;
        }

        if (op>(omax-2)) return ERROR(dstSize_tooSmall);
        *op++ = FSE_GETSYMBOL(&state2);
        if (BIT_reloadDStream(&bitD)==BIT_DStream_overflow) {
            *op++ = FSE_GETSYMBOL(&state1);
            break;
    }   }

    assert(op >= ostart);
    return (size_t)(op-ostart);
}

typedef struct {
    short ncount[FSE_MAX_SYMBOL_VALUE + 1];
} FSE_DecompressWksp;


FORCE_INLINE_TEMPLATE size_t FSE_decompress_wksp_body(
        void* dst, size_t dstCapacity,
        const void* cSrc, size_t cSrcSize,
        unsigned maxLog, void* workSpace, size_t wkspSize,
        int bmi2)
{
    const BYTE* const istart = (const BYTE*)cSrc;
    const BYTE* ip = istart;
    unsigned tableLog;
    unsigned maxSymbolValue = FSE_MAX_SYMBOL_VALUE;
    FSE_DecompressWksp* const wksp = (FSE_DecompressWksp*)workSpace;
    size_t const dtablePos = sizeof(FSE_DecompressWksp) / sizeof(FSE_DTable);
    FSE_DTable* const dtable = (FSE_DTable*)workSpace + dtablePos;

    FSE_STATIC_ASSERT((FSE_MAX_SYMBOL_VALUE + 1) % 2 == 0);
    if (wkspSize < sizeof(*wksp)) return ERROR(GENERIC);

    /* correct offset to dtable depends on this property */
    FSE_STATIC_ASSERT(sizeof(FSE_DecompressWksp) % sizeof(FSE_DTable) == 0);

    /* normal FSE decoding mode */
    {   size_t const NCountLength =
            FSE_readNCount_bmi2(wksp->ncount, &maxSymbolValue, &tableLog, istart, cSrcSize, bmi2);
        if (FSE_isError(NCountLength)) return NCountLength;
        if (tableLog > maxLog) return ERROR(tableLog_tooLarge);
        assert(NCountLength <= cSrcSize);
        ip += NCountLength;
        cSrcSize -= NCountLength;
    }

    if (FSE_DECOMPRESS_WKSP_SIZE(tableLog, maxSymbolValue) > wkspSize) return ERROR(tableLog_tooLarge);
    assert(sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog) <= wkspSize);
    workSpace = (BYTE*)workSpace + sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);
    wkspSize -= sizeof(*wksp) + FSE_DTABLE_SIZE(tableLog);

    CHECK_F( FSE_buildDTable_internal(dtable, wksp->ncount, maxSymbolValue, tableLog, workSpace, wkspSize) );

    {
        const void* ptr = dtable;
        const FSE_DTableHeader* DTableH = (const FSE_DTableHeader*)ptr;
        const U32 fastMode = DTableH->fastMode;

        /* select fast mode (static) */
        if (fastMode) return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 1);
        return FSE_decompress_usingDTable_generic(dst, dstCapacity, ip, cSrcSize, dtable, 0);
    }
}

/* Avoids the FORCE_INLINE of the _body() function. */
static size_t FSE_decompress_wksp_body_default(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
{
    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 0);
}

#if DYNAMIC_BMI2
BMI2_TARGET_ATTRIBUTE static size_t FSE_decompress_wksp_body_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize)
{
    return FSE_decompress_wksp_body(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize, 1);
}
#endif

size_t FSE_decompress_wksp_bmi2(void* dst, size_t dstCapacity, const void* cSrc, size_t cSrcSize, unsigned maxLog, void* workSpace, size_t wkspSize, int bmi2)
{
#if DYNAMIC_BMI2
    if (bmi2) {
        return FSE_decompress_wksp_body_bmi2(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
    }
#endif
    (void)bmi2;
    return FSE_decompress_wksp_body_default(dst, dstCapacity, cSrc, cSrcSize, maxLog, workSpace, wkspSize);
}

#endif   /* FSE_COMMONDEFS_ONLY */
/**** ended inlining common/fse_decompress.c ****/
/**** start inlining common/threading.c ****/
/**
 * Copyright (c) 2016 Tino Reichardt
 * All rights reserved.
 *
 * You can contact the author at:
 * - zstdmt source repository: https://github.com/mcmilk/zstdmt
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**
 * This file will hold wrapper for systems, which do not support pthreads
 */

/**** start inlining threading.h ****/
/**
 * Copyright (c) 2016 Tino Reichardt
 * All rights reserved.
 *
 * You can contact the author at:
 * - zstdmt source repository: https://github.com/mcmilk/zstdmt
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef THREADING_H_938743
#define THREADING_H_938743

/**** skipping file: debug.h ****/

#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)

/**
 * Windows minimalist Pthread Wrapper
 */
#ifdef WINVER
#  undef WINVER
#endif
#define WINVER       0x0600

#ifdef _WIN32_WINNT
#  undef _WIN32_WINNT
#endif
#define _WIN32_WINNT 0x0600

#ifndef WIN32_LEAN_AND_MEAN
#  define WIN32_LEAN_AND_MEAN
#endif

#undef ERROR   /* reported already defined on VS 2015 (Rich Geldreich) */
#include <windows.h>
#undef ERROR
#define ERROR(name) ZSTD_ERROR(name)


/* mutex */
#define ZSTD_pthread_mutex_t           CRITICAL_SECTION
#define ZSTD_pthread_mutex_init(a, b)  ((void)(b), InitializeCriticalSection((a)), 0)
#define ZSTD_pthread_mutex_destroy(a)  DeleteCriticalSection((a))
#define ZSTD_pthread_mutex_lock(a)     EnterCriticalSection((a))
#define ZSTD_pthread_mutex_unlock(a)   LeaveCriticalSection((a))

/* condition variable */
#define ZSTD_pthread_cond_t             CONDITION_VARIABLE
#define ZSTD_pthread_cond_init(a, b)    ((void)(b), InitializeConditionVariable((a)), 0)
#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
#define ZSTD_pthread_cond_wait(a, b)    SleepConditionVariableCS((a), (b), INFINITE)
#define ZSTD_pthread_cond_signal(a)     WakeConditionVariable((a))
#define ZSTD_pthread_cond_broadcast(a)  WakeAllConditionVariable((a))

/* ZSTD_pthread_create() and ZSTD_pthread_join() */
typedef HANDLE ZSTD_pthread_t;

int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
                   void* (*start_routine) (void*), void* arg);

int ZSTD_pthread_join(ZSTD_pthread_t thread);

/**
 * add here more wrappers as required
 */

#elif defined(ZSTD_MULTITHREAD)    /* posix assumed ; need a better detection method */
/* ===   POSIX Systems   === */
#  include <pthread.h>

#if DEBUGLEVEL < 1

#define ZSTD_pthread_mutex_t            pthread_mutex_t
#define ZSTD_pthread_mutex_init(a, b)   pthread_mutex_init((a), (b))
#define ZSTD_pthread_mutex_destroy(a)   pthread_mutex_destroy((a))
#define ZSTD_pthread_mutex_lock(a)      pthread_mutex_lock((a))
#define ZSTD_pthread_mutex_unlock(a)    pthread_mutex_unlock((a))

#define ZSTD_pthread_cond_t             pthread_cond_t
#define ZSTD_pthread_cond_init(a, b)    pthread_cond_init((a), (b))
#define ZSTD_pthread_cond_destroy(a)    pthread_cond_destroy((a))
#define ZSTD_pthread_cond_wait(a, b)    pthread_cond_wait((a), (b))
#define ZSTD_pthread_cond_signal(a)     pthread_cond_signal((a))
#define ZSTD_pthread_cond_broadcast(a)  pthread_cond_broadcast((a))

#define ZSTD_pthread_t                  pthread_t
#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
#define ZSTD_pthread_join(a)         pthread_join((a),NULL)

#else /* DEBUGLEVEL >= 1 */

/* Debug implementation of threading.
 * In this implementation we use pointers for mutexes and condition variables.
 * This way, if we forget to init/destroy them the program will crash or ASAN
 * will report leaks.
 */

#define ZSTD_pthread_mutex_t            pthread_mutex_t*
int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr);
int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex);
#define ZSTD_pthread_mutex_lock(a)      pthread_mutex_lock(*(a))
#define ZSTD_pthread_mutex_unlock(a)    pthread_mutex_unlock(*(a))

#define ZSTD_pthread_cond_t             pthread_cond_t*
int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr);
int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond);
#define ZSTD_pthread_cond_wait(a, b)    pthread_cond_wait(*(a), *(b))
#define ZSTD_pthread_cond_signal(a)     pthread_cond_signal(*(a))
#define ZSTD_pthread_cond_broadcast(a)  pthread_cond_broadcast(*(a))

#define ZSTD_pthread_t                  pthread_t
#define ZSTD_pthread_create(a, b, c, d) pthread_create((a), (b), (c), (d))
#define ZSTD_pthread_join(a)         pthread_join((a),NULL)

#endif

#else  /* ZSTD_MULTITHREAD not defined */
/* No multithreading support */

typedef int ZSTD_pthread_mutex_t;
#define ZSTD_pthread_mutex_init(a, b)   ((void)(a), (void)(b), 0)
#define ZSTD_pthread_mutex_destroy(a)   ((void)(a))
#define ZSTD_pthread_mutex_lock(a)      ((void)(a))
#define ZSTD_pthread_mutex_unlock(a)    ((void)(a))

typedef int ZSTD_pthread_cond_t;
#define ZSTD_pthread_cond_init(a, b)    ((void)(a), (void)(b), 0)
#define ZSTD_pthread_cond_destroy(a)    ((void)(a))
#define ZSTD_pthread_cond_wait(a, b)    ((void)(a), (void)(b))
#define ZSTD_pthread_cond_signal(a)     ((void)(a))
#define ZSTD_pthread_cond_broadcast(a)  ((void)(a))

/* do not use ZSTD_pthread_t */

#endif /* ZSTD_MULTITHREAD */


#endif /* THREADING_H_938743 */
/**** ended inlining threading.h ****/

/* create fake symbol to avoid empty translation unit warning */
int g_ZSTD_threading_useless_symbol;

#if defined(ZSTD_MULTITHREAD) && defined(_WIN32)

/**
 * Windows minimalist Pthread Wrapper
 */


/* ===  Dependencies  === */
#include <process.h>
#include <errno.h>


/* ===  Implementation  === */

typedef struct {
    void* (*start_routine)(void*);
    void* arg;
    int initialized;
    ZSTD_pthread_cond_t initialized_cond;
    ZSTD_pthread_mutex_t initialized_mutex;
} ZSTD_thread_params_t;

static unsigned __stdcall worker(void *arg)
{
    void* (*start_routine)(void*);
    void* thread_arg;

    /* Initialized thread_arg and start_routine and signal main thread that we don't need it
     * to wait any longer.
     */
    {
        ZSTD_thread_params_t*  thread_param = (ZSTD_thread_params_t*)arg;
        thread_arg = thread_param->arg;
        start_routine = thread_param->start_routine;

        /* Signal main thread that we are running and do not depend on its memory anymore */
        ZSTD_pthread_mutex_lock(&thread_param->initialized_mutex);
        thread_param->initialized = 1;
        ZSTD_pthread_cond_signal(&thread_param->initialized_cond);
        ZSTD_pthread_mutex_unlock(&thread_param->initialized_mutex);
    }

    start_routine(thread_arg);

    return 0;
}

int ZSTD_pthread_create(ZSTD_pthread_t* thread, const void* unused,
            void* (*start_routine) (void*), void* arg)
{
    ZSTD_thread_params_t thread_param;
    (void)unused;

    if (thread==NULL) return -1;
    *thread = NULL;

    thread_param.start_routine = start_routine;
    thread_param.arg = arg;
    thread_param.initialized = 0;

    /* Setup thread initialization synchronization */
    if(ZSTD_pthread_cond_init(&thread_param.initialized_cond, NULL)) {
        /* Should never happen on Windows */
        return -1;
    }
    if(ZSTD_pthread_mutex_init(&thread_param.initialized_mutex, NULL)) {
        /* Should never happen on Windows */
        ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
        return -1;
    }

    /* Spawn thread */
    *thread = (HANDLE)_beginthreadex(NULL, 0, worker, &thread_param, 0, NULL);
    if (*thread==NULL) {
        ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
        ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);
        return errno;
    }

    /* Wait for thread to be initialized */
    ZSTD_pthread_mutex_lock(&thread_param.initialized_mutex);
    while(!thread_param.initialized) {
        ZSTD_pthread_cond_wait(&thread_param.initialized_cond, &thread_param.initialized_mutex);
    }
    ZSTD_pthread_mutex_unlock(&thread_param.initialized_mutex);
    ZSTD_pthread_mutex_destroy(&thread_param.initialized_mutex);
    ZSTD_pthread_cond_destroy(&thread_param.initialized_cond);

    return 0;
}

int ZSTD_pthread_join(ZSTD_pthread_t thread)
{
    DWORD result;

    if (!thread) return 0;

    result = WaitForSingleObject(thread, INFINITE);
    CloseHandle(thread);

    switch (result) {
    case WAIT_OBJECT_0:
        return 0;
    case WAIT_ABANDONED:
        return EINVAL;
    default:
        return GetLastError();
    }
}

#endif   /* ZSTD_MULTITHREAD */

#if defined(ZSTD_MULTITHREAD) && DEBUGLEVEL >= 1 && !defined(_WIN32)

#define ZSTD_DEPS_NEED_MALLOC
/**** skipping file: zstd_deps.h ****/

int ZSTD_pthread_mutex_init(ZSTD_pthread_mutex_t* mutex, pthread_mutexattr_t const* attr)
{
    assert(mutex != NULL);
    *mutex = (pthread_mutex_t*)ZSTD_malloc(sizeof(pthread_mutex_t));
    if (!*mutex)
        return 1;
    return pthread_mutex_init(*mutex, attr);
}

int ZSTD_pthread_mutex_destroy(ZSTD_pthread_mutex_t* mutex)
{
    assert(mutex != NULL);
    if (!*mutex)
        return 0;
    {
        int const ret = pthread_mutex_destroy(*mutex);
        ZSTD_free(*mutex);
        return ret;
    }
}

int ZSTD_pthread_cond_init(ZSTD_pthread_cond_t* cond, pthread_condattr_t const* attr)
{
    assert(cond != NULL);
    *cond = (pthread_cond_t*)ZSTD_malloc(sizeof(pthread_cond_t));
    if (!*cond)
        return 1;
    return pthread_cond_init(*cond, attr);
}

int ZSTD_pthread_cond_destroy(ZSTD_pthread_cond_t* cond)
{
    assert(cond != NULL);
    if (!*cond)
        return 0;
    {
        int const ret = pthread_cond_destroy(*cond);
        ZSTD_free(*cond);
        return ret;
    }
}

#endif
/**** ended inlining common/threading.c ****/
/**** start inlining common/pool.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/* ======   Dependencies   ======= */
/**** start inlining ../common/allocations.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* This file provides custom allocation primitives
 */

#define ZSTD_DEPS_NEED_MALLOC
/**** skipping file: zstd_deps.h ****/

/**** skipping file: compiler.h ****/
#define ZSTD_STATIC_LINKING_ONLY
/**** *NOT* inlining ../zstd.h ****/
#include "zstd.h" /* ZSTD_customMem */

#ifndef ZSTD_ALLOCATIONS_H
#define ZSTD_ALLOCATIONS_H

/* custom memory allocation functions */

MEM_STATIC void* ZSTD_customMalloc(size_t size, ZSTD_customMem customMem)
{
    if (customMem.customAlloc)
        return customMem.customAlloc(customMem.opaque, size);
    return ZSTD_malloc(size);
}

MEM_STATIC void* ZSTD_customCalloc(size_t size, ZSTD_customMem customMem)
{
    if (customMem.customAlloc) {
        /* calloc implemented as malloc+memset;
         * not as efficient as calloc, but next best guess for custom malloc */
        void* const ptr = customMem.customAlloc(customMem.opaque, size);
        ZSTD_memset(ptr, 0, size);
        return ptr;
    }
    return ZSTD_calloc(1, size);
}

MEM_STATIC void ZSTD_customFree(void* ptr, ZSTD_customMem customMem)
{
    if (ptr!=NULL) {
        if (customMem.customFree)
            customMem.customFree(customMem.opaque, ptr);
        else
            ZSTD_free(ptr);
    }
}

#endif /* ZSTD_ALLOCATIONS_H */
/**** ended inlining ../common/allocations.h ****/
/**** skipping file: zstd_deps.h ****/
/**** skipping file: debug.h ****/
/**** start inlining pool.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef POOL_H
#define POOL_H


/**** skipping file: zstd_deps.h ****/
#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_customMem */
/**** skipping file: ../zstd.h ****/

typedef struct POOL_ctx_s POOL_ctx;

/*! POOL_create() :
 *  Create a thread pool with at most `numThreads` threads.
 * `numThreads` must be at least 1.
 *  The maximum number of queued jobs before blocking is `queueSize`.
 * @return : POOL_ctx pointer on success, else NULL.
*/
POOL_ctx* POOL_create(size_t numThreads, size_t queueSize);

POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
                               ZSTD_customMem customMem);

/*! POOL_free() :
 *  Free a thread pool returned by POOL_create().
 */
void POOL_free(POOL_ctx* ctx);


/*! POOL_joinJobs() :
 *  Waits for all queued jobs to finish executing.
 */
void POOL_joinJobs(POOL_ctx* ctx);

/*! POOL_resize() :
 *  Expands or shrinks pool's number of threads.
 *  This is more efficient than releasing + creating a new context,
 *  since it tries to preserve and reuse existing threads.
 * `numThreads` must be at least 1.
 * @return : 0 when resize was successful,
 *           !0 (typically 1) if there is an error.
 *    note : only numThreads can be resized, queueSize remains unchanged.
 */
int POOL_resize(POOL_ctx* ctx, size_t numThreads);

/*! POOL_sizeof() :
 * @return threadpool memory usage
 *  note : compatible with NULL (returns 0 in this case)
 */
size_t POOL_sizeof(const POOL_ctx* ctx);

/*! POOL_function :
 *  The function type that can be added to a thread pool.
 */
typedef void (*POOL_function)(void*);

/*! POOL_add() :
 *  Add the job `function(opaque)` to the thread pool. `ctx` must be valid.
 *  Possibly blocks until there is room in the queue.
 *  Note : The function may be executed asynchronously,
 *         therefore, `opaque` must live until function has been completed.
 */
void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque);


/*! POOL_tryAdd() :
 *  Add the job `function(opaque)` to thread pool _if_ a queue slot is available.
 *  Returns immediately even if not (does not block).
 * @return : 1 if successful, 0 if not.
 */
int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque);

#endif
/**** ended inlining pool.h ****/

/* ======   Compiler specifics   ====== */
#if defined(_MSC_VER)
#  pragma warning(disable : 4204)        /* disable: C4204: non-constant aggregate initializer */
#endif


#ifdef ZSTD_MULTITHREAD

/**** skipping file: threading.h ****/

/* A job is a function and an opaque argument */
typedef struct POOL_job_s {
    POOL_function function;
    void *opaque;
} POOL_job;

struct POOL_ctx_s {
    ZSTD_customMem customMem;
    /* Keep track of the threads */
    ZSTD_pthread_t* threads;
    size_t threadCapacity;
    size_t threadLimit;

    /* The queue is a circular buffer */
    POOL_job *queue;
    size_t queueHead;
    size_t queueTail;
    size_t queueSize;

    /* The number of threads working on jobs */
    size_t numThreadsBusy;
    /* Indicates if the queue is empty */
    int queueEmpty;

    /* The mutex protects the queue */
    ZSTD_pthread_mutex_t queueMutex;
    /* Condition variable for pushers to wait on when the queue is full */
    ZSTD_pthread_cond_t queuePushCond;
    /* Condition variables for poppers to wait on when the queue is empty */
    ZSTD_pthread_cond_t queuePopCond;
    /* Indicates if the queue is shutting down */
    int shutdown;
};

/* POOL_thread() :
 * Work thread for the thread pool.
 * Waits for jobs and executes them.
 * @returns : NULL on failure else non-null.
 */
static void* POOL_thread(void* opaque) {
    POOL_ctx* const ctx = (POOL_ctx*)opaque;
    if (!ctx) { return NULL; }
    for (;;) {
        /* Lock the mutex and wait for a non-empty queue or until shutdown */
        ZSTD_pthread_mutex_lock(&ctx->queueMutex);

        while ( ctx->queueEmpty
            || (ctx->numThreadsBusy >= ctx->threadLimit) ) {
            if (ctx->shutdown) {
                /* even if !queueEmpty, (possible if numThreadsBusy >= threadLimit),
                 * a few threads will be shutdown while !queueEmpty,
                 * but enough threads will remain active to finish the queue */
                ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
                return opaque;
            }
            ZSTD_pthread_cond_wait(&ctx->queuePopCond, &ctx->queueMutex);
        }
        /* Pop a job off the queue */
        {   POOL_job const job = ctx->queue[ctx->queueHead];
            ctx->queueHead = (ctx->queueHead + 1) % ctx->queueSize;
            ctx->numThreadsBusy++;
            ctx->queueEmpty = (ctx->queueHead == ctx->queueTail);
            /* Unlock the mutex, signal a pusher, and run the job */
            ZSTD_pthread_cond_signal(&ctx->queuePushCond);
            ZSTD_pthread_mutex_unlock(&ctx->queueMutex);

            job.function(job.opaque);

            /* If the intended queue size was 0, signal after finishing job */
            ZSTD_pthread_mutex_lock(&ctx->queueMutex);
            ctx->numThreadsBusy--;
            ZSTD_pthread_cond_signal(&ctx->queuePushCond);
            ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
        }
    }  /* for (;;) */
    assert(0);  /* Unreachable */
}

/* ZSTD_createThreadPool() : public access point */
POOL_ctx* ZSTD_createThreadPool(size_t numThreads) {
    return POOL_create (numThreads, 0);
}

POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
    return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
}

POOL_ctx* POOL_create_advanced(size_t numThreads, size_t queueSize,
                               ZSTD_customMem customMem)
{
    POOL_ctx* ctx;
    /* Check parameters */
    if (!numThreads) { return NULL; }
    /* Allocate the context and zero initialize */
    ctx = (POOL_ctx*)ZSTD_customCalloc(sizeof(POOL_ctx), customMem);
    if (!ctx) { return NULL; }
    /* Initialize the job queue.
     * It needs one extra space since one space is wasted to differentiate
     * empty and full queues.
     */
    ctx->queueSize = queueSize + 1;
    ctx->queue = (POOL_job*)ZSTD_customCalloc(ctx->queueSize * sizeof(POOL_job), customMem);
    ctx->queueHead = 0;
    ctx->queueTail = 0;
    ctx->numThreadsBusy = 0;
    ctx->queueEmpty = 1;
    {
        int error = 0;
        error |= ZSTD_pthread_mutex_init(&ctx->queueMutex, NULL);
        error |= ZSTD_pthread_cond_init(&ctx->queuePushCond, NULL);
        error |= ZSTD_pthread_cond_init(&ctx->queuePopCond, NULL);
        if (error) { POOL_free(ctx); return NULL; }
    }
    ctx->shutdown = 0;
    /* Allocate space for the thread handles */
    ctx->threads = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), customMem);
    ctx->threadCapacity = 0;
    ctx->customMem = customMem;
    /* Check for errors */
    if (!ctx->threads || !ctx->queue) { POOL_free(ctx); return NULL; }
    /* Initialize the threads */
    {   size_t i;
        for (i = 0; i < numThreads; ++i) {
            if (ZSTD_pthread_create(&ctx->threads[i], NULL, &POOL_thread, ctx)) {
                ctx->threadCapacity = i;
                POOL_free(ctx);
                return NULL;
        }   }
        ctx->threadCapacity = numThreads;
        ctx->threadLimit = numThreads;
    }
    return ctx;
}

/*! POOL_join() :
    Shutdown the queue, wake any sleeping threads, and join all of the threads.
*/
static void POOL_join(POOL_ctx* ctx) {
    /* Shut down the queue */
    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
    ctx->shutdown = 1;
    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
    /* Wake up sleeping threads */
    ZSTD_pthread_cond_broadcast(&ctx->queuePushCond);
    ZSTD_pthread_cond_broadcast(&ctx->queuePopCond);
    /* Join all of the threads */
    {   size_t i;
        for (i = 0; i < ctx->threadCapacity; ++i) {
            ZSTD_pthread_join(ctx->threads[i]);  /* note : could fail */
    }   }
}

void POOL_free(POOL_ctx *ctx) {
    if (!ctx) { return; }
    POOL_join(ctx);
    ZSTD_pthread_mutex_destroy(&ctx->queueMutex);
    ZSTD_pthread_cond_destroy(&ctx->queuePushCond);
    ZSTD_pthread_cond_destroy(&ctx->queuePopCond);
    ZSTD_customFree(ctx->queue, ctx->customMem);
    ZSTD_customFree(ctx->threads, ctx->customMem);
    ZSTD_customFree(ctx, ctx->customMem);
}

/*! POOL_joinJobs() :
 *  Waits for all queued jobs to finish executing.
 */
void POOL_joinJobs(POOL_ctx* ctx) {
    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
    while(!ctx->queueEmpty || ctx->numThreadsBusy > 0) {
        ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
    }
    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
}

void ZSTD_freeThreadPool (ZSTD_threadPool* pool) {
  POOL_free (pool);
}

size_t POOL_sizeof(const POOL_ctx* ctx) {
    if (ctx==NULL) return 0;  /* supports sizeof NULL */
    return sizeof(*ctx)
        + ctx->queueSize * sizeof(POOL_job)
        + ctx->threadCapacity * sizeof(ZSTD_pthread_t);
}


/* @return : 0 on success, 1 on error */
static int POOL_resize_internal(POOL_ctx* ctx, size_t numThreads)
{
    if (numThreads <= ctx->threadCapacity) {
        if (!numThreads) return 1;
        ctx->threadLimit = numThreads;
        return 0;
    }
    /* numThreads > threadCapacity */
    {   ZSTD_pthread_t* const threadPool = (ZSTD_pthread_t*)ZSTD_customCalloc(numThreads * sizeof(ZSTD_pthread_t), ctx->customMem);
        if (!threadPool) return 1;
        /* replace existing thread pool */
        ZSTD_memcpy(threadPool, ctx->threads, ctx->threadCapacity * sizeof(ZSTD_pthread_t));
        ZSTD_customFree(ctx->threads, ctx->customMem);
        ctx->threads = threadPool;
        /* Initialize additional threads */
        {   size_t threadId;
            for (threadId = ctx->threadCapacity; threadId < numThreads; ++threadId) {
                if (ZSTD_pthread_create(&threadPool[threadId], NULL, &POOL_thread, ctx)) {
                    ctx->threadCapacity = threadId;
                    return 1;
            }   }
    }   }
    /* successfully expanded */
    ctx->threadCapacity = numThreads;
    ctx->threadLimit = numThreads;
    return 0;
}

/* @return : 0 on success, 1 on error */
int POOL_resize(POOL_ctx* ctx, size_t numThreads)
{
    int result;
    if (ctx==NULL) return 1;
    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
    result = POOL_resize_internal(ctx, numThreads);
    ZSTD_pthread_cond_broadcast(&ctx->queuePopCond);
    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
    return result;
}

/**
 * Returns 1 if the queue is full and 0 otherwise.
 *
 * When queueSize is 1 (pool was created with an intended queueSize of 0),
 * then a queue is empty if there is a thread free _and_ no job is waiting.
 */
static int isQueueFull(POOL_ctx const* ctx) {
    if (ctx->queueSize > 1) {
        return ctx->queueHead == ((ctx->queueTail + 1) % ctx->queueSize);
    } else {
        return (ctx->numThreadsBusy == ctx->threadLimit) ||
               !ctx->queueEmpty;
    }
}


static void
POOL_add_internal(POOL_ctx* ctx, POOL_function function, void *opaque)
{
    POOL_job job;
    job.function = function;
    job.opaque = opaque;
    assert(ctx != NULL);
    if (ctx->shutdown) return;

    ctx->queueEmpty = 0;
    ctx->queue[ctx->queueTail] = job;
    ctx->queueTail = (ctx->queueTail + 1) % ctx->queueSize;
    ZSTD_pthread_cond_signal(&ctx->queuePopCond);
}

void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque)
{
    assert(ctx != NULL);
    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
    /* Wait until there is space in the queue for the new job */
    while (isQueueFull(ctx) && (!ctx->shutdown)) {
        ZSTD_pthread_cond_wait(&ctx->queuePushCond, &ctx->queueMutex);
    }
    POOL_add_internal(ctx, function, opaque);
    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
}


int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque)
{
    assert(ctx != NULL);
    ZSTD_pthread_mutex_lock(&ctx->queueMutex);
    if (isQueueFull(ctx)) {
        ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
        return 0;
    }
    POOL_add_internal(ctx, function, opaque);
    ZSTD_pthread_mutex_unlock(&ctx->queueMutex);
    return 1;
}


#else  /* ZSTD_MULTITHREAD  not defined */

/* ========================== */
/* No multi-threading support */
/* ========================== */


/* We don't need any data, but if it is empty, malloc() might return NULL. */
struct POOL_ctx_s {
    int dummy;
};
static POOL_ctx g_poolCtx;

POOL_ctx* POOL_create(size_t numThreads, size_t queueSize) {
    return POOL_create_advanced(numThreads, queueSize, ZSTD_defaultCMem);
}

POOL_ctx*
POOL_create_advanced(size_t numThreads, size_t queueSize, ZSTD_customMem customMem)
{
    (void)numThreads;
    (void)queueSize;
    (void)customMem;
    return &g_poolCtx;
}

void POOL_free(POOL_ctx* ctx) {
    assert(!ctx || ctx == &g_poolCtx);
    (void)ctx;
}

void POOL_joinJobs(POOL_ctx* ctx){
    assert(!ctx || ctx == &g_poolCtx);
    (void)ctx;
}

int POOL_resize(POOL_ctx* ctx, size_t numThreads) {
    (void)ctx; (void)numThreads;
    return 0;
}

void POOL_add(POOL_ctx* ctx, POOL_function function, void* opaque) {
    (void)ctx;
    function(opaque);
}

int POOL_tryAdd(POOL_ctx* ctx, POOL_function function, void* opaque) {
    (void)ctx;
    function(opaque);
    return 1;
}

size_t POOL_sizeof(const POOL_ctx* ctx) {
    if (ctx==NULL) return 0;  /* supports sizeof NULL */
    assert(ctx == &g_poolCtx);
    return sizeof(*ctx);
}

#endif  /* ZSTD_MULTITHREAD */
/**** ended inlining common/pool.c ****/
/**** start inlining common/zstd_common.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/*-*************************************
*  Dependencies
***************************************/
#define ZSTD_DEPS_NEED_MALLOC
/**** skipping file: error_private.h ****/
/**** start inlining zstd_internal.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_CCOMMON_H_MODULE
#define ZSTD_CCOMMON_H_MODULE

/* this module contains definitions which must be identical
 * across compression, decompression and dictBuilder.
 * It also contains a few functions useful to at least 2 of them
 * and which benefit from being inlined */

/*-*************************************
*  Dependencies
***************************************/
/**** skipping file: compiler.h ****/
/**** start inlining cpu.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_COMMON_CPU_H
#define ZSTD_COMMON_CPU_H

/**
 * Implementation taken from folly/CpuId.h
 * https://github.com/facebook/folly/blob/master/folly/CpuId.h
 */

/**** skipping file: mem.h ****/

#ifdef _MSC_VER
#include <intrin.h>
#endif

typedef struct {
    U32 f1c;
    U32 f1d;
    U32 f7b;
    U32 f7c;
} ZSTD_cpuid_t;

MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) {
    U32 f1c = 0;
    U32 f1d = 0;
    U32 f7b = 0;
    U32 f7c = 0;
#if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))
#if !defined(_M_X64) || !defined(__clang__) || __clang_major__ >= 16
    int reg[4];
    __cpuid((int*)reg, 0);
    {
        int const n = reg[0];
        if (n >= 1) {
            __cpuid((int*)reg, 1);
            f1c = (U32)reg[2];
            f1d = (U32)reg[3];
        }
        if (n >= 7) {
            __cpuidex((int*)reg, 7, 0);
            f7b = (U32)reg[1];
            f7c = (U32)reg[2];
        }
    }
#else
    /* Clang compiler has a bug (fixed in https://reviews.llvm.org/D101338) in
     * which the `__cpuid` intrinsic does not save and restore `rbx` as it needs
     * to due to being a reserved register. So in that case, do the `cpuid`
     * ourselves. Clang supports inline assembly anyway.
     */
    U32 n;
    __asm__(
        "pushq %%rbx\n\t"
        "cpuid\n\t"
        "popq %%rbx\n\t"
        : "=a"(n)
        : "a"(0)
        : "rcx", "rdx");
    if (n >= 1) {
      U32 f1a;
      __asm__(
          "pushq %%rbx\n\t"
          "cpuid\n\t"
          "popq %%rbx\n\t"
          : "=a"(f1a), "=c"(f1c), "=d"(f1d)
          : "a"(1)
          :);
    }
    if (n >= 7) {
      __asm__(
          "pushq %%rbx\n\t"
          "cpuid\n\t"
          "movq %%rbx, %%rax\n\t"
          "popq %%rbx"
          : "=a"(f7b), "=c"(f7c)
          : "a"(7), "c"(0)
          : "rdx");
    }
#endif
#elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__)
    /* The following block like the normal cpuid branch below, but gcc
     * reserves ebx for use of its pic register so we must specially
     * handle the save and restore to avoid clobbering the register
     */
    U32 n;
    __asm__(
        "pushl %%ebx\n\t"
        "cpuid\n\t"
        "popl %%ebx\n\t"
        : "=a"(n)
        : "a"(0)
        : "ecx", "edx");
    if (n >= 1) {
      U32 f1a;
      __asm__(
          "pushl %%ebx\n\t"
          "cpuid\n\t"
          "popl %%ebx\n\t"
          : "=a"(f1a), "=c"(f1c), "=d"(f1d)
          : "a"(1));
    }
    if (n >= 7) {
      __asm__(
          "pushl %%ebx\n\t"
          "cpuid\n\t"
          "movl %%ebx, %%eax\n\t"
          "popl %%ebx"
          : "=a"(f7b), "=c"(f7c)
          : "a"(7), "c"(0)
          : "edx");
    }
#elif defined(__x86_64__) || defined(_M_X64) || defined(__i386__)
    U32 n;
    __asm__("cpuid" : "=a"(n) : "a"(0) : "ebx", "ecx", "edx");
    if (n >= 1) {
      U32 f1a;
      __asm__("cpuid" : "=a"(f1a), "=c"(f1c), "=d"(f1d) : "a"(1) : "ebx");
    }
    if (n >= 7) {
      U32 f7a;
      __asm__("cpuid"
              : "=a"(f7a), "=b"(f7b), "=c"(f7c)
              : "a"(7), "c"(0)
              : "edx");
    }
#endif
    {
        ZSTD_cpuid_t cpuid;
        cpuid.f1c = f1c;
        cpuid.f1d = f1d;
        cpuid.f7b = f7b;
        cpuid.f7c = f7c;
        return cpuid;
    }
}

#define X(name, r, bit)                                                        \
  MEM_STATIC int ZSTD_cpuid_##name(ZSTD_cpuid_t const cpuid) {                 \
    return ((cpuid.r) & (1U << bit)) != 0;                                     \
  }

/* cpuid(1): Processor Info and Feature Bits. */
#define C(name, bit) X(name, f1c, bit)
  C(sse3, 0)
  C(pclmuldq, 1)
  C(dtes64, 2)
  C(monitor, 3)
  C(dscpl, 4)
  C(vmx, 5)
  C(smx, 6)
  C(eist, 7)
  C(tm2, 8)
  C(ssse3, 9)
  C(cnxtid, 10)
  C(fma, 12)
  C(cx16, 13)
  C(xtpr, 14)
  C(pdcm, 15)
  C(pcid, 17)
  C(dca, 18)
  C(sse41, 19)
  C(sse42, 20)
  C(x2apic, 21)
  C(movbe, 22)
  C(popcnt, 23)
  C(tscdeadline, 24)
  C(aes, 25)
  C(xsave, 26)
  C(osxsave, 27)
  C(avx, 28)
  C(f16c, 29)
  C(rdrand, 30)
#undef C
#define D(name, bit) X(name, f1d, bit)
  D(fpu, 0)
  D(vme, 1)
  D(de, 2)
  D(pse, 3)
  D(tsc, 4)
  D(msr, 5)
  D(pae, 6)
  D(mce, 7)
  D(cx8, 8)
  D(apic, 9)
  D(sep, 11)
  D(mtrr, 12)
  D(pge, 13)
  D(mca, 14)
  D(cmov, 15)
  D(pat, 16)
  D(pse36, 17)
  D(psn, 18)
  D(clfsh, 19)
  D(ds, 21)
  D(acpi, 22)
  D(mmx, 23)
  D(fxsr, 24)
  D(sse, 25)
  D(sse2, 26)
  D(ss, 27)
  D(htt, 28)
  D(tm, 29)
  D(pbe, 31)
#undef D

/* cpuid(7): Extended Features. */
#define B(name, bit) X(name, f7b, bit)
  B(bmi1, 3)
  B(hle, 4)
  B(avx2, 5)
  B(smep, 7)
  B(bmi2, 8)
  B(erms, 9)
  B(invpcid, 10)
  B(rtm, 11)
  B(mpx, 14)
  B(avx512f, 16)
  B(avx512dq, 17)
  B(rdseed, 18)
  B(adx, 19)
  B(smap, 20)
  B(avx512ifma, 21)
  B(pcommit, 22)
  B(clflushopt, 23)
  B(clwb, 24)
  B(avx512pf, 26)
  B(avx512er, 27)
  B(avx512cd, 28)
  B(sha, 29)
  B(avx512bw, 30)
  B(avx512vl, 31)
#undef B
#define C(name, bit) X(name, f7c, bit)
  C(prefetchwt1, 0)
  C(avx512vbmi, 1)
#undef C

#undef X

#endif /* ZSTD_COMMON_CPU_H */
/**** ended inlining cpu.h ****/
/**** skipping file: mem.h ****/
/**** skipping file: debug.h ****/
/**** skipping file: error_private.h ****/
#define ZSTD_STATIC_LINKING_ONLY
/**** skipping file: ../zstd.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: fse.h ****/
/**** skipping file: huf.h ****/
#ifndef XXH_STATIC_LINKING_ONLY
#  define XXH_STATIC_LINKING_ONLY  /* XXH64_state_t */
#endif
/**** start inlining xxhash.h ****/
/*
 * xxHash - Extremely Fast Hash algorithm
 * Header File
 * Copyright (c) Yann Collet - Meta Platforms, Inc
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* Local adaptations for Zstandard */

#ifndef XXH_NO_XXH3
# define XXH_NO_XXH3
#endif

#ifndef XXH_NAMESPACE
# define XXH_NAMESPACE ZSTD_
#endif

/*!
 * @mainpage xxHash
 *
 * xxHash is an extremely fast non-cryptographic hash algorithm, working at RAM speed
 * limits.
 *
 * It is proposed in four flavors, in three families:
 * 1. @ref XXH32_family
 *   - Classic 32-bit hash function. Simple, compact, and runs on almost all
 *     32-bit and 64-bit systems.
 * 2. @ref XXH64_family
 *   - Classic 64-bit adaptation of XXH32. Just as simple, and runs well on most
 *     64-bit systems (but _not_ 32-bit systems).
 * 3. @ref XXH3_family
 *   - Modern 64-bit and 128-bit hash function family which features improved
 *     strength and performance across the board, especially on smaller data.
 *     It benefits greatly from SIMD and 64-bit without requiring it.
 *
 * Benchmarks
 * ---
 * The reference system uses an Intel i7-9700K CPU, and runs Ubuntu x64 20.04.
 * The open source benchmark program is compiled with clang v10.0 using -O3 flag.
 *
 * | Hash Name            | ISA ext | Width | Large Data Speed | Small Data Velocity |
 * | -------------------- | ------- | ----: | ---------------: | ------------------: |
 * | XXH3_64bits()        | @b AVX2 |    64 |        59.4 GB/s |               133.1 |
 * | MeowHash             | AES-NI  |   128 |        58.2 GB/s |                52.5 |
 * | XXH3_128bits()       | @b AVX2 |   128 |        57.9 GB/s |               118.1 |
 * | CLHash               | PCLMUL  |    64 |        37.1 GB/s |                58.1 |
 * | XXH3_64bits()        | @b SSE2 |    64 |        31.5 GB/s |               133.1 |
 * | XXH3_128bits()       | @b SSE2 |   128 |        29.6 GB/s |               118.1 |
 * | RAM sequential read  |         |   N/A |        28.0 GB/s |                 N/A |
 * | ahash                | AES-NI  |    64 |        22.5 GB/s |               107.2 |
 * | City64               |         |    64 |        22.0 GB/s |                76.6 |
 * | T1ha2                |         |    64 |        22.0 GB/s |                99.0 |
 * | City128              |         |   128 |        21.7 GB/s |                57.7 |
 * | FarmHash             | AES-NI  |    64 |        21.3 GB/s |                71.9 |
 * | XXH64()              |         |    64 |        19.4 GB/s |                71.0 |
 * | SpookyHash           |         |    64 |        19.3 GB/s |                53.2 |
 * | Mum                  |         |    64 |        18.0 GB/s |                67.0 |
 * | CRC32C               | SSE4.2  |    32 |        13.0 GB/s |                57.9 |
 * | XXH32()              |         |    32 |         9.7 GB/s |                71.9 |
 * | City32               |         |    32 |         9.1 GB/s |                66.0 |
 * | Blake3*              | @b AVX2 |   256 |         4.4 GB/s |                 8.1 |
 * | Murmur3              |         |    32 |         3.9 GB/s |                56.1 |
 * | SipHash*             |         |    64 |         3.0 GB/s |                43.2 |
 * | Blake3*              | @b SSE2 |   256 |         2.4 GB/s |                 8.1 |
 * | HighwayHash          |         |    64 |         1.4 GB/s |                 6.0 |
 * | FNV64                |         |    64 |         1.2 GB/s |                62.7 |
 * | Blake2*              |         |   256 |         1.1 GB/s |                 5.1 |
 * | SHA1*                |         |   160 |         0.8 GB/s |                 5.6 |
 * | MD5*                 |         |   128 |         0.6 GB/s |                 7.8 |
 * @note
 *   - Hashes which require a specific ISA extension are noted. SSE2 is also noted,
 *     even though it is mandatory on x64.
 *   - Hashes with an asterisk are cryptographic. Note that MD5 is non-cryptographic
 *     by modern standards.
 *   - Small data velocity is a rough average of algorithm's efficiency for small
 *     data. For more accurate information, see the wiki.
 *   - More benchmarks and strength tests are found on the wiki:
 *         https://github.com/Cyan4973/xxHash/wiki
 *
 * Usage
 * ------
 * All xxHash variants use a similar API. Changing the algorithm is a trivial
 * substitution.
 *
 * @pre
 *    For functions which take an input and length parameter, the following
 *    requirements are assumed:
 *    - The range from [`input`, `input + length`) is valid, readable memory.
 *      - The only exception is if the `length` is `0`, `input` may be `NULL`.
 *    - For C++, the objects must have the *TriviallyCopyable* property, as the
 *      functions access bytes directly as if it was an array of `unsigned char`.
 *
 * @anchor single_shot_example
 * **Single Shot**
 *
 * These functions are stateless functions which hash a contiguous block of memory,
 * immediately returning the result. They are the easiest and usually the fastest
 * option.
 *
 * XXH32(), XXH64(), XXH3_64bits(), XXH3_128bits()
 *
 * @code{.c}
 *   #include <string.h>
 *   #include "xxhash.h"
 *
 *   // Example for a function which hashes a null terminated string with XXH32().
 *   XXH32_hash_t hash_string(const char* string, XXH32_hash_t seed)
 *   {
 *       // NULL pointers are only valid if the length is zero
 *       size_t length = (string == NULL) ? 0 : strlen(string);
 *       return XXH32(string, length, seed);
 *   }
 * @endcode
 *
 *
 * @anchor streaming_example
 * **Streaming**
 *
 * These groups of functions allow incremental hashing of unknown size, even
 * more than what would fit in a size_t.
 *
 * XXH32_reset(), XXH64_reset(), XXH3_64bits_reset(), XXH3_128bits_reset()
 *
 * @code{.c}
 *   #include <stdio.h>
 *   #include <assert.h>
 *   #include "xxhash.h"
 *   // Example for a function which hashes a FILE incrementally with XXH3_64bits().
 *   XXH64_hash_t hashFile(FILE* f)
 *   {
 *       // Allocate a state struct. Do not just use malloc() or new.
 *       XXH3_state_t* state = XXH3_createState();
 *       assert(state != NULL && "Out of memory!");
 *       // Reset the state to start a new hashing session.
 *       XXH3_64bits_reset(state);
 *       char buffer[4096];
 *       size_t count;
 *       // Read the file in chunks
 *       while ((count = fread(buffer, 1, sizeof(buffer), f)) != 0) {
 *           // Run update() as many times as necessary to process the data
 *           XXH3_64bits_update(state, buffer, count);
 *       }
 *       // Retrieve the finalized hash. This will not change the state.
 *       XXH64_hash_t result = XXH3_64bits_digest(state);
 *       // Free the state. Do not use free().
 *       XXH3_freeState(state);
 *       return result;
 *   }
 * @endcode
 *
 * Streaming functions generate the xxHash value from an incremental input.
 * This method is slower than single-call functions, due to state management.
 * For small inputs, prefer `XXH32()` and `XXH64()`, which are better optimized.
 *
 * An XXH state must first be allocated using `XXH*_createState()`.
 *
 * Start a new hash by initializing the state with a seed using `XXH*_reset()`.
 *
 * Then, feed the hash state by calling `XXH*_update()` as many times as necessary.
 *
 * The function returns an error code, with 0 meaning OK, and any other value
 * meaning there is an error.
 *
 * Finally, a hash value can be produced anytime, by using `XXH*_digest()`.
 * This function returns the nn-bits hash as an int or long long.
 *
 * It's still possible to continue inserting input into the hash state after a
 * digest, and generate new hash values later on by invoking `XXH*_digest()`.
 *
 * When done, release the state using `XXH*_freeState()`.
 *
 *
 * @anchor canonical_representation_example
 * **Canonical Representation**
 *
 * The default return values from XXH functions are unsigned 32, 64 and 128 bit
 * integers.
 * This the simplest and fastest format for further post-processing.
 *
 * However, this leaves open the question of what is the order on the byte level,
 * since little and big endian conventions will store the same number differently.
 *
 * The canonical representation settles this issue by mandating big-endian
 * convention, the same convention as human-readable numbers (large digits first).
 *
 * When writing hash values to storage, sending them over a network, or printing
 * them, it's highly recommended to use the canonical representation to ensure
 * portability across a wider range of systems, present and future.
 *
 * The following functions allow transformation of hash values to and from
 * canonical format.
 *
 * XXH32_canonicalFromHash(), XXH32_hashFromCanonical(),
 * XXH64_canonicalFromHash(), XXH64_hashFromCanonical(),
 * XXH128_canonicalFromHash(), XXH128_hashFromCanonical(),
 *
 * @code{.c}
 *   #include <stdio.h>
 *   #include "xxhash.h"
 *
 *   // Example for a function which prints XXH32_hash_t in human readable format
 *   void printXxh32(XXH32_hash_t hash)
 *   {
 *       XXH32_canonical_t cano;
 *       XXH32_canonicalFromHash(&cano, hash);
 *       size_t i;
 *       for(i = 0; i < sizeof(cano.digest); ++i) {
 *           printf("%02x", cano.digest[i]);
 *       }
 *       printf("\n");
 *   }
 *
 *   // Example for a function which converts XXH32_canonical_t to XXH32_hash_t
 *   XXH32_hash_t convertCanonicalToXxh32(XXH32_canonical_t cano)
 *   {
 *       XXH32_hash_t hash = XXH32_hashFromCanonical(&cano);
 *       return hash;
 *   }
 * @endcode
 *
 *
 * @file xxhash.h
 * xxHash prototypes and implementation
 */

/* ****************************
 *  INLINE mode
 ******************************/
/*!
 * @defgroup public Public API
 * Contains details on the public xxHash functions.
 * @{
 */
#ifdef XXH_DOXYGEN
/*!
 * @brief Gives access to internal state declaration, required for static allocation.
 *
 * Incompatible with dynamic linking, due to risks of ABI changes.
 *
 * Usage:
 * @code{.c}
 *     #define XXH_STATIC_LINKING_ONLY
 *     #include "xxhash.h"
 * @endcode
 */
#  define XXH_STATIC_LINKING_ONLY
/* Do not undef XXH_STATIC_LINKING_ONLY for Doxygen */

/*!
 * @brief Gives access to internal definitions.
 *
 * Usage:
 * @code{.c}
 *     #define XXH_STATIC_LINKING_ONLY
 *     #define XXH_IMPLEMENTATION
 *     #include "xxhash.h"
 * @endcode
 */
#  define XXH_IMPLEMENTATION
/* Do not undef XXH_IMPLEMENTATION for Doxygen */

/*!
 * @brief Exposes the implementation and marks all functions as `inline`.
 *
 * Use these build macros to inline xxhash into the target unit.
 * Inlining improves performance on small inputs, especially when the length is
 * expressed as a compile-time constant:
 *
 *  https://fastcompression.blogspot.com/2018/03/xxhash-for-small-keys-impressive-power.html
 *
 * It also keeps xxHash symbols private to the unit, so they are not exported.
 *
 * Usage:
 * @code{.c}
 *     #define XXH_INLINE_ALL
 *     #include "xxhash.h"
 * @endcode
 * Do not compile and link xxhash.o as a separate object, as it is not useful.
 */
#  define XXH_INLINE_ALL
#  undef XXH_INLINE_ALL
/*!
 * @brief Exposes the implementation without marking functions as inline.
 */
#  define XXH_PRIVATE_API
#  undef XXH_PRIVATE_API
/*!
 * @brief Emulate a namespace by transparently prefixing all symbols.
 *
 * If you want to include _and expose_ xxHash functions from within your own
 * library, but also want to avoid symbol collisions with other libraries which
 * may also include xxHash, you can use @ref XXH_NAMESPACE to automatically prefix
 * any public symbol from xxhash library with the value of @ref XXH_NAMESPACE
 * (therefore, avoid empty or numeric values).
 *
 * Note that no change is required within the calling program as long as it
 * includes `xxhash.h`: Regular symbol names will be automatically translated
 * by this header.
 */
#  define XXH_NAMESPACE /* YOUR NAME HERE */
#  undef XXH_NAMESPACE
#endif

#if (defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)) \
    && !defined(XXH_INLINE_ALL_31684351384)
   /* this section should be traversed only once */
#  define XXH_INLINE_ALL_31684351384
   /* give access to the advanced API, required to compile implementations */
#  undef XXH_STATIC_LINKING_ONLY   /* avoid macro redef */
#  define XXH_STATIC_LINKING_ONLY
   /* make all functions private */
#  undef XXH_PUBLIC_API
#  if defined(__GNUC__)
#    define XXH_PUBLIC_API static __inline __attribute__((unused))
#  elif defined (__cplusplus) || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */)
#    define XXH_PUBLIC_API static inline
#  elif defined(_MSC_VER)
#    define XXH_PUBLIC_API static __inline
#  else
     /* note: this version may generate warnings for unused static functions */
#    define XXH_PUBLIC_API static
#  endif

   /*
    * This part deals with the special case where a unit wants to inline xxHash,
    * but "xxhash.h" has previously been included without XXH_INLINE_ALL,
    * such as part of some previously included *.h header file.
    * Without further action, the new include would just be ignored,
    * and functions would effectively _not_ be inlined (silent failure).
    * The following macros solve this situation by prefixing all inlined names,
    * avoiding naming collision with previous inclusions.
    */
   /* Before that, we unconditionally #undef all symbols,
    * in case they were already defined with XXH_NAMESPACE.
    * They will then be redefined for XXH_INLINE_ALL
    */
#  undef XXH_versionNumber
    /* XXH32 */
#  undef XXH32
#  undef XXH32_createState
#  undef XXH32_freeState
#  undef XXH32_reset
#  undef XXH32_update
#  undef XXH32_digest
#  undef XXH32_copyState
#  undef XXH32_canonicalFromHash
#  undef XXH32_hashFromCanonical
    /* XXH64 */
#  undef XXH64
#  undef XXH64_createState
#  undef XXH64_freeState
#  undef XXH64_reset
#  undef XXH64_update
#  undef XXH64_digest
#  undef XXH64_copyState
#  undef XXH64_canonicalFromHash
#  undef XXH64_hashFromCanonical
    /* XXH3_64bits */
#  undef XXH3_64bits
#  undef XXH3_64bits_withSecret
#  undef XXH3_64bits_withSeed
#  undef XXH3_64bits_withSecretandSeed
#  undef XXH3_createState
#  undef XXH3_freeState
#  undef XXH3_copyState
#  undef XXH3_64bits_reset
#  undef XXH3_64bits_reset_withSeed
#  undef XXH3_64bits_reset_withSecret
#  undef XXH3_64bits_update
#  undef XXH3_64bits_digest
#  undef XXH3_generateSecret
    /* XXH3_128bits */
#  undef XXH128
#  undef XXH3_128bits
#  undef XXH3_128bits_withSeed
#  undef XXH3_128bits_withSecret
#  undef XXH3_128bits_reset
#  undef XXH3_128bits_reset_withSeed
#  undef XXH3_128bits_reset_withSecret
#  undef XXH3_128bits_reset_withSecretandSeed
#  undef XXH3_128bits_update
#  undef XXH3_128bits_digest
#  undef XXH128_isEqual
#  undef XXH128_cmp
#  undef XXH128_canonicalFromHash
#  undef XXH128_hashFromCanonical
    /* Finally, free the namespace itself */
#  undef XXH_NAMESPACE

    /* employ the namespace for XXH_INLINE_ALL */
#  define XXH_NAMESPACE XXH_INLINE_
   /*
    * Some identifiers (enums, type names) are not symbols,
    * but they must nonetheless be renamed to avoid redeclaration.
    * Alternative solution: do not redeclare them.
    * However, this requires some #ifdefs, and has a more dispersed impact.
    * Meanwhile, renaming can be achieved in a single place.
    */
#  define XXH_IPREF(Id)   XXH_NAMESPACE ## Id
#  define XXH_OK XXH_IPREF(XXH_OK)
#  define XXH_ERROR XXH_IPREF(XXH_ERROR)
#  define XXH_errorcode XXH_IPREF(XXH_errorcode)
#  define XXH32_canonical_t  XXH_IPREF(XXH32_canonical_t)
#  define XXH64_canonical_t  XXH_IPREF(XXH64_canonical_t)
#  define XXH128_canonical_t XXH_IPREF(XXH128_canonical_t)
#  define XXH32_state_s XXH_IPREF(XXH32_state_s)
#  define XXH32_state_t XXH_IPREF(XXH32_state_t)
#  define XXH64_state_s XXH_IPREF(XXH64_state_s)
#  define XXH64_state_t XXH_IPREF(XXH64_state_t)
#  define XXH3_state_s  XXH_IPREF(XXH3_state_s)
#  define XXH3_state_t  XXH_IPREF(XXH3_state_t)
#  define XXH128_hash_t XXH_IPREF(XXH128_hash_t)
   /* Ensure the header is parsed again, even if it was previously included */
#  undef XXHASH_H_5627135585666179
#  undef XXHASH_H_STATIC_13879238742
#endif /* XXH_INLINE_ALL || XXH_PRIVATE_API */

/* ****************************************************************
 *  Stable API
 *****************************************************************/
#ifndef XXHASH_H_5627135585666179
#define XXHASH_H_5627135585666179 1

/*! @brief Marks a global symbol. */
#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
#    ifdef XXH_EXPORT
#      define XXH_PUBLIC_API __declspec(dllexport)
#    elif XXH_IMPORT
#      define XXH_PUBLIC_API __declspec(dllimport)
#    endif
#  else
#    define XXH_PUBLIC_API   /* do nothing */
#  endif
#endif

#ifdef XXH_NAMESPACE
#  define XXH_CAT(A,B) A##B
#  define XXH_NAME2(A,B) XXH_CAT(A,B)
#  define XXH_versionNumber XXH_NAME2(XXH_NAMESPACE, XXH_versionNumber)
/* XXH32 */
#  define XXH32 XXH_NAME2(XXH_NAMESPACE, XXH32)
#  define XXH32_createState XXH_NAME2(XXH_NAMESPACE, XXH32_createState)
#  define XXH32_freeState XXH_NAME2(XXH_NAMESPACE, XXH32_freeState)
#  define XXH32_reset XXH_NAME2(XXH_NAMESPACE, XXH32_reset)
#  define XXH32_update XXH_NAME2(XXH_NAMESPACE, XXH32_update)
#  define XXH32_digest XXH_NAME2(XXH_NAMESPACE, XXH32_digest)
#  define XXH32_copyState XXH_NAME2(XXH_NAMESPACE, XXH32_copyState)
#  define XXH32_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH32_canonicalFromHash)
#  define XXH32_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH32_hashFromCanonical)
/* XXH64 */
#  define XXH64 XXH_NAME2(XXH_NAMESPACE, XXH64)
#  define XXH64_createState XXH_NAME2(XXH_NAMESPACE, XXH64_createState)
#  define XXH64_freeState XXH_NAME2(XXH_NAMESPACE, XXH64_freeState)
#  define XXH64_reset XXH_NAME2(XXH_NAMESPACE, XXH64_reset)
#  define XXH64_update XXH_NAME2(XXH_NAMESPACE, XXH64_update)
#  define XXH64_digest XXH_NAME2(XXH_NAMESPACE, XXH64_digest)
#  define XXH64_copyState XXH_NAME2(XXH_NAMESPACE, XXH64_copyState)
#  define XXH64_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH64_canonicalFromHash)
#  define XXH64_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH64_hashFromCanonical)
/* XXH3_64bits */
#  define XXH3_64bits XXH_NAME2(XXH_NAMESPACE, XXH3_64bits)
#  define XXH3_64bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecret)
#  define XXH3_64bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSeed)
#  define XXH3_64bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_withSecretandSeed)
#  define XXH3_createState XXH_NAME2(XXH_NAMESPACE, XXH3_createState)
#  define XXH3_freeState XXH_NAME2(XXH_NAMESPACE, XXH3_freeState)
#  define XXH3_copyState XXH_NAME2(XXH_NAMESPACE, XXH3_copyState)
#  define XXH3_64bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset)
#  define XXH3_64bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSeed)
#  define XXH3_64bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecret)
#  define XXH3_64bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_reset_withSecretandSeed)
#  define XXH3_64bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_update)
#  define XXH3_64bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_64bits_digest)
#  define XXH3_generateSecret XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret)
#  define XXH3_generateSecret_fromSeed XXH_NAME2(XXH_NAMESPACE, XXH3_generateSecret_fromSeed)
/* XXH3_128bits */
#  define XXH128 XXH_NAME2(XXH_NAMESPACE, XXH128)
#  define XXH3_128bits XXH_NAME2(XXH_NAMESPACE, XXH3_128bits)
#  define XXH3_128bits_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSeed)
#  define XXH3_128bits_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecret)
#  define XXH3_128bits_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_withSecretandSeed)
#  define XXH3_128bits_reset XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset)
#  define XXH3_128bits_reset_withSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSeed)
#  define XXH3_128bits_reset_withSecret XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecret)
#  define XXH3_128bits_reset_withSecretandSeed XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_reset_withSecretandSeed)
#  define XXH3_128bits_update XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_update)
#  define XXH3_128bits_digest XXH_NAME2(XXH_NAMESPACE, XXH3_128bits_digest)
#  define XXH128_isEqual XXH_NAME2(XXH_NAMESPACE, XXH128_isEqual)
#  define XXH128_cmp     XXH_NAME2(XXH_NAMESPACE, XXH128_cmp)
#  define XXH128_canonicalFromHash XXH_NAME2(XXH_NAMESPACE, XXH128_canonicalFromHash)
#  define XXH128_hashFromCanonical XXH_NAME2(XXH_NAMESPACE, XXH128_hashFromCanonical)
#endif


/* *************************************
*  Compiler specifics
***************************************/

/* specific declaration modes for Windows */
#if !defined(XXH_INLINE_ALL) && !defined(XXH_PRIVATE_API)
#  if defined(WIN32) && defined(_MSC_VER) && (defined(XXH_IMPORT) || defined(XXH_EXPORT))
#    ifdef XXH_EXPORT
#      define XXH_PUBLIC_API __declspec(dllexport)
#    elif XXH_IMPORT
#      define XXH_PUBLIC_API __declspec(dllimport)
#    endif
#  else
#    define XXH_PUBLIC_API   /* do nothing */
#  endif
#endif

#if defined (__GNUC__)
# define XXH_CONSTF  __attribute__((const))
# define XXH_PUREF   __attribute__((pure))
# define XXH_MALLOCF __attribute__((malloc))
#else
# define XXH_CONSTF  /* disable */
# define XXH_PUREF
# define XXH_MALLOCF
#endif

/* *************************************
*  Version
***************************************/
#define XXH_VERSION_MAJOR    0
#define XXH_VERSION_MINOR    8
#define XXH_VERSION_RELEASE  2
/*! @brief Version number, encoded as two digits each */
#define XXH_VERSION_NUMBER  (XXH_VERSION_MAJOR *100*100 + XXH_VERSION_MINOR *100 + XXH_VERSION_RELEASE)

#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * @brief Obtains the xxHash version.
 *
 * This is mostly useful when xxHash is compiled as a shared library,
 * since the returned value comes from the library, as opposed to header file.
 *
 * @return @ref XXH_VERSION_NUMBER of the invoked library.
 */
XXH_PUBLIC_API XXH_CONSTF unsigned XXH_versionNumber (void);

#if defined (__cplusplus)
}
#endif

/* ****************************
*  Common basic types
******************************/
#include <stddef.h>   /* size_t */
/*!
 * @brief Exit code for the streaming API.
 */
typedef enum {
    XXH_OK = 0, /*!< OK */
    XXH_ERROR   /*!< Error */
} XXH_errorcode;


/*-**********************************************************************
*  32-bit hash
************************************************************************/
#if defined(XXH_DOXYGEN) /* Don't show <stdint.h> include */
/*!
 * @brief An unsigned 32-bit integer.
 *
 * Not necessarily defined to `uint32_t` but functionally equivalent.
 */
typedef uint32_t XXH32_hash_t;

#elif !defined (__VMS) \
  && (defined (__cplusplus) \
  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
#   ifdef _AIX
#     include <inttypes.h>
#   else
#     include <stdint.h>
#   endif
    typedef uint32_t XXH32_hash_t;

#else
#   include <limits.h>
#   if UINT_MAX == 0xFFFFFFFFUL
      typedef unsigned int XXH32_hash_t;
#   elif ULONG_MAX == 0xFFFFFFFFUL
      typedef unsigned long XXH32_hash_t;
#   else
#     error "unsupported platform: need a 32-bit type"
#   endif
#endif

#if defined (__cplusplus)
extern "C" {
#endif

/*!
 * @}
 *
 * @defgroup XXH32_family XXH32 family
 * @ingroup public
 * Contains functions used in the classic 32-bit xxHash algorithm.
 *
 * @note
 *   XXH32 is useful for older platforms, with no or poor 64-bit performance.
 *   Note that the @ref XXH3_family provides competitive speed for both 32-bit
 *   and 64-bit systems, and offers true 64/128 bit hash results.
 *
 * @see @ref XXH64_family, @ref XXH3_family : Other xxHash families
 * @see @ref XXH32_impl for implementation details
 * @{
 */

/*!
 * @brief Calculates the 32-bit hash of @p input using xxHash32.
 *
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 * @param seed The 32-bit seed to alter the hash's output predictably.
 *
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return The calculated 32-bit xxHash32 value.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32 (const void* input, size_t length, XXH32_hash_t seed);

#ifndef XXH_NO_STREAM
/*!
 * @typedef struct XXH32_state_s XXH32_state_t
 * @brief The opaque state struct for the XXH32 streaming API.
 *
 * @see XXH32_state_s for details.
 */
typedef struct XXH32_state_s XXH32_state_t;

/*!
 * @brief Allocates an @ref XXH32_state_t.
 *
 * @return An allocated pointer of @ref XXH32_state_t on success.
 * @return `NULL` on failure.
 *
 * @note Must be freed with XXH32_freeState().
 */
XXH_PUBLIC_API XXH_MALLOCF XXH32_state_t* XXH32_createState(void);
/*!
 * @brief Frees an @ref XXH32_state_t.
 *
 * @param statePtr A pointer to an @ref XXH32_state_t allocated with @ref XXH32_createState().
 *
 * @return @ref XXH_OK.
 *
 * @note @p statePtr must be allocated with XXH32_createState().
 *
 */
XXH_PUBLIC_API XXH_errorcode  XXH32_freeState(XXH32_state_t* statePtr);
/*!
 * @brief Copies one @ref XXH32_state_t to another.
 *
 * @param dst_state The state to copy to.
 * @param src_state The state to copy from.
 * @pre
 *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
 */
XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dst_state, const XXH32_state_t* src_state);

/*!
 * @brief Resets an @ref XXH32_state_t to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 * @param seed The 32-bit seed to alter the hash result predictably.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note This function resets and seeds a state. Call it before @ref XXH32_update().
 */
XXH_PUBLIC_API XXH_errorcode XXH32_reset  (XXH32_state_t* statePtr, XXH32_hash_t seed);

/*!
 * @brief Consumes a block of @p input to an @ref XXH32_state_t.
 *
 * @param statePtr The state struct to update.
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note Call this to incrementally consume blocks of data.
 */
XXH_PUBLIC_API XXH_errorcode XXH32_update (XXH32_state_t* statePtr, const void* input, size_t length);

/*!
 * @brief Returns the calculated hash value from an @ref XXH32_state_t.
 *
 * @param statePtr The state struct to calculate the hash from.
 *
 * @pre
 *  @p statePtr must not be `NULL`.
 *
 * @return The calculated 32-bit xxHash32 value from that state.
 *
 * @note
 *   Calling XXH32_digest() will not affect @p statePtr, so you can update,
 *   digest, and update again.
 */
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_digest (const XXH32_state_t* statePtr);
#endif /* !XXH_NO_STREAM */

/*******   Canonical representation   *******/

/*!
 * @brief Canonical (big endian) representation of @ref XXH32_hash_t.
 */
typedef struct {
    unsigned char digest[4]; /*!< Hash bytes, big endian */
} XXH32_canonical_t;

/*!
 * @brief Converts an @ref XXH32_hash_t to a big endian @ref XXH32_canonical_t.
 *
 * @param dst  The @ref XXH32_canonical_t pointer to be stored to.
 * @param hash The @ref XXH32_hash_t to be converted.
 *
 * @pre
 *   @p dst must not be `NULL`.
 *
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash);

/*!
 * @brief Converts an @ref XXH32_canonical_t to a native @ref XXH32_hash_t.
 *
 * @param src The @ref XXH32_canonical_t to convert.
 *
 * @pre
 *   @p src must not be `NULL`.
 *
 * @return The converted hash.
 *
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API XXH_PUREF XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src);


/*! @cond Doxygen ignores this part */
#ifdef __has_attribute
# define XXH_HAS_ATTRIBUTE(x) __has_attribute(x)
#else
# define XXH_HAS_ATTRIBUTE(x) 0
#endif
/*! @endcond */

/*! @cond Doxygen ignores this part */
/*
 * C23 __STDC_VERSION__ number hasn't been specified yet. For now
 * leave as `201711L` (C17 + 1).
 * TODO: Update to correct value when its been specified.
 */
#define XXH_C23_VN 201711L
/*! @endcond */

/*! @cond Doxygen ignores this part */
/* C-language Attributes are added in C23. */
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN) && defined(__has_c_attribute)
# define XXH_HAS_C_ATTRIBUTE(x) __has_c_attribute(x)
#else
# define XXH_HAS_C_ATTRIBUTE(x) 0
#endif
/*! @endcond */

/*! @cond Doxygen ignores this part */
#if defined(__cplusplus) && defined(__has_cpp_attribute)
# define XXH_HAS_CPP_ATTRIBUTE(x) __has_cpp_attribute(x)
#else
# define XXH_HAS_CPP_ATTRIBUTE(x) 0
#endif
/*! @endcond */

/*! @cond Doxygen ignores this part */
/*
 * Define XXH_FALLTHROUGH macro for annotating switch case with the 'fallthrough' attribute
 * introduced in CPP17 and C23.
 * CPP17 : https://en.cppreference.com/w/cpp/language/attributes/fallthrough
 * C23   : https://en.cppreference.com/w/c/language/attributes/fallthrough
 */
#if XXH_HAS_C_ATTRIBUTE(fallthrough) || XXH_HAS_CPP_ATTRIBUTE(fallthrough)
# define XXH_FALLTHROUGH [[fallthrough]]
#elif XXH_HAS_ATTRIBUTE(__fallthrough__)
# define XXH_FALLTHROUGH __attribute__ ((__fallthrough__))
#else
# define XXH_FALLTHROUGH /* fallthrough */
#endif
/*! @endcond */

/*! @cond Doxygen ignores this part */
/*
 * Define XXH_NOESCAPE for annotated pointers in public API.
 * https://clang.llvm.org/docs/AttributeReference.html#noescape
 * As of writing this, only supported by clang.
 */
#if XXH_HAS_ATTRIBUTE(noescape)
# define XXH_NOESCAPE __attribute__((noescape))
#else
# define XXH_NOESCAPE
#endif
/*! @endcond */

#if defined (__cplusplus)
} /* end of extern "C" */
#endif

/*!
 * @}
 * @ingroup public
 * @{
 */

#ifndef XXH_NO_LONG_LONG
/*-**********************************************************************
*  64-bit hash
************************************************************************/
#if defined(XXH_DOXYGEN) /* don't include <stdint.h> */
/*!
 * @brief An unsigned 64-bit integer.
 *
 * Not necessarily defined to `uint64_t` but functionally equivalent.
 */
typedef uint64_t XXH64_hash_t;
#elif !defined (__VMS) \
  && (defined (__cplusplus) \
  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
#   ifdef _AIX
#     include <inttypes.h>
#   else
#     include <stdint.h>
#   endif
   typedef uint64_t XXH64_hash_t;
#else
#  include <limits.h>
#  if defined(__LP64__) && ULONG_MAX == 0xFFFFFFFFFFFFFFFFULL
     /* LP64 ABI says uint64_t is unsigned long */
     typedef unsigned long XXH64_hash_t;
#  else
     /* the following type must have a width of 64-bit */
     typedef unsigned long long XXH64_hash_t;
#  endif
#endif

#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * @}
 *
 * @defgroup XXH64_family XXH64 family
 * @ingroup public
 * @{
 * Contains functions used in the classic 64-bit xxHash algorithm.
 *
 * @note
 *   XXH3 provides competitive speed for both 32-bit and 64-bit systems,
 *   and offers true 64/128 bit hash results.
 *   It provides better speed for systems with vector processing capabilities.
 */

/*!
 * @brief Calculates the 64-bit hash of @p input using xxHash64.
 *
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 * @param seed The 64-bit seed to alter the hash's output predictably.
 *
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return The calculated 64-bit xxHash64 value.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);

/*******   Streaming   *******/
#ifndef XXH_NO_STREAM
/*!
 * @brief The opaque state struct for the XXH64 streaming API.
 *
 * @see XXH64_state_s for details.
 */
typedef struct XXH64_state_s XXH64_state_t;   /* incomplete type */

/*!
 * @brief Allocates an @ref XXH64_state_t.
 *
 * @return An allocated pointer of @ref XXH64_state_t on success.
 * @return `NULL` on failure.
 *
 * @note Must be freed with XXH64_freeState().
 */
XXH_PUBLIC_API XXH_MALLOCF XXH64_state_t* XXH64_createState(void);

/*!
 * @brief Frees an @ref XXH64_state_t.
 *
 * @param statePtr A pointer to an @ref XXH64_state_t allocated with @ref XXH64_createState().
 *
 * @return @ref XXH_OK.
 *
 * @note @p statePtr must be allocated with XXH64_createState().
 */
XXH_PUBLIC_API XXH_errorcode  XXH64_freeState(XXH64_state_t* statePtr);

/*!
 * @brief Copies one @ref XXH64_state_t to another.
 *
 * @param dst_state The state to copy to.
 * @param src_state The state to copy from.
 * @pre
 *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
 */
XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dst_state, const XXH64_state_t* src_state);

/*!
 * @brief Resets an @ref XXH64_state_t to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 * @param seed The 64-bit seed to alter the hash result predictably.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note This function resets and seeds a state. Call it before @ref XXH64_update().
 */
XXH_PUBLIC_API XXH_errorcode XXH64_reset  (XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed);

/*!
 * @brief Consumes a block of @p input to an @ref XXH64_state_t.
 *
 * @param statePtr The state struct to update.
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note Call this to incrementally consume blocks of data.
 */
XXH_PUBLIC_API XXH_errorcode XXH64_update (XXH_NOESCAPE XXH64_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);

/*!
 * @brief Returns the calculated hash value from an @ref XXH64_state_t.
 *
 * @param statePtr The state struct to calculate the hash from.
 *
 * @pre
 *  @p statePtr must not be `NULL`.
 *
 * @return The calculated 64-bit xxHash64 value from that state.
 *
 * @note
 *   Calling XXH64_digest() will not affect @p statePtr, so you can update,
 *   digest, and update again.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_digest (XXH_NOESCAPE const XXH64_state_t* statePtr);
#endif /* !XXH_NO_STREAM */
/*******   Canonical representation   *******/

/*!
 * @brief Canonical (big endian) representation of @ref XXH64_hash_t.
 */
typedef struct { unsigned char digest[sizeof(XXH64_hash_t)]; } XXH64_canonical_t;

/*!
 * @brief Converts an @ref XXH64_hash_t to a big endian @ref XXH64_canonical_t.
 *
 * @param dst The @ref XXH64_canonical_t pointer to be stored to.
 * @param hash The @ref XXH64_hash_t to be converted.
 *
 * @pre
 *   @p dst must not be `NULL`.
 *
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash);

/*!
 * @brief Converts an @ref XXH64_canonical_t to a native @ref XXH64_hash_t.
 *
 * @param src The @ref XXH64_canonical_t to convert.
 *
 * @pre
 *   @p src must not be `NULL`.
 *
 * @return The converted hash.
 *
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src);

#ifndef XXH_NO_XXH3

/*!
 * @}
 * ************************************************************************
 * @defgroup XXH3_family XXH3 family
 * @ingroup public
 * @{
 *
 * XXH3 is a more recent hash algorithm featuring:
 *  - Improved speed for both small and large inputs
 *  - True 64-bit and 128-bit outputs
 *  - SIMD acceleration
 *  - Improved 32-bit viability
 *
 * Speed analysis methodology is explained here:
 *
 *    https://fastcompression.blogspot.com/2019/03/presenting-xxh3.html
 *
 * Compared to XXH64, expect XXH3 to run approximately
 * ~2x faster on large inputs and >3x faster on small ones,
 * exact differences vary depending on platform.
 *
 * XXH3's speed benefits greatly from SIMD and 64-bit arithmetic,
 * but does not require it.
 * Most 32-bit and 64-bit targets that can run XXH32 smoothly can run XXH3
 * at competitive speeds, even without vector support. Further details are
 * explained in the implementation.
 *
 * XXH3 has a fast scalar implementation, but it also includes accelerated SIMD
 * implementations for many common platforms:
 *   - AVX512
 *   - AVX2
 *   - SSE2
 *   - ARM NEON
 *   - WebAssembly SIMD128
 *   - POWER8 VSX
 *   - s390x ZVector
 * This can be controlled via the @ref XXH_VECTOR macro, but it automatically
 * selects the best version according to predefined macros. For the x86 family, an
 * automatic runtime dispatcher is included separately in @ref xxh_x86dispatch.c.
 *
 * XXH3 implementation is portable:
 * it has a generic C90 formulation that can be compiled on any platform,
 * all implementations generate exactly the same hash value on all platforms.
 * Starting from v0.8.0, it's also labelled "stable", meaning that
 * any future version will also generate the same hash value.
 *
 * XXH3 offers 2 variants, _64bits and _128bits.
 *
 * When only 64 bits are needed, prefer invoking the _64bits variant, as it
 * reduces the amount of mixing, resulting in faster speed on small inputs.
 * It's also generally simpler to manipulate a scalar return type than a struct.
 *
 * The API supports one-shot hashing, streaming mode, and custom secrets.
 */
/*-**********************************************************************
*  XXH3 64-bit variant
************************************************************************/

/*!
 * @brief Calculates 64-bit unseeded variant of XXH3 hash of @p input.
 *
 * @param input  The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 *
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return The calculated 64-bit XXH3 hash value.
 *
 * @note
 *   This is equivalent to @ref XXH3_64bits_withSeed() with a seed of `0`, however
 *   it may have slightly better performance due to constant propagation of the
 *   defaults.
 *
 * @see
 *    XXH3_64bits_withSeed(), XXH3_64bits_withSecret(): other seeding variants
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length);

/*!
 * @brief Calculates 64-bit seeded variant of XXH3 hash of @p input.
 *
 * @param input  The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 * @param seed   The 64-bit seed to alter the hash result predictably.
 *
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return The calculated 64-bit XXH3 hash value.
 *
 * @note
 *    seed == 0 produces the same results as @ref XXH3_64bits().
 *
 * This variant generates a custom secret on the fly based on default secret
 * altered using the @p seed value.
 *
 * While this operation is decently fast, note that it's not completely free.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed);

/*!
 * The bare minimum size for a custom secret.
 *
 * @see
 *  XXH3_64bits_withSecret(), XXH3_64bits_reset_withSecret(),
 *  XXH3_128bits_withSecret(), XXH3_128bits_reset_withSecret().
 */
#define XXH3_SECRET_SIZE_MIN 136

/*!
 * @brief Calculates 64-bit variant of XXH3 with a custom "secret".
 *
 * @param data       The block of data to be hashed, at least @p len bytes in size.
 * @param len        The length of @p data, in bytes.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 *
 * @return The calculated 64-bit XXH3 hash value.
 *
 * @pre
 *   The memory between @p data and @p data + @p len must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p data may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * It's possible to provide any blob of bytes as a "secret" to generate the hash.
 * This makes it more difficult for an external actor to prepare an intentional collision.
 * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
 * However, the quality of the secret impacts the dispersion of the hash algorithm.
 * Therefore, the secret _must_ look like a bunch of random bytes.
 * Avoid "trivial" or structured data such as repeated sequences or a text document.
 * Whenever in doubt about the "randomness" of the blob of bytes,
 * consider employing @ref XXH3_generateSecret() instead (see below).
 * It will generate a proper high entropy secret derived from the blob of bytes.
 * Another advantage of using XXH3_generateSecret() is that
 * it guarantees that all bits within the initial blob of bytes
 * will impact every bit of the output.
 * This is not necessarily the case when using the blob of bytes directly
 * because, when hashing _small_ inputs, only a portion of the secret is employed.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t XXH3_64bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);


/*******   Streaming   *******/
#ifndef XXH_NO_STREAM
/*
 * Streaming requires state maintenance.
 * This operation costs memory and CPU.
 * As a consequence, streaming is slower than one-shot hashing.
 * For better performance, prefer one-shot functions whenever applicable.
 */

/*!
 * @brief The opaque state struct for the XXH3 streaming API.
 *
 * @see XXH3_state_s for details.
 */
typedef struct XXH3_state_s XXH3_state_t;
XXH_PUBLIC_API XXH_MALLOCF XXH3_state_t* XXH3_createState(void);
XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr);

/*!
 * @brief Copies one @ref XXH3_state_t to another.
 *
 * @param dst_state The state to copy to.
 * @param src_state The state to copy from.
 * @pre
 *   @p dst_state and @p src_state must not be `NULL` and must not overlap.
 */
XXH_PUBLIC_API void XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state);

/*!
 * @brief Resets an @ref XXH3_state_t to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   - This function resets `statePtr` and generate a secret with default parameters.
 *   - Call this function before @ref XXH3_64bits_update().
 *   - Digest will be equivalent to `XXH3_64bits()`.
 *
 */
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);

/*!
 * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 * @param seed     The 64-bit seed to alter the hash result predictably.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   - This function resets `statePtr` and generate a secret from `seed`.
 *   - Call this function before @ref XXH3_64bits_update().
 *   - Digest will be equivalent to `XXH3_64bits_withSeed()`.
 *
 */
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);

/*!
 * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   `secret` is referenced, it _must outlive_ the hash streaming session.
 *
 * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
 * and the quality of produced hash values depends on secret's entropy
 * (secret's content should look like a bunch of random bytes).
 * When in doubt about the randomness of a candidate `secret`,
 * consider employing `XXH3_generateSecret()` instead (see below).
 */
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);

/*!
 * @brief Consumes a block of @p input to an @ref XXH3_state_t.
 *
 * @param statePtr The state struct to update.
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 * @pre
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note Call this to incrementally consume blocks of data.
 */
XXH_PUBLIC_API XXH_errorcode XXH3_64bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);

/*!
 * @brief Returns the calculated XXH3 64-bit hash value from an @ref XXH3_state_t.
 *
 * @param statePtr The state struct to calculate the hash from.
 *
 * @pre
 *  @p statePtr must not be `NULL`.
 *
 * @return The calculated XXH3 64-bit hash value from that state.
 *
 * @note
 *   Calling XXH3_64bits_digest() will not affect @p statePtr, so you can update,
 *   digest, and update again.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t  XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
#endif /* !XXH_NO_STREAM */

/* note : canonical representation of XXH3 is the same as XXH64
 * since they both produce XXH64_hash_t values */


/*-**********************************************************************
*  XXH3 128-bit variant
************************************************************************/

/*!
 * @brief The return value from 128-bit hashes.
 *
 * Stored in little endian order, although the fields themselves are in native
 * endianness.
 */
typedef struct {
    XXH64_hash_t low64;   /*!< `value & 0xFFFFFFFFFFFFFFFF` */
    XXH64_hash_t high64;  /*!< `value >> 64` */
} XXH128_hash_t;

/*!
 * @brief Calculates 128-bit unseeded variant of XXH3 of @p data.
 *
 * @param data The block of data to be hashed, at least @p length bytes in size.
 * @param len  The length of @p data, in bytes.
 *
 * @return The calculated 128-bit variant of XXH3 value.
 *
 * The 128-bit variant of XXH3 has more strength, but it has a bit of overhead
 * for shorter inputs.
 *
 * This is equivalent to @ref XXH3_128bits_withSeed() with a seed of `0`, however
 * it may have slightly better performance due to constant propagation of the
 * defaults.
 *
 * @see XXH3_128bits_withSeed(), XXH3_128bits_withSecret(): other seeding variants
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* data, size_t len);
/*! @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
 *
 * @param data The block of data to be hashed, at least @p length bytes in size.
 * @param len  The length of @p data, in bytes.
 * @param seed The 64-bit seed to alter the hash result predictably.
 *
 * @return The calculated 128-bit variant of XXH3 value.
 *
 * @note
 *    seed == 0 produces the same results as @ref XXH3_64bits().
 *
 * This variant generates a custom secret on the fly based on default secret
 * altered using the @p seed value.
 *
 * While this operation is decently fast, note that it's not completely free.
 *
 * @see XXH3_128bits(), XXH3_128bits_withSecret(): other seeding variants
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSeed(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);
/*!
 * @brief Calculates 128-bit variant of XXH3 with a custom "secret".
 *
 * @param data       The block of data to be hashed, at least @p len bytes in size.
 * @param len        The length of @p data, in bytes.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 *
 * @return The calculated 128-bit variant of XXH3 value.
 *
 * It's possible to provide any blob of bytes as a "secret" to generate the hash.
 * This makes it more difficult for an external actor to prepare an intentional collision.
 * The main condition is that @p secretSize *must* be large enough (>= @ref XXH3_SECRET_SIZE_MIN).
 * However, the quality of the secret impacts the dispersion of the hash algorithm.
 * Therefore, the secret _must_ look like a bunch of random bytes.
 * Avoid "trivial" or structured data such as repeated sequences or a text document.
 * Whenever in doubt about the "randomness" of the blob of bytes,
 * consider employing @ref XXH3_generateSecret() instead (see below).
 * It will generate a proper high entropy secret derived from the blob of bytes.
 * Another advantage of using XXH3_generateSecret() is that
 * it guarantees that all bits within the initial blob of bytes
 * will impact every bit of the output.
 * This is not necessarily the case when using the blob of bytes directly
 * because, when hashing _small_ inputs, only a portion of the secret is employed.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_withSecret(XXH_NOESCAPE const void* data, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize);

/*******   Streaming   *******/
#ifndef XXH_NO_STREAM
/*
 * Streaming requires state maintenance.
 * This operation costs memory and CPU.
 * As a consequence, streaming is slower than one-shot hashing.
 * For better performance, prefer one-shot functions whenever applicable.
 *
 * XXH3_128bits uses the same XXH3_state_t as XXH3_64bits().
 * Use already declared XXH3_createState() and XXH3_freeState().
 *
 * All reset and streaming functions have same meaning as their 64-bit counterpart.
 */

/*!
 * @brief Resets an @ref XXH3_state_t to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   - This function resets `statePtr` and generate a secret with default parameters.
 *   - Call it before @ref XXH3_128bits_update().
 *   - Digest will be equivalent to `XXH3_128bits()`.
 */
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr);

/*!
 * @brief Resets an @ref XXH3_state_t with 64-bit seed to begin a new hash.
 *
 * @param statePtr The state struct to reset.
 * @param seed     The 64-bit seed to alter the hash result predictably.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   - This function resets `statePtr` and generate a secret from `seed`.
 *   - Call it before @ref XXH3_128bits_update().
 *   - Digest will be equivalent to `XXH3_128bits_withSeed()`.
 */
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed);
/*!
 * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
 *
 * @param statePtr   The state struct to reset.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * `secret` is referenced, it _must outlive_ the hash streaming session.
 * Similar to one-shot API, `secretSize` must be >= @ref XXH3_SECRET_SIZE_MIN,
 * and the quality of produced hash values depends on secret's entropy
 * (secret's content should look like a bunch of random bytes).
 * When in doubt about the randomness of a candidate `secret`,
 * consider employing `XXH3_generateSecret()` instead (see below).
 */
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize);

/*!
 * @brief Consumes a block of @p input to an @ref XXH3_state_t.
 *
 * Call this to incrementally consume blocks of data.
 *
 * @param statePtr The state struct to update.
 * @param input The block of data to be hashed, at least @p length bytes in size.
 * @param length The length of @p input, in bytes.
 *
 * @pre
 *   @p statePtr must not be `NULL`.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @note
 *   The memory between @p input and @p input + @p length must be valid,
 *   readable, contiguous memory. However, if @p length is `0`, @p input may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 */
XXH_PUBLIC_API XXH_errorcode XXH3_128bits_update (XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* input, size_t length);

/*!
 * @brief Returns the calculated XXH3 128-bit hash value from an @ref XXH3_state_t.
 *
 * @param statePtr The state struct to calculate the hash from.
 *
 * @pre
 *  @p statePtr must not be `NULL`.
 *
 * @return The calculated XXH3 128-bit hash value from that state.
 *
 * @note
 *   Calling XXH3_128bits_digest() will not affect @p statePtr, so you can update,
 *   digest, and update again.
 *
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* statePtr);
#endif /* !XXH_NO_STREAM */

/* Following helper functions make it possible to compare XXH128_hast_t values.
 * Since XXH128_hash_t is a structure, this capability is not offered by the language.
 * Note: For better performance, these functions can be inlined using XXH_INLINE_ALL */

/*!
 * @brief Check equality of two XXH128_hash_t values
 *
 * @param h1 The 128-bit hash value.
 * @param h2 Another 128-bit hash value.
 *
 * @return `1` if `h1` and `h2` are equal.
 * @return `0` if they are not.
 */
XXH_PUBLIC_API XXH_PUREF int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2);

/*!
 * @brief Compares two @ref XXH128_hash_t
 *
 * This comparator is compatible with stdlib's `qsort()`/`bsearch()`.
 *
 * @param h128_1 Left-hand side value
 * @param h128_2 Right-hand side value
 *
 * @return >0 if @p h128_1  > @p h128_2
 * @return =0 if @p h128_1 == @p h128_2
 * @return <0 if @p h128_1  < @p h128_2
 */
XXH_PUBLIC_API XXH_PUREF int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2);


/*******   Canonical representation   *******/
typedef struct { unsigned char digest[sizeof(XXH128_hash_t)]; } XXH128_canonical_t;


/*!
 * @brief Converts an @ref XXH128_hash_t to a big endian @ref XXH128_canonical_t.
 *
 * @param dst  The @ref XXH128_canonical_t pointer to be stored to.
 * @param hash The @ref XXH128_hash_t to be converted.
 *
 * @pre
 *   @p dst must not be `NULL`.
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API void XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash);

/*!
 * @brief Converts an @ref XXH128_canonical_t to a native @ref XXH128_hash_t.
 *
 * @param src The @ref XXH128_canonical_t to convert.
 *
 * @pre
 *   @p src must not be `NULL`.
 *
 * @return The converted hash.
 * @see @ref canonical_representation_example "Canonical Representation Example"
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src);


#endif  /* !XXH_NO_XXH3 */

#if defined (__cplusplus)
} /* extern "C" */
#endif

#endif  /* XXH_NO_LONG_LONG */

/*!
 * @}
 */
#endif /* XXHASH_H_5627135585666179 */


#if defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742)
#define XXHASH_H_STATIC_13879238742
/* ****************************************************************************
 * This section contains declarations which are not guaranteed to remain stable.
 * They may change in future versions, becoming incompatible with a different
 * version of the library.
 * These declarations should only be used with static linking.
 * Never use them in association with dynamic linking!
 ***************************************************************************** */

/*
 * These definitions are only present to allow static allocation
 * of XXH states, on stack or in a struct, for example.
 * Never **ever** access their members directly.
 */

/*!
 * @internal
 * @brief Structure for XXH32 streaming API.
 *
 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
 * an opaque type. This allows fields to safely be changed.
 *
 * Typedef'd to @ref XXH32_state_t.
 * Do not access the members of this struct directly.
 * @see XXH64_state_s, XXH3_state_s
 */
struct XXH32_state_s {
   XXH32_hash_t total_len_32; /*!< Total length hashed, modulo 2^32 */
   XXH32_hash_t large_len;    /*!< Whether the hash is >= 16 (handles @ref total_len_32 overflow) */
   XXH32_hash_t v[4];         /*!< Accumulator lanes */
   XXH32_hash_t mem32[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[16]. */
   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem32 */
   XXH32_hash_t reserved;     /*!< Reserved field. Do not read nor write to it. */
};   /* typedef'd to XXH32_state_t */


#ifndef XXH_NO_LONG_LONG  /* defined when there is no 64-bit support */

/*!
 * @internal
 * @brief Structure for XXH64 streaming API.
 *
 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined. Otherwise it is
 * an opaque type. This allows fields to safely be changed.
 *
 * Typedef'd to @ref XXH64_state_t.
 * Do not access the members of this struct directly.
 * @see XXH32_state_s, XXH3_state_s
 */
struct XXH64_state_s {
   XXH64_hash_t total_len;    /*!< Total length hashed. This is always 64-bit. */
   XXH64_hash_t v[4];         /*!< Accumulator lanes */
   XXH64_hash_t mem64[4];     /*!< Internal buffer for partial reads. Treated as unsigned char[32]. */
   XXH32_hash_t memsize;      /*!< Amount of data in @ref mem64 */
   XXH32_hash_t reserved32;   /*!< Reserved field, needed for padding anyways*/
   XXH64_hash_t reserved64;   /*!< Reserved field. Do not read or write to it. */
};   /* typedef'd to XXH64_state_t */

#ifndef XXH_NO_XXH3

#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) /* >= C11 */
#  include <stdalign.h>
#  define XXH_ALIGN(n)      alignas(n)
#elif defined(__cplusplus) && (__cplusplus >= 201103L) /* >= C++11 */
/* In C++ alignas() is a keyword */
#  define XXH_ALIGN(n)      alignas(n)
#elif defined(__GNUC__)
#  define XXH_ALIGN(n)      __attribute__ ((aligned(n)))
#elif defined(_MSC_VER)
#  define XXH_ALIGN(n)      __declspec(align(n))
#else
#  define XXH_ALIGN(n)   /* disabled */
#endif

/* Old GCC versions only accept the attribute after the type in structures. */
#if !(defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L))   /* C11+ */ \
    && ! (defined(__cplusplus) && (__cplusplus >= 201103L)) /* >= C++11 */ \
    && defined(__GNUC__)
#   define XXH_ALIGN_MEMBER(align, type) type XXH_ALIGN(align)
#else
#   define XXH_ALIGN_MEMBER(align, type) XXH_ALIGN(align) type
#endif

/*!
 * @brief The size of the internal XXH3 buffer.
 *
 * This is the optimal update size for incremental hashing.
 *
 * @see XXH3_64b_update(), XXH3_128b_update().
 */
#define XXH3_INTERNALBUFFER_SIZE 256

/*!
 * @internal
 * @brief Default size of the secret buffer (and @ref XXH3_kSecret).
 *
 * This is the size used in @ref XXH3_kSecret and the seeded functions.
 *
 * Not to be confused with @ref XXH3_SECRET_SIZE_MIN.
 */
#define XXH3_SECRET_DEFAULT_SIZE 192

/*!
 * @internal
 * @brief Structure for XXH3 streaming API.
 *
 * @note This is only defined when @ref XXH_STATIC_LINKING_ONLY,
 * @ref XXH_INLINE_ALL, or @ref XXH_IMPLEMENTATION is defined.
 * Otherwise it is an opaque type.
 * Never use this definition in combination with dynamic library.
 * This allows fields to safely be changed in the future.
 *
 * @note ** This structure has a strict alignment requirement of 64 bytes!! **
 * Do not allocate this with `malloc()` or `new`,
 * it will not be sufficiently aligned.
 * Use @ref XXH3_createState() and @ref XXH3_freeState(), or stack allocation.
 *
 * Typedef'd to @ref XXH3_state_t.
 * Do never access the members of this struct directly.
 *
 * @see XXH3_INITSTATE() for stack initialization.
 * @see XXH3_createState(), XXH3_freeState().
 * @see XXH32_state_s, XXH64_state_s
 */
struct XXH3_state_s {
   XXH_ALIGN_MEMBER(64, XXH64_hash_t acc[8]);
       /*!< The 8 accumulators. See @ref XXH32_state_s::v and @ref XXH64_state_s::v */
   XXH_ALIGN_MEMBER(64, unsigned char customSecret[XXH3_SECRET_DEFAULT_SIZE]);
       /*!< Used to store a custom secret generated from a seed. */
   XXH_ALIGN_MEMBER(64, unsigned char buffer[XXH3_INTERNALBUFFER_SIZE]);
       /*!< The internal buffer. @see XXH32_state_s::mem32 */
   XXH32_hash_t bufferedSize;
       /*!< The amount of memory in @ref buffer, @see XXH32_state_s::memsize */
   XXH32_hash_t useSeed;
       /*!< Reserved field. Needed for padding on 64-bit. */
   size_t nbStripesSoFar;
       /*!< Number or stripes processed. */
   XXH64_hash_t totalLen;
       /*!< Total length hashed. 64-bit even on 32-bit targets. */
   size_t nbStripesPerBlock;
       /*!< Number of stripes per block. */
   size_t secretLimit;
       /*!< Size of @ref customSecret or @ref extSecret */
   XXH64_hash_t seed;
       /*!< Seed for _withSeed variants. Must be zero otherwise, @see XXH3_INITSTATE() */
   XXH64_hash_t reserved64;
       /*!< Reserved field. */
   const unsigned char* extSecret;
       /*!< Reference to an external secret for the _withSecret variants, NULL
        *   for other variants. */
   /* note: there may be some padding at the end due to alignment on 64 bytes */
}; /* typedef'd to XXH3_state_t */

#undef XXH_ALIGN_MEMBER

/*!
 * @brief Initializes a stack-allocated `XXH3_state_s`.
 *
 * When the @ref XXH3_state_t structure is merely emplaced on stack,
 * it should be initialized with XXH3_INITSTATE() or a memset()
 * in case its first reset uses XXH3_NNbits_reset_withSeed().
 * This init can be omitted if the first reset uses default or _withSecret mode.
 * This operation isn't necessary when the state is created with XXH3_createState().
 * Note that this doesn't prepare the state for a streaming operation,
 * it's still necessary to use XXH3_NNbits_reset*() afterwards.
 */
#define XXH3_INITSTATE(XXH3_state_ptr)                       \
    do {                                                     \
        XXH3_state_t* tmp_xxh3_state_ptr = (XXH3_state_ptr); \
        tmp_xxh3_state_ptr->seed = 0;                        \
        tmp_xxh3_state_ptr->extSecret = NULL;                \
    } while(0)


#if defined (__cplusplus)
extern "C" {
#endif

/*!
 * @brief Calculates the 128-bit hash of @p data using XXH3.
 *
 * @param data The block of data to be hashed, at least @p len bytes in size.
 * @param len  The length of @p data, in bytes.
 * @param seed The 64-bit seed to alter the hash's output predictably.
 *
 * @pre
 *   The memory between @p data and @p data + @p len must be valid,
 *   readable, contiguous memory. However, if @p len is `0`, @p data may be
 *   `NULL`. In C++, this also must be *TriviallyCopyable*.
 *
 * @return The calculated 128-bit XXH3 value.
 *
 * @see @ref single_shot_example "Single Shot Example" for an example.
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t XXH128(XXH_NOESCAPE const void* data, size_t len, XXH64_hash_t seed);


/* ===   Experimental API   === */
/* Symbols defined below must be considered tied to a specific library version. */

/*!
 * @brief Derive a high-entropy secret from any user-defined content, named customSeed.
 *
 * @param secretBuffer    A writable buffer for derived high-entropy secret data.
 * @param secretSize      Size of secretBuffer, in bytes.  Must be >= XXH3_SECRET_DEFAULT_SIZE.
 * @param customSeed      A user-defined content.
 * @param customSeedSize  Size of customSeed, in bytes.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * The generated secret can be used in combination with `*_withSecret()` functions.
 * The `_withSecret()` variants are useful to provide a higher level of protection
 * than 64-bit seed, as it becomes much more difficult for an external actor to
 * guess how to impact the calculation logic.
 *
 * The function accepts as input a custom seed of any length and any content,
 * and derives from it a high-entropy secret of length @p secretSize into an
 * already allocated buffer @p secretBuffer.
 *
 * The generated secret can then be used with any `*_withSecret()` variant.
 * The functions @ref XXH3_128bits_withSecret(), @ref XXH3_64bits_withSecret(),
 * @ref XXH3_128bits_reset_withSecret() and @ref XXH3_64bits_reset_withSecret()
 * are part of this list. They all accept a `secret` parameter
 * which must be large enough for implementation reasons (>= @ref XXH3_SECRET_SIZE_MIN)
 * _and_ feature very high entropy (consist of random-looking bytes).
 * These conditions can be a high bar to meet, so @ref XXH3_generateSecret() can
 * be employed to ensure proper quality.
 *
 * @p customSeed can be anything. It can have any size, even small ones,
 * and its content can be anything, even "poor entropy" sources such as a bunch
 * of zeroes. The resulting `secret` will nonetheless provide all required qualities.
 *
 * @pre
 *   - @p secretSize must be >= @ref XXH3_SECRET_SIZE_MIN
 *   - When @p customSeedSize > 0, supplying NULL as customSeed is undefined behavior.
 *
 * Example code:
 * @code{.c}
 *    #include <stdio.h>
 *    #include <stdlib.h>
 *    #include <string.h>
 *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
 *    #include "xxhash.h"
 *    // Hashes argv[2] using the entropy from argv[1].
 *    int main(int argc, char* argv[])
 *    {
 *        char secret[XXH3_SECRET_SIZE_MIN];
 *        if (argv != 3) { return 1; }
 *        XXH3_generateSecret(secret, sizeof(secret), argv[1], strlen(argv[1]));
 *        XXH64_hash_t h = XXH3_64bits_withSecret(
 *             argv[2], strlen(argv[2]),
 *             secret, sizeof(secret)
 *        );
 *        printf("%016llx\n", (unsigned long long) h);
 *    }
 * @endcode
 */
XXH_PUBLIC_API XXH_errorcode XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize);

/*!
 * @brief Generate the same secret as the _withSeed() variants.
 *
 * @param secretBuffer A writable buffer of @ref XXH3_SECRET_SIZE_MIN bytes
 * @param seed         The 64-bit seed to alter the hash result predictably.
 *
 * The generated secret can be used in combination with
 *`*_withSecret()` and `_withSecretandSeed()` variants.
 *
 * Example C++ `std::string` hash class:
 * @code{.cpp}
 *    #include <string>
 *    #define XXH_STATIC_LINKING_ONLY // expose unstable API
 *    #include "xxhash.h"
 *    // Slow, seeds each time
 *    class HashSlow {
 *        XXH64_hash_t seed;
 *    public:
 *        HashSlow(XXH64_hash_t s) : seed{s} {}
 *        size_t operator()(const std::string& x) const {
 *            return size_t{XXH3_64bits_withSeed(x.c_str(), x.length(), seed)};
 *        }
 *    };
 *    // Fast, caches the seeded secret for future uses.
 *    class HashFast {
 *        unsigned char secret[XXH3_SECRET_SIZE_MIN];
 *    public:
 *        HashFast(XXH64_hash_t s) {
 *            XXH3_generateSecret_fromSeed(secret, seed);
 *        }
 *        size_t operator()(const std::string& x) const {
 *            return size_t{
 *                XXH3_64bits_withSecret(x.c_str(), x.length(), secret, sizeof(secret))
 *            };
 *        }
 *    };
 * @endcode
 */
XXH_PUBLIC_API void XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed);

/*!
 * @brief Calculates 64/128-bit seeded variant of XXH3 hash of @p data.
 *
 * @param data       The block of data to be hashed, at least @p len bytes in size.
 * @param len        The length of @p data, in bytes.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 * @param seed       The 64-bit seed to alter the hash result predictably.
 *
 * These variants generate hash values using either
 * @p seed for "short" keys (< @ref XXH3_MIDSIZE_MAX = 240 bytes)
 * or @p secret for "large" keys (>= @ref XXH3_MIDSIZE_MAX).
 *
 * This generally benefits speed, compared to `_withSeed()` or `_withSecret()`.
 * `_withSeed()` has to generate the secret on the fly for "large" keys.
 * It's fast, but can be perceptible for "not so large" keys (< 1 KB).
 * `_withSecret()` has to generate the masks on the fly for "small" keys,
 * which requires more instructions than _withSeed() variants.
 * Therefore, _withSecretandSeed variant combines the best of both worlds.
 *
 * When @p secret has been generated by XXH3_generateSecret_fromSeed(),
 * this variant produces *exactly* the same results as `_withSeed()` variant,
 * hence offering only a pure speed benefit on "large" input,
 * by skipping the need to regenerate the secret for every large input.
 *
 * Another usage scenario is to hash the secret to a 64-bit hash value,
 * for example with XXH3_64bits(), which then becomes the seed,
 * and then employ both the seed and the secret in _withSecretandSeed().
 * On top of speed, an added benefit is that each bit in the secret
 * has a 50% chance to swap each bit in the output, via its impact to the seed.
 *
 * This is not guaranteed when using the secret directly in "small data" scenarios,
 * because only portions of the secret are employed for small data.
 */
XXH_PUBLIC_API XXH_PUREF XXH64_hash_t
XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* data, size_t len,
                              XXH_NOESCAPE const void* secret, size_t secretSize,
                              XXH64_hash_t seed);
/*!
 * @brief Calculates 128-bit seeded variant of XXH3 hash of @p data.
 *
 * @param input      The block of data to be hashed, at least @p len bytes in size.
 * @param length     The length of @p data, in bytes.
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 * @param seed64     The 64-bit seed to alter the hash result predictably.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @see XXH3_64bits_withSecretandSeed()
 */
XXH_PUBLIC_API XXH_PUREF XXH128_hash_t
XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length,
                               XXH_NOESCAPE const void* secret, size_t secretSize,
                               XXH64_hash_t seed64);
#ifndef XXH_NO_STREAM
/*!
 * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
 *
 * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 * @param seed64     The 64-bit seed to alter the hash result predictably.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @see XXH3_64bits_withSecretandSeed()
 */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
                                    XXH_NOESCAPE const void* secret, size_t secretSize,
                                    XXH64_hash_t seed64);
/*!
 * @brief Resets an @ref XXH3_state_t with secret data to begin a new hash.
 *
 * @param statePtr   A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
 * @param secret     The secret data.
 * @param secretSize The length of @p secret, in bytes.
 * @param seed64     The 64-bit seed to alter the hash result predictably.
 *
 * @return @ref XXH_OK on success.
 * @return @ref XXH_ERROR on failure.
 *
 * @see XXH3_64bits_withSecretandSeed()
 */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr,
                                     XXH_NOESCAPE const void* secret, size_t secretSize,
                                     XXH64_hash_t seed64);
#endif /* !XXH_NO_STREAM */

#if defined (__cplusplus)
} /* extern "C" */
#endif

#endif  /* !XXH_NO_XXH3 */
#endif  /* XXH_NO_LONG_LONG */

#if defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API)
#  define XXH_IMPLEMENTATION
#endif

#endif  /* defined(XXH_STATIC_LINKING_ONLY) && !defined(XXHASH_H_STATIC_13879238742) */


/* ======================================================================== */
/* ======================================================================== */
/* ======================================================================== */


/*-**********************************************************************
 * xxHash implementation
 *-**********************************************************************
 * xxHash's implementation used to be hosted inside xxhash.c.
 *
 * However, inlining requires implementation to be visible to the compiler,
 * hence be included alongside the header.
 * Previously, implementation was hosted inside xxhash.c,
 * which was then #included when inlining was activated.
 * This construction created issues with a few build and install systems,
 * as it required xxhash.c to be stored in /include directory.
 *
 * xxHash implementation is now directly integrated within xxhash.h.
 * As a consequence, xxhash.c is no longer needed in /include.
 *
 * xxhash.c is still available and is still useful.
 * In a "normal" setup, when xxhash is not inlined,
 * xxhash.h only exposes the prototypes and public symbols,
 * while xxhash.c can be built into an object file xxhash.o
 * which can then be linked into the final binary.
 ************************************************************************/

#if ( defined(XXH_INLINE_ALL) || defined(XXH_PRIVATE_API) \
   || defined(XXH_IMPLEMENTATION) ) && !defined(XXH_IMPLEM_13a8737387)
#  define XXH_IMPLEM_13a8737387

/* *************************************
*  Tuning parameters
***************************************/

/*!
 * @defgroup tuning Tuning parameters
 * @{
 *
 * Various macros to control xxHash's behavior.
 */
#ifdef XXH_DOXYGEN
/*!
 * @brief Define this to disable 64-bit code.
 *
 * Useful if only using the @ref XXH32_family and you have a strict C90 compiler.
 */
#  define XXH_NO_LONG_LONG
#  undef XXH_NO_LONG_LONG /* don't actually */
/*!
 * @brief Controls how unaligned memory is accessed.
 *
 * By default, access to unaligned memory is controlled by `memcpy()`, which is
 * safe and portable.
 *
 * Unfortunately, on some target/compiler combinations, the generated assembly
 * is sub-optimal.
 *
 * The below switch allow selection of a different access method
 * in the search for improved performance.
 *
 * @par Possible options:
 *
 *  - `XXH_FORCE_MEMORY_ACCESS=0` (default): `memcpy`
 *   @par
 *     Use `memcpy()`. Safe and portable. Note that most modern compilers will
 *     eliminate the function call and treat it as an unaligned access.
 *
 *  - `XXH_FORCE_MEMORY_ACCESS=1`: `__attribute__((aligned(1)))`
 *   @par
 *     Depends on compiler extensions and is therefore not portable.
 *     This method is safe _if_ your compiler supports it,
 *     and *generally* as fast or faster than `memcpy`.
 *
 *  - `XXH_FORCE_MEMORY_ACCESS=2`: Direct cast
 *  @par
 *     Casts directly and dereferences. This method doesn't depend on the
 *     compiler, but it violates the C standard as it directly dereferences an
 *     unaligned pointer. It can generate buggy code on targets which do not
 *     support unaligned memory accesses, but in some circumstances, it's the
 *     only known way to get the most performance.
 *
 *  - `XXH_FORCE_MEMORY_ACCESS=3`: Byteshift
 *  @par
 *     Also portable. This can generate the best code on old compilers which don't
 *     inline small `memcpy()` calls, and it might also be faster on big-endian
 *     systems which lack a native byteswap instruction. However, some compilers
 *     will emit literal byteshifts even if the target supports unaligned access.
 *
 *
 * @warning
 *   Methods 1 and 2 rely on implementation-defined behavior. Use these with
 *   care, as what works on one compiler/platform/optimization level may cause
 *   another to read garbage data or even crash.
 *
 * See https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html for details.
 *
 * Prefer these methods in priority order (0 > 3 > 1 > 2)
 */
#  define XXH_FORCE_MEMORY_ACCESS 0

/*!
 * @def XXH_SIZE_OPT
 * @brief Controls how much xxHash optimizes for size.
 *
 * xxHash, when compiled, tends to result in a rather large binary size. This
 * is mostly due to heavy usage to forced inlining and constant folding of the
 * @ref XXH3_family to increase performance.
 *
 * However, some developers prefer size over speed. This option can
 * significantly reduce the size of the generated code. When using the `-Os`
 * or `-Oz` options on GCC or Clang, this is defined to 1 by default,
 * otherwise it is defined to 0.
 *
 * Most of these size optimizations can be controlled manually.
 *
 * This is a number from 0-2.
 *  - `XXH_SIZE_OPT` == 0: Default. xxHash makes no size optimizations. Speed
 *    comes first.
 *  - `XXH_SIZE_OPT` == 1: Default for `-Os` and `-Oz`. xxHash is more
 *    conservative and disables hacks that increase code size. It implies the
 *    options @ref XXH_NO_INLINE_HINTS == 1, @ref XXH_FORCE_ALIGN_CHECK == 0,
 *    and @ref XXH3_NEON_LANES == 8 if they are not already defined.
 *  - `XXH_SIZE_OPT` == 2: xxHash tries to make itself as small as possible.
 *    Performance may cry. For example, the single shot functions just use the
 *    streaming API.
 */
#  define XXH_SIZE_OPT 0

/*!
 * @def XXH_FORCE_ALIGN_CHECK
 * @brief If defined to non-zero, adds a special path for aligned inputs (XXH32()
 * and XXH64() only).
 *
 * This is an important performance trick for architectures without decent
 * unaligned memory access performance.
 *
 * It checks for input alignment, and when conditions are met, uses a "fast
 * path" employing direct 32-bit/64-bit reads, resulting in _dramatically
 * faster_ read speed.
 *
 * The check costs one initial branch per hash, which is generally negligible,
 * but not zero.
 *
 * Moreover, it's not useful to generate an additional code path if memory
 * access uses the same instruction for both aligned and unaligned
 * addresses (e.g. x86 and aarch64).
 *
 * In these cases, the alignment check can be removed by setting this macro to 0.
 * Then the code will always use unaligned memory access.
 * Align check is automatically disabled on x86, x64, ARM64, and some ARM chips
 * which are platforms known to offer good unaligned memory accesses performance.
 *
 * It is also disabled by default when @ref XXH_SIZE_OPT >= 1.
 *
 * This option does not affect XXH3 (only XXH32 and XXH64).
 */
#  define XXH_FORCE_ALIGN_CHECK 0

/*!
 * @def XXH_NO_INLINE_HINTS
 * @brief When non-zero, sets all functions to `static`.
 *
 * By default, xxHash tries to force the compiler to inline almost all internal
 * functions.
 *
 * This can usually improve performance due to reduced jumping and improved
 * constant folding, but significantly increases the size of the binary which
 * might not be favorable.
 *
 * Additionally, sometimes the forced inlining can be detrimental to performance,
 * depending on the architecture.
 *
 * XXH_NO_INLINE_HINTS marks all internal functions as static, giving the
 * compiler full control on whether to inline or not.
 *
 * When not optimizing (-O0), using `-fno-inline` with GCC or Clang, or if
 * @ref XXH_SIZE_OPT >= 1, this will automatically be defined.
 */
#  define XXH_NO_INLINE_HINTS 0

/*!
 * @def XXH3_INLINE_SECRET
 * @brief Determines whether to inline the XXH3 withSecret code.
 *
 * When the secret size is known, the compiler can improve the performance
 * of XXH3_64bits_withSecret() and XXH3_128bits_withSecret().
 *
 * However, if the secret size is not known, it doesn't have any benefit. This
 * happens when xxHash is compiled into a global symbol. Therefore, if
 * @ref XXH_INLINE_ALL is *not* defined, this will be defined to 0.
 *
 * Additionally, this defaults to 0 on GCC 12+, which has an issue with function pointers
 * that are *sometimes* force inline on -Og, and it is impossible to automatically
 * detect this optimization level.
 */
#  define XXH3_INLINE_SECRET 0

/*!
 * @def XXH32_ENDJMP
 * @brief Whether to use a jump for `XXH32_finalize`.
 *
 * For performance, `XXH32_finalize` uses multiple branches in the finalizer.
 * This is generally preferable for performance,
 * but depending on exact architecture, a jmp may be preferable.
 *
 * This setting is only possibly making a difference for very small inputs.
 */
#  define XXH32_ENDJMP 0

/*!
 * @internal
 * @brief Redefines old internal names.
 *
 * For compatibility with code that uses xxHash's internals before the names
 * were changed to improve namespacing. There is no other reason to use this.
 */
#  define XXH_OLD_NAMES
#  undef XXH_OLD_NAMES /* don't actually use, it is ugly. */

/*!
 * @def XXH_NO_STREAM
 * @brief Disables the streaming API.
 *
 * When xxHash is not inlined and the streaming functions are not used, disabling
 * the streaming functions can improve code size significantly, especially with
 * the @ref XXH3_family which tends to make constant folded copies of itself.
 */
#  define XXH_NO_STREAM
#  undef XXH_NO_STREAM /* don't actually */
#endif /* XXH_DOXYGEN */
/*!
 * @}
 */

#ifndef XXH_FORCE_MEMORY_ACCESS   /* can be defined externally, on command line for example */
   /* prefer __packed__ structures (method 1) for GCC
    * < ARMv7 with unaligned access (e.g. Raspbian armhf) still uses byte shifting, so we use memcpy
    * which for some reason does unaligned loads. */
#  if defined(__GNUC__) && !(defined(__ARM_ARCH) && __ARM_ARCH < 7 && defined(__ARM_FEATURE_UNALIGNED))
#    define XXH_FORCE_MEMORY_ACCESS 1
#  endif
#endif

#ifndef XXH_SIZE_OPT
   /* default to 1 for -Os or -Oz */
#  if (defined(__GNUC__) || defined(__clang__)) && defined(__OPTIMIZE_SIZE__)
#    define XXH_SIZE_OPT 1
#  else
#    define XXH_SIZE_OPT 0
#  endif
#endif

#ifndef XXH_FORCE_ALIGN_CHECK  /* can be defined externally */
   /* don't check on sizeopt, x86, aarch64, or arm when unaligned access is available */
#  if XXH_SIZE_OPT >= 1 || \
      defined(__i386)  || defined(__x86_64__) || defined(__aarch64__) || defined(__ARM_FEATURE_UNALIGNED) \
   || defined(_M_IX86) || defined(_M_X64)     || defined(_M_ARM64)    || defined(_M_ARM) /* visual */
#    define XXH_FORCE_ALIGN_CHECK 0
#  else
#    define XXH_FORCE_ALIGN_CHECK 1
#  endif
#endif

#ifndef XXH_NO_INLINE_HINTS
#  if XXH_SIZE_OPT >= 1 || defined(__NO_INLINE__)  /* -O0, -fno-inline */
#    define XXH_NO_INLINE_HINTS 1
#  else
#    define XXH_NO_INLINE_HINTS 0
#  endif
#endif

#ifndef XXH3_INLINE_SECRET
#  if (defined(__GNUC__) && !defined(__clang__) && __GNUC__ >= 12) \
     || !defined(XXH_INLINE_ALL)
#    define XXH3_INLINE_SECRET 0
#  else
#    define XXH3_INLINE_SECRET 1
#  endif
#endif

#ifndef XXH32_ENDJMP
/* generally preferable for performance */
#  define XXH32_ENDJMP 0
#endif

/*!
 * @defgroup impl Implementation
 * @{
 */

/* *************************************
*  Includes & Memory related functions
***************************************/
#include <string.h>   /* memcmp, memcpy */
#include <limits.h>   /* ULLONG_MAX */

#if defined(XXH_NO_STREAM)
/* nothing */
#elif defined(XXH_NO_STDLIB)

/* When requesting to disable any mention of stdlib,
 * the library loses the ability to invoked malloc / free.
 * In practice, it means that functions like `XXH*_createState()`
 * will always fail, and return NULL.
 * This flag is useful in situations where
 * xxhash.h is integrated into some kernel, embedded or limited environment
 * without access to dynamic allocation.
 */

#if defined (__cplusplus)
extern "C" {
#endif

static XXH_CONSTF void* XXH_malloc(size_t s) { (void)s; return NULL; }
static void XXH_free(void* p) { (void)p; }

#if defined (__cplusplus)
} /* extern "C" */
#endif

#else

/*
 * Modify the local functions below should you wish to use
 * different memory routines for malloc() and free()
 */
#include <stdlib.h>

#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * @internal
 * @brief Modify this function to use a different routine than malloc().
 */
static XXH_MALLOCF void* XXH_malloc(size_t s) { return malloc(s); }

/*!
 * @internal
 * @brief Modify this function to use a different routine than free().
 */
static void XXH_free(void* p) { free(p); }

#if defined (__cplusplus)
} /* extern "C" */
#endif

#endif  /* XXH_NO_STDLIB */

#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * @internal
 * @brief Modify this function to use a different routine than memcpy().
 */
static void* XXH_memcpy(void* dest, const void* src, size_t size)
{
    return memcpy(dest,src,size);
}

#if defined (__cplusplus)
} /* extern "C" */
#endif

/* *************************************
*  Compiler Specific Options
***************************************/
#ifdef _MSC_VER /* Visual Studio warning fix */
#  pragma warning(disable : 4127) /* disable: C4127: conditional expression is constant */
#endif

#if XXH_NO_INLINE_HINTS  /* disable inlining hints */
#  if defined(__GNUC__) || defined(__clang__)
#    define XXH_FORCE_INLINE static __attribute__((unused))
#  else
#    define XXH_FORCE_INLINE static
#  endif
#  define XXH_NO_INLINE static
/* enable inlining hints */
#elif defined(__GNUC__) || defined(__clang__)
#  define XXH_FORCE_INLINE static __inline__ __attribute__((always_inline, unused))
#  define XXH_NO_INLINE static __attribute__((noinline))
#elif defined(_MSC_VER)  /* Visual Studio */
#  define XXH_FORCE_INLINE static __forceinline
#  define XXH_NO_INLINE static __declspec(noinline)
#elif defined (__cplusplus) \
  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L))   /* C99 */
#  define XXH_FORCE_INLINE static inline
#  define XXH_NO_INLINE static
#else
#  define XXH_FORCE_INLINE static
#  define XXH_NO_INLINE static
#endif

#if XXH3_INLINE_SECRET
#  define XXH3_WITH_SECRET_INLINE XXH_FORCE_INLINE
#else
#  define XXH3_WITH_SECRET_INLINE XXH_NO_INLINE
#endif


/* *************************************
*  Debug
***************************************/
/*!
 * @ingroup tuning
 * @def XXH_DEBUGLEVEL
 * @brief Sets the debugging level.
 *
 * XXH_DEBUGLEVEL is expected to be defined externally, typically via the
 * compiler's command line options. The value must be a number.
 */
#ifndef XXH_DEBUGLEVEL
#  ifdef DEBUGLEVEL /* backwards compat */
#    define XXH_DEBUGLEVEL DEBUGLEVEL
#  else
#    define XXH_DEBUGLEVEL 0
#  endif
#endif

#if (XXH_DEBUGLEVEL>=1)
#  include <assert.h>   /* note: can still be disabled with NDEBUG */
#  define XXH_ASSERT(c)   assert(c)
#else
#  if defined(__INTEL_COMPILER)
#    define XXH_ASSERT(c)   XXH_ASSUME((unsigned char) (c))
#  else
#    define XXH_ASSERT(c)   XXH_ASSUME(c)
#  endif
#endif

/* note: use after variable declarations */
#ifndef XXH_STATIC_ASSERT
#  if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)    /* C11 */
#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { _Static_assert((c),m); } while(0)
#  elif defined(__cplusplus) && (__cplusplus >= 201103L)            /* C++11 */
#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { static_assert((c),m); } while(0)
#  else
#    define XXH_STATIC_ASSERT_WITH_MESSAGE(c,m) do { struct xxh_sa { char x[(c) ? 1 : -1]; }; } while(0)
#  endif
#  define XXH_STATIC_ASSERT(c) XXH_STATIC_ASSERT_WITH_MESSAGE((c),#c)
#endif

/*!
 * @internal
 * @def XXH_COMPILER_GUARD(var)
 * @brief Used to prevent unwanted optimizations for @p var.
 *
 * It uses an empty GCC inline assembly statement with a register constraint
 * which forces @p var into a general purpose register (eg eax, ebx, ecx
 * on x86) and marks it as modified.
 *
 * This is used in a few places to avoid unwanted autovectorization (e.g.
 * XXH32_round()). All vectorization we want is explicit via intrinsics,
 * and _usually_ isn't wanted elsewhere.
 *
 * We also use it to prevent unwanted constant folding for AArch64 in
 * XXH3_initCustomSecret_scalar().
 */
#if defined(__GNUC__) || defined(__clang__)
#  define XXH_COMPILER_GUARD(var) __asm__("" : "+r" (var))
#else
#  define XXH_COMPILER_GUARD(var) ((void)0)
#endif

/* Specifically for NEON vectors which use the "w" constraint, on
 * Clang. */
#if defined(__clang__) && defined(__ARM_ARCH) && !defined(__wasm__)
#  define XXH_COMPILER_GUARD_CLANG_NEON(var) __asm__("" : "+w" (var))
#else
#  define XXH_COMPILER_GUARD_CLANG_NEON(var) ((void)0)
#endif

/* *************************************
*  Basic Types
***************************************/
#if !defined (__VMS) \
 && (defined (__cplusplus) \
 || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
# ifdef _AIX
#   include <inttypes.h>
# else
#   include <stdint.h>
# endif
  typedef uint8_t xxh_u8;
#else
  typedef unsigned char xxh_u8;
#endif
typedef XXH32_hash_t xxh_u32;

#ifdef XXH_OLD_NAMES
#  warning "XXH_OLD_NAMES is planned to be removed starting v0.9. If the program depends on it, consider moving away from it by employing newer type names directly"
#  define BYTE xxh_u8
#  define U8   xxh_u8
#  define U32  xxh_u32
#endif

#if defined (__cplusplus)
extern "C" {
#endif

/* ***   Memory access   *** */

/*!
 * @internal
 * @fn xxh_u32 XXH_read32(const void* ptr)
 * @brief Reads an unaligned 32-bit integer from @p ptr in native endianness.
 *
 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
 *
 * @param ptr The pointer to read from.
 * @return The 32-bit native endian integer from the bytes at @p ptr.
 */

/*!
 * @internal
 * @fn xxh_u32 XXH_readLE32(const void* ptr)
 * @brief Reads an unaligned 32-bit little endian integer from @p ptr.
 *
 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
 *
 * @param ptr The pointer to read from.
 * @return The 32-bit little endian integer from the bytes at @p ptr.
 */

/*!
 * @internal
 * @fn xxh_u32 XXH_readBE32(const void* ptr)
 * @brief Reads an unaligned 32-bit big endian integer from @p ptr.
 *
 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
 *
 * @param ptr The pointer to read from.
 * @return The 32-bit big endian integer from the bytes at @p ptr.
 */

/*!
 * @internal
 * @fn xxh_u32 XXH_readLE32_align(const void* ptr, XXH_alignment align)
 * @brief Like @ref XXH_readLE32(), but has an option for aligned reads.
 *
 * Affected by @ref XXH_FORCE_MEMORY_ACCESS.
 * Note that when @ref XXH_FORCE_ALIGN_CHECK == 0, the @p align parameter is
 * always @ref XXH_alignment::XXH_unaligned.
 *
 * @param ptr The pointer to read from.
 * @param align Whether @p ptr is aligned.
 * @pre
 *   If @p align == @ref XXH_alignment::XXH_aligned, @p ptr must be 4 byte
 *   aligned.
 * @return The 32-bit little endian integer from the bytes at @p ptr.
 */

#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
/*
 * Manual byteshift. Best for old compilers which don't inline memcpy.
 * We actually directly use XXH_readLE32 and XXH_readBE32.
 */
#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))

/*
 * Force direct memory access. Only works on CPU which support unaligned memory
 * access in hardware.
 */
static xxh_u32 XXH_read32(const void* memPtr) { return *(const xxh_u32*) memPtr; }

#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))

/*
 * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
 * documentation claimed that it only increased the alignment, but actually it
 * can decrease it on gcc, clang, and icc:
 * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
 * https://gcc.godbolt.org/z/xYez1j67Y.
 */
#ifdef XXH_OLD_NAMES
typedef union { xxh_u32 u32; } __attribute__((packed)) unalign;
#endif
static xxh_u32 XXH_read32(const void* ptr)
{
    typedef __attribute__((aligned(1))) xxh_u32 xxh_unalign32;
    return *((const xxh_unalign32*)ptr);
}

#else

/*
 * Portable and safe solution. Generally efficient.
 * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
 */
static xxh_u32 XXH_read32(const void* memPtr)
{
    xxh_u32 val;
    XXH_memcpy(&val, memPtr, sizeof(val));
    return val;
}

#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */


/* ***   Endianness   *** */

/*!
 * @ingroup tuning
 * @def XXH_CPU_LITTLE_ENDIAN
 * @brief Whether the target is little endian.
 *
 * Defined to 1 if the target is little endian, or 0 if it is big endian.
 * It can be defined externally, for example on the compiler command line.
 *
 * If it is not defined,
 * a runtime check (which is usually constant folded) is used instead.
 *
 * @note
 *   This is not necessarily defined to an integer constant.
 *
 * @see XXH_isLittleEndian() for the runtime check.
 */
#ifndef XXH_CPU_LITTLE_ENDIAN
/*
 * Try to detect endianness automatically, to avoid the nonstandard behavior
 * in `XXH_isLittleEndian()`
 */
#  if defined(_WIN32) /* Windows is always little endian */ \
     || defined(__LITTLE_ENDIAN__) \
     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__)
#    define XXH_CPU_LITTLE_ENDIAN 1
#  elif defined(__BIG_ENDIAN__) \
     || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
#    define XXH_CPU_LITTLE_ENDIAN 0
#  else
/*!
 * @internal
 * @brief Runtime check for @ref XXH_CPU_LITTLE_ENDIAN.
 *
 * Most compilers will constant fold this.
 */
static int XXH_isLittleEndian(void)
{
    /*
     * Portable and well-defined behavior.
     * Don't use static: it is detrimental to performance.
     */
    const union { xxh_u32 u; xxh_u8 c[4]; } one = { 1 };
    return one.c[0];
}
#   define XXH_CPU_LITTLE_ENDIAN   XXH_isLittleEndian()
#  endif
#endif


/* ****************************************
*  Compiler-specific Functions and Macros
******************************************/
#define XXH_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)

#ifdef __has_builtin
#  define XXH_HAS_BUILTIN(x) __has_builtin(x)
#else
#  define XXH_HAS_BUILTIN(x) 0
#endif


/*
 * C23 and future versions have standard "unreachable()".
 * Once it has been implemented reliably we can add it as an
 * additional case:
 *
 * ```
 * #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= XXH_C23_VN)
 * #  include <stddef.h>
 * #  ifdef unreachable
 * #    define XXH_UNREACHABLE() unreachable()
 * #  endif
 * #endif
 * ```
 *
 * Note C++23 also has std::unreachable() which can be detected
 * as follows:
 * ```
 * #if defined(__cpp_lib_unreachable) && (__cpp_lib_unreachable >= 202202L)
 * #  include <utility>
 * #  define XXH_UNREACHABLE() std::unreachable()
 * #endif
 * ```
 * NB: `__cpp_lib_unreachable` is defined in the `<version>` header.
 * We don't use that as including `<utility>` in `extern "C"` blocks
 * doesn't work on GCC12
 */

#if XXH_HAS_BUILTIN(__builtin_unreachable)
#  define XXH_UNREACHABLE() __builtin_unreachable()

#elif defined(_MSC_VER)
#  define XXH_UNREACHABLE() __assume(0)

#else
#  define XXH_UNREACHABLE()
#endif

#if XXH_HAS_BUILTIN(__builtin_assume)
#  define XXH_ASSUME(c) __builtin_assume(c)
#else
#  define XXH_ASSUME(c) if (!(c)) { XXH_UNREACHABLE(); }
#endif

/*!
 * @internal
 * @def XXH_rotl32(x,r)
 * @brief 32-bit rotate left.
 *
 * @param x The 32-bit integer to be rotated.
 * @param r The number of bits to rotate.
 * @pre
 *   @p r > 0 && @p r < 32
 * @note
 *   @p x and @p r may be evaluated multiple times.
 * @return The rotated result.
 */
#if !defined(NO_CLANG_BUILTIN) && XXH_HAS_BUILTIN(__builtin_rotateleft32) \
                               && XXH_HAS_BUILTIN(__builtin_rotateleft64)
#  define XXH_rotl32 __builtin_rotateleft32
#  define XXH_rotl64 __builtin_rotateleft64
/* Note: although _rotl exists for minGW (GCC under windows), performance seems poor */
#elif defined(_MSC_VER)
#  define XXH_rotl32(x,r) _rotl(x,r)
#  define XXH_rotl64(x,r) _rotl64(x,r)
#else
#  define XXH_rotl32(x,r) (((x) << (r)) | ((x) >> (32 - (r))))
#  define XXH_rotl64(x,r) (((x) << (r)) | ((x) >> (64 - (r))))
#endif

/*!
 * @internal
 * @fn xxh_u32 XXH_swap32(xxh_u32 x)
 * @brief A 32-bit byteswap.
 *
 * @param x The 32-bit integer to byteswap.
 * @return @p x, byteswapped.
 */
#if defined(_MSC_VER)     /* Visual Studio */
#  define XXH_swap32 _byteswap_ulong
#elif XXH_GCC_VERSION >= 403
#  define XXH_swap32 __builtin_bswap32
#else
static xxh_u32 XXH_swap32 (xxh_u32 x)
{
    return  ((x << 24) & 0xff000000 ) |
            ((x <<  8) & 0x00ff0000 ) |
            ((x >>  8) & 0x0000ff00 ) |
            ((x >> 24) & 0x000000ff );
}
#endif


/* ***************************
*  Memory reads
*****************************/

/*!
 * @internal
 * @brief Enum to indicate whether a pointer is aligned.
 */
typedef enum {
    XXH_aligned,  /*!< Aligned */
    XXH_unaligned /*!< Possibly unaligned */
} XXH_alignment;

/*
 * XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load.
 *
 * This is ideal for older compilers which don't inline memcpy.
 */
#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))

XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* memPtr)
{
    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
    return bytePtr[0]
         | ((xxh_u32)bytePtr[1] << 8)
         | ((xxh_u32)bytePtr[2] << 16)
         | ((xxh_u32)bytePtr[3] << 24);
}

XXH_FORCE_INLINE xxh_u32 XXH_readBE32(const void* memPtr)
{
    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
    return bytePtr[3]
         | ((xxh_u32)bytePtr[2] << 8)
         | ((xxh_u32)bytePtr[1] << 16)
         | ((xxh_u32)bytePtr[0] << 24);
}

#else
XXH_FORCE_INLINE xxh_u32 XXH_readLE32(const void* ptr)
{
    return XXH_CPU_LITTLE_ENDIAN ? XXH_read32(ptr) : XXH_swap32(XXH_read32(ptr));
}

static xxh_u32 XXH_readBE32(const void* ptr)
{
    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap32(XXH_read32(ptr)) : XXH_read32(ptr);
}
#endif

XXH_FORCE_INLINE xxh_u32
XXH_readLE32_align(const void* ptr, XXH_alignment align)
{
    if (align==XXH_unaligned) {
        return XXH_readLE32(ptr);
    } else {
        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u32*)ptr : XXH_swap32(*(const xxh_u32*)ptr);
    }
}


/* *************************************
*  Misc
***************************************/
/*! @ingroup public */
XXH_PUBLIC_API unsigned XXH_versionNumber (void) { return XXH_VERSION_NUMBER; }


/* *******************************************************************
*  32-bit hash functions
*********************************************************************/
/*!
 * @}
 * @defgroup XXH32_impl XXH32 implementation
 * @ingroup impl
 *
 * Details on the XXH32 implementation.
 * @{
 */
 /* #define instead of static const, to be used as initializers */
#define XXH_PRIME32_1  0x9E3779B1U  /*!< 0b10011110001101110111100110110001 */
#define XXH_PRIME32_2  0x85EBCA77U  /*!< 0b10000101111010111100101001110111 */
#define XXH_PRIME32_3  0xC2B2AE3DU  /*!< 0b11000010101100101010111000111101 */
#define XXH_PRIME32_4  0x27D4EB2FU  /*!< 0b00100111110101001110101100101111 */
#define XXH_PRIME32_5  0x165667B1U  /*!< 0b00010110010101100110011110110001 */

#ifdef XXH_OLD_NAMES
#  define PRIME32_1 XXH_PRIME32_1
#  define PRIME32_2 XXH_PRIME32_2
#  define PRIME32_3 XXH_PRIME32_3
#  define PRIME32_4 XXH_PRIME32_4
#  define PRIME32_5 XXH_PRIME32_5
#endif

/*!
 * @internal
 * @brief Normal stripe processing routine.
 *
 * This shuffles the bits so that any bit from @p input impacts several bits in
 * @p acc.
 *
 * @param acc The accumulator lane.
 * @param input The stripe of input to mix.
 * @return The mixed accumulator lane.
 */
static xxh_u32 XXH32_round(xxh_u32 acc, xxh_u32 input)
{
    acc += input * XXH_PRIME32_2;
    acc  = XXH_rotl32(acc, 13);
    acc *= XXH_PRIME32_1;
#if (defined(__SSE4_1__) || defined(__aarch64__) || defined(__wasm_simd128__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
    /*
     * UGLY HACK:
     * A compiler fence is the only thing that prevents GCC and Clang from
     * autovectorizing the XXH32 loop (pragmas and attributes don't work for some
     * reason) without globally disabling SSE4.1.
     *
     * The reason we want to avoid vectorization is because despite working on
     * 4 integers at a time, there are multiple factors slowing XXH32 down on
     * SSE4:
     * - There's a ridiculous amount of lag from pmulld (10 cycles of latency on
     *   newer chips!) making it slightly slower to multiply four integers at
     *   once compared to four integers independently. Even when pmulld was
     *   fastest, Sandy/Ivy Bridge, it is still not worth it to go into SSE
     *   just to multiply unless doing a long operation.
     *
     * - Four instructions are required to rotate,
     *      movqda tmp,  v // not required with VEX encoding
     *      pslld  tmp, 13 // tmp <<= 13
     *      psrld  v,   19 // x >>= 19
     *      por    v,  tmp // x |= tmp
     *   compared to one for scalar:
     *      roll   v, 13    // reliably fast across the board
     *      shldl  v, v, 13 // Sandy Bridge and later prefer this for some reason
     *
     * - Instruction level parallelism is actually more beneficial here because
     *   the SIMD actually serializes this operation: While v1 is rotating, v2
     *   can load data, while v3 can multiply. SSE forces them to operate
     *   together.
     *
     * This is also enabled on AArch64, as Clang is *very aggressive* in vectorizing
     * the loop. NEON is only faster on the A53, and with the newer cores, it is less
     * than half the speed.
     *
     * Additionally, this is used on WASM SIMD128 because it JITs to the same
     * SIMD instructions and has the same issue.
     */
    XXH_COMPILER_GUARD(acc);
#endif
    return acc;
}

/*!
 * @internal
 * @brief Mixes all bits to finalize the hash.
 *
 * The final mix ensures that all input bits have a chance to impact any bit in
 * the output digest, resulting in an unbiased distribution.
 *
 * @param hash The hash to avalanche.
 * @return The avalanched hash.
 */
static xxh_u32 XXH32_avalanche(xxh_u32 hash)
{
    hash ^= hash >> 15;
    hash *= XXH_PRIME32_2;
    hash ^= hash >> 13;
    hash *= XXH_PRIME32_3;
    hash ^= hash >> 16;
    return hash;
}

#define XXH_get32bits(p) XXH_readLE32_align(p, align)

/*!
 * @internal
 * @brief Processes the last 0-15 bytes of @p ptr.
 *
 * There may be up to 15 bytes remaining to consume from the input.
 * This final stage will digest them to ensure that all input bytes are present
 * in the final mix.
 *
 * @param hash The hash to finalize.
 * @param ptr The pointer to the remaining input.
 * @param len The remaining length, modulo 16.
 * @param align Whether @p ptr is aligned.
 * @return The finalized hash.
 * @see XXH64_finalize().
 */
static XXH_PUREF xxh_u32
XXH32_finalize(xxh_u32 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
{
#define XXH_PROCESS1 do {                             \
    hash += (*ptr++) * XXH_PRIME32_5;                 \
    hash = XXH_rotl32(hash, 11) * XXH_PRIME32_1;      \
} while (0)

#define XXH_PROCESS4 do {                             \
    hash += XXH_get32bits(ptr) * XXH_PRIME32_3;       \
    ptr += 4;                                         \
    hash  = XXH_rotl32(hash, 17) * XXH_PRIME32_4;     \
} while (0)

    if (ptr==NULL) XXH_ASSERT(len == 0);

    /* Compact rerolled version; generally faster */
    if (!XXH32_ENDJMP) {
        len &= 15;
        while (len >= 4) {
            XXH_PROCESS4;
            len -= 4;
        }
        while (len > 0) {
            XXH_PROCESS1;
            --len;
        }
        return XXH32_avalanche(hash);
    } else {
         switch(len&15) /* or switch(bEnd - p) */ {
           case 12:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 8:       XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 4:       XXH_PROCESS4;
                         return XXH32_avalanche(hash);

           case 13:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 9:       XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 5:       XXH_PROCESS4;
                         XXH_PROCESS1;
                         return XXH32_avalanche(hash);

           case 14:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 10:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 6:       XXH_PROCESS4;
                         XXH_PROCESS1;
                         XXH_PROCESS1;
                         return XXH32_avalanche(hash);

           case 15:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 11:      XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 7:       XXH_PROCESS4;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 3:       XXH_PROCESS1;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 2:       XXH_PROCESS1;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 1:       XXH_PROCESS1;
                         XXH_FALLTHROUGH;  /* fallthrough */
           case 0:       return XXH32_avalanche(hash);
        }
        XXH_ASSERT(0);
        return hash;   /* reaching this point is deemed impossible */
    }
}

#ifdef XXH_OLD_NAMES
#  define PROCESS1 XXH_PROCESS1
#  define PROCESS4 XXH_PROCESS4
#else
#  undef XXH_PROCESS1
#  undef XXH_PROCESS4
#endif

/*!
 * @internal
 * @brief The implementation for @ref XXH32().
 *
 * @param input , len , seed Directly passed from @ref XXH32().
 * @param align Whether @p input is aligned.
 * @return The calculated hash.
 */
XXH_FORCE_INLINE XXH_PUREF xxh_u32
XXH32_endian_align(const xxh_u8* input, size_t len, xxh_u32 seed, XXH_alignment align)
{
    xxh_u32 h32;

    if (input==NULL) XXH_ASSERT(len == 0);

    if (len>=16) {
        const xxh_u8* const bEnd = input + len;
        const xxh_u8* const limit = bEnd - 15;
        xxh_u32 v1 = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
        xxh_u32 v2 = seed + XXH_PRIME32_2;
        xxh_u32 v3 = seed + 0;
        xxh_u32 v4 = seed - XXH_PRIME32_1;

        do {
            v1 = XXH32_round(v1, XXH_get32bits(input)); input += 4;
            v2 = XXH32_round(v2, XXH_get32bits(input)); input += 4;
            v3 = XXH32_round(v3, XXH_get32bits(input)); input += 4;
            v4 = XXH32_round(v4, XXH_get32bits(input)); input += 4;
        } while (input < limit);

        h32 = XXH_rotl32(v1, 1)  + XXH_rotl32(v2, 7)
            + XXH_rotl32(v3, 12) + XXH_rotl32(v4, 18);
    } else {
        h32  = seed + XXH_PRIME32_5;
    }

    h32 += (xxh_u32)len;

    return XXH32_finalize(h32, input, len&15, align);
}

/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH32_hash_t XXH32 (const void* input, size_t len, XXH32_hash_t seed)
{
#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
    XXH32_state_t state;
    XXH32_reset(&state, seed);
    XXH32_update(&state, (const xxh_u8*)input, len);
    return XXH32_digest(&state);
#else
    if (XXH_FORCE_ALIGN_CHECK) {
        if ((((size_t)input) & 3) == 0) {   /* Input is 4-bytes aligned, leverage the speed benefit */
            return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
    }   }

    return XXH32_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);
#endif
}


/*******   Hash streaming   *******/
#ifndef XXH_NO_STREAM
/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH32_state_t* XXH32_createState(void)
{
    return (XXH32_state_t*)XXH_malloc(sizeof(XXH32_state_t));
}
/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH_errorcode XXH32_freeState(XXH32_state_t* statePtr)
{
    XXH_free(statePtr);
    return XXH_OK;
}

/*! @ingroup XXH32_family */
XXH_PUBLIC_API void XXH32_copyState(XXH32_state_t* dstState, const XXH32_state_t* srcState)
{
    XXH_memcpy(dstState, srcState, sizeof(*dstState));
}

/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH_errorcode XXH32_reset(XXH32_state_t* statePtr, XXH32_hash_t seed)
{
    XXH_ASSERT(statePtr != NULL);
    memset(statePtr, 0, sizeof(*statePtr));
    statePtr->v[0] = seed + XXH_PRIME32_1 + XXH_PRIME32_2;
    statePtr->v[1] = seed + XXH_PRIME32_2;
    statePtr->v[2] = seed + 0;
    statePtr->v[3] = seed - XXH_PRIME32_1;
    return XXH_OK;
}


/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH_errorcode
XXH32_update(XXH32_state_t* state, const void* input, size_t len)
{
    if (input==NULL) {
        XXH_ASSERT(len == 0);
        return XXH_OK;
    }

    {   const xxh_u8* p = (const xxh_u8*)input;
        const xxh_u8* const bEnd = p + len;

        state->total_len_32 += (XXH32_hash_t)len;
        state->large_len |= (XXH32_hash_t)((len>=16) | (state->total_len_32>=16));

        if (state->memsize + len < 16)  {   /* fill in tmp buffer */
            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, len);
            state->memsize += (XXH32_hash_t)len;
            return XXH_OK;
        }

        if (state->memsize) {   /* some data left from previous update */
            XXH_memcpy((xxh_u8*)(state->mem32) + state->memsize, input, 16-state->memsize);
            {   const xxh_u32* p32 = state->mem32;
                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p32)); p32++;
                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p32)); p32++;
                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p32)); p32++;
                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p32));
            }
            p += 16-state->memsize;
            state->memsize = 0;
        }

        if (p <= bEnd-16) {
            const xxh_u8* const limit = bEnd - 16;

            do {
                state->v[0] = XXH32_round(state->v[0], XXH_readLE32(p)); p+=4;
                state->v[1] = XXH32_round(state->v[1], XXH_readLE32(p)); p+=4;
                state->v[2] = XXH32_round(state->v[2], XXH_readLE32(p)); p+=4;
                state->v[3] = XXH32_round(state->v[3], XXH_readLE32(p)); p+=4;
            } while (p<=limit);

        }

        if (p < bEnd) {
            XXH_memcpy(state->mem32, p, (size_t)(bEnd-p));
            state->memsize = (unsigned)(bEnd-p);
        }
    }

    return XXH_OK;
}


/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH32_hash_t XXH32_digest(const XXH32_state_t* state)
{
    xxh_u32 h32;

    if (state->large_len) {
        h32 = XXH_rotl32(state->v[0], 1)
            + XXH_rotl32(state->v[1], 7)
            + XXH_rotl32(state->v[2], 12)
            + XXH_rotl32(state->v[3], 18);
    } else {
        h32 = state->v[2] /* == seed */ + XXH_PRIME32_5;
    }

    h32 += state->total_len_32;

    return XXH32_finalize(h32, (const xxh_u8*)state->mem32, state->memsize, XXH_aligned);
}
#endif /* !XXH_NO_STREAM */

/*******   Canonical representation   *******/

/*! @ingroup XXH32_family */
XXH_PUBLIC_API void XXH32_canonicalFromHash(XXH32_canonical_t* dst, XXH32_hash_t hash)
{
    XXH_STATIC_ASSERT(sizeof(XXH32_canonical_t) == sizeof(XXH32_hash_t));
    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap32(hash);
    XXH_memcpy(dst, &hash, sizeof(*dst));
}
/*! @ingroup XXH32_family */
XXH_PUBLIC_API XXH32_hash_t XXH32_hashFromCanonical(const XXH32_canonical_t* src)
{
    return XXH_readBE32(src);
}


#ifndef XXH_NO_LONG_LONG

/* *******************************************************************
*  64-bit hash functions
*********************************************************************/
/*!
 * @}
 * @ingroup impl
 * @{
 */
/*******   Memory access   *******/

typedef XXH64_hash_t xxh_u64;

#ifdef XXH_OLD_NAMES
#  define U64 xxh_u64
#endif

#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))
/*
 * Manual byteshift. Best for old compilers which don't inline memcpy.
 * We actually directly use XXH_readLE64 and XXH_readBE64.
 */
#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==2))

/* Force direct memory access. Only works on CPU which support unaligned memory access in hardware */
static xxh_u64 XXH_read64(const void* memPtr)
{
    return *(const xxh_u64*) memPtr;
}

#elif (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==1))

/*
 * __attribute__((aligned(1))) is supported by gcc and clang. Originally the
 * documentation claimed that it only increased the alignment, but actually it
 * can decrease it on gcc, clang, and icc:
 * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69502,
 * https://gcc.godbolt.org/z/xYez1j67Y.
 */
#ifdef XXH_OLD_NAMES
typedef union { xxh_u32 u32; xxh_u64 u64; } __attribute__((packed)) unalign64;
#endif
static xxh_u64 XXH_read64(const void* ptr)
{
    typedef __attribute__((aligned(1))) xxh_u64 xxh_unalign64;
    return *((const xxh_unalign64*)ptr);
}

#else

/*
 * Portable and safe solution. Generally efficient.
 * see: https://fastcompression.blogspot.com/2015/08/accessing-unaligned-memory.html
 */
static xxh_u64 XXH_read64(const void* memPtr)
{
    xxh_u64 val;
    XXH_memcpy(&val, memPtr, sizeof(val));
    return val;
}

#endif   /* XXH_FORCE_DIRECT_MEMORY_ACCESS */

#if defined(_MSC_VER)     /* Visual Studio */
#  define XXH_swap64 _byteswap_uint64
#elif XXH_GCC_VERSION >= 403
#  define XXH_swap64 __builtin_bswap64
#else
static xxh_u64 XXH_swap64(xxh_u64 x)
{
    return  ((x << 56) & 0xff00000000000000ULL) |
            ((x << 40) & 0x00ff000000000000ULL) |
            ((x << 24) & 0x0000ff0000000000ULL) |
            ((x << 8)  & 0x000000ff00000000ULL) |
            ((x >> 8)  & 0x00000000ff000000ULL) |
            ((x >> 24) & 0x0000000000ff0000ULL) |
            ((x >> 40) & 0x000000000000ff00ULL) |
            ((x >> 56) & 0x00000000000000ffULL);
}
#endif


/* XXH_FORCE_MEMORY_ACCESS==3 is an endian-independent byteshift load. */
#if (defined(XXH_FORCE_MEMORY_ACCESS) && (XXH_FORCE_MEMORY_ACCESS==3))

XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* memPtr)
{
    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
    return bytePtr[0]
         | ((xxh_u64)bytePtr[1] << 8)
         | ((xxh_u64)bytePtr[2] << 16)
         | ((xxh_u64)bytePtr[3] << 24)
         | ((xxh_u64)bytePtr[4] << 32)
         | ((xxh_u64)bytePtr[5] << 40)
         | ((xxh_u64)bytePtr[6] << 48)
         | ((xxh_u64)bytePtr[7] << 56);
}

XXH_FORCE_INLINE xxh_u64 XXH_readBE64(const void* memPtr)
{
    const xxh_u8* bytePtr = (const xxh_u8 *)memPtr;
    return bytePtr[7]
         | ((xxh_u64)bytePtr[6] << 8)
         | ((xxh_u64)bytePtr[5] << 16)
         | ((xxh_u64)bytePtr[4] << 24)
         | ((xxh_u64)bytePtr[3] << 32)
         | ((xxh_u64)bytePtr[2] << 40)
         | ((xxh_u64)bytePtr[1] << 48)
         | ((xxh_u64)bytePtr[0] << 56);
}

#else
XXH_FORCE_INLINE xxh_u64 XXH_readLE64(const void* ptr)
{
    return XXH_CPU_LITTLE_ENDIAN ? XXH_read64(ptr) : XXH_swap64(XXH_read64(ptr));
}

static xxh_u64 XXH_readBE64(const void* ptr)
{
    return XXH_CPU_LITTLE_ENDIAN ? XXH_swap64(XXH_read64(ptr)) : XXH_read64(ptr);
}
#endif

XXH_FORCE_INLINE xxh_u64
XXH_readLE64_align(const void* ptr, XXH_alignment align)
{
    if (align==XXH_unaligned)
        return XXH_readLE64(ptr);
    else
        return XXH_CPU_LITTLE_ENDIAN ? *(const xxh_u64*)ptr : XXH_swap64(*(const xxh_u64*)ptr);
}


/*******   xxh64   *******/
/*!
 * @}
 * @defgroup XXH64_impl XXH64 implementation
 * @ingroup impl
 *
 * Details on the XXH64 implementation.
 * @{
 */
/* #define rather that static const, to be used as initializers */
#define XXH_PRIME64_1  0x9E3779B185EBCA87ULL  /*!< 0b1001111000110111011110011011000110000101111010111100101010000111 */
#define XXH_PRIME64_2  0xC2B2AE3D27D4EB4FULL  /*!< 0b1100001010110010101011100011110100100111110101001110101101001111 */
#define XXH_PRIME64_3  0x165667B19E3779F9ULL  /*!< 0b0001011001010110011001111011000110011110001101110111100111111001 */
#define XXH_PRIME64_4  0x85EBCA77C2B2AE63ULL  /*!< 0b1000010111101011110010100111011111000010101100101010111001100011 */
#define XXH_PRIME64_5  0x27D4EB2F165667C5ULL  /*!< 0b0010011111010100111010110010111100010110010101100110011111000101 */

#ifdef XXH_OLD_NAMES
#  define PRIME64_1 XXH_PRIME64_1
#  define PRIME64_2 XXH_PRIME64_2
#  define PRIME64_3 XXH_PRIME64_3
#  define PRIME64_4 XXH_PRIME64_4
#  define PRIME64_5 XXH_PRIME64_5
#endif

/*! @copydoc XXH32_round */
static xxh_u64 XXH64_round(xxh_u64 acc, xxh_u64 input)
{
    acc += input * XXH_PRIME64_2;
    acc  = XXH_rotl64(acc, 31);
    acc *= XXH_PRIME64_1;
#if (defined(__AVX512F__)) && !defined(XXH_ENABLE_AUTOVECTORIZE)
    /*
     * DISABLE AUTOVECTORIZATION:
     * A compiler fence is used to prevent GCC and Clang from
     * autovectorizing the XXH64 loop (pragmas and attributes don't work for some
     * reason) without globally disabling AVX512.
     *
     * Autovectorization of XXH64 tends to be detrimental,
     * though the exact outcome may change depending on exact cpu and compiler version.
     * For information, it has been reported as detrimental for Skylake-X,
     * but possibly beneficial for Zen4.
     *
     * The default is to disable auto-vectorization,
     * but you can select to enable it instead using `XXH_ENABLE_AUTOVECTORIZE` build variable.
     */
    XXH_COMPILER_GUARD(acc);
#endif
    return acc;
}

static xxh_u64 XXH64_mergeRound(xxh_u64 acc, xxh_u64 val)
{
    val  = XXH64_round(0, val);
    acc ^= val;
    acc  = acc * XXH_PRIME64_1 + XXH_PRIME64_4;
    return acc;
}

/*! @copydoc XXH32_avalanche */
static xxh_u64 XXH64_avalanche(xxh_u64 hash)
{
    hash ^= hash >> 33;
    hash *= XXH_PRIME64_2;
    hash ^= hash >> 29;
    hash *= XXH_PRIME64_3;
    hash ^= hash >> 32;
    return hash;
}


#define XXH_get64bits(p) XXH_readLE64_align(p, align)

/*!
 * @internal
 * @brief Processes the last 0-31 bytes of @p ptr.
 *
 * There may be up to 31 bytes remaining to consume from the input.
 * This final stage will digest them to ensure that all input bytes are present
 * in the final mix.
 *
 * @param hash The hash to finalize.
 * @param ptr The pointer to the remaining input.
 * @param len The remaining length, modulo 32.
 * @param align Whether @p ptr is aligned.
 * @return The finalized hash
 * @see XXH32_finalize().
 */
static XXH_PUREF xxh_u64
XXH64_finalize(xxh_u64 hash, const xxh_u8* ptr, size_t len, XXH_alignment align)
{
    if (ptr==NULL) XXH_ASSERT(len == 0);
    len &= 31;
    while (len >= 8) {
        xxh_u64 const k1 = XXH64_round(0, XXH_get64bits(ptr));
        ptr += 8;
        hash ^= k1;
        hash  = XXH_rotl64(hash,27) * XXH_PRIME64_1 + XXH_PRIME64_4;
        len -= 8;
    }
    if (len >= 4) {
        hash ^= (xxh_u64)(XXH_get32bits(ptr)) * XXH_PRIME64_1;
        ptr += 4;
        hash = XXH_rotl64(hash, 23) * XXH_PRIME64_2 + XXH_PRIME64_3;
        len -= 4;
    }
    while (len > 0) {
        hash ^= (*ptr++) * XXH_PRIME64_5;
        hash = XXH_rotl64(hash, 11) * XXH_PRIME64_1;
        --len;
    }
    return  XXH64_avalanche(hash);
}

#ifdef XXH_OLD_NAMES
#  define PROCESS1_64 XXH_PROCESS1_64
#  define PROCESS4_64 XXH_PROCESS4_64
#  define PROCESS8_64 XXH_PROCESS8_64
#else
#  undef XXH_PROCESS1_64
#  undef XXH_PROCESS4_64
#  undef XXH_PROCESS8_64
#endif

/*!
 * @internal
 * @brief The implementation for @ref XXH64().
 *
 * @param input , len , seed Directly passed from @ref XXH64().
 * @param align Whether @p input is aligned.
 * @return The calculated hash.
 */
XXH_FORCE_INLINE XXH_PUREF xxh_u64
XXH64_endian_align(const xxh_u8* input, size_t len, xxh_u64 seed, XXH_alignment align)
{
    xxh_u64 h64;
    if (input==NULL) XXH_ASSERT(len == 0);

    if (len>=32) {
        const xxh_u8* const bEnd = input + len;
        const xxh_u8* const limit = bEnd - 31;
        xxh_u64 v1 = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
        xxh_u64 v2 = seed + XXH_PRIME64_2;
        xxh_u64 v3 = seed + 0;
        xxh_u64 v4 = seed - XXH_PRIME64_1;

        do {
            v1 = XXH64_round(v1, XXH_get64bits(input)); input+=8;
            v2 = XXH64_round(v2, XXH_get64bits(input)); input+=8;
            v3 = XXH64_round(v3, XXH_get64bits(input)); input+=8;
            v4 = XXH64_round(v4, XXH_get64bits(input)); input+=8;
        } while (input<limit);

        h64 = XXH_rotl64(v1, 1) + XXH_rotl64(v2, 7) + XXH_rotl64(v3, 12) + XXH_rotl64(v4, 18);
        h64 = XXH64_mergeRound(h64, v1);
        h64 = XXH64_mergeRound(h64, v2);
        h64 = XXH64_mergeRound(h64, v3);
        h64 = XXH64_mergeRound(h64, v4);

    } else {
        h64  = seed + XXH_PRIME64_5;
    }

    h64 += (xxh_u64) len;

    return XXH64_finalize(h64, input, len, align);
}


/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH64_hash_t XXH64 (XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
{
#if !defined(XXH_NO_STREAM) && XXH_SIZE_OPT >= 2
    /* Simple version, good for code maintenance, but unfortunately slow for small inputs */
    XXH64_state_t state;
    XXH64_reset(&state, seed);
    XXH64_update(&state, (const xxh_u8*)input, len);
    return XXH64_digest(&state);
#else
    if (XXH_FORCE_ALIGN_CHECK) {
        if ((((size_t)input) & 7)==0) {  /* Input is aligned, let's leverage the speed advantage */
            return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_aligned);
    }   }

    return XXH64_endian_align((const xxh_u8*)input, len, seed, XXH_unaligned);

#endif
}

/*******   Hash Streaming   *******/
#ifndef XXH_NO_STREAM
/*! @ingroup XXH64_family*/
XXH_PUBLIC_API XXH64_state_t* XXH64_createState(void)
{
    return (XXH64_state_t*)XXH_malloc(sizeof(XXH64_state_t));
}
/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH_errorcode XXH64_freeState(XXH64_state_t* statePtr)
{
    XXH_free(statePtr);
    return XXH_OK;
}

/*! @ingroup XXH64_family */
XXH_PUBLIC_API void XXH64_copyState(XXH_NOESCAPE XXH64_state_t* dstState, const XXH64_state_t* srcState)
{
    XXH_memcpy(dstState, srcState, sizeof(*dstState));
}

/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH_errorcode XXH64_reset(XXH_NOESCAPE XXH64_state_t* statePtr, XXH64_hash_t seed)
{
    XXH_ASSERT(statePtr != NULL);
    memset(statePtr, 0, sizeof(*statePtr));
    statePtr->v[0] = seed + XXH_PRIME64_1 + XXH_PRIME64_2;
    statePtr->v[1] = seed + XXH_PRIME64_2;
    statePtr->v[2] = seed + 0;
    statePtr->v[3] = seed - XXH_PRIME64_1;
    return XXH_OK;
}

/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH_errorcode
XXH64_update (XXH_NOESCAPE XXH64_state_t* state, XXH_NOESCAPE const void* input, size_t len)
{
    if (input==NULL) {
        XXH_ASSERT(len == 0);
        return XXH_OK;
    }

    {   const xxh_u8* p = (const xxh_u8*)input;
        const xxh_u8* const bEnd = p + len;

        state->total_len += len;

        if (state->memsize + len < 32) {  /* fill in tmp buffer */
            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, len);
            state->memsize += (xxh_u32)len;
            return XXH_OK;
        }

        if (state->memsize) {   /* tmp buffer is full */
            XXH_memcpy(((xxh_u8*)state->mem64) + state->memsize, input, 32-state->memsize);
            state->v[0] = XXH64_round(state->v[0], XXH_readLE64(state->mem64+0));
            state->v[1] = XXH64_round(state->v[1], XXH_readLE64(state->mem64+1));
            state->v[2] = XXH64_round(state->v[2], XXH_readLE64(state->mem64+2));
            state->v[3] = XXH64_round(state->v[3], XXH_readLE64(state->mem64+3));
            p += 32 - state->memsize;
            state->memsize = 0;
        }

        if (p+32 <= bEnd) {
            const xxh_u8* const limit = bEnd - 32;

            do {
                state->v[0] = XXH64_round(state->v[0], XXH_readLE64(p)); p+=8;
                state->v[1] = XXH64_round(state->v[1], XXH_readLE64(p)); p+=8;
                state->v[2] = XXH64_round(state->v[2], XXH_readLE64(p)); p+=8;
                state->v[3] = XXH64_round(state->v[3], XXH_readLE64(p)); p+=8;
            } while (p<=limit);

        }

        if (p < bEnd) {
            XXH_memcpy(state->mem64, p, (size_t)(bEnd-p));
            state->memsize = (unsigned)(bEnd-p);
        }
    }

    return XXH_OK;
}


/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH64_hash_t XXH64_digest(XXH_NOESCAPE const XXH64_state_t* state)
{
    xxh_u64 h64;

    if (state->total_len >= 32) {
        h64 = XXH_rotl64(state->v[0], 1) + XXH_rotl64(state->v[1], 7) + XXH_rotl64(state->v[2], 12) + XXH_rotl64(state->v[3], 18);
        h64 = XXH64_mergeRound(h64, state->v[0]);
        h64 = XXH64_mergeRound(h64, state->v[1]);
        h64 = XXH64_mergeRound(h64, state->v[2]);
        h64 = XXH64_mergeRound(h64, state->v[3]);
    } else {
        h64  = state->v[2] /*seed*/ + XXH_PRIME64_5;
    }

    h64 += (xxh_u64) state->total_len;

    return XXH64_finalize(h64, (const xxh_u8*)state->mem64, (size_t)state->total_len, XXH_aligned);
}
#endif /* !XXH_NO_STREAM */

/******* Canonical representation   *******/

/*! @ingroup XXH64_family */
XXH_PUBLIC_API void XXH64_canonicalFromHash(XXH_NOESCAPE XXH64_canonical_t* dst, XXH64_hash_t hash)
{
    XXH_STATIC_ASSERT(sizeof(XXH64_canonical_t) == sizeof(XXH64_hash_t));
    if (XXH_CPU_LITTLE_ENDIAN) hash = XXH_swap64(hash);
    XXH_memcpy(dst, &hash, sizeof(*dst));
}

/*! @ingroup XXH64_family */
XXH_PUBLIC_API XXH64_hash_t XXH64_hashFromCanonical(XXH_NOESCAPE const XXH64_canonical_t* src)
{
    return XXH_readBE64(src);
}

#if defined (__cplusplus)
}
#endif

#ifndef XXH_NO_XXH3

/* *********************************************************************
*  XXH3
*  New generation hash designed for speed on small keys and vectorization
************************************************************************ */
/*!
 * @}
 * @defgroup XXH3_impl XXH3 implementation
 * @ingroup impl
 * @{
 */

/* ===   Compiler specifics   === */

#if ((defined(sun) || defined(__sun)) && __cplusplus) /* Solaris includes __STDC_VERSION__ with C++. Tested with GCC 5.5 */
#  define XXH_RESTRICT   /* disable */
#elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 199901L   /* >= C99 */
#  define XXH_RESTRICT   restrict
#elif (defined (__GNUC__) && ((__GNUC__ > 3) || (__GNUC__ == 3 && __GNUC_MINOR__ >= 1))) \
   || (defined (__clang__)) \
   || (defined (_MSC_VER) && (_MSC_VER >= 1400)) \
   || (defined (__INTEL_COMPILER) && (__INTEL_COMPILER >= 1300))
/*
 * There are a LOT more compilers that recognize __restrict but this
 * covers the major ones.
 */
#  define XXH_RESTRICT   __restrict
#else
#  define XXH_RESTRICT   /* disable */
#endif

#if (defined(__GNUC__) && (__GNUC__ >= 3))  \
  || (defined(__INTEL_COMPILER) && (__INTEL_COMPILER >= 800)) \
  || defined(__clang__)
#    define XXH_likely(x) __builtin_expect(x, 1)
#    define XXH_unlikely(x) __builtin_expect(x, 0)
#else
#    define XXH_likely(x) (x)
#    define XXH_unlikely(x) (x)
#endif

#ifndef XXH_HAS_INCLUDE
#  ifdef __has_include
/*
 * Not defined as XXH_HAS_INCLUDE(x) (function-like) because
 * this causes segfaults in Apple Clang 4.2 (on Mac OS X 10.7 Lion)
 */
#    define XXH_HAS_INCLUDE __has_include
#  else
#    define XXH_HAS_INCLUDE(x) 0
#  endif
#endif

#if defined(__GNUC__) || defined(__clang__)
#  if defined(__ARM_FEATURE_SVE)
#    include <arm_sve.h>
#  endif
#  if defined(__ARM_NEON__) || defined(__ARM_NEON) \
   || (defined(_M_ARM) && _M_ARM >= 7) \
   || defined(_M_ARM64) || defined(_M_ARM64EC) \
   || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* WASM SIMD128 via SIMDe */
#    define inline __inline__  /* circumvent a clang bug */
#    include <arm_neon.h>
#    undef inline
#  elif defined(__AVX2__)
#    include <immintrin.h>
#  elif defined(__SSE2__)
#    include <emmintrin.h>
#  endif
#endif

#if defined(_MSC_VER)
#  include <intrin.h>
#endif

/*
 * One goal of XXH3 is to make it fast on both 32-bit and 64-bit, while
 * remaining a true 64-bit/128-bit hash function.
 *
 * This is done by prioritizing a subset of 64-bit operations that can be
 * emulated without too many steps on the average 32-bit machine.
 *
 * For example, these two lines seem similar, and run equally fast on 64-bit:
 *
 *   xxh_u64 x;
 *   x ^= (x >> 47); // good
 *   x ^= (x >> 13); // bad
 *
 * However, to a 32-bit machine, there is a major difference.
 *
 * x ^= (x >> 47) looks like this:
 *
 *   x.lo ^= (x.hi >> (47 - 32));
 *
 * while x ^= (x >> 13) looks like this:
 *
 *   // note: funnel shifts are not usually cheap.
 *   x.lo ^= (x.lo >> 13) | (x.hi << (32 - 13));
 *   x.hi ^= (x.hi >> 13);
 *
 * The first one is significantly faster than the second, simply because the
 * shift is larger than 32. This means:
 *  - All the bits we need are in the upper 32 bits, so we can ignore the lower
 *    32 bits in the shift.
 *  - The shift result will always fit in the lower 32 bits, and therefore,
 *    we can ignore the upper 32 bits in the xor.
 *
 * Thanks to this optimization, XXH3 only requires these features to be efficient:
 *
 *  - Usable unaligned access
 *  - A 32-bit or 64-bit ALU
 *      - If 32-bit, a decent ADC instruction
 *  - A 32 or 64-bit multiply with a 64-bit result
 *  - For the 128-bit variant, a decent byteswap helps short inputs.
 *
 * The first two are already required by XXH32, and almost all 32-bit and 64-bit
 * platforms which can run XXH32 can run XXH3 efficiently.
 *
 * Thumb-1, the classic 16-bit only subset of ARM's instruction set, is one
 * notable exception.
 *
 * First of all, Thumb-1 lacks support for the UMULL instruction which
 * performs the important long multiply. This means numerous __aeabi_lmul
 * calls.
 *
 * Second of all, the 8 functional registers are just not enough.
 * Setup for __aeabi_lmul, byteshift loads, pointers, and all arithmetic need
 * Lo registers, and this shuffling results in thousands more MOVs than A32.
 *
 * A32 and T32 don't have this limitation. They can access all 14 registers,
 * do a 32->64 multiply with UMULL, and the flexible operand allowing free
 * shifts is helpful, too.
 *
 * Therefore, we do a quick sanity check.
 *
 * If compiling Thumb-1 for a target which supports ARM instructions, we will
 * emit a warning, as it is not a "sane" platform to compile for.
 *
 * Usually, if this happens, it is because of an accident and you probably need
 * to specify -march, as you likely meant to compile for a newer architecture.
 *
 * Credit: large sections of the vectorial and asm source code paths
 *         have been contributed by @easyaspi314
 */
#if defined(__thumb__) && !defined(__thumb2__) && defined(__ARM_ARCH_ISA_ARM)
#   warning "XXH3 is highly inefficient without ARM or Thumb-2."
#endif

/* ==========================================
 * Vectorization detection
 * ========================================== */

#ifdef XXH_DOXYGEN
/*!
 * @ingroup tuning
 * @brief Overrides the vectorization implementation chosen for XXH3.
 *
 * Can be defined to 0 to disable SIMD or any of the values mentioned in
 * @ref XXH_VECTOR_TYPE.
 *
 * If this is not defined, it uses predefined macros to determine the best
 * implementation.
 */
#  define XXH_VECTOR XXH_SCALAR
/*!
 * @ingroup tuning
 * @brief Possible values for @ref XXH_VECTOR.
 *
 * Note that these are actually implemented as macros.
 *
 * If this is not defined, it is detected automatically.
 * internal macro XXH_X86DISPATCH overrides this.
 */
enum XXH_VECTOR_TYPE /* fake enum */ {
    XXH_SCALAR = 0,  /*!< Portable scalar version */
    XXH_SSE2   = 1,  /*!<
                      * SSE2 for Pentium 4, Opteron, all x86_64.
                      *
                      * @note SSE2 is also guaranteed on Windows 10, macOS, and
                      * Android x86.
                      */
    XXH_AVX2   = 2,  /*!< AVX2 for Haswell and Bulldozer */
    XXH_AVX512 = 3,  /*!< AVX512 for Skylake and Icelake */
    XXH_NEON   = 4,  /*!<
                       * NEON for most ARMv7-A, all AArch64, and WASM SIMD128
                       * via the SIMDeverywhere polyfill provided with the
                       * Emscripten SDK.
                       */
    XXH_VSX    = 5,  /*!< VSX and ZVector for POWER8/z13 (64-bit) */
    XXH_SVE    = 6,  /*!< SVE for some ARMv8-A and ARMv9-A */
};
/*!
 * @ingroup tuning
 * @brief Selects the minimum alignment for XXH3's accumulators.
 *
 * When using SIMD, this should match the alignment required for said vector
 * type, so, for example, 32 for AVX2.
 *
 * Default: Auto detected.
 */
#  define XXH_ACC_ALIGN 8
#endif

/* Actual definition */
#ifndef XXH_DOXYGEN
#  define XXH_SCALAR 0
#  define XXH_SSE2   1
#  define XXH_AVX2   2
#  define XXH_AVX512 3
#  define XXH_NEON   4
#  define XXH_VSX    5
#  define XXH_SVE    6
#endif

#ifndef XXH_VECTOR    /* can be defined on command line */
#  if defined(__ARM_FEATURE_SVE)
#    define XXH_VECTOR XXH_SVE
#  elif ( \
        defined(__ARM_NEON__) || defined(__ARM_NEON) /* gcc */ \
     || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC) /* msvc */ \
     || (defined(__wasm_simd128__) && XXH_HAS_INCLUDE(<arm_neon.h>)) /* wasm simd128 via SIMDe */ \
   ) && ( \
        defined(_WIN32) || defined(__LITTLE_ENDIAN__) /* little endian only */ \
    || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
   )
#    define XXH_VECTOR XXH_NEON
#  elif defined(__AVX512F__)
#    define XXH_VECTOR XXH_AVX512
#  elif defined(__AVX2__)
#    define XXH_VECTOR XXH_AVX2
#  elif defined(__SSE2__) || defined(_M_X64) || (defined(_M_IX86_FP) && (_M_IX86_FP == 2))
#    define XXH_VECTOR XXH_SSE2
#  elif (defined(__PPC64__) && defined(__POWER8_VECTOR__)) \
     || (defined(__s390x__) && defined(__VEC__)) \
     && defined(__GNUC__) /* TODO: IBM XL */
#    define XXH_VECTOR XXH_VSX
#  else
#    define XXH_VECTOR XXH_SCALAR
#  endif
#endif

/* __ARM_FEATURE_SVE is only supported by GCC & Clang. */
#if (XXH_VECTOR == XXH_SVE) && !defined(__ARM_FEATURE_SVE)
#  ifdef _MSC_VER
#    pragma warning(once : 4606)
#  else
#    warning "__ARM_FEATURE_SVE isn't supported. Use SCALAR instead."
#  endif
#  undef XXH_VECTOR
#  define XXH_VECTOR XXH_SCALAR
#endif

/*
 * Controls the alignment of the accumulator,
 * for compatibility with aligned vector loads, which are usually faster.
 */
#ifndef XXH_ACC_ALIGN
#  if defined(XXH_X86DISPATCH)
#     define XXH_ACC_ALIGN 64  /* for compatibility with avx512 */
#  elif XXH_VECTOR == XXH_SCALAR  /* scalar */
#     define XXH_ACC_ALIGN 8
#  elif XXH_VECTOR == XXH_SSE2  /* sse2 */
#     define XXH_ACC_ALIGN 16
#  elif XXH_VECTOR == XXH_AVX2  /* avx2 */
#     define XXH_ACC_ALIGN 32
#  elif XXH_VECTOR == XXH_NEON  /* neon */
#     define XXH_ACC_ALIGN 16
#  elif XXH_VECTOR == XXH_VSX   /* vsx */
#     define XXH_ACC_ALIGN 16
#  elif XXH_VECTOR == XXH_AVX512  /* avx512 */
#     define XXH_ACC_ALIGN 64
#  elif XXH_VECTOR == XXH_SVE   /* sve */
#     define XXH_ACC_ALIGN 64
#  endif
#endif

#if defined(XXH_X86DISPATCH) || XXH_VECTOR == XXH_SSE2 \
    || XXH_VECTOR == XXH_AVX2 || XXH_VECTOR == XXH_AVX512
#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
#elif XXH_VECTOR == XXH_SVE
#  define XXH_SEC_ALIGN XXH_ACC_ALIGN
#else
#  define XXH_SEC_ALIGN 8
#endif

#if defined(__GNUC__) || defined(__clang__)
#  define XXH_ALIASING __attribute__((may_alias))
#else
#  define XXH_ALIASING /* nothing */
#endif

/*
 * UGLY HACK:
 * GCC usually generates the best code with -O3 for xxHash.
 *
 * However, when targeting AVX2, it is overzealous in its unrolling resulting
 * in code roughly 3/4 the speed of Clang.
 *
 * There are other issues, such as GCC splitting _mm256_loadu_si256 into
 * _mm_loadu_si128 + _mm256_inserti128_si256. This is an optimization which
 * only applies to Sandy and Ivy Bridge... which don't even support AVX2.
 *
 * That is why when compiling the AVX2 version, it is recommended to use either
 *   -O2 -mavx2 -march=haswell
 * or
 *   -O2 -mavx2 -mno-avx256-split-unaligned-load
 * for decent performance, or to use Clang instead.
 *
 * Fortunately, we can control the first one with a pragma that forces GCC into
 * -O2, but the other one we can't control without "failed to inline always
 * inline function due to target mismatch" warnings.
 */
#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
#  pragma GCC push_options
#  pragma GCC optimize("-O2")
#endif

#if defined (__cplusplus)
extern "C" {
#endif

#if XXH_VECTOR == XXH_NEON

/*
 * UGLY HACK: While AArch64 GCC on Linux does not seem to care, on macOS, GCC -O3
 * optimizes out the entire hashLong loop because of the aliasing violation.
 *
 * However, GCC is also inefficient at load-store optimization with vld1q/vst1q,
 * so the only option is to mark it as aliasing.
 */
typedef uint64x2_t xxh_aliasing_uint64x2_t XXH_ALIASING;

/*!
 * @internal
 * @brief `vld1q_u64` but faster and alignment-safe.
 *
 * On AArch64, unaligned access is always safe, but on ARMv7-a, it is only
 * *conditionally* safe (`vld1` has an alignment bit like `movdq[ua]` in x86).
 *
 * GCC for AArch64 sees `vld1q_u8` as an intrinsic instead of a load, so it
 * prohibits load-store optimizations. Therefore, a direct dereference is used.
 *
 * Otherwise, `vld1q_u8` is used with `vreinterpretq_u8_u64` to do a safe
 * unaligned load.
 */
#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__)
XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr) /* silence -Wcast-align */
{
    return *(xxh_aliasing_uint64x2_t const *)ptr;
}
#else
XXH_FORCE_INLINE uint64x2_t XXH_vld1q_u64(void const* ptr)
{
    return vreinterpretq_u64_u8(vld1q_u8((uint8_t const*)ptr));
}
#endif

/*!
 * @internal
 * @brief `vmlal_u32` on low and high halves of a vector.
 *
 * This is a workaround for AArch64 GCC < 11 which implemented arm_neon.h with
 * inline assembly and were therefore incapable of merging the `vget_{low, high}_u32`
 * with `vmlal_u32`.
 */
#if defined(__aarch64__) && defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 11
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{
    /* Inline assembly is the only way */
    __asm__("umlal   %0.2d, %1.2s, %2.2s" : "+w" (acc) : "w" (lhs), "w" (rhs));
    return acc;
}
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{
    /* This intrinsic works as expected */
    return vmlal_high_u32(acc, lhs, rhs);
}
#else
/* Portable intrinsic versions */
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_low_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{
    return vmlal_u32(acc, vget_low_u32(lhs), vget_low_u32(rhs));
}
/*! @copydoc XXH_vmlal_low_u32
 * Assume the compiler converts this to vmlal_high_u32 on aarch64 */
XXH_FORCE_INLINE uint64x2_t
XXH_vmlal_high_u32(uint64x2_t acc, uint32x4_t lhs, uint32x4_t rhs)
{
    return vmlal_u32(acc, vget_high_u32(lhs), vget_high_u32(rhs));
}
#endif

/*!
 * @ingroup tuning
 * @brief Controls the NEON to scalar ratio for XXH3
 *
 * This can be set to 2, 4, 6, or 8.
 *
 * ARM Cortex CPUs are _very_ sensitive to how their pipelines are used.
 *
 * For example, the Cortex-A73 can dispatch 3 micro-ops per cycle, but only 2 of those
 * can be NEON. If you are only using NEON instructions, you are only using 2/3 of the CPU
 * bandwidth.
 *
 * This is even more noticeable on the more advanced cores like the Cortex-A76 which
 * can dispatch 8 micro-ops per cycle, but still only 2 NEON micro-ops at once.
 *
 * Therefore, to make the most out of the pipeline, it is beneficial to run 6 NEON lanes
 * and 2 scalar lanes, which is chosen by default.
 *
 * This does not apply to Apple processors or 32-bit processors, which run better with
 * full NEON. These will default to 8. Additionally, size-optimized builds run 8 lanes.
 *
 * This change benefits CPUs with large micro-op buffers without negatively affecting
 * most other CPUs:
 *
 *  | Chipset               | Dispatch type       | NEON only | 6:2 hybrid | Diff. |
 *  |:----------------------|:--------------------|----------:|-----------:|------:|
 *  | Snapdragon 730 (A76)  | 2 NEON/8 micro-ops  |  8.8 GB/s |  10.1 GB/s |  ~16% |
 *  | Snapdragon 835 (A73)  | 2 NEON/3 micro-ops  |  5.1 GB/s |   5.3 GB/s |   ~5% |
 *  | Marvell PXA1928 (A53) | In-order dual-issue |  1.9 GB/s |   1.9 GB/s |    0% |
 *  | Apple M1              | 4 NEON/8 micro-ops  | 37.3 GB/s |  36.1 GB/s |  ~-3% |
 *
 * It also seems to fix some bad codegen on GCC, making it almost as fast as clang.
 *
 * When using WASM SIMD128, if this is 2 or 6, SIMDe will scalarize 2 of the lanes meaning
 * it effectively becomes worse 4.
 *
 * @see XXH3_accumulate_512_neon()
 */
# ifndef XXH3_NEON_LANES
#  if (defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || defined(_M_ARM64EC)) \
   && !defined(__APPLE__) && XXH_SIZE_OPT <= 0
#   define XXH3_NEON_LANES 6
#  else
#   define XXH3_NEON_LANES XXH_ACC_NB
#  endif
# endif
#endif  /* XXH_VECTOR == XXH_NEON */

#if defined (__cplusplus)
} /* extern "C" */
#endif

/*
 * VSX and Z Vector helpers.
 *
 * This is very messy, and any pull requests to clean this up are welcome.
 *
 * There are a lot of problems with supporting VSX and s390x, due to
 * inconsistent intrinsics, spotty coverage, and multiple endiannesses.
 */
#if XXH_VECTOR == XXH_VSX
/* Annoyingly, these headers _may_ define three macros: `bool`, `vector`,
 * and `pixel`. This is a problem for obvious reasons.
 *
 * These keywords are unnecessary; the spec literally says they are
 * equivalent to `__bool`, `__vector`, and `__pixel` and may be undef'd
 * after including the header.
 *
 * We use pragma push_macro/pop_macro to keep the namespace clean. */
#  pragma push_macro("bool")
#  pragma push_macro("vector")
#  pragma push_macro("pixel")
/* silence potential macro redefined warnings */
#  undef bool
#  undef vector
#  undef pixel

#  if defined(__s390x__)
#    include <s390intrin.h>
#  else
#    include <altivec.h>
#  endif

/* Restore the original macro values, if applicable. */
#  pragma pop_macro("pixel")
#  pragma pop_macro("vector")
#  pragma pop_macro("bool")

typedef __vector unsigned long long xxh_u64x2;
typedef __vector unsigned char xxh_u8x16;
typedef __vector unsigned xxh_u32x4;

/*
 * UGLY HACK: Similar to aarch64 macOS GCC, s390x GCC has the same aliasing issue.
 */
typedef xxh_u64x2 xxh_aliasing_u64x2 XXH_ALIASING;

# ifndef XXH_VSX_BE
#  if defined(__BIG_ENDIAN__) \
  || (defined(__BYTE_ORDER__) && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
#    define XXH_VSX_BE 1
#  elif defined(__VEC_ELEMENT_REG_ORDER__) && __VEC_ELEMENT_REG_ORDER__ == __ORDER_BIG_ENDIAN__
#    warning "-maltivec=be is not recommended. Please use native endianness."
#    define XXH_VSX_BE 1
#  else
#    define XXH_VSX_BE 0
#  endif
# endif /* !defined(XXH_VSX_BE) */

# if XXH_VSX_BE
#  if defined(__POWER9_VECTOR__) || (defined(__clang__) && defined(__s390x__))
#    define XXH_vec_revb vec_revb
#  else
#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * A polyfill for POWER9's vec_revb().
 */
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_revb(xxh_u64x2 val)
{
    xxh_u8x16 const vByteSwap = { 0x07, 0x06, 0x05, 0x04, 0x03, 0x02, 0x01, 0x00,
                                  0x0F, 0x0E, 0x0D, 0x0C, 0x0B, 0x0A, 0x09, 0x08 };
    return vec_perm(val, val, vByteSwap);
}
#if defined (__cplusplus)
} /* extern "C" */
#endif
#  endif
# endif /* XXH_VSX_BE */

#if defined (__cplusplus)
extern "C" {
#endif
/*!
 * Performs an unaligned vector load and byte swaps it on big endian.
 */
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_loadu(const void *ptr)
{
    xxh_u64x2 ret;
    XXH_memcpy(&ret, ptr, sizeof(xxh_u64x2));
# if XXH_VSX_BE
    ret = XXH_vec_revb(ret);
# endif
    return ret;
}

/*
 * vec_mulo and vec_mule are very problematic intrinsics on PowerPC
 *
 * These intrinsics weren't added until GCC 8, despite existing for a while,
 * and they are endian dependent. Also, their meaning swap depending on version.
 * */
# if defined(__s390x__)
 /* s390x is always big endian, no issue on this platform */
#  define XXH_vec_mulo vec_mulo
#  define XXH_vec_mule vec_mule
# elif defined(__clang__) && XXH_HAS_BUILTIN(__builtin_altivec_vmuleuw) && !defined(__ibmxl__)
/* Clang has a better way to control this, we can just use the builtin which doesn't swap. */
 /* The IBM XL Compiler (which defined __clang__) only implements the vec_* operations */
#  define XXH_vec_mulo __builtin_altivec_vmulouw
#  define XXH_vec_mule __builtin_altivec_vmuleuw
# else
/* gcc needs inline assembly */
/* Adapted from https://github.com/google/highwayhash/blob/master/highwayhash/hh_vsx.h. */
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mulo(xxh_u32x4 a, xxh_u32x4 b)
{
    xxh_u64x2 result;
    __asm__("vmulouw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
    return result;
}
XXH_FORCE_INLINE xxh_u64x2 XXH_vec_mule(xxh_u32x4 a, xxh_u32x4 b)
{
    xxh_u64x2 result;
    __asm__("vmuleuw %0, %1, %2" : "=v" (result) : "v" (a), "v" (b));
    return result;
}
# endif /* XXH_vec_mulo, XXH_vec_mule */

#if defined (__cplusplus)
} /* extern "C" */
#endif

#endif /* XXH_VECTOR == XXH_VSX */

#if XXH_VECTOR == XXH_SVE
#define ACCRND(acc, offset) \
do { \
    svuint64_t input_vec = svld1_u64(mask, xinput + offset);         \
    svuint64_t secret_vec = svld1_u64(mask, xsecret + offset);       \
    svuint64_t mixed = sveor_u64_x(mask, secret_vec, input_vec);     \
    svuint64_t swapped = svtbl_u64(input_vec, kSwap);                \
    svuint64_t mixed_lo = svextw_u64_x(mask, mixed);                 \
    svuint64_t mixed_hi = svlsr_n_u64_x(mask, mixed, 32);            \
    svuint64_t mul = svmad_u64_x(mask, mixed_lo, mixed_hi, swapped); \
    acc = svadd_u64_x(mask, acc, mul);                               \
} while (0)
#endif /* XXH_VECTOR == XXH_SVE */

/* prefetch
 * can be disabled, by declaring XXH_NO_PREFETCH build macro */
#if defined(XXH_NO_PREFETCH)
#  define XXH_PREFETCH(ptr)  (void)(ptr)  /* disabled */
#else
#  if XXH_SIZE_OPT >= 1
#    define XXH_PREFETCH(ptr) (void)(ptr)
#  elif defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86))  /* _mm_prefetch() not defined outside of x86/x64 */
#    include <mmintrin.h>   /* https://msdn.microsoft.com/fr-fr/library/84szxsww(v=vs.90).aspx */
#    define XXH_PREFETCH(ptr)  _mm_prefetch((const char*)(ptr), _MM_HINT_T0)
#  elif defined(__GNUC__) && ( (__GNUC__ >= 4) || ( (__GNUC__ == 3) && (__GNUC_MINOR__ >= 1) ) )
#    define XXH_PREFETCH(ptr)  __builtin_prefetch((ptr), 0 /* rw==read */, 3 /* locality */)
#  else
#    define XXH_PREFETCH(ptr) (void)(ptr)  /* disabled */
#  endif
#endif  /* XXH_NO_PREFETCH */

#if defined (__cplusplus)
extern "C" {
#endif
/* ==========================================
 * XXH3 default settings
 * ========================================== */

#define XXH_SECRET_DEFAULT_SIZE 192   /* minimum XXH3_SECRET_SIZE_MIN */

#if (XXH_SECRET_DEFAULT_SIZE < XXH3_SECRET_SIZE_MIN)
#  error "default keyset is not large enough"
#endif

/*! Pseudorandom secret taken directly from FARSH. */
XXH_ALIGN(64) static const xxh_u8 XXH3_kSecret[XXH_SECRET_DEFAULT_SIZE] = {
    0xb8, 0xfe, 0x6c, 0x39, 0x23, 0xa4, 0x4b, 0xbe, 0x7c, 0x01, 0x81, 0x2c, 0xf7, 0x21, 0xad, 0x1c,
    0xde, 0xd4, 0x6d, 0xe9, 0x83, 0x90, 0x97, 0xdb, 0x72, 0x40, 0xa4, 0xa4, 0xb7, 0xb3, 0x67, 0x1f,
    0xcb, 0x79, 0xe6, 0x4e, 0xcc, 0xc0, 0xe5, 0x78, 0x82, 0x5a, 0xd0, 0x7d, 0xcc, 0xff, 0x72, 0x21,
    0xb8, 0x08, 0x46, 0x74, 0xf7, 0x43, 0x24, 0x8e, 0xe0, 0x35, 0x90, 0xe6, 0x81, 0x3a, 0x26, 0x4c,
    0x3c, 0x28, 0x52, 0xbb, 0x91, 0xc3, 0x00, 0xcb, 0x88, 0xd0, 0x65, 0x8b, 0x1b, 0x53, 0x2e, 0xa3,
    0x71, 0x64, 0x48, 0x97, 0xa2, 0x0d, 0xf9, 0x4e, 0x38, 0x19, 0xef, 0x46, 0xa9, 0xde, 0xac, 0xd8,
    0xa8, 0xfa, 0x76, 0x3f, 0xe3, 0x9c, 0x34, 0x3f, 0xf9, 0xdc, 0xbb, 0xc7, 0xc7, 0x0b, 0x4f, 0x1d,
    0x8a, 0x51, 0xe0, 0x4b, 0xcd, 0xb4, 0x59, 0x31, 0xc8, 0x9f, 0x7e, 0xc9, 0xd9, 0x78, 0x73, 0x64,
    0xea, 0xc5, 0xac, 0x83, 0x34, 0xd3, 0xeb, 0xc3, 0xc5, 0x81, 0xa0, 0xff, 0xfa, 0x13, 0x63, 0xeb,
    0x17, 0x0d, 0xdd, 0x51, 0xb7, 0xf0, 0xda, 0x49, 0xd3, 0x16, 0x55, 0x26, 0x29, 0xd4, 0x68, 0x9e,
    0x2b, 0x16, 0xbe, 0x58, 0x7d, 0x47, 0xa1, 0xfc, 0x8f, 0xf8, 0xb8, 0xd1, 0x7a, 0xd0, 0x31, 0xce,
    0x45, 0xcb, 0x3a, 0x8f, 0x95, 0x16, 0x04, 0x28, 0xaf, 0xd7, 0xfb, 0xca, 0xbb, 0x4b, 0x40, 0x7e,
};

static const xxh_u64 PRIME_MX1 = 0x165667919E3779F9ULL;  /*!< 0b0001011001010110011001111001000110011110001101110111100111111001 */
static const xxh_u64 PRIME_MX2 = 0x9FB21C651E98DF25ULL;  /*!< 0b1001111110110010000111000110010100011110100110001101111100100101 */

#ifdef XXH_OLD_NAMES
#  define kSecret XXH3_kSecret
#endif

#ifdef XXH_DOXYGEN
/*!
 * @brief Calculates a 32-bit to 64-bit long multiply.
 *
 * Implemented as a macro.
 *
 * Wraps `__emulu` on MSVC x86 because it tends to call `__allmul` when it doesn't
 * need to (but it shouldn't need to anyways, it is about 7 instructions to do
 * a 64x64 multiply...). Since we know that this will _always_ emit `MULL`, we
 * use that instead of the normal method.
 *
 * If you are compiling for platforms like Thumb-1 and don't have a better option,
 * you may also want to write your own long multiply routine here.
 *
 * @param x, y Numbers to be multiplied
 * @return 64-bit product of the low 32 bits of @p x and @p y.
 */
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64(xxh_u64 x, xxh_u64 y)
{
   return (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF);
}
#elif defined(_MSC_VER) && defined(_M_IX86)
#    define XXH_mult32to64(x, y) __emulu((unsigned)(x), (unsigned)(y))
#else
/*
 * Downcast + upcast is usually better than masking on older compilers like
 * GCC 4.2 (especially 32-bit ones), all without affecting newer compilers.
 *
 * The other method, (x & 0xFFFFFFFF) * (y & 0xFFFFFFFF), will AND both operands
 * and perform a full 64x64 multiply -- entirely redundant on 32-bit.
 */
#    define XXH_mult32to64(x, y) ((xxh_u64)(xxh_u32)(x) * (xxh_u64)(xxh_u32)(y))
#endif

/*!
 * @brief Calculates a 64->128-bit long multiply.
 *
 * Uses `__uint128_t` and `_umul128` if available, otherwise uses a scalar
 * version.
 *
 * @param lhs , rhs The 64-bit integers to be multiplied
 * @return The 128-bit result represented in an @ref XXH128_hash_t.
 */
static XXH128_hash_t
XXH_mult64to128(xxh_u64 lhs, xxh_u64 rhs)
{
    /*
     * GCC/Clang __uint128_t method.
     *
     * On most 64-bit targets, GCC and Clang define a __uint128_t type.
     * This is usually the best way as it usually uses a native long 64-bit
     * multiply, such as MULQ on x86_64 or MUL + UMULH on aarch64.
     *
     * Usually.
     *
     * Despite being a 32-bit platform, Clang (and emscripten) define this type
     * despite not having the arithmetic for it. This results in a laggy
     * compiler builtin call which calculates a full 128-bit multiply.
     * In that case it is best to use the portable one.
     * https://github.com/Cyan4973/xxHash/issues/211#issuecomment-515575677
     */
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__wasm__) \
    && defined(__SIZEOF_INT128__) \
    || (defined(_INTEGRAL_MAX_BITS) && _INTEGRAL_MAX_BITS >= 128)

    __uint128_t const product = (__uint128_t)lhs * (__uint128_t)rhs;
    XXH128_hash_t r128;
    r128.low64  = (xxh_u64)(product);
    r128.high64 = (xxh_u64)(product >> 64);
    return r128;

    /*
     * MSVC for x64's _umul128 method.
     *
     * xxh_u64 _umul128(xxh_u64 Multiplier, xxh_u64 Multiplicand, xxh_u64 *HighProduct);
     *
     * This compiles to single operand MUL on x64.
     */
#elif (defined(_M_X64) || defined(_M_IA64)) && !defined(_M_ARM64EC)

#ifndef _MSC_VER
#   pragma intrinsic(_umul128)
#endif
    xxh_u64 product_high;
    xxh_u64 const product_low = _umul128(lhs, rhs, &product_high);
    XXH128_hash_t r128;
    r128.low64  = product_low;
    r128.high64 = product_high;
    return r128;

    /*
     * MSVC for ARM64's __umulh method.
     *
     * This compiles to the same MUL + UMULH as GCC/Clang's __uint128_t method.
     */
#elif defined(_M_ARM64) || defined(_M_ARM64EC)

#ifndef _MSC_VER
#   pragma intrinsic(__umulh)
#endif
    XXH128_hash_t r128;
    r128.low64  = lhs * rhs;
    r128.high64 = __umulh(lhs, rhs);
    return r128;

#else
    /*
     * Portable scalar method. Optimized for 32-bit and 64-bit ALUs.
     *
     * This is a fast and simple grade school multiply, which is shown below
     * with base 10 arithmetic instead of base 0x100000000.
     *
     *           9 3 // D2 lhs = 93
     *         x 7 5 // D2 rhs = 75
     *     ----------
     *           1 5 // D2 lo_lo = (93 % 10) * (75 % 10) = 15
     *         4 5 | // D2 hi_lo = (93 / 10) * (75 % 10) = 45
     *         2 1 | // D2 lo_hi = (93 % 10) * (75 / 10) = 21
     *     + 6 3 | | // D2 hi_hi = (93 / 10) * (75 / 10) = 63
     *     ---------
     *         2 7 | // D2 cross = (15 / 10) + (45 % 10) + 21 = 27
     *     + 6 7 | | // D2 upper = (27 / 10) + (45 / 10) + 63 = 67
     *     ---------
     *       6 9 7 5 // D4 res = (27 * 10) + (15 % 10) + (67 * 100) = 6975
     *
     * The reasons for adding the products like this are:
     *  1. It avoids manual carry tracking. Just like how
     *     (9 * 9) + 9 + 9 = 99, the same applies with this for UINT64_MAX.
     *     This avoids a lot of complexity.
     *
     *  2. It hints for, and on Clang, compiles to, the powerful UMAAL
     *     instruction available in ARM's Digital Signal Processing extension
     *     in 32-bit ARMv6 and later, which is shown below:
     *
     *         void UMAAL(xxh_u32 *RdLo, xxh_u32 *RdHi, xxh_u32 Rn, xxh_u32 Rm)
     *         {
     *             xxh_u64 product = (xxh_u64)*RdLo * (xxh_u64)*RdHi + Rn + Rm;
     *             *RdLo = (xxh_u32)(product & 0xFFFFFFFF);
     *             *RdHi = (xxh_u32)(product >> 32);
     *         }
     *
     *     This instruction was designed for efficient long multiplication, and
     *     allows this to be calculated in only 4 instructions at speeds
     *     comparable to some 64-bit ALUs.
     *
     *  3. It isn't terrible on other platforms. Usually this will be a couple
     *     of 32-bit ADD/ADCs.
     */

    /* First calculate all of the cross products. */
    xxh_u64 const lo_lo = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs & 0xFFFFFFFF);
    xxh_u64 const hi_lo = XXH_mult32to64(lhs >> 32,        rhs & 0xFFFFFFFF);
    xxh_u64 const lo_hi = XXH_mult32to64(lhs & 0xFFFFFFFF, rhs >> 32);
    xxh_u64 const hi_hi = XXH_mult32to64(lhs >> 32,        rhs >> 32);

    /* Now add the products together. These will never overflow. */
    xxh_u64 const cross = (lo_lo >> 32) + (hi_lo & 0xFFFFFFFF) + lo_hi;
    xxh_u64 const upper = (hi_lo >> 32) + (cross >> 32)        + hi_hi;
    xxh_u64 const lower = (cross << 32) | (lo_lo & 0xFFFFFFFF);

    XXH128_hash_t r128;
    r128.low64  = lower;
    r128.high64 = upper;
    return r128;
#endif
}

/*!
 * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it.
 *
 * The reason for the separate function is to prevent passing too many structs
 * around by value. This will hopefully inline the multiply, but we don't force it.
 *
 * @param lhs , rhs The 64-bit integers to multiply
 * @return The low 64 bits of the product XOR'd by the high 64 bits.
 * @see XXH_mult64to128()
 */
static xxh_u64
XXH3_mul128_fold64(xxh_u64 lhs, xxh_u64 rhs)
{
    XXH128_hash_t product = XXH_mult64to128(lhs, rhs);
    return product.low64 ^ product.high64;
}

/*! Seems to produce slightly better code on GCC for some reason. */
XXH_FORCE_INLINE XXH_CONSTF xxh_u64 XXH_xorshift64(xxh_u64 v64, int shift)
{
    XXH_ASSERT(0 <= shift && shift < 64);
    return v64 ^ (v64 >> shift);
}

/*
 * This is a fast avalanche stage,
 * suitable when input bits are already partially mixed
 */
static XXH64_hash_t XXH3_avalanche(xxh_u64 h64)
{
    h64 = XXH_xorshift64(h64, 37);
    h64 *= PRIME_MX1;
    h64 = XXH_xorshift64(h64, 32);
    return h64;
}

/*
 * This is a stronger avalanche,
 * inspired by Pelle Evensen's rrmxmx
 * preferable when input has not been previously mixed
 */
static XXH64_hash_t XXH3_rrmxmx(xxh_u64 h64, xxh_u64 len)
{
    /* this mix is inspired by Pelle Evensen's rrmxmx */
    h64 ^= XXH_rotl64(h64, 49) ^ XXH_rotl64(h64, 24);
    h64 *= PRIME_MX2;
    h64 ^= (h64 >> 35) + len ;
    h64 *= PRIME_MX2;
    return XXH_xorshift64(h64, 28);
}


/* ==========================================
 * Short keys
 * ==========================================
 * One of the shortcomings of XXH32 and XXH64 was that their performance was
 * sub-optimal on short lengths. It used an iterative algorithm which strongly
 * favored lengths that were a multiple of 4 or 8.
 *
 * Instead of iterating over individual inputs, we use a set of single shot
 * functions which piece together a range of lengths and operate in constant time.
 *
 * Additionally, the number of multiplies has been significantly reduced. This
 * reduces latency, especially when emulating 64-bit multiplies on 32-bit.
 *
 * Depending on the platform, this may or may not be faster than XXH32, but it
 * is almost guaranteed to be faster than XXH64.
 */

/*
 * At very short lengths, there isn't enough input to fully hide secrets, or use
 * the entire secret.
 *
 * There is also only a limited amount of mixing we can do before significantly
 * impacting performance.
 *
 * Therefore, we use different sections of the secret and always mix two secret
 * samples with an XOR. This should have no effect on performance on the
 * seedless or withSeed variants because everything _should_ be constant folded
 * by modern compilers.
 *
 * The XOR mixing hides individual parts of the secret and increases entropy.
 *
 * This adds an extra layer of strength for custom secrets.
 */
XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_1to3_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(1 <= len && len <= 3);
    XXH_ASSERT(secret != NULL);
    /*
     * len = 1: combined = { input[0], 0x01, input[0], input[0] }
     * len = 2: combined = { input[1], 0x02, input[0], input[1] }
     * len = 3: combined = { input[2], 0x03, input[0], input[1] }
     */
    {   xxh_u8  const c1 = input[0];
        xxh_u8  const c2 = input[len >> 1];
        xxh_u8  const c3 = input[len - 1];
        xxh_u32 const combined = ((xxh_u32)c1 << 16) | ((xxh_u32)c2  << 24)
                               | ((xxh_u32)c3 <<  0) | ((xxh_u32)len << 8);
        xxh_u64 const bitflip = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
        xxh_u64 const keyed = (xxh_u64)combined ^ bitflip;
        return XXH64_avalanche(keyed);
    }
}

XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_4to8_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(secret != NULL);
    XXH_ASSERT(4 <= len && len <= 8);
    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
    {   xxh_u32 const input1 = XXH_readLE32(input);
        xxh_u32 const input2 = XXH_readLE32(input + len - 4);
        xxh_u64 const bitflip = (XXH_readLE64(secret+8) ^ XXH_readLE64(secret+16)) - seed;
        xxh_u64 const input64 = input2 + (((xxh_u64)input1) << 32);
        xxh_u64 const keyed = input64 ^ bitflip;
        return XXH3_rrmxmx(keyed, len);
    }
}

XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_9to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(secret != NULL);
    XXH_ASSERT(9 <= len && len <= 16);
    {   xxh_u64 const bitflip1 = (XXH_readLE64(secret+24) ^ XXH_readLE64(secret+32)) + seed;
        xxh_u64 const bitflip2 = (XXH_readLE64(secret+40) ^ XXH_readLE64(secret+48)) - seed;
        xxh_u64 const input_lo = XXH_readLE64(input)           ^ bitflip1;
        xxh_u64 const input_hi = XXH_readLE64(input + len - 8) ^ bitflip2;
        xxh_u64 const acc = len
                          + XXH_swap64(input_lo) + input_hi
                          + XXH3_mul128_fold64(input_lo, input_hi);
        return XXH3_avalanche(acc);
    }
}

XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_0to16_64b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(len <= 16);
    {   if (XXH_likely(len >  8)) return XXH3_len_9to16_64b(input, len, secret, seed);
        if (XXH_likely(len >= 4)) return XXH3_len_4to8_64b(input, len, secret, seed);
        if (len) return XXH3_len_1to3_64b(input, len, secret, seed);
        return XXH64_avalanche(seed ^ (XXH_readLE64(secret+56) ^ XXH_readLE64(secret+64)));
    }
}

/*
 * DISCLAIMER: There are known *seed-dependent* multicollisions here due to
 * multiplication by zero, affecting hashes of lengths 17 to 240.
 *
 * However, they are very unlikely.
 *
 * Keep this in mind when using the unseeded XXH3_64bits() variant: As with all
 * unseeded non-cryptographic hashes, it does not attempt to defend itself
 * against specially crafted inputs, only random inputs.
 *
 * Compared to classic UMAC where a 1 in 2^31 chance of 4 consecutive bytes
 * cancelling out the secret is taken an arbitrary number of times (addressed
 * in XXH3_accumulate_512), this collision is very unlikely with random inputs
 * and/or proper seeding:
 *
 * This only has a 1 in 2^63 chance of 8 consecutive bytes cancelling out, in a
 * function that is only called up to 16 times per hash with up to 240 bytes of
 * input.
 *
 * This is not too bad for a non-cryptographic hash function, especially with
 * only 64 bit outputs.
 *
 * The 128-bit variant (which trades some speed for strength) is NOT affected
 * by this, although it is always a good idea to use a proper seed if you care
 * about strength.
 */
XXH_FORCE_INLINE xxh_u64 XXH3_mix16B(const xxh_u8* XXH_RESTRICT input,
                                     const xxh_u8* XXH_RESTRICT secret, xxh_u64 seed64)
{
#if defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
  && defined(__i386__) && defined(__SSE2__)  /* x86 + SSE2 */ \
  && !defined(XXH_ENABLE_AUTOVECTORIZE)      /* Define to disable like XXH32 hack */
    /*
     * UGLY HACK:
     * GCC for x86 tends to autovectorize the 128-bit multiply, resulting in
     * slower code.
     *
     * By forcing seed64 into a register, we disrupt the cost model and
     * cause it to scalarize. See `XXH32_round()`
     *
     * FIXME: Clang's output is still _much_ faster -- On an AMD Ryzen 3600,
     * XXH3_64bits @ len=240 runs at 4.6 GB/s with Clang 9, but 3.3 GB/s on
     * GCC 9.2, despite both emitting scalar code.
     *
     * GCC generates much better scalar code than Clang for the rest of XXH3,
     * which is why finding a more optimal codepath is an interest.
     */
    XXH_COMPILER_GUARD(seed64);
#endif
    {   xxh_u64 const input_lo = XXH_readLE64(input);
        xxh_u64 const input_hi = XXH_readLE64(input+8);
        return XXH3_mul128_fold64(
            input_lo ^ (XXH_readLE64(secret)   + seed64),
            input_hi ^ (XXH_readLE64(secret+8) - seed64)
        );
    }
}

/* For mid range keys, XXH3 uses a Mum-hash variant. */
XXH_FORCE_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_17to128_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                     const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                     XXH64_hash_t seed)
{
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
    XXH_ASSERT(16 < len && len <= 128);

    {   xxh_u64 acc = len * XXH_PRIME64_1;
#if XXH_SIZE_OPT >= 1
        /* Smaller and cleaner, but slightly slower. */
        unsigned int i = (unsigned int)(len - 1) / 32;
        do {
            acc += XXH3_mix16B(input+16 * i, secret+32*i, seed);
            acc += XXH3_mix16B(input+len-16*(i+1), secret+32*i+16, seed);
        } while (i-- != 0);
#else
        if (len > 32) {
            if (len > 64) {
                if (len > 96) {
                    acc += XXH3_mix16B(input+48, secret+96, seed);
                    acc += XXH3_mix16B(input+len-64, secret+112, seed);
                }
                acc += XXH3_mix16B(input+32, secret+64, seed);
                acc += XXH3_mix16B(input+len-48, secret+80, seed);
            }
            acc += XXH3_mix16B(input+16, secret+32, seed);
            acc += XXH3_mix16B(input+len-32, secret+48, seed);
        }
        acc += XXH3_mix16B(input+0, secret+0, seed);
        acc += XXH3_mix16B(input+len-16, secret+16, seed);
#endif
        return XXH3_avalanche(acc);
    }
}

/*!
 * @brief Maximum size of "short" key in bytes.
 */
#define XXH3_MIDSIZE_MAX 240

XXH_NO_INLINE XXH_PUREF XXH64_hash_t
XXH3_len_129to240_64b(const xxh_u8* XXH_RESTRICT input, size_t len,
                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                      XXH64_hash_t seed)
{
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);

    #define XXH3_MIDSIZE_STARTOFFSET 3
    #define XXH3_MIDSIZE_LASTOFFSET  17

    {   xxh_u64 acc = len * XXH_PRIME64_1;
        xxh_u64 acc_end;
        unsigned int const nbRounds = (unsigned int)len / 16;
        unsigned int i;
        XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);
        for (i=0; i<8; i++) {
            acc += XXH3_mix16B(input+(16*i), secret+(16*i), seed);
        }
        /* last bytes */
        acc_end = XXH3_mix16B(input + len - 16, secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET, seed);
        XXH_ASSERT(nbRounds >= 8);
        acc = XXH3_avalanche(acc);
#if defined(__clang__)                                /* Clang */ \
    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */ \
    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
        /*
         * UGLY HACK:
         * Clang for ARMv7-A tries to vectorize this loop, similar to GCC x86.
         * In everywhere else, it uses scalar code.
         *
         * For 64->128-bit multiplies, even if the NEON was 100% optimal, it
         * would still be slower than UMAAL (see XXH_mult64to128).
         *
         * Unfortunately, Clang doesn't handle the long multiplies properly and
         * converts them to the nonexistent "vmulq_u64" intrinsic, which is then
         * scalarized into an ugly mess of VMOV.32 instructions.
         *
         * This mess is difficult to avoid without turning autovectorization
         * off completely, but they are usually relatively minor and/or not
         * worth it to fix.
         *
         * This loop is the easiest to fix, as unlike XXH32, this pragma
         * _actually works_ because it is a loop vectorization instead of an
         * SLP vectorization.
         */
        #pragma clang loop vectorize(disable)
#endif
        for (i=8 ; i < nbRounds; i++) {
            /*
             * Prevents clang for unrolling the acc loop and interleaving with this one.
             */
            XXH_COMPILER_GUARD(acc);
            acc_end += XXH3_mix16B(input+(16*i), secret+(16*(i-8)) + XXH3_MIDSIZE_STARTOFFSET, seed);
        }
        return XXH3_avalanche(acc + acc_end);
    }
}


/* =======     Long Keys     ======= */

#define XXH_STRIPE_LEN 64
#define XXH_SECRET_CONSUME_RATE 8   /* nb of secret bytes consumed at each accumulation */
#define XXH_ACC_NB (XXH_STRIPE_LEN / sizeof(xxh_u64))

#ifdef XXH_OLD_NAMES
#  define STRIPE_LEN XXH_STRIPE_LEN
#  define ACC_NB XXH_ACC_NB
#endif

#ifndef XXH_PREFETCH_DIST
#  ifdef __clang__
#    define XXH_PREFETCH_DIST 320
#  else
#    if (XXH_VECTOR == XXH_AVX512)
#      define XXH_PREFETCH_DIST 512
#    else
#      define XXH_PREFETCH_DIST 384
#    endif
#  endif  /* __clang__ */
#endif  /* XXH_PREFETCH_DIST */

/*
 * These macros are to generate an XXH3_accumulate() function.
 * The two arguments select the name suffix and target attribute.
 *
 * The name of this symbol is XXH3_accumulate_<name>() and it calls
 * XXH3_accumulate_512_<name>().
 *
 * It may be useful to hand implement this function if the compiler fails to
 * optimize the inline function.
 */
#define XXH3_ACCUMULATE_TEMPLATE(name)                      \
void                                                        \
XXH3_accumulate_##name(xxh_u64* XXH_RESTRICT acc,           \
                       const xxh_u8* XXH_RESTRICT input,    \
                       const xxh_u8* XXH_RESTRICT secret,   \
                       size_t nbStripes)                    \
{                                                           \
    size_t n;                                               \
    for (n = 0; n < nbStripes; n++ ) {                      \
        const xxh_u8* const in = input + n*XXH_STRIPE_LEN;  \
        XXH_PREFETCH(in + XXH_PREFETCH_DIST);               \
        XXH3_accumulate_512_##name(                         \
                 acc,                                       \
                 in,                                        \
                 secret + n*XXH_SECRET_CONSUME_RATE);       \
    }                                                       \
}


XXH_FORCE_INLINE void XXH_writeLE64(void* dst, xxh_u64 v64)
{
    if (!XXH_CPU_LITTLE_ENDIAN) v64 = XXH_swap64(v64);
    XXH_memcpy(dst, &v64, sizeof(v64));
}

/* Several intrinsic functions below are supposed to accept __int64 as argument,
 * as documented in https://software.intel.com/sites/landingpage/IntrinsicsGuide/ .
 * However, several environments do not define __int64 type,
 * requiring a workaround.
 */
#if !defined (__VMS) \
  && (defined (__cplusplus) \
  || (defined (__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) /* C99 */) )
    typedef int64_t xxh_i64;
#else
    /* the following type must have a width of 64-bit */
    typedef long long xxh_i64;
#endif


/*
 * XXH3_accumulate_512 is the tightest loop for long inputs, and it is the most optimized.
 *
 * It is a hardened version of UMAC, based off of FARSH's implementation.
 *
 * This was chosen because it adapts quite well to 32-bit, 64-bit, and SIMD
 * implementations, and it is ridiculously fast.
 *
 * We harden it by mixing the original input to the accumulators as well as the product.
 *
 * This means that in the (relatively likely) case of a multiply by zero, the
 * original input is preserved.
 *
 * On 128-bit inputs, we swap 64-bit pairs when we add the input to improve
 * cross-pollination, as otherwise the upper and lower halves would be
 * essentially independent.
 *
 * This doesn't matter on 64-bit hashes since they all get merged together in
 * the end, so we skip the extra step.
 *
 * Both XXH3_64bits and XXH3_128bits use this subroutine.
 */

#if (XXH_VECTOR == XXH_AVX512) \
     || (defined(XXH_DISPATCH_AVX512) && XXH_DISPATCH_AVX512 != 0)

#ifndef XXH_TARGET_AVX512
# define XXH_TARGET_AVX512  /* disable attribute target */
#endif

XXH_FORCE_INLINE XXH_TARGET_AVX512 void
XXH3_accumulate_512_avx512(void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
                     const void* XXH_RESTRICT secret)
{
    __m512i* const xacc = (__m512i *) acc;
    XXH_ASSERT((((size_t)acc) & 63) == 0);
    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));

    {
        /* data_vec    = input[0]; */
        __m512i const data_vec    = _mm512_loadu_si512   (input);
        /* key_vec     = secret[0]; */
        __m512i const key_vec     = _mm512_loadu_si512   (secret);
        /* data_key    = data_vec ^ key_vec; */
        __m512i const data_key    = _mm512_xor_si512     (data_vec, key_vec);
        /* data_key_lo = data_key >> 32; */
        __m512i const data_key_lo = _mm512_srli_epi64 (data_key, 32);
        /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
        __m512i const product     = _mm512_mul_epu32     (data_key, data_key_lo);
        /* xacc[0] += swap(data_vec); */
        __m512i const data_swap = _mm512_shuffle_epi32(data_vec, (_MM_PERM_ENUM)_MM_SHUFFLE(1, 0, 3, 2));
        __m512i const sum       = _mm512_add_epi64(*xacc, data_swap);
        /* xacc[0] += product; */
        *xacc = _mm512_add_epi64(product, sum);
    }
}
XXH_FORCE_INLINE XXH_TARGET_AVX512 XXH3_ACCUMULATE_TEMPLATE(avx512)

/*
 * XXH3_scrambleAcc: Scrambles the accumulators to improve mixing.
 *
 * Multiplication isn't perfect, as explained by Google in HighwayHash:
 *
 *  // Multiplication mixes/scrambles bytes 0-7 of the 64-bit result to
 *  // varying degrees. In descending order of goodness, bytes
 *  // 3 4 2 5 1 6 0 7 have quality 228 224 164 160 100 96 36 32.
 *  // As expected, the upper and lower bytes are much worse.
 *
 * Source: https://github.com/google/highwayhash/blob/0aaf66b/highwayhash/hh_avx2.h#L291
 *
 * Since our algorithm uses a pseudorandom secret to add some variance into the
 * mix, we don't need to (or want to) mix as often or as much as HighwayHash does.
 *
 * This isn't as tight as XXH3_accumulate, but still written in SIMD to avoid
 * extraction.
 *
 * Both XXH3_64bits and XXH3_128bits use this subroutine.
 */

XXH_FORCE_INLINE XXH_TARGET_AVX512 void
XXH3_scrambleAcc_avx512(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 63) == 0);
    XXH_STATIC_ASSERT(XXH_STRIPE_LEN == sizeof(__m512i));
    {   __m512i* const xacc = (__m512i*) acc;
        const __m512i prime32 = _mm512_set1_epi32((int)XXH_PRIME32_1);

        /* xacc[0] ^= (xacc[0] >> 47) */
        __m512i const acc_vec     = *xacc;
        __m512i const shifted     = _mm512_srli_epi64    (acc_vec, 47);
        /* xacc[0] ^= secret; */
        __m512i const key_vec     = _mm512_loadu_si512   (secret);
        __m512i const data_key    = _mm512_ternarylogic_epi32(key_vec, acc_vec, shifted, 0x96 /* key_vec ^ acc_vec ^ shifted */);

        /* xacc[0] *= XXH_PRIME32_1; */
        __m512i const data_key_hi = _mm512_srli_epi64 (data_key, 32);
        __m512i const prod_lo     = _mm512_mul_epu32     (data_key, prime32);
        __m512i const prod_hi     = _mm512_mul_epu32     (data_key_hi, prime32);
        *xacc = _mm512_add_epi64(prod_lo, _mm512_slli_epi64(prod_hi, 32));
    }
}

XXH_FORCE_INLINE XXH_TARGET_AVX512 void
XXH3_initCustomSecret_avx512(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
{
    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 63) == 0);
    XXH_STATIC_ASSERT(XXH_SEC_ALIGN == 64);
    XXH_ASSERT(((size_t)customSecret & 63) == 0);
    (void)(&XXH_writeLE64);
    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m512i);
        __m512i const seed_pos = _mm512_set1_epi64((xxh_i64)seed64);
        __m512i const seed     = _mm512_mask_sub_epi64(seed_pos, 0xAA, _mm512_set1_epi8(0), seed_pos);

        const __m512i* const src  = (const __m512i*) ((const void*) XXH3_kSecret);
              __m512i* const dest = (      __m512i*) customSecret;
        int i;
        XXH_ASSERT(((size_t)src & 63) == 0); /* control alignment */
        XXH_ASSERT(((size_t)dest & 63) == 0);
        for (i=0; i < nbRounds; ++i) {
            dest[i] = _mm512_add_epi64(_mm512_load_si512(src + i), seed);
    }   }
}

#endif

#if (XXH_VECTOR == XXH_AVX2) \
    || (defined(XXH_DISPATCH_AVX2) && XXH_DISPATCH_AVX2 != 0)

#ifndef XXH_TARGET_AVX2
# define XXH_TARGET_AVX2  /* disable attribute target */
#endif

XXH_FORCE_INLINE XXH_TARGET_AVX2 void
XXH3_accumulate_512_avx2( void* XXH_RESTRICT acc,
                    const void* XXH_RESTRICT input,
                    const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 31) == 0);
    {   __m256i* const xacc    =       (__m256i *) acc;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm256_loadu_si256 requires  a const __m256i * pointer for some reason. */
        const         __m256i* const xinput  = (const __m256i *) input;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
        const         __m256i* const xsecret = (const __m256i *) secret;

        size_t i;
        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
            /* data_vec    = xinput[i]; */
            __m256i const data_vec    = _mm256_loadu_si256    (xinput+i);
            /* key_vec     = xsecret[i]; */
            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
            /* data_key    = data_vec ^ key_vec; */
            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);
            /* data_key_lo = data_key >> 32; */
            __m256i const data_key_lo = _mm256_srli_epi64 (data_key, 32);
            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
            __m256i const product     = _mm256_mul_epu32     (data_key, data_key_lo);
            /* xacc[i] += swap(data_vec); */
            __m256i const data_swap = _mm256_shuffle_epi32(data_vec, _MM_SHUFFLE(1, 0, 3, 2));
            __m256i const sum       = _mm256_add_epi64(xacc[i], data_swap);
            /* xacc[i] += product; */
            xacc[i] = _mm256_add_epi64(product, sum);
    }   }
}
XXH_FORCE_INLINE XXH_TARGET_AVX2 XXH3_ACCUMULATE_TEMPLATE(avx2)

XXH_FORCE_INLINE XXH_TARGET_AVX2 void
XXH3_scrambleAcc_avx2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 31) == 0);
    {   __m256i* const xacc = (__m256i*) acc;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm256_loadu_si256 requires a const __m256i * pointer for some reason. */
        const         __m256i* const xsecret = (const __m256i *) secret;
        const __m256i prime32 = _mm256_set1_epi32((int)XXH_PRIME32_1);

        size_t i;
        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m256i); i++) {
            /* xacc[i] ^= (xacc[i] >> 47) */
            __m256i const acc_vec     = xacc[i];
            __m256i const shifted     = _mm256_srli_epi64    (acc_vec, 47);
            __m256i const data_vec    = _mm256_xor_si256     (acc_vec, shifted);
            /* xacc[i] ^= xsecret; */
            __m256i const key_vec     = _mm256_loadu_si256   (xsecret+i);
            __m256i const data_key    = _mm256_xor_si256     (data_vec, key_vec);

            /* xacc[i] *= XXH_PRIME32_1; */
            __m256i const data_key_hi = _mm256_srli_epi64 (data_key, 32);
            __m256i const prod_lo     = _mm256_mul_epu32     (data_key, prime32);
            __m256i const prod_hi     = _mm256_mul_epu32     (data_key_hi, prime32);
            xacc[i] = _mm256_add_epi64(prod_lo, _mm256_slli_epi64(prod_hi, 32));
        }
    }
}

XXH_FORCE_INLINE XXH_TARGET_AVX2 void XXH3_initCustomSecret_avx2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
{
    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 31) == 0);
    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE / sizeof(__m256i)) == 6);
    XXH_STATIC_ASSERT(XXH_SEC_ALIGN <= 64);
    (void)(&XXH_writeLE64);
    XXH_PREFETCH(customSecret);
    {   __m256i const seed = _mm256_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64, (xxh_i64)(0U - seed64), (xxh_i64)seed64);

        const __m256i* const src  = (const __m256i*) ((const void*) XXH3_kSecret);
              __m256i*       dest = (      __m256i*) customSecret;

#       if defined(__GNUC__) || defined(__clang__)
        /*
         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
         *   - do not extract the secret from sse registers in the internal loop
         *   - use less common registers, and avoid pushing these reg into stack
         */
        XXH_COMPILER_GUARD(dest);
#       endif
        XXH_ASSERT(((size_t)src & 31) == 0); /* control alignment */
        XXH_ASSERT(((size_t)dest & 31) == 0);

        /* GCC -O2 need unroll loop manually */
        dest[0] = _mm256_add_epi64(_mm256_load_si256(src+0), seed);
        dest[1] = _mm256_add_epi64(_mm256_load_si256(src+1), seed);
        dest[2] = _mm256_add_epi64(_mm256_load_si256(src+2), seed);
        dest[3] = _mm256_add_epi64(_mm256_load_si256(src+3), seed);
        dest[4] = _mm256_add_epi64(_mm256_load_si256(src+4), seed);
        dest[5] = _mm256_add_epi64(_mm256_load_si256(src+5), seed);
    }
}

#endif

/* x86dispatch always generates SSE2 */
#if (XXH_VECTOR == XXH_SSE2) || defined(XXH_X86DISPATCH)

#ifndef XXH_TARGET_SSE2
# define XXH_TARGET_SSE2  /* disable attribute target */
#endif

XXH_FORCE_INLINE XXH_TARGET_SSE2 void
XXH3_accumulate_512_sse2( void* XXH_RESTRICT acc,
                    const void* XXH_RESTRICT input,
                    const void* XXH_RESTRICT secret)
{
    /* SSE2 is just a half-scale version of the AVX2 version. */
    XXH_ASSERT((((size_t)acc) & 15) == 0);
    {   __m128i* const xacc    =       (__m128i *) acc;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
        const         __m128i* const xinput  = (const __m128i *) input;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
        const         __m128i* const xsecret = (const __m128i *) secret;

        size_t i;
        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
            /* data_vec    = xinput[i]; */
            __m128i const data_vec    = _mm_loadu_si128   (xinput+i);
            /* key_vec     = xsecret[i]; */
            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
            /* data_key    = data_vec ^ key_vec; */
            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);
            /* data_key_lo = data_key >> 32; */
            __m128i const data_key_lo = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
            /* product     = (data_key & 0xffffffff) * (data_key_lo & 0xffffffff); */
            __m128i const product     = _mm_mul_epu32     (data_key, data_key_lo);
            /* xacc[i] += swap(data_vec); */
            __m128i const data_swap = _mm_shuffle_epi32(data_vec, _MM_SHUFFLE(1,0,3,2));
            __m128i const sum       = _mm_add_epi64(xacc[i], data_swap);
            /* xacc[i] += product; */
            xacc[i] = _mm_add_epi64(product, sum);
    }   }
}
XXH_FORCE_INLINE XXH_TARGET_SSE2 XXH3_ACCUMULATE_TEMPLATE(sse2)

XXH_FORCE_INLINE XXH_TARGET_SSE2 void
XXH3_scrambleAcc_sse2(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 15) == 0);
    {   __m128i* const xacc = (__m128i*) acc;
        /* Unaligned. This is mainly for pointer arithmetic, and because
         * _mm_loadu_si128 requires a const __m128i * pointer for some reason. */
        const         __m128i* const xsecret = (const __m128i *) secret;
        const __m128i prime32 = _mm_set1_epi32((int)XXH_PRIME32_1);

        size_t i;
        for (i=0; i < XXH_STRIPE_LEN/sizeof(__m128i); i++) {
            /* xacc[i] ^= (xacc[i] >> 47) */
            __m128i const acc_vec     = xacc[i];
            __m128i const shifted     = _mm_srli_epi64    (acc_vec, 47);
            __m128i const data_vec    = _mm_xor_si128     (acc_vec, shifted);
            /* xacc[i] ^= xsecret[i]; */
            __m128i const key_vec     = _mm_loadu_si128   (xsecret+i);
            __m128i const data_key    = _mm_xor_si128     (data_vec, key_vec);

            /* xacc[i] *= XXH_PRIME32_1; */
            __m128i const data_key_hi = _mm_shuffle_epi32 (data_key, _MM_SHUFFLE(0, 3, 0, 1));
            __m128i const prod_lo     = _mm_mul_epu32     (data_key, prime32);
            __m128i const prod_hi     = _mm_mul_epu32     (data_key_hi, prime32);
            xacc[i] = _mm_add_epi64(prod_lo, _mm_slli_epi64(prod_hi, 32));
        }
    }
}

XXH_FORCE_INLINE XXH_TARGET_SSE2 void XXH3_initCustomSecret_sse2(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
{
    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);
    (void)(&XXH_writeLE64);
    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / sizeof(__m128i);

#       if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
        /* MSVC 32bit mode does not support _mm_set_epi64x before 2015 */
        XXH_ALIGN(16) const xxh_i64 seed64x2[2] = { (xxh_i64)seed64, (xxh_i64)(0U - seed64) };
        __m128i const seed = _mm_load_si128((__m128i const*)seed64x2);
#       else
        __m128i const seed = _mm_set_epi64x((xxh_i64)(0U - seed64), (xxh_i64)seed64);
#       endif
        int i;

        const void* const src16 = XXH3_kSecret;
        __m128i* dst16 = (__m128i*) customSecret;
#       if defined(__GNUC__) || defined(__clang__)
        /*
         * On GCC & Clang, marking 'dest' as modified will cause the compiler:
         *   - do not extract the secret from sse registers in the internal loop
         *   - use less common registers, and avoid pushing these reg into stack
         */
        XXH_COMPILER_GUARD(dst16);
#       endif
        XXH_ASSERT(((size_t)src16 & 15) == 0); /* control alignment */
        XXH_ASSERT(((size_t)dst16 & 15) == 0);

        for (i=0; i < nbRounds; ++i) {
            dst16[i] = _mm_add_epi64(_mm_load_si128((const __m128i *)src16+i), seed);
    }   }
}

#endif

#if (XXH_VECTOR == XXH_NEON)

/* forward declarations for the scalar routines */
XXH_FORCE_INLINE void
XXH3_scalarRound(void* XXH_RESTRICT acc, void const* XXH_RESTRICT input,
                 void const* XXH_RESTRICT secret, size_t lane);

XXH_FORCE_INLINE void
XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
                         void const* XXH_RESTRICT secret, size_t lane);

/*!
 * @internal
 * @brief The bulk processing loop for NEON and WASM SIMD128.
 *
 * The NEON code path is actually partially scalar when running on AArch64. This
 * is to optimize the pipelining and can have up to 15% speedup depending on the
 * CPU, and it also mitigates some GCC codegen issues.
 *
 * @see XXH3_NEON_LANES for configuring this and details about this optimization.
 *
 * NEON's 32-bit to 64-bit long multiply takes a half vector of 32-bit
 * integers instead of the other platforms which mask full 64-bit vectors,
 * so the setup is more complicated than just shifting right.
 *
 * Additionally, there is an optimization for 4 lanes at once noted below.
 *
 * Since, as stated, the most optimal amount of lanes for Cortexes is 6,
 * there needs to be *three* versions of the accumulate operation used
 * for the remaining 2 lanes.
 *
 * WASM's SIMD128 uses SIMDe's arm_neon.h polyfill because the intrinsics overlap
 * nearly perfectly.
 */

XXH_FORCE_INLINE void
XXH3_accumulate_512_neon( void* XXH_RESTRICT acc,
                    const void* XXH_RESTRICT input,
                    const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 15) == 0);
    XXH_STATIC_ASSERT(XXH3_NEON_LANES > 0 && XXH3_NEON_LANES <= XXH_ACC_NB && XXH3_NEON_LANES % 2 == 0);
    {   /* GCC for darwin arm64 does not like aliasing here */
        xxh_aliasing_uint64x2_t* const xacc = (xxh_aliasing_uint64x2_t*) acc;
        /* We don't use a uint32x4_t pointer because it causes bus errors on ARMv7. */
        uint8_t const* xinput = (const uint8_t *) input;
        uint8_t const* xsecret  = (const uint8_t *) secret;

        size_t i;
#ifdef __wasm_simd128__
        /*
         * On WASM SIMD128, Clang emits direct address loads when XXH3_kSecret
         * is constant propagated, which results in it converting it to this
         * inside the loop:
         *
         *    a = v128.load(XXH3_kSecret +  0 + $secret_offset, offset = 0)
         *    b = v128.load(XXH3_kSecret + 16 + $secret_offset, offset = 0)
         *    ...
         *
         * This requires a full 32-bit address immediate (and therefore a 6 byte
         * instruction) as well as an add for each offset.
         *
         * Putting an asm guard prevents it from folding (at the cost of losing
         * the alignment hint), and uses the free offset in `v128.load` instead
         * of adding secret_offset each time which overall reduces code size by
         * about a kilobyte and improves performance.
         */
        XXH_COMPILER_GUARD(xsecret);
#endif
        /* Scalar lanes use the normal scalarRound routine */
        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
            XXH3_scalarRound(acc, input, secret, i);
        }
        i = 0;
        /* 4 NEON lanes at a time. */
        for (; i+1 < XXH3_NEON_LANES / 2; i+=2) {
            /* data_vec = xinput[i]; */
            uint64x2_t data_vec_1 = XXH_vld1q_u64(xinput  + (i * 16));
            uint64x2_t data_vec_2 = XXH_vld1q_u64(xinput  + ((i+1) * 16));
            /* key_vec  = xsecret[i];  */
            uint64x2_t key_vec_1  = XXH_vld1q_u64(xsecret + (i * 16));
            uint64x2_t key_vec_2  = XXH_vld1q_u64(xsecret + ((i+1) * 16));
            /* data_swap = swap(data_vec) */
            uint64x2_t data_swap_1 = vextq_u64(data_vec_1, data_vec_1, 1);
            uint64x2_t data_swap_2 = vextq_u64(data_vec_2, data_vec_2, 1);
            /* data_key = data_vec ^ key_vec; */
            uint64x2_t data_key_1 = veorq_u64(data_vec_1, key_vec_1);
            uint64x2_t data_key_2 = veorq_u64(data_vec_2, key_vec_2);

            /*
             * If we reinterpret the 64x2 vectors as 32x4 vectors, we can use a
             * de-interleave operation for 4 lanes in 1 step with `vuzpq_u32` to
             * get one vector with the low 32 bits of each lane, and one vector
             * with the high 32 bits of each lane.
             *
             * The intrinsic returns a double vector because the original ARMv7-a
             * instruction modified both arguments in place. AArch64 and SIMD128 emit
             * two instructions from this intrinsic.
             *
             *  [ dk11L | dk11H | dk12L | dk12H ] -> [ dk11L | dk12L | dk21L | dk22L ]
             *  [ dk21L | dk21H | dk22L | dk22H ] -> [ dk11H | dk12H | dk21H | dk22H ]
             */
            uint32x4x2_t unzipped = vuzpq_u32(
                vreinterpretq_u32_u64(data_key_1),
                vreinterpretq_u32_u64(data_key_2)
            );
            /* data_key_lo = data_key & 0xFFFFFFFF */
            uint32x4_t data_key_lo = unzipped.val[0];
            /* data_key_hi = data_key >> 32 */
            uint32x4_t data_key_hi = unzipped.val[1];
            /*
             * Then, we can split the vectors horizontally and multiply which, as for most
             * widening intrinsics, have a variant that works on both high half vectors
             * for free on AArch64. A similar instruction is available on SIMD128.
             *
             * sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi
             */
            uint64x2_t sum_1 = XXH_vmlal_low_u32(data_swap_1, data_key_lo, data_key_hi);
            uint64x2_t sum_2 = XXH_vmlal_high_u32(data_swap_2, data_key_lo, data_key_hi);
            /*
             * Clang reorders
             *    a += b * c;     // umlal   swap.2d, dkl.2s, dkh.2s
             *    c += a;         // add     acc.2d, acc.2d, swap.2d
             * to
             *    c += a;         // add     acc.2d, acc.2d, swap.2d
             *    c += b * c;     // umlal   acc.2d, dkl.2s, dkh.2s
             *
             * While it would make sense in theory since the addition is faster,
             * for reasons likely related to umlal being limited to certain NEON
             * pipelines, this is worse. A compiler guard fixes this.
             */
            XXH_COMPILER_GUARD_CLANG_NEON(sum_1);
            XXH_COMPILER_GUARD_CLANG_NEON(sum_2);
            /* xacc[i] = acc_vec + sum; */
            xacc[i]   = vaddq_u64(xacc[i], sum_1);
            xacc[i+1] = vaddq_u64(xacc[i+1], sum_2);
        }
        /* Operate on the remaining NEON lanes 2 at a time. */
        for (; i < XXH3_NEON_LANES / 2; i++) {
            /* data_vec = xinput[i]; */
            uint64x2_t data_vec = XXH_vld1q_u64(xinput  + (i * 16));
            /* key_vec  = xsecret[i];  */
            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
            /* acc_vec_2 = swap(data_vec) */
            uint64x2_t data_swap = vextq_u64(data_vec, data_vec, 1);
            /* data_key = data_vec ^ key_vec; */
            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
            /* For two lanes, just use VMOVN and VSHRN. */
            /* data_key_lo = data_key & 0xFFFFFFFF; */
            uint32x2_t data_key_lo = vmovn_u64(data_key);
            /* data_key_hi = data_key >> 32; */
            uint32x2_t data_key_hi = vshrn_n_u64(data_key, 32);
            /* sum = data_swap + (u64x2) data_key_lo * (u64x2) data_key_hi; */
            uint64x2_t sum = vmlal_u32(data_swap, data_key_lo, data_key_hi);
            /* Same Clang workaround as before */
            XXH_COMPILER_GUARD_CLANG_NEON(sum);
            /* xacc[i] = acc_vec + sum; */
            xacc[i] = vaddq_u64 (xacc[i], sum);
        }
    }
}
XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(neon)

XXH_FORCE_INLINE void
XXH3_scrambleAcc_neon(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 15) == 0);

    {   xxh_aliasing_uint64x2_t* xacc       = (xxh_aliasing_uint64x2_t*) acc;
        uint8_t const* xsecret = (uint8_t const*) secret;

        size_t i;
        /* WASM uses operator overloads and doesn't need these. */
#ifndef __wasm_simd128__
        /* { prime32_1, prime32_1 } */
        uint32x2_t const kPrimeLo = vdup_n_u32(XXH_PRIME32_1);
        /* { 0, prime32_1, 0, prime32_1 } */
        uint32x4_t const kPrimeHi = vreinterpretq_u32_u64(vdupq_n_u64((xxh_u64)XXH_PRIME32_1 << 32));
#endif

        /* AArch64 uses both scalar and neon at the same time */
        for (i = XXH3_NEON_LANES; i < XXH_ACC_NB; i++) {
            XXH3_scalarScrambleRound(acc, secret, i);
        }
        for (i=0; i < XXH3_NEON_LANES / 2; i++) {
            /* xacc[i] ^= (xacc[i] >> 47); */
            uint64x2_t acc_vec  = xacc[i];
            uint64x2_t shifted  = vshrq_n_u64(acc_vec, 47);
            uint64x2_t data_vec = veorq_u64(acc_vec, shifted);

            /* xacc[i] ^= xsecret[i]; */
            uint64x2_t key_vec  = XXH_vld1q_u64(xsecret + (i * 16));
            uint64x2_t data_key = veorq_u64(data_vec, key_vec);
            /* xacc[i] *= XXH_PRIME32_1 */
#ifdef __wasm_simd128__
            /* SIMD128 has multiply by u64x2, use it instead of expanding and scalarizing */
            xacc[i] = data_key * XXH_PRIME32_1;
#else
            /*
             * Expanded version with portable NEON intrinsics
             *
             *    lo(x) * lo(y) + (hi(x) * lo(y) << 32)
             *
             * prod_hi = hi(data_key) * lo(prime) << 32
             *
             * Since we only need 32 bits of this multiply a trick can be used, reinterpreting the vector
             * as a uint32x4_t and multiplying by { 0, prime, 0, prime } to cancel out the unwanted bits
             * and avoid the shift.
             */
            uint32x4_t prod_hi = vmulq_u32 (vreinterpretq_u32_u64(data_key), kPrimeHi);
            /* Extract low bits for vmlal_u32  */
            uint32x2_t data_key_lo = vmovn_u64(data_key);
            /* xacc[i] = prod_hi + lo(data_key) * XXH_PRIME32_1; */
            xacc[i] = vmlal_u32(vreinterpretq_u64_u32(prod_hi), data_key_lo, kPrimeLo);
#endif
        }
    }
}
#endif

#if (XXH_VECTOR == XXH_VSX)

XXH_FORCE_INLINE void
XXH3_accumulate_512_vsx(  void* XXH_RESTRICT acc,
                    const void* XXH_RESTRICT input,
                    const void* XXH_RESTRICT secret)
{
    /* presumed aligned */
    xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
    xxh_u8 const* const xinput   = (xxh_u8 const*) input;   /* no alignment restriction */
    xxh_u8 const* const xsecret  = (xxh_u8 const*) secret;    /* no alignment restriction */
    xxh_u64x2 const v32 = { 32, 32 };
    size_t i;
    for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
        /* data_vec = xinput[i]; */
        xxh_u64x2 const data_vec = XXH_vec_loadu(xinput + 16*i);
        /* key_vec = xsecret[i]; */
        xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
        xxh_u64x2 const data_key = data_vec ^ key_vec;
        /* shuffled = (data_key << 32) | (data_key >> 32); */
        xxh_u32x4 const shuffled = (xxh_u32x4)vec_rl(data_key, v32);
        /* product = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)shuffled & 0xFFFFFFFF); */
        xxh_u64x2 const product  = XXH_vec_mulo((xxh_u32x4)data_key, shuffled);
        /* acc_vec = xacc[i]; */
        xxh_u64x2 acc_vec        = xacc[i];
        acc_vec += product;

        /* swap high and low halves */
#ifdef __s390x__
        acc_vec += vec_permi(data_vec, data_vec, 2);
#else
        acc_vec += vec_xxpermdi(data_vec, data_vec, 2);
#endif
        xacc[i] = acc_vec;
    }
}
XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(vsx)

XXH_FORCE_INLINE void
XXH3_scrambleAcc_vsx(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    XXH_ASSERT((((size_t)acc) & 15) == 0);

    {   xxh_aliasing_u64x2* const xacc = (xxh_aliasing_u64x2*) acc;
        const xxh_u8* const xsecret = (const xxh_u8*) secret;
        /* constants */
        xxh_u64x2 const v32  = { 32, 32 };
        xxh_u64x2 const v47 = { 47, 47 };
        xxh_u32x4 const prime = { XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1, XXH_PRIME32_1 };
        size_t i;
        for (i = 0; i < XXH_STRIPE_LEN / sizeof(xxh_u64x2); i++) {
            /* xacc[i] ^= (xacc[i] >> 47); */
            xxh_u64x2 const acc_vec  = xacc[i];
            xxh_u64x2 const data_vec = acc_vec ^ (acc_vec >> v47);

            /* xacc[i] ^= xsecret[i]; */
            xxh_u64x2 const key_vec  = XXH_vec_loadu(xsecret + 16*i);
            xxh_u64x2 const data_key = data_vec ^ key_vec;

            /* xacc[i] *= XXH_PRIME32_1 */
            /* prod_lo = ((xxh_u64x2)data_key & 0xFFFFFFFF) * ((xxh_u64x2)prime & 0xFFFFFFFF);  */
            xxh_u64x2 const prod_even  = XXH_vec_mule((xxh_u32x4)data_key, prime);
            /* prod_hi = ((xxh_u64x2)data_key >> 32) * ((xxh_u64x2)prime >> 32);  */
            xxh_u64x2 const prod_odd  = XXH_vec_mulo((xxh_u32x4)data_key, prime);
            xacc[i] = prod_odd + (prod_even << v32);
    }   }
}

#endif

#if (XXH_VECTOR == XXH_SVE)

XXH_FORCE_INLINE void
XXH3_accumulate_512_sve( void* XXH_RESTRICT acc,
                   const void* XXH_RESTRICT input,
                   const void* XXH_RESTRICT secret)
{
    uint64_t *xacc = (uint64_t *)acc;
    const uint64_t *xinput = (const uint64_t *)(const void *)input;
    const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
    svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
    uint64_t element_count = svcntd();
    if (element_count >= 8) {
        svbool_t mask = svptrue_pat_b64(SV_VL8);
        svuint64_t vacc = svld1_u64(mask, xacc);
        ACCRND(vacc, 0);
        svst1_u64(mask, xacc, vacc);
    } else if (element_count == 2) {   /* sve128 */
        svbool_t mask = svptrue_pat_b64(SV_VL2);
        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
        svuint64_t acc1 = svld1_u64(mask, xacc + 2);
        svuint64_t acc2 = svld1_u64(mask, xacc + 4);
        svuint64_t acc3 = svld1_u64(mask, xacc + 6);
        ACCRND(acc0, 0);
        ACCRND(acc1, 2);
        ACCRND(acc2, 4);
        ACCRND(acc3, 6);
        svst1_u64(mask, xacc + 0, acc0);
        svst1_u64(mask, xacc + 2, acc1);
        svst1_u64(mask, xacc + 4, acc2);
        svst1_u64(mask, xacc + 6, acc3);
    } else {
        svbool_t mask = svptrue_pat_b64(SV_VL4);
        svuint64_t acc0 = svld1_u64(mask, xacc + 0);
        svuint64_t acc1 = svld1_u64(mask, xacc + 4);
        ACCRND(acc0, 0);
        ACCRND(acc1, 4);
        svst1_u64(mask, xacc + 0, acc0);
        svst1_u64(mask, xacc + 4, acc1);
    }
}

XXH_FORCE_INLINE void
XXH3_accumulate_sve(xxh_u64* XXH_RESTRICT acc,
               const xxh_u8* XXH_RESTRICT input,
               const xxh_u8* XXH_RESTRICT secret,
               size_t nbStripes)
{
    if (nbStripes != 0) {
        uint64_t *xacc = (uint64_t *)acc;
        const uint64_t *xinput = (const uint64_t *)(const void *)input;
        const uint64_t *xsecret = (const uint64_t *)(const void *)secret;
        svuint64_t kSwap = sveor_n_u64_z(svptrue_b64(), svindex_u64(0, 1), 1);
        uint64_t element_count = svcntd();
        if (element_count >= 8) {
            svbool_t mask = svptrue_pat_b64(SV_VL8);
            svuint64_t vacc = svld1_u64(mask, xacc + 0);
            do {
                /* svprfd(svbool_t, void *, enum svfprop); */
                svprfd(mask, xinput + 128, SV_PLDL1STRM);
                ACCRND(vacc, 0);
                xinput += 8;
                xsecret += 1;
                nbStripes--;
           } while (nbStripes != 0);

           svst1_u64(mask, xacc + 0, vacc);
        } else if (element_count == 2) { /* sve128 */
            svbool_t mask = svptrue_pat_b64(SV_VL2);
            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
            svuint64_t acc1 = svld1_u64(mask, xacc + 2);
            svuint64_t acc2 = svld1_u64(mask, xacc + 4);
            svuint64_t acc3 = svld1_u64(mask, xacc + 6);
            do {
                svprfd(mask, xinput + 128, SV_PLDL1STRM);
                ACCRND(acc0, 0);
                ACCRND(acc1, 2);
                ACCRND(acc2, 4);
                ACCRND(acc3, 6);
                xinput += 8;
                xsecret += 1;
                nbStripes--;
           } while (nbStripes != 0);

           svst1_u64(mask, xacc + 0, acc0);
           svst1_u64(mask, xacc + 2, acc1);
           svst1_u64(mask, xacc + 4, acc2);
           svst1_u64(mask, xacc + 6, acc3);
        } else {
            svbool_t mask = svptrue_pat_b64(SV_VL4);
            svuint64_t acc0 = svld1_u64(mask, xacc + 0);
            svuint64_t acc1 = svld1_u64(mask, xacc + 4);
            do {
                svprfd(mask, xinput + 128, SV_PLDL1STRM);
                ACCRND(acc0, 0);
                ACCRND(acc1, 4);
                xinput += 8;
                xsecret += 1;
                nbStripes--;
           } while (nbStripes != 0);

           svst1_u64(mask, xacc + 0, acc0);
           svst1_u64(mask, xacc + 4, acc1);
       }
    }
}

#endif

/* scalar variants - universal */

#if defined(__aarch64__) && (defined(__GNUC__) || defined(__clang__))
/*
 * In XXH3_scalarRound(), GCC and Clang have a similar codegen issue, where they
 * emit an excess mask and a full 64-bit multiply-add (MADD X-form).
 *
 * While this might not seem like much, as AArch64 is a 64-bit architecture, only
 * big Cortex designs have a full 64-bit multiplier.
 *
 * On the little cores, the smaller 32-bit multiplier is used, and full 64-bit
 * multiplies expand to 2-3 multiplies in microcode. This has a major penalty
 * of up to 4 latency cycles and 2 stall cycles in the multiply pipeline.
 *
 * Thankfully, AArch64 still provides the 32-bit long multiply-add (UMADDL) which does
 * not have this penalty and does the mask automatically.
 */
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
{
    xxh_u64 ret;
    /* note: %x = 64-bit register, %w = 32-bit register */
    __asm__("umaddl %x0, %w1, %w2, %x3" : "=r" (ret) : "r" (lhs), "r" (rhs), "r" (acc));
    return ret;
}
#else
XXH_FORCE_INLINE xxh_u64
XXH_mult32to64_add64(xxh_u64 lhs, xxh_u64 rhs, xxh_u64 acc)
{
    return XXH_mult32to64((xxh_u32)lhs, (xxh_u32)rhs) + acc;
}
#endif

/*!
 * @internal
 * @brief Scalar round for @ref XXH3_accumulate_512_scalar().
 *
 * This is extracted to its own function because the NEON path uses a combination
 * of NEON and scalar.
 */
XXH_FORCE_INLINE void
XXH3_scalarRound(void* XXH_RESTRICT acc,
                 void const* XXH_RESTRICT input,
                 void const* XXH_RESTRICT secret,
                 size_t lane)
{
    xxh_u64* xacc = (xxh_u64*) acc;
    xxh_u8 const* xinput  = (xxh_u8 const*) input;
    xxh_u8 const* xsecret = (xxh_u8 const*) secret;
    XXH_ASSERT(lane < XXH_ACC_NB);
    XXH_ASSERT(((size_t)acc & (XXH_ACC_ALIGN-1)) == 0);
    {
        xxh_u64 const data_val = XXH_readLE64(xinput + lane * 8);
        xxh_u64 const data_key = data_val ^ XXH_readLE64(xsecret + lane * 8);
        xacc[lane ^ 1] += data_val; /* swap adjacent lanes */
        xacc[lane] = XXH_mult32to64_add64(data_key /* & 0xFFFFFFFF */, data_key >> 32, xacc[lane]);
    }
}

/*!
 * @internal
 * @brief Processes a 64 byte block of data using the scalar path.
 */
XXH_FORCE_INLINE void
XXH3_accumulate_512_scalar(void* XXH_RESTRICT acc,
                     const void* XXH_RESTRICT input,
                     const void* XXH_RESTRICT secret)
{
    size_t i;
    /* ARM GCC refuses to unroll this loop, resulting in a 24% slowdown on ARMv6. */
#if defined(__GNUC__) && !defined(__clang__) \
  && (defined(__arm__) || defined(__thumb2__)) \
  && defined(__ARM_FEATURE_UNALIGNED) /* no unaligned access just wastes bytes */ \
  && XXH_SIZE_OPT <= 0
#  pragma GCC unroll 8
#endif
    for (i=0; i < XXH_ACC_NB; i++) {
        XXH3_scalarRound(acc, input, secret, i);
    }
}
XXH_FORCE_INLINE XXH3_ACCUMULATE_TEMPLATE(scalar)

/*!
 * @internal
 * @brief Scalar scramble step for @ref XXH3_scrambleAcc_scalar().
 *
 * This is extracted to its own function because the NEON path uses a combination
 * of NEON and scalar.
 */
XXH_FORCE_INLINE void
XXH3_scalarScrambleRound(void* XXH_RESTRICT acc,
                         void const* XXH_RESTRICT secret,
                         size_t lane)
{
    xxh_u64* const xacc = (xxh_u64*) acc;   /* presumed aligned */
    const xxh_u8* const xsecret = (const xxh_u8*) secret;   /* no alignment restriction */
    XXH_ASSERT((((size_t)acc) & (XXH_ACC_ALIGN-1)) == 0);
    XXH_ASSERT(lane < XXH_ACC_NB);
    {
        xxh_u64 const key64 = XXH_readLE64(xsecret + lane * 8);
        xxh_u64 acc64 = xacc[lane];
        acc64 = XXH_xorshift64(acc64, 47);
        acc64 ^= key64;
        acc64 *= XXH_PRIME32_1;
        xacc[lane] = acc64;
    }
}

/*!
 * @internal
 * @brief Scrambles the accumulators after a large chunk has been read
 */
XXH_FORCE_INLINE void
XXH3_scrambleAcc_scalar(void* XXH_RESTRICT acc, const void* XXH_RESTRICT secret)
{
    size_t i;
    for (i=0; i < XXH_ACC_NB; i++) {
        XXH3_scalarScrambleRound(acc, secret, i);
    }
}

XXH_FORCE_INLINE void
XXH3_initCustomSecret_scalar(void* XXH_RESTRICT customSecret, xxh_u64 seed64)
{
    /*
     * We need a separate pointer for the hack below,
     * which requires a non-const pointer.
     * Any decent compiler will optimize this out otherwise.
     */
    const xxh_u8* kSecretPtr = XXH3_kSecret;
    XXH_STATIC_ASSERT((XXH_SECRET_DEFAULT_SIZE & 15) == 0);

#if defined(__GNUC__) && defined(__aarch64__)
    /*
     * UGLY HACK:
     * GCC and Clang generate a bunch of MOV/MOVK pairs for aarch64, and they are
     * placed sequentially, in order, at the top of the unrolled loop.
     *
     * While MOVK is great for generating constants (2 cycles for a 64-bit
     * constant compared to 4 cycles for LDR), it fights for bandwidth with
     * the arithmetic instructions.
     *
     *   I   L   S
     * MOVK
     * MOVK
     * MOVK
     * MOVK
     * ADD
     * SUB      STR
     *          STR
     * By forcing loads from memory (as the asm line causes the compiler to assume
     * that XXH3_kSecretPtr has been changed), the pipelines are used more
     * efficiently:
     *   I   L   S
     *      LDR
     *  ADD LDR
     *  SUB     STR
     *          STR
     *
     * See XXH3_NEON_LANES for details on the pipsline.
     *
     * XXH3_64bits_withSeed, len == 256, Snapdragon 835
     *   without hack: 2654.4 MB/s
     *   with hack:    3202.9 MB/s
     */
    XXH_COMPILER_GUARD(kSecretPtr);
#endif
    {   int const nbRounds = XXH_SECRET_DEFAULT_SIZE / 16;
        int i;
        for (i=0; i < nbRounds; i++) {
            /*
             * The asm hack causes the compiler to assume that kSecretPtr aliases with
             * customSecret, and on aarch64, this prevented LDP from merging two
             * loads together for free. Putting the loads together before the stores
             * properly generates LDP.
             */
            xxh_u64 lo = XXH_readLE64(kSecretPtr + 16*i)     + seed64;
            xxh_u64 hi = XXH_readLE64(kSecretPtr + 16*i + 8) - seed64;
            XXH_writeLE64((xxh_u8*)customSecret + 16*i,     lo);
            XXH_writeLE64((xxh_u8*)customSecret + 16*i + 8, hi);
    }   }
}


typedef void (*XXH3_f_accumulate)(xxh_u64* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, const xxh_u8* XXH_RESTRICT, size_t);
typedef void (*XXH3_f_scrambleAcc)(void* XXH_RESTRICT, const void*);
typedef void (*XXH3_f_initCustomSecret)(void* XXH_RESTRICT, xxh_u64);


#if (XXH_VECTOR == XXH_AVX512)

#define XXH3_accumulate_512 XXH3_accumulate_512_avx512
#define XXH3_accumulate     XXH3_accumulate_avx512
#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx512
#define XXH3_initCustomSecret XXH3_initCustomSecret_avx512

#elif (XXH_VECTOR == XXH_AVX2)

#define XXH3_accumulate_512 XXH3_accumulate_512_avx2
#define XXH3_accumulate     XXH3_accumulate_avx2
#define XXH3_scrambleAcc    XXH3_scrambleAcc_avx2
#define XXH3_initCustomSecret XXH3_initCustomSecret_avx2

#elif (XXH_VECTOR == XXH_SSE2)

#define XXH3_accumulate_512 XXH3_accumulate_512_sse2
#define XXH3_accumulate     XXH3_accumulate_sse2
#define XXH3_scrambleAcc    XXH3_scrambleAcc_sse2
#define XXH3_initCustomSecret XXH3_initCustomSecret_sse2

#elif (XXH_VECTOR == XXH_NEON)

#define XXH3_accumulate_512 XXH3_accumulate_512_neon
#define XXH3_accumulate     XXH3_accumulate_neon
#define XXH3_scrambleAcc    XXH3_scrambleAcc_neon
#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar

#elif (XXH_VECTOR == XXH_VSX)

#define XXH3_accumulate_512 XXH3_accumulate_512_vsx
#define XXH3_accumulate     XXH3_accumulate_vsx
#define XXH3_scrambleAcc    XXH3_scrambleAcc_vsx
#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar

#elif (XXH_VECTOR == XXH_SVE)
#define XXH3_accumulate_512 XXH3_accumulate_512_sve
#define XXH3_accumulate     XXH3_accumulate_sve
#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar

#else /* scalar */

#define XXH3_accumulate_512 XXH3_accumulate_512_scalar
#define XXH3_accumulate     XXH3_accumulate_scalar
#define XXH3_scrambleAcc    XXH3_scrambleAcc_scalar
#define XXH3_initCustomSecret XXH3_initCustomSecret_scalar

#endif

#if XXH_SIZE_OPT >= 1 /* don't do SIMD for initialization */
#  undef XXH3_initCustomSecret
#  define XXH3_initCustomSecret XXH3_initCustomSecret_scalar
#endif

XXH_FORCE_INLINE void
XXH3_hashLong_internal_loop(xxh_u64* XXH_RESTRICT acc,
                      const xxh_u8* XXH_RESTRICT input, size_t len,
                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                            XXH3_f_accumulate f_acc,
                            XXH3_f_scrambleAcc f_scramble)
{
    size_t const nbStripesPerBlock = (secretSize - XXH_STRIPE_LEN) / XXH_SECRET_CONSUME_RATE;
    size_t const block_len = XXH_STRIPE_LEN * nbStripesPerBlock;
    size_t const nb_blocks = (len - 1) / block_len;

    size_t n;

    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);

    for (n = 0; n < nb_blocks; n++) {
        f_acc(acc, input + n*block_len, secret, nbStripesPerBlock);
        f_scramble(acc, secret + secretSize - XXH_STRIPE_LEN);
    }

    /* last partial block */
    XXH_ASSERT(len > XXH_STRIPE_LEN);
    {   size_t const nbStripes = ((len - 1) - (block_len * nb_blocks)) / XXH_STRIPE_LEN;
        XXH_ASSERT(nbStripes <= (secretSize / XXH_SECRET_CONSUME_RATE));
        f_acc(acc, input + nb_blocks*block_len, secret, nbStripes);

        /* last stripe */
        {   const xxh_u8* const p = input + len - XXH_STRIPE_LEN;
#define XXH_SECRET_LASTACC_START 7  /* not aligned on 8, last secret is different from acc & scrambler */
            XXH3_accumulate_512(acc, p, secret + secretSize - XXH_STRIPE_LEN - XXH_SECRET_LASTACC_START);
    }   }
}

XXH_FORCE_INLINE xxh_u64
XXH3_mix2Accs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret)
{
    return XXH3_mul128_fold64(
               acc[0] ^ XXH_readLE64(secret),
               acc[1] ^ XXH_readLE64(secret+8) );
}

static XXH64_hash_t
XXH3_mergeAccs(const xxh_u64* XXH_RESTRICT acc, const xxh_u8* XXH_RESTRICT secret, xxh_u64 start)
{
    xxh_u64 result64 = start;
    size_t i = 0;

    for (i = 0; i < 4; i++) {
        result64 += XXH3_mix2Accs(acc+2*i, secret + 16*i);
#if defined(__clang__)                                /* Clang */ \
    && (defined(__arm__) || defined(__thumb__))       /* ARMv7 */ \
    && (defined(__ARM_NEON) || defined(__ARM_NEON__)) /* NEON */  \
    && !defined(XXH_ENABLE_AUTOVECTORIZE)             /* Define to disable */
        /*
         * UGLY HACK:
         * Prevent autovectorization on Clang ARMv7-a. Exact same problem as
         * the one in XXH3_len_129to240_64b. Speeds up shorter keys > 240b.
         * XXH3_64bits, len == 256, Snapdragon 835:
         *   without hack: 2063.7 MB/s
         *   with hack:    2560.7 MB/s
         */
        XXH_COMPILER_GUARD(result64);
#endif
    }

    return XXH3_avalanche(result64);
}

#define XXH3_INIT_ACC { XXH_PRIME32_3, XXH_PRIME64_1, XXH_PRIME64_2, XXH_PRIME64_3, \
                        XXH_PRIME64_4, XXH_PRIME32_2, XXH_PRIME64_5, XXH_PRIME32_1 }

XXH_FORCE_INLINE XXH64_hash_t
XXH3_hashLong_64b_internal(const void* XXH_RESTRICT input, size_t len,
                           const void* XXH_RESTRICT secret, size_t secretSize,
                           XXH3_f_accumulate f_acc,
                           XXH3_f_scrambleAcc f_scramble)
{
    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;

    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, (const xxh_u8*)secret, secretSize, f_acc, f_scramble);

    /* converge into final hash */
    XXH_STATIC_ASSERT(sizeof(acc) == 64);
    /* do not align on 8, so that the secret is different from the accumulator */
#define XXH_SECRET_MERGEACCS_START 11
    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
    return XXH3_mergeAccs(acc, (const xxh_u8*)secret + XXH_SECRET_MERGEACCS_START, (xxh_u64)len * XXH_PRIME64_1);
}

/*
 * It's important for performance to transmit secret's size (when it's static)
 * so that the compiler can properly optimize the vectorized loop.
 * This makes a big performance difference for "medium" keys (<1 KB) when using AVX instruction set.
 * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
 * breaks -Og, this is XXH_NO_INLINE.
 */
XXH3_WITH_SECRET_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSecret(const void* XXH_RESTRICT input, size_t len,
                             XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
    (void)seed64;
    return XXH3_hashLong_64b_internal(input, len, secret, secretLen, XXH3_accumulate, XXH3_scrambleAcc);
}

/*
 * It's preferable for performance that XXH3_hashLong is not inlined,
 * as it results in a smaller function for small data, easier to the instruction cache.
 * Note that inside this no_inline function, we do inline the internal loop,
 * and provide a statically defined secret size to allow optimization of vector loop.
 */
XXH_NO_INLINE XXH_PUREF XXH64_hash_t
XXH3_hashLong_64b_default(const void* XXH_RESTRICT input, size_t len,
                          XXH64_hash_t seed64, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
    (void)seed64; (void)secret; (void)secretLen;
    return XXH3_hashLong_64b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_accumulate, XXH3_scrambleAcc);
}

/*
 * XXH3_hashLong_64b_withSeed():
 * Generate a custom key based on alteration of default XXH3_kSecret with the seed,
 * and then use this key for long mode hashing.
 *
 * This operation is decently fast but nonetheless costs a little bit of time.
 * Try to avoid it whenever possible (typically when seed==0).
 *
 * It's important for performance that XXH3_hashLong is not inlined. Not sure
 * why (uop cache maybe?), but the difference is large and easily measurable.
 */
XXH_FORCE_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSeed_internal(const void* input, size_t len,
                                    XXH64_hash_t seed,
                                    XXH3_f_accumulate f_acc,
                                    XXH3_f_scrambleAcc f_scramble,
                                    XXH3_f_initCustomSecret f_initSec)
{
#if XXH_SIZE_OPT <= 0
    if (seed == 0)
        return XXH3_hashLong_64b_internal(input, len,
                                          XXH3_kSecret, sizeof(XXH3_kSecret),
                                          f_acc, f_scramble);
#endif
    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
        f_initSec(secret, seed);
        return XXH3_hashLong_64b_internal(input, len, secret, sizeof(secret),
                                          f_acc, f_scramble);
    }
}

/*
 * It's important for performance that XXH3_hashLong is not inlined.
 */
XXH_NO_INLINE XXH64_hash_t
XXH3_hashLong_64b_withSeed(const void* XXH_RESTRICT input, size_t len,
                           XXH64_hash_t seed, const xxh_u8* XXH_RESTRICT secret, size_t secretLen)
{
    (void)secret; (void)secretLen;
    return XXH3_hashLong_64b_withSeed_internal(input, len, seed,
                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
}


typedef XXH64_hash_t (*XXH3_hashLong64_f)(const void* XXH_RESTRICT, size_t,
                                          XXH64_hash_t, const xxh_u8* XXH_RESTRICT, size_t);

XXH_FORCE_INLINE XXH64_hash_t
XXH3_64bits_internal(const void* XXH_RESTRICT input, size_t len,
                     XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
                     XXH3_hashLong64_f f_hashLong)
{
    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
    /*
     * If an action is to be taken if `secretLen` condition is not respected,
     * it should be done here.
     * For now, it's a contract pre-condition.
     * Adding a check and a branch here would cost performance at every hash.
     * Also, note that function signature doesn't offer room to return an error.
     */
    if (len <= 16)
        return XXH3_len_0to16_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
    if (len <= 128)
        return XXH3_len_17to128_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
    if (len <= XXH3_MIDSIZE_MAX)
        return XXH3_len_129to240_64b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
    return f_hashLong(input, len, seed64, (const xxh_u8*)secret, secretLen);
}


/* ===   Public entry point   === */

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH64_hash_t XXH3_64bits(XXH_NOESCAPE const void* input, size_t length)
{
    return XXH3_64bits_internal(input, length, 0, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_default);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH64_hash_t
XXH3_64bits_withSecret(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize)
{
    return XXH3_64bits_internal(input, length, 0, secret, secretSize, XXH3_hashLong_64b_withSecret);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH64_hash_t
XXH3_64bits_withSeed(XXH_NOESCAPE const void* input, size_t length, XXH64_hash_t seed)
{
    return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), XXH3_hashLong_64b_withSeed);
}

XXH_PUBLIC_API XXH64_hash_t
XXH3_64bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t length, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
{
    if (length <= XXH3_MIDSIZE_MAX)
        return XXH3_64bits_internal(input, length, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
    return XXH3_hashLong_64b_withSecret(input, length, seed, (const xxh_u8*)secret, secretSize);
}


/* ===   XXH3 streaming   === */
#ifndef XXH_NO_STREAM
/*
 * Malloc's a pointer that is always aligned to align.
 *
 * This must be freed with `XXH_alignedFree()`.
 *
 * malloc typically guarantees 16 byte alignment on 64-bit systems and 8 byte
 * alignment on 32-bit. This isn't enough for the 32 byte aligned loads in AVX2
 * or on 32-bit, the 16 byte aligned loads in SSE2 and NEON.
 *
 * This underalignment previously caused a rather obvious crash which went
 * completely unnoticed due to XXH3_createState() not actually being tested.
 * Credit to RedSpah for noticing this bug.
 *
 * The alignment is done manually: Functions like posix_memalign or _mm_malloc
 * are avoided: To maintain portability, we would have to write a fallback
 * like this anyways, and besides, testing for the existence of library
 * functions without relying on external build tools is impossible.
 *
 * The method is simple: Overallocate, manually align, and store the offset
 * to the original behind the returned pointer.
 *
 * Align must be a power of 2 and 8 <= align <= 128.
 */
static XXH_MALLOCF void* XXH_alignedMalloc(size_t s, size_t align)
{
    XXH_ASSERT(align <= 128 && align >= 8); /* range check */
    XXH_ASSERT((align & (align-1)) == 0);   /* power of 2 */
    XXH_ASSERT(s != 0 && s < (s + align));  /* empty/overflow */
    {   /* Overallocate to make room for manual realignment and an offset byte */
        xxh_u8* base = (xxh_u8*)XXH_malloc(s + align);
        if (base != NULL) {
            /*
             * Get the offset needed to align this pointer.
             *
             * Even if the returned pointer is aligned, there will always be
             * at least one byte to store the offset to the original pointer.
             */
            size_t offset = align - ((size_t)base & (align - 1)); /* base % align */
            /* Add the offset for the now-aligned pointer */
            xxh_u8* ptr = base + offset;

            XXH_ASSERT((size_t)ptr % align == 0);

            /* Store the offset immediately before the returned pointer. */
            ptr[-1] = (xxh_u8)offset;
            return ptr;
        }
        return NULL;
    }
}
/*
 * Frees an aligned pointer allocated by XXH_alignedMalloc(). Don't pass
 * normal malloc'd pointers, XXH_alignedMalloc has a specific data layout.
 */
static void XXH_alignedFree(void* p)
{
    if (p != NULL) {
        xxh_u8* ptr = (xxh_u8*)p;
        /* Get the offset byte we added in XXH_malloc. */
        xxh_u8 offset = ptr[-1];
        /* Free the original malloc'd pointer */
        xxh_u8* base = ptr - offset;
        XXH_free(base);
    }
}
/*! @ingroup XXH3_family */
/*!
 * @brief Allocate an @ref XXH3_state_t.
 *
 * @return An allocated pointer of @ref XXH3_state_t on success.
 * @return `NULL` on failure.
 *
 * @note Must be freed with XXH3_freeState().
 */
XXH_PUBLIC_API XXH3_state_t* XXH3_createState(void)
{
    XXH3_state_t* const state = (XXH3_state_t*)XXH_alignedMalloc(sizeof(XXH3_state_t), 64);
    if (state==NULL) return NULL;
    XXH3_INITSTATE(state);
    return state;
}

/*! @ingroup XXH3_family */
/*!
 * @brief Frees an @ref XXH3_state_t.
 *
 * @param statePtr A pointer to an @ref XXH3_state_t allocated with @ref XXH3_createState().
 *
 * @return @ref XXH_OK.
 *
 * @note Must be allocated with XXH3_createState().
 */
XXH_PUBLIC_API XXH_errorcode XXH3_freeState(XXH3_state_t* statePtr)
{
    XXH_alignedFree(statePtr);
    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API void
XXH3_copyState(XXH_NOESCAPE XXH3_state_t* dst_state, XXH_NOESCAPE const XXH3_state_t* src_state)
{
    XXH_memcpy(dst_state, src_state, sizeof(*dst_state));
}

static void
XXH3_reset_internal(XXH3_state_t* statePtr,
                    XXH64_hash_t seed,
                    const void* secret, size_t secretSize)
{
    size_t const initStart = offsetof(XXH3_state_t, bufferedSize);
    size_t const initLength = offsetof(XXH3_state_t, nbStripesPerBlock) - initStart;
    XXH_ASSERT(offsetof(XXH3_state_t, nbStripesPerBlock) > initStart);
    XXH_ASSERT(statePtr != NULL);
    /* set members from bufferedSize to nbStripesPerBlock (excluded) to 0 */
    memset((char*)statePtr + initStart, 0, initLength);
    statePtr->acc[0] = XXH_PRIME32_3;
    statePtr->acc[1] = XXH_PRIME64_1;
    statePtr->acc[2] = XXH_PRIME64_2;
    statePtr->acc[3] = XXH_PRIME64_3;
    statePtr->acc[4] = XXH_PRIME64_4;
    statePtr->acc[5] = XXH_PRIME32_2;
    statePtr->acc[6] = XXH_PRIME64_5;
    statePtr->acc[7] = XXH_PRIME32_1;
    statePtr->seed = seed;
    statePtr->useSeed = (seed != 0);
    statePtr->extSecret = (const unsigned char*)secret;
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
    statePtr->secretLimit = secretSize - XXH_STRIPE_LEN;
    statePtr->nbStripesPerBlock = statePtr->secretLimit / XXH_SECRET_CONSUME_RATE;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
{
    if (statePtr == NULL) return XXH_ERROR;
    XXH3_reset_internal(statePtr, 0, XXH3_kSecret, XXH_SECRET_DEFAULT_SIZE);
    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
{
    if (statePtr == NULL) return XXH_ERROR;
    XXH3_reset_internal(statePtr, 0, secret, secretSize);
    if (secret == NULL) return XXH_ERROR;
    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
{
    if (statePtr == NULL) return XXH_ERROR;
    if (seed==0) return XXH3_64bits_reset(statePtr);
    if ((seed != statePtr->seed) || (statePtr->extSecret != NULL))
        XXH3_initCustomSecret(statePtr->customSecret, seed);
    XXH3_reset_internal(statePtr, seed, NULL, XXH_SECRET_DEFAULT_SIZE);
    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed64)
{
    if (statePtr == NULL) return XXH_ERROR;
    if (secret == NULL) return XXH_ERROR;
    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
    XXH3_reset_internal(statePtr, seed64, secret, secretSize);
    statePtr->useSeed = 1; /* always, even if seed64==0 */
    return XXH_OK;
}

/*!
 * @internal
 * @brief Processes a large input for XXH3_update() and XXH3_digest_long().
 *
 * Unlike XXH3_hashLong_internal_loop(), this can process data that overlaps a block.
 *
 * @param acc                Pointer to the 8 accumulator lanes
 * @param nbStripesSoFarPtr  In/out pointer to the number of leftover stripes in the block*
 * @param nbStripesPerBlock  Number of stripes in a block
 * @param input              Input pointer
 * @param nbStripes          Number of stripes to process
 * @param secret             Secret pointer
 * @param secretLimit        Offset of the last block in @p secret
 * @param f_acc              Pointer to an XXH3_accumulate implementation
 * @param f_scramble         Pointer to an XXH3_scrambleAcc implementation
 * @return                   Pointer past the end of @p input after processing
 */
XXH_FORCE_INLINE const xxh_u8 *
XXH3_consumeStripes(xxh_u64* XXH_RESTRICT acc,
                    size_t* XXH_RESTRICT nbStripesSoFarPtr, size_t nbStripesPerBlock,
                    const xxh_u8* XXH_RESTRICT input, size_t nbStripes,
                    const xxh_u8* XXH_RESTRICT secret, size_t secretLimit,
                    XXH3_f_accumulate f_acc,
                    XXH3_f_scrambleAcc f_scramble)
{
    const xxh_u8* initialSecret = secret + *nbStripesSoFarPtr * XXH_SECRET_CONSUME_RATE;
    /* Process full blocks */
    if (nbStripes >= (nbStripesPerBlock - *nbStripesSoFarPtr)) {
        /* Process the initial partial block... */
        size_t nbStripesThisIter = nbStripesPerBlock - *nbStripesSoFarPtr;

        do {
            /* Accumulate and scramble */
            f_acc(acc, input, initialSecret, nbStripesThisIter);
            f_scramble(acc, secret + secretLimit);
            input += nbStripesThisIter * XXH_STRIPE_LEN;
            nbStripes -= nbStripesThisIter;
            /* Then continue the loop with the full block size */
            nbStripesThisIter = nbStripesPerBlock;
            initialSecret = secret;
        } while (nbStripes >= nbStripesPerBlock);
        *nbStripesSoFarPtr = 0;
    }
    /* Process a partial block */
    if (nbStripes > 0) {
        f_acc(acc, input, initialSecret, nbStripes);
        input += nbStripes * XXH_STRIPE_LEN;
        *nbStripesSoFarPtr += nbStripes;
    }
    /* Return end pointer */
    return input;
}

#ifndef XXH3_STREAM_USE_STACK
# if XXH_SIZE_OPT <= 0 && !defined(__clang__) /* clang doesn't need additional stack space */
#   define XXH3_STREAM_USE_STACK 1
# endif
#endif
/*
 * Both XXH3_64bits_update and XXH3_128bits_update use this routine.
 */
XXH_FORCE_INLINE XXH_errorcode
XXH3_update(XXH3_state_t* XXH_RESTRICT const state,
            const xxh_u8* XXH_RESTRICT input, size_t len,
            XXH3_f_accumulate f_acc,
            XXH3_f_scrambleAcc f_scramble)
{
    if (input==NULL) {
        XXH_ASSERT(len == 0);
        return XXH_OK;
    }

    XXH_ASSERT(state != NULL);
    {   const xxh_u8* const bEnd = input + len;
        const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
        /* For some reason, gcc and MSVC seem to suffer greatly
         * when operating accumulators directly into state.
         * Operating into stack space seems to enable proper optimization.
         * clang, on the other hand, doesn't seem to need this trick */
        XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[8];
        XXH_memcpy(acc, state->acc, sizeof(acc));
#else
        xxh_u64* XXH_RESTRICT const acc = state->acc;
#endif
        state->totalLen += len;
        XXH_ASSERT(state->bufferedSize <= XXH3_INTERNALBUFFER_SIZE);

        /* small input : just fill in tmp buffer */
        if (len <= XXH3_INTERNALBUFFER_SIZE - state->bufferedSize) {
            XXH_memcpy(state->buffer + state->bufferedSize, input, len);
            state->bufferedSize += (XXH32_hash_t)len;
            return XXH_OK;
        }

        /* total input is now > XXH3_INTERNALBUFFER_SIZE */
        #define XXH3_INTERNALBUFFER_STRIPES (XXH3_INTERNALBUFFER_SIZE / XXH_STRIPE_LEN)
        XXH_STATIC_ASSERT(XXH3_INTERNALBUFFER_SIZE % XXH_STRIPE_LEN == 0);   /* clean multiple */

        /*
         * Internal buffer is partially filled (always, except at beginning)
         * Complete it, then consume it.
         */
        if (state->bufferedSize) {
            size_t const loadSize = XXH3_INTERNALBUFFER_SIZE - state->bufferedSize;
            XXH_memcpy(state->buffer + state->bufferedSize, input, loadSize);
            input += loadSize;
            XXH3_consumeStripes(acc,
                               &state->nbStripesSoFar, state->nbStripesPerBlock,
                                state->buffer, XXH3_INTERNALBUFFER_STRIPES,
                                secret, state->secretLimit,
                                f_acc, f_scramble);
            state->bufferedSize = 0;
        }
        XXH_ASSERT(input < bEnd);
        if (bEnd - input > XXH3_INTERNALBUFFER_SIZE) {
            size_t nbStripes = (size_t)(bEnd - 1 - input) / XXH_STRIPE_LEN;
            input = XXH3_consumeStripes(acc,
                                       &state->nbStripesSoFar, state->nbStripesPerBlock,
                                       input, nbStripes,
                                       secret, state->secretLimit,
                                       f_acc, f_scramble);
            XXH_memcpy(state->buffer + sizeof(state->buffer) - XXH_STRIPE_LEN, input - XXH_STRIPE_LEN, XXH_STRIPE_LEN);

        }
        /* Some remaining input (always) : buffer it */
        XXH_ASSERT(input < bEnd);
        XXH_ASSERT(bEnd - input <= XXH3_INTERNALBUFFER_SIZE);
        XXH_ASSERT(state->bufferedSize == 0);
        XXH_memcpy(state->buffer, input, (size_t)(bEnd-input));
        state->bufferedSize = (XXH32_hash_t)(bEnd-input);
#if defined(XXH3_STREAM_USE_STACK) && XXH3_STREAM_USE_STACK >= 1
        /* save stack accumulators into state */
        XXH_memcpy(state->acc, acc, sizeof(acc));
#endif
    }

    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_64bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
{
    return XXH3_update(state, (const xxh_u8*)input, len,
                       XXH3_accumulate, XXH3_scrambleAcc);
}


XXH_FORCE_INLINE void
XXH3_digest_long (XXH64_hash_t* acc,
                  const XXH3_state_t* state,
                  const unsigned char* secret)
{
    xxh_u8 lastStripe[XXH_STRIPE_LEN];
    const xxh_u8* lastStripePtr;

    /*
     * Digest on a local copy. This way, the state remains unaltered, and it can
     * continue ingesting more input afterwards.
     */
    XXH_memcpy(acc, state->acc, sizeof(state->acc));
    if (state->bufferedSize >= XXH_STRIPE_LEN) {
        /* Consume remaining stripes then point to remaining data in buffer */
        size_t const nbStripes = (state->bufferedSize - 1) / XXH_STRIPE_LEN;
        size_t nbStripesSoFar = state->nbStripesSoFar;
        XXH3_consumeStripes(acc,
                           &nbStripesSoFar, state->nbStripesPerBlock,
                            state->buffer, nbStripes,
                            secret, state->secretLimit,
                            XXH3_accumulate, XXH3_scrambleAcc);
        lastStripePtr = state->buffer + state->bufferedSize - XXH_STRIPE_LEN;
    } else {  /* bufferedSize < XXH_STRIPE_LEN */
        /* Copy to temp buffer */
        size_t const catchupSize = XXH_STRIPE_LEN - state->bufferedSize;
        XXH_ASSERT(state->bufferedSize > 0);  /* there is always some input buffered */
        XXH_memcpy(lastStripe, state->buffer + sizeof(state->buffer) - catchupSize, catchupSize);
        XXH_memcpy(lastStripe + catchupSize, state->buffer, state->bufferedSize);
        lastStripePtr = lastStripe;
    }
    /* Last stripe */
    XXH3_accumulate_512(acc,
                        lastStripePtr,
                        secret + state->secretLimit - XXH_SECRET_LASTACC_START);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH64_hash_t XXH3_64bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
{
    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
    if (state->totalLen > XXH3_MIDSIZE_MAX) {
        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
        XXH3_digest_long(acc, state, secret);
        return XXH3_mergeAccs(acc,
                              secret + XXH_SECRET_MERGEACCS_START,
                              (xxh_u64)state->totalLen * XXH_PRIME64_1);
    }
    /* totalLen <= XXH3_MIDSIZE_MAX: digesting a short input */
    if (state->useSeed)
        return XXH3_64bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
    return XXH3_64bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                  secret, state->secretLimit + XXH_STRIPE_LEN);
}
#endif /* !XXH_NO_STREAM */


/* ==========================================
 * XXH3 128 bits (a.k.a XXH128)
 * ==========================================
 * XXH3's 128-bit variant has better mixing and strength than the 64-bit variant,
 * even without counting the significantly larger output size.
 *
 * For example, extra steps are taken to avoid the seed-dependent collisions
 * in 17-240 byte inputs (See XXH3_mix16B and XXH128_mix32B).
 *
 * This strength naturally comes at the cost of some speed, especially on short
 * lengths. Note that longer hashes are about as fast as the 64-bit version
 * due to it using only a slight modification of the 64-bit loop.
 *
 * XXH128 is also more oriented towards 64-bit machines. It is still extremely
 * fast for a _128-bit_ hash on 32-bit (it usually clears XXH64).
 */

XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_1to3_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    /* A doubled version of 1to3_64b with different constants. */
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(1 <= len && len <= 3);
    XXH_ASSERT(secret != NULL);
    /*
     * len = 1: combinedl = { input[0], 0x01, input[0], input[0] }
     * len = 2: combinedl = { input[1], 0x02, input[0], input[1] }
     * len = 3: combinedl = { input[2], 0x03, input[0], input[1] }
     */
    {   xxh_u8 const c1 = input[0];
        xxh_u8 const c2 = input[len >> 1];
        xxh_u8 const c3 = input[len - 1];
        xxh_u32 const combinedl = ((xxh_u32)c1 <<16) | ((xxh_u32)c2 << 24)
                                | ((xxh_u32)c3 << 0) | ((xxh_u32)len << 8);
        xxh_u32 const combinedh = XXH_rotl32(XXH_swap32(combinedl), 13);
        xxh_u64 const bitflipl = (XXH_readLE32(secret) ^ XXH_readLE32(secret+4)) + seed;
        xxh_u64 const bitfliph = (XXH_readLE32(secret+8) ^ XXH_readLE32(secret+12)) - seed;
        xxh_u64 const keyed_lo = (xxh_u64)combinedl ^ bitflipl;
        xxh_u64 const keyed_hi = (xxh_u64)combinedh ^ bitfliph;
        XXH128_hash_t h128;
        h128.low64  = XXH64_avalanche(keyed_lo);
        h128.high64 = XXH64_avalanche(keyed_hi);
        return h128;
    }
}

XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_4to8_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(secret != NULL);
    XXH_ASSERT(4 <= len && len <= 8);
    seed ^= (xxh_u64)XXH_swap32((xxh_u32)seed) << 32;
    {   xxh_u32 const input_lo = XXH_readLE32(input);
        xxh_u32 const input_hi = XXH_readLE32(input + len - 4);
        xxh_u64 const input_64 = input_lo + ((xxh_u64)input_hi << 32);
        xxh_u64 const bitflip = (XXH_readLE64(secret+16) ^ XXH_readLE64(secret+24)) + seed;
        xxh_u64 const keyed = input_64 ^ bitflip;

        /* Shift len to the left to ensure it is even, this avoids even multiplies. */
        XXH128_hash_t m128 = XXH_mult64to128(keyed, XXH_PRIME64_1 + (len << 2));

        m128.high64 += (m128.low64 << 1);
        m128.low64  ^= (m128.high64 >> 3);

        m128.low64   = XXH_xorshift64(m128.low64, 35);
        m128.low64  *= PRIME_MX2;
        m128.low64   = XXH_xorshift64(m128.low64, 28);
        m128.high64  = XXH3_avalanche(m128.high64);
        return m128;
    }
}

XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_9to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(input != NULL);
    XXH_ASSERT(secret != NULL);
    XXH_ASSERT(9 <= len && len <= 16);
    {   xxh_u64 const bitflipl = (XXH_readLE64(secret+32) ^ XXH_readLE64(secret+40)) - seed;
        xxh_u64 const bitfliph = (XXH_readLE64(secret+48) ^ XXH_readLE64(secret+56)) + seed;
        xxh_u64 const input_lo = XXH_readLE64(input);
        xxh_u64       input_hi = XXH_readLE64(input + len - 8);
        XXH128_hash_t m128 = XXH_mult64to128(input_lo ^ input_hi ^ bitflipl, XXH_PRIME64_1);
        /*
         * Put len in the middle of m128 to ensure that the length gets mixed to
         * both the low and high bits in the 128x64 multiply below.
         */
        m128.low64 += (xxh_u64)(len - 1) << 54;
        input_hi   ^= bitfliph;
        /*
         * Add the high 32 bits of input_hi to the high 32 bits of m128, then
         * add the long product of the low 32 bits of input_hi and XXH_PRIME32_2 to
         * the high 64 bits of m128.
         *
         * The best approach to this operation is different on 32-bit and 64-bit.
         */
        if (sizeof(void *) < sizeof(xxh_u64)) { /* 32-bit */
            /*
             * 32-bit optimized version, which is more readable.
             *
             * On 32-bit, it removes an ADC and delays a dependency between the two
             * halves of m128.high64, but it generates an extra mask on 64-bit.
             */
            m128.high64 += (input_hi & 0xFFFFFFFF00000000ULL) + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2);
        } else {
            /*
             * 64-bit optimized (albeit more confusing) version.
             *
             * Uses some properties of addition and multiplication to remove the mask:
             *
             * Let:
             *    a = input_hi.lo = (input_hi & 0x00000000FFFFFFFF)
             *    b = input_hi.hi = (input_hi & 0xFFFFFFFF00000000)
             *    c = XXH_PRIME32_2
             *
             *    a + (b * c)
             * Inverse Property: x + y - x == y
             *    a + (b * (1 + c - 1))
             * Distributive Property: x * (y + z) == (x * y) + (x * z)
             *    a + (b * 1) + (b * (c - 1))
             * Identity Property: x * 1 == x
             *    a + b + (b * (c - 1))
             *
             * Substitute a, b, and c:
             *    input_hi.hi + input_hi.lo + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
             *
             * Since input_hi.hi + input_hi.lo == input_hi, we get this:
             *    input_hi + ((xxh_u64)input_hi.lo * (XXH_PRIME32_2 - 1))
             */
            m128.high64 += input_hi + XXH_mult32to64((xxh_u32)input_hi, XXH_PRIME32_2 - 1);
        }
        /* m128 ^= XXH_swap64(m128 >> 64); */
        m128.low64  ^= XXH_swap64(m128.high64);

        {   /* 128x64 multiply: h128 = m128 * XXH_PRIME64_2; */
            XXH128_hash_t h128 = XXH_mult64to128(m128.low64, XXH_PRIME64_2);
            h128.high64 += m128.high64 * XXH_PRIME64_2;

            h128.low64   = XXH3_avalanche(h128.low64);
            h128.high64  = XXH3_avalanche(h128.high64);
            return h128;
    }   }
}

/*
 * Assumption: `secret` size is >= XXH3_SECRET_SIZE_MIN
 */
XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_0to16_128b(const xxh_u8* input, size_t len, const xxh_u8* secret, XXH64_hash_t seed)
{
    XXH_ASSERT(len <= 16);
    {   if (len > 8) return XXH3_len_9to16_128b(input, len, secret, seed);
        if (len >= 4) return XXH3_len_4to8_128b(input, len, secret, seed);
        if (len) return XXH3_len_1to3_128b(input, len, secret, seed);
        {   XXH128_hash_t h128;
            xxh_u64 const bitflipl = XXH_readLE64(secret+64) ^ XXH_readLE64(secret+72);
            xxh_u64 const bitfliph = XXH_readLE64(secret+80) ^ XXH_readLE64(secret+88);
            h128.low64 = XXH64_avalanche(seed ^ bitflipl);
            h128.high64 = XXH64_avalanche( seed ^ bitfliph);
            return h128;
    }   }
}

/*
 * A bit slower than XXH3_mix16B, but handles multiply by zero better.
 */
XXH_FORCE_INLINE XXH128_hash_t
XXH128_mix32B(XXH128_hash_t acc, const xxh_u8* input_1, const xxh_u8* input_2,
              const xxh_u8* secret, XXH64_hash_t seed)
{
    acc.low64  += XXH3_mix16B (input_1, secret+0, seed);
    acc.low64  ^= XXH_readLE64(input_2) + XXH_readLE64(input_2 + 8);
    acc.high64 += XXH3_mix16B (input_2, secret+16, seed);
    acc.high64 ^= XXH_readLE64(input_1) + XXH_readLE64(input_1 + 8);
    return acc;
}


XXH_FORCE_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_17to128_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                      const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                      XXH64_hash_t seed)
{
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
    XXH_ASSERT(16 < len && len <= 128);

    {   XXH128_hash_t acc;
        acc.low64 = len * XXH_PRIME64_1;
        acc.high64 = 0;

#if XXH_SIZE_OPT >= 1
        {
            /* Smaller, but slightly slower. */
            unsigned int i = (unsigned int)(len - 1) / 32;
            do {
                acc = XXH128_mix32B(acc, input+16*i, input+len-16*(i+1), secret+32*i, seed);
            } while (i-- != 0);
        }
#else
        if (len > 32) {
            if (len > 64) {
                if (len > 96) {
                    acc = XXH128_mix32B(acc, input+48, input+len-64, secret+96, seed);
                }
                acc = XXH128_mix32B(acc, input+32, input+len-48, secret+64, seed);
            }
            acc = XXH128_mix32B(acc, input+16, input+len-32, secret+32, seed);
        }
        acc = XXH128_mix32B(acc, input, input+len-16, secret, seed);
#endif
        {   XXH128_hash_t h128;
            h128.low64  = acc.low64 + acc.high64;
            h128.high64 = (acc.low64    * XXH_PRIME64_1)
                        + (acc.high64   * XXH_PRIME64_4)
                        + ((len - seed) * XXH_PRIME64_2);
            h128.low64  = XXH3_avalanche(h128.low64);
            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
            return h128;
        }
    }
}

XXH_NO_INLINE XXH_PUREF XXH128_hash_t
XXH3_len_129to240_128b(const xxh_u8* XXH_RESTRICT input, size_t len,
                       const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                       XXH64_hash_t seed)
{
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN); (void)secretSize;
    XXH_ASSERT(128 < len && len <= XXH3_MIDSIZE_MAX);

    {   XXH128_hash_t acc;
        unsigned i;
        acc.low64 = len * XXH_PRIME64_1;
        acc.high64 = 0;
        /*
         *  We set as `i` as offset + 32. We do this so that unchanged
         * `len` can be used as upper bound. This reaches a sweet spot
         * where both x86 and aarch64 get simple agen and good codegen
         * for the loop.
         */
        for (i = 32; i < 160; i += 32) {
            acc = XXH128_mix32B(acc,
                                input  + i - 32,
                                input  + i - 16,
                                secret + i - 32,
                                seed);
        }
        acc.low64 = XXH3_avalanche(acc.low64);
        acc.high64 = XXH3_avalanche(acc.high64);
        /*
         * NB: `i <= len` will duplicate the last 32-bytes if
         * len % 32 was zero. This is an unfortunate necessity to keep
         * the hash result stable.
         */
        for (i=160; i <= len; i += 32) {
            acc = XXH128_mix32B(acc,
                                input + i - 32,
                                input + i - 16,
                                secret + XXH3_MIDSIZE_STARTOFFSET + i - 160,
                                seed);
        }
        /* last bytes */
        acc = XXH128_mix32B(acc,
                            input + len - 16,
                            input + len - 32,
                            secret + XXH3_SECRET_SIZE_MIN - XXH3_MIDSIZE_LASTOFFSET - 16,
                            (XXH64_hash_t)0 - seed);

        {   XXH128_hash_t h128;
            h128.low64  = acc.low64 + acc.high64;
            h128.high64 = (acc.low64    * XXH_PRIME64_1)
                        + (acc.high64   * XXH_PRIME64_4)
                        + ((len - seed) * XXH_PRIME64_2);
            h128.low64  = XXH3_avalanche(h128.low64);
            h128.high64 = (XXH64_hash_t)0 - XXH3_avalanche(h128.high64);
            return h128;
        }
    }
}

XXH_FORCE_INLINE XXH128_hash_t
XXH3_hashLong_128b_internal(const void* XXH_RESTRICT input, size_t len,
                            const xxh_u8* XXH_RESTRICT secret, size_t secretSize,
                            XXH3_f_accumulate f_acc,
                            XXH3_f_scrambleAcc f_scramble)
{
    XXH_ALIGN(XXH_ACC_ALIGN) xxh_u64 acc[XXH_ACC_NB] = XXH3_INIT_ACC;

    XXH3_hashLong_internal_loop(acc, (const xxh_u8*)input, len, secret, secretSize, f_acc, f_scramble);

    /* converge into final hash */
    XXH_STATIC_ASSERT(sizeof(acc) == 64);
    XXH_ASSERT(secretSize >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
    {   XXH128_hash_t h128;
        h128.low64  = XXH3_mergeAccs(acc,
                                     secret + XXH_SECRET_MERGEACCS_START,
                                     (xxh_u64)len * XXH_PRIME64_1);
        h128.high64 = XXH3_mergeAccs(acc,
                                     secret + secretSize
                                            - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
                                     ~((xxh_u64)len * XXH_PRIME64_2));
        return h128;
    }
}

/*
 * It's important for performance that XXH3_hashLong() is not inlined.
 */
XXH_NO_INLINE XXH_PUREF XXH128_hash_t
XXH3_hashLong_128b_default(const void* XXH_RESTRICT input, size_t len,
                           XXH64_hash_t seed64,
                           const void* XXH_RESTRICT secret, size_t secretLen)
{
    (void)seed64; (void)secret; (void)secretLen;
    return XXH3_hashLong_128b_internal(input, len, XXH3_kSecret, sizeof(XXH3_kSecret),
                                       XXH3_accumulate, XXH3_scrambleAcc);
}

/*
 * It's important for performance to pass @p secretLen (when it's static)
 * to the compiler, so that it can properly optimize the vectorized loop.
 *
 * When the secret size is unknown, or on GCC 12 where the mix of NO_INLINE and FORCE_INLINE
 * breaks -Og, this is XXH_NO_INLINE.
 */
XXH3_WITH_SECRET_INLINE XXH128_hash_t
XXH3_hashLong_128b_withSecret(const void* XXH_RESTRICT input, size_t len,
                              XXH64_hash_t seed64,
                              const void* XXH_RESTRICT secret, size_t secretLen)
{
    (void)seed64;
    return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, secretLen,
                                       XXH3_accumulate, XXH3_scrambleAcc);
}

XXH_FORCE_INLINE XXH128_hash_t
XXH3_hashLong_128b_withSeed_internal(const void* XXH_RESTRICT input, size_t len,
                                XXH64_hash_t seed64,
                                XXH3_f_accumulate f_acc,
                                XXH3_f_scrambleAcc f_scramble,
                                XXH3_f_initCustomSecret f_initSec)
{
    if (seed64 == 0)
        return XXH3_hashLong_128b_internal(input, len,
                                           XXH3_kSecret, sizeof(XXH3_kSecret),
                                           f_acc, f_scramble);
    {   XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
        f_initSec(secret, seed64);
        return XXH3_hashLong_128b_internal(input, len, (const xxh_u8*)secret, sizeof(secret),
                                           f_acc, f_scramble);
    }
}

/*
 * It's important for performance that XXH3_hashLong is not inlined.
 */
XXH_NO_INLINE XXH128_hash_t
XXH3_hashLong_128b_withSeed(const void* input, size_t len,
                            XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen)
{
    (void)secret; (void)secretLen;
    return XXH3_hashLong_128b_withSeed_internal(input, len, seed64,
                XXH3_accumulate, XXH3_scrambleAcc, XXH3_initCustomSecret);
}

typedef XXH128_hash_t (*XXH3_hashLong128_f)(const void* XXH_RESTRICT, size_t,
                                            XXH64_hash_t, const void* XXH_RESTRICT, size_t);

XXH_FORCE_INLINE XXH128_hash_t
XXH3_128bits_internal(const void* input, size_t len,
                      XXH64_hash_t seed64, const void* XXH_RESTRICT secret, size_t secretLen,
                      XXH3_hashLong128_f f_hl128)
{
    XXH_ASSERT(secretLen >= XXH3_SECRET_SIZE_MIN);
    /*
     * If an action is to be taken if `secret` conditions are not respected,
     * it should be done here.
     * For now, it's a contract pre-condition.
     * Adding a check and a branch here would cost performance at every hash.
     */
    if (len <= 16)
        return XXH3_len_0to16_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, seed64);
    if (len <= 128)
        return XXH3_len_17to128_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
    if (len <= XXH3_MIDSIZE_MAX)
        return XXH3_len_129to240_128b((const xxh_u8*)input, len, (const xxh_u8*)secret, secretLen, seed64);
    return f_hl128(input, len, seed64, secret, secretLen);
}


/* ===   Public XXH128 API   === */

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t XXH3_128bits(XXH_NOESCAPE const void* input, size_t len)
{
    return XXH3_128bits_internal(input, len, 0,
                                 XXH3_kSecret, sizeof(XXH3_kSecret),
                                 XXH3_hashLong_128b_default);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t
XXH3_128bits_withSecret(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize)
{
    return XXH3_128bits_internal(input, len, 0,
                                 (const xxh_u8*)secret, secretSize,
                                 XXH3_hashLong_128b_withSecret);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t
XXH3_128bits_withSeed(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
{
    return XXH3_128bits_internal(input, len, seed,
                                 XXH3_kSecret, sizeof(XXH3_kSecret),
                                 XXH3_hashLong_128b_withSeed);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t
XXH3_128bits_withSecretandSeed(XXH_NOESCAPE const void* input, size_t len, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
{
    if (len <= XXH3_MIDSIZE_MAX)
        return XXH3_128bits_internal(input, len, seed, XXH3_kSecret, sizeof(XXH3_kSecret), NULL);
    return XXH3_hashLong_128b_withSecret(input, len, seed, secret, secretSize);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t
XXH128(XXH_NOESCAPE const void* input, size_t len, XXH64_hash_t seed)
{
    return XXH3_128bits_withSeed(input, len, seed);
}


/* ===   XXH3 128-bit streaming   === */
#ifndef XXH_NO_STREAM
/*
 * All initialization and update functions are identical to 64-bit streaming variant.
 * The only difference is the finalization routine.
 */

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset(XXH_NOESCAPE XXH3_state_t* statePtr)
{
    return XXH3_64bits_reset(statePtr);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset_withSecret(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize)
{
    return XXH3_64bits_reset_withSecret(statePtr, secret, secretSize);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset_withSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH64_hash_t seed)
{
    return XXH3_64bits_reset_withSeed(statePtr, seed);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_reset_withSecretandSeed(XXH_NOESCAPE XXH3_state_t* statePtr, XXH_NOESCAPE const void* secret, size_t secretSize, XXH64_hash_t seed)
{
    return XXH3_64bits_reset_withSecretandSeed(statePtr, secret, secretSize, seed);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_128bits_update(XXH_NOESCAPE XXH3_state_t* state, XXH_NOESCAPE const void* input, size_t len)
{
    return XXH3_64bits_update(state, input, len);
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t XXH3_128bits_digest (XXH_NOESCAPE const XXH3_state_t* state)
{
    const unsigned char* const secret = (state->extSecret == NULL) ? state->customSecret : state->extSecret;
    if (state->totalLen > XXH3_MIDSIZE_MAX) {
        XXH_ALIGN(XXH_ACC_ALIGN) XXH64_hash_t acc[XXH_ACC_NB];
        XXH3_digest_long(acc, state, secret);
        XXH_ASSERT(state->secretLimit + XXH_STRIPE_LEN >= sizeof(acc) + XXH_SECRET_MERGEACCS_START);
        {   XXH128_hash_t h128;
            h128.low64  = XXH3_mergeAccs(acc,
                                         secret + XXH_SECRET_MERGEACCS_START,
                                         (xxh_u64)state->totalLen * XXH_PRIME64_1);
            h128.high64 = XXH3_mergeAccs(acc,
                                         secret + state->secretLimit + XXH_STRIPE_LEN
                                                - sizeof(acc) - XXH_SECRET_MERGEACCS_START,
                                         ~((xxh_u64)state->totalLen * XXH_PRIME64_2));
            return h128;
        }
    }
    /* len <= XXH3_MIDSIZE_MAX : short code */
    if (state->seed)
        return XXH3_128bits_withSeed(state->buffer, (size_t)state->totalLen, state->seed);
    return XXH3_128bits_withSecret(state->buffer, (size_t)(state->totalLen),
                                   secret, state->secretLimit + XXH_STRIPE_LEN);
}
#endif /* !XXH_NO_STREAM */
/* 128-bit utility functions */

/* return : 1 is equal, 0 if different */
/*! @ingroup XXH3_family */
XXH_PUBLIC_API int XXH128_isEqual(XXH128_hash_t h1, XXH128_hash_t h2)
{
    /* note : XXH128_hash_t is compact, it has no padding byte */
    return !(memcmp(&h1, &h2, sizeof(h1)));
}

/* This prototype is compatible with stdlib's qsort().
 * @return : >0 if *h128_1  > *h128_2
 *           <0 if *h128_1  < *h128_2
 *           =0 if *h128_1 == *h128_2  */
/*! @ingroup XXH3_family */
XXH_PUBLIC_API int XXH128_cmp(XXH_NOESCAPE const void* h128_1, XXH_NOESCAPE const void* h128_2)
{
    XXH128_hash_t const h1 = *(const XXH128_hash_t*)h128_1;
    XXH128_hash_t const h2 = *(const XXH128_hash_t*)h128_2;
    int const hcmp = (h1.high64 > h2.high64) - (h2.high64 > h1.high64);
    /* note : bets that, in most cases, hash values are different */
    if (hcmp) return hcmp;
    return (h1.low64 > h2.low64) - (h2.low64 > h1.low64);
}


/*======   Canonical representation   ======*/
/*! @ingroup XXH3_family */
XXH_PUBLIC_API void
XXH128_canonicalFromHash(XXH_NOESCAPE XXH128_canonical_t* dst, XXH128_hash_t hash)
{
    XXH_STATIC_ASSERT(sizeof(XXH128_canonical_t) == sizeof(XXH128_hash_t));
    if (XXH_CPU_LITTLE_ENDIAN) {
        hash.high64 = XXH_swap64(hash.high64);
        hash.low64  = XXH_swap64(hash.low64);
    }
    XXH_memcpy(dst, &hash.high64, sizeof(hash.high64));
    XXH_memcpy((char*)dst + sizeof(hash.high64), &hash.low64, sizeof(hash.low64));
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH128_hash_t
XXH128_hashFromCanonical(XXH_NOESCAPE const XXH128_canonical_t* src)
{
    XXH128_hash_t h;
    h.high64 = XXH_readBE64(src);
    h.low64  = XXH_readBE64(src->digest + 8);
    return h;
}


/* ==========================================
 * Secret generators
 * ==========================================
 */
#define XXH_MIN(x, y) (((x) > (y)) ? (y) : (x))

XXH_FORCE_INLINE void XXH3_combine16(void* dst, XXH128_hash_t h128)
{
    XXH_writeLE64( dst, XXH_readLE64(dst) ^ h128.low64 );
    XXH_writeLE64( (char*)dst+8, XXH_readLE64((char*)dst+8) ^ h128.high64 );
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API XXH_errorcode
XXH3_generateSecret(XXH_NOESCAPE void* secretBuffer, size_t secretSize, XXH_NOESCAPE const void* customSeed, size_t customSeedSize)
{
#if (XXH_DEBUGLEVEL >= 1)
    XXH_ASSERT(secretBuffer != NULL);
    XXH_ASSERT(secretSize >= XXH3_SECRET_SIZE_MIN);
#else
    /* production mode, assert() are disabled */
    if (secretBuffer == NULL) return XXH_ERROR;
    if (secretSize < XXH3_SECRET_SIZE_MIN) return XXH_ERROR;
#endif

    if (customSeedSize == 0) {
        customSeed = XXH3_kSecret;
        customSeedSize = XXH_SECRET_DEFAULT_SIZE;
    }
#if (XXH_DEBUGLEVEL >= 1)
    XXH_ASSERT(customSeed != NULL);
#else
    if (customSeed == NULL) return XXH_ERROR;
#endif

    /* Fill secretBuffer with a copy of customSeed - repeat as needed */
    {   size_t pos = 0;
        while (pos < secretSize) {
            size_t const toCopy = XXH_MIN((secretSize - pos), customSeedSize);
            memcpy((char*)secretBuffer + pos, customSeed, toCopy);
            pos += toCopy;
    }   }

    {   size_t const nbSeg16 = secretSize / 16;
        size_t n;
        XXH128_canonical_t scrambler;
        XXH128_canonicalFromHash(&scrambler, XXH128(customSeed, customSeedSize, 0));
        for (n=0; n<nbSeg16; n++) {
            XXH128_hash_t const h128 = XXH128(&scrambler, sizeof(scrambler), n);
            XXH3_combine16((char*)secretBuffer + n*16, h128);
        }
        /* last segment */
        XXH3_combine16((char*)secretBuffer + secretSize - 16, XXH128_hashFromCanonical(&scrambler));
    }
    return XXH_OK;
}

/*! @ingroup XXH3_family */
XXH_PUBLIC_API void
XXH3_generateSecret_fromSeed(XXH_NOESCAPE void* secretBuffer, XXH64_hash_t seed)
{
    XXH_ALIGN(XXH_SEC_ALIGN) xxh_u8 secret[XXH_SECRET_DEFAULT_SIZE];
    XXH3_initCustomSecret(secret, seed);
    XXH_ASSERT(secretBuffer != NULL);
    memcpy(secretBuffer, secret, XXH_SECRET_DEFAULT_SIZE);
}


/* Pop our optimization override from above */
#if XXH_VECTOR == XXH_AVX2 /* AVX2 */ \
  && defined(__GNUC__) && !defined(__clang__) /* GCC, not Clang */ \
  && defined(__OPTIMIZE__) && XXH_SIZE_OPT <= 0 /* respect -O0 and -Os */
#  pragma GCC pop_options
#endif


#if defined (__cplusplus)
} /* extern "C" */
#endif

#endif  /* XXH_NO_LONG_LONG */
#endif  /* XXH_NO_XXH3 */

/*!
 * @}
 */
#endif  /* XXH_IMPLEMENTATION */
/**** ended inlining xxhash.h ****/
#ifndef ZSTD_NO_TRACE
/**** start inlining zstd_trace.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_TRACE_H
#define ZSTD_TRACE_H

#include <stddef.h>

/* weak symbol support
 * For now, enable conservatively:
 * - Only GNUC
 * - Only ELF
 * - Only x86-64, i386, aarch64 and risc-v.
 * Also, explicitly disable on platforms known not to work so they aren't
 * forgotten in the future.
 */
#if !defined(ZSTD_HAVE_WEAK_SYMBOLS) && \
    defined(__GNUC__) && defined(__ELF__) && \
    (defined(__x86_64__) || defined(_M_X64) || defined(__i386__) || \
     defined(_M_IX86) || defined(__aarch64__) || defined(__riscv)) && \
    !defined(__APPLE__) && !defined(_WIN32) && !defined(__MINGW32__) && \
    !defined(__CYGWIN__) && !defined(_AIX)
#  define ZSTD_HAVE_WEAK_SYMBOLS 1
#else
#  define ZSTD_HAVE_WEAK_SYMBOLS 0
#endif
#if ZSTD_HAVE_WEAK_SYMBOLS
#  define ZSTD_WEAK_ATTR __attribute__((__weak__))
#else
#  define ZSTD_WEAK_ATTR
#endif

/* Only enable tracing when weak symbols are available. */
#ifndef ZSTD_TRACE
#  define ZSTD_TRACE ZSTD_HAVE_WEAK_SYMBOLS
#endif

#if ZSTD_TRACE

struct ZSTD_CCtx_s;
struct ZSTD_DCtx_s;
struct ZSTD_CCtx_params_s;

typedef struct {
    /**
     * ZSTD_VERSION_NUMBER
     *
     * This is guaranteed to be the first member of ZSTD_trace.
     * Otherwise, this struct is not stable between versions. If
     * the version number does not match your expectation, you
     * should not interpret the rest of the struct.
     */
    unsigned version;
    /**
     * Non-zero if streaming (de)compression is used.
     */
    int streaming;
    /**
     * The dictionary ID.
     */
    unsigned dictionaryID;
    /**
     * Is the dictionary cold?
     * Only set on decompression.
     */
    int dictionaryIsCold;
    /**
     * The dictionary size or zero if no dictionary.
     */
    size_t dictionarySize;
    /**
     * The uncompressed size of the data.
     */
    size_t uncompressedSize;
    /**
     * The compressed size of the data.
     */
    size_t compressedSize;
    /**
     * The fully resolved CCtx parameters (NULL on decompression).
     */
    struct ZSTD_CCtx_params_s const* params;
    /**
     * The ZSTD_CCtx pointer (NULL on decompression).
     */
    struct ZSTD_CCtx_s const* cctx;
    /**
     * The ZSTD_DCtx pointer (NULL on compression).
     */
    struct ZSTD_DCtx_s const* dctx;
} ZSTD_Trace;

/**
 * A tracing context. It must be 0 when tracing is disabled.
 * Otherwise, any non-zero value returned by a tracing begin()
 * function is presented to any subsequent calls to end().
 *
 * Any non-zero value is treated as tracing is enabled and not
 * interpreted by the library.
 *
 * Two possible uses are:
 * * A timestamp for when the begin() function was called.
 * * A unique key identifying the (de)compression, like the
 *   address of the [dc]ctx pointer if you need to track
 *   more information than just a timestamp.
 */
typedef unsigned long long ZSTD_TraceCtx;

/**
 * Trace the beginning of a compression call.
 * @param cctx The dctx pointer for the compression.
 *             It can be used as a key to map begin() to end().
 * @returns Non-zero if tracing is enabled. The return value is
 *          passed to ZSTD_trace_compress_end().
 */
ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_compress_begin(
    struct ZSTD_CCtx_s const* cctx);

/**
 * Trace the end of a compression call.
 * @param ctx The return value of ZSTD_trace_compress_begin().
 * @param trace The zstd tracing info.
 */
ZSTD_WEAK_ATTR void ZSTD_trace_compress_end(
    ZSTD_TraceCtx ctx,
    ZSTD_Trace const* trace);

/**
 * Trace the beginning of a decompression call.
 * @param dctx The dctx pointer for the decompression.
 *             It can be used as a key to map begin() to end().
 * @returns Non-zero if tracing is enabled. The return value is
 *          passed to ZSTD_trace_compress_end().
 */
ZSTD_WEAK_ATTR ZSTD_TraceCtx ZSTD_trace_decompress_begin(
    struct ZSTD_DCtx_s const* dctx);

/**
 * Trace the end of a decompression call.
 * @param ctx The return value of ZSTD_trace_decompress_begin().
 * @param trace The zstd tracing info.
 */
ZSTD_WEAK_ATTR void ZSTD_trace_decompress_end(
    ZSTD_TraceCtx ctx,
    ZSTD_Trace const* trace);

#endif /* ZSTD_TRACE */

#endif /* ZSTD_TRACE_H */
/**** ended inlining zstd_trace.h ****/
#else
#  define ZSTD_TRACE 0
#endif

/* ---- static assert (debug) --- */
#define ZSTD_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)
#define ZSTD_isError ERR_isError   /* for inlining */
#define FSE_isError  ERR_isError
#define HUF_isError  ERR_isError


/*-*************************************
*  shared macros
***************************************/
#undef MIN
#undef MAX
#define MIN(a,b) ((a)<(b) ? (a) : (b))
#define MAX(a,b) ((a)>(b) ? (a) : (b))
#define BOUNDED(min,val,max) (MAX(min,MIN(val,max)))


/*-*************************************
*  Common constants
***************************************/
#define ZSTD_OPT_NUM    (1<<12)

#define ZSTD_REP_NUM      3                 /* number of repcodes */
static UNUSED_ATTR const U32 repStartValue[ZSTD_REP_NUM] = { 1, 4, 8 };

#define KB *(1 <<10)
#define MB *(1 <<20)
#define GB *(1U<<30)

#define BIT7 128
#define BIT6  64
#define BIT5  32
#define BIT4  16
#define BIT1   2
#define BIT0   1

#define ZSTD_WINDOWLOG_ABSOLUTEMIN 10
static UNUSED_ATTR const size_t ZSTD_fcs_fieldSize[4] = { 0, 2, 4, 8 };
static UNUSED_ATTR const size_t ZSTD_did_fieldSize[4] = { 0, 1, 2, 4 };

#define ZSTD_FRAMEIDSIZE 4   /* magic number size */

#define ZSTD_BLOCKHEADERSIZE 3   /* C standard doesn't allow `static const` variable to be init using another `static const` variable */
static UNUSED_ATTR const size_t ZSTD_blockHeaderSize = ZSTD_BLOCKHEADERSIZE;
typedef enum { bt_raw, bt_rle, bt_compressed, bt_reserved } blockType_e;

#define ZSTD_FRAMECHECKSUMSIZE 4

#define MIN_SEQUENCES_SIZE 1 /* nbSeq==0 */
#define MIN_CBLOCK_SIZE (1 /*litCSize*/ + 1 /* RLE or RAW */)   /* for a non-null block */
#define MIN_LITERALS_FOR_4_STREAMS 6

typedef enum { set_basic, set_rle, set_compressed, set_repeat } SymbolEncodingType_e;

#define LONGNBSEQ 0x7F00

#define MINMATCH 3

#define Litbits  8
#define LitHufLog 11
#define MaxLit ((1<<Litbits) - 1)
#define MaxML   52
#define MaxLL   35
#define DefaultMaxOff 28
#define MaxOff  31
#define MaxSeq MAX(MaxLL, MaxML)   /* Assumption : MaxOff < MaxLL,MaxML */
#define MLFSELog    9
#define LLFSELog    9
#define OffFSELog   8
#define MaxFSELog  MAX(MAX(MLFSELog, LLFSELog), OffFSELog)
#define MaxMLBits 16
#define MaxLLBits 16

#define ZSTD_MAX_HUF_HEADER_SIZE 128 /* header + <= 127 byte tree description */
/* Each table cannot take more than #symbols * FSELog bits */
#define ZSTD_MAX_FSE_HEADERS_SIZE (((MaxML + 1) * MLFSELog + (MaxLL + 1) * LLFSELog + (MaxOff + 1) * OffFSELog + 7) / 8)

static UNUSED_ATTR const U8 LL_bits[MaxLL+1] = {
     0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0,
     1, 1, 1, 1, 2, 2, 3, 3,
     4, 6, 7, 8, 9,10,11,12,
    13,14,15,16
};
static UNUSED_ATTR const S16 LL_defaultNorm[MaxLL+1] = {
     4, 3, 2, 2, 2, 2, 2, 2,
     2, 2, 2, 2, 2, 1, 1, 1,
     2, 2, 2, 2, 2, 2, 2, 2,
     2, 3, 2, 1, 1, 1, 1, 1,
    -1,-1,-1,-1
};
#define LL_DEFAULTNORMLOG 6  /* for static allocation */
static UNUSED_ATTR const U32 LL_defaultNormLog = LL_DEFAULTNORMLOG;

static UNUSED_ATTR const U8 ML_bits[MaxML+1] = {
     0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0,
     0, 0, 0, 0, 0, 0, 0, 0,
     1, 1, 1, 1, 2, 2, 3, 3,
     4, 4, 5, 7, 8, 9,10,11,
    12,13,14,15,16
};
static UNUSED_ATTR const S16 ML_defaultNorm[MaxML+1] = {
     1, 4, 3, 2, 2, 2, 2, 2,
     2, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1,-1,-1,
    -1,-1,-1,-1,-1
};
#define ML_DEFAULTNORMLOG 6  /* for static allocation */
static UNUSED_ATTR const U32 ML_defaultNormLog = ML_DEFAULTNORMLOG;

static UNUSED_ATTR const S16 OF_defaultNorm[DefaultMaxOff+1] = {
     1, 1, 1, 1, 1, 1, 2, 2,
     2, 1, 1, 1, 1, 1, 1, 1,
     1, 1, 1, 1, 1, 1, 1, 1,
    -1,-1,-1,-1,-1
};
#define OF_DEFAULTNORMLOG 5  /* for static allocation */
static UNUSED_ATTR const U32 OF_defaultNormLog = OF_DEFAULTNORMLOG;


/*-*******************************************
*  Shared functions to include for inlining
*********************************************/
static void ZSTD_copy8(void* dst, const void* src) {
#if defined(ZSTD_ARCH_ARM_NEON)
    vst1_u8((uint8_t*)dst, vld1_u8((const uint8_t*)src));
#else
    ZSTD_memcpy(dst, src, 8);
#endif
}
#define COPY8(d,s) do { ZSTD_copy8(d,s); d+=8; s+=8; } while (0)

/* Need to use memmove here since the literal buffer can now be located within
   the dst buffer. In circumstances where the op "catches up" to where the
   literal buffer is, there can be partial overlaps in this call on the final
   copy if the literal is being shifted by less than 16 bytes. */
static void ZSTD_copy16(void* dst, const void* src) {
#if defined(ZSTD_ARCH_ARM_NEON)
    vst1q_u8((uint8_t*)dst, vld1q_u8((const uint8_t*)src));
#elif defined(ZSTD_ARCH_X86_SSE2)
    _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));
#elif defined(__clang__)
    ZSTD_memmove(dst, src, 16);
#else
    /* ZSTD_memmove is not inlined properly by gcc */
    BYTE copy16_buf[16];
    ZSTD_memcpy(copy16_buf, src, 16);
    ZSTD_memcpy(dst, copy16_buf, 16);
#endif
}
#define COPY16(d,s) do { ZSTD_copy16(d,s); d+=16; s+=16; } while (0)

#define WILDCOPY_OVERLENGTH 32
#define WILDCOPY_VECLEN 16

typedef enum {
    ZSTD_no_overlap,
    ZSTD_overlap_src_before_dst
    /*  ZSTD_overlap_dst_before_src, */
} ZSTD_overlap_e;

/*! ZSTD_wildcopy() :
 *  Custom version of ZSTD_memcpy(), can over read/write up to WILDCOPY_OVERLENGTH bytes (if length==0)
 *  @param ovtype controls the overlap detection
 *         - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
 *         - ZSTD_overlap_src_before_dst: The src and dst may overlap, but they MUST be at least 8 bytes apart.
 *           The src buffer must be before the dst buffer.
 */
MEM_STATIC FORCE_INLINE_ATTR
void ZSTD_wildcopy(void* dst, const void* src, ptrdiff_t length, ZSTD_overlap_e const ovtype)
{
    ptrdiff_t diff = (BYTE*)dst - (const BYTE*)src;
    const BYTE* ip = (const BYTE*)src;
    BYTE* op = (BYTE*)dst;
    BYTE* const oend = op + length;

    if (ovtype == ZSTD_overlap_src_before_dst && diff < WILDCOPY_VECLEN) {
        /* Handle short offset copies. */
        do {
            COPY8(op, ip);
        } while (op < oend);
    } else {
        assert(diff >= WILDCOPY_VECLEN || diff <= -WILDCOPY_VECLEN);
        /* Separate out the first COPY16() call because the copy length is
         * almost certain to be short, so the branches have different
         * probabilities. Since it is almost certain to be short, only do
         * one COPY16() in the first call. Then, do two calls per loop since
         * at that point it is more likely to have a high trip count.
         */
        ZSTD_copy16(op, ip);
        if (16 >= length) return;
        op += 16;
        ip += 16;
        do {
            COPY16(op, ip);
            COPY16(op, ip);
        }
        while (op < oend);
    }
}

MEM_STATIC size_t ZSTD_limitCopy(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    size_t const length = MIN(dstCapacity, srcSize);
    if (length > 0) {
        ZSTD_memcpy(dst, src, length);
    }
    return length;
}

/* define "workspace is too large" as this number of times larger than needed */
#define ZSTD_WORKSPACETOOLARGE_FACTOR 3

/* when workspace is continuously too large
 * during at least this number of times,
 * context's memory usage is considered wasteful,
 * because it's sized to handle a worst case scenario which rarely happens.
 * In which case, resize it down to free some memory */
#define ZSTD_WORKSPACETOOLARGE_MAXDURATION 128

/* Controls whether the input/output buffer is buffered or stable. */
typedef enum {
    ZSTD_bm_buffered = 0,  /* Buffer the input/output */
    ZSTD_bm_stable = 1     /* ZSTD_inBuffer/ZSTD_outBuffer is stable */
} ZSTD_bufferMode_e;


/*-*******************************************
*  Private declarations
*********************************************/

/**
 * Contains the compressed frame size and an upper-bound for the decompressed frame size.
 * Note: before using `compressedSize`, check for errors using ZSTD_isError().
 *       similarly, before using `decompressedBound`, check for errors using:
 *          `decompressedBound != ZSTD_CONTENTSIZE_ERROR`
 */
typedef struct {
    size_t nbBlocks;
    size_t compressedSize;
    unsigned long long decompressedBound;
} ZSTD_frameSizeInfo;   /* decompress & legacy */

/* ZSTD_invalidateRepCodes() :
 * ensures next compression will not use repcodes from previous block.
 * Note : only works with regular variant;
 *        do not use with extDict variant ! */
void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx);   /* zstdmt, adaptive_compression (shouldn't get this definition from here) */


typedef struct {
    blockType_e blockType;
    U32 lastBlock;
    U32 origSize;
} blockProperties_t;   /* declared here for decompress and fullbench */

/*! ZSTD_getcBlockSize() :
 *  Provides the size of compressed block from block header `src` */
/*  Used by: decompress, fullbench */
size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
                          blockProperties_t* bpPtr);

/*! ZSTD_decodeSeqHeaders() :
 *  decode sequence header from src */
/*  Used by: zstd_decompress_block, fullbench */
size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                       const void* src, size_t srcSize);

/**
 * @returns true iff the CPU supports dynamic BMI2 dispatch.
 */
MEM_STATIC int ZSTD_cpuSupportsBmi2(void)
{
    ZSTD_cpuid_t cpuid = ZSTD_cpuid();
    return ZSTD_cpuid_bmi1(cpuid) && ZSTD_cpuid_bmi2(cpuid);
}

#endif   /* ZSTD_CCOMMON_H_MODULE */
/**** ended inlining zstd_internal.h ****/


/*-****************************************
*  Version
******************************************/
unsigned ZSTD_versionNumber(void) { return ZSTD_VERSION_NUMBER; }

const char* ZSTD_versionString(void) { return ZSTD_VERSION_STRING; }


/*-****************************************
*  ZSTD Error Management
******************************************/
#undef ZSTD_isError   /* defined within zstd_internal.h */
/*! ZSTD_isError() :
 *  tells if a return value is an error code
 *  symbol is required for external callers */
unsigned ZSTD_isError(size_t code) { return ERR_isError(code); }

/*! ZSTD_getErrorName() :
 *  provides error code string from function result (useful for debugging) */
const char* ZSTD_getErrorName(size_t code) { return ERR_getErrorName(code); }

/*! ZSTD_getError() :
 *  convert a `size_t` function result into a proper ZSTD_errorCode enum */
ZSTD_ErrorCode ZSTD_getErrorCode(size_t code) { return ERR_getErrorCode(code); }

/*! ZSTD_getErrorString() :
 *  provides error code string from enum */
const char* ZSTD_getErrorString(ZSTD_ErrorCode code) { return ERR_getErrorString(code); }
/**** ended inlining common/zstd_common.c ****/

/**** start inlining compress/fse_compress.c ****/
/* ******************************************************************
 * FSE : Finite State Entropy encoder
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* **************************************************************
*  Includes
****************************************************************/
/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/debug.h ****/
/**** start inlining hist.h ****/
/* ******************************************************************
 * hist : Histogram functions
 * part of Finite State Entropy project
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* --- dependencies --- */
/**** skipping file: ../common/zstd_deps.h ****/


/* --- simple histogram functions --- */

/*! HIST_count():
 *  Provides the precise count of each byte within a table 'count'.
 * 'count' is a table of unsigned int, of minimum size (*maxSymbolValuePtr+1).
 *  Updates *maxSymbolValuePtr with actual largest symbol value detected.
 * @return : count of the most frequent symbol (which isn't identified).
 *           or an error code, which can be tested using HIST_isError().
 *           note : if return == srcSize, there is only one symbol.
 */
size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr,
                  const void* src, size_t srcSize);

unsigned HIST_isError(size_t code);  /**< tells if a return value is an error code */


/* --- advanced histogram functions --- */

#define HIST_WKSP_SIZE_U32 1024
#define HIST_WKSP_SIZE    (HIST_WKSP_SIZE_U32 * sizeof(unsigned))
/** HIST_count_wksp() :
 *  Same as HIST_count(), but using an externally provided scratch buffer.
 *  Benefit is this function will use very little stack space.
 * `workSpace` is a writable buffer which must be 4-bytes aligned,
 * `workSpaceSize` must be >= HIST_WKSP_SIZE
 */
size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
                       const void* src, size_t srcSize,
                       void* workSpace, size_t workSpaceSize);

/** HIST_countFast() :
 *  same as HIST_count(), but blindly trusts that all byte values within src are <= *maxSymbolValuePtr.
 *  This function is unsafe, and will segfault if any value within `src` is `> *maxSymbolValuePtr`
 */
size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
                      const void* src, size_t srcSize);

/** HIST_countFast_wksp() :
 *  Same as HIST_countFast(), but using an externally provided scratch buffer.
 * `workSpace` is a writable buffer which must be 4-bytes aligned,
 * `workSpaceSize` must be >= HIST_WKSP_SIZE
 */
size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
                           const void* src, size_t srcSize,
                           void* workSpace, size_t workSpaceSize);

/*! HIST_count_simple() :
 *  Same as HIST_countFast(), this function is unsafe,
 *  and will segfault if any value within `src` is `> *maxSymbolValuePtr`.
 *  It is also a bit slower for large inputs.
 *  However, it does not need any additional memory (not even on stack).
 * @return : count of the most frequent symbol.
 *  Note this function doesn't produce any error (i.e. it must succeed).
 */
unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
                           const void* src, size_t srcSize);

/*! HIST_add() :
 *  Lowest level: just add nb of occurrences of characters from @src into @count.
 *  @count is not reset. @count array is presumed large enough (i.e. 1 KB).
 @  This function does not need any additional stack memory.
 */
void HIST_add(unsigned* count, const void* src, size_t srcSize);
/**** ended inlining hist.h ****/
/**** skipping file: ../common/bitstream.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/error_private.h ****/
#define ZSTD_DEPS_NEED_MALLOC
#define ZSTD_DEPS_NEED_MATH64
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/bits.h ****/


/* **************************************************************
*  Error Management
****************************************************************/
#define FSE_isError ERR_isError


/* **************************************************************
*  Templates
****************************************************************/
/*
  designed to be included
  for type-specific functions (template emulation in C)
  Objective is to write these functions only once, for improved maintenance
*/

/* safety checks */
#ifndef FSE_FUNCTION_EXTENSION
#  error "FSE_FUNCTION_EXTENSION must be defined"
#endif
#ifndef FSE_FUNCTION_TYPE
#  error "FSE_FUNCTION_TYPE must be defined"
#endif

/* Function names */
#define FSE_CAT(X,Y) X##Y
#define FSE_FUNCTION_NAME(X,Y) FSE_CAT(X,Y)
#define FSE_TYPE_NAME(X,Y) FSE_CAT(X,Y)


/* Function templates */

/* FSE_buildCTable_wksp() :
 * Same as FSE_buildCTable(), but using an externally allocated scratch buffer (`workSpace`).
 * wkspSize should be sized to handle worst case situation, which is `1<<max_tableLog * sizeof(FSE_FUNCTION_TYPE)`
 * workSpace must also be properly aligned with FSE_FUNCTION_TYPE requirements
 */
size_t FSE_buildCTable_wksp(FSE_CTable* ct,
                      const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
                            void* workSpace, size_t wkspSize)
{
    U32 const tableSize = 1 << tableLog;
    U32 const tableMask = tableSize - 1;
    void* const ptr = ct;
    U16* const tableU16 = ( (U16*) ptr) + 2;
    void* const FSCT = ((U32*)ptr) + 1 /* header */ + (tableLog ? tableSize>>1 : 1) ;
    FSE_symbolCompressionTransform* const symbolTT = (FSE_symbolCompressionTransform*) (FSCT);
    U32 const step = FSE_TABLESTEP(tableSize);
    U32 const maxSV1 = maxSymbolValue+1;

    U16* cumul = (U16*)workSpace;   /* size = maxSV1 */
    FSE_FUNCTION_TYPE* const tableSymbol = (FSE_FUNCTION_TYPE*)(cumul + (maxSV1+1));  /* size = tableSize */

    U32 highThreshold = tableSize-1;

    assert(((size_t)workSpace & 1) == 0);  /* Must be 2 bytes-aligned */
    if (FSE_BUILD_CTABLE_WORKSPACE_SIZE(maxSymbolValue, tableLog) > wkspSize) return ERROR(tableLog_tooLarge);
    /* CTable header */
    tableU16[-2] = (U16) tableLog;
    tableU16[-1] = (U16) maxSymbolValue;
    assert(tableLog < 16);   /* required for threshold strategy to work */

    /* For explanations on how to distribute symbol values over the table :
     * https://fastcompression.blogspot.fr/2014/02/fse-distributing-symbol-values.html */

     #ifdef __clang_analyzer__
     ZSTD_memset(tableSymbol, 0, sizeof(*tableSymbol) * tableSize);   /* useless initialization, just to keep scan-build happy */
     #endif

    /* symbol start positions */
    {   U32 u;
        cumul[0] = 0;
        for (u=1; u <= maxSV1; u++) {
            if (normalizedCounter[u-1]==-1) {  /* Low proba symbol */
                cumul[u] = cumul[u-1] + 1;
                tableSymbol[highThreshold--] = (FSE_FUNCTION_TYPE)(u-1);
            } else {
                assert(normalizedCounter[u-1] >= 0);
                cumul[u] = cumul[u-1] + (U16)normalizedCounter[u-1];
                assert(cumul[u] >= cumul[u-1]);  /* no overflow */
        }   }
        cumul[maxSV1] = (U16)(tableSize+1);
    }

    /* Spread symbols */
    if (highThreshold == tableSize - 1) {
        /* Case for no low prob count symbols. Lay down 8 bytes at a time
         * to reduce branch misses since we are operating on a small block
         */
        BYTE* const spread = tableSymbol + tableSize; /* size = tableSize + 8 (may write beyond tableSize) */
        {   U64 const add = 0x0101010101010101ull;
            size_t pos = 0;
            U64 sv = 0;
            U32 s;
            for (s=0; s<maxSV1; ++s, sv += add) {
                int i;
                int const n = normalizedCounter[s];
                MEM_write64(spread + pos, sv);
                for (i = 8; i < n; i += 8) {
                    MEM_write64(spread + pos + i, sv);
                }
                assert(n>=0);
                pos += (size_t)n;
            }
        }
        /* Spread symbols across the table. Lack of lowprob symbols means that
         * we don't need variable sized inner loop, so we can unroll the loop and
         * reduce branch misses.
         */
        {   size_t position = 0;
            size_t s;
            size_t const unroll = 2; /* Experimentally determined optimal unroll */
            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
            for (s = 0; s < (size_t)tableSize; s += unroll) {
                size_t u;
                for (u = 0; u < unroll; ++u) {
                    size_t const uPosition = (position + (u * step)) & tableMask;
                    tableSymbol[uPosition] = spread[s + u];
                }
                position = (position + (unroll * step)) & tableMask;
            }
            assert(position == 0);   /* Must have initialized all positions */
        }
    } else {
        U32 position = 0;
        U32 symbol;
        for (symbol=0; symbol<maxSV1; symbol++) {
            int nbOccurrences;
            int const freq = normalizedCounter[symbol];
            for (nbOccurrences=0; nbOccurrences<freq; nbOccurrences++) {
                tableSymbol[position] = (FSE_FUNCTION_TYPE)symbol;
                position = (position + step) & tableMask;
                while (position > highThreshold)
                    position = (position + step) & tableMask;   /* Low proba area */
        }   }
        assert(position==0);  /* Must have initialized all positions */
    }

    /* Build table */
    {   U32 u; for (u=0; u<tableSize; u++) {
        FSE_FUNCTION_TYPE s = tableSymbol[u];   /* note : static analyzer may not understand tableSymbol is properly initialized */
        tableU16[cumul[s]++] = (U16) (tableSize+u);   /* TableU16 : sorted by symbol order; gives next state value */
    }   }

    /* Build Symbol Transformation Table */
    {   unsigned total = 0;
        unsigned s;
        for (s=0; s<=maxSymbolValue; s++) {
            switch (normalizedCounter[s])
            {
            case  0:
                /* filling nonetheless, for compatibility with FSE_getMaxNbBits() */
                symbolTT[s].deltaNbBits = ((tableLog+1) << 16) - (1<<tableLog);
                break;

            case -1:
            case  1:
                symbolTT[s].deltaNbBits = (tableLog << 16) - (1<<tableLog);
                assert(total <= INT_MAX);
                symbolTT[s].deltaFindState = (int)(total - 1);
                total ++;
                break;
            default :
                assert(normalizedCounter[s] > 1);
                {   U32 const maxBitsOut = tableLog - ZSTD_highbit32 ((U32)normalizedCounter[s]-1);
                    U32 const minStatePlus = (U32)normalizedCounter[s] << maxBitsOut;
                    symbolTT[s].deltaNbBits = (maxBitsOut << 16) - minStatePlus;
                    symbolTT[s].deltaFindState = (int)(total - (unsigned)normalizedCounter[s]);
                    total +=  (unsigned)normalizedCounter[s];
    }   }   }   }

#if 0  /* debug : symbol costs */
    DEBUGLOG(5, "\n --- table statistics : ");
    {   U32 symbol;
        for (symbol=0; symbol<=maxSymbolValue; symbol++) {
            DEBUGLOG(5, "%3u: w=%3i,   maxBits=%u, fracBits=%.2f",
                symbol, normalizedCounter[symbol],
                FSE_getMaxNbBits(symbolTT, symbol),
                (double)FSE_bitCost(symbolTT, tableLog, symbol, 8) / 256);
    }   }
#endif

    return 0;
}


#ifndef FSE_COMMONDEFS_ONLY

/*-**************************************************************
*  FSE NCount encoding
****************************************************************/
size_t FSE_NCountWriteBound(unsigned maxSymbolValue, unsigned tableLog)
{
    size_t const maxHeaderSize = (((maxSymbolValue+1) * tableLog
                                   + 4 /* bitCount initialized at 4 */
                                   + 2 /* first two symbols may use one additional bit each */) / 8)
                                   + 1 /* round up to whole nb bytes */
                                   + 2 /* additional two bytes for bitstream flush */;
    return maxSymbolValue ? maxHeaderSize : FSE_NCOUNTBOUND;  /* maxSymbolValue==0 ? use default */
}

static size_t
FSE_writeNCount_generic (void* header, size_t headerBufferSize,
                   const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog,
                         unsigned writeIsSafe)
{
    BYTE* const ostart = (BYTE*) header;
    BYTE* out = ostart;
    BYTE* const oend = ostart + headerBufferSize;
    int nbBits;
    const int tableSize = 1 << tableLog;
    int remaining;
    int threshold;
    U32 bitStream = 0;
    int bitCount = 0;
    unsigned symbol = 0;
    unsigned const alphabetSize = maxSymbolValue + 1;
    int previousIs0 = 0;

    /* Table Size */
    bitStream += (tableLog-FSE_MIN_TABLELOG) << bitCount;
    bitCount  += 4;

    /* Init */
    remaining = tableSize+1;   /* +1 for extra accuracy */
    threshold = tableSize;
    nbBits = (int)tableLog+1;

    while ((symbol < alphabetSize) && (remaining>1)) {  /* stops at 1 */
        if (previousIs0) {
            unsigned start = symbol;
            while ((symbol < alphabetSize) && !normalizedCounter[symbol]) symbol++;
            if (symbol == alphabetSize) break;   /* incorrect distribution */
            while (symbol >= start+24) {
                start+=24;
                bitStream += 0xFFFFU << bitCount;
                if ((!writeIsSafe) && (out > oend-2))
                    return ERROR(dstSize_tooSmall);   /* Buffer overflow */
                out[0] = (BYTE) bitStream;
                out[1] = (BYTE)(bitStream>>8);
                out+=2;
                bitStream>>=16;
            }
            while (symbol >= start+3) {
                start+=3;
                bitStream += 3U << bitCount;
                bitCount += 2;
            }
            bitStream += (symbol-start) << bitCount;
            bitCount += 2;
            if (bitCount>16) {
                if ((!writeIsSafe) && (out > oend - 2))
                    return ERROR(dstSize_tooSmall);   /* Buffer overflow */
                out[0] = (BYTE)bitStream;
                out[1] = (BYTE)(bitStream>>8);
                out += 2;
                bitStream >>= 16;
                bitCount -= 16;
        }   }
        {   int count = normalizedCounter[symbol++];
            int const max = (2*threshold-1) - remaining;
            remaining -= count < 0 ? -count : count;
            count++;   /* +1 for extra accuracy */
            if (count>=threshold)
                count += max;   /* [0..max[ [max..threshold[ (...) [threshold+max 2*threshold[ */
            bitStream += (U32)count << bitCount;
            bitCount  += nbBits;
            bitCount  -= (count<max);
            previousIs0  = (count==1);
            if (remaining<1) return ERROR(GENERIC);
            while (remaining<threshold) { nbBits--; threshold>>=1; }
        }
        if (bitCount>16) {
            if ((!writeIsSafe) && (out > oend - 2))
                return ERROR(dstSize_tooSmall);   /* Buffer overflow */
            out[0] = (BYTE)bitStream;
            out[1] = (BYTE)(bitStream>>8);
            out += 2;
            bitStream >>= 16;
            bitCount -= 16;
    }   }

    if (remaining != 1)
        return ERROR(GENERIC);  /* incorrect normalized distribution */
    assert(symbol <= alphabetSize);

    /* flush remaining bitStream */
    if ((!writeIsSafe) && (out > oend - 2))
        return ERROR(dstSize_tooSmall);   /* Buffer overflow */
    out[0] = (BYTE)bitStream;
    out[1] = (BYTE)(bitStream>>8);
    out+= (bitCount+7) /8;

    assert(out >= ostart);
    return (size_t)(out-ostart);
}


size_t FSE_writeNCount (void* buffer, size_t bufferSize,
                  const short* normalizedCounter, unsigned maxSymbolValue, unsigned tableLog)
{
    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported */
    if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported */

    if (bufferSize < FSE_NCountWriteBound(maxSymbolValue, tableLog))
        return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 0);

    return FSE_writeNCount_generic(buffer, bufferSize, normalizedCounter, maxSymbolValue, tableLog, 1 /* write in buffer is safe */);
}


/*-**************************************************************
*  FSE Compression Code
****************************************************************/

/* provides the minimum logSize to safely represent a distribution */
static unsigned FSE_minTableLog(size_t srcSize, unsigned maxSymbolValue)
{
    U32 minBitsSrc = ZSTD_highbit32((U32)(srcSize)) + 1;
    U32 minBitsSymbols = ZSTD_highbit32(maxSymbolValue) + 2;
    U32 minBits = minBitsSrc < minBitsSymbols ? minBitsSrc : minBitsSymbols;
    assert(srcSize > 1); /* Not supported, RLE should be used instead */
    return minBits;
}

unsigned FSE_optimalTableLog_internal(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue, unsigned minus)
{
    U32 maxBitsSrc = ZSTD_highbit32((U32)(srcSize - 1)) - minus;
    U32 tableLog = maxTableLog;
    U32 minBits = FSE_minTableLog(srcSize, maxSymbolValue);
    assert(srcSize > 1); /* Not supported, RLE should be used instead */
    if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
    if (maxBitsSrc < tableLog) tableLog = maxBitsSrc;   /* Accuracy can be reduced */
    if (minBits > tableLog) tableLog = minBits;   /* Need a minimum to safely represent all symbol values */
    if (tableLog < FSE_MIN_TABLELOG) tableLog = FSE_MIN_TABLELOG;
    if (tableLog > FSE_MAX_TABLELOG) tableLog = FSE_MAX_TABLELOG;
    return tableLog;
}

unsigned FSE_optimalTableLog(unsigned maxTableLog, size_t srcSize, unsigned maxSymbolValue)
{
    return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 2);
}

/* Secondary normalization method.
   To be used when primary method fails. */

static size_t FSE_normalizeM2(short* norm, U32 tableLog, const unsigned* count, size_t total, U32 maxSymbolValue, short lowProbCount)
{
    short const NOT_YET_ASSIGNED = -2;
    U32 s;
    U32 distributed = 0;
    U32 ToDistribute;

    /* Init */
    U32 const lowThreshold = (U32)(total >> tableLog);
    U32 lowOne = (U32)((total * 3) >> (tableLog + 1));

    for (s=0; s<=maxSymbolValue; s++) {
        if (count[s] == 0) {
            norm[s]=0;
            continue;
        }
        if (count[s] <= lowThreshold) {
            norm[s] = lowProbCount;
            distributed++;
            total -= count[s];
            continue;
        }
        if (count[s] <= lowOne) {
            norm[s] = 1;
            distributed++;
            total -= count[s];
            continue;
        }

        norm[s]=NOT_YET_ASSIGNED;
    }
    ToDistribute = (1 << tableLog) - distributed;

    if (ToDistribute == 0)
        return 0;

    if ((total / ToDistribute) > lowOne) {
        /* risk of rounding to zero */
        lowOne = (U32)((total * 3) / (ToDistribute * 2));
        for (s=0; s<=maxSymbolValue; s++) {
            if ((norm[s] == NOT_YET_ASSIGNED) && (count[s] <= lowOne)) {
                norm[s] = 1;
                distributed++;
                total -= count[s];
                continue;
        }   }
        ToDistribute = (1 << tableLog) - distributed;
    }

    if (distributed == maxSymbolValue+1) {
        /* all values are pretty poor;
           probably incompressible data (should have already been detected);
           find max, then give all remaining points to max */
        U32 maxV = 0, maxC = 0;
        for (s=0; s<=maxSymbolValue; s++)
            if (count[s] > maxC) { maxV=s; maxC=count[s]; }
        norm[maxV] += (short)ToDistribute;
        return 0;
    }

    if (total == 0) {
        /* all of the symbols were low enough for the lowOne or lowThreshold */
        for (s=0; ToDistribute > 0; s = (s+1)%(maxSymbolValue+1))
            if (norm[s] > 0) { ToDistribute--; norm[s]++; }
        return 0;
    }

    {   U64 const vStepLog = 62 - tableLog;
        U64 const mid = (1ULL << (vStepLog-1)) - 1;
        U64 const rStep = ZSTD_div64((((U64)1<<vStepLog) * ToDistribute) + mid, (U32)total);   /* scale on remaining */
        U64 tmpTotal = mid;
        for (s=0; s<=maxSymbolValue; s++) {
            if (norm[s]==NOT_YET_ASSIGNED) {
                U64 const end = tmpTotal + (count[s] * rStep);
                U32 const sStart = (U32)(tmpTotal >> vStepLog);
                U32 const sEnd = (U32)(end >> vStepLog);
                U32 const weight = sEnd - sStart;
                if (weight < 1)
                    return ERROR(GENERIC);
                norm[s] = (short)weight;
                tmpTotal = end;
    }   }   }

    return 0;
}

size_t FSE_normalizeCount (short* normalizedCounter, unsigned tableLog,
                           const unsigned* count, size_t total,
                           unsigned maxSymbolValue, unsigned useLowProbCount)
{
    /* Sanity checks */
    if (tableLog==0) tableLog = FSE_DEFAULT_TABLELOG;
    if (tableLog < FSE_MIN_TABLELOG) return ERROR(GENERIC);   /* Unsupported size */
    if (tableLog > FSE_MAX_TABLELOG) return ERROR(tableLog_tooLarge);   /* Unsupported size */
    if (tableLog < FSE_minTableLog(total, maxSymbolValue)) return ERROR(GENERIC);   /* Too small tableLog, compression potentially impossible */

    {   static U32 const rtbTable[] = {     0, 473195, 504333, 520860, 550000, 700000, 750000, 830000 };
        short const lowProbCount = useLowProbCount ? -1 : 1;
        U64 const scale = 62 - tableLog;
        U64 const step = ZSTD_div64((U64)1<<62, (U32)total);   /* <== here, one division ! */
        U64 const vStep = 1ULL<<(scale-20);
        int stillToDistribute = 1<<tableLog;
        unsigned s;
        unsigned largest=0;
        short largestP=0;
        U32 lowThreshold = (U32)(total >> tableLog);

        for (s=0; s<=maxSymbolValue; s++) {
            if (count[s] == total) return 0;   /* rle special case */
            if (count[s] == 0) { normalizedCounter[s]=0; continue; }
            if (count[s] <= lowThreshold) {
                normalizedCounter[s] = lowProbCount;
                stillToDistribute--;
            } else {
                short proba = (short)((count[s]*step) >> scale);
                if (proba<8) {
                    U64 restToBeat = vStep * rtbTable[proba];
                    proba += (count[s]*step) - ((U64)proba<<scale) > restToBeat;
                }
                if (proba > largestP) { largestP=proba; largest=s; }
                normalizedCounter[s] = proba;
                stillToDistribute -= proba;
        }   }
        if (-stillToDistribute >= (normalizedCounter[largest] >> 1)) {
            /* corner case, need another normalization method */
            size_t const errorCode = FSE_normalizeM2(normalizedCounter, tableLog, count, total, maxSymbolValue, lowProbCount);
            if (FSE_isError(errorCode)) return errorCode;
        }
        else normalizedCounter[largest] += (short)stillToDistribute;
    }

#if 0
    {   /* Print Table (debug) */
        U32 s;
        U32 nTotal = 0;
        for (s=0; s<=maxSymbolValue; s++)
            RAWLOG(2, "%3i: %4i \n", s, normalizedCounter[s]);
        for (s=0; s<=maxSymbolValue; s++)
            nTotal += abs(normalizedCounter[s]);
        if (nTotal != (1U<<tableLog))
            RAWLOG(2, "Warning !!! Total == %u != %u !!!", nTotal, 1U<<tableLog);
        getchar();
    }
#endif

    return tableLog;
}

/* fake FSE_CTable, for rle input (always same symbol) */
size_t FSE_buildCTable_rle (FSE_CTable* ct, BYTE symbolValue)
{
    void* ptr = ct;
    U16* tableU16 = ( (U16*) ptr) + 2;
    void* FSCTptr = (U32*)ptr + 2;
    FSE_symbolCompressionTransform* symbolTT = (FSE_symbolCompressionTransform*) FSCTptr;

    /* header */
    tableU16[-2] = (U16) 0;
    tableU16[-1] = (U16) symbolValue;

    /* Build table */
    tableU16[0] = 0;
    tableU16[1] = 0;   /* just in case */

    /* Build Symbol Transformation Table */
    symbolTT[symbolValue].deltaNbBits = 0;
    symbolTT[symbolValue].deltaFindState = 0;

    return 0;
}


static size_t FSE_compress_usingCTable_generic (void* dst, size_t dstSize,
                           const void* src, size_t srcSize,
                           const FSE_CTable* ct, const unsigned fast)
{
    const BYTE* const istart = (const BYTE*) src;
    const BYTE* const iend = istart + srcSize;
    const BYTE* ip=iend;

    BIT_CStream_t bitC;
    FSE_CState_t CState1, CState2;

    /* init */
    if (srcSize <= 2) return 0;
    { size_t const initError = BIT_initCStream(&bitC, dst, dstSize);
      if (FSE_isError(initError)) return 0; /* not enough space available to write a bitstream */ }

#define FSE_FLUSHBITS(s)  (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))

    if (srcSize & 1) {
        FSE_initCState2(&CState1, ct, *--ip);
        FSE_initCState2(&CState2, ct, *--ip);
        FSE_encodeSymbol(&bitC, &CState1, *--ip);
        FSE_FLUSHBITS(&bitC);
    } else {
        FSE_initCState2(&CState2, ct, *--ip);
        FSE_initCState2(&CState1, ct, *--ip);
    }

    /* join to mod 4 */
    srcSize -= 2;
    if ((sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) && (srcSize & 2)) {  /* test bit 2 */
        FSE_encodeSymbol(&bitC, &CState2, *--ip);
        FSE_encodeSymbol(&bitC, &CState1, *--ip);
        FSE_FLUSHBITS(&bitC);
    }

    /* 2 or 4 encoding per loop */
    while ( ip>istart ) {

        FSE_encodeSymbol(&bitC, &CState2, *--ip);

        if (sizeof(bitC.bitContainer)*8 < FSE_MAX_TABLELOG*2+7 )   /* this test must be static */
            FSE_FLUSHBITS(&bitC);

        FSE_encodeSymbol(&bitC, &CState1, *--ip);

        if (sizeof(bitC.bitContainer)*8 > FSE_MAX_TABLELOG*4+7 ) {  /* this test must be static */
            FSE_encodeSymbol(&bitC, &CState2, *--ip);
            FSE_encodeSymbol(&bitC, &CState1, *--ip);
        }

        FSE_FLUSHBITS(&bitC);
    }

    FSE_flushCState(&bitC, &CState2);
    FSE_flushCState(&bitC, &CState1);
    return BIT_closeCStream(&bitC);
}

size_t FSE_compress_usingCTable (void* dst, size_t dstSize,
                           const void* src, size_t srcSize,
                           const FSE_CTable* ct)
{
    unsigned const fast = (dstSize >= FSE_BLOCKBOUND(srcSize));

    if (fast)
        return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 1);
    else
        return FSE_compress_usingCTable_generic(dst, dstSize, src, srcSize, ct, 0);
}


size_t FSE_compressBound(size_t size) { return FSE_COMPRESSBOUND(size); }

#endif   /* FSE_COMMONDEFS_ONLY */
/**** ended inlining compress/fse_compress.c ****/
/**** start inlining compress/hist.c ****/
/* ******************************************************************
 * hist : Histogram functions
 * part of Finite State Entropy project
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* --- dependencies --- */
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/debug.h ****/
/**** skipping file: ../common/error_private.h ****/
/**** skipping file: hist.h ****/


/* --- Error management --- */
unsigned HIST_isError(size_t code) { return ERR_isError(code); }

/*-**************************************************************
 *  Histogram functions
 ****************************************************************/
void HIST_add(unsigned* count, const void* src, size_t srcSize)
{
    const BYTE* ip = (const BYTE*)src;
    const BYTE* const end = ip + srcSize;

    while (ip<end) {
        count[*ip++]++;
    }
}

unsigned HIST_count_simple(unsigned* count, unsigned* maxSymbolValuePtr,
                           const void* src, size_t srcSize)
{
    const BYTE* ip = (const BYTE*)src;
    const BYTE* const end = ip + srcSize;
    unsigned maxSymbolValue = *maxSymbolValuePtr;
    unsigned largestCount=0;

    ZSTD_memset(count, 0, (maxSymbolValue+1) * sizeof(*count));
    if (srcSize==0) { *maxSymbolValuePtr = 0; return 0; }

    while (ip<end) {
        assert(*ip <= maxSymbolValue);
        count[*ip++]++;
    }

    while (!count[maxSymbolValue]) maxSymbolValue--;
    *maxSymbolValuePtr = maxSymbolValue;

    {   U32 s;
        for (s=0; s<=maxSymbolValue; s++)
            if (count[s] > largestCount) largestCount = count[s];
    }

    return largestCount;
}

typedef enum { trustInput, checkMaxSymbolValue } HIST_checkInput_e;

/* HIST_count_parallel_wksp() :
 * store histogram into 4 intermediate tables, recombined at the end.
 * this design makes better use of OoO cpus,
 * and is noticeably faster when some values are heavily repeated.
 * But it needs some additional workspace for intermediate tables.
 * `workSpace` must be a U32 table of size >= HIST_WKSP_SIZE_U32.
 * @return : largest histogram frequency,
 *           or an error code (notably when histogram's alphabet is larger than *maxSymbolValuePtr) */
static size_t HIST_count_parallel_wksp(
                                unsigned* count, unsigned* maxSymbolValuePtr,
                                const void* source, size_t sourceSize,
                                HIST_checkInput_e check,
                                U32* const workSpace)
{
    const BYTE* ip = (const BYTE*)source;
    const BYTE* const iend = ip+sourceSize;
    size_t const countSize = (*maxSymbolValuePtr + 1) * sizeof(*count);
    unsigned max=0;
    U32* const Counting1 = workSpace;
    U32* const Counting2 = Counting1 + 256;
    U32* const Counting3 = Counting2 + 256;
    U32* const Counting4 = Counting3 + 256;

    /* safety checks */
    assert(*maxSymbolValuePtr <= 255);
    if (!sourceSize) {
        ZSTD_memset(count, 0, countSize);
        *maxSymbolValuePtr = 0;
        return 0;
    }
    ZSTD_memset(workSpace, 0, 4*256*sizeof(unsigned));

    /* by stripes of 16 bytes */
    {   U32 cached = MEM_read32(ip); ip += 4;
        while (ip < iend-15) {
            U32 c = cached; cached = MEM_read32(ip); ip += 4;
            Counting1[(BYTE) c     ]++;
            Counting2[(BYTE)(c>>8) ]++;
            Counting3[(BYTE)(c>>16)]++;
            Counting4[       c>>24 ]++;
            c = cached; cached = MEM_read32(ip); ip += 4;
            Counting1[(BYTE) c     ]++;
            Counting2[(BYTE)(c>>8) ]++;
            Counting3[(BYTE)(c>>16)]++;
            Counting4[       c>>24 ]++;
            c = cached; cached = MEM_read32(ip); ip += 4;
            Counting1[(BYTE) c     ]++;
            Counting2[(BYTE)(c>>8) ]++;
            Counting3[(BYTE)(c>>16)]++;
            Counting4[       c>>24 ]++;
            c = cached; cached = MEM_read32(ip); ip += 4;
            Counting1[(BYTE) c     ]++;
            Counting2[(BYTE)(c>>8) ]++;
            Counting3[(BYTE)(c>>16)]++;
            Counting4[       c>>24 ]++;
        }
        ip-=4;
    }

    /* finish last symbols */
    while (ip<iend) Counting1[*ip++]++;

    {   U32 s;
        for (s=0; s<256; s++) {
            Counting1[s] += Counting2[s] + Counting3[s] + Counting4[s];
            if (Counting1[s] > max) max = Counting1[s];
    }   }

    {   unsigned maxSymbolValue = 255;
        while (!Counting1[maxSymbolValue]) maxSymbolValue--;
        if (check && maxSymbolValue > *maxSymbolValuePtr) return ERROR(maxSymbolValue_tooSmall);
        *maxSymbolValuePtr = maxSymbolValue;
        ZSTD_memmove(count, Counting1, countSize);   /* in case count & Counting1 are overlapping */
    }
    return (size_t)max;
}

/* HIST_countFast_wksp() :
 * Same as HIST_countFast(), but using an externally provided scratch buffer.
 * `workSpace` is a writable buffer which must be 4-bytes aligned,
 * `workSpaceSize` must be >= HIST_WKSP_SIZE
 */
size_t HIST_countFast_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
                          const void* source, size_t sourceSize,
                          void* workSpace, size_t workSpaceSize)
{
    if (sourceSize < 1500) /* heuristic threshold */
        return HIST_count_simple(count, maxSymbolValuePtr, source, sourceSize);
    if ((size_t)workSpace & 3) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
    if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall);
    return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, trustInput, (U32*)workSpace);
}

/* HIST_count_wksp() :
 * Same as HIST_count(), but using an externally provided scratch buffer.
 * `workSpace` size must be table of >= HIST_WKSP_SIZE_U32 unsigned */
size_t HIST_count_wksp(unsigned* count, unsigned* maxSymbolValuePtr,
                       const void* source, size_t sourceSize,
                       void* workSpace, size_t workSpaceSize)
{
    if ((size_t)workSpace & 3) return ERROR(GENERIC);  /* must be aligned on 4-bytes boundaries */
    if (workSpaceSize < HIST_WKSP_SIZE) return ERROR(workSpace_tooSmall);
    if (*maxSymbolValuePtr < 255)
        return HIST_count_parallel_wksp(count, maxSymbolValuePtr, source, sourceSize, checkMaxSymbolValue, (U32*)workSpace);
    *maxSymbolValuePtr = 255;
    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, workSpace, workSpaceSize);
}

#ifndef ZSTD_NO_UNUSED_FUNCTIONS
/* fast variant (unsafe : won't check if src contains values beyond count[] limit) */
size_t HIST_countFast(unsigned* count, unsigned* maxSymbolValuePtr,
                     const void* source, size_t sourceSize)
{
    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
    return HIST_countFast_wksp(count, maxSymbolValuePtr, source, sourceSize, tmpCounters, sizeof(tmpCounters));
}

size_t HIST_count(unsigned* count, unsigned* maxSymbolValuePtr,
                 const void* src, size_t srcSize)
{
    unsigned tmpCounters[HIST_WKSP_SIZE_U32];
    return HIST_count_wksp(count, maxSymbolValuePtr, src, srcSize, tmpCounters, sizeof(tmpCounters));
}
#endif
/**** ended inlining compress/hist.c ****/
/**** start inlining compress/huf_compress.c ****/
/* ******************************************************************
 * Huffman encoder, part of New Generation Entropy library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *  - Public forum : https://groups.google.com/forum/#!forum/lz4c
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* **************************************************************
*  Compiler specifics
****************************************************************/
#ifdef _MSC_VER    /* Visual Studio */
#  pragma warning(disable : 4127)        /* disable: C4127: conditional expression is constant */
#endif


/* **************************************************************
*  Includes
****************************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/bitstream.h ****/
/**** skipping file: hist.h ****/
#define FSE_STATIC_LINKING_ONLY   /* FSE_optimalTableLog_internal */
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: ../common/error_private.h ****/
/**** skipping file: ../common/bits.h ****/


/* **************************************************************
*  Error Management
****************************************************************/
#define HUF_isError ERR_isError
#define HUF_STATIC_ASSERT(c) DEBUG_STATIC_ASSERT(c)   /* use only *after* variable declarations */


/* **************************************************************
*  Required declarations
****************************************************************/
typedef struct nodeElt_s {
    U32 count;
    U16 parent;
    BYTE byte;
    BYTE nbBits;
} nodeElt;


/* **************************************************************
*  Debug Traces
****************************************************************/

#if DEBUGLEVEL >= 2

static size_t showU32(const U32* arr, size_t size)
{
    size_t u;
    for (u=0; u<size; u++) {
        RAWLOG(6, " %u", arr[u]); (void)arr;
    }
    RAWLOG(6, " \n");
    return size;
}

static size_t HUF_getNbBits(HUF_CElt elt);

static size_t showCTableBits(const HUF_CElt* ctable, size_t size)
{
    size_t u;
    for (u=0; u<size; u++) {
        RAWLOG(6, " %zu", HUF_getNbBits(ctable[u])); (void)ctable;
    }
    RAWLOG(6, " \n");
    return size;

}

static size_t showHNodeSymbols(const nodeElt* hnode, size_t size)
{
    size_t u;
    for (u=0; u<size; u++) {
        RAWLOG(6, " %u", hnode[u].byte); (void)hnode;
    }
    RAWLOG(6, " \n");
    return size;
}

static size_t showHNodeBits(const nodeElt* hnode, size_t size)
{
    size_t u;
    for (u=0; u<size; u++) {
        RAWLOG(6, " %u", hnode[u].nbBits); (void)hnode;
    }
    RAWLOG(6, " \n");
    return size;
}

#endif


/* *******************************************************
*  HUF : Huffman block compression
*********************************************************/
#define HUF_WORKSPACE_MAX_ALIGNMENT 8

static void* HUF_alignUpWorkspace(void* workspace, size_t* workspaceSizePtr, size_t align)
{
    size_t const mask = align - 1;
    size_t const rem = (size_t)workspace & mask;
    size_t const add = (align - rem) & mask;
    BYTE* const aligned = (BYTE*)workspace + add;
    assert((align & (align - 1)) == 0); /* pow 2 */
    assert(align <= HUF_WORKSPACE_MAX_ALIGNMENT);
    if (*workspaceSizePtr >= add) {
        assert(add < align);
        assert(((size_t)aligned & mask) == 0);
        *workspaceSizePtr -= add;
        return aligned;
    } else {
        *workspaceSizePtr = 0;
        return NULL;
    }
}


/* HUF_compressWeights() :
 * Same as FSE_compress(), but dedicated to huff0's weights compression.
 * The use case needs much less stack memory.
 * Note : all elements within weightTable are supposed to be <= HUF_TABLELOG_MAX.
 */
#define MAX_FSE_TABLELOG_FOR_HUFF_HEADER 6

typedef struct {
    FSE_CTable CTable[FSE_CTABLE_SIZE_U32(MAX_FSE_TABLELOG_FOR_HUFF_HEADER, HUF_TABLELOG_MAX)];
    U32 scratchBuffer[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(HUF_TABLELOG_MAX, MAX_FSE_TABLELOG_FOR_HUFF_HEADER)];
    unsigned count[HUF_TABLELOG_MAX+1];
    S16 norm[HUF_TABLELOG_MAX+1];
} HUF_CompressWeightsWksp;

static size_t
HUF_compressWeights(void* dst, size_t dstSize,
              const void* weightTable, size_t wtSize,
                    void* workspace, size_t workspaceSize)
{
    BYTE* const ostart = (BYTE*) dst;
    BYTE* op = ostart;
    BYTE* const oend = ostart + dstSize;

    unsigned maxSymbolValue = HUF_TABLELOG_MAX;
    U32 tableLog = MAX_FSE_TABLELOG_FOR_HUFF_HEADER;
    HUF_CompressWeightsWksp* wksp = (HUF_CompressWeightsWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));

    if (workspaceSize < sizeof(HUF_CompressWeightsWksp)) return ERROR(GENERIC);

    /* init conditions */
    if (wtSize <= 1) return 0;  /* Not compressible */

    /* Scan input and build symbol stats */
    {   unsigned const maxCount = HIST_count_simple(wksp->count, &maxSymbolValue, weightTable, wtSize);   /* never fails */
        if (maxCount == wtSize) return 1;   /* only a single symbol in src : rle */
        if (maxCount == 1) return 0;        /* each symbol present maximum once => not compressible */
    }

    tableLog = FSE_optimalTableLog(tableLog, wtSize, maxSymbolValue);
    CHECK_F( FSE_normalizeCount(wksp->norm, tableLog, wksp->count, wtSize, maxSymbolValue, /* useLowProbCount */ 0) );

    /* Write table description header */
    {   CHECK_V_F(hSize, FSE_writeNCount(op, (size_t)(oend-op), wksp->norm, maxSymbolValue, tableLog) );
        op += hSize;
    }

    /* Compress */
    CHECK_F( FSE_buildCTable_wksp(wksp->CTable, wksp->norm, maxSymbolValue, tableLog, wksp->scratchBuffer, sizeof(wksp->scratchBuffer)) );
    {   CHECK_V_F(cSize, FSE_compress_usingCTable(op, (size_t)(oend - op), weightTable, wtSize, wksp->CTable) );
        if (cSize == 0) return 0;   /* not enough space for compressed data */
        op += cSize;
    }

    return (size_t)(op-ostart);
}

static size_t HUF_getNbBits(HUF_CElt elt)
{
    return elt & 0xFF;
}

static size_t HUF_getNbBitsFast(HUF_CElt elt)
{
    return elt;
}

static size_t HUF_getValue(HUF_CElt elt)
{
    return elt & ~(size_t)0xFF;
}

static size_t HUF_getValueFast(HUF_CElt elt)
{
    return elt;
}

static void HUF_setNbBits(HUF_CElt* elt, size_t nbBits)
{
    assert(nbBits <= HUF_TABLELOG_ABSOLUTEMAX);
    *elt = nbBits;
}

static void HUF_setValue(HUF_CElt* elt, size_t value)
{
    size_t const nbBits = HUF_getNbBits(*elt);
    if (nbBits > 0) {
        assert((value >> nbBits) == 0);
        *elt |= value << (sizeof(HUF_CElt) * 8 - nbBits);
    }
}

HUF_CTableHeader HUF_readCTableHeader(HUF_CElt const* ctable)
{
    HUF_CTableHeader header;
    ZSTD_memcpy(&header, ctable, sizeof(header));
    return header;
}

static void HUF_writeCTableHeader(HUF_CElt* ctable, U32 tableLog, U32 maxSymbolValue)
{
    HUF_CTableHeader header;
    HUF_STATIC_ASSERT(sizeof(ctable[0]) == sizeof(header));
    ZSTD_memset(&header, 0, sizeof(header));
    assert(tableLog < 256);
    header.tableLog = (BYTE)tableLog;
    assert(maxSymbolValue < 256);
    header.maxSymbolValue = (BYTE)maxSymbolValue;
    ZSTD_memcpy(ctable, &header, sizeof(header));
}

typedef struct {
    HUF_CompressWeightsWksp wksp;
    BYTE bitsToWeight[HUF_TABLELOG_MAX + 1];   /* precomputed conversion table */
    BYTE huffWeight[HUF_SYMBOLVALUE_MAX];
} HUF_WriteCTableWksp;

size_t HUF_writeCTable_wksp(void* dst, size_t maxDstSize,
                            const HUF_CElt* CTable, unsigned maxSymbolValue, unsigned huffLog,
                            void* workspace, size_t workspaceSize)
{
    HUF_CElt const* const ct = CTable + 1;
    BYTE* op = (BYTE*)dst;
    U32 n;
    HUF_WriteCTableWksp* wksp = (HUF_WriteCTableWksp*)HUF_alignUpWorkspace(workspace, &workspaceSize, ZSTD_ALIGNOF(U32));

    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE >= sizeof(HUF_WriteCTableWksp));

    assert(HUF_readCTableHeader(CTable).maxSymbolValue == maxSymbolValue);
    assert(HUF_readCTableHeader(CTable).tableLog == huffLog);

    /* check conditions */
    if (workspaceSize < sizeof(HUF_WriteCTableWksp)) return ERROR(GENERIC);
    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);

    /* convert to weight */
    wksp->bitsToWeight[0] = 0;
    for (n=1; n<huffLog+1; n++)
        wksp->bitsToWeight[n] = (BYTE)(huffLog + 1 - n);
    for (n=0; n<maxSymbolValue; n++)
        wksp->huffWeight[n] = wksp->bitsToWeight[HUF_getNbBits(ct[n])];

    /* attempt weights compression by FSE */
    if (maxDstSize < 1) return ERROR(dstSize_tooSmall);
    {   CHECK_V_F(hSize, HUF_compressWeights(op+1, maxDstSize-1, wksp->huffWeight, maxSymbolValue, &wksp->wksp, sizeof(wksp->wksp)) );
        if ((hSize>1) & (hSize < maxSymbolValue/2)) {   /* FSE compressed */
            op[0] = (BYTE)hSize;
            return hSize+1;
    }   }

    /* write raw values as 4-bits (max : 15) */
    if (maxSymbolValue > (256-128)) return ERROR(GENERIC);   /* should not happen : likely means source cannot be compressed */
    if (((maxSymbolValue+1)/2) + 1 > maxDstSize) return ERROR(dstSize_tooSmall);   /* not enough space within dst buffer */
    op[0] = (BYTE)(128 /*special case*/ + (maxSymbolValue-1));
    wksp->huffWeight[maxSymbolValue] = 0;   /* to be sure it doesn't cause msan issue in final combination */
    for (n=0; n<maxSymbolValue; n+=2)
        op[(n/2)+1] = (BYTE)((wksp->huffWeight[n] << 4) + wksp->huffWeight[n+1]);
    return ((maxSymbolValue+1)/2) + 1;
}


size_t HUF_readCTable (HUF_CElt* CTable, unsigned* maxSymbolValuePtr, const void* src, size_t srcSize, unsigned* hasZeroWeights)
{
    BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];   /* init not required, even though some static analyzer may complain */
    U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];   /* large enough for values from 0 to 16 */
    U32 tableLog = 0;
    U32 nbSymbols = 0;
    HUF_CElt* const ct = CTable + 1;

    /* get symbol weights */
    CHECK_V_F(readSize, HUF_readStats(huffWeight, HUF_SYMBOLVALUE_MAX+1, rankVal, &nbSymbols, &tableLog, src, srcSize));
    *hasZeroWeights = (rankVal[0] > 0);

    /* check result */
    if (tableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
    if (nbSymbols > *maxSymbolValuePtr+1) return ERROR(maxSymbolValue_tooSmall);

    *maxSymbolValuePtr = nbSymbols - 1;

    HUF_writeCTableHeader(CTable, tableLog, *maxSymbolValuePtr);

    /* Prepare base value per rank */
    {   U32 n, nextRankStart = 0;
        for (n=1; n<=tableLog; n++) {
            U32 curr = nextRankStart;
            nextRankStart += (rankVal[n] << (n-1));
            rankVal[n] = curr;
    }   }

    /* fill nbBits */
    {   U32 n; for (n=0; n<nbSymbols; n++) {
            const U32 w = huffWeight[n];
            HUF_setNbBits(ct + n, (BYTE)(tableLog + 1 - w) & -(w != 0));
    }   }

    /* fill val */
    {   U16 nbPerRank[HUF_TABLELOG_MAX+2]  = {0};  /* support w=0=>n=tableLog+1 */
        U16 valPerRank[HUF_TABLELOG_MAX+2] = {0};
        { U32 n; for (n=0; n<nbSymbols; n++) nbPerRank[HUF_getNbBits(ct[n])]++; }
        /* determine stating value per rank */
        valPerRank[tableLog+1] = 0;   /* for w==0 */
        {   U16 min = 0;
            U32 n; for (n=tableLog; n>0; n--) {  /* start at n=tablelog <-> w=1 */
                valPerRank[n] = min;     /* get starting value within each rank */
                min += nbPerRank[n];
                min >>= 1;
        }   }
        /* assign value within rank, symbol order */
        { U32 n; for (n=0; n<nbSymbols; n++) HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++); }
    }

    return readSize;
}

U32 HUF_getNbBitsFromCTable(HUF_CElt const* CTable, U32 symbolValue)
{
    const HUF_CElt* const ct = CTable + 1;
    assert(symbolValue <= HUF_SYMBOLVALUE_MAX);
    if (symbolValue > HUF_readCTableHeader(CTable).maxSymbolValue)
        return 0;
    return (U32)HUF_getNbBits(ct[symbolValue]);
}


/**
 * HUF_setMaxHeight():
 * Try to enforce @targetNbBits on the Huffman tree described in @huffNode.
 *
 * It attempts to convert all nodes with nbBits > @targetNbBits
 * to employ @targetNbBits instead. Then it adjusts the tree
 * so that it remains a valid canonical Huffman tree.
 *
 * @pre               The sum of the ranks of each symbol == 2^largestBits,
 *                    where largestBits == huffNode[lastNonNull].nbBits.
 * @post              The sum of the ranks of each symbol == 2^largestBits,
 *                    where largestBits is the return value (expected <= targetNbBits).
 *
 * @param huffNode    The Huffman tree modified in place to enforce targetNbBits.
 *                    It's presumed sorted, from most frequent to rarest symbol.
 * @param lastNonNull The symbol with the lowest count in the Huffman tree.
 * @param targetNbBits  The allowed number of bits, which the Huffman tree
 *                    may not respect. After this function the Huffman tree will
 *                    respect targetNbBits.
 * @return            The maximum number of bits of the Huffman tree after adjustment.
 */
static U32 HUF_setMaxHeight(nodeElt* huffNode, U32 lastNonNull, U32 targetNbBits)
{
    const U32 largestBits = huffNode[lastNonNull].nbBits;
    /* early exit : no elt > targetNbBits, so the tree is already valid. */
    if (largestBits <= targetNbBits) return largestBits;

    DEBUGLOG(5, "HUF_setMaxHeight (targetNbBits = %u)", targetNbBits);

    /* there are several too large elements (at least >= 2) */
    {   int totalCost = 0;
        const U32 baseCost = 1 << (largestBits - targetNbBits);
        int n = (int)lastNonNull;

        /* Adjust any ranks > targetNbBits to targetNbBits.
         * Compute totalCost, which is how far the sum of the ranks is
         * we are over 2^largestBits after adjust the offending ranks.
         */
        while (huffNode[n].nbBits > targetNbBits) {
            totalCost += baseCost - (1 << (largestBits - huffNode[n].nbBits));
            huffNode[n].nbBits = (BYTE)targetNbBits;
            n--;
        }
        /* n stops at huffNode[n].nbBits <= targetNbBits */
        assert(huffNode[n].nbBits <= targetNbBits);
        /* n end at index of smallest symbol using < targetNbBits */
        while (huffNode[n].nbBits == targetNbBits) --n;

        /* renorm totalCost from 2^largestBits to 2^targetNbBits
         * note : totalCost is necessarily a multiple of baseCost */
        assert(((U32)totalCost & (baseCost - 1)) == 0);
        totalCost >>= (largestBits - targetNbBits);
        assert(totalCost > 0);

        /* repay normalized cost */
        {   U32 const noSymbol = 0xF0F0F0F0;
            U32 rankLast[HUF_TABLELOG_MAX+2];

            /* Get pos of last (smallest = lowest cum. count) symbol per rank */
            ZSTD_memset(rankLast, 0xF0, sizeof(rankLast));
            {   U32 currentNbBits = targetNbBits;
                int pos;
                for (pos=n ; pos >= 0; pos--) {
                    if (huffNode[pos].nbBits >= currentNbBits) continue;
                    currentNbBits = huffNode[pos].nbBits;   /* < targetNbBits */
                    rankLast[targetNbBits-currentNbBits] = (U32)pos;
            }   }

            while (totalCost > 0) {
                /* Try to reduce the next power of 2 above totalCost because we
                 * gain back half the rank.
                 */
                U32 nBitsToDecrease = ZSTD_highbit32((U32)totalCost) + 1;
                for ( ; nBitsToDecrease > 1; nBitsToDecrease--) {
                    U32 const highPos = rankLast[nBitsToDecrease];
                    U32 const lowPos = rankLast[nBitsToDecrease-1];
                    if (highPos == noSymbol) continue;
                    /* Decrease highPos if no symbols of lowPos or if it is
                     * not cheaper to remove 2 lowPos than highPos.
                     */
                    if (lowPos == noSymbol) break;
                    {   U32 const highTotal = huffNode[highPos].count;
                        U32 const lowTotal = 2 * huffNode[lowPos].count;
                        if (highTotal <= lowTotal) break;
                }   }
                /* only triggered when no more rank 1 symbol left => find closest one (note : there is necessarily at least one !) */
                assert(rankLast[nBitsToDecrease] != noSymbol || nBitsToDecrease == 1);
                /* HUF_MAX_TABLELOG test just to please gcc 5+; but it should not be necessary */
                while ((nBitsToDecrease<=HUF_TABLELOG_MAX) && (rankLast[nBitsToDecrease] == noSymbol))
                    nBitsToDecrease++;
                assert(rankLast[nBitsToDecrease] != noSymbol);
                /* Increase the number of bits to gain back half the rank cost. */
                totalCost -= 1 << (nBitsToDecrease-1);
                huffNode[rankLast[nBitsToDecrease]].nbBits++;

                /* Fix up the new rank.
                 * If the new rank was empty, this symbol is now its smallest.
                 * Otherwise, this symbol will be the largest in the new rank so no adjustment.
                 */
                if (rankLast[nBitsToDecrease-1] == noSymbol)
                    rankLast[nBitsToDecrease-1] = rankLast[nBitsToDecrease];
                /* Fix up the old rank.
                 * If the symbol was at position 0, meaning it was the highest weight symbol in the tree,
                 * it must be the only symbol in its rank, so the old rank now has no symbols.
                 * Otherwise, since the Huffman nodes are sorted by count, the previous position is now
                 * the smallest node in the rank. If the previous position belongs to a different rank,
                 * then the rank is now empty.
                 */
                if (rankLast[nBitsToDecrease] == 0)    /* special case, reached largest symbol */
                    rankLast[nBitsToDecrease] = noSymbol;
                else {
                    rankLast[nBitsToDecrease]--;
                    if (huffNode[rankLast[nBitsToDecrease]].nbBits != targetNbBits-nBitsToDecrease)
                        rankLast[nBitsToDecrease] = noSymbol;   /* this rank is now empty */
                }
            }   /* while (totalCost > 0) */

            /* If we've removed too much weight, then we have to add it back.
             * To avoid overshooting again, we only adjust the smallest rank.
             * We take the largest nodes from the lowest rank 0 and move them
             * to rank 1. There's guaranteed to be enough rank 0 symbols because
             * TODO.
             */
            while (totalCost < 0) {  /* Sometimes, cost correction overshoot */
                /* special case : no rank 1 symbol (using targetNbBits-1);
                 * let's create one from largest rank 0 (using targetNbBits).
                 */
                if (rankLast[1] == noSymbol) {
                    while (huffNode[n].nbBits == targetNbBits) n--;
                    huffNode[n+1].nbBits--;
                    assert(n >= 0);
                    rankLast[1] = (U32)(n+1);
                    totalCost++;
                    continue;
                }
                huffNode[ rankLast[1] + 1 ].nbBits--;
                rankLast[1]++;
                totalCost ++;
            }
        }   /* repay normalized cost */
    }   /* there are several too large elements (at least >= 2) */

    return targetNbBits;
}

typedef struct {
    U16 base;
    U16 curr;
} rankPos;

typedef nodeElt huffNodeTable[2 * (HUF_SYMBOLVALUE_MAX + 1)];

/* Number of buckets available for HUF_sort() */
#define RANK_POSITION_TABLE_SIZE 192

typedef struct {
  huffNodeTable huffNodeTbl;
  rankPos rankPosition[RANK_POSITION_TABLE_SIZE];
} HUF_buildCTable_wksp_tables;

/* RANK_POSITION_DISTINCT_COUNT_CUTOFF == Cutoff point in HUF_sort() buckets for which we use log2 bucketing.
 * Strategy is to use as many buckets as possible for representing distinct
 * counts while using the remainder to represent all "large" counts.
 *
 * To satisfy this requirement for 192 buckets, we can do the following:
 * Let buckets 0-166 represent distinct counts of [0, 166]
 * Let buckets 166 to 192 represent all remaining counts up to RANK_POSITION_MAX_COUNT_LOG using log2 bucketing.
 */
#define RANK_POSITION_MAX_COUNT_LOG 32
#define RANK_POSITION_LOG_BUCKETS_BEGIN ((RANK_POSITION_TABLE_SIZE - 1) - RANK_POSITION_MAX_COUNT_LOG - 1 /* == 158 */)
#define RANK_POSITION_DISTINCT_COUNT_CUTOFF (RANK_POSITION_LOG_BUCKETS_BEGIN + ZSTD_highbit32(RANK_POSITION_LOG_BUCKETS_BEGIN) /* == 166 */)

/* Return the appropriate bucket index for a given count. See definition of
 * RANK_POSITION_DISTINCT_COUNT_CUTOFF for explanation of bucketing strategy.
 */
static U32 HUF_getIndex(U32 const count) {
    return (count < RANK_POSITION_DISTINCT_COUNT_CUTOFF)
        ? count
        : ZSTD_highbit32(count) + RANK_POSITION_LOG_BUCKETS_BEGIN;
}

/* Helper swap function for HUF_quickSortPartition() */
static void HUF_swapNodes(nodeElt* a, nodeElt* b) {
	nodeElt tmp = *a;
	*a = *b;
	*b = tmp;
}

/* Returns 0 if the huffNode array is not sorted by descending count */
MEM_STATIC int HUF_isSorted(nodeElt huffNode[], U32 const maxSymbolValue1) {
    U32 i;
    for (i = 1; i < maxSymbolValue1; ++i) {
        if (huffNode[i].count > huffNode[i-1].count) {
            return 0;
        }
    }
    return 1;
}

/* Insertion sort by descending order */
HINT_INLINE void HUF_insertionSort(nodeElt huffNode[], int const low, int const high) {
    int i;
    int const size = high-low+1;
    huffNode += low;
    for (i = 1; i < size; ++i) {
        nodeElt const key = huffNode[i];
        int j = i - 1;
        while (j >= 0 && huffNode[j].count < key.count) {
            huffNode[j + 1] = huffNode[j];
            j--;
        }
        huffNode[j + 1] = key;
    }
}

/* Pivot helper function for quicksort. */
static int HUF_quickSortPartition(nodeElt arr[], int const low, int const high) {
    /* Simply select rightmost element as pivot. "Better" selectors like
     * median-of-three don't experimentally appear to have any benefit.
     */
    U32 const pivot = arr[high].count;
    int i = low - 1;
    int j = low;
    for ( ; j < high; j++) {
        if (arr[j].count > pivot) {
            i++;
            HUF_swapNodes(&arr[i], &arr[j]);
        }
    }
    HUF_swapNodes(&arr[i + 1], &arr[high]);
    return i + 1;
}

/* Classic quicksort by descending with partially iterative calls
 * to reduce worst case callstack size.
 */
static void HUF_simpleQuickSort(nodeElt arr[], int low, int high) {
    int const kInsertionSortThreshold = 8;
    if (high - low < kInsertionSortThreshold) {
        HUF_insertionSort(arr, low, high);
        return;
    }
    while (low < high) {
        int const idx = HUF_quickSortPartition(arr, low, high);
        if (idx - low < high - idx) {
            HUF_simpleQuickSort(arr, low, idx - 1);
            low = idx + 1;
        } else {
            HUF_simpleQuickSort(arr, idx + 1, high);
            high = idx - 1;
        }
    }
}

/**
 * HUF_sort():
 * Sorts the symbols [0, maxSymbolValue] by count[symbol] in decreasing order.
 * This is a typical bucket sorting strategy that uses either quicksort or insertion sort to sort each bucket.
 *
 * @param[out] huffNode       Sorted symbols by decreasing count. Only members `.count` and `.byte` are filled.
 *                            Must have (maxSymbolValue + 1) entries.
 * @param[in]  count          Histogram of the symbols.
 * @param[in]  maxSymbolValue Maximum symbol value.
 * @param      rankPosition   This is a scratch workspace. Must have RANK_POSITION_TABLE_SIZE entries.
 */
static void HUF_sort(nodeElt huffNode[], const unsigned count[], U32 const maxSymbolValue, rankPos rankPosition[]) {
    U32 n;
    U32 const maxSymbolValue1 = maxSymbolValue+1;

    /* Compute base and set curr to base.
     * For symbol s let lowerRank = HUF_getIndex(count[n]) and rank = lowerRank + 1.
     * See HUF_getIndex to see bucketing strategy.
     * We attribute each symbol to lowerRank's base value, because we want to know where
     * each rank begins in the output, so for rank R we want to count ranks R+1 and above.
     */
    ZSTD_memset(rankPosition, 0, sizeof(*rankPosition) * RANK_POSITION_TABLE_SIZE);
    for (n = 0; n < maxSymbolValue1; ++n) {
        U32 lowerRank = HUF_getIndex(count[n]);
        assert(lowerRank < RANK_POSITION_TABLE_SIZE - 1);
        rankPosition[lowerRank].base++;
    }

    assert(rankPosition[RANK_POSITION_TABLE_SIZE - 1].base == 0);
    /* Set up the rankPosition table */
    for (n = RANK_POSITION_TABLE_SIZE - 1; n > 0; --n) {
        rankPosition[n-1].base += rankPosition[n].base;
        rankPosition[n-1].curr = rankPosition[n-1].base;
    }

    /* Insert each symbol into their appropriate bucket, setting up rankPosition table. */
    for (n = 0; n < maxSymbolValue1; ++n) {
        U32 const c = count[n];
        U32 const r = HUF_getIndex(c) + 1;
        U32 const pos = rankPosition[r].curr++;
        assert(pos < maxSymbolValue1);
        huffNode[pos].count = c;
        huffNode[pos].byte  = (BYTE)n;
    }

    /* Sort each bucket. */
    for (n = RANK_POSITION_DISTINCT_COUNT_CUTOFF; n < RANK_POSITION_TABLE_SIZE - 1; ++n) {
        int const bucketSize = rankPosition[n].curr - rankPosition[n].base;
        U32 const bucketStartIdx = rankPosition[n].base;
        if (bucketSize > 1) {
            assert(bucketStartIdx < maxSymbolValue1);
            HUF_simpleQuickSort(huffNode + bucketStartIdx, 0, bucketSize-1);
        }
    }

    assert(HUF_isSorted(huffNode, maxSymbolValue1));
}


/** HUF_buildCTable_wksp() :
 *  Same as HUF_buildCTable(), but using externally allocated scratch buffer.
 *  `workSpace` must be aligned on 4-bytes boundaries, and be at least as large as sizeof(HUF_buildCTable_wksp_tables).
 */
#define STARTNODE (HUF_SYMBOLVALUE_MAX+1)

/* HUF_buildTree():
 * Takes the huffNode array sorted by HUF_sort() and builds an unlimited-depth Huffman tree.
 *
 * @param huffNode        The array sorted by HUF_sort(). Builds the Huffman tree in this array.
 * @param maxSymbolValue  The maximum symbol value.
 * @return                The smallest node in the Huffman tree (by count).
 */
static int HUF_buildTree(nodeElt* huffNode, U32 maxSymbolValue)
{
    nodeElt* const huffNode0 = huffNode - 1;
    int nonNullRank;
    int lowS, lowN;
    int nodeNb = STARTNODE;
    int n, nodeRoot;
    DEBUGLOG(5, "HUF_buildTree (alphabet size = %u)", maxSymbolValue + 1);
    /* init for parents */
    nonNullRank = (int)maxSymbolValue;
    while(huffNode[nonNullRank].count == 0) nonNullRank--;
    lowS = nonNullRank; nodeRoot = nodeNb + lowS - 1; lowN = nodeNb;
    huffNode[nodeNb].count = huffNode[lowS].count + huffNode[lowS-1].count;
    huffNode[lowS].parent = huffNode[lowS-1].parent = (U16)nodeNb;
    nodeNb++; lowS-=2;
    for (n=nodeNb; n<=nodeRoot; n++) huffNode[n].count = (U32)(1U<<30);
    huffNode0[0].count = (U32)(1U<<31);  /* fake entry, strong barrier */

    /* create parents */
    while (nodeNb <= nodeRoot) {
        int const n1 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
        int const n2 = (huffNode[lowS].count < huffNode[lowN].count) ? lowS-- : lowN++;
        huffNode[nodeNb].count = huffNode[n1].count + huffNode[n2].count;
        huffNode[n1].parent = huffNode[n2].parent = (U16)nodeNb;
        nodeNb++;
    }

    /* distribute weights (unlimited tree height) */
    huffNode[nodeRoot].nbBits = 0;
    for (n=nodeRoot-1; n>=STARTNODE; n--)
        huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;
    for (n=0; n<=nonNullRank; n++)
        huffNode[n].nbBits = huffNode[ huffNode[n].parent ].nbBits + 1;

    DEBUGLOG(6, "Initial distribution of bits completed (%zu sorted symbols)", showHNodeBits(huffNode, maxSymbolValue+1));

    return nonNullRank;
}

/**
 * HUF_buildCTableFromTree():
 * Build the CTable given the Huffman tree in huffNode.
 *
 * @param[out] CTable         The output Huffman CTable.
 * @param      huffNode       The Huffman tree.
 * @param      nonNullRank    The last and smallest node in the Huffman tree.
 * @param      maxSymbolValue The maximum symbol value.
 * @param      maxNbBits      The exact maximum number of bits used in the Huffman tree.
 */
static void HUF_buildCTableFromTree(HUF_CElt* CTable, nodeElt const* huffNode, int nonNullRank, U32 maxSymbolValue, U32 maxNbBits)
{
    HUF_CElt* const ct = CTable + 1;
    /* fill result into ctable (val, nbBits) */
    int n;
    U16 nbPerRank[HUF_TABLELOG_MAX+1] = {0};
    U16 valPerRank[HUF_TABLELOG_MAX+1] = {0};
    int const alphabetSize = (int)(maxSymbolValue + 1);
    for (n=0; n<=nonNullRank; n++)
        nbPerRank[huffNode[n].nbBits]++;
    /* determine starting value per rank */
    {   U16 min = 0;
        for (n=(int)maxNbBits; n>0; n--) {
            valPerRank[n] = min;      /* get starting value within each rank */
            min += nbPerRank[n];
            min >>= 1;
    }   }
    for (n=0; n<alphabetSize; n++)
        HUF_setNbBits(ct + huffNode[n].byte, huffNode[n].nbBits);   /* push nbBits per symbol, symbol order */
    for (n=0; n<alphabetSize; n++)
        HUF_setValue(ct + n, valPerRank[HUF_getNbBits(ct[n])]++);   /* assign value within rank, symbol order */

    HUF_writeCTableHeader(CTable, maxNbBits, maxSymbolValue);
}

size_t
HUF_buildCTable_wksp(HUF_CElt* CTable, const unsigned* count, U32 maxSymbolValue, U32 maxNbBits,
                     void* workSpace, size_t wkspSize)
{
    HUF_buildCTable_wksp_tables* const wksp_tables =
        (HUF_buildCTable_wksp_tables*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(U32));
    nodeElt* const huffNode0 = wksp_tables->huffNodeTbl;
    nodeElt* const huffNode = huffNode0+1;
    int nonNullRank;

    HUF_STATIC_ASSERT(HUF_CTABLE_WORKSPACE_SIZE == sizeof(HUF_buildCTable_wksp_tables));

    DEBUGLOG(5, "HUF_buildCTable_wksp (alphabet size = %u)", maxSymbolValue+1);

    /* safety checks */
    if (wkspSize < sizeof(HUF_buildCTable_wksp_tables))
        return ERROR(workSpace_tooSmall);
    if (maxNbBits == 0) maxNbBits = HUF_TABLELOG_DEFAULT;
    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX)
        return ERROR(maxSymbolValue_tooLarge);
    ZSTD_memset(huffNode0, 0, sizeof(huffNodeTable));

    /* sort, decreasing order */
    HUF_sort(huffNode, count, maxSymbolValue, wksp_tables->rankPosition);
    DEBUGLOG(6, "sorted symbols completed (%zu symbols)", showHNodeSymbols(huffNode, maxSymbolValue+1));

    /* build tree */
    nonNullRank = HUF_buildTree(huffNode, maxSymbolValue);

    /* determine and enforce maxTableLog */
    maxNbBits = HUF_setMaxHeight(huffNode, (U32)nonNullRank, maxNbBits);
    if (maxNbBits > HUF_TABLELOG_MAX) return ERROR(GENERIC);   /* check fit into table */

    HUF_buildCTableFromTree(CTable, huffNode, nonNullRank, maxSymbolValue, maxNbBits);

    return maxNbBits;
}

size_t HUF_estimateCompressedSize(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue)
{
    HUF_CElt const* ct = CTable + 1;
    size_t nbBits = 0;
    int s;
    for (s = 0; s <= (int)maxSymbolValue; ++s) {
        nbBits += HUF_getNbBits(ct[s]) * count[s];
    }
    return nbBits >> 3;
}

int HUF_validateCTable(const HUF_CElt* CTable, const unsigned* count, unsigned maxSymbolValue) {
    HUF_CTableHeader header = HUF_readCTableHeader(CTable);
    HUF_CElt const* ct = CTable + 1;
    int bad = 0;
    int s;

    assert(header.tableLog <= HUF_TABLELOG_ABSOLUTEMAX);

    if (header.maxSymbolValue < maxSymbolValue)
        return 0;

    for (s = 0; s <= (int)maxSymbolValue; ++s) {
        bad |= (count[s] != 0) & (HUF_getNbBits(ct[s]) == 0);
    }
    return !bad;
}

size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }

/** HUF_CStream_t:
 * Huffman uses its own BIT_CStream_t implementation.
 * There are three major differences from BIT_CStream_t:
 *   1. HUF_addBits() takes a HUF_CElt (size_t) which is
 *      the pair (nbBits, value) in the format:
 *      format:
 *        - Bits [0, 4)            = nbBits
 *        - Bits [4, 64 - nbBits)  = 0
 *        - Bits [64 - nbBits, 64) = value
 *   2. The bitContainer is built from the upper bits and
 *      right shifted. E.g. to add a new value of N bits
 *      you right shift the bitContainer by N, then or in
 *      the new value into the N upper bits.
 *   3. The bitstream has two bit containers. You can add
 *      bits to the second container and merge them into
 *      the first container.
 */

#define HUF_BITS_IN_CONTAINER (sizeof(size_t) * 8)

typedef struct {
    size_t bitContainer[2];
    size_t bitPos[2];

    BYTE* startPtr;
    BYTE* ptr;
    BYTE* endPtr;
} HUF_CStream_t;

/**! HUF_initCStream():
 * Initializes the bitstream.
 * @returns 0 or an error code.
 */
static size_t HUF_initCStream(HUF_CStream_t* bitC,
                                  void* startPtr, size_t dstCapacity)
{
    ZSTD_memset(bitC, 0, sizeof(*bitC));
    bitC->startPtr = (BYTE*)startPtr;
    bitC->ptr = bitC->startPtr;
    bitC->endPtr = bitC->startPtr + dstCapacity - sizeof(bitC->bitContainer[0]);
    if (dstCapacity <= sizeof(bitC->bitContainer[0])) return ERROR(dstSize_tooSmall);
    return 0;
}

/*! HUF_addBits():
 * Adds the symbol stored in HUF_CElt elt to the bitstream.
 *
 * @param elt   The element we're adding. This is a (nbBits, value) pair.
 *              See the HUF_CStream_t docs for the format.
 * @param idx   Insert into the bitstream at this idx.
 * @param kFast This is a template parameter. If the bitstream is guaranteed
 *              to have at least 4 unused bits after this call it may be 1,
 *              otherwise it must be 0. HUF_addBits() is faster when fast is set.
 */
FORCE_INLINE_TEMPLATE void HUF_addBits(HUF_CStream_t* bitC, HUF_CElt elt, int idx, int kFast)
{
    assert(idx <= 1);
    assert(HUF_getNbBits(elt) <= HUF_TABLELOG_ABSOLUTEMAX);
    /* This is efficient on x86-64 with BMI2 because shrx
     * only reads the low 6 bits of the register. The compiler
     * knows this and elides the mask. When fast is set,
     * every operation can use the same value loaded from elt.
     */
    bitC->bitContainer[idx] >>= HUF_getNbBits(elt);
    bitC->bitContainer[idx] |= kFast ? HUF_getValueFast(elt) : HUF_getValue(elt);
    /* We only read the low 8 bits of bitC->bitPos[idx] so it
     * doesn't matter that the high bits have noise from the value.
     */
    bitC->bitPos[idx] += HUF_getNbBitsFast(elt);
    assert((bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
    /* The last 4-bits of elt are dirty if fast is set,
     * so we must not be overwriting bits that have already been
     * inserted into the bit container.
     */
#if DEBUGLEVEL >= 1
    {
        size_t const nbBits = HUF_getNbBits(elt);
        size_t const dirtyBits = nbBits == 0 ? 0 : ZSTD_highbit32((U32)nbBits) + 1;
        (void)dirtyBits;
        /* Middle bits are 0. */
        assert(((elt >> dirtyBits) << (dirtyBits + nbBits)) == 0);
        /* We didn't overwrite any bits in the bit container. */
        assert(!kFast || (bitC->bitPos[idx] & 0xFF) <= HUF_BITS_IN_CONTAINER);
        (void)dirtyBits;
    }
#endif
}

FORCE_INLINE_TEMPLATE void HUF_zeroIndex1(HUF_CStream_t* bitC)
{
    bitC->bitContainer[1] = 0;
    bitC->bitPos[1] = 0;
}

/*! HUF_mergeIndex1() :
 * Merges the bit container @ index 1 into the bit container @ index 0
 * and zeros the bit container @ index 1.
 */
FORCE_INLINE_TEMPLATE void HUF_mergeIndex1(HUF_CStream_t* bitC)
{
    assert((bitC->bitPos[1] & 0xFF) < HUF_BITS_IN_CONTAINER);
    bitC->bitContainer[0] >>= (bitC->bitPos[1] & 0xFF);
    bitC->bitContainer[0] |= bitC->bitContainer[1];
    bitC->bitPos[0] += bitC->bitPos[1];
    assert((bitC->bitPos[0] & 0xFF) <= HUF_BITS_IN_CONTAINER);
}

/*! HUF_flushBits() :
* Flushes the bits in the bit container @ index 0.
*
* @post bitPos will be < 8.
* @param kFast If kFast is set then we must know a-priori that
*              the bit container will not overflow.
*/
FORCE_INLINE_TEMPLATE void HUF_flushBits(HUF_CStream_t* bitC, int kFast)
{
    /* The upper bits of bitPos are noisy, so we must mask by 0xFF. */
    size_t const nbBits = bitC->bitPos[0] & 0xFF;
    size_t const nbBytes = nbBits >> 3;
    /* The top nbBits bits of bitContainer are the ones we need. */
    size_t const bitContainer = bitC->bitContainer[0] >> (HUF_BITS_IN_CONTAINER - nbBits);
    /* Mask bitPos to account for the bytes we consumed. */
    bitC->bitPos[0] &= 7;
    assert(nbBits > 0);
    assert(nbBits <= sizeof(bitC->bitContainer[0]) * 8);
    assert(bitC->ptr <= bitC->endPtr);
    MEM_writeLEST(bitC->ptr, bitContainer);
    bitC->ptr += nbBytes;
    assert(!kFast || bitC->ptr <= bitC->endPtr);
    if (!kFast && bitC->ptr > bitC->endPtr) bitC->ptr = bitC->endPtr;
    /* bitContainer doesn't need to be modified because the leftover
     * bits are already the top bitPos bits. And we don't care about
     * noise in the lower values.
     */
}

/*! HUF_endMark()
 * @returns The Huffman stream end mark: A 1-bit value = 1.
 */
static HUF_CElt HUF_endMark(void)
{
    HUF_CElt endMark;
    HUF_setNbBits(&endMark, 1);
    HUF_setValue(&endMark, 1);
    return endMark;
}

/*! HUF_closeCStream() :
 *  @return Size of CStream, in bytes,
 *          or 0 if it could not fit into dstBuffer */
static size_t HUF_closeCStream(HUF_CStream_t* bitC)
{
    HUF_addBits(bitC, HUF_endMark(), /* idx */ 0, /* kFast */ 0);
    HUF_flushBits(bitC, /* kFast */ 0);
    {
        size_t const nbBits = bitC->bitPos[0] & 0xFF;
        if (bitC->ptr >= bitC->endPtr) return 0; /* overflow detected */
        return (size_t)(bitC->ptr - bitC->startPtr) + (nbBits > 0);
    }
}

FORCE_INLINE_TEMPLATE void
HUF_encodeSymbol(HUF_CStream_t* bitCPtr, U32 symbol, const HUF_CElt* CTable, int idx, int fast)
{
    HUF_addBits(bitCPtr, CTable[symbol], idx, fast);
}

FORCE_INLINE_TEMPLATE void
HUF_compress1X_usingCTable_internal_body_loop(HUF_CStream_t* bitC,
                                   const BYTE* ip, size_t srcSize,
                                   const HUF_CElt* ct,
                                   int kUnroll, int kFastFlush, int kLastFast)
{
    /* Join to kUnroll */
    int n = (int)srcSize;
    int rem = n % kUnroll;
    if (rem > 0) {
        for (; rem > 0; --rem) {
            HUF_encodeSymbol(bitC, ip[--n], ct, 0, /* fast */ 0);
        }
        HUF_flushBits(bitC, kFastFlush);
    }
    assert(n % kUnroll == 0);

    /* Join to 2 * kUnroll */
    if (n % (2 * kUnroll)) {
        int u;
        for (u = 1; u < kUnroll; ++u) {
            HUF_encodeSymbol(bitC, ip[n - u], ct, 0, 1);
        }
        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, 0, kLastFast);
        HUF_flushBits(bitC, kFastFlush);
        n -= kUnroll;
    }
    assert(n % (2 * kUnroll) == 0);

    for (; n>0; n-= 2 * kUnroll) {
        /* Encode kUnroll symbols into the bitstream @ index 0. */
        int u;
        for (u = 1; u < kUnroll; ++u) {
            HUF_encodeSymbol(bitC, ip[n - u], ct, /* idx */ 0, /* fast */ 1);
        }
        HUF_encodeSymbol(bitC, ip[n - kUnroll], ct, /* idx */ 0, /* fast */ kLastFast);
        HUF_flushBits(bitC, kFastFlush);
        /* Encode kUnroll symbols into the bitstream @ index 1.
         * This allows us to start filling the bit container
         * without any data dependencies.
         */
        HUF_zeroIndex1(bitC);
        for (u = 1; u < kUnroll; ++u) {
            HUF_encodeSymbol(bitC, ip[n - kUnroll - u], ct, /* idx */ 1, /* fast */ 1);
        }
        HUF_encodeSymbol(bitC, ip[n - kUnroll - kUnroll], ct, /* idx */ 1, /* fast */ kLastFast);
        /* Merge bitstream @ index 1 into the bitstream @ index 0 */
        HUF_mergeIndex1(bitC);
        HUF_flushBits(bitC, kFastFlush);
    }
    assert(n == 0);

}

/**
 * Returns a tight upper bound on the output space needed by Huffman
 * with 8 bytes buffer to handle over-writes. If the output is at least
 * this large we don't need to do bounds checks during Huffman encoding.
 */
static size_t HUF_tightCompressBound(size_t srcSize, size_t tableLog)
{
    return ((srcSize * tableLog) >> 3) + 8;
}


FORCE_INLINE_TEMPLATE size_t
HUF_compress1X_usingCTable_internal_body(void* dst, size_t dstSize,
                                   const void* src, size_t srcSize,
                                   const HUF_CElt* CTable)
{
    U32 const tableLog = HUF_readCTableHeader(CTable).tableLog;
    HUF_CElt const* ct = CTable + 1;
    const BYTE* ip = (const BYTE*) src;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstSize;
    HUF_CStream_t bitC;

    /* init */
    if (dstSize < 8) return 0;   /* not enough space to compress */
    { BYTE* op = ostart;
      size_t const initErr = HUF_initCStream(&bitC, op, (size_t)(oend-op));
      if (HUF_isError(initErr)) return 0; }

    if (dstSize < HUF_tightCompressBound(srcSize, (size_t)tableLog) || tableLog > 11)
        HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ MEM_32bits() ? 2 : 4, /* kFast */ 0, /* kLastFast */ 0);
    else {
        if (MEM_32bits()) {
            switch (tableLog) {
            case 11:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 0);
                break;
            case 10: ZSTD_FALLTHROUGH;
            case 9: ZSTD_FALLTHROUGH;
            case 8:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 2, /* kFastFlush */ 1, /* kLastFast */ 1);
                break;
            case 7: ZSTD_FALLTHROUGH;
            default:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 3, /* kFastFlush */ 1, /* kLastFast */ 1);
                break;
            }
        } else {
            switch (tableLog) {
            case 11:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 0);
                break;
            case 10:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 5, /* kFastFlush */ 1, /* kLastFast */ 1);
                break;
            case 9:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 6, /* kFastFlush */ 1, /* kLastFast */ 0);
                break;
            case 8:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 7, /* kFastFlush */ 1, /* kLastFast */ 0);
                break;
            case 7:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 8, /* kFastFlush */ 1, /* kLastFast */ 0);
                break;
            case 6: ZSTD_FALLTHROUGH;
            default:
                HUF_compress1X_usingCTable_internal_body_loop(&bitC, ip, srcSize, ct, /* kUnroll */ 9, /* kFastFlush */ 1, /* kLastFast */ 1);
                break;
            }
        }
    }
    assert(bitC.ptr <= bitC.endPtr);

    return HUF_closeCStream(&bitC);
}

#if DYNAMIC_BMI2

static BMI2_TARGET_ATTRIBUTE size_t
HUF_compress1X_usingCTable_internal_bmi2(void* dst, size_t dstSize,
                                   const void* src, size_t srcSize,
                                   const HUF_CElt* CTable)
{
    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
}

static size_t
HUF_compress1X_usingCTable_internal_default(void* dst, size_t dstSize,
                                      const void* src, size_t srcSize,
                                      const HUF_CElt* CTable)
{
    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
}

static size_t
HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                              const void* src, size_t srcSize,
                              const HUF_CElt* CTable, const int flags)
{
    if (flags & HUF_flags_bmi2) {
        return HUF_compress1X_usingCTable_internal_bmi2(dst, dstSize, src, srcSize, CTable);
    }
    return HUF_compress1X_usingCTable_internal_default(dst, dstSize, src, srcSize, CTable);
}

#else

static size_t
HUF_compress1X_usingCTable_internal(void* dst, size_t dstSize,
                              const void* src, size_t srcSize,
                              const HUF_CElt* CTable, const int flags)
{
    (void)flags;
    return HUF_compress1X_usingCTable_internal_body(dst, dstSize, src, srcSize, CTable);
}

#endif

size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
{
    return HUF_compress1X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
}

static size_t
HUF_compress4X_usingCTable_internal(void* dst, size_t dstSize,
                              const void* src, size_t srcSize,
                              const HUF_CElt* CTable, int flags)
{
    size_t const segmentSize = (srcSize+3)/4;   /* first 3 segments */
    const BYTE* ip = (const BYTE*) src;
    const BYTE* const iend = ip + srcSize;
    BYTE* const ostart = (BYTE*) dst;
    BYTE* const oend = ostart + dstSize;
    BYTE* op = ostart;

    if (dstSize < 6 + 1 + 1 + 1 + 8) return 0;   /* minimum space to compress successfully */
    if (srcSize < 12) return 0;   /* no saving possible : too small input */
    op += 6;   /* jumpTable */

    assert(op <= oend);
    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
        if (cSize == 0 || cSize > 65535) return 0;
        MEM_writeLE16(ostart, (U16)cSize);
        op += cSize;
    }

    ip += segmentSize;
    assert(op <= oend);
    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
        if (cSize == 0 || cSize > 65535) return 0;
        MEM_writeLE16(ostart+2, (U16)cSize);
        op += cSize;
    }

    ip += segmentSize;
    assert(op <= oend);
    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, segmentSize, CTable, flags) );
        if (cSize == 0 || cSize > 65535) return 0;
        MEM_writeLE16(ostart+4, (U16)cSize);
        op += cSize;
    }

    ip += segmentSize;
    assert(op <= oend);
    assert(ip <= iend);
    {   CHECK_V_F(cSize, HUF_compress1X_usingCTable_internal(op, (size_t)(oend-op), ip, (size_t)(iend-ip), CTable, flags) );
        if (cSize == 0 || cSize > 65535) return 0;
        op += cSize;
    }

    return (size_t)(op-ostart);
}

size_t HUF_compress4X_usingCTable(void* dst, size_t dstSize, const void* src, size_t srcSize, const HUF_CElt* CTable, int flags)
{
    return HUF_compress4X_usingCTable_internal(dst, dstSize, src, srcSize, CTable, flags);
}

typedef enum { HUF_singleStream, HUF_fourStreams } HUF_nbStreams_e;

static size_t HUF_compressCTable_internal(
                BYTE* const ostart, BYTE* op, BYTE* const oend,
                const void* src, size_t srcSize,
                HUF_nbStreams_e nbStreams, const HUF_CElt* CTable, const int flags)
{
    size_t const cSize = (nbStreams==HUF_singleStream) ?
                         HUF_compress1X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags) :
                         HUF_compress4X_usingCTable_internal(op, (size_t)(oend - op), src, srcSize, CTable, flags);
    if (HUF_isError(cSize)) { return cSize; }
    if (cSize==0) { return 0; }   /* uncompressible */
    op += cSize;
    /* check compressibility */
    assert(op >= ostart);
    if ((size_t)(op-ostart) >= srcSize-1) { return 0; }
    return (size_t)(op-ostart);
}

typedef struct {
    unsigned count[HUF_SYMBOLVALUE_MAX + 1];
    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(HUF_SYMBOLVALUE_MAX)];
    union {
        HUF_buildCTable_wksp_tables buildCTable_wksp;
        HUF_WriteCTableWksp writeCTable_wksp;
        U32 hist_wksp[HIST_WKSP_SIZE_U32];
    } wksps;
} HUF_compress_tables_t;

#define SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE 4096
#define SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO 10  /* Must be >= 2 */

unsigned HUF_cardinality(const unsigned* count, unsigned maxSymbolValue)
{
    unsigned cardinality = 0;
    unsigned i;

    for (i = 0; i < maxSymbolValue + 1; i++) {
        if (count[i] != 0) cardinality += 1;
    }

    return cardinality;
}

unsigned HUF_minTableLog(unsigned symbolCardinality)
{
    U32 minBitsSymbols = ZSTD_highbit32(symbolCardinality) + 1;
    return minBitsSymbols;
}

unsigned HUF_optimalTableLog(
            unsigned maxTableLog,
            size_t srcSize,
            unsigned maxSymbolValue,
            void* workSpace, size_t wkspSize,
            HUF_CElt* table,
      const unsigned* count,
            int flags)
{
    assert(srcSize > 1); /* Not supported, RLE should be used instead */
    assert(wkspSize >= sizeof(HUF_buildCTable_wksp_tables));

    if (!(flags & HUF_flags_optimalDepth)) {
        /* cheap evaluation, based on FSE */
        return FSE_optimalTableLog_internal(maxTableLog, srcSize, maxSymbolValue, 1);
    }

    {   BYTE* dst = (BYTE*)workSpace + sizeof(HUF_WriteCTableWksp);
        size_t dstSize = wkspSize - sizeof(HUF_WriteCTableWksp);
        size_t hSize, newSize;
        const unsigned symbolCardinality = HUF_cardinality(count, maxSymbolValue);
        const unsigned minTableLog = HUF_minTableLog(symbolCardinality);
        size_t optSize = ((size_t) ~0) - 1;
        unsigned optLog = maxTableLog, optLogGuess;

        DEBUGLOG(6, "HUF_optimalTableLog: probing huf depth (srcSize=%zu)", srcSize);

        /* Search until size increases */
        for (optLogGuess = minTableLog; optLogGuess <= maxTableLog; optLogGuess++) {
            DEBUGLOG(7, "checking for huffLog=%u", optLogGuess);

            {   size_t maxBits = HUF_buildCTable_wksp(table, count, maxSymbolValue, optLogGuess, workSpace, wkspSize);
                if (ERR_isError(maxBits)) continue;

                if (maxBits < optLogGuess && optLogGuess > minTableLog) break;

                hSize = HUF_writeCTable_wksp(dst, dstSize, table, maxSymbolValue, (U32)maxBits, workSpace, wkspSize);
            }

            if (ERR_isError(hSize)) continue;

            newSize = HUF_estimateCompressedSize(table, count, maxSymbolValue) + hSize;

            if (newSize > optSize + 1) {
                break;
            }

            if (newSize < optSize) {
                optSize = newSize;
                optLog = optLogGuess;
            }
        }
        assert(optLog <= HUF_TABLELOG_MAX);
        return optLog;
    }
}

/* HUF_compress_internal() :
 * `workSpace_align4` must be aligned on 4-bytes boundaries,
 * and occupies the same space as a table of HUF_WORKSPACE_SIZE_U64 unsigned */
static size_t
HUF_compress_internal (void* dst, size_t dstSize,
                 const void* src, size_t srcSize,
                       unsigned maxSymbolValue, unsigned huffLog,
                       HUF_nbStreams_e nbStreams,
                       void* workSpace, size_t wkspSize,
                       HUF_CElt* oldHufTable, HUF_repeat* repeat, int flags)
{
    HUF_compress_tables_t* const table = (HUF_compress_tables_t*)HUF_alignUpWorkspace(workSpace, &wkspSize, ZSTD_ALIGNOF(size_t));
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstSize;
    BYTE* op = ostart;

    DEBUGLOG(5, "HUF_compress_internal (srcSize=%zu)", srcSize);
    HUF_STATIC_ASSERT(sizeof(*table) + HUF_WORKSPACE_MAX_ALIGNMENT <= HUF_WORKSPACE_SIZE);

    /* checks & inits */
    if (wkspSize < sizeof(*table)) return ERROR(workSpace_tooSmall);
    if (!srcSize) return 0;  /* Uncompressed */
    if (!dstSize) return 0;  /* cannot fit anything within dst budget */
    if (srcSize > HUF_BLOCKSIZE_MAX) return ERROR(srcSize_wrong);   /* current block size limit */
    if (huffLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
    if (maxSymbolValue > HUF_SYMBOLVALUE_MAX) return ERROR(maxSymbolValue_tooLarge);
    if (!maxSymbolValue) maxSymbolValue = HUF_SYMBOLVALUE_MAX;
    if (!huffLog) huffLog = HUF_TABLELOG_DEFAULT;

    /* Heuristic : If old table is valid, use it for small inputs */
    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat == HUF_repeat_valid) {
        return HUF_compressCTable_internal(ostart, op, oend,
                                           src, srcSize,
                                           nbStreams, oldHufTable, flags);
    }

    /* If uncompressible data is suspected, do a smaller sampling first */
    DEBUG_STATIC_ASSERT(SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO >= 2);
    if ((flags & HUF_flags_suspectUncompressible) && srcSize >= (SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE * SUSPECT_INCOMPRESSIBLE_SAMPLE_RATIO)) {
        size_t largestTotal = 0;
        DEBUGLOG(5, "input suspected incompressible : sampling to check");
        {   unsigned maxSymbolValueBegin = maxSymbolValue;
            CHECK_V_F(largestBegin, HIST_count_simple (table->count, &maxSymbolValueBegin, (const BYTE*)src, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
            largestTotal += largestBegin;
        }
        {   unsigned maxSymbolValueEnd = maxSymbolValue;
            CHECK_V_F(largestEnd, HIST_count_simple (table->count, &maxSymbolValueEnd, (const BYTE*)src + srcSize - SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE, SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) );
            largestTotal += largestEnd;
        }
        if (largestTotal <= ((2 * SUSPECT_INCOMPRESSIBLE_SAMPLE_SIZE) >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
    }

    /* Scan input and build symbol stats */
    {   CHECK_V_F(largest, HIST_count_wksp (table->count, &maxSymbolValue, (const BYTE*)src, srcSize, table->wksps.hist_wksp, sizeof(table->wksps.hist_wksp)) );
        if (largest == srcSize) { *ostart = ((const BYTE*)src)[0]; return 1; }   /* single symbol, rle */
        if (largest <= (srcSize >> 7)+4) return 0;   /* heuristic : probably not compressible enough */
    }
    DEBUGLOG(6, "histogram detail completed (%zu symbols)", showU32(table->count, maxSymbolValue+1));

    /* Check validity of previous table */
    if ( repeat
      && *repeat == HUF_repeat_check
      && !HUF_validateCTable(oldHufTable, table->count, maxSymbolValue)) {
        *repeat = HUF_repeat_none;
    }
    /* Heuristic : use existing table for small inputs */
    if ((flags & HUF_flags_preferRepeat) && repeat && *repeat != HUF_repeat_none) {
        return HUF_compressCTable_internal(ostart, op, oend,
                                           src, srcSize,
                                           nbStreams, oldHufTable, flags);
    }

    /* Build Huffman Tree */
    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, &table->wksps, sizeof(table->wksps), table->CTable, table->count, flags);
    {   size_t const maxBits = HUF_buildCTable_wksp(table->CTable, table->count,
                                            maxSymbolValue, huffLog,
                                            &table->wksps.buildCTable_wksp, sizeof(table->wksps.buildCTable_wksp));
        CHECK_F(maxBits);
        huffLog = (U32)maxBits;
        DEBUGLOG(6, "bit distribution completed (%zu symbols)", showCTableBits(table->CTable + 1, maxSymbolValue+1));
    }

    /* Write table description header */
    {   CHECK_V_F(hSize, HUF_writeCTable_wksp(op, dstSize, table->CTable, maxSymbolValue, huffLog,
                                              &table->wksps.writeCTable_wksp, sizeof(table->wksps.writeCTable_wksp)) );
        /* Check if using previous huffman table is beneficial */
        if (repeat && *repeat != HUF_repeat_none) {
            size_t const oldSize = HUF_estimateCompressedSize(oldHufTable, table->count, maxSymbolValue);
            size_t const newSize = HUF_estimateCompressedSize(table->CTable, table->count, maxSymbolValue);
            if (oldSize <= hSize + newSize || hSize + 12 >= srcSize) {
                return HUF_compressCTable_internal(ostart, op, oend,
                                                   src, srcSize,
                                                   nbStreams, oldHufTable, flags);
        }   }

        /* Use the new huffman table */
        if (hSize + 12ul >= srcSize) { return 0; }
        op += hSize;
        if (repeat) { *repeat = HUF_repeat_none; }
        if (oldHufTable)
            ZSTD_memcpy(oldHufTable, table->CTable, sizeof(table->CTable));  /* Save new table */
    }
    return HUF_compressCTable_internal(ostart, op, oend,
                                       src, srcSize,
                                       nbStreams, table->CTable, flags);
}

size_t HUF_compress1X_repeat (void* dst, size_t dstSize,
                      const void* src, size_t srcSize,
                      unsigned maxSymbolValue, unsigned huffLog,
                      void* workSpace, size_t wkspSize,
                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
{
    DEBUGLOG(5, "HUF_compress1X_repeat (srcSize = %zu)", srcSize);
    return HUF_compress_internal(dst, dstSize, src, srcSize,
                                 maxSymbolValue, huffLog, HUF_singleStream,
                                 workSpace, wkspSize, hufTable,
                                 repeat, flags);
}

/* HUF_compress4X_repeat():
 * compress input using 4 streams.
 * consider skipping quickly
 * reuse an existing huffman compression table */
size_t HUF_compress4X_repeat (void* dst, size_t dstSize,
                      const void* src, size_t srcSize,
                      unsigned maxSymbolValue, unsigned huffLog,
                      void* workSpace, size_t wkspSize,
                      HUF_CElt* hufTable, HUF_repeat* repeat, int flags)
{
    DEBUGLOG(5, "HUF_compress4X_repeat (srcSize = %zu)", srcSize);
    return HUF_compress_internal(dst, dstSize, src, srcSize,
                                 maxSymbolValue, huffLog, HUF_fourStreams,
                                 workSpace, wkspSize,
                                 hufTable, repeat, flags);
}
/**** ended inlining compress/huf_compress.c ****/
/**** start inlining compress/zstd_compress_literals.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

 /*-*************************************
 *  Dependencies
 ***************************************/
/**** start inlining zstd_compress_literals.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_COMPRESS_LITERALS_H
#define ZSTD_COMPRESS_LITERALS_H

/**** start inlining zstd_compress_internal.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* This header contains definitions
 * that shall **only** be used by modules within lib/compress.
 */

#ifndef ZSTD_COMPRESS_H
#define ZSTD_COMPRESS_H

/*-*************************************
*  Dependencies
***************************************/
/**** skipping file: ../common/zstd_internal.h ****/
/**** start inlining zstd_cwksp.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_CWKSP_H
#define ZSTD_CWKSP_H

/*-*************************************
*  Dependencies
***************************************/
/**** skipping file: ../common/allocations.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../common/portability_macros.h ****/
/**** skipping file: ../common/compiler.h ****/

/*-*************************************
*  Constants
***************************************/

/* Since the workspace is effectively its own little malloc implementation /
 * arena, when we run under ASAN, we should similarly insert redzones between
 * each internal element of the workspace, so ASAN will catch overruns that
 * reach outside an object but that stay inside the workspace.
 *
 * This defines the size of that redzone.
 */
#ifndef ZSTD_CWKSP_ASAN_REDZONE_SIZE
#define ZSTD_CWKSP_ASAN_REDZONE_SIZE 128
#endif


/* Set our tables and aligneds to align by 64 bytes */
#define ZSTD_CWKSP_ALIGNMENT_BYTES 64

/*-*************************************
*  Structures
***************************************/
typedef enum {
    ZSTD_cwksp_alloc_objects,
    ZSTD_cwksp_alloc_aligned_init_once,
    ZSTD_cwksp_alloc_aligned,
    ZSTD_cwksp_alloc_buffers
} ZSTD_cwksp_alloc_phase_e;

/**
 * Used to describe whether the workspace is statically allocated (and will not
 * necessarily ever be freed), or if it's dynamically allocated and we can
 * expect a well-formed caller to free this.
 */
typedef enum {
    ZSTD_cwksp_dynamic_alloc,
    ZSTD_cwksp_static_alloc
} ZSTD_cwksp_static_alloc_e;

/**
 * Zstd fits all its internal datastructures into a single continuous buffer,
 * so that it only needs to perform a single OS allocation (or so that a buffer
 * can be provided to it and it can perform no allocations at all). This buffer
 * is called the workspace.
 *
 * Several optimizations complicate that process of allocating memory ranges
 * from this workspace for each internal datastructure:
 *
 * - These different internal datastructures have different setup requirements:
 *
 *   - The static objects need to be cleared once and can then be trivially
 *     reused for each compression.
 *
 *   - Various buffers don't need to be initialized at all--they are always
 *     written into before they're read.
 *
 *   - The matchstate tables have a unique requirement that they don't need
 *     their memory to be totally cleared, but they do need the memory to have
 *     some bound, i.e., a guarantee that all values in the memory they've been
 *     allocated is less than some maximum value (which is the starting value
 *     for the indices that they will then use for compression). When this
 *     guarantee is provided to them, they can use the memory without any setup
 *     work. When it can't, they have to clear the area.
 *
 * - These buffers also have different alignment requirements.
 *
 * - We would like to reuse the objects in the workspace for multiple
 *   compressions without having to perform any expensive reallocation or
 *   reinitialization work.
 *
 * - We would like to be able to efficiently reuse the workspace across
 *   multiple compressions **even when the compression parameters change** and
 *   we need to resize some of the objects (where possible).
 *
 * To attempt to manage this buffer, given these constraints, the ZSTD_cwksp
 * abstraction was created. It works as follows:
 *
 * Workspace Layout:
 *
 * [                        ... workspace ...                           ]
 * [objects][tables ->] free space [<- buffers][<- aligned][<- init once]
 *
 * The various objects that live in the workspace are divided into the
 * following categories, and are allocated separately:
 *
 * - Static objects: this is optionally the enclosing ZSTD_CCtx or ZSTD_CDict,
 *   so that literally everything fits in a single buffer. Note: if present,
 *   this must be the first object in the workspace, since ZSTD_customFree{CCtx,
 *   CDict}() rely on a pointer comparison to see whether one or two frees are
 *   required.
 *
 * - Fixed size objects: these are fixed-size, fixed-count objects that are
 *   nonetheless "dynamically" allocated in the workspace so that we can
 *   control how they're initialized separately from the broader ZSTD_CCtx.
 *   Examples:
 *   - Entropy Workspace
 *   - 2 x ZSTD_compressedBlockState_t
 *   - CDict dictionary contents
 *
 * - Tables: these are any of several different datastructures (hash tables,
 *   chain tables, binary trees) that all respect a common format: they are
 *   uint32_t arrays, all of whose values are between 0 and (nextSrc - base).
 *   Their sizes depend on the cparams. These tables are 64-byte aligned.
 *
 * - Init once: these buffers require to be initialized at least once before
 *   use. They should be used when we want to skip memory initialization
 *   while not triggering memory checkers (like Valgrind) when reading from
 *   from this memory without writing to it first.
 *   These buffers should be used carefully as they might contain data
 *   from previous compressions.
 *   Buffers are aligned to 64 bytes.
 *
 * - Aligned: these buffers don't require any initialization before they're
 *   used. The user of the buffer should make sure they write into a buffer
 *   location before reading from it.
 *   Buffers are aligned to 64 bytes.
 *
 * - Buffers: these buffers are used for various purposes that don't require
 *   any alignment or initialization before they're used. This means they can
 *   be moved around at no cost for a new compression.
 *
 * Allocating Memory:
 *
 * The various types of objects must be allocated in order, so they can be
 * correctly packed into the workspace buffer. That order is:
 *
 * 1. Objects
 * 2. Init once / Tables
 * 3. Aligned / Tables
 * 4. Buffers / Tables
 *
 * Attempts to reserve objects of different types out of order will fail.
 */
typedef struct {
    void* workspace;
    void* workspaceEnd;

    void* objectEnd;
    void* tableEnd;
    void* tableValidEnd;
    void* allocStart;
    void* initOnceStart;

    BYTE allocFailed;
    int workspaceOversizedDuration;
    ZSTD_cwksp_alloc_phase_e phase;
    ZSTD_cwksp_static_alloc_e isStatic;
} ZSTD_cwksp;

/*-*************************************
*  Functions
***************************************/

MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws);
MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws);

MEM_STATIC void ZSTD_cwksp_assert_internal_consistency(ZSTD_cwksp* ws) {
    (void)ws;
    assert(ws->workspace <= ws->objectEnd);
    assert(ws->objectEnd <= ws->tableEnd);
    assert(ws->objectEnd <= ws->tableValidEnd);
    assert(ws->tableEnd <= ws->allocStart);
    assert(ws->tableValidEnd <= ws->allocStart);
    assert(ws->allocStart <= ws->workspaceEnd);
    assert(ws->initOnceStart <= ZSTD_cwksp_initialAllocStart(ws));
    assert(ws->workspace <= ws->initOnceStart);
#if ZSTD_MEMORY_SANITIZER
    {
        intptr_t const offset = __msan_test_shadow(ws->initOnceStart,
            (U8*)ZSTD_cwksp_initialAllocStart(ws) - (U8*)ws->initOnceStart);
        (void)offset;
#if defined(ZSTD_MSAN_PRINT)
        if(offset!=-1) {
            __msan_print_shadow((U8*)ws->initOnceStart + offset - 8, 32);
        }
#endif
        assert(offset==-1);
    };
#endif
}

/**
 * Align must be a power of 2.
 */
MEM_STATIC size_t ZSTD_cwksp_align(size_t size, size_t align) {
    size_t const mask = align - 1;
    assert(ZSTD_isPower2(align));
    return (size + mask) & ~mask;
}

/**
 * Use this to determine how much space in the workspace we will consume to
 * allocate this object. (Normally it should be exactly the size of the object,
 * but under special conditions, like ASAN, where we pad each object, it might
 * be larger.)
 *
 * Since tables aren't currently redzoned, you don't need to call through this
 * to figure out how much space you need for the matchState tables. Everything
 * else is though.
 *
 * Do not use for sizing aligned buffers. Instead, use ZSTD_cwksp_aligned64_alloc_size().
 */
MEM_STATIC size_t ZSTD_cwksp_alloc_size(size_t size) {
    if (size == 0)
        return 0;
#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    return size + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
#else
    return size;
#endif
}

MEM_STATIC size_t ZSTD_cwksp_aligned_alloc_size(size_t size, size_t alignment) {
    return ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(size, alignment));
}

/**
 * Returns an adjusted alloc size that is the nearest larger multiple of 64 bytes.
 * Used to determine the number of bytes required for a given "aligned".
 */
MEM_STATIC size_t ZSTD_cwksp_aligned64_alloc_size(size_t size) {
    return ZSTD_cwksp_aligned_alloc_size(size, ZSTD_CWKSP_ALIGNMENT_BYTES);
}

/**
 * Returns the amount of additional space the cwksp must allocate
 * for internal purposes (currently only alignment).
 */
MEM_STATIC size_t ZSTD_cwksp_slack_space_required(void) {
    /* For alignment, the wksp will always allocate an additional 2*ZSTD_CWKSP_ALIGNMENT_BYTES
     * bytes to align the beginning of tables section and end of buffers;
     */
    size_t const slackSpace = ZSTD_CWKSP_ALIGNMENT_BYTES * 2;
    return slackSpace;
}


/**
 * Return the number of additional bytes required to align a pointer to the given number of bytes.
 * alignBytes must be a power of two.
 */
MEM_STATIC size_t ZSTD_cwksp_bytes_to_align_ptr(void* ptr, const size_t alignBytes) {
    size_t const alignBytesMask = alignBytes - 1;
    size_t const bytes = (alignBytes - ((size_t)ptr & (alignBytesMask))) & alignBytesMask;
    assert(ZSTD_isPower2(alignBytes));
    assert(bytes < alignBytes);
    return bytes;
}

/**
 * Returns the initial value for allocStart which is used to determine the position from
 * which we can allocate from the end of the workspace.
 */
MEM_STATIC void*  ZSTD_cwksp_initialAllocStart(ZSTD_cwksp* ws)
{
    char* endPtr = (char*)ws->workspaceEnd;
    assert(ZSTD_isPower2(ZSTD_CWKSP_ALIGNMENT_BYTES));
    endPtr = endPtr - ((size_t)endPtr % ZSTD_CWKSP_ALIGNMENT_BYTES);
    return (void*)endPtr;
}

/**
 * Internal function. Do not use directly.
 * Reserves the given number of bytes within the aligned/buffer segment of the wksp,
 * which counts from the end of the wksp (as opposed to the object/table segment).
 *
 * Returns a pointer to the beginning of that space.
 */
MEM_STATIC void*
ZSTD_cwksp_reserve_internal_buffer_space(ZSTD_cwksp* ws, size_t const bytes)
{
    void* const alloc = (BYTE*)ws->allocStart - bytes;
    void* const bottom = ws->tableEnd;
    DEBUGLOG(5, "cwksp: reserving [0x%p]:%zd bytes; %zd bytes remaining",
        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
    ZSTD_cwksp_assert_internal_consistency(ws);
    assert(alloc >= bottom);
    if (alloc < bottom) {
        DEBUGLOG(4, "cwksp: alloc failed!");
        ws->allocFailed = 1;
        return NULL;
    }
    /* the area is reserved from the end of wksp.
     * If it overlaps with tableValidEnd, it voids guarantees on values' range */
    if (alloc < ws->tableValidEnd) {
        ws->tableValidEnd = alloc;
    }
    ws->allocStart = alloc;
    return alloc;
}

/**
 * Moves the cwksp to the next phase, and does any necessary allocations.
 * cwksp initialization must necessarily go through each phase in order.
 * Returns a 0 on success, or zstd error
 */
MEM_STATIC size_t
ZSTD_cwksp_internal_advance_phase(ZSTD_cwksp* ws, ZSTD_cwksp_alloc_phase_e phase)
{
    assert(phase >= ws->phase);
    if (phase > ws->phase) {
        /* Going from allocating objects to allocating initOnce / tables */
        if (ws->phase < ZSTD_cwksp_alloc_aligned_init_once &&
            phase >= ZSTD_cwksp_alloc_aligned_init_once) {
            ws->tableValidEnd = ws->objectEnd;
            ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);

            {   /* Align the start of the tables to 64 bytes. Use [0, 63] bytes */
                void *const alloc = ws->objectEnd;
                size_t const bytesToAlign = ZSTD_cwksp_bytes_to_align_ptr(alloc, ZSTD_CWKSP_ALIGNMENT_BYTES);
                void *const objectEnd = (BYTE *) alloc + bytesToAlign;
                DEBUGLOG(5, "reserving table alignment addtl space: %zu", bytesToAlign);
                RETURN_ERROR_IF(objectEnd > ws->workspaceEnd, memory_allocation,
                                "table phase - alignment initial allocation failed!");
                ws->objectEnd = objectEnd;
                ws->tableEnd = objectEnd;  /* table area starts being empty */
                if (ws->tableValidEnd < ws->tableEnd) {
                    ws->tableValidEnd = ws->tableEnd;
                }
            }
        }
        ws->phase = phase;
        ZSTD_cwksp_assert_internal_consistency(ws);
    }
    return 0;
}

/**
 * Returns whether this object/buffer/etc was allocated in this workspace.
 */
MEM_STATIC int ZSTD_cwksp_owns_buffer(const ZSTD_cwksp* ws, const void* ptr)
{
    return (ptr != NULL) && (ws->workspace <= ptr) && (ptr < ws->workspaceEnd);
}

/**
 * Internal function. Do not use directly.
 */
MEM_STATIC void*
ZSTD_cwksp_reserve_internal(ZSTD_cwksp* ws, size_t bytes, ZSTD_cwksp_alloc_phase_e phase)
{
    void* alloc;
    if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase)) || bytes == 0) {
        return NULL;
    }

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* over-reserve space */
    bytes += 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
#endif

    alloc = ZSTD_cwksp_reserve_internal_buffer_space(ws, bytes);

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
     * either size. */
    if (alloc) {
        alloc = (BYTE *)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
        if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
            /* We need to keep the redzone poisoned while unpoisoning the bytes that
             * are actually allocated. */
            __asan_unpoison_memory_region(alloc, bytes - 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE);
        }
    }
#endif

    return alloc;
}

/**
 * Reserves and returns unaligned memory.
 */
MEM_STATIC BYTE* ZSTD_cwksp_reserve_buffer(ZSTD_cwksp* ws, size_t bytes)
{
    return (BYTE*)ZSTD_cwksp_reserve_internal(ws, bytes, ZSTD_cwksp_alloc_buffers);
}

/**
 * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
 * This memory has been initialized at least once in the past.
 * This doesn't mean it has been initialized this time, and it might contain data from previous
 * operations.
 * The main usage is for algorithms that might need read access into uninitialized memory.
 * The algorithm must maintain safety under these conditions and must make sure it doesn't
 * leak any of the past data (directly or in side channels).
 */
MEM_STATIC void* ZSTD_cwksp_reserve_aligned_init_once(ZSTD_cwksp* ws, size_t bytes)
{
    size_t const alignedBytes = ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES);
    void* ptr = ZSTD_cwksp_reserve_internal(ws, alignedBytes, ZSTD_cwksp_alloc_aligned_init_once);
    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
    if(ptr && ptr < ws->initOnceStart) {
        /* We assume the memory following the current allocation is either:
         * 1. Not usable as initOnce memory (end of workspace)
         * 2. Another initOnce buffer that has been allocated before (and so was previously memset)
         * 3. An ASAN redzone, in which case we don't want to write on it
         * For these reasons it should be fine to not explicitly zero every byte up to ws->initOnceStart.
         * Note that we assume here that MSAN and ASAN cannot run in the same time. */
        ZSTD_memset(ptr, 0, MIN((size_t)((U8*)ws->initOnceStart - (U8*)ptr), alignedBytes));
        ws->initOnceStart = ptr;
    }
#if ZSTD_MEMORY_SANITIZER
    assert(__msan_test_shadow(ptr, bytes) == -1);
#endif
    return ptr;
}

/**
 * Reserves and returns memory sized on and aligned on ZSTD_CWKSP_ALIGNMENT_BYTES (64 bytes).
 */
MEM_STATIC void* ZSTD_cwksp_reserve_aligned64(ZSTD_cwksp* ws, size_t bytes)
{
    void* const ptr = ZSTD_cwksp_reserve_internal(ws,
                        ZSTD_cwksp_align(bytes, ZSTD_CWKSP_ALIGNMENT_BYTES),
                        ZSTD_cwksp_alloc_aligned);
    assert(((size_t)ptr & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
    return ptr;
}

/**
 * Aligned on 64 bytes. These buffers have the special property that
 * their values remain constrained, allowing us to reuse them without
 * memset()-ing them.
 */
MEM_STATIC void* ZSTD_cwksp_reserve_table(ZSTD_cwksp* ws, size_t bytes)
{
    const ZSTD_cwksp_alloc_phase_e phase = ZSTD_cwksp_alloc_aligned_init_once;
    void* alloc;
    void* end;
    void* top;

    /* We can only start allocating tables after we are done reserving space for objects at the
     * start of the workspace */
    if(ws->phase < phase) {
        if (ZSTD_isError(ZSTD_cwksp_internal_advance_phase(ws, phase))) {
            return NULL;
        }
    }
    alloc = ws->tableEnd;
    end = (BYTE *)alloc + bytes;
    top = ws->allocStart;

    DEBUGLOG(5, "cwksp: reserving %p table %zd bytes, %zd bytes remaining",
        alloc, bytes, ZSTD_cwksp_available_space(ws) - bytes);
    assert((bytes & (sizeof(U32)-1)) == 0);
    ZSTD_cwksp_assert_internal_consistency(ws);
    assert(end <= top);
    if (end > top) {
        DEBUGLOG(4, "cwksp: table alloc failed!");
        ws->allocFailed = 1;
        return NULL;
    }
    ws->tableEnd = end;

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
        __asan_unpoison_memory_region(alloc, bytes);
    }
#endif

    assert((bytes & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
    assert(((size_t)alloc & (ZSTD_CWKSP_ALIGNMENT_BYTES-1)) == 0);
    return alloc;
}

/**
 * Aligned on sizeof(void*).
 * Note : should happen only once, at workspace first initialization
 */
MEM_STATIC void* ZSTD_cwksp_reserve_object(ZSTD_cwksp* ws, size_t bytes)
{
    size_t const roundedBytes = ZSTD_cwksp_align(bytes, sizeof(void*));
    void* alloc = ws->objectEnd;
    void* end = (BYTE*)alloc + roundedBytes;

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* over-reserve space */
    end = (BYTE *)end + 2 * ZSTD_CWKSP_ASAN_REDZONE_SIZE;
#endif

    DEBUGLOG(4,
        "cwksp: reserving %p object %zd bytes (rounded to %zd), %zd bytes remaining",
        alloc, bytes, roundedBytes, ZSTD_cwksp_available_space(ws) - roundedBytes);
    assert((size_t)alloc % ZSTD_ALIGNOF(void*) == 0);
    assert(bytes % ZSTD_ALIGNOF(void*) == 0);
    ZSTD_cwksp_assert_internal_consistency(ws);
    /* we must be in the first phase, no advance is possible */
    if (ws->phase != ZSTD_cwksp_alloc_objects || end > ws->workspaceEnd) {
        DEBUGLOG(3, "cwksp: object alloc failed!");
        ws->allocFailed = 1;
        return NULL;
    }
    ws->objectEnd = end;
    ws->tableEnd = end;
    ws->tableValidEnd = end;

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* Move alloc so there's ZSTD_CWKSP_ASAN_REDZONE_SIZE unused space on
     * either size. */
    alloc = (BYTE*)alloc + ZSTD_CWKSP_ASAN_REDZONE_SIZE;
    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
        __asan_unpoison_memory_region(alloc, bytes);
    }
#endif

    return alloc;
}
/**
 * with alignment control
 * Note : should happen only once, at workspace first initialization
 */
MEM_STATIC void* ZSTD_cwksp_reserve_object_aligned(ZSTD_cwksp* ws, size_t byteSize, size_t alignment)
{
    size_t const mask = alignment - 1;
    size_t const surplus = (alignment > sizeof(void*)) ? alignment - sizeof(void*) : 0;
    void* const start = ZSTD_cwksp_reserve_object(ws, byteSize + surplus);
    if (start == NULL) return NULL;
    if (surplus == 0) return start;
    assert(ZSTD_isPower2(alignment));
    return (void*)(((size_t)start + surplus) & ~mask);
}

MEM_STATIC void ZSTD_cwksp_mark_tables_dirty(ZSTD_cwksp* ws)
{
    DEBUGLOG(4, "cwksp: ZSTD_cwksp_mark_tables_dirty");

#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
    /* To validate that the table reuse logic is sound, and that we don't
     * access table space that we haven't cleaned, we re-"poison" the table
     * space every time we mark it dirty.
     * Since tableValidEnd space and initOnce space may overlap we don't poison
     * the initOnce portion as it break its promise. This means that this poisoning
     * check isn't always applied fully. */
    {
        size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
        assert(__msan_test_shadow(ws->objectEnd, size) == -1);
        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
            __msan_poison(ws->objectEnd, size);
        } else {
            assert(ws->initOnceStart >= ws->objectEnd);
            __msan_poison(ws->objectEnd, (BYTE*)ws->initOnceStart - (BYTE*)ws->objectEnd);
        }
    }
#endif

    assert(ws->tableValidEnd >= ws->objectEnd);
    assert(ws->tableValidEnd <= ws->allocStart);
    ws->tableValidEnd = ws->objectEnd;
    ZSTD_cwksp_assert_internal_consistency(ws);
}

MEM_STATIC void ZSTD_cwksp_mark_tables_clean(ZSTD_cwksp* ws) {
    DEBUGLOG(4, "cwksp: ZSTD_cwksp_mark_tables_clean");
    assert(ws->tableValidEnd >= ws->objectEnd);
    assert(ws->tableValidEnd <= ws->allocStart);
    if (ws->tableValidEnd < ws->tableEnd) {
        ws->tableValidEnd = ws->tableEnd;
    }
    ZSTD_cwksp_assert_internal_consistency(ws);
}

/**
 * Zero the part of the allocated tables not already marked clean.
 */
MEM_STATIC void ZSTD_cwksp_clean_tables(ZSTD_cwksp* ws) {
    DEBUGLOG(4, "cwksp: ZSTD_cwksp_clean_tables");
    assert(ws->tableValidEnd >= ws->objectEnd);
    assert(ws->tableValidEnd <= ws->allocStart);
    if (ws->tableValidEnd < ws->tableEnd) {
        ZSTD_memset(ws->tableValidEnd, 0, (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->tableValidEnd));
    }
    ZSTD_cwksp_mark_tables_clean(ws);
}

/**
 * Invalidates table allocations.
 * All other allocations remain valid.
 */
MEM_STATIC void ZSTD_cwksp_clear_tables(ZSTD_cwksp* ws)
{
    DEBUGLOG(4, "cwksp: clearing tables!");

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* We don't do this when the workspace is statically allocated, because
     * when that is the case, we have no capability to hook into the end of the
     * workspace's lifecycle to unpoison the memory.
     */
    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
        size_t size = (BYTE*)ws->tableValidEnd - (BYTE*)ws->objectEnd;
        __asan_poison_memory_region(ws->objectEnd, size);
    }
#endif

    ws->tableEnd = ws->objectEnd;
    ZSTD_cwksp_assert_internal_consistency(ws);
}

/**
 * Invalidates all buffer, aligned, and table allocations.
 * Object allocations remain valid.
 */
MEM_STATIC void ZSTD_cwksp_clear(ZSTD_cwksp* ws) {
    DEBUGLOG(4, "cwksp: clearing!");

#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
    /* To validate that the context reuse logic is sound, and that we don't
     * access stuff that this compression hasn't initialized, we re-"poison"
     * the workspace except for the areas in which we expect memory reuse
     * without initialization (objects, valid tables area and init once
     * memory). */
    {
        if((BYTE*)ws->tableValidEnd < (BYTE*)ws->initOnceStart) {
            size_t size = (BYTE*)ws->initOnceStart - (BYTE*)ws->tableValidEnd;
            __msan_poison(ws->tableValidEnd, size);
        }
    }
#endif

#if ZSTD_ADDRESS_SANITIZER && !defined (ZSTD_ASAN_DONT_POISON_WORKSPACE)
    /* We don't do this when the workspace is statically allocated, because
     * when that is the case, we have no capability to hook into the end of the
     * workspace's lifecycle to unpoison the memory.
     */
    if (ws->isStatic == ZSTD_cwksp_dynamic_alloc) {
        size_t size = (BYTE*)ws->workspaceEnd - (BYTE*)ws->objectEnd;
        __asan_poison_memory_region(ws->objectEnd, size);
    }
#endif

    ws->tableEnd = ws->objectEnd;
    ws->allocStart = ZSTD_cwksp_initialAllocStart(ws);
    ws->allocFailed = 0;
    if (ws->phase > ZSTD_cwksp_alloc_aligned_init_once) {
        ws->phase = ZSTD_cwksp_alloc_aligned_init_once;
    }
    ZSTD_cwksp_assert_internal_consistency(ws);
}

MEM_STATIC size_t ZSTD_cwksp_sizeof(const ZSTD_cwksp* ws) {
    return (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->workspace);
}

MEM_STATIC size_t ZSTD_cwksp_used(const ZSTD_cwksp* ws) {
    return (size_t)((BYTE*)ws->tableEnd - (BYTE*)ws->workspace)
         + (size_t)((BYTE*)ws->workspaceEnd - (BYTE*)ws->allocStart);
}

/**
 * The provided workspace takes ownership of the buffer [start, start+size).
 * Any existing values in the workspace are ignored (the previously managed
 * buffer, if present, must be separately freed).
 */
MEM_STATIC void ZSTD_cwksp_init(ZSTD_cwksp* ws, void* start, size_t size, ZSTD_cwksp_static_alloc_e isStatic) {
    DEBUGLOG(4, "cwksp: init'ing workspace with %zd bytes", size);
    assert(((size_t)start & (sizeof(void*)-1)) == 0); /* ensure correct alignment */
    ws->workspace = start;
    ws->workspaceEnd = (BYTE*)start + size;
    ws->objectEnd = ws->workspace;
    ws->tableValidEnd = ws->objectEnd;
    ws->initOnceStart = ZSTD_cwksp_initialAllocStart(ws);
    ws->phase = ZSTD_cwksp_alloc_objects;
    ws->isStatic = isStatic;
    ZSTD_cwksp_clear(ws);
    ws->workspaceOversizedDuration = 0;
    ZSTD_cwksp_assert_internal_consistency(ws);
}

MEM_STATIC size_t ZSTD_cwksp_create(ZSTD_cwksp* ws, size_t size, ZSTD_customMem customMem) {
    void* workspace = ZSTD_customMalloc(size, customMem);
    DEBUGLOG(4, "cwksp: creating new workspace with %zd bytes", size);
    RETURN_ERROR_IF(workspace == NULL, memory_allocation, "NULL pointer!");
    ZSTD_cwksp_init(ws, workspace, size, ZSTD_cwksp_dynamic_alloc);
    return 0;
}

MEM_STATIC void ZSTD_cwksp_free(ZSTD_cwksp* ws, ZSTD_customMem customMem) {
    void *ptr = ws->workspace;
    DEBUGLOG(4, "cwksp: freeing workspace");
#if ZSTD_MEMORY_SANITIZER && !defined(ZSTD_MSAN_DONT_POISON_WORKSPACE)
    if (ptr != NULL && customMem.customFree != NULL) {
        __msan_unpoison(ptr, ZSTD_cwksp_sizeof(ws));
    }
#endif
    ZSTD_memset(ws, 0, sizeof(ZSTD_cwksp));
    ZSTD_customFree(ptr, customMem);
}

/**
 * Moves the management of a workspace from one cwksp to another. The src cwksp
 * is left in an invalid state (src must be re-init()'ed before it's used again).
 */
MEM_STATIC void ZSTD_cwksp_move(ZSTD_cwksp* dst, ZSTD_cwksp* src) {
    *dst = *src;
    ZSTD_memset(src, 0, sizeof(ZSTD_cwksp));
}

MEM_STATIC int ZSTD_cwksp_reserve_failed(const ZSTD_cwksp* ws) {
    return ws->allocFailed;
}

/*-*************************************
*  Functions Checking Free Space
***************************************/

/* ZSTD_alignmentSpaceWithinBounds() :
 * Returns if the estimated space needed for a wksp is within an acceptable limit of the
 * actual amount of space used.
 */
MEM_STATIC int ZSTD_cwksp_estimated_space_within_bounds(const ZSTD_cwksp *const ws, size_t const estimatedSpace) {
    /* We have an alignment space between objects and tables between tables and buffers, so we can have up to twice
     * the alignment bytes difference between estimation and actual usage */
    return (estimatedSpace - ZSTD_cwksp_slack_space_required()) <= ZSTD_cwksp_used(ws) &&
           ZSTD_cwksp_used(ws) <= estimatedSpace;
}


MEM_STATIC size_t ZSTD_cwksp_available_space(ZSTD_cwksp* ws) {
    return (size_t)((BYTE*)ws->allocStart - (BYTE*)ws->tableEnd);
}

MEM_STATIC int ZSTD_cwksp_check_available(ZSTD_cwksp* ws, size_t additionalNeededSpace) {
    return ZSTD_cwksp_available_space(ws) >= additionalNeededSpace;
}

MEM_STATIC int ZSTD_cwksp_check_too_large(ZSTD_cwksp* ws, size_t additionalNeededSpace) {
    return ZSTD_cwksp_check_available(
        ws, additionalNeededSpace * ZSTD_WORKSPACETOOLARGE_FACTOR);
}

MEM_STATIC int ZSTD_cwksp_check_wasteful(ZSTD_cwksp* ws, size_t additionalNeededSpace) {
    return ZSTD_cwksp_check_too_large(ws, additionalNeededSpace)
        && ws->workspaceOversizedDuration > ZSTD_WORKSPACETOOLARGE_MAXDURATION;
}

MEM_STATIC void ZSTD_cwksp_bump_oversized_duration(
        ZSTD_cwksp* ws, size_t additionalNeededSpace) {
    if (ZSTD_cwksp_check_too_large(ws, additionalNeededSpace)) {
        ws->workspaceOversizedDuration++;
    } else {
        ws->workspaceOversizedDuration = 0;
    }
}

#endif /* ZSTD_CWKSP_H */
/**** ended inlining zstd_cwksp.h ****/
#ifdef ZSTD_MULTITHREAD
/**** start inlining zstdmt_compress.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

 #ifndef ZSTDMT_COMPRESS_H
 #define ZSTDMT_COMPRESS_H

/* ===   Dependencies   === */
/**** skipping file: ../common/zstd_deps.h ****/
#define ZSTD_STATIC_LINKING_ONLY   /* ZSTD_parameters */
/**** skipping file: ../zstd.h ****/

/* Note : This is an internal API.
 *        These APIs used to be exposed with ZSTDLIB_API,
 *        because it used to be the only way to invoke MT compression.
 *        Now, you must use ZSTD_compress2 and ZSTD_compressStream2() instead.
 *
 *        This API requires ZSTD_MULTITHREAD to be defined during compilation,
 *        otherwise ZSTDMT_createCCtx*() will fail.
 */

/* ===   Constants   === */
#ifndef ZSTDMT_NBWORKERS_MAX /* a different value can be selected at compile time */
#  define ZSTDMT_NBWORKERS_MAX ((sizeof(void*)==4) /*32-bit*/ ? 64 : 256)
#endif
#ifndef ZSTDMT_JOBSIZE_MIN   /* a different value can be selected at compile time */
#  define ZSTDMT_JOBSIZE_MIN (512 KB)
#endif
#define ZSTDMT_JOBLOG_MAX   (MEM_32bits() ? 29 : 30)
#define ZSTDMT_JOBSIZE_MAX  (MEM_32bits() ? (512 MB) : (1024 MB))


/* ========================================================
 * ===  Private interface, for use by ZSTD_compress.c   ===
 * ===  Not exposed in libzstd. Never invoke directly   ===
 * ======================================================== */

/* ===   Memory management   === */
typedef struct ZSTDMT_CCtx_s ZSTDMT_CCtx;
/* Requires ZSTD_MULTITHREAD to be defined during compilation, otherwise it will return NULL. */
ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers,
                                        ZSTD_customMem cMem,
					ZSTD_threadPool *pool);
size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx);

size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx);

/* ===   Streaming functions   === */

size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx);

/*! ZSTDMT_initCStream_internal() :
 *  Private use only. Init streaming operation.
 *  expects params to be valid.
 *  must receive dict, or cdict, or none, but not both.
 *  mtctx can be freshly constructed or reused from a prior compression.
 *  If mtctx is reused, memory allocations from the prior compression may not be freed,
 *  even if they are not needed for the current compression.
 *  @return : 0, or an error code */
size_t ZSTDMT_initCStream_internal(ZSTDMT_CCtx* mtctx,
                    const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
                    const ZSTD_CDict* cdict,
                    ZSTD_CCtx_params params, unsigned long long pledgedSrcSize);

/*! ZSTDMT_compressStream_generic() :
 *  Combines ZSTDMT_compressStream() with optional ZSTDMT_flushStream() or ZSTDMT_endStream()
 *  depending on flush directive.
 * @return : minimum amount of data still to be flushed
 *           0 if fully flushed
 *           or an error code
 *  note : needs to be init using any ZSTD_initCStream*() variant */
size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
                                     ZSTD_outBuffer* output,
                                     ZSTD_inBuffer* input,
                                     ZSTD_EndDirective endOp);

 /*! ZSTDMT_toFlushNow()
  *  Tell how many bytes are ready to be flushed immediately.
  *  Probe the oldest active job (not yet entirely flushed) and check its output buffer.
  *  If return 0, it means there is no active job,
  *  or, it means oldest job is still active, but everything produced has been flushed so far,
  *  therefore flushing is limited by speed of oldest job. */
size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx);

/*! ZSTDMT_updateCParams_whileCompressing() :
 *  Updates only a selected set of compression parameters, to remain compatible with current frame.
 *  New parameters will be applied to next compression job. */
void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams);

/*! ZSTDMT_getFrameProgression():
 *  tells how much data has been consumed (input) and produced (output) for current frame.
 *  able to count progression inside worker threads.
 */
ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx);

#endif   /* ZSTDMT_COMPRESS_H */
/**** ended inlining zstdmt_compress.h ****/
#endif
/**** skipping file: ../common/bits.h ****/
/**** start inlining zstd_preSplit.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_PRESPLIT_H
#define ZSTD_PRESPLIT_H

#include <stddef.h>  /* size_t */

#define ZSTD_SLIPBLOCK_WORKSPACESIZE 8208

/* ZSTD_splitBlock():
 * @level must be a value between 0 and 4.
 *        higher levels spend more energy to detect block boundaries.
 * @workspace must be aligned for size_t.
 * @wkspSize must be at least >= ZSTD_SLIPBLOCK_WORKSPACESIZE
 * note:
 * For the time being, this function only accepts full 128 KB blocks.
 * Therefore, @blockSize must be == 128 KB.
 * While this could be extended to smaller sizes in the future,
 * it is not yet clear if this would be useful. TBD.
 */
size_t ZSTD_splitBlock(const void* blockStart, size_t blockSize,
                    int level,
                    void* workspace, size_t wkspSize);

#endif /* ZSTD_PRESPLIT_H */
/**** ended inlining zstd_preSplit.h ****/

/*-*************************************
*  Constants
***************************************/
#define kSearchStrength      8
#define HASH_READ_SIZE       8
#define ZSTD_DUBT_UNSORTED_MARK 1   /* For btlazy2 strategy, index ZSTD_DUBT_UNSORTED_MARK==1 means "unsorted".
                                       It could be confused for a real successor at index "1", if sorted as larger than its predecessor.
                                       It's not a big deal though : candidate will just be sorted again.
                                       Additionally, candidate position 1 will be lost.
                                       But candidate 1 cannot hide a large tree of candidates, so it's a minimal loss.
                                       The benefit is that ZSTD_DUBT_UNSORTED_MARK cannot be mishandled after table reuse with a different strategy.
                                       This constant is required by ZSTD_compressBlock_btlazy2() and ZSTD_reduceTable_internal() */


/*-*************************************
*  Context memory management
***************************************/
typedef enum { ZSTDcs_created=0, ZSTDcs_init, ZSTDcs_ongoing, ZSTDcs_ending } ZSTD_compressionStage_e;
typedef enum { zcss_init=0, zcss_load, zcss_flush } ZSTD_cStreamStage;

typedef struct ZSTD_prefixDict_s {
    const void* dict;
    size_t dictSize;
    ZSTD_dictContentType_e dictContentType;
} ZSTD_prefixDict;

typedef struct {
    void* dictBuffer;
    void const* dict;
    size_t dictSize;
    ZSTD_dictContentType_e dictContentType;
    ZSTD_CDict* cdict;
} ZSTD_localDict;

typedef struct {
    HUF_CElt CTable[HUF_CTABLE_SIZE_ST(255)];
    HUF_repeat repeatMode;
} ZSTD_hufCTables_t;

typedef struct {
    FSE_CTable offcodeCTable[FSE_CTABLE_SIZE_U32(OffFSELog, MaxOff)];
    FSE_CTable matchlengthCTable[FSE_CTABLE_SIZE_U32(MLFSELog, MaxML)];
    FSE_CTable litlengthCTable[FSE_CTABLE_SIZE_U32(LLFSELog, MaxLL)];
    FSE_repeat offcode_repeatMode;
    FSE_repeat matchlength_repeatMode;
    FSE_repeat litlength_repeatMode;
} ZSTD_fseCTables_t;

typedef struct {
    ZSTD_hufCTables_t huf;
    ZSTD_fseCTables_t fse;
} ZSTD_entropyCTables_t;

/***********************************************
*  Sequences *
***********************************************/
typedef struct SeqDef_s {
    U32 offBase;   /* offBase == Offset + ZSTD_REP_NUM, or repcode 1,2,3 */
    U16 litLength;
    U16 mlBase;    /* mlBase == matchLength - MINMATCH */
} SeqDef;

/* Controls whether seqStore has a single "long" litLength or matchLength. See SeqStore_t. */
typedef enum {
    ZSTD_llt_none = 0,             /* no longLengthType */
    ZSTD_llt_literalLength = 1,    /* represents a long literal */
    ZSTD_llt_matchLength = 2       /* represents a long match */
} ZSTD_longLengthType_e;

typedef struct {
    SeqDef* sequencesStart;
    SeqDef* sequences;      /* ptr to end of sequences */
    BYTE*  litStart;
    BYTE*  lit;             /* ptr to end of literals */
    BYTE*  llCode;
    BYTE*  mlCode;
    BYTE*  ofCode;
    size_t maxNbSeq;
    size_t maxNbLit;

    /* longLengthPos and longLengthType to allow us to represent either a single litLength or matchLength
     * in the seqStore that has a value larger than U16 (if it exists). To do so, we increment
     * the existing value of the litLength or matchLength by 0x10000.
     */
    ZSTD_longLengthType_e longLengthType;
    U32                   longLengthPos;  /* Index of the sequence to apply long length modification to */
} SeqStore_t;

typedef struct {
    U32 litLength;
    U32 matchLength;
} ZSTD_SequenceLength;

/**
 * Returns the ZSTD_SequenceLength for the given sequences. It handles the decoding of long sequences
 * indicated by longLengthPos and longLengthType, and adds MINMATCH back to matchLength.
 */
MEM_STATIC ZSTD_SequenceLength ZSTD_getSequenceLength(SeqStore_t const* seqStore, SeqDef const* seq)
{
    ZSTD_SequenceLength seqLen;
    seqLen.litLength = seq->litLength;
    seqLen.matchLength = seq->mlBase + MINMATCH;
    if (seqStore->longLengthPos == (U32)(seq - seqStore->sequencesStart)) {
        if (seqStore->longLengthType == ZSTD_llt_literalLength) {
            seqLen.litLength += 0x10000;
        }
        if (seqStore->longLengthType == ZSTD_llt_matchLength) {
            seqLen.matchLength += 0x10000;
        }
    }
    return seqLen;
}

const SeqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx);   /* compress & dictBuilder */
int ZSTD_seqToCodes(const SeqStore_t* seqStorePtr);   /* compress, dictBuilder, decodeCorpus (shouldn't get its definition from here) */


/***********************************************
*  Entropy buffer statistics structs and funcs *
***********************************************/
/** ZSTD_hufCTablesMetadata_t :
 *  Stores Literals Block Type for a super-block in hType, and
 *  huffman tree description in hufDesBuffer.
 *  hufDesSize refers to the size of huffman tree description in bytes.
 *  This metadata is populated in ZSTD_buildBlockEntropyStats_literals() */
typedef struct {
    SymbolEncodingType_e hType;
    BYTE hufDesBuffer[ZSTD_MAX_HUF_HEADER_SIZE];
    size_t hufDesSize;
} ZSTD_hufCTablesMetadata_t;

/** ZSTD_fseCTablesMetadata_t :
 *  Stores symbol compression modes for a super-block in {ll, ol, ml}Type, and
 *  fse tables in fseTablesBuffer.
 *  fseTablesSize refers to the size of fse tables in bytes.
 *  This metadata is populated in ZSTD_buildBlockEntropyStats_sequences() */
typedef struct {
    SymbolEncodingType_e llType;
    SymbolEncodingType_e ofType;
    SymbolEncodingType_e mlType;
    BYTE fseTablesBuffer[ZSTD_MAX_FSE_HEADERS_SIZE];
    size_t fseTablesSize;
    size_t lastCountSize; /* This is to account for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
} ZSTD_fseCTablesMetadata_t;

typedef struct {
    ZSTD_hufCTablesMetadata_t hufMetadata;
    ZSTD_fseCTablesMetadata_t fseMetadata;
} ZSTD_entropyCTablesMetadata_t;

/** ZSTD_buildBlockEntropyStats() :
 *  Builds entropy for the block.
 *  @return : 0 on success or error code */
size_t ZSTD_buildBlockEntropyStats(
                    const SeqStore_t* seqStorePtr,
                    const ZSTD_entropyCTables_t* prevEntropy,
                          ZSTD_entropyCTables_t* nextEntropy,
                    const ZSTD_CCtx_params* cctxParams,
                          ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                          void* workspace, size_t wkspSize);

/*********************************
*  Compression internals structs *
*********************************/

typedef struct {
    U32 off;            /* Offset sumtype code for the match, using ZSTD_storeSeq() format */
    U32 len;            /* Raw length of match */
} ZSTD_match_t;

typedef struct {
    U32 offset;         /* Offset of sequence */
    U32 litLength;      /* Length of literals prior to match */
    U32 matchLength;    /* Raw length of match */
} rawSeq;

typedef struct {
  rawSeq* seq;          /* The start of the sequences */
  size_t pos;           /* The index in seq where reading stopped. pos <= size. */
  size_t posInSequence; /* The position within the sequence at seq[pos] where reading
                           stopped. posInSequence <= seq[pos].litLength + seq[pos].matchLength */
  size_t size;          /* The number of sequences. <= capacity. */
  size_t capacity;      /* The capacity starting from `seq` pointer */
} RawSeqStore_t;

UNUSED_ATTR static const RawSeqStore_t kNullRawSeqStore = {NULL, 0, 0, 0, 0};

typedef struct {
    int price;  /* price from beginning of segment to this position */
    U32 off;    /* offset of previous match */
    U32 mlen;   /* length of previous match */
    U32 litlen; /* nb of literals since previous match */
    U32 rep[ZSTD_REP_NUM];  /* offset history after previous match */
} ZSTD_optimal_t;

typedef enum { zop_dynamic=0, zop_predef } ZSTD_OptPrice_e;

#define ZSTD_OPT_SIZE (ZSTD_OPT_NUM+3)
typedef struct {
    /* All tables are allocated inside cctx->workspace by ZSTD_resetCCtx_internal() */
    unsigned* litFreq;           /* table of literals statistics, of size 256 */
    unsigned* litLengthFreq;     /* table of litLength statistics, of size (MaxLL+1) */
    unsigned* matchLengthFreq;   /* table of matchLength statistics, of size (MaxML+1) */
    unsigned* offCodeFreq;       /* table of offCode statistics, of size (MaxOff+1) */
    ZSTD_match_t* matchTable;    /* list of found matches, of size ZSTD_OPT_SIZE */
    ZSTD_optimal_t* priceTable;  /* All positions tracked by optimal parser, of size ZSTD_OPT_SIZE */

    U32  litSum;                 /* nb of literals */
    U32  litLengthSum;           /* nb of litLength codes */
    U32  matchLengthSum;         /* nb of matchLength codes */
    U32  offCodeSum;             /* nb of offset codes */
    U32  litSumBasePrice;        /* to compare to log2(litfreq) */
    U32  litLengthSumBasePrice;  /* to compare to log2(llfreq)  */
    U32  matchLengthSumBasePrice;/* to compare to log2(mlfreq)  */
    U32  offCodeSumBasePrice;    /* to compare to log2(offreq)  */
    ZSTD_OptPrice_e priceType;   /* prices can be determined dynamically, or follow a pre-defined cost structure */
    const ZSTD_entropyCTables_t* symbolCosts;  /* pre-calculated dictionary statistics */
    ZSTD_ParamSwitch_e literalCompressionMode;
} optState_t;

typedef struct {
  ZSTD_entropyCTables_t entropy;
  U32 rep[ZSTD_REP_NUM];
} ZSTD_compressedBlockState_t;

typedef struct {
    BYTE const* nextSrc;       /* next block here to continue on current prefix */
    BYTE const* base;          /* All regular indexes relative to this position */
    BYTE const* dictBase;      /* extDict indexes relative to this position */
    U32 dictLimit;             /* below that point, need extDict */
    U32 lowLimit;              /* below that point, no more valid data */
    U32 nbOverflowCorrections; /* Number of times overflow correction has run since
                                * ZSTD_window_init(). Useful for debugging coredumps
                                * and for ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY.
                                */
} ZSTD_window_t;

#define ZSTD_WINDOW_START_INDEX 2

typedef struct ZSTD_MatchState_t ZSTD_MatchState_t;

#define ZSTD_ROW_HASH_CACHE_SIZE 8       /* Size of prefetching hash cache for row-based matchfinder */

struct ZSTD_MatchState_t {
    ZSTD_window_t window;   /* State for window round buffer management */
    U32 loadedDictEnd;      /* index of end of dictionary, within context's referential.
                             * When loadedDictEnd != 0, a dictionary is in use, and still valid.
                             * This relies on a mechanism to set loadedDictEnd=0 when dictionary is no longer within distance.
                             * Such mechanism is provided within ZSTD_window_enforceMaxDist() and ZSTD_checkDictValidity().
                             * When dict referential is copied into active context (i.e. not attached),
                             * loadedDictEnd == dictSize, since referential starts from zero.
                             */
    U32 nextToUpdate;       /* index from which to continue table update */
    U32 hashLog3;           /* dispatch table for matches of len==3 : larger == faster, more memory */

    U32 rowHashLog;                          /* For row-based matchfinder: Hashlog based on nb of rows in the hashTable.*/
    BYTE* tagTable;                          /* For row-based matchFinder: A row-based table containing the hashes and head index. */
    U32 hashCache[ZSTD_ROW_HASH_CACHE_SIZE]; /* For row-based matchFinder: a cache of hashes to improve speed */
    U64 hashSalt;                            /* For row-based matchFinder: salts the hash for reuse of tag table */
    U32 hashSaltEntropy;                     /* For row-based matchFinder: collects entropy for salt generation */

    U32* hashTable;
    U32* hashTable3;
    U32* chainTable;

    int forceNonContiguous; /* Non-zero if we should force non-contiguous load for the next window update. */

    int dedicatedDictSearch;  /* Indicates whether this matchState is using the
                               * dedicated dictionary search structure.
                               */
    optState_t opt;         /* optimal parser state */
    const ZSTD_MatchState_t* dictMatchState;
    ZSTD_compressionParameters cParams;
    const RawSeqStore_t* ldmSeqStore;

    /* Controls prefetching in some dictMatchState matchfinders.
     * This behavior is controlled from the cctx ms.
     * This parameter has no effect in the cdict ms. */
    int prefetchCDictTables;

    /* When == 0, lazy match finders insert every position.
     * When != 0, lazy match finders only insert positions they search.
     * This allows them to skip much faster over incompressible data,
     * at a small cost to compression ratio.
     */
    int lazySkipping;
};

typedef struct {
    ZSTD_compressedBlockState_t* prevCBlock;
    ZSTD_compressedBlockState_t* nextCBlock;
    ZSTD_MatchState_t matchState;
} ZSTD_blockState_t;

typedef struct {
    U32 offset;
    U32 checksum;
} ldmEntry_t;

typedef struct {
    BYTE const* split;
    U32 hash;
    U32 checksum;
    ldmEntry_t* bucket;
} ldmMatchCandidate_t;

#define LDM_BATCH_SIZE 64

typedef struct {
    ZSTD_window_t window;   /* State for the window round buffer management */
    ldmEntry_t* hashTable;
    U32 loadedDictEnd;
    BYTE* bucketOffsets;    /* Next position in bucket to insert entry */
    size_t splitIndices[LDM_BATCH_SIZE];
    ldmMatchCandidate_t matchCandidates[LDM_BATCH_SIZE];
} ldmState_t;

typedef struct {
    ZSTD_ParamSwitch_e enableLdm; /* ZSTD_ps_enable to enable LDM. ZSTD_ps_auto by default */
    U32 hashLog;            /* Log size of hashTable */
    U32 bucketSizeLog;      /* Log bucket size for collision resolution, at most 8 */
    U32 minMatchLength;     /* Minimum match length */
    U32 hashRateLog;       /* Log number of entries to skip */
    U32 windowLog;          /* Window log for the LDM */
} ldmParams_t;

typedef struct {
    int collectSequences;
    ZSTD_Sequence* seqStart;
    size_t seqIndex;
    size_t maxSequences;
} SeqCollector;

struct ZSTD_CCtx_params_s {
    ZSTD_format_e format;
    ZSTD_compressionParameters cParams;
    ZSTD_frameParameters fParams;

    int compressionLevel;
    int forceWindow;           /* force back-references to respect limit of
                                * 1<<wLog, even for dictionary */
    size_t targetCBlockSize;   /* Tries to fit compressed block size to be around targetCBlockSize.
                                * No target when targetCBlockSize == 0.
                                * There is no guarantee on compressed block size */
    int srcSizeHint;           /* User's best guess of source size.
                                * Hint is not valid when srcSizeHint == 0.
                                * There is no guarantee that hint is close to actual source size */

    ZSTD_dictAttachPref_e attachDictPref;
    ZSTD_ParamSwitch_e literalCompressionMode;

    /* Multithreading: used to pass parameters to mtctx */
    int nbWorkers;
    size_t jobSize;
    int overlapLog;
    int rsyncable;

    /* Long distance matching parameters */
    ldmParams_t ldmParams;

    /* Dedicated dict search algorithm trigger */
    int enableDedicatedDictSearch;

    /* Input/output buffer modes */
    ZSTD_bufferMode_e inBufferMode;
    ZSTD_bufferMode_e outBufferMode;

    /* Sequence compression API */
    ZSTD_SequenceFormat_e blockDelimiters;
    int validateSequences;

    /* Block splitting
     * @postBlockSplitter executes split analysis after sequences are produced,
     * it's more accurate but consumes more resources.
     * @preBlockSplitter_level splits before knowing sequences,
     * it's more approximative but also cheaper.
     * Valid @preBlockSplitter_level values range from 0 to 6 (included).
     * 0 means auto, 1 means do not split,
     * then levels are sorted in increasing cpu budget, from 2 (fastest) to 6 (slowest).
     * Highest @preBlockSplitter_level combines well with @postBlockSplitter.
     */
    ZSTD_ParamSwitch_e postBlockSplitter;
    int preBlockSplitter_level;

    /* Adjust the max block size*/
    size_t maxBlockSize;

    /* Param for deciding whether to use row-based matchfinder */
    ZSTD_ParamSwitch_e useRowMatchFinder;

    /* Always load a dictionary in ext-dict mode (not prefix mode)? */
    int deterministicRefPrefix;

    /* Internal use, for createCCtxParams() and freeCCtxParams() only */
    ZSTD_customMem customMem;

    /* Controls prefetching in some dictMatchState matchfinders */
    ZSTD_ParamSwitch_e prefetchCDictTables;

    /* Controls whether zstd will fall back to an internal matchfinder
     * if the external matchfinder returns an error code. */
    int enableMatchFinderFallback;

    /* Parameters for the external sequence producer API.
     * Users set these parameters through ZSTD_registerSequenceProducer().
     * It is not possible to set these parameters individually through the public API. */
    void* extSeqProdState;
    ZSTD_sequenceProducer_F extSeqProdFunc;

    /* Controls repcode search in external sequence parsing */
    ZSTD_ParamSwitch_e searchForExternalRepcodes;
};  /* typedef'd to ZSTD_CCtx_params within "zstd.h" */

#define COMPRESS_SEQUENCES_WORKSPACE_SIZE (sizeof(unsigned) * (MaxSeq + 2))
#define ENTROPY_WORKSPACE_SIZE (HUF_WORKSPACE_SIZE + COMPRESS_SEQUENCES_WORKSPACE_SIZE)
#define TMP_WORKSPACE_SIZE (MAX(ENTROPY_WORKSPACE_SIZE, ZSTD_SLIPBLOCK_WORKSPACESIZE))

/**
 * Indicates whether this compression proceeds directly from user-provided
 * source buffer to user-provided destination buffer (ZSTDb_not_buffered), or
 * whether the context needs to buffer the input/output (ZSTDb_buffered).
 */
typedef enum {
    ZSTDb_not_buffered,
    ZSTDb_buffered
} ZSTD_buffered_policy_e;

/**
 * Struct that contains all elements of block splitter that should be allocated
 * in a wksp.
 */
#define ZSTD_MAX_NB_BLOCK_SPLITS 196
typedef struct {
    SeqStore_t fullSeqStoreChunk;
    SeqStore_t firstHalfSeqStore;
    SeqStore_t secondHalfSeqStore;
    SeqStore_t currSeqStore;
    SeqStore_t nextSeqStore;

    U32 partitions[ZSTD_MAX_NB_BLOCK_SPLITS];
    ZSTD_entropyCTablesMetadata_t entropyMetadata;
} ZSTD_blockSplitCtx;

struct ZSTD_CCtx_s {
    ZSTD_compressionStage_e stage;
    int cParamsChanged;                  /* == 1 if cParams(except wlog) or compression level are changed in requestedParams. Triggers transmission of new params to ZSTDMT (if available) then reset to 0. */
    int bmi2;                            /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
    ZSTD_CCtx_params requestedParams;
    ZSTD_CCtx_params appliedParams;
    ZSTD_CCtx_params simpleApiParams;    /* Param storage used by the simple API - not sticky. Must only be used in top-level simple API functions for storage. */
    U32   dictID;
    size_t dictContentSize;

    ZSTD_cwksp workspace; /* manages buffer for dynamic allocations */
    size_t blockSizeMax;
    unsigned long long pledgedSrcSizePlusOne;  /* this way, 0 (default) == unknown */
    unsigned long long consumedSrcSize;
    unsigned long long producedCSize;
    XXH64_state_t xxhState;
    ZSTD_customMem customMem;
    ZSTD_threadPool* pool;
    size_t staticSize;
    SeqCollector seqCollector;
    int isFirstBlock;
    int initialized;

    SeqStore_t seqStore;      /* sequences storage ptrs */
    ldmState_t ldmState;      /* long distance matching state */
    rawSeq* ldmSequences;     /* Storage for the ldm output sequences */
    size_t maxNbLdmSequences;
    RawSeqStore_t externSeqStore; /* Mutable reference to external sequences */
    ZSTD_blockState_t blockState;
    void* tmpWorkspace;  /* used as substitute of stack space - must be aligned for S64 type */
    size_t tmpWkspSize;

    /* Whether we are streaming or not */
    ZSTD_buffered_policy_e bufferedPolicy;

    /* streaming */
    char*  inBuff;
    size_t inBuffSize;
    size_t inToCompress;
    size_t inBuffPos;
    size_t inBuffTarget;
    char*  outBuff;
    size_t outBuffSize;
    size_t outBuffContentSize;
    size_t outBuffFlushedSize;
    ZSTD_cStreamStage streamStage;
    U32    frameEnded;

    /* Stable in/out buffer verification */
    ZSTD_inBuffer expectedInBuffer;
    size_t stableIn_notConsumed; /* nb bytes within stable input buffer that are said to be consumed but are not */
    size_t expectedOutBufferSize;

    /* Dictionary */
    ZSTD_localDict localDict;
    const ZSTD_CDict* cdict;
    ZSTD_prefixDict prefixDict;   /* single-usage dictionary */

    /* Multi-threading */
#ifdef ZSTD_MULTITHREAD
    ZSTDMT_CCtx* mtctx;
#endif

    /* Tracing */
#if ZSTD_TRACE
    ZSTD_TraceCtx traceCtx;
#endif

    /* Workspace for block splitter */
    ZSTD_blockSplitCtx blockSplitCtx;

    /* Buffer for output from external sequence producer */
    ZSTD_Sequence* extSeqBuf;
    size_t extSeqBufCapacity;
};

typedef enum { ZSTD_dtlm_fast, ZSTD_dtlm_full } ZSTD_dictTableLoadMethod_e;
typedef enum { ZSTD_tfp_forCCtx, ZSTD_tfp_forCDict } ZSTD_tableFillPurpose_e;

typedef enum {
    ZSTD_noDict = 0,
    ZSTD_extDict = 1,
    ZSTD_dictMatchState = 2,
    ZSTD_dedicatedDictSearch = 3
} ZSTD_dictMode_e;

typedef enum {
    ZSTD_cpm_noAttachDict = 0,  /* Compression with ZSTD_noDict or ZSTD_extDict.
                                 * In this mode we use both the srcSize and the dictSize
                                 * when selecting and adjusting parameters.
                                 */
    ZSTD_cpm_attachDict = 1,    /* Compression with ZSTD_dictMatchState or ZSTD_dedicatedDictSearch.
                                 * In this mode we only take the srcSize into account when selecting
                                 * and adjusting parameters.
                                 */
    ZSTD_cpm_createCDict = 2,   /* Creating a CDict.
                                 * In this mode we take both the source size and the dictionary size
                                 * into account when selecting and adjusting the parameters.
                                 */
    ZSTD_cpm_unknown = 3        /* ZSTD_getCParams, ZSTD_getParams, ZSTD_adjustParams.
                                 * We don't know what these parameters are for. We default to the legacy
                                 * behavior of taking both the source size and the dict size into account
                                 * when selecting and adjusting parameters.
                                 */
} ZSTD_CParamMode_e;

typedef size_t (*ZSTD_BlockCompressor_f) (
        ZSTD_MatchState_t* bs, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
ZSTD_BlockCompressor_f ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_ParamSwitch_e rowMatchfinderMode, ZSTD_dictMode_e dictMode);


MEM_STATIC U32 ZSTD_LLcode(U32 litLength)
{
    static const BYTE LL_Code[64] = {  0,  1,  2,  3,  4,  5,  6,  7,
                                       8,  9, 10, 11, 12, 13, 14, 15,
                                      16, 16, 17, 17, 18, 18, 19, 19,
                                      20, 20, 20, 20, 21, 21, 21, 21,
                                      22, 22, 22, 22, 22, 22, 22, 22,
                                      23, 23, 23, 23, 23, 23, 23, 23,
                                      24, 24, 24, 24, 24, 24, 24, 24,
                                      24, 24, 24, 24, 24, 24, 24, 24 };
    static const U32 LL_deltaCode = 19;
    return (litLength > 63) ? ZSTD_highbit32(litLength) + LL_deltaCode : LL_Code[litLength];
}

/* ZSTD_MLcode() :
 * note : mlBase = matchLength - MINMATCH;
 *        because it's the format it's stored in seqStore->sequences */
MEM_STATIC U32 ZSTD_MLcode(U32 mlBase)
{
    static const BYTE ML_Code[128] = { 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15,
                                      16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31,
                                      32, 32, 33, 33, 34, 34, 35, 35, 36, 36, 36, 36, 37, 37, 37, 37,
                                      38, 38, 38, 38, 38, 38, 38, 38, 39, 39, 39, 39, 39, 39, 39, 39,
                                      40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40, 40,
                                      41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41, 41,
                                      42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42,
                                      42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42, 42 };
    static const U32 ML_deltaCode = 36;
    return (mlBase > 127) ? ZSTD_highbit32(mlBase) + ML_deltaCode : ML_Code[mlBase];
}

/* ZSTD_cParam_withinBounds:
 * @return 1 if value is within cParam bounds,
 * 0 otherwise */
MEM_STATIC int ZSTD_cParam_withinBounds(ZSTD_cParameter cParam, int value)
{
    ZSTD_bounds const bounds = ZSTD_cParam_getBounds(cParam);
    if (ZSTD_isError(bounds.error)) return 0;
    if (value < bounds.lowerBound) return 0;
    if (value > bounds.upperBound) return 0;
    return 1;
}

/* ZSTD_selectAddr:
 * @return index >= lowLimit ? candidate : backup,
 * tries to force branchless codegen. */
MEM_STATIC const BYTE*
ZSTD_selectAddr(U32 index, U32 lowLimit, const BYTE* candidate, const BYTE* backup)
{
#if defined(__GNUC__) && defined(__x86_64__)
    __asm__ (
        "cmp %1, %2\n"
        "cmova %3, %0\n"
        : "+r"(candidate)
        : "r"(index), "r"(lowLimit), "r"(backup)
        );
    return candidate;
#else
    return index >= lowLimit ? candidate : backup;
#endif
}

/* ZSTD_noCompressBlock() :
 * Writes uncompressed block to dst buffer from given src.
 * Returns the size of the block */
MEM_STATIC size_t
ZSTD_noCompressBlock(void* dst, size_t dstCapacity, const void* src, size_t srcSize, U32 lastBlock)
{
    U32 const cBlockHeader24 = lastBlock + (((U32)bt_raw)<<1) + (U32)(srcSize << 3);
    DEBUGLOG(5, "ZSTD_noCompressBlock (srcSize=%zu, dstCapacity=%zu)", srcSize, dstCapacity);
    RETURN_ERROR_IF(srcSize + ZSTD_blockHeaderSize > dstCapacity,
                    dstSize_tooSmall, "dst buf too small for uncompressed block");
    MEM_writeLE24(dst, cBlockHeader24);
    ZSTD_memcpy((BYTE*)dst + ZSTD_blockHeaderSize, src, srcSize);
    return ZSTD_blockHeaderSize + srcSize;
}

MEM_STATIC size_t
ZSTD_rleCompressBlock(void* dst, size_t dstCapacity, BYTE src, size_t srcSize, U32 lastBlock)
{
    BYTE* const op = (BYTE*)dst;
    U32 const cBlockHeader = lastBlock + (((U32)bt_rle)<<1) + (U32)(srcSize << 3);
    RETURN_ERROR_IF(dstCapacity < 4, dstSize_tooSmall, "");
    MEM_writeLE24(op, cBlockHeader);
    op[3] = src;
    return 4;
}


/* ZSTD_minGain() :
 * minimum compression required
 * to generate a compress block or a compressed literals section.
 * note : use same formula for both situations */
MEM_STATIC size_t ZSTD_minGain(size_t srcSize, ZSTD_strategy strat)
{
    U32 const minlog = (strat>=ZSTD_btultra) ? (U32)(strat) - 1 : 6;
    ZSTD_STATIC_ASSERT(ZSTD_btultra == 8);
    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
    return (srcSize >> minlog) + 2;
}

MEM_STATIC int ZSTD_literalsCompressionIsDisabled(const ZSTD_CCtx_params* cctxParams)
{
    switch (cctxParams->literalCompressionMode) {
    case ZSTD_ps_enable:
        return 0;
    case ZSTD_ps_disable:
        return 1;
    default:
        assert(0 /* impossible: pre-validated */);
        ZSTD_FALLTHROUGH;
    case ZSTD_ps_auto:
        return (cctxParams->cParams.strategy == ZSTD_fast) && (cctxParams->cParams.targetLength > 0);
    }
}

/*! ZSTD_safecopyLiterals() :
 *  memcpy() function that won't read beyond more than WILDCOPY_OVERLENGTH bytes past ilimit_w.
 *  Only called when the sequence ends past ilimit_w, so it only needs to be optimized for single
 *  large copies.
 */
static void
ZSTD_safecopyLiterals(BYTE* op, BYTE const* ip, BYTE const* const iend, BYTE const* ilimit_w)
{
    assert(iend > ilimit_w);
    if (ip <= ilimit_w) {
        ZSTD_wildcopy(op, ip, ilimit_w - ip, ZSTD_no_overlap);
        op += ilimit_w - ip;
        ip = ilimit_w;
    }
    while (ip < iend) *op++ = *ip++;
}


#define REPCODE1_TO_OFFBASE REPCODE_TO_OFFBASE(1)
#define REPCODE2_TO_OFFBASE REPCODE_TO_OFFBASE(2)
#define REPCODE3_TO_OFFBASE REPCODE_TO_OFFBASE(3)
#define REPCODE_TO_OFFBASE(r) (assert((r)>=1), assert((r)<=ZSTD_REP_NUM), (r)) /* accepts IDs 1,2,3 */
#define OFFSET_TO_OFFBASE(o)  (assert((o)>0), o + ZSTD_REP_NUM)
#define OFFBASE_IS_OFFSET(o)  ((o) > ZSTD_REP_NUM)
#define OFFBASE_IS_REPCODE(o) ( 1 <= (o) && (o) <= ZSTD_REP_NUM)
#define OFFBASE_TO_OFFSET(o)  (assert(OFFBASE_IS_OFFSET(o)), (o) - ZSTD_REP_NUM)
#define OFFBASE_TO_REPCODE(o) (assert(OFFBASE_IS_REPCODE(o)), (o))  /* returns ID 1,2,3 */

/*! ZSTD_storeSeqOnly() :
 *  Store a sequence (litlen, litPtr, offBase and matchLength) into SeqStore_t.
 *  Literals themselves are not copied, but @litPtr is updated.
 *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
 *  @matchLength : must be >= MINMATCH
*/
HINT_INLINE UNUSED_ATTR void
ZSTD_storeSeqOnly(SeqStore_t* seqStorePtr,
              size_t litLength,
              U32 offBase,
              size_t matchLength)
{
    assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);

    /* literal Length */
    assert(litLength <= ZSTD_BLOCKSIZE_MAX);
    if (UNLIKELY(litLength>0xFFFF)) {
        assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
        seqStorePtr->longLengthType = ZSTD_llt_literalLength;
        seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
    }
    seqStorePtr->sequences[0].litLength = (U16)litLength;

    /* match offset */
    seqStorePtr->sequences[0].offBase = offBase;

    /* match Length */
    assert(matchLength <= ZSTD_BLOCKSIZE_MAX);
    assert(matchLength >= MINMATCH);
    {   size_t const mlBase = matchLength - MINMATCH;
        if (UNLIKELY(mlBase>0xFFFF)) {
            assert(seqStorePtr->longLengthType == ZSTD_llt_none); /* there can only be a single long length */
            seqStorePtr->longLengthType = ZSTD_llt_matchLength;
            seqStorePtr->longLengthPos = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
        }
        seqStorePtr->sequences[0].mlBase = (U16)mlBase;
    }

    seqStorePtr->sequences++;
}

/*! ZSTD_storeSeq() :
 *  Store a sequence (litlen, litPtr, offBase and matchLength) into SeqStore_t.
 *  @offBase : Users should employ macros REPCODE_TO_OFFBASE() and OFFSET_TO_OFFBASE().
 *  @matchLength : must be >= MINMATCH
 *  Allowed to over-read literals up to litLimit.
*/
HINT_INLINE UNUSED_ATTR void
ZSTD_storeSeq(SeqStore_t* seqStorePtr,
              size_t litLength, const BYTE* literals, const BYTE* litLimit,
              U32 offBase,
              size_t matchLength)
{
    BYTE const* const litLimit_w = litLimit - WILDCOPY_OVERLENGTH;
    BYTE const* const litEnd = literals + litLength;
#if defined(DEBUGLEVEL) && (DEBUGLEVEL >= 6)
    static const BYTE* g_start = NULL;
    if (g_start==NULL) g_start = (const BYTE*)literals;  /* note : index only works for compression within a single segment */
    {   U32 const pos = (U32)((const BYTE*)literals - g_start);
        DEBUGLOG(6, "Cpos%7u :%3u literals, match%4u bytes at offBase%7u",
               pos, (U32)litLength, (U32)matchLength, (U32)offBase);
    }
#endif
    assert((size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart) < seqStorePtr->maxNbSeq);
    /* copy Literals */
    assert(seqStorePtr->maxNbLit <= 128 KB);
    assert(seqStorePtr->lit + litLength <= seqStorePtr->litStart + seqStorePtr->maxNbLit);
    assert(literals + litLength <= litLimit);
    if (litEnd <= litLimit_w) {
        /* Common case we can use wildcopy.
         * First copy 16 bytes, because literals are likely short.
         */
        ZSTD_STATIC_ASSERT(WILDCOPY_OVERLENGTH >= 16);
        ZSTD_copy16(seqStorePtr->lit, literals);
        if (litLength > 16) {
            ZSTD_wildcopy(seqStorePtr->lit+16, literals+16, (ptrdiff_t)litLength-16, ZSTD_no_overlap);
        }
    } else {
        ZSTD_safecopyLiterals(seqStorePtr->lit, literals, litEnd, litLimit_w);
    }
    seqStorePtr->lit += litLength;

    ZSTD_storeSeqOnly(seqStorePtr, litLength, offBase, matchLength);
}

/* ZSTD_updateRep() :
 * updates in-place @rep (array of repeat offsets)
 * @offBase : sum-type, using numeric representation of ZSTD_storeSeq()
 */
MEM_STATIC void
ZSTD_updateRep(U32 rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
{
    if (OFFBASE_IS_OFFSET(offBase)) {  /* full offset */
        rep[2] = rep[1];
        rep[1] = rep[0];
        rep[0] = OFFBASE_TO_OFFSET(offBase);
    } else {   /* repcode */
        U32 const repCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;
        if (repCode > 0) {  /* note : if repCode==0, no change */
            U32 const currentOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
            rep[2] = (repCode >= 2) ? rep[1] : rep[2];
            rep[1] = rep[0];
            rep[0] = currentOffset;
        } else {   /* repCode == 0 */
            /* nothing to do */
        }
    }
}

typedef struct repcodes_s {
    U32 rep[3];
} Repcodes_t;

MEM_STATIC Repcodes_t
ZSTD_newRep(U32 const rep[ZSTD_REP_NUM], U32 const offBase, U32 const ll0)
{
    Repcodes_t newReps;
    ZSTD_memcpy(&newReps, rep, sizeof(newReps));
    ZSTD_updateRep(newReps.rep, offBase, ll0);
    return newReps;
}


/*-*************************************
*  Match length counter
***************************************/
MEM_STATIC size_t ZSTD_count(const BYTE* pIn, const BYTE* pMatch, const BYTE* const pInLimit)
{
    const BYTE* const pStart = pIn;
    const BYTE* const pInLoopLimit = pInLimit - (sizeof(size_t)-1);

    if (pIn < pInLoopLimit) {
        { size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
          if (diff) return ZSTD_NbCommonBytes(diff); }
        pIn+=sizeof(size_t); pMatch+=sizeof(size_t);
        while (pIn < pInLoopLimit) {
            size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
            if (!diff) { pIn+=sizeof(size_t); pMatch+=sizeof(size_t); continue; }
            pIn += ZSTD_NbCommonBytes(diff);
            return (size_t)(pIn - pStart);
    }   }
    if (MEM_64bits() && (pIn<(pInLimit-3)) && (MEM_read32(pMatch) == MEM_read32(pIn))) { pIn+=4; pMatch+=4; }
    if ((pIn<(pInLimit-1)) && (MEM_read16(pMatch) == MEM_read16(pIn))) { pIn+=2; pMatch+=2; }
    if ((pIn<pInLimit) && (*pMatch == *pIn)) pIn++;
    return (size_t)(pIn - pStart);
}

/** ZSTD_count_2segments() :
 *  can count match length with `ip` & `match` in 2 different segments.
 *  convention : on reaching mEnd, match count continue starting from iStart
 */
MEM_STATIC size_t
ZSTD_count_2segments(const BYTE* ip, const BYTE* match,
                     const BYTE* iEnd, const BYTE* mEnd, const BYTE* iStart)
{
    const BYTE* const vEnd = MIN( ip + (mEnd - match), iEnd);
    size_t const matchLength = ZSTD_count(ip, match, vEnd);
    if (match + matchLength != mEnd) return matchLength;
    DEBUGLOG(7, "ZSTD_count_2segments: found a 2-parts match (current length==%zu)", matchLength);
    DEBUGLOG(7, "distance from match beginning to end dictionary = %i", (int)(mEnd - match));
    DEBUGLOG(7, "distance from current pos to end buffer = %i", (int)(iEnd - ip));
    DEBUGLOG(7, "next byte : ip==%02X, istart==%02X", ip[matchLength], *iStart);
    DEBUGLOG(7, "final match length = %zu", matchLength + ZSTD_count(ip+matchLength, iStart, iEnd));
    return matchLength + ZSTD_count(ip+matchLength, iStart, iEnd);
}


/*-*************************************
 *  Hashes
 ***************************************/
static const U32 prime3bytes = 506832829U;
static U32    ZSTD_hash3(U32 u, U32 h, U32 s) { assert(h <= 32); return (((u << (32-24)) * prime3bytes) ^ s)  >> (32-h) ; }
MEM_STATIC size_t ZSTD_hash3Ptr(const void* ptr, U32 h) { return ZSTD_hash3(MEM_readLE32(ptr), h, 0); } /* only in zstd_opt.h */
MEM_STATIC size_t ZSTD_hash3PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash3(MEM_readLE32(ptr), h, s); }

static const U32 prime4bytes = 2654435761U;
static U32    ZSTD_hash4(U32 u, U32 h, U32 s) { assert(h <= 32); return ((u * prime4bytes) ^ s) >> (32-h) ; }
static size_t ZSTD_hash4Ptr(const void* ptr, U32 h) { return ZSTD_hash4(MEM_readLE32(ptr), h, 0); }
static size_t ZSTD_hash4PtrS(const void* ptr, U32 h, U32 s) { return ZSTD_hash4(MEM_readLE32(ptr), h, s); }

static const U64 prime5bytes = 889523592379ULL;
static size_t ZSTD_hash5(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-40)) * prime5bytes) ^ s) >> (64-h)) ; }
static size_t ZSTD_hash5Ptr(const void* p, U32 h) { return ZSTD_hash5(MEM_readLE64(p), h, 0); }
static size_t ZSTD_hash5PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash5(MEM_readLE64(p), h, s); }

static const U64 prime6bytes = 227718039650203ULL;
static size_t ZSTD_hash6(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-48)) * prime6bytes) ^ s) >> (64-h)) ; }
static size_t ZSTD_hash6Ptr(const void* p, U32 h) { return ZSTD_hash6(MEM_readLE64(p), h, 0); }
static size_t ZSTD_hash6PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash6(MEM_readLE64(p), h, s); }

static const U64 prime7bytes = 58295818150454627ULL;
static size_t ZSTD_hash7(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u  << (64-56)) * prime7bytes) ^ s) >> (64-h)) ; }
static size_t ZSTD_hash7Ptr(const void* p, U32 h) { return ZSTD_hash7(MEM_readLE64(p), h, 0); }
static size_t ZSTD_hash7PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash7(MEM_readLE64(p), h, s); }

static const U64 prime8bytes = 0xCF1BBCDCB7A56463ULL;
static size_t ZSTD_hash8(U64 u, U32 h, U64 s) { assert(h <= 64); return (size_t)((((u) * prime8bytes)  ^ s) >> (64-h)) ; }
static size_t ZSTD_hash8Ptr(const void* p, U32 h) { return ZSTD_hash8(MEM_readLE64(p), h, 0); }
static size_t ZSTD_hash8PtrS(const void* p, U32 h, U64 s) { return ZSTD_hash8(MEM_readLE64(p), h, s); }


MEM_STATIC FORCE_INLINE_ATTR
size_t ZSTD_hashPtr(const void* p, U32 hBits, U32 mls)
{
    /* Although some of these hashes do support hBits up to 64, some do not.
     * To be on the safe side, always avoid hBits > 32. */
    assert(hBits <= 32);

    switch(mls)
    {
    default:
    case 4: return ZSTD_hash4Ptr(p, hBits);
    case 5: return ZSTD_hash5Ptr(p, hBits);
    case 6: return ZSTD_hash6Ptr(p, hBits);
    case 7: return ZSTD_hash7Ptr(p, hBits);
    case 8: return ZSTD_hash8Ptr(p, hBits);
    }
}

MEM_STATIC FORCE_INLINE_ATTR
size_t ZSTD_hashPtrSalted(const void* p, U32 hBits, U32 mls, const U64 hashSalt) {
    /* Although some of these hashes do support hBits up to 64, some do not.
     * To be on the safe side, always avoid hBits > 32. */
    assert(hBits <= 32);

    switch(mls)
    {
        default:
        case 4: return ZSTD_hash4PtrS(p, hBits, (U32)hashSalt);
        case 5: return ZSTD_hash5PtrS(p, hBits, hashSalt);
        case 6: return ZSTD_hash6PtrS(p, hBits, hashSalt);
        case 7: return ZSTD_hash7PtrS(p, hBits, hashSalt);
        case 8: return ZSTD_hash8PtrS(p, hBits, hashSalt);
    }
}


/** ZSTD_ipow() :
 * Return base^exponent.
 */
static U64 ZSTD_ipow(U64 base, U64 exponent)
{
    U64 power = 1;
    while (exponent) {
      if (exponent & 1) power *= base;
      exponent >>= 1;
      base *= base;
    }
    return power;
}

#define ZSTD_ROLL_HASH_CHAR_OFFSET 10

/** ZSTD_rollingHash_append() :
 * Add the buffer to the hash value.
 */
static U64 ZSTD_rollingHash_append(U64 hash, void const* buf, size_t size)
{
    BYTE const* istart = (BYTE const*)buf;
    size_t pos;
    for (pos = 0; pos < size; ++pos) {
        hash *= prime8bytes;
        hash += istart[pos] + ZSTD_ROLL_HASH_CHAR_OFFSET;
    }
    return hash;
}

/** ZSTD_rollingHash_compute() :
 * Compute the rolling hash value of the buffer.
 */
MEM_STATIC U64 ZSTD_rollingHash_compute(void const* buf, size_t size)
{
    return ZSTD_rollingHash_append(0, buf, size);
}

/** ZSTD_rollingHash_primePower() :
 * Compute the primePower to be passed to ZSTD_rollingHash_rotate() for a hash
 * over a window of length bytes.
 */
MEM_STATIC U64 ZSTD_rollingHash_primePower(U32 length)
{
    return ZSTD_ipow(prime8bytes, length - 1);
}

/** ZSTD_rollingHash_rotate() :
 * Rotate the rolling hash by one byte.
 */
MEM_STATIC U64 ZSTD_rollingHash_rotate(U64 hash, BYTE toRemove, BYTE toAdd, U64 primePower)
{
    hash -= (toRemove + ZSTD_ROLL_HASH_CHAR_OFFSET) * primePower;
    hash *= prime8bytes;
    hash += toAdd + ZSTD_ROLL_HASH_CHAR_OFFSET;
    return hash;
}

/*-*************************************
*  Round buffer management
***************************************/
/* Max @current value allowed:
 * In 32-bit mode: we want to avoid crossing the 2 GB limit,
 *                 reducing risks of side effects in case of signed operations on indexes.
 * In 64-bit mode: we want to ensure that adding the maximum job size (512 MB)
 *                 doesn't overflow U32 index capacity (4 GB) */
#define ZSTD_CURRENT_MAX (MEM_64bits() ? 3500U MB : 2000U MB)
/* Maximum chunk size before overflow correction needs to be called again */
#define ZSTD_CHUNKSIZE_MAX                                                     \
    ( ((U32)-1)                  /* Maximum ending current index */            \
    - ZSTD_CURRENT_MAX)          /* Maximum beginning lowLimit */

/**
 * ZSTD_window_clear():
 * Clears the window containing the history by simply setting it to empty.
 */
MEM_STATIC void ZSTD_window_clear(ZSTD_window_t* window)
{
    size_t const endT = (size_t)(window->nextSrc - window->base);
    U32 const end = (U32)endT;

    window->lowLimit = end;
    window->dictLimit = end;
}

MEM_STATIC U32 ZSTD_window_isEmpty(ZSTD_window_t const window)
{
    return window.dictLimit == ZSTD_WINDOW_START_INDEX &&
           window.lowLimit == ZSTD_WINDOW_START_INDEX &&
           (window.nextSrc - window.base) == ZSTD_WINDOW_START_INDEX;
}

/**
 * ZSTD_window_hasExtDict():
 * Returns non-zero if the window has a non-empty extDict.
 */
MEM_STATIC U32 ZSTD_window_hasExtDict(ZSTD_window_t const window)
{
    return window.lowLimit < window.dictLimit;
}

/**
 * ZSTD_matchState_dictMode():
 * Inspects the provided matchState and figures out what dictMode should be
 * passed to the compressor.
 */
MEM_STATIC ZSTD_dictMode_e ZSTD_matchState_dictMode(const ZSTD_MatchState_t *ms)
{
    return ZSTD_window_hasExtDict(ms->window) ?
        ZSTD_extDict :
        ms->dictMatchState != NULL ?
            (ms->dictMatchState->dedicatedDictSearch ? ZSTD_dedicatedDictSearch : ZSTD_dictMatchState) :
            ZSTD_noDict;
}

/* Defining this macro to non-zero tells zstd to run the overflow correction
 * code much more frequently. This is very inefficient, and should only be
 * used for tests and fuzzers.
 */
#ifndef ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY
#  ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 1
#  else
#    define ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY 0
#  endif
#endif

/**
 * ZSTD_window_canOverflowCorrect():
 * Returns non-zero if the indices are large enough for overflow correction
 * to work correctly without impacting compression ratio.
 */
MEM_STATIC U32 ZSTD_window_canOverflowCorrect(ZSTD_window_t const window,
                                              U32 cycleLog,
                                              U32 maxDist,
                                              U32 loadedDictEnd,
                                              void const* src)
{
    U32 const cycleSize = 1u << cycleLog;
    U32 const curr = (U32)((BYTE const*)src - window.base);
    U32 const minIndexToOverflowCorrect = cycleSize
                                        + MAX(maxDist, cycleSize)
                                        + ZSTD_WINDOW_START_INDEX;

    /* Adjust the min index to backoff the overflow correction frequency,
     * so we don't waste too much CPU in overflow correction. If this
     * computation overflows we don't really care, we just need to make
     * sure it is at least minIndexToOverflowCorrect.
     */
    U32 const adjustment = window.nbOverflowCorrections + 1;
    U32 const adjustedIndex = MAX(minIndexToOverflowCorrect * adjustment,
                                  minIndexToOverflowCorrect);
    U32 const indexLargeEnough = curr > adjustedIndex;

    /* Only overflow correct early if the dictionary is invalidated already,
     * so we don't hurt compression ratio.
     */
    U32 const dictionaryInvalidated = curr > maxDist + loadedDictEnd;

    return indexLargeEnough && dictionaryInvalidated;
}

/**
 * ZSTD_window_needOverflowCorrection():
 * Returns non-zero if the indices are getting too large and need overflow
 * protection.
 */
MEM_STATIC U32 ZSTD_window_needOverflowCorrection(ZSTD_window_t const window,
                                                  U32 cycleLog,
                                                  U32 maxDist,
                                                  U32 loadedDictEnd,
                                                  void const* src,
                                                  void const* srcEnd)
{
    U32 const curr = (U32)((BYTE const*)srcEnd - window.base);
    if (ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
        if (ZSTD_window_canOverflowCorrect(window, cycleLog, maxDist, loadedDictEnd, src)) {
            return 1;
        }
    }
    return curr > ZSTD_CURRENT_MAX;
}

/**
 * ZSTD_window_correctOverflow():
 * Reduces the indices to protect from index overflow.
 * Returns the correction made to the indices, which must be applied to every
 * stored index.
 *
 * The least significant cycleLog bits of the indices must remain the same,
 * which may be 0. Every index up to maxDist in the past must be valid.
 */
MEM_STATIC
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_window_correctOverflow(ZSTD_window_t* window, U32 cycleLog,
                                           U32 maxDist, void const* src)
{
    /* preemptive overflow correction:
     * 1. correction is large enough:
     *    lowLimit > (3<<29) ==> current > 3<<29 + 1<<windowLog
     *    1<<windowLog <= newCurrent < 1<<chainLog + 1<<windowLog
     *
     *    current - newCurrent
     *    > (3<<29 + 1<<windowLog) - (1<<windowLog + 1<<chainLog)
     *    > (3<<29) - (1<<chainLog)
     *    > (3<<29) - (1<<30)             (NOTE: chainLog <= 30)
     *    > 1<<29
     *
     * 2. (ip+ZSTD_CHUNKSIZE_MAX - cctx->base) doesn't overflow:
     *    After correction, current is less than (1<<chainLog + 1<<windowLog).
     *    In 64-bit mode we are safe, because we have 64-bit ptrdiff_t.
     *    In 32-bit mode we are safe, because (chainLog <= 29), so
     *    ip+ZSTD_CHUNKSIZE_MAX - cctx->base < 1<<32.
     * 3. (cctx->lowLimit + 1<<windowLog) < 1<<32:
     *    windowLog <= 31 ==> 3<<29 + 1<<windowLog < 7<<29 < 1<<32.
     */
    U32 const cycleSize = 1u << cycleLog;
    U32 const cycleMask = cycleSize - 1;
    U32 const curr = (U32)((BYTE const*)src - window->base);
    U32 const currentCycle = curr & cycleMask;
    /* Ensure newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX. */
    U32 const currentCycleCorrection = currentCycle < ZSTD_WINDOW_START_INDEX
                                     ? MAX(cycleSize, ZSTD_WINDOW_START_INDEX)
                                     : 0;
    U32 const newCurrent = currentCycle
                         + currentCycleCorrection
                         + MAX(maxDist, cycleSize);
    U32 const correction = curr - newCurrent;
    /* maxDist must be a power of two so that:
     *   (newCurrent & cycleMask) == (curr & cycleMask)
     * This is required to not corrupt the chains / binary tree.
     */
    assert((maxDist & (maxDist - 1)) == 0);
    assert((curr & cycleMask) == (newCurrent & cycleMask));
    assert(curr > newCurrent);
    if (!ZSTD_WINDOW_OVERFLOW_CORRECT_FREQUENTLY) {
        /* Loose bound, should be around 1<<29 (see above) */
        assert(correction > 1<<28);
    }

    window->base += correction;
    window->dictBase += correction;
    if (window->lowLimit < correction + ZSTD_WINDOW_START_INDEX) {
        window->lowLimit = ZSTD_WINDOW_START_INDEX;
    } else {
        window->lowLimit -= correction;
    }
    if (window->dictLimit < correction + ZSTD_WINDOW_START_INDEX) {
        window->dictLimit = ZSTD_WINDOW_START_INDEX;
    } else {
        window->dictLimit -= correction;
    }

    /* Ensure we can still reference the full window. */
    assert(newCurrent >= maxDist);
    assert(newCurrent - maxDist >= ZSTD_WINDOW_START_INDEX);
    /* Ensure that lowLimit and dictLimit didn't underflow. */
    assert(window->lowLimit <= newCurrent);
    assert(window->dictLimit <= newCurrent);

    ++window->nbOverflowCorrections;

    DEBUGLOG(4, "Correction of 0x%x bytes to lowLimit=0x%x", correction,
             window->lowLimit);
    return correction;
}

/**
 * ZSTD_window_enforceMaxDist():
 * Updates lowLimit so that:
 *    (srcEnd - base) - lowLimit == maxDist + loadedDictEnd
 *
 * It ensures index is valid as long as index >= lowLimit.
 * This must be called before a block compression call.
 *
 * loadedDictEnd is only defined if a dictionary is in use for current compression.
 * As the name implies, loadedDictEnd represents the index at end of dictionary.
 * The value lies within context's referential, it can be directly compared to blockEndIdx.
 *
 * If loadedDictEndPtr is NULL, no dictionary is in use, and we use loadedDictEnd == 0.
 * If loadedDictEndPtr is not NULL, we set it to zero after updating lowLimit.
 * This is because dictionaries are allowed to be referenced fully
 * as long as the last byte of the dictionary is in the window.
 * Once input has progressed beyond window size, dictionary cannot be referenced anymore.
 *
 * In normal dict mode, the dictionary lies between lowLimit and dictLimit.
 * In dictMatchState mode, lowLimit and dictLimit are the same,
 * and the dictionary is below them.
 * forceWindow and dictMatchState are therefore incompatible.
 */
MEM_STATIC void
ZSTD_window_enforceMaxDist(ZSTD_window_t* window,
                     const void* blockEnd,
                           U32   maxDist,
                           U32*  loadedDictEndPtr,
                     const ZSTD_MatchState_t** dictMatchStatePtr)
{
    U32 const blockEndIdx = (U32)((BYTE const*)blockEnd - window->base);
    U32 const loadedDictEnd = (loadedDictEndPtr != NULL) ? *loadedDictEndPtr : 0;
    DEBUGLOG(5, "ZSTD_window_enforceMaxDist: blockEndIdx=%u, maxDist=%u, loadedDictEnd=%u",
                (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);

    /* - When there is no dictionary : loadedDictEnd == 0.
         In which case, the test (blockEndIdx > maxDist) is merely to avoid
         overflowing next operation `newLowLimit = blockEndIdx - maxDist`.
       - When there is a standard dictionary :
         Index referential is copied from the dictionary,
         which means it starts from 0.
         In which case, loadedDictEnd == dictSize,
         and it makes sense to compare `blockEndIdx > maxDist + dictSize`
         since `blockEndIdx` also starts from zero.
       - When there is an attached dictionary :
         loadedDictEnd is expressed within the referential of the context,
         so it can be directly compared against blockEndIdx.
    */
    if (blockEndIdx > maxDist + loadedDictEnd) {
        U32 const newLowLimit = blockEndIdx - maxDist;
        if (window->lowLimit < newLowLimit) window->lowLimit = newLowLimit;
        if (window->dictLimit < window->lowLimit) {
            DEBUGLOG(5, "Update dictLimit to match lowLimit, from %u to %u",
                        (unsigned)window->dictLimit, (unsigned)window->lowLimit);
            window->dictLimit = window->lowLimit;
        }
        /* On reaching window size, dictionaries are invalidated */
        if (loadedDictEndPtr) *loadedDictEndPtr = 0;
        if (dictMatchStatePtr) *dictMatchStatePtr = NULL;
    }
}

/* Similar to ZSTD_window_enforceMaxDist(),
 * but only invalidates dictionary
 * when input progresses beyond window size.
 * assumption : loadedDictEndPtr and dictMatchStatePtr are valid (non NULL)
 *              loadedDictEnd uses same referential as window->base
 *              maxDist is the window size */
MEM_STATIC void
ZSTD_checkDictValidity(const ZSTD_window_t* window,
                       const void* blockEnd,
                             U32   maxDist,
                             U32*  loadedDictEndPtr,
                       const ZSTD_MatchState_t** dictMatchStatePtr)
{
    assert(loadedDictEndPtr != NULL);
    assert(dictMatchStatePtr != NULL);
    {   U32 const blockEndIdx = (U32)((BYTE const*)blockEnd - window->base);
        U32 const loadedDictEnd = *loadedDictEndPtr;
        DEBUGLOG(5, "ZSTD_checkDictValidity: blockEndIdx=%u, maxDist=%u, loadedDictEnd=%u",
                    (unsigned)blockEndIdx, (unsigned)maxDist, (unsigned)loadedDictEnd);
        assert(blockEndIdx >= loadedDictEnd);

        if (blockEndIdx > loadedDictEnd + maxDist || loadedDictEnd != window->dictLimit) {
            /* On reaching window size, dictionaries are invalidated.
             * For simplification, if window size is reached anywhere within next block,
             * the dictionary is invalidated for the full block.
             *
             * We also have to invalidate the dictionary if ZSTD_window_update() has detected
             * non-contiguous segments, which means that loadedDictEnd != window->dictLimit.
             * loadedDictEnd may be 0, if forceWindow is true, but in that case we never use
             * dictMatchState, so setting it to NULL is not a problem.
             */
            DEBUGLOG(6, "invalidating dictionary for current block (distance > windowSize)");
            *loadedDictEndPtr = 0;
            *dictMatchStatePtr = NULL;
        } else {
            if (*loadedDictEndPtr != 0) {
                DEBUGLOG(6, "dictionary considered valid for current block");
    }   }   }
}

MEM_STATIC void ZSTD_window_init(ZSTD_window_t* window) {
    ZSTD_memset(window, 0, sizeof(*window));
    window->base = (BYTE const*)" ";
    window->dictBase = (BYTE const*)" ";
    ZSTD_STATIC_ASSERT(ZSTD_DUBT_UNSORTED_MARK < ZSTD_WINDOW_START_INDEX); /* Start above ZSTD_DUBT_UNSORTED_MARK */
    window->dictLimit = ZSTD_WINDOW_START_INDEX;    /* start from >0, so that 1st position is valid */
    window->lowLimit = ZSTD_WINDOW_START_INDEX;     /* it ensures first and later CCtx usages compress the same */
    window->nextSrc = window->base + ZSTD_WINDOW_START_INDEX;   /* see issue #1241 */
    window->nbOverflowCorrections = 0;
}

/**
 * ZSTD_window_update():
 * Updates the window by appending [src, src + srcSize) to the window.
 * If it is not contiguous, the current prefix becomes the extDict, and we
 * forget about the extDict. Handles overlap of the prefix and extDict.
 * Returns non-zero if the segment is contiguous.
 */
MEM_STATIC
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_window_update(ZSTD_window_t* window,
                 const void* src, size_t srcSize,
                       int forceNonContiguous)
{
    BYTE const* const ip = (BYTE const*)src;
    U32 contiguous = 1;
    DEBUGLOG(5, "ZSTD_window_update");
    if (srcSize == 0)
        return contiguous;
    assert(window->base != NULL);
    assert(window->dictBase != NULL);
    /* Check if blocks follow each other */
    if (src != window->nextSrc || forceNonContiguous) {
        /* not contiguous */
        size_t const distanceFromBase = (size_t)(window->nextSrc - window->base);
        DEBUGLOG(5, "Non contiguous blocks, new segment starts at %u", window->dictLimit);
        window->lowLimit = window->dictLimit;
        assert(distanceFromBase == (size_t)(U32)distanceFromBase);  /* should never overflow */
        window->dictLimit = (U32)distanceFromBase;
        window->dictBase = window->base;
        window->base = ip - distanceFromBase;
        /* ms->nextToUpdate = window->dictLimit; */
        if (window->dictLimit - window->lowLimit < HASH_READ_SIZE) window->lowLimit = window->dictLimit;   /* too small extDict */
        contiguous = 0;
    }
    window->nextSrc = ip + srcSize;
    /* if input and dictionary overlap : reduce dictionary (area presumed modified by input) */
    if ( (ip+srcSize > window->dictBase + window->lowLimit)
       & (ip < window->dictBase + window->dictLimit)) {
        size_t const highInputIdx = (size_t)((ip + srcSize) - window->dictBase);
        U32 const lowLimitMax = (highInputIdx > (size_t)window->dictLimit) ? window->dictLimit : (U32)highInputIdx;
        assert(highInputIdx < UINT_MAX);
        window->lowLimit = lowLimitMax;
        DEBUGLOG(5, "Overlapping extDict and input : new lowLimit = %u", window->lowLimit);
    }
    return contiguous;
}

/**
 * Returns the lowest allowed match index. It may either be in the ext-dict or the prefix.
 */
MEM_STATIC U32 ZSTD_getLowestMatchIndex(const ZSTD_MatchState_t* ms, U32 curr, unsigned windowLog)
{
    U32 const maxDistance = 1U << windowLog;
    U32 const lowestValid = ms->window.lowLimit;
    U32 const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
    U32 const isDictionary = (ms->loadedDictEnd != 0);
    /* When using a dictionary the entire dictionary is valid if a single byte of the dictionary
     * is within the window. We invalidate the dictionary (and set loadedDictEnd to 0) when it isn't
     * valid for the entire block. So this check is sufficient to find the lowest valid match index.
     */
    U32 const matchLowest = isDictionary ? lowestValid : withinWindow;
    return matchLowest;
}

/**
 * Returns the lowest allowed match index in the prefix.
 */
MEM_STATIC U32 ZSTD_getLowestPrefixIndex(const ZSTD_MatchState_t* ms, U32 curr, unsigned windowLog)
{
    U32    const maxDistance = 1U << windowLog;
    U32    const lowestValid = ms->window.dictLimit;
    U32    const withinWindow = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
    U32    const isDictionary = (ms->loadedDictEnd != 0);
    /* When computing the lowest prefix index we need to take the dictionary into account to handle
     * the edge case where the dictionary and the source are contiguous in memory.
     */
    U32    const matchLowest = isDictionary ? lowestValid : withinWindow;
    return matchLowest;
}

/* index_safety_check:
 * intentional underflow : ensure repIndex isn't overlapping dict + prefix
 * @return 1 if values are not overlapping,
 * 0 otherwise */
MEM_STATIC int ZSTD_index_overlap_check(const U32 prefixLowestIndex, const U32 repIndex) {
    return ((U32)((prefixLowestIndex-1)  - repIndex) >= 3);
}


/* debug functions */
#if (DEBUGLEVEL>=2)

MEM_STATIC double ZSTD_fWeight(U32 rawStat)
{
    U32 const fp_accuracy = 8;
    U32 const fp_multiplier = (1 << fp_accuracy);
    U32 const newStat = rawStat + 1;
    U32 const hb = ZSTD_highbit32(newStat);
    U32 const BWeight = hb * fp_multiplier;
    U32 const FWeight = (newStat << fp_accuracy) >> hb;
    U32 const weight = BWeight + FWeight;
    assert(hb + fp_accuracy < 31);
    return (double)weight / fp_multiplier;
}

/* display a table content,
 * listing each element, its frequency, and its predicted bit cost */
MEM_STATIC void ZSTD_debugTable(const U32* table, U32 max)
{
    unsigned u, sum;
    for (u=0, sum=0; u<=max; u++) sum += table[u];
    DEBUGLOG(2, "total nb elts: %u", sum);
    for (u=0; u<=max; u++) {
        DEBUGLOG(2, "%2u: %5u  (%.2f)",
                u, table[u], ZSTD_fWeight(sum) - ZSTD_fWeight(table[u]) );
    }
}

#endif

/* Short Cache */

/* Normally, zstd matchfinders follow this flow:
 *     1. Compute hash at ip
 *     2. Load index from hashTable[hash]
 *     3. Check if *ip == *(base + index)
 * In dictionary compression, loading *(base + index) is often an L2 or even L3 miss.
 *
 * Short cache is an optimization which allows us to avoid step 3 most of the time
 * when the data doesn't actually match. With short cache, the flow becomes:
 *     1. Compute (hash, currentTag) at ip. currentTag is an 8-bit independent hash at ip.
 *     2. Load (index, matchTag) from hashTable[hash]. See ZSTD_writeTaggedIndex to understand how this works.
 *     3. Only if currentTag == matchTag, check *ip == *(base + index). Otherwise, continue.
 *
 * Currently, short cache is only implemented in CDict hashtables. Thus, its use is limited to
 * dictMatchState matchfinders.
 */
#define ZSTD_SHORT_CACHE_TAG_BITS 8
#define ZSTD_SHORT_CACHE_TAG_MASK ((1u << ZSTD_SHORT_CACHE_TAG_BITS) - 1)

/* Helper function for ZSTD_fillHashTable and ZSTD_fillDoubleHashTable.
 * Unpacks hashAndTag into (hash, tag), then packs (index, tag) into hashTable[hash]. */
MEM_STATIC void ZSTD_writeTaggedIndex(U32* const hashTable, size_t hashAndTag, U32 index) {
    size_t const hash = hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
    U32 const tag = (U32)(hashAndTag & ZSTD_SHORT_CACHE_TAG_MASK);
    assert(index >> (32 - ZSTD_SHORT_CACHE_TAG_BITS) == 0);
    hashTable[hash] = (index << ZSTD_SHORT_CACHE_TAG_BITS) | tag;
}

/* Helper function for short cache matchfinders.
 * Unpacks tag1 and tag2 from lower bits of packedTag1 and packedTag2, then checks if the tags match. */
MEM_STATIC int ZSTD_comparePackedTags(size_t packedTag1, size_t packedTag2) {
    U32 const tag1 = packedTag1 & ZSTD_SHORT_CACHE_TAG_MASK;
    U32 const tag2 = packedTag2 & ZSTD_SHORT_CACHE_TAG_MASK;
    return tag1 == tag2;
}

/* ===============================================================
 * Shared internal declarations
 * These prototypes may be called from sources not in lib/compress
 * =============================================================== */

/* ZSTD_loadCEntropy() :
 * dict : must point at beginning of a valid zstd dictionary.
 * return : size of dictionary header (size of magic number + dict ID + entropy tables)
 * assumptions : magic number supposed already checked
 *               and dictSize >= 8 */
size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
                         const void* const dict, size_t dictSize);

void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs);

typedef struct {
    U32 idx;            /* Index in array of ZSTD_Sequence */
    U32 posInSequence;  /* Position within sequence at idx */
    size_t posInSrc;    /* Number of bytes given by sequences provided so far */
} ZSTD_SequencePosition;

/* for benchmark */
size_t ZSTD_convertBlockSequences(ZSTD_CCtx* cctx,
                        const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                        int const repcodeResolution);

typedef struct {
    size_t nbSequences;
    size_t blockSize;
    size_t litSize;
} BlockSummary;

BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs);

/* ==============================================================
 * Private declarations
 * These prototypes shall only be called from within lib/compress
 * ============================================================== */

/* ZSTD_getCParamsFromCCtxParams() :
 * cParams are built depending on compressionLevel, src size hints,
 * LDM and manually set compression parameters.
 * Note: srcSizeHint == 0 means 0!
 */
ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);

/*! ZSTD_initCStream_internal() :
 *  Private use only. Init streaming operation.
 *  expects params to be valid.
 *  must receive dict, or cdict, or none, but not both.
 *  @return : 0, or an error code */
size_t ZSTD_initCStream_internal(ZSTD_CStream* zcs,
                     const void* dict, size_t dictSize,
                     const ZSTD_CDict* cdict,
                     const ZSTD_CCtx_params* params, unsigned long long pledgedSrcSize);

void ZSTD_resetSeqStore(SeqStore_t* ssPtr);

/*! ZSTD_getCParamsFromCDict() :
 *  as the name implies */
ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict);

/* ZSTD_compressBegin_advanced_internal() :
 * Private use only. To be called from zstdmt_compress.c. */
size_t ZSTD_compressBegin_advanced_internal(ZSTD_CCtx* cctx,
                                    const void* dict, size_t dictSize,
                                    ZSTD_dictContentType_e dictContentType,
                                    ZSTD_dictTableLoadMethod_e dtlm,
                                    const ZSTD_CDict* cdict,
                                    const ZSTD_CCtx_params* params,
                                    unsigned long long pledgedSrcSize);

/* ZSTD_compress_advanced_internal() :
 * Private use only. To be called from zstdmt_compress.c. */
size_t ZSTD_compress_advanced_internal(ZSTD_CCtx* cctx,
                                       void* dst, size_t dstCapacity,
                                 const void* src, size_t srcSize,
                                 const void* dict,size_t dictSize,
                                 const ZSTD_CCtx_params* params);


/* ZSTD_writeLastEmptyBlock() :
 * output an empty Block with end-of-frame mark to complete a frame
 * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
 *           or an error code if `dstCapacity` is too small (<ZSTD_blockHeaderSize)
 */
size_t ZSTD_writeLastEmptyBlock(void* dst, size_t dstCapacity);


/* ZSTD_referenceExternalSequences() :
 * Must be called before starting a compression operation.
 * seqs must parse a prefix of the source.
 * This cannot be used when long range matching is enabled.
 * Zstd will use these sequences, and pass the literals to a secondary block
 * compressor.
 * NOTE: seqs are not verified! Invalid sequences can cause out-of-bounds memory
 * access and data corruption.
 */
void ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq);

/** ZSTD_cycleLog() :
 *  condition for correct operation : hashLog > 1 */
U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat);

/** ZSTD_CCtx_trace() :
 *  Trace the end of a compression call.
 */
void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize);

/* Returns 1 if an external sequence producer is registered, otherwise returns 0. */
MEM_STATIC int ZSTD_hasExtSeqProd(const ZSTD_CCtx_params* params) {
    return params->extSeqProdFunc != NULL;
}

/* ===============================================================
 * Deprecated definitions that are still used internally to avoid
 * deprecation warnings. These functions are exactly equivalent to
 * their public variants, but avoid the deprecation warnings.
 * =============================================================== */

size_t ZSTD_compressBegin_usingCDict_deprecated(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict);

size_t ZSTD_compressContinue_public(ZSTD_CCtx* cctx,
                                    void* dst, size_t dstCapacity,
                              const void* src, size_t srcSize);

size_t ZSTD_compressEnd_public(ZSTD_CCtx* cctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize);

size_t ZSTD_compressBlock_deprecated(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize);


#endif /* ZSTD_COMPRESS_H */
/**** ended inlining zstd_compress_internal.h ****/


size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize);

/* ZSTD_compressRleLiteralsBlock() :
 * Conditions :
 * - All bytes in @src are identical
 * - dstCapacity >= 4 */
size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize);

/* ZSTD_compressLiterals():
 * @entropyWorkspace: must be aligned on 4-bytes boundaries
 * @entropyWorkspaceSize : must be >= HUF_WORKSPACE_SIZE
 * @suspectUncompressible: sampling checks, to potentially skip huffman coding
 */
size_t ZSTD_compressLiterals (void* dst, size_t dstCapacity,
                        const void* src, size_t srcSize,
                              void* entropyWorkspace, size_t entropyWorkspaceSize,
                        const ZSTD_hufCTables_t* prevHuf,
                              ZSTD_hufCTables_t* nextHuf,
                              ZSTD_strategy strategy, int disableLiteralCompression,
                              int suspectUncompressible,
                              int bmi2);

#endif /* ZSTD_COMPRESS_LITERALS_H */
/**** ended inlining zstd_compress_literals.h ****/


/* **************************************************************
*  Debug Traces
****************************************************************/
#if DEBUGLEVEL >= 2

static size_t showHexa(const void* src, size_t srcSize)
{
    const BYTE* const ip = (const BYTE*)src;
    size_t u;
    for (u=0; u<srcSize; u++) {
        RAWLOG(5, " %02X", ip[u]); (void)ip;
    }
    RAWLOG(5, " \n");
    return srcSize;
}

#endif


/* **************************************************************
*  Literals compression - special cases
****************************************************************/
size_t ZSTD_noCompressLiterals (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    BYTE* const ostart = (BYTE*)dst;
    U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);

    DEBUGLOG(5, "ZSTD_noCompressLiterals: srcSize=%zu, dstCapacity=%zu", srcSize, dstCapacity);

    RETURN_ERROR_IF(srcSize + flSize > dstCapacity, dstSize_tooSmall, "");

    switch(flSize)
    {
        case 1: /* 2 - 1 - 5 */
            ostart[0] = (BYTE)((U32)set_basic + (srcSize<<3));
            break;
        case 2: /* 2 - 2 - 12 */
            MEM_writeLE16(ostart, (U16)((U32)set_basic + (1<<2) + (srcSize<<4)));
            break;
        case 3: /* 2 - 2 - 20 */
            MEM_writeLE32(ostart, (U32)((U32)set_basic + (3<<2) + (srcSize<<4)));
            break;
        default:   /* not necessary : flSize is {1,2,3} */
            assert(0);
    }

    ZSTD_memcpy(ostart + flSize, src, srcSize);
    DEBUGLOG(5, "Raw (uncompressed) literals: %u -> %u", (U32)srcSize, (U32)(srcSize + flSize));
    return srcSize + flSize;
}

static int allBytesIdentical(const void* src, size_t srcSize)
{
    assert(srcSize >= 1);
    assert(src != NULL);
    {   const BYTE b = ((const BYTE*)src)[0];
        size_t p;
        for (p=1; p<srcSize; p++) {
            if (((const BYTE*)src)[p] != b) return 0;
        }
        return 1;
    }
}

size_t ZSTD_compressRleLiteralsBlock (void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    BYTE* const ostart = (BYTE*)dst;
    U32   const flSize = 1 + (srcSize>31) + (srcSize>4095);

    assert(dstCapacity >= 4); (void)dstCapacity;
    assert(allBytesIdentical(src, srcSize));

    switch(flSize)
    {
        case 1: /* 2 - 1 - 5 */
            ostart[0] = (BYTE)((U32)set_rle + (srcSize<<3));
            break;
        case 2: /* 2 - 2 - 12 */
            MEM_writeLE16(ostart, (U16)((U32)set_rle + (1<<2) + (srcSize<<4)));
            break;
        case 3: /* 2 - 2 - 20 */
            MEM_writeLE32(ostart, (U32)((U32)set_rle + (3<<2) + (srcSize<<4)));
            break;
        default:   /* not necessary : flSize is {1,2,3} */
            assert(0);
    }

    ostart[flSize] = *(const BYTE*)src;
    DEBUGLOG(5, "RLE : Repeated Literal (%02X: %u times) -> %u bytes encoded", ((const BYTE*)src)[0], (U32)srcSize, (U32)flSize + 1);
    return flSize+1;
}

/* ZSTD_minLiteralsToCompress() :
 * returns minimal amount of literals
 * for literal compression to even be attempted.
 * Minimum is made tighter as compression strategy increases.
 */
static size_t
ZSTD_minLiteralsToCompress(ZSTD_strategy strategy, HUF_repeat huf_repeat)
{
    assert((int)strategy >= 0);
    assert((int)strategy <= 9);
    /* btultra2 : min 8 bytes;
     * then 2x larger for each successive compression strategy
     * max threshold 64 bytes */
    {   int const shift = MIN(9-(int)strategy, 3);
        size_t const mintc = (huf_repeat == HUF_repeat_valid) ? 6 : (size_t)8 << shift;
        DEBUGLOG(7, "minLiteralsToCompress = %zu", mintc);
        return mintc;
    }
}

size_t ZSTD_compressLiterals (
                  void* dst, size_t dstCapacity,
            const void* src, size_t srcSize,
                  void* entropyWorkspace, size_t entropyWorkspaceSize,
            const ZSTD_hufCTables_t* prevHuf,
                  ZSTD_hufCTables_t* nextHuf,
                  ZSTD_strategy strategy,
                  int disableLiteralCompression,
                  int suspectUncompressible,
                  int bmi2)
{
    size_t const lhSize = 3 + (srcSize >= 1 KB) + (srcSize >= 16 KB);
    BYTE*  const ostart = (BYTE*)dst;
    U32 singleStream = srcSize < 256;
    SymbolEncodingType_e hType = set_compressed;
    size_t cLitSize;

    DEBUGLOG(5,"ZSTD_compressLiterals (disableLiteralCompression=%i, srcSize=%u, dstCapacity=%zu)",
                disableLiteralCompression, (U32)srcSize, dstCapacity);

    DEBUGLOG(6, "Completed literals listing (%zu bytes)", showHexa(src, srcSize));

    /* Prepare nextEntropy assuming reusing the existing table */
    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));

    if (disableLiteralCompression)
        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);

    /* if too small, don't even attempt compression (speed opt) */
    if (srcSize < ZSTD_minLiteralsToCompress(strategy, prevHuf->repeatMode))
        return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);

    RETURN_ERROR_IF(dstCapacity < lhSize+1, dstSize_tooSmall, "not enough space for compression");
    {   HUF_repeat repeat = prevHuf->repeatMode;
        int const flags = 0
            | (bmi2 ? HUF_flags_bmi2 : 0)
            | (strategy < ZSTD_lazy && srcSize <= 1024 ? HUF_flags_preferRepeat : 0)
            | (strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD ? HUF_flags_optimalDepth : 0)
            | (suspectUncompressible ? HUF_flags_suspectUncompressible : 0);

        typedef size_t (*huf_compress_f)(void*, size_t, const void*, size_t, unsigned, unsigned, void*, size_t, HUF_CElt*, HUF_repeat*, int);
        huf_compress_f huf_compress;
        if (repeat == HUF_repeat_valid && lhSize == 3) singleStream = 1;
        huf_compress = singleStream ? HUF_compress1X_repeat : HUF_compress4X_repeat;
        cLitSize = huf_compress(ostart+lhSize, dstCapacity-lhSize,
                                src, srcSize,
                                HUF_SYMBOLVALUE_MAX, LitHufLog,
                                entropyWorkspace, entropyWorkspaceSize,
                                (HUF_CElt*)nextHuf->CTable,
                                &repeat, flags);
        DEBUGLOG(5, "%zu literals compressed into %zu bytes (before header)", srcSize, cLitSize);
        if (repeat != HUF_repeat_none) {
            /* reused the existing table */
            DEBUGLOG(5, "reusing statistics from previous huffman block");
            hType = set_repeat;
        }
    }

    {   size_t const minGain = ZSTD_minGain(srcSize, strategy);
        if ((cLitSize==0) || (cLitSize >= srcSize - minGain) || ERR_isError(cLitSize)) {
            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
            return ZSTD_noCompressLiterals(dst, dstCapacity, src, srcSize);
    }   }
    if (cLitSize==1) {
        /* A return value of 1 signals that the alphabet consists of a single symbol.
         * However, in some rare circumstances, it could be the compressed size (a single byte).
         * For that outcome to have a chance to happen, it's necessary that `srcSize < 8`.
         * (it's also necessary to not generate statistics).
         * Therefore, in such a case, actively check that all bytes are identical. */
        if ((srcSize >= 8) || allBytesIdentical(src, srcSize)) {
            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
            return ZSTD_compressRleLiteralsBlock(dst, dstCapacity, src, srcSize);
    }   }

    if (hType == set_compressed) {
        /* using a newly constructed table */
        nextHuf->repeatMode = HUF_repeat_check;
    }

    /* Build header */
    switch(lhSize)
    {
    case 3: /* 2 - 2 - 10 - 10 */
        if (!singleStream) assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<14);
            MEM_writeLE24(ostart, lhc);
            break;
        }
    case 4: /* 2 - 2 - 14 - 14 */
        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
        {   U32 const lhc = hType + (2 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<18);
            MEM_writeLE32(ostart, lhc);
            break;
        }
    case 5: /* 2 - 2 - 18 - 18 */
        assert(srcSize >= MIN_LITERALS_FOR_4_STREAMS);
        {   U32 const lhc = hType + (3 << 2) + ((U32)srcSize<<4) + ((U32)cLitSize<<22);
            MEM_writeLE32(ostart, lhc);
            ostart[4] = (BYTE)(cLitSize >> 10);
            break;
        }
    default:  /* not possible : lhSize is {3,4,5} */
        assert(0);
    }
    DEBUGLOG(5, "Compressed literals: %u -> %u", (U32)srcSize, (U32)(lhSize+cLitSize));
    return lhSize+cLitSize;
}
/**** ended inlining compress/zstd_compress_literals.c ****/
/**** start inlining compress/zstd_compress_sequences.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

 /*-*************************************
 *  Dependencies
 ***************************************/
/**** start inlining zstd_compress_sequences.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_COMPRESS_SEQUENCES_H
#define ZSTD_COMPRESS_SEQUENCES_H

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/zstd_internal.h ****/

typedef enum {
    ZSTD_defaultDisallowed = 0,
    ZSTD_defaultAllowed = 1
} ZSTD_DefaultPolicy_e;

SymbolEncodingType_e
ZSTD_selectEncodingType(
        FSE_repeat* repeatMode, unsigned const* count, unsigned const max,
        size_t const mostFrequent, size_t nbSeq, unsigned const FSELog,
        FSE_CTable const* prevCTable,
        short const* defaultNorm, U32 defaultNormLog,
        ZSTD_DefaultPolicy_e const isDefaultAllowed,
        ZSTD_strategy const strategy);

size_t
ZSTD_buildCTable(void* dst, size_t dstCapacity,
                FSE_CTable* nextCTable, U32 FSELog, SymbolEncodingType_e type,
                unsigned* count, U32 max,
                const BYTE* codeTable, size_t nbSeq,
                const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                const FSE_CTable* prevCTable, size_t prevCTableSize,
                void* entropyWorkspace, size_t entropyWorkspaceSize);

size_t ZSTD_encodeSequences(
            void* dst, size_t dstCapacity,
            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
            SeqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2);

size_t ZSTD_fseBitCost(
    FSE_CTable const* ctable,
    unsigned const* count,
    unsigned const max);

size_t ZSTD_crossEntropyCost(short const* norm, unsigned accuracyLog,
                             unsigned const* count, unsigned const max);
#endif /* ZSTD_COMPRESS_SEQUENCES_H */
/**** ended inlining zstd_compress_sequences.h ****/

/**
 * -log2(x / 256) lookup table for x in [0, 256).
 * If x == 0: Return 0
 * Else: Return floor(-log2(x / 256) * 256)
 */
static unsigned const kInverseProbabilityLog256[256] = {
    0,    2048, 1792, 1642, 1536, 1453, 1386, 1329, 1280, 1236, 1197, 1162,
    1130, 1100, 1073, 1047, 1024, 1001, 980,  960,  941,  923,  906,  889,
    874,  859,  844,  830,  817,  804,  791,  779,  768,  756,  745,  734,
    724,  714,  704,  694,  685,  676,  667,  658,  650,  642,  633,  626,
    618,  610,  603,  595,  588,  581,  574,  567,  561,  554,  548,  542,
    535,  529,  523,  517,  512,  506,  500,  495,  489,  484,  478,  473,
    468,  463,  458,  453,  448,  443,  438,  434,  429,  424,  420,  415,
    411,  407,  402,  398,  394,  390,  386,  382,  377,  373,  370,  366,
    362,  358,  354,  350,  347,  343,  339,  336,  332,  329,  325,  322,
    318,  315,  311,  308,  305,  302,  298,  295,  292,  289,  286,  282,
    279,  276,  273,  270,  267,  264,  261,  258,  256,  253,  250,  247,
    244,  241,  239,  236,  233,  230,  228,  225,  222,  220,  217,  215,
    212,  209,  207,  204,  202,  199,  197,  194,  192,  190,  187,  185,
    182,  180,  178,  175,  173,  171,  168,  166,  164,  162,  159,  157,
    155,  153,  151,  149,  146,  144,  142,  140,  138,  136,  134,  132,
    130,  128,  126,  123,  121,  119,  117,  115,  114,  112,  110,  108,
    106,  104,  102,  100,  98,   96,   94,   93,   91,   89,   87,   85,
    83,   82,   80,   78,   76,   74,   73,   71,   69,   67,   66,   64,
    62,   61,   59,   57,   55,   54,   52,   50,   49,   47,   46,   44,
    42,   41,   39,   37,   36,   34,   33,   31,   30,   28,   26,   25,
    23,   22,   20,   19,   17,   16,   14,   13,   11,   10,   8,    7,
    5,    4,    2,    1,
};

static unsigned ZSTD_getFSEMaxSymbolValue(FSE_CTable const* ctable) {
  void const* ptr = ctable;
  U16 const* u16ptr = (U16 const*)ptr;
  U32 const maxSymbolValue = MEM_read16(u16ptr + 1);
  return maxSymbolValue;
}

/**
 * Returns true if we should use ncount=-1 else we should
 * use ncount=1 for low probability symbols instead.
 */
static unsigned ZSTD_useLowProbCount(size_t const nbSeq)
{
    /* Heuristic: This should cover most blocks <= 16K and
     * start to fade out after 16K to about 32K depending on
     * compressibility.
     */
    return nbSeq >= 2048;
}

/**
 * Returns the cost in bytes of encoding the normalized count header.
 * Returns an error if any of the helper functions return an error.
 */
static size_t ZSTD_NCountCost(unsigned const* count, unsigned const max,
                              size_t const nbSeq, unsigned const FSELog)
{
    BYTE wksp[FSE_NCOUNTBOUND];
    S16 norm[MaxSeq + 1];
    const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
    FORWARD_IF_ERROR(FSE_normalizeCount(norm, tableLog, count, nbSeq, max, ZSTD_useLowProbCount(nbSeq)), "");
    return FSE_writeNCount(wksp, sizeof(wksp), norm, max, tableLog);
}

/**
 * Returns the cost in bits of encoding the distribution described by count
 * using the entropy bound.
 */
static size_t ZSTD_entropyCost(unsigned const* count, unsigned const max, size_t const total)
{
    unsigned cost = 0;
    unsigned s;

    assert(total > 0);
    for (s = 0; s <= max; ++s) {
        unsigned norm = (unsigned)((256 * count[s]) / total);
        if (count[s] != 0 && norm == 0)
            norm = 1;
        assert(count[s] < total);
        cost += count[s] * kInverseProbabilityLog256[norm];
    }
    return cost >> 8;
}

/**
 * Returns the cost in bits of encoding the distribution in count using ctable.
 * Returns an error if ctable cannot represent all the symbols in count.
 */
size_t ZSTD_fseBitCost(
    FSE_CTable const* ctable,
    unsigned const* count,
    unsigned const max)
{
    unsigned const kAccuracyLog = 8;
    size_t cost = 0;
    unsigned s;
    FSE_CState_t cstate;
    FSE_initCState(&cstate, ctable);
    if (ZSTD_getFSEMaxSymbolValue(ctable) < max) {
        DEBUGLOG(5, "Repeat FSE_CTable has maxSymbolValue %u < %u",
                    ZSTD_getFSEMaxSymbolValue(ctable), max);
        return ERROR(GENERIC);
    }
    for (s = 0; s <= max; ++s) {
        unsigned const tableLog = cstate.stateLog;
        unsigned const badCost = (tableLog + 1) << kAccuracyLog;
        unsigned const bitCost = FSE_bitCost(cstate.symbolTT, tableLog, s, kAccuracyLog);
        if (count[s] == 0)
            continue;
        if (bitCost >= badCost) {
            DEBUGLOG(5, "Repeat FSE_CTable has Prob[%u] == 0", s);
            return ERROR(GENERIC);
        }
        cost += (size_t)count[s] * bitCost;
    }
    return cost >> kAccuracyLog;
}

/**
 * Returns the cost in bits of encoding the distribution in count using the
 * table described by norm. The max symbol support by norm is assumed >= max.
 * norm must be valid for every symbol with non-zero probability in count.
 */
size_t ZSTD_crossEntropyCost(short const* norm, unsigned accuracyLog,
                             unsigned const* count, unsigned const max)
{
    unsigned const shift = 8 - accuracyLog;
    size_t cost = 0;
    unsigned s;
    assert(accuracyLog <= 8);
    for (s = 0; s <= max; ++s) {
        unsigned const normAcc = (norm[s] != -1) ? (unsigned)norm[s] : 1;
        unsigned const norm256 = normAcc << shift;
        assert(norm256 > 0);
        assert(norm256 < 256);
        cost += count[s] * kInverseProbabilityLog256[norm256];
    }
    return cost >> 8;
}

SymbolEncodingType_e
ZSTD_selectEncodingType(
        FSE_repeat* repeatMode, unsigned const* count, unsigned const max,
        size_t const mostFrequent, size_t nbSeq, unsigned const FSELog,
        FSE_CTable const* prevCTable,
        short const* defaultNorm, U32 defaultNormLog,
        ZSTD_DefaultPolicy_e const isDefaultAllowed,
        ZSTD_strategy const strategy)
{
    ZSTD_STATIC_ASSERT(ZSTD_defaultDisallowed == 0 && ZSTD_defaultAllowed != 0);
    if (mostFrequent == nbSeq) {
        *repeatMode = FSE_repeat_none;
        if (isDefaultAllowed && nbSeq <= 2) {
            /* Prefer set_basic over set_rle when there are 2 or fewer symbols,
             * since RLE uses 1 byte, but set_basic uses 5-6 bits per symbol.
             * If basic encoding isn't possible, always choose RLE.
             */
            DEBUGLOG(5, "Selected set_basic");
            return set_basic;
        }
        DEBUGLOG(5, "Selected set_rle");
        return set_rle;
    }
    if (strategy < ZSTD_lazy) {
        if (isDefaultAllowed) {
            size_t const staticFse_nbSeq_max = 1000;
            size_t const mult = 10 - strategy;
            size_t const baseLog = 3;
            size_t const dynamicFse_nbSeq_min = (((size_t)1 << defaultNormLog) * mult) >> baseLog;  /* 28-36 for offset, 56-72 for lengths */
            assert(defaultNormLog >= 5 && defaultNormLog <= 6);  /* xx_DEFAULTNORMLOG */
            assert(mult <= 9 && mult >= 7);
            if ( (*repeatMode == FSE_repeat_valid)
              && (nbSeq < staticFse_nbSeq_max) ) {
                DEBUGLOG(5, "Selected set_repeat");
                return set_repeat;
            }
            if ( (nbSeq < dynamicFse_nbSeq_min)
              || (mostFrequent < (nbSeq >> (defaultNormLog-1))) ) {
                DEBUGLOG(5, "Selected set_basic");
                /* The format allows default tables to be repeated, but it isn't useful.
                 * When using simple heuristics to select encoding type, we don't want
                 * to confuse these tables with dictionaries. When running more careful
                 * analysis, we don't need to waste time checking both repeating tables
                 * and default tables.
                 */
                *repeatMode = FSE_repeat_none;
                return set_basic;
            }
        }
    } else {
        size_t const basicCost = isDefaultAllowed ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, count, max) : ERROR(GENERIC);
        size_t const repeatCost = *repeatMode != FSE_repeat_none ? ZSTD_fseBitCost(prevCTable, count, max) : ERROR(GENERIC);
        size_t const NCountCost = ZSTD_NCountCost(count, max, nbSeq, FSELog);
        size_t const compressedCost = (NCountCost << 3) + ZSTD_entropyCost(count, max, nbSeq);

        if (isDefaultAllowed) {
            assert(!ZSTD_isError(basicCost));
            assert(!(*repeatMode == FSE_repeat_valid && ZSTD_isError(repeatCost)));
        }
        assert(!ZSTD_isError(NCountCost));
        assert(compressedCost < ERROR(maxCode));
        DEBUGLOG(5, "Estimated bit costs: basic=%u\trepeat=%u\tcompressed=%u",
                    (unsigned)basicCost, (unsigned)repeatCost, (unsigned)compressedCost);
        if (basicCost <= repeatCost && basicCost <= compressedCost) {
            DEBUGLOG(5, "Selected set_basic");
            assert(isDefaultAllowed);
            *repeatMode = FSE_repeat_none;
            return set_basic;
        }
        if (repeatCost <= compressedCost) {
            DEBUGLOG(5, "Selected set_repeat");
            assert(!ZSTD_isError(repeatCost));
            return set_repeat;
        }
        assert(compressedCost < basicCost && compressedCost < repeatCost);
    }
    DEBUGLOG(5, "Selected set_compressed");
    *repeatMode = FSE_repeat_check;
    return set_compressed;
}

typedef struct {
    S16 norm[MaxSeq + 1];
    U32 wksp[FSE_BUILD_CTABLE_WORKSPACE_SIZE_U32(MaxSeq, MaxFSELog)];
} ZSTD_BuildCTableWksp;

size_t
ZSTD_buildCTable(void* dst, size_t dstCapacity,
                FSE_CTable* nextCTable, U32 FSELog, SymbolEncodingType_e type,
                unsigned* count, U32 max,
                const BYTE* codeTable, size_t nbSeq,
                const S16* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                const FSE_CTable* prevCTable, size_t prevCTableSize,
                void* entropyWorkspace, size_t entropyWorkspaceSize)
{
    BYTE* op = (BYTE*)dst;
    const BYTE* const oend = op + dstCapacity;
    DEBUGLOG(6, "ZSTD_buildCTable (dstCapacity=%u)", (unsigned)dstCapacity);

    switch (type) {
    case set_rle:
        FORWARD_IF_ERROR(FSE_buildCTable_rle(nextCTable, (BYTE)max), "");
        RETURN_ERROR_IF(dstCapacity==0, dstSize_tooSmall, "not enough space");
        *op = codeTable[0];
        return 1;
    case set_repeat:
        ZSTD_memcpy(nextCTable, prevCTable, prevCTableSize);
        return 0;
    case set_basic:
        FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, defaultNorm, defaultMax, defaultNormLog, entropyWorkspace, entropyWorkspaceSize), "");  /* note : could be pre-calculated */
        return 0;
    case set_compressed: {
        ZSTD_BuildCTableWksp* wksp = (ZSTD_BuildCTableWksp*)entropyWorkspace;
        size_t nbSeq_1 = nbSeq;
        const U32 tableLog = FSE_optimalTableLog(FSELog, nbSeq, max);
        if (count[codeTable[nbSeq-1]] > 1) {
            count[codeTable[nbSeq-1]]--;
            nbSeq_1--;
        }
        assert(nbSeq_1 > 1);
        assert(entropyWorkspaceSize >= sizeof(ZSTD_BuildCTableWksp));
        (void)entropyWorkspaceSize;
        FORWARD_IF_ERROR(FSE_normalizeCount(wksp->norm, tableLog, count, nbSeq_1, max, ZSTD_useLowProbCount(nbSeq_1)), "FSE_normalizeCount failed");
        assert(oend >= op);
        {   size_t const NCountSize = FSE_writeNCount(op, (size_t)(oend - op), wksp->norm, max, tableLog);   /* overflow protected */
            FORWARD_IF_ERROR(NCountSize, "FSE_writeNCount failed");
            FORWARD_IF_ERROR(FSE_buildCTable_wksp(nextCTable, wksp->norm, max, tableLog, wksp->wksp, sizeof(wksp->wksp)), "FSE_buildCTable_wksp failed");
            return NCountSize;
        }
    }
    default: assert(0); RETURN_ERROR(GENERIC, "impossible to reach");
    }
}

FORCE_INLINE_TEMPLATE size_t
ZSTD_encodeSequences_body(
            void* dst, size_t dstCapacity,
            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
            SeqDef const* sequences, size_t nbSeq, int longOffsets)
{
    BIT_CStream_t blockStream;
    FSE_CState_t  stateMatchLength;
    FSE_CState_t  stateOffsetBits;
    FSE_CState_t  stateLitLength;

    RETURN_ERROR_IF(
        ERR_isError(BIT_initCStream(&blockStream, dst, dstCapacity)),
        dstSize_tooSmall, "not enough space remaining");
    DEBUGLOG(6, "available space for bitstream : %i  (dstCapacity=%u)",
                (int)(blockStream.endPtr - blockStream.startPtr),
                (unsigned)dstCapacity);

    /* first symbols */
    FSE_initCState2(&stateMatchLength, CTable_MatchLength, mlCodeTable[nbSeq-1]);
    FSE_initCState2(&stateOffsetBits,  CTable_OffsetBits,  ofCodeTable[nbSeq-1]);
    FSE_initCState2(&stateLitLength,   CTable_LitLength,   llCodeTable[nbSeq-1]);
    BIT_addBits(&blockStream, sequences[nbSeq-1].litLength, LL_bits[llCodeTable[nbSeq-1]]);
    if (MEM_32bits()) BIT_flushBits(&blockStream);
    BIT_addBits(&blockStream, sequences[nbSeq-1].mlBase, ML_bits[mlCodeTable[nbSeq-1]]);
    if (MEM_32bits()) BIT_flushBits(&blockStream);
    if (longOffsets) {
        U32 const ofBits = ofCodeTable[nbSeq-1];
        unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
        if (extraBits) {
            BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, extraBits);
            BIT_flushBits(&blockStream);
        }
        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase >> extraBits,
                    ofBits - extraBits);
    } else {
        BIT_addBits(&blockStream, sequences[nbSeq-1].offBase, ofCodeTable[nbSeq-1]);
    }
    BIT_flushBits(&blockStream);

    {   size_t n;
        for (n=nbSeq-2 ; n<nbSeq ; n--) {      /* intentional underflow */
            BYTE const llCode = llCodeTable[n];
            BYTE const ofCode = ofCodeTable[n];
            BYTE const mlCode = mlCodeTable[n];
            U32  const llBits = LL_bits[llCode];
            U32  const ofBits = ofCode;
            U32  const mlBits = ML_bits[mlCode];
            DEBUGLOG(6, "encoding: litlen:%2u - matchlen:%2u - offCode:%7u",
                        (unsigned)sequences[n].litLength,
                        (unsigned)sequences[n].mlBase + MINMATCH,
                        (unsigned)sequences[n].offBase);
                                                                            /* 32b*/  /* 64b*/
                                                                            /* (7)*/  /* (7)*/
            FSE_encodeSymbol(&blockStream, &stateOffsetBits, ofCode);       /* 15 */  /* 15 */
            FSE_encodeSymbol(&blockStream, &stateMatchLength, mlCode);      /* 24 */  /* 24 */
            if (MEM_32bits()) BIT_flushBits(&blockStream);                  /* (7)*/
            FSE_encodeSymbol(&blockStream, &stateLitLength, llCode);        /* 16 */  /* 33 */
            if (MEM_32bits() || (ofBits+mlBits+llBits >= 64-7-(LLFSELog+MLFSELog+OffFSELog)))
                BIT_flushBits(&blockStream);                                /* (7)*/
            BIT_addBits(&blockStream, sequences[n].litLength, llBits);
            if (MEM_32bits() && ((llBits+mlBits)>24)) BIT_flushBits(&blockStream);
            BIT_addBits(&blockStream, sequences[n].mlBase, mlBits);
            if (MEM_32bits() || (ofBits+mlBits+llBits > 56)) BIT_flushBits(&blockStream);
            if (longOffsets) {
                unsigned const extraBits = ofBits - MIN(ofBits, STREAM_ACCUMULATOR_MIN-1);
                if (extraBits) {
                    BIT_addBits(&blockStream, sequences[n].offBase, extraBits);
                    BIT_flushBits(&blockStream);                            /* (7)*/
                }
                BIT_addBits(&blockStream, sequences[n].offBase >> extraBits,
                            ofBits - extraBits);                            /* 31 */
            } else {
                BIT_addBits(&blockStream, sequences[n].offBase, ofBits);     /* 31 */
            }
            BIT_flushBits(&blockStream);                                    /* (7)*/
            DEBUGLOG(7, "remaining space : %i", (int)(blockStream.endPtr - blockStream.ptr));
    }   }

    DEBUGLOG(6, "ZSTD_encodeSequences: flushing ML state with %u bits", stateMatchLength.stateLog);
    FSE_flushCState(&blockStream, &stateMatchLength);
    DEBUGLOG(6, "ZSTD_encodeSequences: flushing Off state with %u bits", stateOffsetBits.stateLog);
    FSE_flushCState(&blockStream, &stateOffsetBits);
    DEBUGLOG(6, "ZSTD_encodeSequences: flushing LL state with %u bits", stateLitLength.stateLog);
    FSE_flushCState(&blockStream, &stateLitLength);

    {   size_t const streamSize = BIT_closeCStream(&blockStream);
        RETURN_ERROR_IF(streamSize==0, dstSize_tooSmall, "not enough space");
        return streamSize;
    }
}

static size_t
ZSTD_encodeSequences_default(
            void* dst, size_t dstCapacity,
            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
            SeqDef const* sequences, size_t nbSeq, int longOffsets)
{
    return ZSTD_encodeSequences_body(dst, dstCapacity,
                                    CTable_MatchLength, mlCodeTable,
                                    CTable_OffsetBits, ofCodeTable,
                                    CTable_LitLength, llCodeTable,
                                    sequences, nbSeq, longOffsets);
}


#if DYNAMIC_BMI2

static BMI2_TARGET_ATTRIBUTE size_t
ZSTD_encodeSequences_bmi2(
            void* dst, size_t dstCapacity,
            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
            SeqDef const* sequences, size_t nbSeq, int longOffsets)
{
    return ZSTD_encodeSequences_body(dst, dstCapacity,
                                    CTable_MatchLength, mlCodeTable,
                                    CTable_OffsetBits, ofCodeTable,
                                    CTable_LitLength, llCodeTable,
                                    sequences, nbSeq, longOffsets);
}

#endif

size_t ZSTD_encodeSequences(
            void* dst, size_t dstCapacity,
            FSE_CTable const* CTable_MatchLength, BYTE const* mlCodeTable,
            FSE_CTable const* CTable_OffsetBits, BYTE const* ofCodeTable,
            FSE_CTable const* CTable_LitLength, BYTE const* llCodeTable,
            SeqDef const* sequences, size_t nbSeq, int longOffsets, int bmi2)
{
    DEBUGLOG(5, "ZSTD_encodeSequences: dstCapacity = %u", (unsigned)dstCapacity);
#if DYNAMIC_BMI2
    if (bmi2) {
        return ZSTD_encodeSequences_bmi2(dst, dstCapacity,
                                         CTable_MatchLength, mlCodeTable,
                                         CTable_OffsetBits, ofCodeTable,
                                         CTable_LitLength, llCodeTable,
                                         sequences, nbSeq, longOffsets);
    }
#endif
    (void)bmi2;
    return ZSTD_encodeSequences_default(dst, dstCapacity,
                                        CTable_MatchLength, mlCodeTable,
                                        CTable_OffsetBits, ofCodeTable,
                                        CTable_LitLength, llCodeTable,
                                        sequences, nbSeq, longOffsets);
}
/**** ended inlining compress/zstd_compress_sequences.c ****/
/**** start inlining compress/zstd_compress_superblock.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

 /*-*************************************
 *  Dependencies
 ***************************************/
/**** start inlining zstd_compress_superblock.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_COMPRESS_ADVANCED_H
#define ZSTD_COMPRESS_ADVANCED_H

/*-*************************************
*  Dependencies
***************************************/

/**** skipping file: ../zstd.h ****/

/*-*************************************
*  Target Compressed Block Size
***************************************/

/* ZSTD_compressSuperBlock() :
 * Used to compress a super block when targetCBlockSize is being used.
 * The given block will be compressed into multiple sub blocks that are around targetCBlockSize. */
size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                               void* dst, size_t dstCapacity,
                               void const* src, size_t srcSize,
                               unsigned lastBlock);

#endif /* ZSTD_COMPRESS_ADVANCED_H */
/**** ended inlining zstd_compress_superblock.h ****/

/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: hist.h ****/
/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_compress_sequences.h ****/
/**** skipping file: zstd_compress_literals.h ****/

/** ZSTD_compressSubBlock_literal() :
 *  Compresses literals section for a sub-block.
 *  When we have to write the Huffman table we will sometimes choose a header
 *  size larger than necessary. This is because we have to pick the header size
 *  before we know the table size + compressed size, so we have a bound on the
 *  table size. If we guessed incorrectly, we fall back to uncompressed literals.
 *
 *  We write the header when writeEntropy=1 and set entropyWritten=1 when we succeeded
 *  in writing the header, otherwise it is set to 0.
 *
 *  hufMetadata->hType has literals block type info.
 *      If it is set_basic, all sub-blocks literals section will be Raw_Literals_Block.
 *      If it is set_rle, all sub-blocks literals section will be RLE_Literals_Block.
 *      If it is set_compressed, first sub-block's literals section will be Compressed_Literals_Block
 *      If it is set_compressed, first sub-block's literals section will be Treeless_Literals_Block
 *      and the following sub-blocks' literals sections will be Treeless_Literals_Block.
 *  @return : compressed size of literals section of a sub-block
 *            Or 0 if unable to compress.
 *            Or error code */
static size_t
ZSTD_compressSubBlock_literal(const HUF_CElt* hufTable,
                              const ZSTD_hufCTablesMetadata_t* hufMetadata,
                              const BYTE* literals, size_t litSize,
                              void* dst, size_t dstSize,
                              const int bmi2, int writeEntropy, int* entropyWritten)
{
    size_t const header = writeEntropy ? 200 : 0;
    size_t const lhSize = 3 + (litSize >= (1 KB - header)) + (litSize >= (16 KB - header));
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstSize;
    BYTE* op = ostart + lhSize;
    U32 const singleStream = lhSize == 3;
    SymbolEncodingType_e hType = writeEntropy ? hufMetadata->hType : set_repeat;
    size_t cLitSize = 0;

    DEBUGLOG(5, "ZSTD_compressSubBlock_literal (litSize=%zu, lhSize=%zu, writeEntropy=%d)", litSize, lhSize, writeEntropy);

    *entropyWritten = 0;
    if (litSize == 0 || hufMetadata->hType == set_basic) {
      DEBUGLOG(5, "ZSTD_compressSubBlock_literal using raw literal");
      return ZSTD_noCompressLiterals(dst, dstSize, literals, litSize);
    } else if (hufMetadata->hType == set_rle) {
      DEBUGLOG(5, "ZSTD_compressSubBlock_literal using rle literal");
      return ZSTD_compressRleLiteralsBlock(dst, dstSize, literals, litSize);
    }

    assert(litSize > 0);
    assert(hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat);

    if (writeEntropy && hufMetadata->hType == set_compressed) {
        ZSTD_memcpy(op, hufMetadata->hufDesBuffer, hufMetadata->hufDesSize);
        op += hufMetadata->hufDesSize;
        cLitSize += hufMetadata->hufDesSize;
        DEBUGLOG(5, "ZSTD_compressSubBlock_literal (hSize=%zu)", hufMetadata->hufDesSize);
    }

    {   int const flags = bmi2 ? HUF_flags_bmi2 : 0;
        const size_t cSize = singleStream ? HUF_compress1X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags)
                                          : HUF_compress4X_usingCTable(op, (size_t)(oend-op), literals, litSize, hufTable, flags);
        op += cSize;
        cLitSize += cSize;
        if (cSize == 0 || ERR_isError(cSize)) {
            DEBUGLOG(5, "Failed to write entropy tables %s", ZSTD_getErrorName(cSize));
            return 0;
        }
        /* If we expand and we aren't writing a header then emit uncompressed */
        if (!writeEntropy && cLitSize >= litSize) {
            DEBUGLOG(5, "ZSTD_compressSubBlock_literal using raw literal because uncompressible");
            return ZSTD_noCompressLiterals(dst, dstSize, literals, litSize);
        }
        /* If we are writing headers then allow expansion that doesn't change our header size. */
        if (lhSize < (size_t)(3 + (cLitSize >= 1 KB) + (cLitSize >= 16 KB))) {
            assert(cLitSize > litSize);
            DEBUGLOG(5, "Literals expanded beyond allowed header size");
            return ZSTD_noCompressLiterals(dst, dstSize, literals, litSize);
        }
        DEBUGLOG(5, "ZSTD_compressSubBlock_literal (cSize=%zu)", cSize);
    }

    /* Build header */
    switch(lhSize)
    {
    case 3: /* 2 - 2 - 10 - 10 */
        {   U32 const lhc = hType + ((U32)(!singleStream) << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<14);
            MEM_writeLE24(ostart, lhc);
            break;
        }
    case 4: /* 2 - 2 - 14 - 14 */
        {   U32 const lhc = hType + (2 << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<18);
            MEM_writeLE32(ostart, lhc);
            break;
        }
    case 5: /* 2 - 2 - 18 - 18 */
        {   U32 const lhc = hType + (3 << 2) + ((U32)litSize<<4) + ((U32)cLitSize<<22);
            MEM_writeLE32(ostart, lhc);
            ostart[4] = (BYTE)(cLitSize >> 10);
            break;
        }
    default:  /* not possible : lhSize is {3,4,5} */
        assert(0);
    }
    *entropyWritten = 1;
    DEBUGLOG(5, "Compressed literals: %u -> %u", (U32)litSize, (U32)(op-ostart));
    return (size_t)(op-ostart);
}

static size_t
ZSTD_seqDecompressedSize(SeqStore_t const* seqStore,
                   const SeqDef* sequences, size_t nbSeqs,
                         size_t litSize, int lastSubBlock)
{
    size_t matchLengthSum = 0;
    size_t litLengthSum = 0;
    size_t n;
    for (n=0; n<nbSeqs; n++) {
        const ZSTD_SequenceLength seqLen = ZSTD_getSequenceLength(seqStore, sequences+n);
        litLengthSum += seqLen.litLength;
        matchLengthSum += seqLen.matchLength;
    }
    DEBUGLOG(5, "ZSTD_seqDecompressedSize: %u sequences from %p: %u literals + %u matchlength",
                (unsigned)nbSeqs, (const void*)sequences,
                (unsigned)litLengthSum, (unsigned)matchLengthSum);
    if (!lastSubBlock)
        assert(litLengthSum == litSize);
    else
        assert(litLengthSum <= litSize);
    (void)litLengthSum;
    return matchLengthSum + litSize;
}

/** ZSTD_compressSubBlock_sequences() :
 *  Compresses sequences section for a sub-block.
 *  fseMetadata->llType, fseMetadata->ofType, and fseMetadata->mlType have
 *  symbol compression modes for the super-block.
 *  The first successfully compressed block will have these in its header.
 *  We set entropyWritten=1 when we succeed in compressing the sequences.
 *  The following sub-blocks will always have repeat mode.
 *  @return : compressed size of sequences section of a sub-block
 *            Or 0 if it is unable to compress
 *            Or error code. */
static size_t
ZSTD_compressSubBlock_sequences(const ZSTD_fseCTables_t* fseTables,
                                const ZSTD_fseCTablesMetadata_t* fseMetadata,
                                const SeqDef* sequences, size_t nbSeq,
                                const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
                                const ZSTD_CCtx_params* cctxParams,
                                void* dst, size_t dstCapacity,
                                const int bmi2, int writeEntropy, int* entropyWritten)
{
    const int longOffsets = cctxParams->cParams.windowLog > STREAM_ACCUMULATOR_MIN;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstCapacity;
    BYTE* op = ostart;
    BYTE* seqHead;

    DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (nbSeq=%zu, writeEntropy=%d, longOffsets=%d)", nbSeq, writeEntropy, longOffsets);

    *entropyWritten = 0;
    /* Sequences Header */
    RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/,
                    dstSize_tooSmall, "");
    if (nbSeq < 128)
        *op++ = (BYTE)nbSeq;
    else if (nbSeq < LONGNBSEQ)
        op[0] = (BYTE)((nbSeq>>8) + 0x80), op[1] = (BYTE)nbSeq, op+=2;
    else
        op[0]=0xFF, MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ)), op+=3;
    if (nbSeq==0) {
        return (size_t)(op - ostart);
    }

    /* seqHead : flags for FSE encoding type */
    seqHead = op++;

    DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (seqHeadSize=%u)", (unsigned)(op-ostart));

    if (writeEntropy) {
        const U32 LLtype = fseMetadata->llType;
        const U32 Offtype = fseMetadata->ofType;
        const U32 MLtype = fseMetadata->mlType;
        DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (fseTablesSize=%zu)", fseMetadata->fseTablesSize);
        *seqHead = (BYTE)((LLtype<<6) + (Offtype<<4) + (MLtype<<2));
        ZSTD_memcpy(op, fseMetadata->fseTablesBuffer, fseMetadata->fseTablesSize);
        op += fseMetadata->fseTablesSize;
    } else {
        const U32 repeat = set_repeat;
        *seqHead = (BYTE)((repeat<<6) + (repeat<<4) + (repeat<<2));
    }

    {   size_t const bitstreamSize = ZSTD_encodeSequences(
                                        op, (size_t)(oend - op),
                                        fseTables->matchlengthCTable, mlCode,
                                        fseTables->offcodeCTable, ofCode,
                                        fseTables->litlengthCTable, llCode,
                                        sequences, nbSeq,
                                        longOffsets, bmi2);
        FORWARD_IF_ERROR(bitstreamSize, "ZSTD_encodeSequences failed");
        op += bitstreamSize;
        /* zstd versions <= 1.3.4 mistakenly report corruption when
         * FSE_readNCount() receives a buffer < 4 bytes.
         * Fixed by https://github.com/facebook/zstd/pull/1146.
         * This can happen when the last set_compressed table present is 2
         * bytes and the bitstream is only one byte.
         * In this exceedingly rare case, we will simply emit an uncompressed
         * block, since it isn't worth optimizing.
         */
#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
        if (writeEntropy && fseMetadata->lastCountSize && fseMetadata->lastCountSize + bitstreamSize < 4) {
            /* NCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
            assert(fseMetadata->lastCountSize + bitstreamSize == 3);
            DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.3.4 by "
                        "emitting an uncompressed block.");
            return 0;
        }
#endif
        DEBUGLOG(5, "ZSTD_compressSubBlock_sequences (bitstreamSize=%zu)", bitstreamSize);
    }

    /* zstd versions <= 1.4.0 mistakenly report error when
     * sequences section body size is less than 3 bytes.
     * Fixed by https://github.com/facebook/zstd/pull/1664.
     * This can happen when the previous sequences section block is compressed
     * with rle mode and the current block's sequences section is compressed
     * with repeat mode where sequences section body size can be 1 byte.
     */
#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    if (op-seqHead < 4) {
        DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.4.0 by emitting "
                    "an uncompressed block when sequences are < 4 bytes");
        return 0;
    }
#endif

    *entropyWritten = 1;
    return (size_t)(op - ostart);
}

/** ZSTD_compressSubBlock() :
 *  Compresses a single sub-block.
 *  @return : compressed size of the sub-block
 *            Or 0 if it failed to compress. */
static size_t ZSTD_compressSubBlock(const ZSTD_entropyCTables_t* entropy,
                                    const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                                    const SeqDef* sequences, size_t nbSeq,
                                    const BYTE* literals, size_t litSize,
                                    const BYTE* llCode, const BYTE* mlCode, const BYTE* ofCode,
                                    const ZSTD_CCtx_params* cctxParams,
                                    void* dst, size_t dstCapacity,
                                    const int bmi2,
                                    int writeLitEntropy, int writeSeqEntropy,
                                    int* litEntropyWritten, int* seqEntropyWritten,
                                    U32 lastBlock)
{
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstCapacity;
    BYTE* op = ostart + ZSTD_blockHeaderSize;
    DEBUGLOG(5, "ZSTD_compressSubBlock (litSize=%zu, nbSeq=%zu, writeLitEntropy=%d, writeSeqEntropy=%d, lastBlock=%d)",
                litSize, nbSeq, writeLitEntropy, writeSeqEntropy, lastBlock);
    {   size_t cLitSize = ZSTD_compressSubBlock_literal((const HUF_CElt*)entropy->huf.CTable,
                                                        &entropyMetadata->hufMetadata, literals, litSize,
                                                        op, (size_t)(oend-op),
                                                        bmi2, writeLitEntropy, litEntropyWritten);
        FORWARD_IF_ERROR(cLitSize, "ZSTD_compressSubBlock_literal failed");
        if (cLitSize == 0) return 0;
        op += cLitSize;
    }
    {   size_t cSeqSize = ZSTD_compressSubBlock_sequences(&entropy->fse,
                                                  &entropyMetadata->fseMetadata,
                                                  sequences, nbSeq,
                                                  llCode, mlCode, ofCode,
                                                  cctxParams,
                                                  op, (size_t)(oend-op),
                                                  bmi2, writeSeqEntropy, seqEntropyWritten);
        FORWARD_IF_ERROR(cSeqSize, "ZSTD_compressSubBlock_sequences failed");
        if (cSeqSize == 0) return 0;
        op += cSeqSize;
    }
    /* Write block header */
    {   size_t cSize = (size_t)(op-ostart) - ZSTD_blockHeaderSize;
        U32 const cBlockHeader24 = lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
        MEM_writeLE24(ostart, cBlockHeader24);
    }
    return (size_t)(op-ostart);
}

static size_t ZSTD_estimateSubBlockSize_literal(const BYTE* literals, size_t litSize,
                                                const ZSTD_hufCTables_t* huf,
                                                const ZSTD_hufCTablesMetadata_t* hufMetadata,
                                                void* workspace, size_t wkspSize,
                                                int writeEntropy)
{
    unsigned* const countWksp = (unsigned*)workspace;
    unsigned maxSymbolValue = 255;
    size_t literalSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */

    if (hufMetadata->hType == set_basic) return litSize;
    else if (hufMetadata->hType == set_rle) return 1;
    else if (hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat) {
        size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)literals, litSize, workspace, wkspSize);
        if (ZSTD_isError(largest)) return litSize;
        {   size_t cLitSizeEstimate = HUF_estimateCompressedSize((const HUF_CElt*)huf->CTable, countWksp, maxSymbolValue);
            if (writeEntropy) cLitSizeEstimate += hufMetadata->hufDesSize;
            return cLitSizeEstimate + literalSectionHeaderSize;
    }   }
    assert(0); /* impossible */
    return 0;
}

static size_t ZSTD_estimateSubBlockSize_symbolType(SymbolEncodingType_e type,
                        const BYTE* codeTable, unsigned maxCode,
                        size_t nbSeq, const FSE_CTable* fseCTable,
                        const U8* additionalBits,
                        short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                        void* workspace, size_t wkspSize)
{
    unsigned* const countWksp = (unsigned*)workspace;
    const BYTE* ctp = codeTable;
    const BYTE* const ctStart = ctp;
    const BYTE* const ctEnd = ctStart + nbSeq;
    size_t cSymbolTypeSizeEstimateInBits = 0;
    unsigned max = maxCode;

    HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
    if (type == set_basic) {
        /* We selected this encoding type, so it must be valid. */
        assert(max <= defaultMax);
        cSymbolTypeSizeEstimateInBits = max <= defaultMax
                ? ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max)
                : ERROR(GENERIC);
    } else if (type == set_rle) {
        cSymbolTypeSizeEstimateInBits = 0;
    } else if (type == set_compressed || type == set_repeat) {
        cSymbolTypeSizeEstimateInBits = ZSTD_fseBitCost(fseCTable, countWksp, max);
    }
    if (ZSTD_isError(cSymbolTypeSizeEstimateInBits)) return nbSeq * 10;
    while (ctp < ctEnd) {
        if (additionalBits) cSymbolTypeSizeEstimateInBits += additionalBits[*ctp];
        else cSymbolTypeSizeEstimateInBits += *ctp; /* for offset, offset code is also the number of additional bits */
        ctp++;
    }
    return cSymbolTypeSizeEstimateInBits / 8;
}

static size_t ZSTD_estimateSubBlockSize_sequences(const BYTE* ofCodeTable,
                                                  const BYTE* llCodeTable,
                                                  const BYTE* mlCodeTable,
                                                  size_t nbSeq,
                                                  const ZSTD_fseCTables_t* fseTables,
                                                  const ZSTD_fseCTablesMetadata_t* fseMetadata,
                                                  void* workspace, size_t wkspSize,
                                                  int writeEntropy)
{
    size_t const sequencesSectionHeaderSize = 3; /* Use hard coded size of 3 bytes */
    size_t cSeqSizeEstimate = 0;
    if (nbSeq == 0) return sequencesSectionHeaderSize;
    cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, MaxOff,
                                         nbSeq, fseTables->offcodeCTable, NULL,
                                         OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                                         workspace, wkspSize);
    cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->llType, llCodeTable, MaxLL,
                                         nbSeq, fseTables->litlengthCTable, LL_bits,
                                         LL_defaultNorm, LL_defaultNormLog, MaxLL,
                                         workspace, wkspSize);
    cSeqSizeEstimate += ZSTD_estimateSubBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, MaxML,
                                         nbSeq, fseTables->matchlengthCTable, ML_bits,
                                         ML_defaultNorm, ML_defaultNormLog, MaxML,
                                         workspace, wkspSize);
    if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
    return cSeqSizeEstimate + sequencesSectionHeaderSize;
}

typedef struct {
    size_t estLitSize;
    size_t estBlockSize;
} EstimatedBlockSize;
static EstimatedBlockSize ZSTD_estimateSubBlockSize(const BYTE* literals, size_t litSize,
                                        const BYTE* ofCodeTable,
                                        const BYTE* llCodeTable,
                                        const BYTE* mlCodeTable,
                                        size_t nbSeq,
                                        const ZSTD_entropyCTables_t* entropy,
                                        const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                                        void* workspace, size_t wkspSize,
                                        int writeLitEntropy, int writeSeqEntropy)
{
    EstimatedBlockSize ebs;
    ebs.estLitSize = ZSTD_estimateSubBlockSize_literal(literals, litSize,
                                                        &entropy->huf, &entropyMetadata->hufMetadata,
                                                        workspace, wkspSize, writeLitEntropy);
    ebs.estBlockSize = ZSTD_estimateSubBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
                                                         nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
                                                         workspace, wkspSize, writeSeqEntropy);
    ebs.estBlockSize += ebs.estLitSize + ZSTD_blockHeaderSize;
    return ebs;
}

static int ZSTD_needSequenceEntropyTables(ZSTD_fseCTablesMetadata_t const* fseMetadata)
{
    if (fseMetadata->llType == set_compressed || fseMetadata->llType == set_rle)
        return 1;
    if (fseMetadata->mlType == set_compressed || fseMetadata->mlType == set_rle)
        return 1;
    if (fseMetadata->ofType == set_compressed || fseMetadata->ofType == set_rle)
        return 1;
    return 0;
}

static size_t countLiterals(SeqStore_t const* seqStore, const SeqDef* sp, size_t seqCount)
{
    size_t n, total = 0;
    assert(sp != NULL);
    for (n=0; n<seqCount; n++) {
        total += ZSTD_getSequenceLength(seqStore, sp+n).litLength;
    }
    DEBUGLOG(6, "countLiterals for %zu sequences from %p => %zu bytes", seqCount, (const void*)sp, total);
    return total;
}

#define BYTESCALE 256

static size_t sizeBlockSequences(const SeqDef* sp, size_t nbSeqs,
                size_t targetBudget, size_t avgLitCost, size_t avgSeqCost,
                int firstSubBlock)
{
    size_t n, budget = 0, inSize=0;
    /* entropy headers */
    size_t const headerSize = (size_t)firstSubBlock * 120 * BYTESCALE; /* generous estimate */
    assert(firstSubBlock==0 || firstSubBlock==1);
    budget += headerSize;

    /* first sequence => at least one sequence*/
    budget += sp[0].litLength * avgLitCost + avgSeqCost;
    if (budget > targetBudget) return 1;
    inSize = sp[0].litLength + (sp[0].mlBase+MINMATCH);

    /* loop over sequences */
    for (n=1; n<nbSeqs; n++) {
        size_t currentCost = sp[n].litLength * avgLitCost + avgSeqCost;
        budget += currentCost;
        inSize += sp[n].litLength + (sp[n].mlBase+MINMATCH);
        /* stop when sub-block budget is reached */
        if ( (budget > targetBudget)
            /* though continue to expand until the sub-block is deemed compressible */
          && (budget < inSize * BYTESCALE) )
            break;
    }

    return n;
}

/** ZSTD_compressSubBlock_multi() :
 *  Breaks super-block into multiple sub-blocks and compresses them.
 *  Entropy will be written into the first block.
 *  The following blocks use repeat_mode to compress.
 *  Sub-blocks are all compressed, except the last one when beneficial.
 *  @return : compressed size of the super block (which features multiple ZSTD blocks)
 *            or 0 if it failed to compress. */
static size_t ZSTD_compressSubBlock_multi(const SeqStore_t* seqStorePtr,
                            const ZSTD_compressedBlockState_t* prevCBlock,
                            ZSTD_compressedBlockState_t* nextCBlock,
                            const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                            const ZSTD_CCtx_params* cctxParams,
                                  void* dst, size_t dstCapacity,
                            const void* src, size_t srcSize,
                            const int bmi2, U32 lastBlock,
                            void* workspace, size_t wkspSize)
{
    const SeqDef* const sstart = seqStorePtr->sequencesStart;
    const SeqDef* const send = seqStorePtr->sequences;
    const SeqDef* sp = sstart; /* tracks progresses within seqStorePtr->sequences */
    size_t const nbSeqs = (size_t)(send - sstart);
    const BYTE* const lstart = seqStorePtr->litStart;
    const BYTE* const lend = seqStorePtr->lit;
    const BYTE* lp = lstart;
    size_t const nbLiterals = (size_t)(lend - lstart);
    BYTE const* ip = (BYTE const*)src;
    BYTE const* const iend = ip + srcSize;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstCapacity;
    BYTE* op = ostart;
    const BYTE* llCodePtr = seqStorePtr->llCode;
    const BYTE* mlCodePtr = seqStorePtr->mlCode;
    const BYTE* ofCodePtr = seqStorePtr->ofCode;
    size_t const minTarget = ZSTD_TARGETCBLOCKSIZE_MIN; /* enforce minimum size, to reduce undesirable side effects */
    size_t const targetCBlockSize = MAX(minTarget, cctxParams->targetCBlockSize);
    int writeLitEntropy = (entropyMetadata->hufMetadata.hType == set_compressed);
    int writeSeqEntropy = 1;

    DEBUGLOG(5, "ZSTD_compressSubBlock_multi (srcSize=%u, litSize=%u, nbSeq=%u)",
               (unsigned)srcSize, (unsigned)(lend-lstart), (unsigned)(send-sstart));

        /* let's start by a general estimation for the full block */
    if (nbSeqs > 0) {
        EstimatedBlockSize const ebs =
                ZSTD_estimateSubBlockSize(lp, nbLiterals,
                                        ofCodePtr, llCodePtr, mlCodePtr, nbSeqs,
                                        &nextCBlock->entropy, entropyMetadata,
                                        workspace, wkspSize,
                                        writeLitEntropy, writeSeqEntropy);
        /* quick estimation */
        size_t const avgLitCost = nbLiterals ? (ebs.estLitSize * BYTESCALE) / nbLiterals : BYTESCALE;
        size_t const avgSeqCost = ((ebs.estBlockSize - ebs.estLitSize) * BYTESCALE) / nbSeqs;
        const size_t nbSubBlocks = MAX((ebs.estBlockSize + (targetCBlockSize/2)) / targetCBlockSize, 1);
        size_t n, avgBlockBudget, blockBudgetSupp=0;
        avgBlockBudget = (ebs.estBlockSize * BYTESCALE) / nbSubBlocks;
        DEBUGLOG(5, "estimated fullblock size=%u bytes ; avgLitCost=%.2f ; avgSeqCost=%.2f ; targetCBlockSize=%u, nbSubBlocks=%u ; avgBlockBudget=%.0f bytes",
                    (unsigned)ebs.estBlockSize, (double)avgLitCost/BYTESCALE, (double)avgSeqCost/BYTESCALE,
                    (unsigned)targetCBlockSize, (unsigned)nbSubBlocks, (double)avgBlockBudget/BYTESCALE);
        /* simplification: if estimates states that the full superblock doesn't compress, just bail out immediately
         * this will result in the production of a single uncompressed block covering @srcSize.*/
        if (ebs.estBlockSize > srcSize) return 0;

        /* compress and write sub-blocks */
        assert(nbSubBlocks>0);
        for (n=0; n < nbSubBlocks-1; n++) {
            /* determine nb of sequences for current sub-block + nbLiterals from next sequence */
            size_t const seqCount = sizeBlockSequences(sp, (size_t)(send-sp),
                                        avgBlockBudget + blockBudgetSupp, avgLitCost, avgSeqCost, n==0);
            /* if reached last sequence : break to last sub-block (simplification) */
            assert(seqCount <= (size_t)(send-sp));
            if (sp + seqCount == send) break;
            assert(seqCount > 0);
            /* compress sub-block */
            {   int litEntropyWritten = 0;
                int seqEntropyWritten = 0;
                size_t litSize = countLiterals(seqStorePtr, sp, seqCount);
                const size_t decompressedSize =
                        ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 0);
                size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata,
                                                sp, seqCount,
                                                lp, litSize,
                                                llCodePtr, mlCodePtr, ofCodePtr,
                                                cctxParams,
                                                op, (size_t)(oend-op),
                                                bmi2, writeLitEntropy, writeSeqEntropy,
                                                &litEntropyWritten, &seqEntropyWritten,
                                                0);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed");

                /* check compressibility, update state components */
                if (cSize > 0 && cSize < decompressedSize) {
                    DEBUGLOG(5, "Committed sub-block compressing %u bytes => %u bytes",
                                (unsigned)decompressedSize, (unsigned)cSize);
                    assert(ip + decompressedSize <= iend);
                    ip += decompressedSize;
                    lp += litSize;
                    op += cSize;
                    llCodePtr += seqCount;
                    mlCodePtr += seqCount;
                    ofCodePtr += seqCount;
                    /* Entropy only needs to be written once */
                    if (litEntropyWritten) {
                        writeLitEntropy = 0;
                    }
                    if (seqEntropyWritten) {
                        writeSeqEntropy = 0;
                    }
                    sp += seqCount;
                    blockBudgetSupp = 0;
            }   }
            /* otherwise : do not compress yet, coalesce current sub-block with following one */
        }
    } /* if (nbSeqs > 0) */

    /* write last block */
    DEBUGLOG(5, "Generate last sub-block: %u sequences remaining", (unsigned)(send - sp));
    {   int litEntropyWritten = 0;
        int seqEntropyWritten = 0;
        size_t litSize = (size_t)(lend - lp);
        size_t seqCount = (size_t)(send - sp);
        const size_t decompressedSize =
                ZSTD_seqDecompressedSize(seqStorePtr, sp, seqCount, litSize, 1);
        size_t const cSize = ZSTD_compressSubBlock(&nextCBlock->entropy, entropyMetadata,
                                            sp, seqCount,
                                            lp, litSize,
                                            llCodePtr, mlCodePtr, ofCodePtr,
                                            cctxParams,
                                            op, (size_t)(oend-op),
                                            bmi2, writeLitEntropy, writeSeqEntropy,
                                            &litEntropyWritten, &seqEntropyWritten,
                                            lastBlock);
        FORWARD_IF_ERROR(cSize, "ZSTD_compressSubBlock failed");

        /* update pointers, the nb of literals borrowed from next sequence must be preserved */
        if (cSize > 0 && cSize < decompressedSize) {
            DEBUGLOG(5, "Last sub-block compressed %u bytes => %u bytes",
                        (unsigned)decompressedSize, (unsigned)cSize);
            assert(ip + decompressedSize <= iend);
            ip += decompressedSize;
            lp += litSize;
            op += cSize;
            llCodePtr += seqCount;
            mlCodePtr += seqCount;
            ofCodePtr += seqCount;
            /* Entropy only needs to be written once */
            if (litEntropyWritten) {
                writeLitEntropy = 0;
            }
            if (seqEntropyWritten) {
                writeSeqEntropy = 0;
            }
            sp += seqCount;
        }
    }


    if (writeLitEntropy) {
        DEBUGLOG(5, "Literal entropy tables were never written");
        ZSTD_memcpy(&nextCBlock->entropy.huf, &prevCBlock->entropy.huf, sizeof(prevCBlock->entropy.huf));
    }
    if (writeSeqEntropy && ZSTD_needSequenceEntropyTables(&entropyMetadata->fseMetadata)) {
        /* If we haven't written our entropy tables, then we've violated our contract and
         * must emit an uncompressed block.
         */
        DEBUGLOG(5, "Sequence entropy tables were never written => cancel, emit an uncompressed block");
        return 0;
    }

    if (ip < iend) {
        /* some data left : last part of the block sent uncompressed */
        size_t const rSize = (size_t)((iend - ip));
        size_t const cSize = ZSTD_noCompressBlock(op, (size_t)(oend - op), ip, rSize, lastBlock);
        DEBUGLOG(5, "Generate last uncompressed sub-block of %u bytes", (unsigned)(rSize));
        FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
        assert(cSize != 0);
        op += cSize;
        /* We have to regenerate the repcodes because we've skipped some sequences */
        if (sp < send) {
            const SeqDef* seq;
            Repcodes_t rep;
            ZSTD_memcpy(&rep, prevCBlock->rep, sizeof(rep));
            for (seq = sstart; seq < sp; ++seq) {
                ZSTD_updateRep(rep.rep, seq->offBase, ZSTD_getSequenceLength(seqStorePtr, seq).litLength == 0);
            }
            ZSTD_memcpy(nextCBlock->rep, &rep, sizeof(rep));
        }
    }

    DEBUGLOG(5, "ZSTD_compressSubBlock_multi compressed all subBlocks: total compressed size = %u",
                (unsigned)(op-ostart));
    return (size_t)(op-ostart);
}

size_t ZSTD_compressSuperBlock(ZSTD_CCtx* zc,
                               void* dst, size_t dstCapacity,
                               const void* src, size_t srcSize,
                               unsigned lastBlock)
{
    ZSTD_entropyCTablesMetadata_t entropyMetadata;

    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(&zc->seqStore,
          &zc->blockState.prevCBlock->entropy,
          &zc->blockState.nextCBlock->entropy,
          &zc->appliedParams,
          &entropyMetadata,
          zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */), "");

    return ZSTD_compressSubBlock_multi(&zc->seqStore,
            zc->blockState.prevCBlock,
            zc->blockState.nextCBlock,
            &entropyMetadata,
            &zc->appliedParams,
            dst, dstCapacity,
            src, srcSize,
            zc->bmi2, lastBlock,
            zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */);
}
/**** ended inlining compress/zstd_compress_superblock.c ****/
/**** start inlining compress/zstd_preSplit.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: hist.h ****/
/**** skipping file: zstd_preSplit.h ****/


#define BLOCKSIZE_MIN 3500
#define THRESHOLD_PENALTY_RATE 16
#define THRESHOLD_BASE (THRESHOLD_PENALTY_RATE - 2)
#define THRESHOLD_PENALTY 3

#define HASHLENGTH 2
#define HASHLOG_MAX 10
#define HASHTABLESIZE (1 << HASHLOG_MAX)
#define HASHMASK (HASHTABLESIZE - 1)
#define KNUTH 0x9e3779b9

/* for hashLog > 8, hash 2 bytes.
 * for hashLog == 8, just take the byte, no hashing.
 * The speed of this method relies on compile-time constant propagation */
FORCE_INLINE_TEMPLATE unsigned hash2(const void *p, unsigned hashLog)
{
    assert(hashLog >= 8);
    if (hashLog == 8) return (U32)((const BYTE*)p)[0];
    assert(hashLog <= HASHLOG_MAX);
    return (U32)(MEM_read16(p)) * KNUTH >> (32 - hashLog);
}


typedef struct {
  unsigned events[HASHTABLESIZE];
  size_t nbEvents;
} Fingerprint;
typedef struct {
    Fingerprint pastEvents;
    Fingerprint newEvents;
} FPStats;

static void initStats(FPStats* fpstats)
{
    ZSTD_memset(fpstats, 0, sizeof(FPStats));
}

FORCE_INLINE_TEMPLATE void
addEvents_generic(Fingerprint* fp, const void* src, size_t srcSize, size_t samplingRate, unsigned hashLog)
{
    const char* p = (const char*)src;
    size_t limit = srcSize - HASHLENGTH + 1;
    size_t n;
    assert(srcSize >= HASHLENGTH);
    for (n = 0; n < limit; n+=samplingRate) {
        fp->events[hash2(p+n, hashLog)]++;
    }
    fp->nbEvents += limit/samplingRate;
}

FORCE_INLINE_TEMPLATE void
recordFingerprint_generic(Fingerprint* fp, const void* src, size_t srcSize, size_t samplingRate, unsigned hashLog)
{
    ZSTD_memset(fp, 0, sizeof(unsigned) * ((size_t)1 << hashLog));
    fp->nbEvents = 0;
    addEvents_generic(fp, src, srcSize, samplingRate, hashLog);
}

typedef void (*RecordEvents_f)(Fingerprint* fp, const void* src, size_t srcSize);

#define FP_RECORD(_rate) ZSTD_recordFingerprint_##_rate

#define ZSTD_GEN_RECORD_FINGERPRINT(_rate, _hSize)                                 \
    static void FP_RECORD(_rate)(Fingerprint* fp, const void* src, size_t srcSize) \
    {                                                                              \
        recordFingerprint_generic(fp, src, srcSize, _rate, _hSize);                \
    }

ZSTD_GEN_RECORD_FINGERPRINT(1, 10)
ZSTD_GEN_RECORD_FINGERPRINT(5, 10)
ZSTD_GEN_RECORD_FINGERPRINT(11, 9)
ZSTD_GEN_RECORD_FINGERPRINT(43, 8)


static U64 abs64(S64 s64) { return (U64)((s64 < 0) ? -s64 : s64); }

static U64 fpDistance(const Fingerprint* fp1, const Fingerprint* fp2, unsigned hashLog)
{
    U64 distance = 0;
    size_t n;
    assert(hashLog <= HASHLOG_MAX);
    for (n = 0; n < ((size_t)1 << hashLog); n++) {
        distance +=
            abs64((S64)fp1->events[n] * (S64)fp2->nbEvents - (S64)fp2->events[n] * (S64)fp1->nbEvents);
    }
    return distance;
}

/* Compare newEvents with pastEvents
 * return 1 when considered "too different"
 */
static int compareFingerprints(const Fingerprint* ref,
                            const Fingerprint* newfp,
                            int penalty,
                            unsigned hashLog)
{
    assert(ref->nbEvents > 0);
    assert(newfp->nbEvents > 0);
    {   U64 p50 = (U64)ref->nbEvents * (U64)newfp->nbEvents;
        U64 deviation = fpDistance(ref, newfp, hashLog);
        U64 threshold = p50 * (U64)(THRESHOLD_BASE + penalty) / THRESHOLD_PENALTY_RATE;
        return deviation >= threshold;
    }
}

static void mergeEvents(Fingerprint* acc, const Fingerprint* newfp)
{
    size_t n;
    for (n = 0; n < HASHTABLESIZE; n++) {
        acc->events[n] += newfp->events[n];
    }
    acc->nbEvents += newfp->nbEvents;
}

static void flushEvents(FPStats* fpstats)
{
    size_t n;
    for (n = 0; n < HASHTABLESIZE; n++) {
        fpstats->pastEvents.events[n] = fpstats->newEvents.events[n];
    }
    fpstats->pastEvents.nbEvents = fpstats->newEvents.nbEvents;
    ZSTD_memset(&fpstats->newEvents, 0, sizeof(fpstats->newEvents));
}

static void removeEvents(Fingerprint* acc, const Fingerprint* slice)
{
    size_t n;
    for (n = 0; n < HASHTABLESIZE; n++) {
        assert(acc->events[n] >= slice->events[n]);
        acc->events[n] -= slice->events[n];
    }
    acc->nbEvents -= slice->nbEvents;
}

#define CHUNKSIZE (8 << 10)
static size_t ZSTD_splitBlock_byChunks(const void* blockStart, size_t blockSize,
                        int level,
                        void* workspace, size_t wkspSize)
{
    static const RecordEvents_f records_fs[] = {
        FP_RECORD(43), FP_RECORD(11), FP_RECORD(5), FP_RECORD(1)
    };
    static const unsigned hashParams[] = { 8, 9, 10, 10 };
    const RecordEvents_f record_f = (assert(0<=level && level<=3), records_fs[level]);
    FPStats* const fpstats = (FPStats*)workspace;
    const char* p = (const char*)blockStart;
    int penalty = THRESHOLD_PENALTY;
    size_t pos = 0;
    assert(blockSize == (128 << 10));
    assert(workspace != NULL);
    assert((size_t)workspace % ZSTD_ALIGNOF(FPStats) == 0);
    ZSTD_STATIC_ASSERT(ZSTD_SLIPBLOCK_WORKSPACESIZE >= sizeof(FPStats));
    assert(wkspSize >= sizeof(FPStats)); (void)wkspSize;

    initStats(fpstats);
    record_f(&fpstats->pastEvents, p, CHUNKSIZE);
    for (pos = CHUNKSIZE; pos <= blockSize - CHUNKSIZE; pos += CHUNKSIZE) {
        record_f(&fpstats->newEvents, p + pos, CHUNKSIZE);
        if (compareFingerprints(&fpstats->pastEvents, &fpstats->newEvents, penalty, hashParams[level])) {
            return pos;
        } else {
            mergeEvents(&fpstats->pastEvents, &fpstats->newEvents);
            if (penalty > 0) penalty--;
        }
    }
    assert(pos == blockSize);
    return blockSize;
    (void)flushEvents; (void)removeEvents;
}

/* ZSTD_splitBlock_fromBorders(): very fast strategy :
 * compare fingerprint from beginning and end of the block,
 * derive from their difference if it's preferable to split in the middle,
 * repeat the process a second time, for finer grained decision.
 * 3 times did not brought improvements, so I stopped at 2.
 * Benefits are good enough for a cheap heuristic.
 * More accurate splitting saves more, but speed impact is also more perceptible.
 * For better accuracy, use more elaborate variant *_byChunks.
 */
static size_t ZSTD_splitBlock_fromBorders(const void* blockStart, size_t blockSize,
                        void* workspace, size_t wkspSize)
{
#define SEGMENT_SIZE 512
    FPStats* const fpstats = (FPStats*)workspace;
    Fingerprint* middleEvents = (Fingerprint*)(void*)((char*)workspace + 512 * sizeof(unsigned));
    assert(blockSize == (128 << 10));
    assert(workspace != NULL);
    assert((size_t)workspace % ZSTD_ALIGNOF(FPStats) == 0);
    ZSTD_STATIC_ASSERT(ZSTD_SLIPBLOCK_WORKSPACESIZE >= sizeof(FPStats));
    assert(wkspSize >= sizeof(FPStats)); (void)wkspSize;

    initStats(fpstats);
    HIST_add(fpstats->pastEvents.events, blockStart, SEGMENT_SIZE);
    HIST_add(fpstats->newEvents.events, (const char*)blockStart + blockSize - SEGMENT_SIZE, SEGMENT_SIZE);
    fpstats->pastEvents.nbEvents = fpstats->newEvents.nbEvents = SEGMENT_SIZE;
    if (!compareFingerprints(&fpstats->pastEvents, &fpstats->newEvents, 0, 8))
        return blockSize;

    HIST_add(middleEvents->events, (const char*)blockStart + blockSize/2 - SEGMENT_SIZE/2, SEGMENT_SIZE);
    middleEvents->nbEvents = SEGMENT_SIZE;
    {   U64 const distFromBegin = fpDistance(&fpstats->pastEvents, middleEvents, 8);
        U64 const distFromEnd = fpDistance(&fpstats->newEvents, middleEvents, 8);
        U64 const minDistance = SEGMENT_SIZE * SEGMENT_SIZE / 3;
        if (abs64((S64)distFromBegin - (S64)distFromEnd) < minDistance)
            return 64 KB;
        return (distFromBegin > distFromEnd) ? 32 KB : 96 KB;
    }
}

size_t ZSTD_splitBlock(const void* blockStart, size_t blockSize,
                    int level,
                    void* workspace, size_t wkspSize)
{
    DEBUGLOG(6, "ZSTD_splitBlock (level=%i)", level);
    assert(0<=level && level<=4);
    if (level == 0)
        return ZSTD_splitBlock_fromBorders(blockStart, blockSize, workspace, wkspSize);
    /* level >= 1*/
    return ZSTD_splitBlock_byChunks(blockStart, blockSize, level-1, workspace, wkspSize);
}
/**** ended inlining compress/zstd_preSplit.c ****/
/**** start inlining compress/zstd_compress.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/*-*************************************
*  Dependencies
***************************************/
/**** skipping file: ../common/allocations.h ****/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/error_private.h ****/
/**** skipping file: hist.h ****/
#define FSE_STATIC_LINKING_ONLY   /* FSE_encodeSymbol */
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_compress_sequences.h ****/
/**** skipping file: zstd_compress_literals.h ****/
/**** start inlining zstd_fast.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_FAST_H
#define ZSTD_FAST_H

/**** skipping file: ../common/mem.h ****/
/**** skipping file: zstd_compress_internal.h ****/

void ZSTD_fillHashTable(ZSTD_MatchState_t* ms,
                        void const* end, ZSTD_dictTableLoadMethod_e dtlm,
                        ZSTD_tableFillPurpose_e tfp);
size_t ZSTD_compressBlock_fast(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_fast_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_fast_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#endif /* ZSTD_FAST_H */
/**** ended inlining zstd_fast.h ****/
/**** start inlining zstd_double_fast.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_DOUBLE_FAST_H
#define ZSTD_DOUBLE_FAST_H

/**** skipping file: ../common/mem.h ****/
/**** skipping file: zstd_compress_internal.h ****/

#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR

void ZSTD_fillDoubleHashTable(ZSTD_MatchState_t* ms,
                              void const* end, ZSTD_dictTableLoadMethod_e dtlm,
                              ZSTD_tableFillPurpose_e tfp);

size_t ZSTD_compressBlock_doubleFast(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_doubleFast_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_doubleFast_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_DOUBLEFAST ZSTD_compressBlock_doubleFast
#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE ZSTD_compressBlock_doubleFast_dictMatchState
#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT ZSTD_compressBlock_doubleFast_extDict
#else
#define ZSTD_COMPRESSBLOCK_DOUBLEFAST NULL
#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT NULL
#endif /* ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR */

#endif /* ZSTD_DOUBLE_FAST_H */
/**** ended inlining zstd_double_fast.h ****/
/**** start inlining zstd_lazy.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_LAZY_H
#define ZSTD_LAZY_H

/**** skipping file: zstd_compress_internal.h ****/

/**
 * Dedicated Dictionary Search Structure bucket log. In the
 * ZSTD_dedicatedDictSearch mode, the hashTable has
 * 2 ** ZSTD_LAZY_DDSS_BUCKET_LOG entries in each bucket, rather than just
 * one.
 */
#define ZSTD_LAZY_DDSS_BUCKET_LOG 2

#define ZSTD_ROW_HASH_TAG_BITS 8        /* nb bits to use for the tag */

#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
U32 ZSTD_insertAndFindFirstIndex(ZSTD_MatchState_t* ms, const BYTE* ip);
void ZSTD_row_update(ZSTD_MatchState_t* const ms, const BYTE* ip);

void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_MatchState_t* ms, const BYTE* const ip);

void ZSTD_preserveUnsortedMark (U32* const table, U32 const size, U32 const reducerValue);  /*! used in ZSTD_reduceIndex(). preemptively increase value of ZSTD_DUBT_UNSORTED_MARK */
#endif

#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_greedy(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_greedy_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_GREEDY ZSTD_compressBlock_greedy
#define ZSTD_COMPRESSBLOCK_GREEDY_ROW ZSTD_compressBlock_greedy_row
#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE ZSTD_compressBlock_greedy_dictMatchState
#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW ZSTD_compressBlock_greedy_dictMatchState_row
#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH ZSTD_compressBlock_greedy_dedicatedDictSearch
#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_greedy_dedicatedDictSearch_row
#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT ZSTD_compressBlock_greedy_extDict
#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW ZSTD_compressBlock_greedy_extDict_row
#else
#define ZSTD_COMPRESSBLOCK_GREEDY NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_ROW NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT NULL
#define ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW NULL
#endif

#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_LAZY ZSTD_compressBlock_lazy
#define ZSTD_COMPRESSBLOCK_LAZY_ROW ZSTD_compressBlock_lazy_row
#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE ZSTD_compressBlock_lazy_dictMatchState
#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW ZSTD_compressBlock_lazy_dictMatchState_row
#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH ZSTD_compressBlock_lazy_dedicatedDictSearch
#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_lazy_dedicatedDictSearch_row
#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT ZSTD_compressBlock_lazy_extDict
#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW ZSTD_compressBlock_lazy_extDict_row
#else
#define ZSTD_COMPRESSBLOCK_LAZY NULL
#define ZSTD_COMPRESSBLOCK_LAZY_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH NULL
#define ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT NULL
#define ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW NULL
#endif

#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_lazy2_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_LAZY2 ZSTD_compressBlock_lazy2
#define ZSTD_COMPRESSBLOCK_LAZY2_ROW ZSTD_compressBlock_lazy2_row
#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE ZSTD_compressBlock_lazy2_dictMatchState
#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW ZSTD_compressBlock_lazy2_dictMatchState_row
#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH ZSTD_compressBlock_lazy2_dedicatedDictSearch
#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW ZSTD_compressBlock_lazy2_dedicatedDictSearch_row
#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT ZSTD_compressBlock_lazy2_extDict
#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW ZSTD_compressBlock_lazy2_extDict_row
#else
#define ZSTD_COMPRESSBLOCK_LAZY2 NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT NULL
#define ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW NULL
#endif

#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btlazy2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btlazy2_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btlazy2_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_BTLAZY2 ZSTD_compressBlock_btlazy2
#define ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE ZSTD_compressBlock_btlazy2_dictMatchState
#define ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT ZSTD_compressBlock_btlazy2_extDict
#else
#define ZSTD_COMPRESSBLOCK_BTLAZY2 NULL
#define ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT NULL
#endif

#endif /* ZSTD_LAZY_H */
/**** ended inlining zstd_lazy.h ****/
/**** start inlining zstd_opt.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_OPT_H
#define ZSTD_OPT_H

/**** skipping file: zstd_compress_internal.h ****/

#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)
/* used in ZSTD_loadDictionaryContent() */
void ZSTD_updateTree(ZSTD_MatchState_t* ms, const BYTE* ip, const BYTE* iend);
#endif

#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btopt(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btopt_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btopt_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_BTOPT ZSTD_compressBlock_btopt
#define ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE ZSTD_compressBlock_btopt_dictMatchState
#define ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT ZSTD_compressBlock_btopt_extDict
#else
#define ZSTD_COMPRESSBLOCK_BTOPT NULL
#define ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT NULL
#endif

#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btultra(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btultra_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);
size_t ZSTD_compressBlock_btultra_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

        /* note : no btultra2 variant for extDict nor dictMatchState,
         * because btultra2 is not meant to work with dictionaries
         * and is only specific for the first block (no prefix) */
size_t ZSTD_compressBlock_btultra2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize);

#define ZSTD_COMPRESSBLOCK_BTULTRA ZSTD_compressBlock_btultra
#define ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE ZSTD_compressBlock_btultra_dictMatchState
#define ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT ZSTD_compressBlock_btultra_extDict
#define ZSTD_COMPRESSBLOCK_BTULTRA2 ZSTD_compressBlock_btultra2
#else
#define ZSTD_COMPRESSBLOCK_BTULTRA NULL
#define ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE NULL
#define ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT NULL
#define ZSTD_COMPRESSBLOCK_BTULTRA2 NULL
#endif

#endif /* ZSTD_OPT_H */
/**** ended inlining zstd_opt.h ****/
/**** start inlining zstd_ldm.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_LDM_H
#define ZSTD_LDM_H

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: ../zstd.h ****/

/*-*************************************
*  Long distance matching
***************************************/

#define ZSTD_LDM_DEFAULT_WINDOW_LOG ZSTD_WINDOWLOG_LIMIT_DEFAULT

void ZSTD_ldm_fillHashTable(
            ldmState_t* state, const BYTE* ip,
            const BYTE* iend, ldmParams_t const* params);

/**
 * ZSTD_ldm_generateSequences():
 *
 * Generates the sequences using the long distance match finder.
 * Generates long range matching sequences in `sequences`, which parse a prefix
 * of the source. `sequences` must be large enough to store every sequence,
 * which can be checked with `ZSTD_ldm_getMaxNbSeq()`.
 * @returns 0 or an error code.
 *
 * NOTE: The user must have called ZSTD_window_update() for all of the input
 * they have, even if they pass it to ZSTD_ldm_generateSequences() in chunks.
 * NOTE: This function returns an error if it runs out of space to store
 *       sequences.
 */
size_t ZSTD_ldm_generateSequences(
            ldmState_t* ldms, RawSeqStore_t* sequences,
            ldmParams_t const* params, void const* src, size_t srcSize);

/**
 * ZSTD_ldm_blockCompress():
 *
 * Compresses a block using the predefined sequences, along with a secondary
 * block compressor. The literals section of every sequence is passed to the
 * secondary block compressor, and those sequences are interspersed with the
 * predefined sequences. Returns the length of the last literals.
 * Updates `rawSeqStore.pos` to indicate how many sequences have been consumed.
 * `rawSeqStore.seq` may also be updated to split the last sequence between two
 * blocks.
 * @return The length of the last literals.
 *
 * NOTE: The source must be at most the maximum block size, but the predefined
 * sequences can be any size, and may be longer than the block. In the case that
 * they are longer than the block, the last sequences may need to be split into
 * two. We handle that case correctly, and update `rawSeqStore` appropriately.
 * NOTE: This function does not return any errors.
 */
size_t ZSTD_ldm_blockCompress(RawSeqStore_t* rawSeqStore,
            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
            ZSTD_ParamSwitch_e useRowMatchFinder,
            void const* src, size_t srcSize);

/**
 * ZSTD_ldm_skipSequences():
 *
 * Skip past `srcSize` bytes worth of sequences in `rawSeqStore`.
 * Avoids emitting matches less than `minMatch` bytes.
 * Must be called for data that is not passed to ZSTD_ldm_blockCompress().
 */
void ZSTD_ldm_skipSequences(RawSeqStore_t* rawSeqStore, size_t srcSize,
    U32 const minMatch);

/* ZSTD_ldm_skipRawSeqStoreBytes():
 * Moves forward in rawSeqStore by nbBytes, updating fields 'pos' and 'posInSequence'.
 * Not to be used in conjunction with ZSTD_ldm_skipSequences().
 * Must be called for data with is not passed to ZSTD_ldm_blockCompress().
 */
void ZSTD_ldm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes);

/** ZSTD_ldm_getTableSize() :
 *  Estimate the space needed for long distance matching tables or 0 if LDM is
 *  disabled.
 */
size_t ZSTD_ldm_getTableSize(ldmParams_t params);

/** ZSTD_ldm_getSeqSpace() :
 *  Return an upper bound on the number of sequences that can be produced by
 *  the long distance matcher, or 0 if LDM is disabled.
 */
size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize);

/** ZSTD_ldm_adjustParameters() :
 *  If the params->hashRateLog is not set, set it to its default value based on
 *  windowLog and params->hashLog.
 *
 *  Ensures that params->bucketSizeLog is <= params->hashLog (setting it to
 *  params->hashLog if it is not).
 *
 *  Ensures that the minMatchLength >= targetLength during optimal parsing.
 */
void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                               ZSTD_compressionParameters const* cParams);

#endif /* ZSTD_FAST_H */
/**** ended inlining zstd_ldm.h ****/
/**** skipping file: zstd_compress_superblock.h ****/
/**** skipping file: ../common/bits.h ****/

/* ***************************************************************
*  Tuning parameters
*****************************************************************/
/*!
 * COMPRESS_HEAPMODE :
 * Select how default decompression function ZSTD_compress() allocates its context,
 * on stack (0, default), or into heap (1).
 * Note that functions with explicit context such as ZSTD_compressCCtx() are unaffected.
 */
#ifndef ZSTD_COMPRESS_HEAPMODE
#  define ZSTD_COMPRESS_HEAPMODE 0
#endif

/*!
 * ZSTD_HASHLOG3_MAX :
 * Maximum size of the hash table dedicated to find 3-bytes matches,
 * in log format, aka 17 => 1 << 17 == 128Ki positions.
 * This structure is only used in zstd_opt.
 * Since allocation is centralized for all strategies, it has to be known here.
 * The actual (selected) size of the hash table is then stored in ZSTD_MatchState_t.hashLog3,
 * so that zstd_opt.c doesn't need to know about this constant.
 */
#ifndef ZSTD_HASHLOG3_MAX
#  define ZSTD_HASHLOG3_MAX 17
#endif

/*-*************************************
*  Helper functions
***************************************/
/* ZSTD_compressBound()
 * Note that the result from this function is only valid for
 * the one-pass compression functions.
 * When employing the streaming mode,
 * if flushes are frequently altering the size of blocks,
 * the overhead from block headers can make the compressed data larger
 * than the return value of ZSTD_compressBound().
 */
size_t ZSTD_compressBound(size_t srcSize) {
    size_t const r = ZSTD_COMPRESSBOUND(srcSize);
    if (r==0) return ERROR(srcSize_wrong);
    return r;
}


/*-*************************************
*  Context memory management
***************************************/
struct ZSTD_CDict_s {
    const void* dictContent;
    size_t dictContentSize;
    ZSTD_dictContentType_e dictContentType; /* The dictContentType the CDict was created with */
    U32* entropyWorkspace; /* entropy workspace of HUF_WORKSPACE_SIZE bytes */
    ZSTD_cwksp workspace;
    ZSTD_MatchState_t matchState;
    ZSTD_compressedBlockState_t cBlockState;
    ZSTD_customMem customMem;
    U32 dictID;
    int compressionLevel; /* 0 indicates that advanced API was used to select CDict params */
    ZSTD_ParamSwitch_e useRowMatchFinder; /* Indicates whether the CDict was created with params that would use
                                           * row-based matchfinder. Unless the cdict is reloaded, we will use
                                           * the same greedy/lazy matchfinder at compression time.
                                           */
};  /* typedef'd to ZSTD_CDict within "zstd.h" */

ZSTD_CCtx* ZSTD_createCCtx(void)
{
    return ZSTD_createCCtx_advanced(ZSTD_defaultCMem);
}

static void ZSTD_initCCtx(ZSTD_CCtx* cctx, ZSTD_customMem memManager)
{
    assert(cctx != NULL);
    ZSTD_memset(cctx, 0, sizeof(*cctx));
    cctx->customMem = memManager;
    cctx->bmi2 = ZSTD_cpuSupportsBmi2();
    {   size_t const err = ZSTD_CCtx_reset(cctx, ZSTD_reset_parameters);
        assert(!ZSTD_isError(err));
        (void)err;
    }
}

ZSTD_CCtx* ZSTD_createCCtx_advanced(ZSTD_customMem customMem)
{
    ZSTD_STATIC_ASSERT(zcss_init==0);
    ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN==(0ULL - 1));
    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
    {   ZSTD_CCtx* const cctx = (ZSTD_CCtx*)ZSTD_customMalloc(sizeof(ZSTD_CCtx), customMem);
        if (!cctx) return NULL;
        ZSTD_initCCtx(cctx, customMem);
        return cctx;
    }
}

ZSTD_CCtx* ZSTD_initStaticCCtx(void* workspace, size_t workspaceSize)
{
    ZSTD_cwksp ws;
    ZSTD_CCtx* cctx;
    if (workspaceSize <= sizeof(ZSTD_CCtx)) return NULL;  /* minimum size */
    if ((size_t)workspace & 7) return NULL;  /* must be 8-aligned */
    ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc);

    cctx = (ZSTD_CCtx*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CCtx));
    if (cctx == NULL) return NULL;

    ZSTD_memset(cctx, 0, sizeof(ZSTD_CCtx));
    ZSTD_cwksp_move(&cctx->workspace, &ws);
    cctx->staticSize = workspaceSize;

    /* statically sized space. tmpWorkspace never moves (but prev/next block swap places) */
    if (!ZSTD_cwksp_check_available(&cctx->workspace, TMP_WORKSPACE_SIZE + 2 * sizeof(ZSTD_compressedBlockState_t))) return NULL;
    cctx->blockState.prevCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
    cctx->blockState.nextCBlock = (ZSTD_compressedBlockState_t*)ZSTD_cwksp_reserve_object(&cctx->workspace, sizeof(ZSTD_compressedBlockState_t));
    cctx->tmpWorkspace = ZSTD_cwksp_reserve_object(&cctx->workspace, TMP_WORKSPACE_SIZE);
    cctx->tmpWkspSize = TMP_WORKSPACE_SIZE;
    cctx->bmi2 = ZSTD_cpuid_bmi2(ZSTD_cpuid());
    return cctx;
}

/**
 * Clears and frees all of the dictionaries in the CCtx.
 */
static void ZSTD_clearAllDicts(ZSTD_CCtx* cctx)
{
    ZSTD_customFree(cctx->localDict.dictBuffer, cctx->customMem);
    ZSTD_freeCDict(cctx->localDict.cdict);
    ZSTD_memset(&cctx->localDict, 0, sizeof(cctx->localDict));
    ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));
    cctx->cdict = NULL;
}

static size_t ZSTD_sizeof_localDict(ZSTD_localDict dict)
{
    size_t const bufferSize = dict.dictBuffer != NULL ? dict.dictSize : 0;
    size_t const cdictSize = ZSTD_sizeof_CDict(dict.cdict);
    return bufferSize + cdictSize;
}

static void ZSTD_freeCCtxContent(ZSTD_CCtx* cctx)
{
    assert(cctx != NULL);
    assert(cctx->staticSize == 0);
    ZSTD_clearAllDicts(cctx);
#ifdef ZSTD_MULTITHREAD
    ZSTDMT_freeCCtx(cctx->mtctx); cctx->mtctx = NULL;
#endif
    ZSTD_cwksp_free(&cctx->workspace, cctx->customMem);
}

size_t ZSTD_freeCCtx(ZSTD_CCtx* cctx)
{
    DEBUGLOG(3, "ZSTD_freeCCtx (address: %p)", (void*)cctx);
    if (cctx==NULL) return 0;   /* support free on NULL */
    RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
                    "not compatible with static CCtx");
    {   int cctxInWorkspace = ZSTD_cwksp_owns_buffer(&cctx->workspace, cctx);
        ZSTD_freeCCtxContent(cctx);
        if (!cctxInWorkspace) ZSTD_customFree(cctx, cctx->customMem);
    }
    return 0;
}


static size_t ZSTD_sizeof_mtctx(const ZSTD_CCtx* cctx)
{
#ifdef ZSTD_MULTITHREAD
    return ZSTDMT_sizeof_CCtx(cctx->mtctx);
#else
    (void)cctx;
    return 0;
#endif
}


size_t ZSTD_sizeof_CCtx(const ZSTD_CCtx* cctx)
{
    if (cctx==NULL) return 0;   /* support sizeof on NULL */
    /* cctx may be in the workspace */
    return (cctx->workspace.workspace == cctx ? 0 : sizeof(*cctx))
           + ZSTD_cwksp_sizeof(&cctx->workspace)
           + ZSTD_sizeof_localDict(cctx->localDict)
           + ZSTD_sizeof_mtctx(cctx);
}

size_t ZSTD_sizeof_CStream(const ZSTD_CStream* zcs)
{
    return ZSTD_sizeof_CCtx(zcs);  /* same object */
}

/* private API call, for dictBuilder only */
const SeqStore_t* ZSTD_getSeqStore(const ZSTD_CCtx* ctx) { return &(ctx->seqStore); }

/* Returns true if the strategy supports using a row based matchfinder */
static int ZSTD_rowMatchFinderSupported(const ZSTD_strategy strategy) {
    return (strategy >= ZSTD_greedy && strategy <= ZSTD_lazy2);
}

/* Returns true if the strategy and useRowMatchFinder mode indicate that we will use the row based matchfinder
 * for this compression.
 */
static int ZSTD_rowMatchFinderUsed(const ZSTD_strategy strategy, const ZSTD_ParamSwitch_e mode) {
    assert(mode != ZSTD_ps_auto);
    return ZSTD_rowMatchFinderSupported(strategy) && (mode == ZSTD_ps_enable);
}

/* Returns row matchfinder usage given an initial mode and cParams */
static ZSTD_ParamSwitch_e ZSTD_resolveRowMatchFinderMode(ZSTD_ParamSwitch_e mode,
                                                         const ZSTD_compressionParameters* const cParams) {
    if (mode != ZSTD_ps_auto) return mode; /* if requested enabled, but no SIMD, we still will use row matchfinder */
    mode = ZSTD_ps_disable;
    if (!ZSTD_rowMatchFinderSupported(cParams->strategy)) return mode;
    if (cParams->windowLog > 14) mode = ZSTD_ps_enable;
    return mode;
}

/* Returns block splitter usage (generally speaking, when using slower/stronger compression modes) */
static ZSTD_ParamSwitch_e ZSTD_resolveBlockSplitterMode(ZSTD_ParamSwitch_e mode,
                                                        const ZSTD_compressionParameters* const cParams) {
    if (mode != ZSTD_ps_auto) return mode;
    return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 17) ? ZSTD_ps_enable : ZSTD_ps_disable;
}

/* Returns 1 if the arguments indicate that we should allocate a chainTable, 0 otherwise */
static int ZSTD_allocateChainTable(const ZSTD_strategy strategy,
                                   const ZSTD_ParamSwitch_e useRowMatchFinder,
                                   const U32 forDDSDict) {
    assert(useRowMatchFinder != ZSTD_ps_auto);
    /* We always should allocate a chaintable if we are allocating a matchstate for a DDS dictionary matchstate.
     * We do not allocate a chaintable if we are using ZSTD_fast, or are using the row-based matchfinder.
     */
    return forDDSDict || ((strategy != ZSTD_fast) && !ZSTD_rowMatchFinderUsed(strategy, useRowMatchFinder));
}

/* Returns ZSTD_ps_enable if compression parameters are such that we should
 * enable long distance matching (wlog >= 27, strategy >= btopt).
 * Returns ZSTD_ps_disable otherwise.
 */
static ZSTD_ParamSwitch_e ZSTD_resolveEnableLdm(ZSTD_ParamSwitch_e mode,
                                 const ZSTD_compressionParameters* const cParams) {
    if (mode != ZSTD_ps_auto) return mode;
    return (cParams->strategy >= ZSTD_btopt && cParams->windowLog >= 27) ? ZSTD_ps_enable : ZSTD_ps_disable;
}

static int ZSTD_resolveExternalSequenceValidation(int mode) {
    return mode;
}

/* Resolves maxBlockSize to the default if no value is present. */
static size_t ZSTD_resolveMaxBlockSize(size_t maxBlockSize) {
    if (maxBlockSize == 0) {
        return ZSTD_BLOCKSIZE_MAX;
    } else {
        return maxBlockSize;
    }
}

static ZSTD_ParamSwitch_e ZSTD_resolveExternalRepcodeSearch(ZSTD_ParamSwitch_e value, int cLevel) {
    if (value != ZSTD_ps_auto) return value;
    if (cLevel < 10) {
        return ZSTD_ps_disable;
    } else {
        return ZSTD_ps_enable;
    }
}

/* Returns 1 if compression parameters are such that CDict hashtable and chaintable indices are tagged.
 * If so, the tags need to be removed in ZSTD_resetCCtx_byCopyingCDict. */
static int ZSTD_CDictIndicesAreTagged(const ZSTD_compressionParameters* const cParams) {
    return cParams->strategy == ZSTD_fast || cParams->strategy == ZSTD_dfast;
}

static ZSTD_CCtx_params ZSTD_makeCCtxParamsFromCParams(
        ZSTD_compressionParameters cParams)
{
    ZSTD_CCtx_params cctxParams;
    /* should not matter, as all cParams are presumed properly defined */
    ZSTD_CCtxParams_init(&cctxParams, ZSTD_CLEVEL_DEFAULT);
    cctxParams.cParams = cParams;

    /* Adjust advanced params according to cParams */
    cctxParams.ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams.ldmParams.enableLdm, &cParams);
    if (cctxParams.ldmParams.enableLdm == ZSTD_ps_enable) {
        ZSTD_ldm_adjustParameters(&cctxParams.ldmParams, &cParams);
        assert(cctxParams.ldmParams.hashLog >= cctxParams.ldmParams.bucketSizeLog);
        assert(cctxParams.ldmParams.hashRateLog < 32);
    }
    cctxParams.postBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams.postBlockSplitter, &cParams);
    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);
    cctxParams.validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams.validateSequences);
    cctxParams.maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams.maxBlockSize);
    cctxParams.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams.searchForExternalRepcodes,
                                                                             cctxParams.compressionLevel);
    assert(!ZSTD_checkCParams(cParams));
    return cctxParams;
}

static ZSTD_CCtx_params* ZSTD_createCCtxParams_advanced(
        ZSTD_customMem customMem)
{
    ZSTD_CCtx_params* params;
    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
    params = (ZSTD_CCtx_params*)ZSTD_customCalloc(
            sizeof(ZSTD_CCtx_params), customMem);
    if (!params) { return NULL; }
    ZSTD_CCtxParams_init(params, ZSTD_CLEVEL_DEFAULT);
    params->customMem = customMem;
    return params;
}

ZSTD_CCtx_params* ZSTD_createCCtxParams(void)
{
    return ZSTD_createCCtxParams_advanced(ZSTD_defaultCMem);
}

size_t ZSTD_freeCCtxParams(ZSTD_CCtx_params* params)
{
    if (params == NULL) { return 0; }
    ZSTD_customFree(params, params->customMem);
    return 0;
}

size_t ZSTD_CCtxParams_reset(ZSTD_CCtx_params* params)
{
    return ZSTD_CCtxParams_init(params, ZSTD_CLEVEL_DEFAULT);
}

size_t ZSTD_CCtxParams_init(ZSTD_CCtx_params* cctxParams, int compressionLevel) {
    RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!");
    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
    cctxParams->compressionLevel = compressionLevel;
    cctxParams->fParams.contentSizeFlag = 1;
    return 0;
}

#define ZSTD_NO_CLEVEL 0

/**
 * Initializes `cctxParams` from `params` and `compressionLevel`.
 * @param compressionLevel If params are derived from a compression level then that compression level, otherwise ZSTD_NO_CLEVEL.
 */
static void
ZSTD_CCtxParams_init_internal(ZSTD_CCtx_params* cctxParams,
                        const ZSTD_parameters* params,
                              int compressionLevel)
{
    assert(!ZSTD_checkCParams(params->cParams));
    ZSTD_memset(cctxParams, 0, sizeof(*cctxParams));
    cctxParams->cParams = params->cParams;
    cctxParams->fParams = params->fParams;
    /* Should not matter, as all cParams are presumed properly defined.
     * But, set it for tracing anyway.
     */
    cctxParams->compressionLevel = compressionLevel;
    cctxParams->useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams->useRowMatchFinder, &params->cParams);
    cctxParams->postBlockSplitter = ZSTD_resolveBlockSplitterMode(cctxParams->postBlockSplitter, &params->cParams);
    cctxParams->ldmParams.enableLdm = ZSTD_resolveEnableLdm(cctxParams->ldmParams.enableLdm, &params->cParams);
    cctxParams->validateSequences = ZSTD_resolveExternalSequenceValidation(cctxParams->validateSequences);
    cctxParams->maxBlockSize = ZSTD_resolveMaxBlockSize(cctxParams->maxBlockSize);
    cctxParams->searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(cctxParams->searchForExternalRepcodes, compressionLevel);
    DEBUGLOG(4, "ZSTD_CCtxParams_init_internal: useRowMatchFinder=%d, useBlockSplitter=%d ldm=%d",
                cctxParams->useRowMatchFinder, cctxParams->postBlockSplitter, cctxParams->ldmParams.enableLdm);
}

size_t ZSTD_CCtxParams_init_advanced(ZSTD_CCtx_params* cctxParams, ZSTD_parameters params)
{
    RETURN_ERROR_IF(!cctxParams, GENERIC, "NULL pointer!");
    FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
    ZSTD_CCtxParams_init_internal(cctxParams, &params, ZSTD_NO_CLEVEL);
    return 0;
}

/**
 * Sets cctxParams' cParams and fParams from params, but otherwise leaves them alone.
 * @param params Validated zstd parameters.
 */
static void ZSTD_CCtxParams_setZstdParams(
        ZSTD_CCtx_params* cctxParams, const ZSTD_parameters* params)
{
    assert(!ZSTD_checkCParams(params->cParams));
    cctxParams->cParams = params->cParams;
    cctxParams->fParams = params->fParams;
    /* Should not matter, as all cParams are presumed properly defined.
     * But, set it for tracing anyway.
     */
    cctxParams->compressionLevel = ZSTD_NO_CLEVEL;
}

ZSTD_bounds ZSTD_cParam_getBounds(ZSTD_cParameter param)
{
    ZSTD_bounds bounds = { 0, 0, 0 };

    switch(param)
    {
    case ZSTD_c_compressionLevel:
        bounds.lowerBound = ZSTD_minCLevel();
        bounds.upperBound = ZSTD_maxCLevel();
        return bounds;

    case ZSTD_c_windowLog:
        bounds.lowerBound = ZSTD_WINDOWLOG_MIN;
        bounds.upperBound = ZSTD_WINDOWLOG_MAX;
        return bounds;

    case ZSTD_c_hashLog:
        bounds.lowerBound = ZSTD_HASHLOG_MIN;
        bounds.upperBound = ZSTD_HASHLOG_MAX;
        return bounds;

    case ZSTD_c_chainLog:
        bounds.lowerBound = ZSTD_CHAINLOG_MIN;
        bounds.upperBound = ZSTD_CHAINLOG_MAX;
        return bounds;

    case ZSTD_c_searchLog:
        bounds.lowerBound = ZSTD_SEARCHLOG_MIN;
        bounds.upperBound = ZSTD_SEARCHLOG_MAX;
        return bounds;

    case ZSTD_c_minMatch:
        bounds.lowerBound = ZSTD_MINMATCH_MIN;
        bounds.upperBound = ZSTD_MINMATCH_MAX;
        return bounds;

    case ZSTD_c_targetLength:
        bounds.lowerBound = ZSTD_TARGETLENGTH_MIN;
        bounds.upperBound = ZSTD_TARGETLENGTH_MAX;
        return bounds;

    case ZSTD_c_strategy:
        bounds.lowerBound = ZSTD_STRATEGY_MIN;
        bounds.upperBound = ZSTD_STRATEGY_MAX;
        return bounds;

    case ZSTD_c_contentSizeFlag:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_checksumFlag:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_dictIDFlag:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_nbWorkers:
        bounds.lowerBound = 0;
#ifdef ZSTD_MULTITHREAD
        bounds.upperBound = ZSTDMT_NBWORKERS_MAX;
#else
        bounds.upperBound = 0;
#endif
        return bounds;

    case ZSTD_c_jobSize:
        bounds.lowerBound = 0;
#ifdef ZSTD_MULTITHREAD
        bounds.upperBound = ZSTDMT_JOBSIZE_MAX;
#else
        bounds.upperBound = 0;
#endif
        return bounds;

    case ZSTD_c_overlapLog:
#ifdef ZSTD_MULTITHREAD
        bounds.lowerBound = ZSTD_OVERLAPLOG_MIN;
        bounds.upperBound = ZSTD_OVERLAPLOG_MAX;
#else
        bounds.lowerBound = 0;
        bounds.upperBound = 0;
#endif
        return bounds;

    case ZSTD_c_enableDedicatedDictSearch:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_enableLongDistanceMatching:
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    case ZSTD_c_ldmHashLog:
        bounds.lowerBound = ZSTD_LDM_HASHLOG_MIN;
        bounds.upperBound = ZSTD_LDM_HASHLOG_MAX;
        return bounds;

    case ZSTD_c_ldmMinMatch:
        bounds.lowerBound = ZSTD_LDM_MINMATCH_MIN;
        bounds.upperBound = ZSTD_LDM_MINMATCH_MAX;
        return bounds;

    case ZSTD_c_ldmBucketSizeLog:
        bounds.lowerBound = ZSTD_LDM_BUCKETSIZELOG_MIN;
        bounds.upperBound = ZSTD_LDM_BUCKETSIZELOG_MAX;
        return bounds;

    case ZSTD_c_ldmHashRateLog:
        bounds.lowerBound = ZSTD_LDM_HASHRATELOG_MIN;
        bounds.upperBound = ZSTD_LDM_HASHRATELOG_MAX;
        return bounds;

    /* experimental parameters */
    case ZSTD_c_rsyncable:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_forceMaxWindow :
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_format:
        ZSTD_STATIC_ASSERT(ZSTD_f_zstd1 < ZSTD_f_zstd1_magicless);
        bounds.lowerBound = ZSTD_f_zstd1;
        bounds.upperBound = ZSTD_f_zstd1_magicless;   /* note : how to ensure at compile time that this is the highest value enum ? */
        return bounds;

    case ZSTD_c_forceAttachDict:
        ZSTD_STATIC_ASSERT(ZSTD_dictDefaultAttach < ZSTD_dictForceLoad);
        bounds.lowerBound = ZSTD_dictDefaultAttach;
        bounds.upperBound = ZSTD_dictForceLoad;       /* note : how to ensure at compile time that this is the highest value enum ? */
        return bounds;

    case ZSTD_c_literalCompressionMode:
        ZSTD_STATIC_ASSERT(ZSTD_ps_auto < ZSTD_ps_enable && ZSTD_ps_enable < ZSTD_ps_disable);
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    case ZSTD_c_targetCBlockSize:
        bounds.lowerBound = ZSTD_TARGETCBLOCKSIZE_MIN;
        bounds.upperBound = ZSTD_TARGETCBLOCKSIZE_MAX;
        return bounds;

    case ZSTD_c_srcSizeHint:
        bounds.lowerBound = ZSTD_SRCSIZEHINT_MIN;
        bounds.upperBound = ZSTD_SRCSIZEHINT_MAX;
        return bounds;

    case ZSTD_c_stableInBuffer:
    case ZSTD_c_stableOutBuffer:
        bounds.lowerBound = (int)ZSTD_bm_buffered;
        bounds.upperBound = (int)ZSTD_bm_stable;
        return bounds;

    case ZSTD_c_blockDelimiters:
        bounds.lowerBound = (int)ZSTD_sf_noBlockDelimiters;
        bounds.upperBound = (int)ZSTD_sf_explicitBlockDelimiters;
        return bounds;

    case ZSTD_c_validateSequences:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_splitAfterSequences:
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    case ZSTD_c_blockSplitterLevel:
        bounds.lowerBound = 0;
        bounds.upperBound = ZSTD_BLOCKSPLITTER_LEVEL_MAX;
        return bounds;

    case ZSTD_c_useRowMatchFinder:
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    case ZSTD_c_deterministicRefPrefix:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_prefetchCDictTables:
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    case ZSTD_c_enableSeqProducerFallback:
        bounds.lowerBound = 0;
        bounds.upperBound = 1;
        return bounds;

    case ZSTD_c_maxBlockSize:
        bounds.lowerBound = ZSTD_BLOCKSIZE_MAX_MIN;
        bounds.upperBound = ZSTD_BLOCKSIZE_MAX;
        return bounds;

    case ZSTD_c_repcodeResolution:
        bounds.lowerBound = (int)ZSTD_ps_auto;
        bounds.upperBound = (int)ZSTD_ps_disable;
        return bounds;

    default:
        bounds.error = ERROR(parameter_unsupported);
        return bounds;
    }
}

/* ZSTD_cParam_clampBounds:
 * Clamps the value into the bounded range.
 */
static size_t ZSTD_cParam_clampBounds(ZSTD_cParameter cParam, int* value)
{
    ZSTD_bounds const bounds = ZSTD_cParam_getBounds(cParam);
    if (ZSTD_isError(bounds.error)) return bounds.error;
    if (*value < bounds.lowerBound) *value = bounds.lowerBound;
    if (*value > bounds.upperBound) *value = bounds.upperBound;
    return 0;
}

#define BOUNDCHECK(cParam, val)                                       \
    do {                                                              \
        RETURN_ERROR_IF(!ZSTD_cParam_withinBounds(cParam,val),        \
                        parameter_outOfBound, "Param out of bounds"); \
    } while (0)


static int ZSTD_isUpdateAuthorized(ZSTD_cParameter param)
{
    switch(param)
    {
    case ZSTD_c_compressionLevel:
    case ZSTD_c_hashLog:
    case ZSTD_c_chainLog:
    case ZSTD_c_searchLog:
    case ZSTD_c_minMatch:
    case ZSTD_c_targetLength:
    case ZSTD_c_strategy:
    case ZSTD_c_blockSplitterLevel:
        return 1;

    case ZSTD_c_format:
    case ZSTD_c_windowLog:
    case ZSTD_c_contentSizeFlag:
    case ZSTD_c_checksumFlag:
    case ZSTD_c_dictIDFlag:
    case ZSTD_c_forceMaxWindow :
    case ZSTD_c_nbWorkers:
    case ZSTD_c_jobSize:
    case ZSTD_c_overlapLog:
    case ZSTD_c_rsyncable:
    case ZSTD_c_enableDedicatedDictSearch:
    case ZSTD_c_enableLongDistanceMatching:
    case ZSTD_c_ldmHashLog:
    case ZSTD_c_ldmMinMatch:
    case ZSTD_c_ldmBucketSizeLog:
    case ZSTD_c_ldmHashRateLog:
    case ZSTD_c_forceAttachDict:
    case ZSTD_c_literalCompressionMode:
    case ZSTD_c_targetCBlockSize:
    case ZSTD_c_srcSizeHint:
    case ZSTD_c_stableInBuffer:
    case ZSTD_c_stableOutBuffer:
    case ZSTD_c_blockDelimiters:
    case ZSTD_c_validateSequences:
    case ZSTD_c_splitAfterSequences:
    case ZSTD_c_useRowMatchFinder:
    case ZSTD_c_deterministicRefPrefix:
    case ZSTD_c_prefetchCDictTables:
    case ZSTD_c_enableSeqProducerFallback:
    case ZSTD_c_maxBlockSize:
    case ZSTD_c_repcodeResolution:
    default:
        return 0;
    }
}

size_t ZSTD_CCtx_setParameter(ZSTD_CCtx* cctx, ZSTD_cParameter param, int value)
{
    DEBUGLOG(4, "ZSTD_CCtx_setParameter (%i, %i)", (int)param, value);
    if (cctx->streamStage != zcss_init) {
        if (ZSTD_isUpdateAuthorized(param)) {
            cctx->cParamsChanged = 1;
        } else {
            RETURN_ERROR(stage_wrong, "can only set params in cctx init stage");
    }   }

    switch(param)
    {
    case ZSTD_c_nbWorkers:
        RETURN_ERROR_IF((value!=0) && cctx->staticSize, parameter_unsupported,
                        "MT not compatible with static alloc");
        break;

    case ZSTD_c_compressionLevel:
    case ZSTD_c_windowLog:
    case ZSTD_c_hashLog:
    case ZSTD_c_chainLog:
    case ZSTD_c_searchLog:
    case ZSTD_c_minMatch:
    case ZSTD_c_targetLength:
    case ZSTD_c_strategy:
    case ZSTD_c_ldmHashRateLog:
    case ZSTD_c_format:
    case ZSTD_c_contentSizeFlag:
    case ZSTD_c_checksumFlag:
    case ZSTD_c_dictIDFlag:
    case ZSTD_c_forceMaxWindow:
    case ZSTD_c_forceAttachDict:
    case ZSTD_c_literalCompressionMode:
    case ZSTD_c_jobSize:
    case ZSTD_c_overlapLog:
    case ZSTD_c_rsyncable:
    case ZSTD_c_enableDedicatedDictSearch:
    case ZSTD_c_enableLongDistanceMatching:
    case ZSTD_c_ldmHashLog:
    case ZSTD_c_ldmMinMatch:
    case ZSTD_c_ldmBucketSizeLog:
    case ZSTD_c_targetCBlockSize:
    case ZSTD_c_srcSizeHint:
    case ZSTD_c_stableInBuffer:
    case ZSTD_c_stableOutBuffer:
    case ZSTD_c_blockDelimiters:
    case ZSTD_c_validateSequences:
    case ZSTD_c_splitAfterSequences:
    case ZSTD_c_blockSplitterLevel:
    case ZSTD_c_useRowMatchFinder:
    case ZSTD_c_deterministicRefPrefix:
    case ZSTD_c_prefetchCDictTables:
    case ZSTD_c_enableSeqProducerFallback:
    case ZSTD_c_maxBlockSize:
    case ZSTD_c_repcodeResolution:
        break;

    default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
    }
    return ZSTD_CCtxParams_setParameter(&cctx->requestedParams, param, value);
}

size_t ZSTD_CCtxParams_setParameter(ZSTD_CCtx_params* CCtxParams,
                                    ZSTD_cParameter param, int value)
{
    DEBUGLOG(4, "ZSTD_CCtxParams_setParameter (%i, %i)", (int)param, value);
    switch(param)
    {
    case ZSTD_c_format :
        BOUNDCHECK(ZSTD_c_format, value);
        CCtxParams->format = (ZSTD_format_e)value;
        return (size_t)CCtxParams->format;

    case ZSTD_c_compressionLevel : {
        FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(param, &value), "");
        if (value == 0)
            CCtxParams->compressionLevel = ZSTD_CLEVEL_DEFAULT; /* 0 == default */
        else
            CCtxParams->compressionLevel = value;
        if (CCtxParams->compressionLevel >= 0) return (size_t)CCtxParams->compressionLevel;
        return 0;  /* return type (size_t) cannot represent negative values */
    }

    case ZSTD_c_windowLog :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_windowLog, value);
        CCtxParams->cParams.windowLog = (U32)value;
        return CCtxParams->cParams.windowLog;

    case ZSTD_c_hashLog :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_hashLog, value);
        CCtxParams->cParams.hashLog = (U32)value;
        return CCtxParams->cParams.hashLog;

    case ZSTD_c_chainLog :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_chainLog, value);
        CCtxParams->cParams.chainLog = (U32)value;
        return CCtxParams->cParams.chainLog;

    case ZSTD_c_searchLog :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_searchLog, value);
        CCtxParams->cParams.searchLog = (U32)value;
        return (size_t)value;

    case ZSTD_c_minMatch :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_minMatch, value);
        CCtxParams->cParams.minMatch = (U32)value;
        return CCtxParams->cParams.minMatch;

    case ZSTD_c_targetLength :
        BOUNDCHECK(ZSTD_c_targetLength, value);
        CCtxParams->cParams.targetLength = (U32)value;
        return CCtxParams->cParams.targetLength;

    case ZSTD_c_strategy :
        if (value!=0)   /* 0 => use default */
            BOUNDCHECK(ZSTD_c_strategy, value);
        CCtxParams->cParams.strategy = (ZSTD_strategy)value;
        return (size_t)CCtxParams->cParams.strategy;

    case ZSTD_c_contentSizeFlag :
        /* Content size written in frame header _when known_ (default:1) */
        DEBUGLOG(4, "set content size flag = %u", (value!=0));
        CCtxParams->fParams.contentSizeFlag = value != 0;
        return (size_t)CCtxParams->fParams.contentSizeFlag;

    case ZSTD_c_checksumFlag :
        /* A 32-bits content checksum will be calculated and written at end of frame (default:0) */
        CCtxParams->fParams.checksumFlag = value != 0;
        return (size_t)CCtxParams->fParams.checksumFlag;

    case ZSTD_c_dictIDFlag : /* When applicable, dictionary's dictID is provided in frame header (default:1) */
        DEBUGLOG(4, "set dictIDFlag = %u", (value!=0));
        CCtxParams->fParams.noDictIDFlag = !value;
        return !CCtxParams->fParams.noDictIDFlag;

    case ZSTD_c_forceMaxWindow :
        CCtxParams->forceWindow = (value != 0);
        return (size_t)CCtxParams->forceWindow;

    case ZSTD_c_forceAttachDict : {
        const ZSTD_dictAttachPref_e pref = (ZSTD_dictAttachPref_e)value;
        BOUNDCHECK(ZSTD_c_forceAttachDict, (int)pref);
        CCtxParams->attachDictPref = pref;
        return CCtxParams->attachDictPref;
    }

    case ZSTD_c_literalCompressionMode : {
        const ZSTD_ParamSwitch_e lcm = (ZSTD_ParamSwitch_e)value;
        BOUNDCHECK(ZSTD_c_literalCompressionMode, (int)lcm);
        CCtxParams->literalCompressionMode = lcm;
        return CCtxParams->literalCompressionMode;
    }

    case ZSTD_c_nbWorkers :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR_IF(value!=0, parameter_unsupported, "not compiled with multithreading");
        return 0;
#else
        FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(param, &value), "");
        CCtxParams->nbWorkers = value;
        return (size_t)(CCtxParams->nbWorkers);
#endif

    case ZSTD_c_jobSize :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR_IF(value!=0, parameter_unsupported, "not compiled with multithreading");
        return 0;
#else
        /* Adjust to the minimum non-default value. */
        if (value != 0 && value < ZSTDMT_JOBSIZE_MIN)
            value = ZSTDMT_JOBSIZE_MIN;
        FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(param, &value), "");
        assert(value >= 0);
        CCtxParams->jobSize = (size_t)value;
        return CCtxParams->jobSize;
#endif

    case ZSTD_c_overlapLog :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR_IF(value!=0, parameter_unsupported, "not compiled with multithreading");
        return 0;
#else
        FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(ZSTD_c_overlapLog, &value), "");
        CCtxParams->overlapLog = value;
        return (size_t)CCtxParams->overlapLog;
#endif

    case ZSTD_c_rsyncable :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR_IF(value!=0, parameter_unsupported, "not compiled with multithreading");
        return 0;
#else
        FORWARD_IF_ERROR(ZSTD_cParam_clampBounds(ZSTD_c_overlapLog, &value), "");
        CCtxParams->rsyncable = value;
        return (size_t)CCtxParams->rsyncable;
#endif

    case ZSTD_c_enableDedicatedDictSearch :
        CCtxParams->enableDedicatedDictSearch = (value!=0);
        return (size_t)CCtxParams->enableDedicatedDictSearch;

    case ZSTD_c_enableLongDistanceMatching :
        BOUNDCHECK(ZSTD_c_enableLongDistanceMatching, value);
        CCtxParams->ldmParams.enableLdm = (ZSTD_ParamSwitch_e)value;
        return CCtxParams->ldmParams.enableLdm;

    case ZSTD_c_ldmHashLog :
        if (value!=0)   /* 0 ==> auto */
            BOUNDCHECK(ZSTD_c_ldmHashLog, value);
        CCtxParams->ldmParams.hashLog = (U32)value;
        return CCtxParams->ldmParams.hashLog;

    case ZSTD_c_ldmMinMatch :
        if (value!=0)   /* 0 ==> default */
            BOUNDCHECK(ZSTD_c_ldmMinMatch, value);
        CCtxParams->ldmParams.minMatchLength = (U32)value;
        return CCtxParams->ldmParams.minMatchLength;

    case ZSTD_c_ldmBucketSizeLog :
        if (value!=0)   /* 0 ==> default */
            BOUNDCHECK(ZSTD_c_ldmBucketSizeLog, value);
        CCtxParams->ldmParams.bucketSizeLog = (U32)value;
        return CCtxParams->ldmParams.bucketSizeLog;

    case ZSTD_c_ldmHashRateLog :
        if (value!=0)   /* 0 ==> default */
            BOUNDCHECK(ZSTD_c_ldmHashRateLog, value);
        CCtxParams->ldmParams.hashRateLog = (U32)value;
        return CCtxParams->ldmParams.hashRateLog;

    case ZSTD_c_targetCBlockSize :
        if (value!=0) {  /* 0 ==> default */
            value = MAX(value, ZSTD_TARGETCBLOCKSIZE_MIN);
            BOUNDCHECK(ZSTD_c_targetCBlockSize, value);
        }
        CCtxParams->targetCBlockSize = (U32)value;
        return CCtxParams->targetCBlockSize;

    case ZSTD_c_srcSizeHint :
        if (value!=0)    /* 0 ==> default */
            BOUNDCHECK(ZSTD_c_srcSizeHint, value);
        CCtxParams->srcSizeHint = value;
        return (size_t)CCtxParams->srcSizeHint;

    case ZSTD_c_stableInBuffer:
        BOUNDCHECK(ZSTD_c_stableInBuffer, value);
        CCtxParams->inBufferMode = (ZSTD_bufferMode_e)value;
        return CCtxParams->inBufferMode;

    case ZSTD_c_stableOutBuffer:
        BOUNDCHECK(ZSTD_c_stableOutBuffer, value);
        CCtxParams->outBufferMode = (ZSTD_bufferMode_e)value;
        return CCtxParams->outBufferMode;

    case ZSTD_c_blockDelimiters:
        BOUNDCHECK(ZSTD_c_blockDelimiters, value);
        CCtxParams->blockDelimiters = (ZSTD_SequenceFormat_e)value;
        return CCtxParams->blockDelimiters;

    case ZSTD_c_validateSequences:
        BOUNDCHECK(ZSTD_c_validateSequences, value);
        CCtxParams->validateSequences = value;
        return (size_t)CCtxParams->validateSequences;

    case ZSTD_c_splitAfterSequences:
        BOUNDCHECK(ZSTD_c_splitAfterSequences, value);
        CCtxParams->postBlockSplitter = (ZSTD_ParamSwitch_e)value;
        return CCtxParams->postBlockSplitter;

    case ZSTD_c_blockSplitterLevel:
        BOUNDCHECK(ZSTD_c_blockSplitterLevel, value);
        CCtxParams->preBlockSplitter_level = value;
        return (size_t)CCtxParams->preBlockSplitter_level;

    case ZSTD_c_useRowMatchFinder:
        BOUNDCHECK(ZSTD_c_useRowMatchFinder, value);
        CCtxParams->useRowMatchFinder = (ZSTD_ParamSwitch_e)value;
        return CCtxParams->useRowMatchFinder;

    case ZSTD_c_deterministicRefPrefix:
        BOUNDCHECK(ZSTD_c_deterministicRefPrefix, value);
        CCtxParams->deterministicRefPrefix = !!value;
        return (size_t)CCtxParams->deterministicRefPrefix;

    case ZSTD_c_prefetchCDictTables:
        BOUNDCHECK(ZSTD_c_prefetchCDictTables, value);
        CCtxParams->prefetchCDictTables = (ZSTD_ParamSwitch_e)value;
        return CCtxParams->prefetchCDictTables;

    case ZSTD_c_enableSeqProducerFallback:
        BOUNDCHECK(ZSTD_c_enableSeqProducerFallback, value);
        CCtxParams->enableMatchFinderFallback = value;
        return (size_t)CCtxParams->enableMatchFinderFallback;

    case ZSTD_c_maxBlockSize:
        if (value!=0)    /* 0 ==> default */
            BOUNDCHECK(ZSTD_c_maxBlockSize, value);
        assert(value>=0);
        CCtxParams->maxBlockSize = (size_t)value;
        return CCtxParams->maxBlockSize;

    case ZSTD_c_repcodeResolution:
        BOUNDCHECK(ZSTD_c_repcodeResolution, value);
        CCtxParams->searchForExternalRepcodes = (ZSTD_ParamSwitch_e)value;
        return CCtxParams->searchForExternalRepcodes;

    default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
    }
}

size_t ZSTD_CCtx_getParameter(ZSTD_CCtx const* cctx, ZSTD_cParameter param, int* value)
{
    return ZSTD_CCtxParams_getParameter(&cctx->requestedParams, param, value);
}

size_t ZSTD_CCtxParams_getParameter(
        ZSTD_CCtx_params const* CCtxParams, ZSTD_cParameter param, int* value)
{
    switch(param)
    {
    case ZSTD_c_format :
        *value = (int)CCtxParams->format;
        break;
    case ZSTD_c_compressionLevel :
        *value = CCtxParams->compressionLevel;
        break;
    case ZSTD_c_windowLog :
        *value = (int)CCtxParams->cParams.windowLog;
        break;
    case ZSTD_c_hashLog :
        *value = (int)CCtxParams->cParams.hashLog;
        break;
    case ZSTD_c_chainLog :
        *value = (int)CCtxParams->cParams.chainLog;
        break;
    case ZSTD_c_searchLog :
        *value = (int)CCtxParams->cParams.searchLog;
        break;
    case ZSTD_c_minMatch :
        *value = (int)CCtxParams->cParams.minMatch;
        break;
    case ZSTD_c_targetLength :
        *value = (int)CCtxParams->cParams.targetLength;
        break;
    case ZSTD_c_strategy :
        *value = (int)CCtxParams->cParams.strategy;
        break;
    case ZSTD_c_contentSizeFlag :
        *value = CCtxParams->fParams.contentSizeFlag;
        break;
    case ZSTD_c_checksumFlag :
        *value = CCtxParams->fParams.checksumFlag;
        break;
    case ZSTD_c_dictIDFlag :
        *value = !CCtxParams->fParams.noDictIDFlag;
        break;
    case ZSTD_c_forceMaxWindow :
        *value = CCtxParams->forceWindow;
        break;
    case ZSTD_c_forceAttachDict :
        *value = (int)CCtxParams->attachDictPref;
        break;
    case ZSTD_c_literalCompressionMode :
        *value = (int)CCtxParams->literalCompressionMode;
        break;
    case ZSTD_c_nbWorkers :
#ifndef ZSTD_MULTITHREAD
        assert(CCtxParams->nbWorkers == 0);
#endif
        *value = CCtxParams->nbWorkers;
        break;
    case ZSTD_c_jobSize :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR(parameter_unsupported, "not compiled with multithreading");
#else
        assert(CCtxParams->jobSize <= INT_MAX);
        *value = (int)CCtxParams->jobSize;
        break;
#endif
    case ZSTD_c_overlapLog :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR(parameter_unsupported, "not compiled with multithreading");
#else
        *value = CCtxParams->overlapLog;
        break;
#endif
    case ZSTD_c_rsyncable :
#ifndef ZSTD_MULTITHREAD
        RETURN_ERROR(parameter_unsupported, "not compiled with multithreading");
#else
        *value = CCtxParams->rsyncable;
        break;
#endif
    case ZSTD_c_enableDedicatedDictSearch :
        *value = CCtxParams->enableDedicatedDictSearch;
        break;
    case ZSTD_c_enableLongDistanceMatching :
        *value = (int)CCtxParams->ldmParams.enableLdm;
        break;
    case ZSTD_c_ldmHashLog :
        *value = (int)CCtxParams->ldmParams.hashLog;
        break;
    case ZSTD_c_ldmMinMatch :
        *value = (int)CCtxParams->ldmParams.minMatchLength;
        break;
    case ZSTD_c_ldmBucketSizeLog :
        *value = (int)CCtxParams->ldmParams.bucketSizeLog;
        break;
    case ZSTD_c_ldmHashRateLog :
        *value = (int)CCtxParams->ldmParams.hashRateLog;
        break;
    case ZSTD_c_targetCBlockSize :
        *value = (int)CCtxParams->targetCBlockSize;
        break;
    case ZSTD_c_srcSizeHint :
        *value = (int)CCtxParams->srcSizeHint;
        break;
    case ZSTD_c_stableInBuffer :
        *value = (int)CCtxParams->inBufferMode;
        break;
    case ZSTD_c_stableOutBuffer :
        *value = (int)CCtxParams->outBufferMode;
        break;
    case ZSTD_c_blockDelimiters :
        *value = (int)CCtxParams->blockDelimiters;
        break;
    case ZSTD_c_validateSequences :
        *value = (int)CCtxParams->validateSequences;
        break;
    case ZSTD_c_splitAfterSequences :
        *value = (int)CCtxParams->postBlockSplitter;
        break;
    case ZSTD_c_blockSplitterLevel :
        *value = CCtxParams->preBlockSplitter_level;
        break;
    case ZSTD_c_useRowMatchFinder :
        *value = (int)CCtxParams->useRowMatchFinder;
        break;
    case ZSTD_c_deterministicRefPrefix:
        *value = (int)CCtxParams->deterministicRefPrefix;
        break;
    case ZSTD_c_prefetchCDictTables:
        *value = (int)CCtxParams->prefetchCDictTables;
        break;
    case ZSTD_c_enableSeqProducerFallback:
        *value = CCtxParams->enableMatchFinderFallback;
        break;
    case ZSTD_c_maxBlockSize:
        *value = (int)CCtxParams->maxBlockSize;
        break;
    case ZSTD_c_repcodeResolution:
        *value = (int)CCtxParams->searchForExternalRepcodes;
        break;
    default: RETURN_ERROR(parameter_unsupported, "unknown parameter");
    }
    return 0;
}

/** ZSTD_CCtx_setParametersUsingCCtxParams() :
 *  just applies `params` into `cctx`
 *  no action is performed, parameters are merely stored.
 *  If ZSTDMT is enabled, parameters are pushed to cctx->mtctx.
 *    This is possible even if a compression is ongoing.
 *    In which case, new parameters will be applied on the fly, starting with next compression job.
 */
size_t ZSTD_CCtx_setParametersUsingCCtxParams(
        ZSTD_CCtx* cctx, const ZSTD_CCtx_params* params)
{
    DEBUGLOG(4, "ZSTD_CCtx_setParametersUsingCCtxParams");
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "The context is in the wrong stage!");
    RETURN_ERROR_IF(cctx->cdict, stage_wrong,
                    "Can't override parameters with cdict attached (some must "
                    "be inherited from the cdict).");

    cctx->requestedParams = *params;
    return 0;
}

size_t ZSTD_CCtx_setCParams(ZSTD_CCtx* cctx, ZSTD_compressionParameters cparams)
{
    ZSTD_STATIC_ASSERT(sizeof(cparams) == 7 * 4 /* all params are listed below */);
    DEBUGLOG(4, "ZSTD_CCtx_setCParams");
    /* only update if all parameters are valid */
    FORWARD_IF_ERROR(ZSTD_checkCParams(cparams), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_windowLog, (int)cparams.windowLog), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_chainLog, (int)cparams.chainLog), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_hashLog, (int)cparams.hashLog), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_searchLog, (int)cparams.searchLog), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_minMatch, (int)cparams.minMatch), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_targetLength, (int)cparams.targetLength), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_strategy, (int)cparams.strategy), "");
    return 0;
}

size_t ZSTD_CCtx_setFParams(ZSTD_CCtx* cctx, ZSTD_frameParameters fparams)
{
    ZSTD_STATIC_ASSERT(sizeof(fparams) == 3 * 4 /* all params are listed below */);
    DEBUGLOG(4, "ZSTD_CCtx_setFParams");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_contentSizeFlag, fparams.contentSizeFlag != 0), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_checksumFlag, fparams.checksumFlag != 0), "");
    FORWARD_IF_ERROR(ZSTD_CCtx_setParameter(cctx, ZSTD_c_dictIDFlag, fparams.noDictIDFlag == 0), "");
    return 0;
}

size_t ZSTD_CCtx_setParams(ZSTD_CCtx* cctx, ZSTD_parameters params)
{
    DEBUGLOG(4, "ZSTD_CCtx_setParams");
    /* First check cParams, because we want to update all or none. */
    FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
    /* Next set fParams, because this could fail if the cctx isn't in init stage. */
    FORWARD_IF_ERROR(ZSTD_CCtx_setFParams(cctx, params.fParams), "");
    /* Finally set cParams, which should succeed. */
    FORWARD_IF_ERROR(ZSTD_CCtx_setCParams(cctx, params.cParams), "");
    return 0;
}

size_t ZSTD_CCtx_setPledgedSrcSize(ZSTD_CCtx* cctx, unsigned long long pledgedSrcSize)
{
    DEBUGLOG(4, "ZSTD_CCtx_setPledgedSrcSize to %llu bytes", pledgedSrcSize);
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "Can't set pledgedSrcSize when not in init stage.");
    cctx->pledgedSrcSizePlusOne = pledgedSrcSize+1;
    return 0;
}

static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(
        int const compressionLevel,
        size_t const dictSize);
static int ZSTD_dedicatedDictSearch_isSupported(
        const ZSTD_compressionParameters* cParams);
static void ZSTD_dedicatedDictSearch_revertCParams(
        ZSTD_compressionParameters* cParams);

/**
 * Initializes the local dictionary using requested parameters.
 * NOTE: Initialization does not employ the pledged src size,
 * because the dictionary may be used for multiple compressions.
 */
static size_t ZSTD_initLocalDict(ZSTD_CCtx* cctx)
{
    ZSTD_localDict* const dl = &cctx->localDict;
    if (dl->dict == NULL) {
        /* No local dictionary. */
        assert(dl->dictBuffer == NULL);
        assert(dl->cdict == NULL);
        assert(dl->dictSize == 0);
        return 0;
    }
    if (dl->cdict != NULL) {
        /* Local dictionary already initialized. */
        assert(cctx->cdict == dl->cdict);
        return 0;
    }
    assert(dl->dictSize > 0);
    assert(cctx->cdict == NULL);
    assert(cctx->prefixDict.dict == NULL);

    dl->cdict = ZSTD_createCDict_advanced2(
            dl->dict,
            dl->dictSize,
            ZSTD_dlm_byRef,
            dl->dictContentType,
            &cctx->requestedParams,
            cctx->customMem);
    RETURN_ERROR_IF(!dl->cdict, memory_allocation, "ZSTD_createCDict_advanced failed");
    cctx->cdict = dl->cdict;
    return 0;
}

size_t ZSTD_CCtx_loadDictionary_advanced(
        ZSTD_CCtx* cctx,
        const void* dict, size_t dictSize,
        ZSTD_dictLoadMethod_e dictLoadMethod,
        ZSTD_dictContentType_e dictContentType)
{
    DEBUGLOG(4, "ZSTD_CCtx_loadDictionary_advanced (size: %u)", (U32)dictSize);
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "Can't load a dictionary when cctx is not in init stage.");
    ZSTD_clearAllDicts(cctx);  /* erase any previously set dictionary */
    if (dict == NULL || dictSize == 0)  /* no dictionary */
        return 0;
    if (dictLoadMethod == ZSTD_dlm_byRef) {
        cctx->localDict.dict = dict;
    } else {
        /* copy dictionary content inside CCtx to own its lifetime */
        void* dictBuffer;
        RETURN_ERROR_IF(cctx->staticSize, memory_allocation,
                        "static CCtx can't allocate for an internal copy of dictionary");
        dictBuffer = ZSTD_customMalloc(dictSize, cctx->customMem);
        RETURN_ERROR_IF(dictBuffer==NULL, memory_allocation,
                        "allocation failed for dictionary content");
        ZSTD_memcpy(dictBuffer, dict, dictSize);
        cctx->localDict.dictBuffer = dictBuffer;  /* owned ptr to free */
        cctx->localDict.dict = dictBuffer;        /* read-only reference */
    }
    cctx->localDict.dictSize = dictSize;
    cctx->localDict.dictContentType = dictContentType;
    return 0;
}

size_t ZSTD_CCtx_loadDictionary_byReference(
      ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
{
    return ZSTD_CCtx_loadDictionary_advanced(
            cctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto);
}

size_t ZSTD_CCtx_loadDictionary(ZSTD_CCtx* cctx, const void* dict, size_t dictSize)
{
    return ZSTD_CCtx_loadDictionary_advanced(
            cctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto);
}


size_t ZSTD_CCtx_refCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
{
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "Can't ref a dict when ctx not in init stage.");
    /* Free the existing local cdict (if any) to save memory. */
    ZSTD_clearAllDicts(cctx);
    cctx->cdict = cdict;
    return 0;
}

size_t ZSTD_CCtx_refThreadPool(ZSTD_CCtx* cctx, ZSTD_threadPool* pool)
{
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "Can't ref a pool when ctx not in init stage.");
    cctx->pool = pool;
    return 0;
}

size_t ZSTD_CCtx_refPrefix(ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize)
{
    return ZSTD_CCtx_refPrefix_advanced(cctx, prefix, prefixSize, ZSTD_dct_rawContent);
}

size_t ZSTD_CCtx_refPrefix_advanced(
        ZSTD_CCtx* cctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType)
{
    RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                    "Can't ref a prefix when ctx not in init stage.");
    ZSTD_clearAllDicts(cctx);
    if (prefix != NULL && prefixSize > 0) {
        cctx->prefixDict.dict = prefix;
        cctx->prefixDict.dictSize = prefixSize;
        cctx->prefixDict.dictContentType = dictContentType;
    }
    return 0;
}

/*! ZSTD_CCtx_reset() :
 *  Also dumps dictionary */
size_t ZSTD_CCtx_reset(ZSTD_CCtx* cctx, ZSTD_ResetDirective reset)
{
    if ( (reset == ZSTD_reset_session_only)
      || (reset == ZSTD_reset_session_and_parameters) ) {
        cctx->streamStage = zcss_init;
        cctx->pledgedSrcSizePlusOne = 0;
    }
    if ( (reset == ZSTD_reset_parameters)
      || (reset == ZSTD_reset_session_and_parameters) ) {
        RETURN_ERROR_IF(cctx->streamStage != zcss_init, stage_wrong,
                        "Reset parameters is only possible during init stage.");
        ZSTD_clearAllDicts(cctx);
        return ZSTD_CCtxParams_reset(&cctx->requestedParams);
    }
    return 0;
}


/** ZSTD_checkCParams() :
    control CParam values remain within authorized range.
    @return : 0, or an error code if one value is beyond authorized range */
size_t ZSTD_checkCParams(ZSTD_compressionParameters cParams)
{
    BOUNDCHECK(ZSTD_c_windowLog, (int)cParams.windowLog);
    BOUNDCHECK(ZSTD_c_chainLog,  (int)cParams.chainLog);
    BOUNDCHECK(ZSTD_c_hashLog,   (int)cParams.hashLog);
    BOUNDCHECK(ZSTD_c_searchLog, (int)cParams.searchLog);
    BOUNDCHECK(ZSTD_c_minMatch,  (int)cParams.minMatch);
    BOUNDCHECK(ZSTD_c_targetLength,(int)cParams.targetLength);
    BOUNDCHECK(ZSTD_c_strategy,  (int)cParams.strategy);
    return 0;
}

/** ZSTD_clampCParams() :
 *  make CParam values within valid range.
 *  @return : valid CParams */
static ZSTD_compressionParameters
ZSTD_clampCParams(ZSTD_compressionParameters cParams)
{
#   define CLAMP_TYPE(cParam, val, type)                                      \
        do {                                                                  \
            ZSTD_bounds const bounds = ZSTD_cParam_getBounds(cParam);         \
            if ((int)val<bounds.lowerBound) val=(type)bounds.lowerBound;      \
            else if ((int)val>bounds.upperBound) val=(type)bounds.upperBound; \
        } while (0)
#   define CLAMP(cParam, val) CLAMP_TYPE(cParam, val, unsigned)
    CLAMP(ZSTD_c_windowLog, cParams.windowLog);
    CLAMP(ZSTD_c_chainLog,  cParams.chainLog);
    CLAMP(ZSTD_c_hashLog,   cParams.hashLog);
    CLAMP(ZSTD_c_searchLog, cParams.searchLog);
    CLAMP(ZSTD_c_minMatch,  cParams.minMatch);
    CLAMP(ZSTD_c_targetLength,cParams.targetLength);
    CLAMP_TYPE(ZSTD_c_strategy,cParams.strategy, ZSTD_strategy);
    return cParams;
}

/** ZSTD_cycleLog() :
 *  condition for correct operation : hashLog > 1 */
U32 ZSTD_cycleLog(U32 hashLog, ZSTD_strategy strat)
{
    U32 const btScale = ((U32)strat >= (U32)ZSTD_btlazy2);
    return hashLog - btScale;
}

/** ZSTD_dictAndWindowLog() :
 * Returns an adjusted window log that is large enough to fit the source and the dictionary.
 * The zstd format says that the entire dictionary is valid if one byte of the dictionary
 * is within the window. So the hashLog and chainLog should be large enough to reference both
 * the dictionary and the window. So we must use this adjusted dictAndWindowLog when downsizing
 * the hashLog and windowLog.
 * NOTE: srcSize must not be ZSTD_CONTENTSIZE_UNKNOWN.
 */
static U32 ZSTD_dictAndWindowLog(U32 windowLog, U64 srcSize, U64 dictSize)
{
    const U64 maxWindowSize = 1ULL << ZSTD_WINDOWLOG_MAX;
    /* No dictionary ==> No change */
    if (dictSize == 0) {
        return windowLog;
    }
    assert(windowLog <= ZSTD_WINDOWLOG_MAX);
    assert(srcSize != ZSTD_CONTENTSIZE_UNKNOWN); /* Handled in ZSTD_adjustCParams_internal() */
    {
        U64 const windowSize = 1ULL << windowLog;
        U64 const dictAndWindowSize = dictSize + windowSize;
        /* If the window size is already large enough to fit both the source and the dictionary
         * then just use the window size. Otherwise adjust so that it fits the dictionary and
         * the window.
         */
        if (windowSize >= dictSize + srcSize) {
            return windowLog; /* Window size large enough already */
        } else if (dictAndWindowSize >= maxWindowSize) {
            return ZSTD_WINDOWLOG_MAX; /* Larger than max window log */
        } else  {
            return ZSTD_highbit32((U32)dictAndWindowSize - 1) + 1;
        }
    }
}

/** ZSTD_adjustCParams_internal() :
 *  optimize `cPar` for a specified input (`srcSize` and `dictSize`).
 *  mostly downsize to reduce memory consumption and initialization latency.
 * `srcSize` can be ZSTD_CONTENTSIZE_UNKNOWN when not known.
 * `mode` is the mode for parameter adjustment. See docs for `ZSTD_CParamMode_e`.
 *  note : `srcSize==0` means 0!
 *  condition : cPar is presumed validated (can be checked using ZSTD_checkCParams()). */
static ZSTD_compressionParameters
ZSTD_adjustCParams_internal(ZSTD_compressionParameters cPar,
                            unsigned long long srcSize,
                            size_t dictSize,
                            ZSTD_CParamMode_e mode,
                            ZSTD_ParamSwitch_e useRowMatchFinder)
{
    const U64 minSrcSize = 513; /* (1<<9) + 1 */
    const U64 maxWindowResize = 1ULL << (ZSTD_WINDOWLOG_MAX-1);
    assert(ZSTD_checkCParams(cPar)==0);

    /* Cascade the selected strategy down to the next-highest one built into
     * this binary. */
#ifdef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_btultra2) {
        cPar.strategy = ZSTD_btultra;
    }
    if (cPar.strategy == ZSTD_btultra) {
        cPar.strategy = ZSTD_btopt;
    }
#endif
#ifdef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_btopt) {
        cPar.strategy = ZSTD_btlazy2;
    }
#endif
#ifdef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_btlazy2) {
        cPar.strategy = ZSTD_lazy2;
    }
#endif
#ifdef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_lazy2) {
        cPar.strategy = ZSTD_lazy;
    }
#endif
#ifdef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_lazy) {
        cPar.strategy = ZSTD_greedy;
    }
#endif
#ifdef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_greedy) {
        cPar.strategy = ZSTD_dfast;
    }
#endif
#ifdef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
    if (cPar.strategy == ZSTD_dfast) {
        cPar.strategy = ZSTD_fast;
        cPar.targetLength = 0;
    }
#endif

    switch (mode) {
    case ZSTD_cpm_unknown:
    case ZSTD_cpm_noAttachDict:
        /* If we don't know the source size, don't make any
         * assumptions about it. We will already have selected
         * smaller parameters if a dictionary is in use.
         */
        break;
    case ZSTD_cpm_createCDict:
        /* Assume a small source size when creating a dictionary
         * with an unknown source size.
         */
        if (dictSize && srcSize == ZSTD_CONTENTSIZE_UNKNOWN)
            srcSize = minSrcSize;
        break;
    case ZSTD_cpm_attachDict:
        /* Dictionary has its own dedicated parameters which have
         * already been selected. We are selecting parameters
         * for only the source.
         */
        dictSize = 0;
        break;
    default:
        assert(0);
        break;
    }

    /* resize windowLog if input is small enough, to use less memory */
    if ( (srcSize <= maxWindowResize)
      && (dictSize <= maxWindowResize) )  {
        U32 const tSize = (U32)(srcSize + dictSize);
        static U32 const hashSizeMin = 1 << ZSTD_HASHLOG_MIN;
        U32 const srcLog = (tSize < hashSizeMin) ? ZSTD_HASHLOG_MIN :
                            ZSTD_highbit32(tSize-1) + 1;
        if (cPar.windowLog > srcLog) cPar.windowLog = srcLog;
    }
    if (srcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
        U32 const dictAndWindowLog = ZSTD_dictAndWindowLog(cPar.windowLog, (U64)srcSize, (U64)dictSize);
        U32 const cycleLog = ZSTD_cycleLog(cPar.chainLog, cPar.strategy);
        if (cPar.hashLog > dictAndWindowLog+1) cPar.hashLog = dictAndWindowLog+1;
        if (cycleLog > dictAndWindowLog)
            cPar.chainLog -= (cycleLog - dictAndWindowLog);
    }

    if (cPar.windowLog < ZSTD_WINDOWLOG_ABSOLUTEMIN)
        cPar.windowLog = ZSTD_WINDOWLOG_ABSOLUTEMIN;  /* minimum wlog required for valid frame header */

    /* We can't use more than 32 bits of hash in total, so that means that we require:
     * (hashLog + 8) <= 32 && (chainLog + 8) <= 32
     */
    if (mode == ZSTD_cpm_createCDict && ZSTD_CDictIndicesAreTagged(&cPar)) {
        U32 const maxShortCacheHashLog = 32 - ZSTD_SHORT_CACHE_TAG_BITS;
        if (cPar.hashLog > maxShortCacheHashLog) {
            cPar.hashLog = maxShortCacheHashLog;
        }
        if (cPar.chainLog > maxShortCacheHashLog) {
            cPar.chainLog = maxShortCacheHashLog;
        }
    }


    /* At this point, we aren't 100% sure if we are using the row match finder.
     * Unless it is explicitly disabled, conservatively assume that it is enabled.
     * In this case it will only be disabled for small sources, so shrinking the
     * hash log a little bit shouldn't result in any ratio loss.
     */
    if (useRowMatchFinder == ZSTD_ps_auto)
        useRowMatchFinder = ZSTD_ps_enable;

    /* We can't hash more than 32-bits in total. So that means that we require:
     * (hashLog - rowLog + 8) <= 32
     */
    if (ZSTD_rowMatchFinderUsed(cPar.strategy, useRowMatchFinder)) {
        /* Switch to 32-entry rows if searchLog is 5 (or more) */
        U32 const rowLog = BOUNDED(4, cPar.searchLog, 6);
        U32 const maxRowHashLog = 32 - ZSTD_ROW_HASH_TAG_BITS;
        U32 const maxHashLog = maxRowHashLog + rowLog;
        assert(cPar.hashLog >= rowLog);
        if (cPar.hashLog > maxHashLog) {
            cPar.hashLog = maxHashLog;
        }
    }

    return cPar;
}

ZSTD_compressionParameters
ZSTD_adjustCParams(ZSTD_compressionParameters cPar,
                   unsigned long long srcSize,
                   size_t dictSize)
{
    cPar = ZSTD_clampCParams(cPar);   /* resulting cPar is necessarily valid (all parameters within range) */
    if (srcSize == 0) srcSize = ZSTD_CONTENTSIZE_UNKNOWN;
    return ZSTD_adjustCParams_internal(cPar, srcSize, dictSize, ZSTD_cpm_unknown, ZSTD_ps_auto);
}

static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);
static ZSTD_parameters ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode);

static void ZSTD_overrideCParams(
              ZSTD_compressionParameters* cParams,
        const ZSTD_compressionParameters* overrides)
{
    if (overrides->windowLog)    cParams->windowLog    = overrides->windowLog;
    if (overrides->hashLog)      cParams->hashLog      = overrides->hashLog;
    if (overrides->chainLog)     cParams->chainLog     = overrides->chainLog;
    if (overrides->searchLog)    cParams->searchLog    = overrides->searchLog;
    if (overrides->minMatch)     cParams->minMatch     = overrides->minMatch;
    if (overrides->targetLength) cParams->targetLength = overrides->targetLength;
    if (overrides->strategy)     cParams->strategy     = overrides->strategy;
}

ZSTD_compressionParameters ZSTD_getCParamsFromCCtxParams(
        const ZSTD_CCtx_params* CCtxParams, U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
{
    ZSTD_compressionParameters cParams;
    if (srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN && CCtxParams->srcSizeHint > 0) {
        assert(CCtxParams->srcSizeHint>=0);
        srcSizeHint = (U64)CCtxParams->srcSizeHint;
    }
    cParams = ZSTD_getCParams_internal(CCtxParams->compressionLevel, srcSizeHint, dictSize, mode);
    if (CCtxParams->ldmParams.enableLdm == ZSTD_ps_enable) cParams.windowLog = ZSTD_LDM_DEFAULT_WINDOW_LOG;
    ZSTD_overrideCParams(&cParams, &CCtxParams->cParams);
    assert(!ZSTD_checkCParams(cParams));
    /* srcSizeHint == 0 means 0 */
    return ZSTD_adjustCParams_internal(cParams, srcSizeHint, dictSize, mode, CCtxParams->useRowMatchFinder);
}

static size_t
ZSTD_sizeof_matchState(const ZSTD_compressionParameters* const cParams,
                       const ZSTD_ParamSwitch_e useRowMatchFinder,
                       const int enableDedicatedDictSearch,
                       const U32 forCCtx)
{
    /* chain table size should be 0 for fast or row-hash strategies */
    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder, enableDedicatedDictSearch && !forCCtx)
                                ? ((size_t)1 << cParams->chainLog)
                                : 0;
    size_t const hSize = ((size_t)1) << cParams->hashLog;
    U32    const hashLog3 = (forCCtx && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
    size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;
    /* We don't use ZSTD_cwksp_alloc_size() here because the tables aren't
     * surrounded by redzones in ASAN. */
    size_t const tableSpace = chainSize * sizeof(U32)
                            + hSize * sizeof(U32)
                            + h3Size * sizeof(U32);
    size_t const optPotentialSpace =
        ZSTD_cwksp_aligned64_alloc_size((MaxML+1) * sizeof(U32))
      + ZSTD_cwksp_aligned64_alloc_size((MaxLL+1) * sizeof(U32))
      + ZSTD_cwksp_aligned64_alloc_size((MaxOff+1) * sizeof(U32))
      + ZSTD_cwksp_aligned64_alloc_size((1<<Litbits) * sizeof(U32))
      + ZSTD_cwksp_aligned64_alloc_size(ZSTD_OPT_SIZE * sizeof(ZSTD_match_t))
      + ZSTD_cwksp_aligned64_alloc_size(ZSTD_OPT_SIZE * sizeof(ZSTD_optimal_t));
    size_t const lazyAdditionalSpace = ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)
                                            ? ZSTD_cwksp_aligned64_alloc_size(hSize)
                                            : 0;
    size_t const optSpace = (forCCtx && (cParams->strategy >= ZSTD_btopt))
                                ? optPotentialSpace
                                : 0;
    size_t const slackSpace = ZSTD_cwksp_slack_space_required();

    /* tables are guaranteed to be sized in multiples of 64 bytes (or 16 uint32_t) */
    ZSTD_STATIC_ASSERT(ZSTD_HASHLOG_MIN >= 4 && ZSTD_WINDOWLOG_MIN >= 4 && ZSTD_CHAINLOG_MIN >= 4);
    assert(useRowMatchFinder != ZSTD_ps_auto);

    DEBUGLOG(4, "chainSize: %u - hSize: %u - h3Size: %u",
                (U32)chainSize, (U32)hSize, (U32)h3Size);
    return tableSpace + optSpace + slackSpace + lazyAdditionalSpace;
}

/* Helper function for calculating memory requirements.
 * Gives a tighter bound than ZSTD_sequenceBound() by taking minMatch into account. */
static size_t ZSTD_maxNbSeq(size_t blockSize, unsigned minMatch, int useSequenceProducer) {
    U32 const divider = (minMatch==3 || useSequenceProducer) ? 3 : 4;
    return blockSize / divider;
}

static size_t ZSTD_estimateCCtxSize_usingCCtxParams_internal(
        const ZSTD_compressionParameters* cParams,
        const ldmParams_t* ldmParams,
        const int isStatic,
        const ZSTD_ParamSwitch_e useRowMatchFinder,
        const size_t buffInSize,
        const size_t buffOutSize,
        const U64 pledgedSrcSize,
        int useSequenceProducer,
        size_t maxBlockSize)
{
    size_t const windowSize = (size_t) BOUNDED(1ULL, 1ULL << cParams->windowLog, pledgedSrcSize);
    size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(maxBlockSize), windowSize);
    size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, cParams->minMatch, useSequenceProducer);
    size_t const tokenSpace = ZSTD_cwksp_alloc_size(WILDCOPY_OVERLENGTH + blockSize)
                            + ZSTD_cwksp_aligned64_alloc_size(maxNbSeq * sizeof(SeqDef))
                            + 3 * ZSTD_cwksp_alloc_size(maxNbSeq * sizeof(BYTE));
    size_t const tmpWorkSpace = ZSTD_cwksp_alloc_size(TMP_WORKSPACE_SIZE);
    size_t const blockStateSpace = 2 * ZSTD_cwksp_alloc_size(sizeof(ZSTD_compressedBlockState_t));
    size_t const matchStateSize = ZSTD_sizeof_matchState(cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 0, /* forCCtx */ 1);

    size_t const ldmSpace = ZSTD_ldm_getTableSize(*ldmParams);
    size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(*ldmParams, blockSize);
    size_t const ldmSeqSpace = ldmParams->enableLdm == ZSTD_ps_enable ?
        ZSTD_cwksp_aligned64_alloc_size(maxNbLdmSeq * sizeof(rawSeq)) : 0;


    size_t const bufferSpace = ZSTD_cwksp_alloc_size(buffInSize)
                             + ZSTD_cwksp_alloc_size(buffOutSize);

    size_t const cctxSpace = isStatic ? ZSTD_cwksp_alloc_size(sizeof(ZSTD_CCtx)) : 0;

    size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
    size_t const externalSeqSpace = useSequenceProducer
        ? ZSTD_cwksp_aligned64_alloc_size(maxNbExternalSeq * sizeof(ZSTD_Sequence))
        : 0;

    size_t const neededSpace =
        cctxSpace +
        tmpWorkSpace +
        blockStateSpace +
        ldmSpace +
        ldmSeqSpace +
        matchStateSize +
        tokenSpace +
        bufferSpace +
        externalSeqSpace;

    DEBUGLOG(5, "estimate workspace : %u", (U32)neededSpace);
    return neededSpace;
}

size_t ZSTD_estimateCCtxSize_usingCCtxParams(const ZSTD_CCtx_params* params)
{
    ZSTD_compressionParameters const cParams =
                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
    ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder,
                                                                               &cParams);

    RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
    /* estimateCCtxSize is for one-shot compression. So no buffers should
     * be needed. However, we still allocate two 0-sized buffers, which can
     * take space under ASAN. */
    return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
        &cParams, &params->ldmParams, 1, useRowMatchFinder, 0, 0, ZSTD_CONTENTSIZE_UNKNOWN, ZSTD_hasExtSeqProd(params), params->maxBlockSize);
}

size_t ZSTD_estimateCCtxSize_usingCParams(ZSTD_compressionParameters cParams)
{
    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
        size_t noRowCCtxSize;
        size_t rowCCtxSize;
        initialParams.useRowMatchFinder = ZSTD_ps_disable;
        noRowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
        initialParams.useRowMatchFinder = ZSTD_ps_enable;
        rowCCtxSize = ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
        return MAX(noRowCCtxSize, rowCCtxSize);
    } else {
        return ZSTD_estimateCCtxSize_usingCCtxParams(&initialParams);
    }
}

static size_t ZSTD_estimateCCtxSize_internal(int compressionLevel)
{
    int tier = 0;
    size_t largestSize = 0;
    static const unsigned long long srcSizeTiers[4] = {16 KB, 128 KB, 256 KB, ZSTD_CONTENTSIZE_UNKNOWN};
    for (; tier < 4; ++tier) {
        /* Choose the set of cParams for a given level across all srcSizes that give the largest cctxSize */
        ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeTiers[tier], 0, ZSTD_cpm_noAttachDict);
        largestSize = MAX(ZSTD_estimateCCtxSize_usingCParams(cParams), largestSize);
    }
    return largestSize;
}

size_t ZSTD_estimateCCtxSize(int compressionLevel)
{
    int level;
    size_t memBudget = 0;
    for (level=MIN(compressionLevel, 1); level<=compressionLevel; level++) {
        /* Ensure monotonically increasing memory usage as compression level increases */
        size_t const newMB = ZSTD_estimateCCtxSize_internal(level);
        if (newMB > memBudget) memBudget = newMB;
    }
    return memBudget;
}

size_t ZSTD_estimateCStreamSize_usingCCtxParams(const ZSTD_CCtx_params* params)
{
    RETURN_ERROR_IF(params->nbWorkers > 0, GENERIC, "Estimate CCtx size is supported for single-threaded compression only.");
    {   ZSTD_compressionParameters const cParams =
                ZSTD_getCParamsFromCCtxParams(params, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
        size_t const blockSize = MIN(ZSTD_resolveMaxBlockSize(params->maxBlockSize), (size_t)1 << cParams.windowLog);
        size_t const inBuffSize = (params->inBufferMode == ZSTD_bm_buffered)
                ? ((size_t)1 << cParams.windowLog) + blockSize
                : 0;
        size_t const outBuffSize = (params->outBufferMode == ZSTD_bm_buffered)
                ? ZSTD_compressBound(blockSize) + 1
                : 0;
        ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params->useRowMatchFinder, &params->cParams);

        return ZSTD_estimateCCtxSize_usingCCtxParams_internal(
            &cParams, &params->ldmParams, 1, useRowMatchFinder, inBuffSize, outBuffSize,
            ZSTD_CONTENTSIZE_UNKNOWN, ZSTD_hasExtSeqProd(params), params->maxBlockSize);
    }
}

size_t ZSTD_estimateCStreamSize_usingCParams(ZSTD_compressionParameters cParams)
{
    ZSTD_CCtx_params initialParams = ZSTD_makeCCtxParamsFromCParams(cParams);
    if (ZSTD_rowMatchFinderSupported(cParams.strategy)) {
        /* Pick bigger of not using and using row-based matchfinder for greedy and lazy strategies */
        size_t noRowCCtxSize;
        size_t rowCCtxSize;
        initialParams.useRowMatchFinder = ZSTD_ps_disable;
        noRowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
        initialParams.useRowMatchFinder = ZSTD_ps_enable;
        rowCCtxSize = ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
        return MAX(noRowCCtxSize, rowCCtxSize);
    } else {
        return ZSTD_estimateCStreamSize_usingCCtxParams(&initialParams);
    }
}

static size_t ZSTD_estimateCStreamSize_internal(int compressionLevel)
{
    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
    return ZSTD_estimateCStreamSize_usingCParams(cParams);
}

size_t ZSTD_estimateCStreamSize(int compressionLevel)
{
    int level;
    size_t memBudget = 0;
    for (level=MIN(compressionLevel, 1); level<=compressionLevel; level++) {
        size_t const newMB = ZSTD_estimateCStreamSize_internal(level);
        if (newMB > memBudget) memBudget = newMB;
    }
    return memBudget;
}

/* ZSTD_getFrameProgression():
 * tells how much data has been consumed (input) and produced (output) for current frame.
 * able to count progression inside worker threads (non-blocking mode).
 */
ZSTD_frameProgression ZSTD_getFrameProgression(const ZSTD_CCtx* cctx)
{
#ifdef ZSTD_MULTITHREAD
    if (cctx->appliedParams.nbWorkers > 0) {
        return ZSTDMT_getFrameProgression(cctx->mtctx);
    }
#endif
    {   ZSTD_frameProgression fp;
        size_t const buffered = (cctx->inBuff == NULL) ? 0 :
                                cctx->inBuffPos - cctx->inToCompress;
        if (buffered) assert(cctx->inBuffPos >= cctx->inToCompress);
        assert(buffered <= ZSTD_BLOCKSIZE_MAX);
        fp.ingested = cctx->consumedSrcSize + buffered;
        fp.consumed = cctx->consumedSrcSize;
        fp.produced = cctx->producedCSize;
        fp.flushed  = cctx->producedCSize;   /* simplified; some data might still be left within streaming output buffer */
        fp.currentJobID = 0;
        fp.nbActiveWorkers = 0;
        return fp;
}   }

/*! ZSTD_toFlushNow()
 *  Only useful for multithreading scenarios currently (nbWorkers >= 1).
 */
size_t ZSTD_toFlushNow(ZSTD_CCtx* cctx)
{
#ifdef ZSTD_MULTITHREAD
    if (cctx->appliedParams.nbWorkers > 0) {
        return ZSTDMT_toFlushNow(cctx->mtctx);
    }
#endif
    (void)cctx;
    return 0;   /* over-simplification; could also check if context is currently running in streaming mode, and in which case, report how many bytes are left to be flushed within output buffer */
}

static void ZSTD_assertEqualCParams(ZSTD_compressionParameters cParams1,
                                    ZSTD_compressionParameters cParams2)
{
    (void)cParams1;
    (void)cParams2;
    assert(cParams1.windowLog    == cParams2.windowLog);
    assert(cParams1.chainLog     == cParams2.chainLog);
    assert(cParams1.hashLog      == cParams2.hashLog);
    assert(cParams1.searchLog    == cParams2.searchLog);
    assert(cParams1.minMatch     == cParams2.minMatch);
    assert(cParams1.targetLength == cParams2.targetLength);
    assert(cParams1.strategy     == cParams2.strategy);
}

void ZSTD_reset_compressedBlockState(ZSTD_compressedBlockState_t* bs)
{
    int i;
    for (i = 0; i < ZSTD_REP_NUM; ++i)
        bs->rep[i] = repStartValue[i];
    bs->entropy.huf.repeatMode = HUF_repeat_none;
    bs->entropy.fse.offcode_repeatMode = FSE_repeat_none;
    bs->entropy.fse.matchlength_repeatMode = FSE_repeat_none;
    bs->entropy.fse.litlength_repeatMode = FSE_repeat_none;
}

/*! ZSTD_invalidateMatchState()
 *  Invalidate all the matches in the match finder tables.
 *  Requires nextSrc and base to be set (can be NULL).
 */
static void ZSTD_invalidateMatchState(ZSTD_MatchState_t* ms)
{
    ZSTD_window_clear(&ms->window);

    ms->nextToUpdate = ms->window.dictLimit;
    ms->loadedDictEnd = 0;
    ms->opt.litLengthSum = 0;  /* force reset of btopt stats */
    ms->dictMatchState = NULL;
}

/**
 * Controls, for this matchState reset, whether the tables need to be cleared /
 * prepared for the coming compression (ZSTDcrp_makeClean), or whether the
 * tables can be left unclean (ZSTDcrp_leaveDirty), because we know that a
 * subsequent operation will overwrite the table space anyways (e.g., copying
 * the matchState contents in from a CDict).
 */
typedef enum {
    ZSTDcrp_makeClean,
    ZSTDcrp_leaveDirty
} ZSTD_compResetPolicy_e;

/**
 * Controls, for this matchState reset, whether indexing can continue where it
 * left off (ZSTDirp_continue), or whether it needs to be restarted from zero
 * (ZSTDirp_reset).
 */
typedef enum {
    ZSTDirp_continue,
    ZSTDirp_reset
} ZSTD_indexResetPolicy_e;

typedef enum {
    ZSTD_resetTarget_CDict,
    ZSTD_resetTarget_CCtx
} ZSTD_resetTarget_e;

/* Mixes bits in a 64 bits in a value, based on XXH3_rrmxmx */
static U64 ZSTD_bitmix(U64 val, U64 len) {
    val ^= ZSTD_rotateRight_U64(val, 49) ^ ZSTD_rotateRight_U64(val, 24);
    val *= 0x9FB21C651E98DF25ULL;
    val ^= (val >> 35) + len ;
    val *= 0x9FB21C651E98DF25ULL;
    return val ^ (val >> 28);
}

/* Mixes in the hashSalt and hashSaltEntropy to create a new hashSalt */
static void ZSTD_advanceHashSalt(ZSTD_MatchState_t* ms) {
    ms->hashSalt = ZSTD_bitmix(ms->hashSalt, 8) ^ ZSTD_bitmix((U64) ms->hashSaltEntropy, 4);
}

static size_t
ZSTD_reset_matchState(ZSTD_MatchState_t* ms,
                      ZSTD_cwksp* ws,
                const ZSTD_compressionParameters* cParams,
                const ZSTD_ParamSwitch_e useRowMatchFinder,
                const ZSTD_compResetPolicy_e crp,
                const ZSTD_indexResetPolicy_e forceResetIndex,
                const ZSTD_resetTarget_e forWho)
{
    /* disable chain table allocation for fast or row-based strategies */
    size_t const chainSize = ZSTD_allocateChainTable(cParams->strategy, useRowMatchFinder,
                                                     ms->dedicatedDictSearch && (forWho == ZSTD_resetTarget_CDict))
                                ? ((size_t)1 << cParams->chainLog)
                                : 0;
    size_t const hSize = ((size_t)1) << cParams->hashLog;
    U32    const hashLog3 = ((forWho == ZSTD_resetTarget_CCtx) && cParams->minMatch==3) ? MIN(ZSTD_HASHLOG3_MAX, cParams->windowLog) : 0;
    size_t const h3Size = hashLog3 ? ((size_t)1) << hashLog3 : 0;

    DEBUGLOG(4, "reset indices : %u", forceResetIndex == ZSTDirp_reset);
    assert(useRowMatchFinder != ZSTD_ps_auto);
    if (forceResetIndex == ZSTDirp_reset) {
        ZSTD_window_init(&ms->window);
        ZSTD_cwksp_mark_tables_dirty(ws);
    }

    ms->hashLog3 = hashLog3;
    ms->lazySkipping = 0;

    ZSTD_invalidateMatchState(ms);

    assert(!ZSTD_cwksp_reserve_failed(ws)); /* check that allocation hasn't already failed */

    ZSTD_cwksp_clear_tables(ws);

    DEBUGLOG(5, "reserving table space");
    /* table Space */
    ms->hashTable = (U32*)ZSTD_cwksp_reserve_table(ws, hSize * sizeof(U32));
    ms->chainTable = (U32*)ZSTD_cwksp_reserve_table(ws, chainSize * sizeof(U32));
    ms->hashTable3 = (U32*)ZSTD_cwksp_reserve_table(ws, h3Size * sizeof(U32));
    RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation,
                    "failed a workspace allocation in ZSTD_reset_matchState");

    DEBUGLOG(4, "reset table : %u", crp!=ZSTDcrp_leaveDirty);
    if (crp!=ZSTDcrp_leaveDirty) {
        /* reset tables only */
        ZSTD_cwksp_clean_tables(ws);
    }

    if (ZSTD_rowMatchFinderUsed(cParams->strategy, useRowMatchFinder)) {
        /* Row match finder needs an additional table of hashes ("tags") */
        size_t const tagTableSize = hSize;
        /* We want to generate a new salt in case we reset a Cctx, but we always want to use
         * 0 when we reset a Cdict */
        if(forWho == ZSTD_resetTarget_CCtx) {
            ms->tagTable = (BYTE*) ZSTD_cwksp_reserve_aligned_init_once(ws, tagTableSize);
            ZSTD_advanceHashSalt(ms);
        } else {
            /* When we are not salting we want to always memset the memory */
            ms->tagTable = (BYTE*) ZSTD_cwksp_reserve_aligned64(ws, tagTableSize);
            ZSTD_memset(ms->tagTable, 0, tagTableSize);
            ms->hashSalt = 0;
        }
        {   /* Switch to 32-entry rows if searchLog is 5 (or more) */
            U32 const rowLog = BOUNDED(4, cParams->searchLog, 6);
            assert(cParams->hashLog >= rowLog);
            ms->rowHashLog = cParams->hashLog - rowLog;
        }
    }

    /* opt parser space */
    if ((forWho == ZSTD_resetTarget_CCtx) && (cParams->strategy >= ZSTD_btopt)) {
        DEBUGLOG(4, "reserving optimal parser space");
        ms->opt.litFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (1<<Litbits) * sizeof(unsigned));
        ms->opt.litLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxLL+1) * sizeof(unsigned));
        ms->opt.matchLengthFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxML+1) * sizeof(unsigned));
        ms->opt.offCodeFreq = (unsigned*)ZSTD_cwksp_reserve_aligned64(ws, (MaxOff+1) * sizeof(unsigned));
        ms->opt.matchTable = (ZSTD_match_t*)ZSTD_cwksp_reserve_aligned64(ws, ZSTD_OPT_SIZE * sizeof(ZSTD_match_t));
        ms->opt.priceTable = (ZSTD_optimal_t*)ZSTD_cwksp_reserve_aligned64(ws, ZSTD_OPT_SIZE * sizeof(ZSTD_optimal_t));
    }

    ms->cParams = *cParams;

    RETURN_ERROR_IF(ZSTD_cwksp_reserve_failed(ws), memory_allocation,
                    "failed a workspace allocation in ZSTD_reset_matchState");
    return 0;
}

/* ZSTD_indexTooCloseToMax() :
 * minor optimization : prefer memset() rather than reduceIndex()
 * which is measurably slow in some circumstances (reported for Visual Studio).
 * Works when re-using a context for a lot of smallish inputs :
 * if all inputs are smaller than ZSTD_INDEXOVERFLOW_MARGIN,
 * memset() will be triggered before reduceIndex().
 */
#define ZSTD_INDEXOVERFLOW_MARGIN (16 MB)
static int ZSTD_indexTooCloseToMax(ZSTD_window_t w)
{
    return (size_t)(w.nextSrc - w.base) > (ZSTD_CURRENT_MAX - ZSTD_INDEXOVERFLOW_MARGIN);
}

/** ZSTD_dictTooBig():
 * When dictionaries are larger than ZSTD_CHUNKSIZE_MAX they can't be loaded in
 * one go generically. So we ensure that in that case we reset the tables to zero,
 * so that we can load as much of the dictionary as possible.
 */
static int ZSTD_dictTooBig(size_t const loadedDictSize)
{
    return loadedDictSize > ZSTD_CHUNKSIZE_MAX;
}

/*! ZSTD_resetCCtx_internal() :
 * @param loadedDictSize The size of the dictionary to be loaded
 * into the context, if any. If no dictionary is used, or the
 * dictionary is being attached / copied, then pass 0.
 * note : `params` are assumed fully validated at this stage.
 */
static size_t ZSTD_resetCCtx_internal(ZSTD_CCtx* zc,
                                      ZSTD_CCtx_params const* params,
                                      U64 const pledgedSrcSize,
                                      size_t const loadedDictSize,
                                      ZSTD_compResetPolicy_e const crp,
                                      ZSTD_buffered_policy_e const zbuff)
{
    ZSTD_cwksp* const ws = &zc->workspace;
    DEBUGLOG(4, "ZSTD_resetCCtx_internal: pledgedSrcSize=%u, wlog=%u, useRowMatchFinder=%d useBlockSplitter=%d",
                (U32)pledgedSrcSize, params->cParams.windowLog, (int)params->useRowMatchFinder, (int)params->postBlockSplitter);
    assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));

    zc->isFirstBlock = 1;

    /* Set applied params early so we can modify them for LDM,
     * and point params at the applied params.
     */
    zc->appliedParams = *params;
    params = &zc->appliedParams;

    assert(params->useRowMatchFinder != ZSTD_ps_auto);
    assert(params->postBlockSplitter != ZSTD_ps_auto);
    assert(params->ldmParams.enableLdm != ZSTD_ps_auto);
    assert(params->maxBlockSize != 0);
    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
        /* Adjust long distance matching parameters */
        ZSTD_ldm_adjustParameters(&zc->appliedParams.ldmParams, &params->cParams);
        assert(params->ldmParams.hashLog >= params->ldmParams.bucketSizeLog);
        assert(params->ldmParams.hashRateLog < 32);
    }

    {   size_t const windowSize = MAX(1, (size_t)MIN(((U64)1 << params->cParams.windowLog), pledgedSrcSize));
        size_t const blockSize = MIN(params->maxBlockSize, windowSize);
        size_t const maxNbSeq = ZSTD_maxNbSeq(blockSize, params->cParams.minMatch, ZSTD_hasExtSeqProd(params));
        size_t const buffOutSize = (zbuff == ZSTDb_buffered && params->outBufferMode == ZSTD_bm_buffered)
                ? ZSTD_compressBound(blockSize) + 1
                : 0;
        size_t const buffInSize = (zbuff == ZSTDb_buffered && params->inBufferMode == ZSTD_bm_buffered)
                ? windowSize + blockSize
                : 0;
        size_t const maxNbLdmSeq = ZSTD_ldm_getMaxNbSeq(params->ldmParams, blockSize);

        int const indexTooClose = ZSTD_indexTooCloseToMax(zc->blockState.matchState.window);
        int const dictTooBig = ZSTD_dictTooBig(loadedDictSize);
        ZSTD_indexResetPolicy_e needsIndexReset =
            (indexTooClose || dictTooBig || !zc->initialized) ? ZSTDirp_reset : ZSTDirp_continue;

        size_t const neededSpace =
            ZSTD_estimateCCtxSize_usingCCtxParams_internal(
                &params->cParams, &params->ldmParams, zc->staticSize != 0, params->useRowMatchFinder,
                buffInSize, buffOutSize, pledgedSrcSize, ZSTD_hasExtSeqProd(params), params->maxBlockSize);

        FORWARD_IF_ERROR(neededSpace, "cctx size estimate failed!");

        if (!zc->staticSize) ZSTD_cwksp_bump_oversized_duration(ws, 0);

        {   /* Check if workspace is large enough, alloc a new one if needed */
            int const workspaceTooSmall = ZSTD_cwksp_sizeof(ws) < neededSpace;
            int const workspaceWasteful = ZSTD_cwksp_check_wasteful(ws, neededSpace);
            int resizeWorkspace = workspaceTooSmall || workspaceWasteful;
            DEBUGLOG(4, "Need %zu B workspace", neededSpace);
            DEBUGLOG(4, "windowSize: %zu - blockSize: %zu", windowSize, blockSize);

            if (resizeWorkspace) {
                DEBUGLOG(4, "Resize workspaceSize from %zuKB to %zuKB",
                            ZSTD_cwksp_sizeof(ws) >> 10,
                            neededSpace >> 10);

                RETURN_ERROR_IF(zc->staticSize, memory_allocation, "static cctx : no resize");

                needsIndexReset = ZSTDirp_reset;

                ZSTD_cwksp_free(ws, zc->customMem);
                FORWARD_IF_ERROR(ZSTD_cwksp_create(ws, neededSpace, zc->customMem), "");

                DEBUGLOG(5, "reserving object space");
                /* Statically sized space.
                 * tmpWorkspace never moves,
                 * though prev/next block swap places */
                assert(ZSTD_cwksp_check_available(ws, 2 * sizeof(ZSTD_compressedBlockState_t)));
                zc->blockState.prevCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t));
                RETURN_ERROR_IF(zc->blockState.prevCBlock == NULL, memory_allocation, "couldn't allocate prevCBlock");
                zc->blockState.nextCBlock = (ZSTD_compressedBlockState_t*) ZSTD_cwksp_reserve_object(ws, sizeof(ZSTD_compressedBlockState_t));
                RETURN_ERROR_IF(zc->blockState.nextCBlock == NULL, memory_allocation, "couldn't allocate nextCBlock");
                zc->tmpWorkspace = ZSTD_cwksp_reserve_object(ws, TMP_WORKSPACE_SIZE);
                RETURN_ERROR_IF(zc->tmpWorkspace == NULL, memory_allocation, "couldn't allocate tmpWorkspace");
                zc->tmpWkspSize = TMP_WORKSPACE_SIZE;
        }   }

        ZSTD_cwksp_clear(ws);

        /* init params */
        zc->blockState.matchState.cParams = params->cParams;
        zc->blockState.matchState.prefetchCDictTables = params->prefetchCDictTables == ZSTD_ps_enable;
        zc->pledgedSrcSizePlusOne = pledgedSrcSize+1;
        zc->consumedSrcSize = 0;
        zc->producedCSize = 0;
        if (pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN)
            zc->appliedParams.fParams.contentSizeFlag = 0;
        DEBUGLOG(4, "pledged content size : %u ; flag : %u",
            (unsigned)pledgedSrcSize, zc->appliedParams.fParams.contentSizeFlag);
        zc->blockSizeMax = blockSize;

        XXH64_reset(&zc->xxhState, 0);
        zc->stage = ZSTDcs_init;
        zc->dictID = 0;
        zc->dictContentSize = 0;

        ZSTD_reset_compressedBlockState(zc->blockState.prevCBlock);

        FORWARD_IF_ERROR(ZSTD_reset_matchState(
                &zc->blockState.matchState,
                ws,
                &params->cParams,
                params->useRowMatchFinder,
                crp,
                needsIndexReset,
                ZSTD_resetTarget_CCtx), "");

        zc->seqStore.sequencesStart = (SeqDef*)ZSTD_cwksp_reserve_aligned64(ws, maxNbSeq * sizeof(SeqDef));

        /* ldm hash table */
        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
            /* TODO: avoid memset? */
            size_t const ldmHSize = ((size_t)1) << params->ldmParams.hashLog;
            zc->ldmState.hashTable = (ldmEntry_t*)ZSTD_cwksp_reserve_aligned64(ws, ldmHSize * sizeof(ldmEntry_t));
            ZSTD_memset(zc->ldmState.hashTable, 0, ldmHSize * sizeof(ldmEntry_t));
            zc->ldmSequences = (rawSeq*)ZSTD_cwksp_reserve_aligned64(ws, maxNbLdmSeq * sizeof(rawSeq));
            zc->maxNbLdmSequences = maxNbLdmSeq;

            ZSTD_window_init(&zc->ldmState.window);
            zc->ldmState.loadedDictEnd = 0;
        }

        /* reserve space for block-level external sequences */
        if (ZSTD_hasExtSeqProd(params)) {
            size_t const maxNbExternalSeq = ZSTD_sequenceBound(blockSize);
            zc->extSeqBufCapacity = maxNbExternalSeq;
            zc->extSeqBuf =
                (ZSTD_Sequence*)ZSTD_cwksp_reserve_aligned64(ws, maxNbExternalSeq * sizeof(ZSTD_Sequence));
        }

        /* buffers */

        /* ZSTD_wildcopy() is used to copy into the literals buffer,
         * so we have to oversize the buffer by WILDCOPY_OVERLENGTH bytes.
         */
        zc->seqStore.litStart = ZSTD_cwksp_reserve_buffer(ws, blockSize + WILDCOPY_OVERLENGTH);
        zc->seqStore.maxNbLit = blockSize;

        zc->bufferedPolicy = zbuff;
        zc->inBuffSize = buffInSize;
        zc->inBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffInSize);
        zc->outBuffSize = buffOutSize;
        zc->outBuff = (char*)ZSTD_cwksp_reserve_buffer(ws, buffOutSize);

        /* ldm bucketOffsets table */
        if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
            /* TODO: avoid memset? */
            size_t const numBuckets =
                  ((size_t)1) << (params->ldmParams.hashLog -
                                  params->ldmParams.bucketSizeLog);
            zc->ldmState.bucketOffsets = ZSTD_cwksp_reserve_buffer(ws, numBuckets);
            ZSTD_memset(zc->ldmState.bucketOffsets, 0, numBuckets);
        }

        /* sequences storage */
        ZSTD_referenceExternalSequences(zc, NULL, 0);
        zc->seqStore.maxNbSeq = maxNbSeq;
        zc->seqStore.llCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));
        zc->seqStore.mlCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));
        zc->seqStore.ofCode = ZSTD_cwksp_reserve_buffer(ws, maxNbSeq * sizeof(BYTE));

        DEBUGLOG(3, "wksp: finished allocating, %zd bytes remain available", ZSTD_cwksp_available_space(ws));
        assert(ZSTD_cwksp_estimated_space_within_bounds(ws, neededSpace));

        zc->initialized = 1;

        return 0;
    }
}

/* ZSTD_invalidateRepCodes() :
 * ensures next compression will not use repcodes from previous block.
 * Note : only works with regular variant;
 *        do not use with extDict variant ! */
void ZSTD_invalidateRepCodes(ZSTD_CCtx* cctx) {
    int i;
    for (i=0; i<ZSTD_REP_NUM; i++) cctx->blockState.prevCBlock->rep[i] = 0;
    assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
}

/* These are the approximate sizes for each strategy past which copying the
 * dictionary tables into the working context is faster than using them
 * in-place.
 */
static const size_t attachDictSizeCutoffs[ZSTD_STRATEGY_MAX+1] = {
    8 KB,  /* unused */
    8 KB,  /* ZSTD_fast */
    16 KB, /* ZSTD_dfast */
    32 KB, /* ZSTD_greedy */
    32 KB, /* ZSTD_lazy */
    32 KB, /* ZSTD_lazy2 */
    32 KB, /* ZSTD_btlazy2 */
    32 KB, /* ZSTD_btopt */
    8 KB,  /* ZSTD_btultra */
    8 KB   /* ZSTD_btultra2 */
};

static int ZSTD_shouldAttachDict(const ZSTD_CDict* cdict,
                                 const ZSTD_CCtx_params* params,
                                 U64 pledgedSrcSize)
{
    size_t cutoff = attachDictSizeCutoffs[cdict->matchState.cParams.strategy];
    int const dedicatedDictSearch = cdict->matchState.dedicatedDictSearch;
    return dedicatedDictSearch
        || ( ( pledgedSrcSize <= cutoff
            || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
            || params->attachDictPref == ZSTD_dictForceAttach )
          && params->attachDictPref != ZSTD_dictForceCopy
          && !params->forceWindow ); /* dictMatchState isn't correctly
                                      * handled in _enforceMaxDist */
}

static size_t
ZSTD_resetCCtx_byAttachingCDict(ZSTD_CCtx* cctx,
                        const ZSTD_CDict* cdict,
                        ZSTD_CCtx_params params,
                        U64 pledgedSrcSize,
                        ZSTD_buffered_policy_e zbuff)
{
    DEBUGLOG(4, "ZSTD_resetCCtx_byAttachingCDict() pledgedSrcSize=%llu",
                (unsigned long long)pledgedSrcSize);
    {
        ZSTD_compressionParameters adjusted_cdict_cParams = cdict->matchState.cParams;
        unsigned const windowLog = params.cParams.windowLog;
        assert(windowLog != 0);
        /* Resize working context table params for input only, since the dict
         * has its own tables. */
        /* pledgedSrcSize == 0 means 0! */

        if (cdict->matchState.dedicatedDictSearch) {
            ZSTD_dedicatedDictSearch_revertCParams(&adjusted_cdict_cParams);
        }

        params.cParams = ZSTD_adjustCParams_internal(adjusted_cdict_cParams, pledgedSrcSize,
                                                     cdict->dictContentSize, ZSTD_cpm_attachDict,
                                                     params.useRowMatchFinder);
        params.cParams.windowLog = windowLog;
        params.useRowMatchFinder = cdict->useRowMatchFinder;    /* cdict overrides */
        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
                                                 /* loadedDictSize */ 0,
                                                 ZSTDcrp_makeClean, zbuff), "");
        assert(cctx->appliedParams.cParams.strategy == adjusted_cdict_cParams.strategy);
    }

    {   const U32 cdictEnd = (U32)( cdict->matchState.window.nextSrc
                                  - cdict->matchState.window.base);
        const U32 cdictLen = cdictEnd - cdict->matchState.window.dictLimit;
        if (cdictLen == 0) {
            /* don't even attach dictionaries with no contents */
            DEBUGLOG(4, "skipping attaching empty dictionary");
        } else {
            DEBUGLOG(4, "attaching dictionary into context");
            cctx->blockState.matchState.dictMatchState = &cdict->matchState;

            /* prep working match state so dict matches never have negative indices
             * when they are translated to the working context's index space. */
            if (cctx->blockState.matchState.window.dictLimit < cdictEnd) {
                cctx->blockState.matchState.window.nextSrc =
                    cctx->blockState.matchState.window.base + cdictEnd;
                ZSTD_window_clear(&cctx->blockState.matchState.window);
            }
            /* loadedDictEnd is expressed within the referential of the active context */
            cctx->blockState.matchState.loadedDictEnd = cctx->blockState.matchState.window.dictLimit;
    }   }

    cctx->dictID = cdict->dictID;
    cctx->dictContentSize = cdict->dictContentSize;

    /* copy block state */
    ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));

    return 0;
}

static void ZSTD_copyCDictTableIntoCCtx(U32* dst, U32 const* src, size_t tableSize,
                                        ZSTD_compressionParameters const* cParams) {
    if (ZSTD_CDictIndicesAreTagged(cParams)){
        /* Remove tags from the CDict table if they are present.
         * See docs on "short cache" in zstd_compress_internal.h for context. */
        size_t i;
        for (i = 0; i < tableSize; i++) {
            U32 const taggedIndex = src[i];
            U32 const index = taggedIndex >> ZSTD_SHORT_CACHE_TAG_BITS;
            dst[i] = index;
        }
    } else {
        ZSTD_memcpy(dst, src, tableSize * sizeof(U32));
    }
}

static size_t ZSTD_resetCCtx_byCopyingCDict(ZSTD_CCtx* cctx,
                            const ZSTD_CDict* cdict,
                            ZSTD_CCtx_params params,
                            U64 pledgedSrcSize,
                            ZSTD_buffered_policy_e zbuff)
{
    const ZSTD_compressionParameters *cdict_cParams = &cdict->matchState.cParams;

    assert(!cdict->matchState.dedicatedDictSearch);
    DEBUGLOG(4, "ZSTD_resetCCtx_byCopyingCDict() pledgedSrcSize=%llu",
                (unsigned long long)pledgedSrcSize);

    {   unsigned const windowLog = params.cParams.windowLog;
        assert(windowLog != 0);
        /* Copy only compression parameters related to tables. */
        params.cParams = *cdict_cParams;
        params.cParams.windowLog = windowLog;
        params.useRowMatchFinder = cdict->useRowMatchFinder;
        FORWARD_IF_ERROR(ZSTD_resetCCtx_internal(cctx, &params, pledgedSrcSize,
                                                 /* loadedDictSize */ 0,
                                                 ZSTDcrp_leaveDirty, zbuff), "");
        assert(cctx->appliedParams.cParams.strategy == cdict_cParams->strategy);
        assert(cctx->appliedParams.cParams.hashLog == cdict_cParams->hashLog);
        assert(cctx->appliedParams.cParams.chainLog == cdict_cParams->chainLog);
    }

    ZSTD_cwksp_mark_tables_dirty(&cctx->workspace);
    assert(params.useRowMatchFinder != ZSTD_ps_auto);

    /* copy tables */
    {   size_t const chainSize = ZSTD_allocateChainTable(cdict_cParams->strategy, cdict->useRowMatchFinder, 0 /* DDS guaranteed disabled */)
                                                            ? ((size_t)1 << cdict_cParams->chainLog)
                                                            : 0;
        size_t const hSize =  (size_t)1 << cdict_cParams->hashLog;

        ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.hashTable,
                                cdict->matchState.hashTable,
                                hSize, cdict_cParams);

        /* Do not copy cdict's chainTable if cctx has parameters such that it would not use chainTable */
        if (ZSTD_allocateChainTable(cctx->appliedParams.cParams.strategy, cctx->appliedParams.useRowMatchFinder, 0 /* forDDSDict */)) {
            ZSTD_copyCDictTableIntoCCtx(cctx->blockState.matchState.chainTable,
                                    cdict->matchState.chainTable,
                                    chainSize, cdict_cParams);
        }
        /* copy tag table */
        if (ZSTD_rowMatchFinderUsed(cdict_cParams->strategy, cdict->useRowMatchFinder)) {
            size_t const tagTableSize = hSize;
            ZSTD_memcpy(cctx->blockState.matchState.tagTable,
                        cdict->matchState.tagTable,
                        tagTableSize);
            cctx->blockState.matchState.hashSalt = cdict->matchState.hashSalt;
        }
    }

    /* Zero the hashTable3, since the cdict never fills it */
    assert(cctx->blockState.matchState.hashLog3 <= 31);
    {   U32 const h3log = cctx->blockState.matchState.hashLog3;
        size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;
        assert(cdict->matchState.hashLog3 == 0);
        ZSTD_memset(cctx->blockState.matchState.hashTable3, 0, h3Size * sizeof(U32));
    }

    ZSTD_cwksp_mark_tables_clean(&cctx->workspace);

    /* copy dictionary offsets */
    {   ZSTD_MatchState_t const* srcMatchState = &cdict->matchState;
        ZSTD_MatchState_t* dstMatchState = &cctx->blockState.matchState;
        dstMatchState->window       = srcMatchState->window;
        dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
        dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
    }

    cctx->dictID = cdict->dictID;
    cctx->dictContentSize = cdict->dictContentSize;

    /* copy block state */
    ZSTD_memcpy(cctx->blockState.prevCBlock, &cdict->cBlockState, sizeof(cdict->cBlockState));

    return 0;
}

/* We have a choice between copying the dictionary context into the working
 * context, or referencing the dictionary context from the working context
 * in-place. We decide here which strategy to use. */
static size_t ZSTD_resetCCtx_usingCDict(ZSTD_CCtx* cctx,
                            const ZSTD_CDict* cdict,
                            const ZSTD_CCtx_params* params,
                            U64 pledgedSrcSize,
                            ZSTD_buffered_policy_e zbuff)
{

    DEBUGLOG(4, "ZSTD_resetCCtx_usingCDict (pledgedSrcSize=%u)",
                (unsigned)pledgedSrcSize);

    if (ZSTD_shouldAttachDict(cdict, params, pledgedSrcSize)) {
        return ZSTD_resetCCtx_byAttachingCDict(
            cctx, cdict, *params, pledgedSrcSize, zbuff);
    } else {
        return ZSTD_resetCCtx_byCopyingCDict(
            cctx, cdict, *params, pledgedSrcSize, zbuff);
    }
}

/*! ZSTD_copyCCtx_internal() :
 *  Duplicate an existing context `srcCCtx` into another one `dstCCtx`.
 *  Only works during stage ZSTDcs_init (i.e. after creation, but before first call to ZSTD_compressContinue()).
 *  The "context", in this case, refers to the hash and chain tables,
 *  entropy tables, and dictionary references.
 * `windowLog` value is enforced if != 0, otherwise value is copied from srcCCtx.
 * @return : 0, or an error code */
static size_t ZSTD_copyCCtx_internal(ZSTD_CCtx* dstCCtx,
                            const ZSTD_CCtx* srcCCtx,
                            ZSTD_frameParameters fParams,
                            U64 pledgedSrcSize,
                            ZSTD_buffered_policy_e zbuff)
{
    RETURN_ERROR_IF(srcCCtx->stage!=ZSTDcs_init, stage_wrong,
                    "Can't copy a ctx that's not in init stage.");
    DEBUGLOG(5, "ZSTD_copyCCtx_internal");
    ZSTD_memcpy(&dstCCtx->customMem, &srcCCtx->customMem, sizeof(ZSTD_customMem));
    {   ZSTD_CCtx_params params = dstCCtx->requestedParams;
        /* Copy only compression parameters related to tables. */
        params.cParams = srcCCtx->appliedParams.cParams;
        assert(srcCCtx->appliedParams.useRowMatchFinder != ZSTD_ps_auto);
        assert(srcCCtx->appliedParams.postBlockSplitter != ZSTD_ps_auto);
        assert(srcCCtx->appliedParams.ldmParams.enableLdm != ZSTD_ps_auto);
        params.useRowMatchFinder = srcCCtx->appliedParams.useRowMatchFinder;
        params.postBlockSplitter = srcCCtx->appliedParams.postBlockSplitter;
        params.ldmParams = srcCCtx->appliedParams.ldmParams;
        params.fParams = fParams;
        params.maxBlockSize = srcCCtx->appliedParams.maxBlockSize;
        ZSTD_resetCCtx_internal(dstCCtx, &params, pledgedSrcSize,
                                /* loadedDictSize */ 0,
                                ZSTDcrp_leaveDirty, zbuff);
        assert(dstCCtx->appliedParams.cParams.windowLog == srcCCtx->appliedParams.cParams.windowLog);
        assert(dstCCtx->appliedParams.cParams.strategy == srcCCtx->appliedParams.cParams.strategy);
        assert(dstCCtx->appliedParams.cParams.hashLog == srcCCtx->appliedParams.cParams.hashLog);
        assert(dstCCtx->appliedParams.cParams.chainLog == srcCCtx->appliedParams.cParams.chainLog);
        assert(dstCCtx->blockState.matchState.hashLog3 == srcCCtx->blockState.matchState.hashLog3);
    }

    ZSTD_cwksp_mark_tables_dirty(&dstCCtx->workspace);

    /* copy tables */
    {   size_t const chainSize = ZSTD_allocateChainTable(srcCCtx->appliedParams.cParams.strategy,
                                                         srcCCtx->appliedParams.useRowMatchFinder,
                                                         0 /* forDDSDict */)
                                    ? ((size_t)1 << srcCCtx->appliedParams.cParams.chainLog)
                                    : 0;
        size_t const hSize =  (size_t)1 << srcCCtx->appliedParams.cParams.hashLog;
        U32 const h3log = srcCCtx->blockState.matchState.hashLog3;
        size_t const h3Size = h3log ? ((size_t)1 << h3log) : 0;

        ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable,
               srcCCtx->blockState.matchState.hashTable,
               hSize * sizeof(U32));
        ZSTD_memcpy(dstCCtx->blockState.matchState.chainTable,
               srcCCtx->blockState.matchState.chainTable,
               chainSize * sizeof(U32));
        ZSTD_memcpy(dstCCtx->blockState.matchState.hashTable3,
               srcCCtx->blockState.matchState.hashTable3,
               h3Size * sizeof(U32));
    }

    ZSTD_cwksp_mark_tables_clean(&dstCCtx->workspace);

    /* copy dictionary offsets */
    {
        const ZSTD_MatchState_t* srcMatchState = &srcCCtx->blockState.matchState;
        ZSTD_MatchState_t* dstMatchState = &dstCCtx->blockState.matchState;
        dstMatchState->window       = srcMatchState->window;
        dstMatchState->nextToUpdate = srcMatchState->nextToUpdate;
        dstMatchState->loadedDictEnd= srcMatchState->loadedDictEnd;
    }
    dstCCtx->dictID = srcCCtx->dictID;
    dstCCtx->dictContentSize = srcCCtx->dictContentSize;

    /* copy block state */
    ZSTD_memcpy(dstCCtx->blockState.prevCBlock, srcCCtx->blockState.prevCBlock, sizeof(*srcCCtx->blockState.prevCBlock));

    return 0;
}

/*! ZSTD_copyCCtx() :
 *  Duplicate an existing context `srcCCtx` into another one `dstCCtx`.
 *  Only works during stage ZSTDcs_init (i.e. after creation, but before first call to ZSTD_compressContinue()).
 *  pledgedSrcSize==0 means "unknown".
*   @return : 0, or an error code */
size_t ZSTD_copyCCtx(ZSTD_CCtx* dstCCtx, const ZSTD_CCtx* srcCCtx, unsigned long long pledgedSrcSize)
{
    ZSTD_frameParameters fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
    ZSTD_buffered_policy_e const zbuff = srcCCtx->bufferedPolicy;
    ZSTD_STATIC_ASSERT((U32)ZSTDb_buffered==1);
    if (pledgedSrcSize==0) pledgedSrcSize = ZSTD_CONTENTSIZE_UNKNOWN;
    fParams.contentSizeFlag = (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN);

    return ZSTD_copyCCtx_internal(dstCCtx, srcCCtx,
                                fParams, pledgedSrcSize,
                                zbuff);
}


#define ZSTD_ROWSIZE 16
/*! ZSTD_reduceTable() :
 *  reduce table indexes by `reducerValue`, or squash to zero.
 *  PreserveMark preserves "unsorted mark" for btlazy2 strategy.
 *  It must be set to a clear 0/1 value, to remove branch during inlining.
 *  Presume table size is a multiple of ZSTD_ROWSIZE
 *  to help auto-vectorization */
FORCE_INLINE_TEMPLATE void
ZSTD_reduceTable_internal (U32* const table, U32 const size, U32 const reducerValue, int const preserveMark)
{
    int const nbRows = (int)size / ZSTD_ROWSIZE;
    int cellNb = 0;
    int rowNb;
    /* Protect special index values < ZSTD_WINDOW_START_INDEX. */
    U32 const reducerThreshold = reducerValue + ZSTD_WINDOW_START_INDEX;
    assert((size & (ZSTD_ROWSIZE-1)) == 0);  /* multiple of ZSTD_ROWSIZE */
    assert(size < (1U<<31));   /* can be cast to int */

#if ZSTD_MEMORY_SANITIZER && !defined (ZSTD_MSAN_DONT_POISON_WORKSPACE)
    /* To validate that the table reuse logic is sound, and that we don't
     * access table space that we haven't cleaned, we re-"poison" the table
     * space every time we mark it dirty.
     *
     * This function however is intended to operate on those dirty tables and
     * re-clean them. So when this function is used correctly, we can unpoison
     * the memory it operated on. This introduces a blind spot though, since
     * if we now try to operate on __actually__ poisoned memory, we will not
     * detect that. */
    __msan_unpoison(table, size * sizeof(U32));
#endif

    for (rowNb=0 ; rowNb < nbRows ; rowNb++) {
        int column;
        for (column=0; column<ZSTD_ROWSIZE; column++) {
            U32 newVal;
            if (preserveMark && table[cellNb] == ZSTD_DUBT_UNSORTED_MARK) {
                /* This write is pointless, but is required(?) for the compiler
                 * to auto-vectorize the loop. */
                newVal = ZSTD_DUBT_UNSORTED_MARK;
            } else if (table[cellNb] < reducerThreshold) {
                newVal = 0;
            } else {
                newVal = table[cellNb] - reducerValue;
            }
            table[cellNb] = newVal;
            cellNb++;
    }   }
}

static void ZSTD_reduceTable(U32* const table, U32 const size, U32 const reducerValue)
{
    ZSTD_reduceTable_internal(table, size, reducerValue, 0);
}

static void ZSTD_reduceTable_btlazy2(U32* const table, U32 const size, U32 const reducerValue)
{
    ZSTD_reduceTable_internal(table, size, reducerValue, 1);
}

/*! ZSTD_reduceIndex() :
*   rescale all indexes to avoid future overflow (indexes are U32) */
static void ZSTD_reduceIndex (ZSTD_MatchState_t* ms, ZSTD_CCtx_params const* params, const U32 reducerValue)
{
    {   U32 const hSize = (U32)1 << params->cParams.hashLog;
        ZSTD_reduceTable(ms->hashTable, hSize, reducerValue);
    }

    if (ZSTD_allocateChainTable(params->cParams.strategy, params->useRowMatchFinder, (U32)ms->dedicatedDictSearch)) {
        U32 const chainSize = (U32)1 << params->cParams.chainLog;
        if (params->cParams.strategy == ZSTD_btlazy2)
            ZSTD_reduceTable_btlazy2(ms->chainTable, chainSize, reducerValue);
        else
            ZSTD_reduceTable(ms->chainTable, chainSize, reducerValue);
    }

    if (ms->hashLog3) {
        U32 const h3Size = (U32)1 << ms->hashLog3;
        ZSTD_reduceTable(ms->hashTable3, h3Size, reducerValue);
    }
}


/*-*******************************************************
*  Block entropic compression
*********************************************************/

/* See doc/zstd_compression_format.md for detailed format description */

int ZSTD_seqToCodes(const SeqStore_t* seqStorePtr)
{
    const SeqDef* const sequences = seqStorePtr->sequencesStart;
    BYTE* const llCodeTable = seqStorePtr->llCode;
    BYTE* const ofCodeTable = seqStorePtr->ofCode;
    BYTE* const mlCodeTable = seqStorePtr->mlCode;
    U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
    U32 u;
    int longOffsets = 0;
    assert(nbSeq <= seqStorePtr->maxNbSeq);
    for (u=0; u<nbSeq; u++) {
        U32 const llv = sequences[u].litLength;
        U32 const ofCode = ZSTD_highbit32(sequences[u].offBase);
        U32 const mlv = sequences[u].mlBase;
        llCodeTable[u] = (BYTE)ZSTD_LLcode(llv);
        ofCodeTable[u] = (BYTE)ofCode;
        mlCodeTable[u] = (BYTE)ZSTD_MLcode(mlv);
        assert(!(MEM_64bits() && ofCode >= STREAM_ACCUMULATOR_MIN));
        if (MEM_32bits() && ofCode >= STREAM_ACCUMULATOR_MIN)
            longOffsets = 1;
    }
    if (seqStorePtr->longLengthType==ZSTD_llt_literalLength)
        llCodeTable[seqStorePtr->longLengthPos] = MaxLL;
    if (seqStorePtr->longLengthType==ZSTD_llt_matchLength)
        mlCodeTable[seqStorePtr->longLengthPos] = MaxML;
    return longOffsets;
}

/* ZSTD_useTargetCBlockSize():
 * Returns if target compressed block size param is being used.
 * If used, compression will do best effort to make a compressed block size to be around targetCBlockSize.
 * Returns 1 if true, 0 otherwise. */
static int ZSTD_useTargetCBlockSize(const ZSTD_CCtx_params* cctxParams)
{
    DEBUGLOG(5, "ZSTD_useTargetCBlockSize (targetCBlockSize=%zu)", cctxParams->targetCBlockSize);
    return (cctxParams->targetCBlockSize != 0);
}

/* ZSTD_blockSplitterEnabled():
 * Returns if block splitting param is being used
 * If used, compression will do best effort to split a block in order to improve compression ratio.
 * At the time this function is called, the parameter must be finalized.
 * Returns 1 if true, 0 otherwise. */
static int ZSTD_blockSplitterEnabled(ZSTD_CCtx_params* cctxParams)
{
    DEBUGLOG(5, "ZSTD_blockSplitterEnabled (postBlockSplitter=%d)", cctxParams->postBlockSplitter);
    assert(cctxParams->postBlockSplitter != ZSTD_ps_auto);
    return (cctxParams->postBlockSplitter == ZSTD_ps_enable);
}

/* Type returned by ZSTD_buildSequencesStatistics containing finalized symbol encoding types
 * and size of the sequences statistics
 */
typedef struct {
    U32 LLtype;
    U32 Offtype;
    U32 MLtype;
    size_t size;
    size_t lastCountSize; /* Accounts for bug in 1.3.4. More detail in ZSTD_entropyCompressSeqStore_internal() */
    int longOffsets;
} ZSTD_symbolEncodingTypeStats_t;

/* ZSTD_buildSequencesStatistics():
 * Returns a ZSTD_symbolEncodingTypeStats_t, or a zstd error code in the `size` field.
 * Modifies `nextEntropy` to have the appropriate values as a side effect.
 * nbSeq must be greater than 0.
 *
 * entropyWkspSize must be of size at least ENTROPY_WORKSPACE_SIZE - (MaxSeq + 1)*sizeof(U32)
 */
static ZSTD_symbolEncodingTypeStats_t
ZSTD_buildSequencesStatistics(
                const SeqStore_t* seqStorePtr, size_t nbSeq,
                const ZSTD_fseCTables_t* prevEntropy, ZSTD_fseCTables_t* nextEntropy,
                      BYTE* dst, const BYTE* const dstEnd,
                      ZSTD_strategy strategy, unsigned* countWorkspace,
                      void* entropyWorkspace, size_t entropyWkspSize)
{
    BYTE* const ostart = dst;
    const BYTE* const oend = dstEnd;
    BYTE* op = ostart;
    FSE_CTable* CTable_LitLength = nextEntropy->litlengthCTable;
    FSE_CTable* CTable_OffsetBits = nextEntropy->offcodeCTable;
    FSE_CTable* CTable_MatchLength = nextEntropy->matchlengthCTable;
    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
    const BYTE* const llCodeTable = seqStorePtr->llCode;
    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
    ZSTD_symbolEncodingTypeStats_t stats;

    stats.lastCountSize = 0;
    /* convert length/distances into codes */
    stats.longOffsets = ZSTD_seqToCodes(seqStorePtr);
    assert(op <= oend);
    assert(nbSeq != 0); /* ZSTD_selectEncodingType() divides by nbSeq */
    /* build CTable for Literal Lengths */
    {   unsigned max = MaxLL;
        size_t const mostFrequent = HIST_countFast_wksp(countWorkspace, &max, llCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
        DEBUGLOG(5, "Building LL table");
        nextEntropy->litlength_repeatMode = prevEntropy->litlength_repeatMode;
        stats.LLtype = ZSTD_selectEncodingType(&nextEntropy->litlength_repeatMode,
                                        countWorkspace, max, mostFrequent, nbSeq,
                                        LLFSELog, prevEntropy->litlengthCTable,
                                        LL_defaultNorm, LL_defaultNormLog,
                                        ZSTD_defaultAllowed, strategy);
        assert(set_basic < set_compressed && set_rle < set_compressed);
        assert(!(stats.LLtype < set_compressed && nextEntropy->litlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
        {   size_t const countSize = ZSTD_buildCTable(
                op, (size_t)(oend - op),
                CTable_LitLength, LLFSELog, (SymbolEncodingType_e)stats.LLtype,
                countWorkspace, max, llCodeTable, nbSeq,
                LL_defaultNorm, LL_defaultNormLog, MaxLL,
                prevEntropy->litlengthCTable,
                sizeof(prevEntropy->litlengthCTable),
                entropyWorkspace, entropyWkspSize);
            if (ZSTD_isError(countSize)) {
                DEBUGLOG(3, "ZSTD_buildCTable for LitLens failed");
                stats.size = countSize;
                return stats;
            }
            if (stats.LLtype == set_compressed)
                stats.lastCountSize = countSize;
            op += countSize;
            assert(op <= oend);
    }   }
    /* build CTable for Offsets */
    {   unsigned max = MaxOff;
        size_t const mostFrequent = HIST_countFast_wksp(
            countWorkspace, &max, ofCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);  /* can't fail */
        /* We can only use the basic table if max <= DefaultMaxOff, otherwise the offsets are too large */
        ZSTD_DefaultPolicy_e const defaultPolicy = (max <= DefaultMaxOff) ? ZSTD_defaultAllowed : ZSTD_defaultDisallowed;
        DEBUGLOG(5, "Building OF table");
        nextEntropy->offcode_repeatMode = prevEntropy->offcode_repeatMode;
        stats.Offtype = ZSTD_selectEncodingType(&nextEntropy->offcode_repeatMode,
                                        countWorkspace, max, mostFrequent, nbSeq,
                                        OffFSELog, prevEntropy->offcodeCTable,
                                        OF_defaultNorm, OF_defaultNormLog,
                                        defaultPolicy, strategy);
        assert(!(stats.Offtype < set_compressed && nextEntropy->offcode_repeatMode != FSE_repeat_none)); /* We don't copy tables */
        {   size_t const countSize = ZSTD_buildCTable(
                op, (size_t)(oend - op),
                CTable_OffsetBits, OffFSELog, (SymbolEncodingType_e)stats.Offtype,
                countWorkspace, max, ofCodeTable, nbSeq,
                OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                prevEntropy->offcodeCTable,
                sizeof(prevEntropy->offcodeCTable),
                entropyWorkspace, entropyWkspSize);
            if (ZSTD_isError(countSize)) {
                DEBUGLOG(3, "ZSTD_buildCTable for Offsets failed");
                stats.size = countSize;
                return stats;
            }
            if (stats.Offtype == set_compressed)
                stats.lastCountSize = countSize;
            op += countSize;
            assert(op <= oend);
    }   }
    /* build CTable for MatchLengths */
    {   unsigned max = MaxML;
        size_t const mostFrequent = HIST_countFast_wksp(
            countWorkspace, &max, mlCodeTable, nbSeq, entropyWorkspace, entropyWkspSize);   /* can't fail */
        DEBUGLOG(5, "Building ML table (remaining space : %i)", (int)(oend-op));
        nextEntropy->matchlength_repeatMode = prevEntropy->matchlength_repeatMode;
        stats.MLtype = ZSTD_selectEncodingType(&nextEntropy->matchlength_repeatMode,
                                        countWorkspace, max, mostFrequent, nbSeq,
                                        MLFSELog, prevEntropy->matchlengthCTable,
                                        ML_defaultNorm, ML_defaultNormLog,
                                        ZSTD_defaultAllowed, strategy);
        assert(!(stats.MLtype < set_compressed && nextEntropy->matchlength_repeatMode != FSE_repeat_none)); /* We don't copy tables */
        {   size_t const countSize = ZSTD_buildCTable(
                op, (size_t)(oend - op),
                CTable_MatchLength, MLFSELog, (SymbolEncodingType_e)stats.MLtype,
                countWorkspace, max, mlCodeTable, nbSeq,
                ML_defaultNorm, ML_defaultNormLog, MaxML,
                prevEntropy->matchlengthCTable,
                sizeof(prevEntropy->matchlengthCTable),
                entropyWorkspace, entropyWkspSize);
            if (ZSTD_isError(countSize)) {
                DEBUGLOG(3, "ZSTD_buildCTable for MatchLengths failed");
                stats.size = countSize;
                return stats;
            }
            if (stats.MLtype == set_compressed)
                stats.lastCountSize = countSize;
            op += countSize;
            assert(op <= oend);
    }   }
    stats.size = (size_t)(op-ostart);
    return stats;
}

/* ZSTD_entropyCompressSeqStore_internal():
 * compresses both literals and sequences
 * Returns compressed size of block, or a zstd error.
 */
#define SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO 20
MEM_STATIC size_t
ZSTD_entropyCompressSeqStore_internal(
                              void* dst, size_t dstCapacity,
                        const void* literals, size_t litSize,
                        const SeqStore_t* seqStorePtr,
                        const ZSTD_entropyCTables_t* prevEntropy,
                              ZSTD_entropyCTables_t* nextEntropy,
                        const ZSTD_CCtx_params* cctxParams,
                              void* entropyWorkspace, size_t entropyWkspSize,
                        const int bmi2)
{
    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
    unsigned* count = (unsigned*)entropyWorkspace;
    FSE_CTable* CTable_LitLength = nextEntropy->fse.litlengthCTable;
    FSE_CTable* CTable_OffsetBits = nextEntropy->fse.offcodeCTable;
    FSE_CTable* CTable_MatchLength = nextEntropy->fse.matchlengthCTable;
    const SeqDef* const sequences = seqStorePtr->sequencesStart;
    const size_t nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
    const BYTE* const ofCodeTable = seqStorePtr->ofCode;
    const BYTE* const llCodeTable = seqStorePtr->llCode;
    const BYTE* const mlCodeTable = seqStorePtr->mlCode;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ostart + dstCapacity;
    BYTE* op = ostart;
    size_t lastCountSize;
    int longOffsets = 0;

    entropyWorkspace = count + (MaxSeq + 1);
    entropyWkspSize -= (MaxSeq + 1) * sizeof(*count);

    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore_internal (nbSeq=%zu, dstCapacity=%zu)", nbSeq, dstCapacity);
    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
    assert(entropyWkspSize >= HUF_WORKSPACE_SIZE);

    /* Compress literals */
    {   size_t const numSequences = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
        /* Base suspicion of uncompressibility on ratio of literals to sequences */
        int const suspectUncompressible = (numSequences == 0) || (litSize / numSequences >= SUSPECT_UNCOMPRESSIBLE_LITERAL_RATIO);

        size_t const cSize = ZSTD_compressLiterals(
                                    op, dstCapacity,
                                    literals, litSize,
                                    entropyWorkspace, entropyWkspSize,
                                    &prevEntropy->huf, &nextEntropy->huf,
                                    cctxParams->cParams.strategy,
                                    ZSTD_literalsCompressionIsDisabled(cctxParams),
                                    suspectUncompressible, bmi2);
        FORWARD_IF_ERROR(cSize, "ZSTD_compressLiterals failed");
        assert(cSize <= dstCapacity);
        op += cSize;
    }

    /* Sequences Header */
    RETURN_ERROR_IF((oend-op) < 3 /*max nbSeq Size*/ + 1 /*seqHead*/,
                    dstSize_tooSmall, "Can't fit seq hdr in output buf!");
    if (nbSeq < 128) {
        *op++ = (BYTE)nbSeq;
    } else if (nbSeq < LONGNBSEQ) {
        op[0] = (BYTE)((nbSeq>>8) + 0x80);
        op[1] = (BYTE)nbSeq;
        op+=2;
    } else {
        op[0]=0xFF;
        MEM_writeLE16(op+1, (U16)(nbSeq - LONGNBSEQ));
        op+=3;
    }
    assert(op <= oend);
    if (nbSeq==0) {
        /* Copy the old tables over as if we repeated them */
        ZSTD_memcpy(&nextEntropy->fse, &prevEntropy->fse, sizeof(prevEntropy->fse));
        return (size_t)(op - ostart);
    }
    {   BYTE* const seqHead = op++;
        /* build stats for sequences */
        const ZSTD_symbolEncodingTypeStats_t stats =
                ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
                                             &prevEntropy->fse, &nextEntropy->fse,
                                              op, oend,
                                              strategy, count,
                                              entropyWorkspace, entropyWkspSize);
        FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
        *seqHead = (BYTE)((stats.LLtype<<6) + (stats.Offtype<<4) + (stats.MLtype<<2));
        lastCountSize = stats.lastCountSize;
        op += stats.size;
        longOffsets = stats.longOffsets;
    }

    {   size_t const bitstreamSize = ZSTD_encodeSequences(
                                        op, (size_t)(oend - op),
                                        CTable_MatchLength, mlCodeTable,
                                        CTable_OffsetBits, ofCodeTable,
                                        CTable_LitLength, llCodeTable,
                                        sequences, nbSeq,
                                        longOffsets, bmi2);
        FORWARD_IF_ERROR(bitstreamSize, "ZSTD_encodeSequences failed");
        op += bitstreamSize;
        assert(op <= oend);
        /* zstd versions <= 1.3.4 mistakenly report corruption when
         * FSE_readNCount() receives a buffer < 4 bytes.
         * Fixed by https://github.com/facebook/zstd/pull/1146.
         * This can happen when the last set_compressed table present is 2
         * bytes and the bitstream is only one byte.
         * In this exceedingly rare case, we will simply emit an uncompressed
         * block, since it isn't worth optimizing.
         */
        if (lastCountSize && (lastCountSize + bitstreamSize) < 4) {
            /* lastCountSize >= 2 && bitstreamSize > 0 ==> lastCountSize == 3 */
            assert(lastCountSize + bitstreamSize == 3);
            DEBUGLOG(5, "Avoiding bug in zstd decoder in versions <= 1.3.4 by "
                        "emitting an uncompressed block.");
            return 0;
        }
    }

    DEBUGLOG(5, "compressed block size : %u", (unsigned)(op - ostart));
    return (size_t)(op - ostart);
}

static size_t
ZSTD_entropyCompressSeqStore_wExtLitBuffer(
                          void* dst, size_t dstCapacity,
                    const void* literals, size_t litSize,
                          size_t blockSize,
                    const SeqStore_t* seqStorePtr,
                    const ZSTD_entropyCTables_t* prevEntropy,
                          ZSTD_entropyCTables_t* nextEntropy,
                    const ZSTD_CCtx_params* cctxParams,
                          void* entropyWorkspace, size_t entropyWkspSize,
                          int bmi2)
{
    size_t const cSize = ZSTD_entropyCompressSeqStore_internal(
                            dst, dstCapacity,
                            literals, litSize,
                            seqStorePtr, prevEntropy, nextEntropy, cctxParams,
                            entropyWorkspace, entropyWkspSize, bmi2);
    if (cSize == 0) return 0;
    /* When srcSize <= dstCapacity, there is enough space to write a raw uncompressed block.
     * Since we ran out of space, block must be not compressible, so fall back to raw uncompressed block.
     */
    if ((cSize == ERROR(dstSize_tooSmall)) & (blockSize <= dstCapacity)) {
        DEBUGLOG(4, "not enough dstCapacity (%zu) for ZSTD_entropyCompressSeqStore_internal()=> do not compress block", dstCapacity);
        return 0;  /* block not compressed */
    }
    FORWARD_IF_ERROR(cSize, "ZSTD_entropyCompressSeqStore_internal failed");

    /* Check compressibility */
    {   size_t const maxCSize = blockSize - ZSTD_minGain(blockSize, cctxParams->cParams.strategy);
        if (cSize >= maxCSize) return 0;  /* block not compressed */
    }
    DEBUGLOG(5, "ZSTD_entropyCompressSeqStore() cSize: %zu", cSize);
    /* libzstd decoder before  > v1.5.4 is not compatible with compressed blocks of size ZSTD_BLOCKSIZE_MAX exactly.
     * This restriction is indirectly already fulfilled by respecting ZSTD_minGain() condition above.
     */
    assert(cSize < ZSTD_BLOCKSIZE_MAX);
    return cSize;
}

static size_t
ZSTD_entropyCompressSeqStore(
                    const SeqStore_t* seqStorePtr,
                    const ZSTD_entropyCTables_t* prevEntropy,
                          ZSTD_entropyCTables_t* nextEntropy,
                    const ZSTD_CCtx_params* cctxParams,
                          void* dst, size_t dstCapacity,
                          size_t srcSize,
                          void* entropyWorkspace, size_t entropyWkspSize,
                          int bmi2)
{
    return ZSTD_entropyCompressSeqStore_wExtLitBuffer(
                dst, dstCapacity,
                seqStorePtr->litStart, (size_t)(seqStorePtr->lit - seqStorePtr->litStart),
                srcSize,
                seqStorePtr,
                prevEntropy, nextEntropy,
                cctxParams,
                entropyWorkspace, entropyWkspSize,
                bmi2);
}

/* ZSTD_selectBlockCompressor() :
 * Not static, but internal use only (used by long distance matcher)
 * assumption : strat is a valid strategy */
ZSTD_BlockCompressor_f ZSTD_selectBlockCompressor(ZSTD_strategy strat, ZSTD_ParamSwitch_e useRowMatchFinder, ZSTD_dictMode_e dictMode)
{
    static const ZSTD_BlockCompressor_f blockCompressor[4][ZSTD_STRATEGY_MAX+1] = {
        { ZSTD_compressBlock_fast  /* default for 0 */,
          ZSTD_compressBlock_fast,
          ZSTD_COMPRESSBLOCK_DOUBLEFAST,
          ZSTD_COMPRESSBLOCK_GREEDY,
          ZSTD_COMPRESSBLOCK_LAZY,
          ZSTD_COMPRESSBLOCK_LAZY2,
          ZSTD_COMPRESSBLOCK_BTLAZY2,
          ZSTD_COMPRESSBLOCK_BTOPT,
          ZSTD_COMPRESSBLOCK_BTULTRA,
          ZSTD_COMPRESSBLOCK_BTULTRA2
        },
        { ZSTD_compressBlock_fast_extDict  /* default for 0 */,
          ZSTD_compressBlock_fast_extDict,
          ZSTD_COMPRESSBLOCK_DOUBLEFAST_EXTDICT,
          ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT,
          ZSTD_COMPRESSBLOCK_LAZY_EXTDICT,
          ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT,
          ZSTD_COMPRESSBLOCK_BTLAZY2_EXTDICT,
          ZSTD_COMPRESSBLOCK_BTOPT_EXTDICT,
          ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT,
          ZSTD_COMPRESSBLOCK_BTULTRA_EXTDICT
        },
        { ZSTD_compressBlock_fast_dictMatchState  /* default for 0 */,
          ZSTD_compressBlock_fast_dictMatchState,
          ZSTD_COMPRESSBLOCK_DOUBLEFAST_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_BTLAZY2_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_BTOPT_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE,
          ZSTD_COMPRESSBLOCK_BTULTRA_DICTMATCHSTATE
        },
        { NULL  /* default for 0 */,
          NULL,
          NULL,
          ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH,
          ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH,
          ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH,
          NULL,
          NULL,
          NULL,
          NULL }
    };
    ZSTD_BlockCompressor_f selectedCompressor;
    ZSTD_STATIC_ASSERT((unsigned)ZSTD_fast == 1);

    assert(ZSTD_cParam_withinBounds(ZSTD_c_strategy, (int)strat));
    DEBUGLOG(5, "Selected block compressor: dictMode=%d strat=%d rowMatchfinder=%d", (int)dictMode, (int)strat, (int)useRowMatchFinder);
    if (ZSTD_rowMatchFinderUsed(strat, useRowMatchFinder)) {
        static const ZSTD_BlockCompressor_f rowBasedBlockCompressors[4][3] = {
            {
                ZSTD_COMPRESSBLOCK_GREEDY_ROW,
                ZSTD_COMPRESSBLOCK_LAZY_ROW,
                ZSTD_COMPRESSBLOCK_LAZY2_ROW
            },
            {
                ZSTD_COMPRESSBLOCK_GREEDY_EXTDICT_ROW,
                ZSTD_COMPRESSBLOCK_LAZY_EXTDICT_ROW,
                ZSTD_COMPRESSBLOCK_LAZY2_EXTDICT_ROW
            },
            {
                ZSTD_COMPRESSBLOCK_GREEDY_DICTMATCHSTATE_ROW,
                ZSTD_COMPRESSBLOCK_LAZY_DICTMATCHSTATE_ROW,
                ZSTD_COMPRESSBLOCK_LAZY2_DICTMATCHSTATE_ROW
            },
            {
                ZSTD_COMPRESSBLOCK_GREEDY_DEDICATEDDICTSEARCH_ROW,
                ZSTD_COMPRESSBLOCK_LAZY_DEDICATEDDICTSEARCH_ROW,
                ZSTD_COMPRESSBLOCK_LAZY2_DEDICATEDDICTSEARCH_ROW
            }
        };
        DEBUGLOG(5, "Selecting a row-based matchfinder");
        assert(useRowMatchFinder != ZSTD_ps_auto);
        selectedCompressor = rowBasedBlockCompressors[(int)dictMode][(int)strat - (int)ZSTD_greedy];
    } else {
        selectedCompressor = blockCompressor[(int)dictMode][(int)strat];
    }
    assert(selectedCompressor != NULL);
    return selectedCompressor;
}

static void ZSTD_storeLastLiterals(SeqStore_t* seqStorePtr,
                                   const BYTE* anchor, size_t lastLLSize)
{
    ZSTD_memcpy(seqStorePtr->lit, anchor, lastLLSize);
    seqStorePtr->lit += lastLLSize;
}

void ZSTD_resetSeqStore(SeqStore_t* ssPtr)
{
    ssPtr->lit = ssPtr->litStart;
    ssPtr->sequences = ssPtr->sequencesStart;
    ssPtr->longLengthType = ZSTD_llt_none;
}

/* ZSTD_postProcessSequenceProducerResult() :
 * Validates and post-processes sequences obtained through the external matchfinder API:
 *   - Checks whether nbExternalSeqs represents an error condition.
 *   - Appends a block delimiter to outSeqs if one is not already present.
 *     See zstd.h for context regarding block delimiters.
 * Returns the number of sequences after post-processing, or an error code. */
static size_t ZSTD_postProcessSequenceProducerResult(
    ZSTD_Sequence* outSeqs, size_t nbExternalSeqs, size_t outSeqsCapacity, size_t srcSize
) {
    RETURN_ERROR_IF(
        nbExternalSeqs > outSeqsCapacity,
        sequenceProducer_failed,
        "External sequence producer returned error code %lu",
        (unsigned long)nbExternalSeqs
    );

    RETURN_ERROR_IF(
        nbExternalSeqs == 0 && srcSize > 0,
        sequenceProducer_failed,
        "Got zero sequences from external sequence producer for a non-empty src buffer!"
    );

    if (srcSize == 0) {
        ZSTD_memset(&outSeqs[0], 0, sizeof(ZSTD_Sequence));
        return 1;
    }

    {
        ZSTD_Sequence const lastSeq = outSeqs[nbExternalSeqs - 1];

        /* We can return early if lastSeq is already a block delimiter. */
        if (lastSeq.offset == 0 && lastSeq.matchLength == 0) {
            return nbExternalSeqs;
        }

        /* This error condition is only possible if the external matchfinder
         * produced an invalid parse, by definition of ZSTD_sequenceBound(). */
        RETURN_ERROR_IF(
            nbExternalSeqs == outSeqsCapacity,
            sequenceProducer_failed,
            "nbExternalSeqs == outSeqsCapacity but lastSeq is not a block delimiter!"
        );

        /* lastSeq is not a block delimiter, so we need to append one. */
        ZSTD_memset(&outSeqs[nbExternalSeqs], 0, sizeof(ZSTD_Sequence));
        return nbExternalSeqs + 1;
    }
}

/* ZSTD_fastSequenceLengthSum() :
 * Returns sum(litLen) + sum(matchLen) + lastLits for *seqBuf*.
 * Similar to another function in zstd_compress.c (determine_blockSize),
 * except it doesn't check for a block delimiter to end summation.
 * Removing the early exit allows the compiler to auto-vectorize (https://godbolt.org/z/cY1cajz9P).
 * This function can be deleted and replaced by determine_blockSize after we resolve issue #3456. */
static size_t ZSTD_fastSequenceLengthSum(ZSTD_Sequence const* seqBuf, size_t seqBufSize) {
    size_t matchLenSum, litLenSum, i;
    matchLenSum = 0;
    litLenSum = 0;
    for (i = 0; i < seqBufSize; i++) {
        litLenSum += seqBuf[i].litLength;
        matchLenSum += seqBuf[i].matchLength;
    }
    return litLenSum + matchLenSum;
}

/**
 * Function to validate sequences produced by a block compressor.
 */
static void ZSTD_validateSeqStore(const SeqStore_t* seqStore, const ZSTD_compressionParameters* cParams)
{
#if DEBUGLEVEL >= 1
    const SeqDef* seq = seqStore->sequencesStart;
    const SeqDef* const seqEnd = seqStore->sequences;
    size_t const matchLenLowerBound = cParams->minMatch == 3 ? 3 : 4;
    for (; seq < seqEnd; ++seq) {
        const ZSTD_SequenceLength seqLength = ZSTD_getSequenceLength(seqStore, seq);
        assert(seqLength.matchLength >= matchLenLowerBound);
        (void)seqLength;
        (void)matchLenLowerBound;
    }
#else
    (void)seqStore;
    (void)cParams;
#endif
}

static size_t
ZSTD_transferSequences_wBlockDelim(ZSTD_CCtx* cctx,
                                   ZSTD_SequencePosition* seqPos,
                             const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
                             const void* src, size_t blockSize,
                                   ZSTD_ParamSwitch_e externalRepSearch);

typedef enum { ZSTDbss_compress, ZSTDbss_noCompress } ZSTD_BuildSeqStore_e;

static size_t ZSTD_buildSeqStore(ZSTD_CCtx* zc, const void* src, size_t srcSize)
{
    ZSTD_MatchState_t* const ms = &zc->blockState.matchState;
    DEBUGLOG(5, "ZSTD_buildSeqStore (srcSize=%zu)", srcSize);
    assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
    /* Assert that we have correctly flushed the ctx params into the ms's copy */
    ZSTD_assertEqualCParams(zc->appliedParams.cParams, ms->cParams);
    /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
     * additional 1. We need to revisit and change this logic to be more consistent */
    if (srcSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
        if (zc->appliedParams.cParams.strategy >= ZSTD_btopt) {
            ZSTD_ldm_skipRawSeqStoreBytes(&zc->externSeqStore, srcSize);
        } else {
            ZSTD_ldm_skipSequences(&zc->externSeqStore, srcSize, zc->appliedParams.cParams.minMatch);
        }
        return ZSTDbss_noCompress; /* don't even attempt compression below a certain srcSize */
    }
    ZSTD_resetSeqStore(&(zc->seqStore));
    /* required for optimal parser to read stats from dictionary */
    ms->opt.symbolCosts = &zc->blockState.prevCBlock->entropy;
    /* tell the optimal parser how we expect to compress literals */
    ms->opt.literalCompressionMode = zc->appliedParams.literalCompressionMode;
    /* a gap between an attached dict and the current window is not safe,
     * they must remain adjacent,
     * and when that stops being the case, the dict must be unset */
    assert(ms->dictMatchState == NULL || ms->loadedDictEnd == ms->window.dictLimit);

    /* limited update after a very long match */
    {   const BYTE* const base = ms->window.base;
        const BYTE* const istart = (const BYTE*)src;
        const U32 curr = (U32)(istart-base);
        if (sizeof(ptrdiff_t)==8) assert(istart - base < (ptrdiff_t)(U32)(-1));   /* ensure no overflow */
        if (curr > ms->nextToUpdate + 384)
            ms->nextToUpdate = curr - MIN(192, (U32)(curr - ms->nextToUpdate - 384));
    }

    /* select and store sequences */
    {   ZSTD_dictMode_e const dictMode = ZSTD_matchState_dictMode(ms);
        size_t lastLLSize;
        {   int i;
            for (i = 0; i < ZSTD_REP_NUM; ++i)
                zc->blockState.nextCBlock->rep[i] = zc->blockState.prevCBlock->rep[i];
        }
        if (zc->externSeqStore.pos < zc->externSeqStore.size) {
            assert(zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_disable);

            /* External matchfinder + LDM is technically possible, just not implemented yet.
             * We need to revisit soon and implement it. */
            RETURN_ERROR_IF(
                ZSTD_hasExtSeqProd(&zc->appliedParams),
                parameter_combination_unsupported,
                "Long-distance matching with external sequence producer enabled is not currently supported."
            );

            /* Updates ldmSeqStore.pos */
            lastLLSize =
                ZSTD_ldm_blockCompress(&zc->externSeqStore,
                                       ms, &zc->seqStore,
                                       zc->blockState.nextCBlock->rep,
                                       zc->appliedParams.useRowMatchFinder,
                                       src, srcSize);
            assert(zc->externSeqStore.pos <= zc->externSeqStore.size);
        } else if (zc->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
            RawSeqStore_t ldmSeqStore = kNullRawSeqStore;

            /* External matchfinder + LDM is technically possible, just not implemented yet.
             * We need to revisit soon and implement it. */
            RETURN_ERROR_IF(
                ZSTD_hasExtSeqProd(&zc->appliedParams),
                parameter_combination_unsupported,
                "Long-distance matching with external sequence producer enabled is not currently supported."
            );

            ldmSeqStore.seq = zc->ldmSequences;
            ldmSeqStore.capacity = zc->maxNbLdmSequences;
            /* Updates ldmSeqStore.size */
            FORWARD_IF_ERROR(ZSTD_ldm_generateSequences(&zc->ldmState, &ldmSeqStore,
                                               &zc->appliedParams.ldmParams,
                                               src, srcSize), "");
            /* Updates ldmSeqStore.pos */
            lastLLSize =
                ZSTD_ldm_blockCompress(&ldmSeqStore,
                                       ms, &zc->seqStore,
                                       zc->blockState.nextCBlock->rep,
                                       zc->appliedParams.useRowMatchFinder,
                                       src, srcSize);
            assert(ldmSeqStore.pos == ldmSeqStore.size);
        } else if (ZSTD_hasExtSeqProd(&zc->appliedParams)) {
            assert(
                zc->extSeqBufCapacity >= ZSTD_sequenceBound(srcSize)
            );
            assert(zc->appliedParams.extSeqProdFunc != NULL);

            {   U32 const windowSize = (U32)1 << zc->appliedParams.cParams.windowLog;

                size_t const nbExternalSeqs = (zc->appliedParams.extSeqProdFunc)(
                    zc->appliedParams.extSeqProdState,
                    zc->extSeqBuf,
                    zc->extSeqBufCapacity,
                    src, srcSize,
                    NULL, 0,  /* dict and dictSize, currently not supported */
                    zc->appliedParams.compressionLevel,
                    windowSize
                );

                size_t const nbPostProcessedSeqs = ZSTD_postProcessSequenceProducerResult(
                    zc->extSeqBuf,
                    nbExternalSeqs,
                    zc->extSeqBufCapacity,
                    srcSize
                );

                /* Return early if there is no error, since we don't need to worry about last literals */
                if (!ZSTD_isError(nbPostProcessedSeqs)) {
                    ZSTD_SequencePosition seqPos = {0,0,0};
                    size_t const seqLenSum = ZSTD_fastSequenceLengthSum(zc->extSeqBuf, nbPostProcessedSeqs);
                    RETURN_ERROR_IF(seqLenSum > srcSize, externalSequences_invalid, "External sequences imply too large a block!");
                    FORWARD_IF_ERROR(
                        ZSTD_transferSequences_wBlockDelim(
                            zc, &seqPos,
                            zc->extSeqBuf, nbPostProcessedSeqs,
                            src, srcSize,
                            zc->appliedParams.searchForExternalRepcodes
                        ),
                        "Failed to copy external sequences to seqStore!"
                    );
                    ms->ldmSeqStore = NULL;
                    DEBUGLOG(5, "Copied %lu sequences from external sequence producer to internal seqStore.", (unsigned long)nbExternalSeqs);
                    return ZSTDbss_compress;
                }

                /* Propagate the error if fallback is disabled */
                if (!zc->appliedParams.enableMatchFinderFallback) {
                    return nbPostProcessedSeqs;
                }

                /* Fallback to software matchfinder */
                {   ZSTD_BlockCompressor_f const blockCompressor =
                        ZSTD_selectBlockCompressor(
                            zc->appliedParams.cParams.strategy,
                            zc->appliedParams.useRowMatchFinder,
                            dictMode);
                    ms->ldmSeqStore = NULL;
                    DEBUGLOG(
                        5,
                        "External sequence producer returned error code %lu. Falling back to internal parser.",
                        (unsigned long)nbExternalSeqs
                    );
                    lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
            }   }
        } else {   /* not long range mode and no external matchfinder */
            ZSTD_BlockCompressor_f const blockCompressor = ZSTD_selectBlockCompressor(
                    zc->appliedParams.cParams.strategy,
                    zc->appliedParams.useRowMatchFinder,
                    dictMode);
            ms->ldmSeqStore = NULL;
            lastLLSize = blockCompressor(ms, &zc->seqStore, zc->blockState.nextCBlock->rep, src, srcSize);
        }
        {   const BYTE* const lastLiterals = (const BYTE*)src + srcSize - lastLLSize;
            ZSTD_storeLastLiterals(&zc->seqStore, lastLiterals, lastLLSize);
    }   }
    ZSTD_validateSeqStore(&zc->seqStore, &zc->appliedParams.cParams);
    return ZSTDbss_compress;
}

static size_t ZSTD_copyBlockSequences(SeqCollector* seqCollector, const SeqStore_t* seqStore, const U32 prevRepcodes[ZSTD_REP_NUM])
{
    const SeqDef* inSeqs = seqStore->sequencesStart;
    const size_t nbInSequences = (size_t)(seqStore->sequences - inSeqs);
    const size_t nbInLiterals = (size_t)(seqStore->lit - seqStore->litStart);

    ZSTD_Sequence* outSeqs = seqCollector->seqIndex == 0 ? seqCollector->seqStart : seqCollector->seqStart + seqCollector->seqIndex;
    const size_t nbOutSequences = nbInSequences + 1;
    size_t nbOutLiterals = 0;
    Repcodes_t repcodes;
    size_t i;

    /* Bounds check that we have enough space for every input sequence
     * and the block delimiter
     */
    assert(seqCollector->seqIndex <= seqCollector->maxSequences);
    RETURN_ERROR_IF(
        nbOutSequences > (size_t)(seqCollector->maxSequences - seqCollector->seqIndex),
        dstSize_tooSmall,
        "Not enough space to copy sequences");

    ZSTD_memcpy(&repcodes, prevRepcodes, sizeof(repcodes));
    for (i = 0; i < nbInSequences; ++i) {
        U32 rawOffset;
        outSeqs[i].litLength = inSeqs[i].litLength;
        outSeqs[i].matchLength = inSeqs[i].mlBase + MINMATCH;
        outSeqs[i].rep = 0;

        /* Handle the possible single length >= 64K
         * There can only be one because we add MINMATCH to every match length,
         * and blocks are at most 128K.
         */
        if (i == seqStore->longLengthPos) {
            if (seqStore->longLengthType == ZSTD_llt_literalLength) {
                outSeqs[i].litLength += 0x10000;
            } else if (seqStore->longLengthType == ZSTD_llt_matchLength) {
                outSeqs[i].matchLength += 0x10000;
            }
        }

        /* Determine the raw offset given the offBase, which may be a repcode. */
        if (OFFBASE_IS_REPCODE(inSeqs[i].offBase)) {
            const U32 repcode = OFFBASE_TO_REPCODE(inSeqs[i].offBase);
            assert(repcode > 0);
            outSeqs[i].rep = repcode;
            if (outSeqs[i].litLength != 0) {
                rawOffset = repcodes.rep[repcode - 1];
            } else {
                if (repcode == 3) {
                    assert(repcodes.rep[0] > 1);
                    rawOffset = repcodes.rep[0] - 1;
                } else {
                    rawOffset = repcodes.rep[repcode];
                }
            }
        } else {
            rawOffset = OFFBASE_TO_OFFSET(inSeqs[i].offBase);
        }
        outSeqs[i].offset = rawOffset;

        /* Update repcode history for the sequence */
        ZSTD_updateRep(repcodes.rep,
                       inSeqs[i].offBase,
                       inSeqs[i].litLength == 0);

        nbOutLiterals += outSeqs[i].litLength;
    }
    /* Insert last literals (if any exist) in the block as a sequence with ml == off == 0.
     * If there are no last literals, then we'll emit (of: 0, ml: 0, ll: 0), which is a marker
     * for the block boundary, according to the API.
     */
    assert(nbInLiterals >= nbOutLiterals);
    {
        const size_t lastLLSize = nbInLiterals - nbOutLiterals;
        outSeqs[nbInSequences].litLength = (U32)lastLLSize;
        outSeqs[nbInSequences].matchLength = 0;
        outSeqs[nbInSequences].offset = 0;
        assert(nbOutSequences == nbInSequences + 1);
    }
    seqCollector->seqIndex += nbOutSequences;
    assert(seqCollector->seqIndex <= seqCollector->maxSequences);

    return 0;
}

size_t ZSTD_sequenceBound(size_t srcSize) {
    const size_t maxNbSeq = (srcSize / ZSTD_MINMATCH_MIN) + 1;
    const size_t maxNbDelims = (srcSize / ZSTD_BLOCKSIZE_MAX_MIN) + 1;
    return maxNbSeq + maxNbDelims;
}

size_t ZSTD_generateSequences(ZSTD_CCtx* zc, ZSTD_Sequence* outSeqs,
                              size_t outSeqsSize, const void* src, size_t srcSize)
{
    const size_t dstCapacity = ZSTD_compressBound(srcSize);
    void* dst; /* Make C90 happy. */
    SeqCollector seqCollector;
    {
        int targetCBlockSize;
        FORWARD_IF_ERROR(ZSTD_CCtx_getParameter(zc, ZSTD_c_targetCBlockSize, &targetCBlockSize), "");
        RETURN_ERROR_IF(targetCBlockSize != 0, parameter_unsupported, "targetCBlockSize != 0");
    }
    {
        int nbWorkers;
        FORWARD_IF_ERROR(ZSTD_CCtx_getParameter(zc, ZSTD_c_nbWorkers, &nbWorkers), "");
        RETURN_ERROR_IF(nbWorkers != 0, parameter_unsupported, "nbWorkers != 0");
    }

    dst = ZSTD_customMalloc(dstCapacity, ZSTD_defaultCMem);
    RETURN_ERROR_IF(dst == NULL, memory_allocation, "NULL pointer!");

    seqCollector.collectSequences = 1;
    seqCollector.seqStart = outSeqs;
    seqCollector.seqIndex = 0;
    seqCollector.maxSequences = outSeqsSize;
    zc->seqCollector = seqCollector;

    {
        const size_t ret = ZSTD_compress2(zc, dst, dstCapacity, src, srcSize);
        ZSTD_customFree(dst, ZSTD_defaultCMem);
        FORWARD_IF_ERROR(ret, "ZSTD_compress2 failed");
    }
    assert(zc->seqCollector.seqIndex <= ZSTD_sequenceBound(srcSize));
    return zc->seqCollector.seqIndex;
}

size_t ZSTD_mergeBlockDelimiters(ZSTD_Sequence* sequences, size_t seqsSize) {
    size_t in = 0;
    size_t out = 0;
    for (; in < seqsSize; ++in) {
        if (sequences[in].offset == 0 && sequences[in].matchLength == 0) {
            if (in != seqsSize - 1) {
                sequences[in+1].litLength += sequences[in].litLength;
            }
        } else {
            sequences[out] = sequences[in];
            ++out;
        }
    }
    return out;
}

/* Unrolled loop to read four size_ts of input at a time. Returns 1 if is RLE, 0 if not. */
static int ZSTD_isRLE(const BYTE* src, size_t length) {
    const BYTE* ip = src;
    const BYTE value = ip[0];
    const size_t valueST = (size_t)((U64)value * 0x0101010101010101ULL);
    const size_t unrollSize = sizeof(size_t) * 4;
    const size_t unrollMask = unrollSize - 1;
    const size_t prefixLength = length & unrollMask;
    size_t i;
    if (length == 1) return 1;
    /* Check if prefix is RLE first before using unrolled loop */
    if (prefixLength && ZSTD_count(ip+1, ip, ip+prefixLength) != prefixLength-1) {
        return 0;
    }
    for (i = prefixLength; i != length; i += unrollSize) {
        size_t u;
        for (u = 0; u < unrollSize; u += sizeof(size_t)) {
            if (MEM_readST(ip + i + u) != valueST) {
                return 0;
    }   }   }
    return 1;
}

/* Returns true if the given block may be RLE.
 * This is just a heuristic based on the compressibility.
 * It may return both false positives and false negatives.
 */
static int ZSTD_maybeRLE(SeqStore_t const* seqStore)
{
    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
    size_t const nbLits = (size_t)(seqStore->lit - seqStore->litStart);

    return nbSeqs < 4 && nbLits < 10;
}

static void
ZSTD_blockState_confirmRepcodesAndEntropyTables(ZSTD_blockState_t* const bs)
{
    ZSTD_compressedBlockState_t* const tmp = bs->prevCBlock;
    bs->prevCBlock = bs->nextCBlock;
    bs->nextCBlock = tmp;
}

/* Writes the block header */
static void
writeBlockHeader(void* op, size_t cSize, size_t blockSize, U32 lastBlock)
{
    U32 const cBlockHeader = cSize == 1 ?
                        lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) :
                        lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
    MEM_writeLE24(op, cBlockHeader);
    DEBUGLOG(5, "writeBlockHeader: cSize: %zu blockSize: %zu lastBlock: %u", cSize, blockSize, lastBlock);
}

/** ZSTD_buildBlockEntropyStats_literals() :
 *  Builds entropy for the literals.
 *  Stores literals block type (raw, rle, compressed, repeat) and
 *  huffman description table to hufMetadata.
 *  Requires ENTROPY_WORKSPACE_SIZE workspace
 * @return : size of huffman description table, or an error code
 */
static size_t
ZSTD_buildBlockEntropyStats_literals(void* const src, size_t srcSize,
                               const ZSTD_hufCTables_t* prevHuf,
                                     ZSTD_hufCTables_t* nextHuf,
                                     ZSTD_hufCTablesMetadata_t* hufMetadata,
                               const int literalsCompressionIsDisabled,
                                     void* workspace, size_t wkspSize,
                                     int hufFlags)
{
    BYTE* const wkspStart = (BYTE*)workspace;
    BYTE* const wkspEnd = wkspStart + wkspSize;
    BYTE* const countWkspStart = wkspStart;
    unsigned* const countWksp = (unsigned*)workspace;
    const size_t countWkspSize = (HUF_SYMBOLVALUE_MAX + 1) * sizeof(unsigned);
    BYTE* const nodeWksp = countWkspStart + countWkspSize;
    const size_t nodeWkspSize = (size_t)(wkspEnd - nodeWksp);
    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
    unsigned huffLog = LitHufLog;
    HUF_repeat repeat = prevHuf->repeatMode;
    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_literals (srcSize=%zu)", srcSize);

    /* Prepare nextEntropy assuming reusing the existing table */
    ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));

    if (literalsCompressionIsDisabled) {
        DEBUGLOG(5, "set_basic - disabled");
        hufMetadata->hType = set_basic;
        return 0;
    }

    /* small ? don't even attempt compression (speed opt) */
#ifndef COMPRESS_LITERALS_SIZE_MIN
# define COMPRESS_LITERALS_SIZE_MIN 63  /* heuristic */
#endif
    {   size_t const minLitSize = (prevHuf->repeatMode == HUF_repeat_valid) ? 6 : COMPRESS_LITERALS_SIZE_MIN;
        if (srcSize <= minLitSize) {
            DEBUGLOG(5, "set_basic - too small");
            hufMetadata->hType = set_basic;
            return 0;
    }   }

    /* Scan input and build symbol stats */
    {   size_t const largest =
            HIST_count_wksp (countWksp, &maxSymbolValue,
                            (const BYTE*)src, srcSize,
                            workspace, wkspSize);
        FORWARD_IF_ERROR(largest, "HIST_count_wksp failed");
        if (largest == srcSize) {
            /* only one literal symbol */
            DEBUGLOG(5, "set_rle");
            hufMetadata->hType = set_rle;
            return 0;
        }
        if (largest <= (srcSize >> 7)+4) {
            /* heuristic: likely not compressible */
            DEBUGLOG(5, "set_basic - no gain");
            hufMetadata->hType = set_basic;
            return 0;
    }   }

    /* Validate the previous Huffman table */
    if (repeat == HUF_repeat_check
      && !HUF_validateCTable((HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue)) {
        repeat = HUF_repeat_none;
    }

    /* Build Huffman Tree */
    ZSTD_memset(nextHuf->CTable, 0, sizeof(nextHuf->CTable));
    huffLog = HUF_optimalTableLog(huffLog, srcSize, maxSymbolValue, nodeWksp, nodeWkspSize, nextHuf->CTable, countWksp, hufFlags);
    assert(huffLog <= LitHufLog);
    {   size_t const maxBits = HUF_buildCTable_wksp((HUF_CElt*)nextHuf->CTable, countWksp,
                                                    maxSymbolValue, huffLog,
                                                    nodeWksp, nodeWkspSize);
        FORWARD_IF_ERROR(maxBits, "HUF_buildCTable_wksp");
        huffLog = (U32)maxBits;
    }
    {   /* Build and write the CTable */
        size_t const newCSize = HUF_estimateCompressedSize(
                (HUF_CElt*)nextHuf->CTable, countWksp, maxSymbolValue);
        size_t const hSize = HUF_writeCTable_wksp(
                hufMetadata->hufDesBuffer, sizeof(hufMetadata->hufDesBuffer),
                (HUF_CElt*)nextHuf->CTable, maxSymbolValue, huffLog,
                nodeWksp, nodeWkspSize);
        /* Check against repeating the previous CTable */
        if (repeat != HUF_repeat_none) {
            size_t const oldCSize = HUF_estimateCompressedSize(
                    (HUF_CElt const*)prevHuf->CTable, countWksp, maxSymbolValue);
            if (oldCSize < srcSize && (oldCSize <= hSize + newCSize || hSize + 12 >= srcSize)) {
                DEBUGLOG(5, "set_repeat - smaller");
                ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
                hufMetadata->hType = set_repeat;
                return 0;
        }   }
        if (newCSize + hSize >= srcSize) {
            DEBUGLOG(5, "set_basic - no gains");
            ZSTD_memcpy(nextHuf, prevHuf, sizeof(*prevHuf));
            hufMetadata->hType = set_basic;
            return 0;
        }
        DEBUGLOG(5, "set_compressed (hSize=%u)", (U32)hSize);
        hufMetadata->hType = set_compressed;
        nextHuf->repeatMode = HUF_repeat_check;
        return hSize;
    }
}


/* ZSTD_buildDummySequencesStatistics():
 * Returns a ZSTD_symbolEncodingTypeStats_t with all encoding types as set_basic,
 * and updates nextEntropy to the appropriate repeatMode.
 */
static ZSTD_symbolEncodingTypeStats_t
ZSTD_buildDummySequencesStatistics(ZSTD_fseCTables_t* nextEntropy)
{
    ZSTD_symbolEncodingTypeStats_t stats = {set_basic, set_basic, set_basic, 0, 0, 0};
    nextEntropy->litlength_repeatMode = FSE_repeat_none;
    nextEntropy->offcode_repeatMode = FSE_repeat_none;
    nextEntropy->matchlength_repeatMode = FSE_repeat_none;
    return stats;
}

/** ZSTD_buildBlockEntropyStats_sequences() :
 *  Builds entropy for the sequences.
 *  Stores symbol compression modes and fse table to fseMetadata.
 *  Requires ENTROPY_WORKSPACE_SIZE wksp.
 * @return : size of fse tables or error code */
static size_t
ZSTD_buildBlockEntropyStats_sequences(
                const SeqStore_t* seqStorePtr,
                const ZSTD_fseCTables_t* prevEntropy,
                      ZSTD_fseCTables_t* nextEntropy,
                const ZSTD_CCtx_params* cctxParams,
                      ZSTD_fseCTablesMetadata_t* fseMetadata,
                      void* workspace, size_t wkspSize)
{
    ZSTD_strategy const strategy = cctxParams->cParams.strategy;
    size_t const nbSeq = (size_t)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
    BYTE* const ostart = fseMetadata->fseTablesBuffer;
    BYTE* const oend = ostart + sizeof(fseMetadata->fseTablesBuffer);
    BYTE* op = ostart;
    unsigned* countWorkspace = (unsigned*)workspace;
    unsigned* entropyWorkspace = countWorkspace + (MaxSeq + 1);
    size_t entropyWorkspaceSize = wkspSize - (MaxSeq + 1) * sizeof(*countWorkspace);
    ZSTD_symbolEncodingTypeStats_t stats;

    DEBUGLOG(5, "ZSTD_buildBlockEntropyStats_sequences (nbSeq=%zu)", nbSeq);
    stats = nbSeq != 0 ? ZSTD_buildSequencesStatistics(seqStorePtr, nbSeq,
                                          prevEntropy, nextEntropy, op, oend,
                                          strategy, countWorkspace,
                                          entropyWorkspace, entropyWorkspaceSize)
                       : ZSTD_buildDummySequencesStatistics(nextEntropy);
    FORWARD_IF_ERROR(stats.size, "ZSTD_buildSequencesStatistics failed!");
    fseMetadata->llType = (SymbolEncodingType_e) stats.LLtype;
    fseMetadata->ofType = (SymbolEncodingType_e) stats.Offtype;
    fseMetadata->mlType = (SymbolEncodingType_e) stats.MLtype;
    fseMetadata->lastCountSize = stats.lastCountSize;
    return stats.size;
}


/** ZSTD_buildBlockEntropyStats() :
 *  Builds entropy for the block.
 *  Requires workspace size ENTROPY_WORKSPACE_SIZE
 * @return : 0 on success, or an error code
 *  Note : also employed in superblock
 */
size_t ZSTD_buildBlockEntropyStats(
            const SeqStore_t* seqStorePtr,
            const ZSTD_entropyCTables_t* prevEntropy,
                  ZSTD_entropyCTables_t* nextEntropy,
            const ZSTD_CCtx_params* cctxParams,
                  ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                  void* workspace, size_t wkspSize)
{
    size_t const litSize = (size_t)(seqStorePtr->lit - seqStorePtr->litStart);
    int const huf_useOptDepth = (cctxParams->cParams.strategy >= HUF_OPTIMAL_DEPTH_THRESHOLD);
    int const hufFlags = huf_useOptDepth ? HUF_flags_optimalDepth : 0;

    entropyMetadata->hufMetadata.hufDesSize =
        ZSTD_buildBlockEntropyStats_literals(seqStorePtr->litStart, litSize,
                                            &prevEntropy->huf, &nextEntropy->huf,
                                            &entropyMetadata->hufMetadata,
                                            ZSTD_literalsCompressionIsDisabled(cctxParams),
                                            workspace, wkspSize, hufFlags);

    FORWARD_IF_ERROR(entropyMetadata->hufMetadata.hufDesSize, "ZSTD_buildBlockEntropyStats_literals failed");
    entropyMetadata->fseMetadata.fseTablesSize =
        ZSTD_buildBlockEntropyStats_sequences(seqStorePtr,
                                              &prevEntropy->fse, &nextEntropy->fse,
                                              cctxParams,
                                              &entropyMetadata->fseMetadata,
                                              workspace, wkspSize);
    FORWARD_IF_ERROR(entropyMetadata->fseMetadata.fseTablesSize, "ZSTD_buildBlockEntropyStats_sequences failed");
    return 0;
}

/* Returns the size estimate for the literals section (header + content) of a block */
static size_t
ZSTD_estimateBlockSize_literal(const BYTE* literals, size_t litSize,
                               const ZSTD_hufCTables_t* huf,
                               const ZSTD_hufCTablesMetadata_t* hufMetadata,
                               void* workspace, size_t wkspSize,
                               int writeEntropy)
{
    unsigned* const countWksp = (unsigned*)workspace;
    unsigned maxSymbolValue = HUF_SYMBOLVALUE_MAX;
    size_t literalSectionHeaderSize = 3 + (litSize >= 1 KB) + (litSize >= 16 KB);
    U32 singleStream = litSize < 256;

    if (hufMetadata->hType == set_basic) return litSize;
    else if (hufMetadata->hType == set_rle) return 1;
    else if (hufMetadata->hType == set_compressed || hufMetadata->hType == set_repeat) {
        size_t const largest = HIST_count_wksp (countWksp, &maxSymbolValue, (const BYTE*)literals, litSize, workspace, wkspSize);
        if (ZSTD_isError(largest)) return litSize;
        {   size_t cLitSizeEstimate = HUF_estimateCompressedSize((const HUF_CElt*)huf->CTable, countWksp, maxSymbolValue);
            if (writeEntropy) cLitSizeEstimate += hufMetadata->hufDesSize;
            if (!singleStream) cLitSizeEstimate += 6; /* multi-stream huffman uses 6-byte jump table */
            return cLitSizeEstimate + literalSectionHeaderSize;
    }   }
    assert(0); /* impossible */
    return 0;
}

/* Returns the size estimate for the FSE-compressed symbols (of, ml, ll) of a block */
static size_t
ZSTD_estimateBlockSize_symbolType(SymbolEncodingType_e type,
                    const BYTE* codeTable, size_t nbSeq, unsigned maxCode,
                    const FSE_CTable* fseCTable,
                    const U8* additionalBits,
                    short const* defaultNorm, U32 defaultNormLog, U32 defaultMax,
                    void* workspace, size_t wkspSize)
{
    unsigned* const countWksp = (unsigned*)workspace;
    const BYTE* ctp = codeTable;
    const BYTE* const ctStart = ctp;
    const BYTE* const ctEnd = ctStart + nbSeq;
    size_t cSymbolTypeSizeEstimateInBits = 0;
    unsigned max = maxCode;

    HIST_countFast_wksp(countWksp, &max, codeTable, nbSeq, workspace, wkspSize);  /* can't fail */
    if (type == set_basic) {
        /* We selected this encoding type, so it must be valid. */
        assert(max <= defaultMax);
        (void)defaultMax;
        cSymbolTypeSizeEstimateInBits = ZSTD_crossEntropyCost(defaultNorm, defaultNormLog, countWksp, max);
    } else if (type == set_rle) {
        cSymbolTypeSizeEstimateInBits = 0;
    } else if (type == set_compressed || type == set_repeat) {
        cSymbolTypeSizeEstimateInBits = ZSTD_fseBitCost(fseCTable, countWksp, max);
    }
    if (ZSTD_isError(cSymbolTypeSizeEstimateInBits)) {
        return nbSeq * 10;
    }
    while (ctp < ctEnd) {
        if (additionalBits) cSymbolTypeSizeEstimateInBits += additionalBits[*ctp];
        else cSymbolTypeSizeEstimateInBits += *ctp; /* for offset, offset code is also the number of additional bits */
        ctp++;
    }
    return cSymbolTypeSizeEstimateInBits >> 3;
}

/* Returns the size estimate for the sequences section (header + content) of a block */
static size_t
ZSTD_estimateBlockSize_sequences(const BYTE* ofCodeTable,
                                 const BYTE* llCodeTable,
                                 const BYTE* mlCodeTable,
                                 size_t nbSeq,
                                 const ZSTD_fseCTables_t* fseTables,
                                 const ZSTD_fseCTablesMetadata_t* fseMetadata,
                                 void* workspace, size_t wkspSize,
                                 int writeEntropy)
{
    size_t sequencesSectionHeaderSize = 1 /* seqHead */ + 1 /* min seqSize size */ + (nbSeq >= 128) + (nbSeq >= LONGNBSEQ);
    size_t cSeqSizeEstimate = 0;
    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->ofType, ofCodeTable, nbSeq, MaxOff,
                                    fseTables->offcodeCTable, NULL,
                                    OF_defaultNorm, OF_defaultNormLog, DefaultMaxOff,
                                    workspace, wkspSize);
    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->llType, llCodeTable, nbSeq, MaxLL,
                                    fseTables->litlengthCTable, LL_bits,
                                    LL_defaultNorm, LL_defaultNormLog, MaxLL,
                                    workspace, wkspSize);
    cSeqSizeEstimate += ZSTD_estimateBlockSize_symbolType(fseMetadata->mlType, mlCodeTable, nbSeq, MaxML,
                                    fseTables->matchlengthCTable, ML_bits,
                                    ML_defaultNorm, ML_defaultNormLog, MaxML,
                                    workspace, wkspSize);
    if (writeEntropy) cSeqSizeEstimate += fseMetadata->fseTablesSize;
    return cSeqSizeEstimate + sequencesSectionHeaderSize;
}

/* Returns the size estimate for a given stream of literals, of, ll, ml */
static size_t
ZSTD_estimateBlockSize(const BYTE* literals, size_t litSize,
                       const BYTE* ofCodeTable,
                       const BYTE* llCodeTable,
                       const BYTE* mlCodeTable,
                       size_t nbSeq,
                       const ZSTD_entropyCTables_t* entropy,
                       const ZSTD_entropyCTablesMetadata_t* entropyMetadata,
                       void* workspace, size_t wkspSize,
                       int writeLitEntropy, int writeSeqEntropy)
{
    size_t const literalsSize = ZSTD_estimateBlockSize_literal(literals, litSize,
                                    &entropy->huf, &entropyMetadata->hufMetadata,
                                    workspace, wkspSize, writeLitEntropy);
    size_t const seqSize = ZSTD_estimateBlockSize_sequences(ofCodeTable, llCodeTable, mlCodeTable,
                                    nbSeq, &entropy->fse, &entropyMetadata->fseMetadata,
                                    workspace, wkspSize, writeSeqEntropy);
    return seqSize + literalsSize + ZSTD_blockHeaderSize;
}

/* Builds entropy statistics and uses them for blocksize estimation.
 *
 * @return: estimated compressed size of the seqStore, or a zstd error.
 */
static size_t
ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(SeqStore_t* seqStore, ZSTD_CCtx* zc)
{
    ZSTD_entropyCTablesMetadata_t* const entropyMetadata = &zc->blockSplitCtx.entropyMetadata;
    DEBUGLOG(6, "ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize()");
    FORWARD_IF_ERROR(ZSTD_buildBlockEntropyStats(seqStore,
                    &zc->blockState.prevCBlock->entropy,
                    &zc->blockState.nextCBlock->entropy,
                    &zc->appliedParams,
                    entropyMetadata,
                    zc->tmpWorkspace, zc->tmpWkspSize), "");
    return ZSTD_estimateBlockSize(
                    seqStore->litStart, (size_t)(seqStore->lit - seqStore->litStart),
                    seqStore->ofCode, seqStore->llCode, seqStore->mlCode,
                    (size_t)(seqStore->sequences - seqStore->sequencesStart),
                    &zc->blockState.nextCBlock->entropy,
                    entropyMetadata,
                    zc->tmpWorkspace, zc->tmpWkspSize,
                    (int)(entropyMetadata->hufMetadata.hType == set_compressed), 1);
}

/* Returns literals bytes represented in a seqStore */
static size_t ZSTD_countSeqStoreLiteralsBytes(const SeqStore_t* const seqStore)
{
    size_t literalsBytes = 0;
    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
    size_t i;
    for (i = 0; i < nbSeqs; ++i) {
        SeqDef const seq = seqStore->sequencesStart[i];
        literalsBytes += seq.litLength;
        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_literalLength) {
            literalsBytes += 0x10000;
    }   }
    return literalsBytes;
}

/* Returns match bytes represented in a seqStore */
static size_t ZSTD_countSeqStoreMatchBytes(const SeqStore_t* const seqStore)
{
    size_t matchBytes = 0;
    size_t const nbSeqs = (size_t)(seqStore->sequences - seqStore->sequencesStart);
    size_t i;
    for (i = 0; i < nbSeqs; ++i) {
        SeqDef seq = seqStore->sequencesStart[i];
        matchBytes += seq.mlBase + MINMATCH;
        if (i == seqStore->longLengthPos && seqStore->longLengthType == ZSTD_llt_matchLength) {
            matchBytes += 0x10000;
    }   }
    return matchBytes;
}

/* Derives the seqStore that is a chunk of the originalSeqStore from [startIdx, endIdx).
 * Stores the result in resultSeqStore.
 */
static void ZSTD_deriveSeqStoreChunk(SeqStore_t* resultSeqStore,
                               const SeqStore_t* originalSeqStore,
                                     size_t startIdx, size_t endIdx)
{
    *resultSeqStore = *originalSeqStore;
    if (startIdx > 0) {
        resultSeqStore->sequences = originalSeqStore->sequencesStart + startIdx;
        resultSeqStore->litStart += ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
    }

    /* Move longLengthPos into the correct position if necessary */
    if (originalSeqStore->longLengthType != ZSTD_llt_none) {
        if (originalSeqStore->longLengthPos < startIdx || originalSeqStore->longLengthPos > endIdx) {
            resultSeqStore->longLengthType = ZSTD_llt_none;
        } else {
            resultSeqStore->longLengthPos -= (U32)startIdx;
        }
    }
    resultSeqStore->sequencesStart = originalSeqStore->sequencesStart + startIdx;
    resultSeqStore->sequences = originalSeqStore->sequencesStart + endIdx;
    if (endIdx == (size_t)(originalSeqStore->sequences - originalSeqStore->sequencesStart)) {
        /* This accounts for possible last literals if the derived chunk reaches the end of the block */
        assert(resultSeqStore->lit == originalSeqStore->lit);
    } else {
        size_t const literalsBytes = ZSTD_countSeqStoreLiteralsBytes(resultSeqStore);
        resultSeqStore->lit = resultSeqStore->litStart + literalsBytes;
    }
    resultSeqStore->llCode += startIdx;
    resultSeqStore->mlCode += startIdx;
    resultSeqStore->ofCode += startIdx;
}

/**
 * Returns the raw offset represented by the combination of offBase, ll0, and repcode history.
 * offBase must represent a repcode in the numeric representation of ZSTD_storeSeq().
 */
static U32
ZSTD_resolveRepcodeToRawOffset(const U32 rep[ZSTD_REP_NUM], const U32 offBase, const U32 ll0)
{
    U32 const adjustedRepCode = OFFBASE_TO_REPCODE(offBase) - 1 + ll0;  /* [ 0 - 3 ] */
    assert(OFFBASE_IS_REPCODE(offBase));
    if (adjustedRepCode == ZSTD_REP_NUM) {
        assert(ll0);
        /* litlength == 0 and offCode == 2 implies selection of first repcode - 1
         * This is only valid if it results in a valid offset value, aka > 0.
         * Note : it may happen that `rep[0]==1` in exceptional circumstances.
         * In which case this function will return 0, which is an invalid offset.
         * It's not an issue though, since this value will be
         * compared and discarded within ZSTD_seqStore_resolveOffCodes().
         */
        return rep[0] - 1;
    }
    return rep[adjustedRepCode];
}

/**
 * ZSTD_seqStore_resolveOffCodes() reconciles any possible divergences in offset history that may arise
 * due to emission of RLE/raw blocks that disturb the offset history,
 * and replaces any repcodes within the seqStore that may be invalid.
 *
 * dRepcodes are updated as would be on the decompression side.
 * cRepcodes are updated exactly in accordance with the seqStore.
 *
 * Note : this function assumes seq->offBase respects the following numbering scheme :
 *        0 : invalid
 *        1-3 : repcode 1-3
 *        4+ : real_offset+3
 */
static void
ZSTD_seqStore_resolveOffCodes(Repcodes_t* const dRepcodes, Repcodes_t* const cRepcodes,
                        const SeqStore_t* const seqStore, U32 const nbSeq)
{
    U32 idx = 0;
    U32 const longLitLenIdx = seqStore->longLengthType == ZSTD_llt_literalLength ? seqStore->longLengthPos : nbSeq;
    for (; idx < nbSeq; ++idx) {
        SeqDef* const seq = seqStore->sequencesStart + idx;
        U32 const ll0 = (seq->litLength == 0) && (idx != longLitLenIdx);
        U32 const offBase = seq->offBase;
        assert(offBase > 0);
        if (OFFBASE_IS_REPCODE(offBase)) {
            U32 const dRawOffset = ZSTD_resolveRepcodeToRawOffset(dRepcodes->rep, offBase, ll0);
            U32 const cRawOffset = ZSTD_resolveRepcodeToRawOffset(cRepcodes->rep, offBase, ll0);
            /* Adjust simulated decompression repcode history if we come across a mismatch. Replace
             * the repcode with the offset it actually references, determined by the compression
             * repcode history.
             */
            if (dRawOffset != cRawOffset) {
                seq->offBase = OFFSET_TO_OFFBASE(cRawOffset);
            }
        }
        /* Compression repcode history is always updated with values directly from the unmodified seqStore.
         * Decompression repcode history may use modified seq->offset value taken from compression repcode history.
         */
        ZSTD_updateRep(dRepcodes->rep, seq->offBase, ll0);
        ZSTD_updateRep(cRepcodes->rep, offBase, ll0);
    }
}

/* ZSTD_compressSeqStore_singleBlock():
 * Compresses a seqStore into a block with a block header, into the buffer dst.
 *
 * Returns the total size of that block (including header) or a ZSTD error code.
 */
static size_t
ZSTD_compressSeqStore_singleBlock(ZSTD_CCtx* zc,
                            const SeqStore_t* const seqStore,
                                  Repcodes_t* const dRep, Repcodes_t* const cRep,
                                  void* dst, size_t dstCapacity,
                            const void* src, size_t srcSize,
                                  U32 lastBlock, U32 isPartition)
{
    const U32 rleMaxLength = 25;
    BYTE* op = (BYTE*)dst;
    const BYTE* ip = (const BYTE*)src;
    size_t cSize;
    size_t cSeqsSize;

    /* In case of an RLE or raw block, the simulated decompression repcode history must be reset */
    Repcodes_t const dRepOriginal = *dRep;
    DEBUGLOG(5, "ZSTD_compressSeqStore_singleBlock");
    if (isPartition)
        ZSTD_seqStore_resolveOffCodes(dRep, cRep, seqStore, (U32)(seqStore->sequences - seqStore->sequencesStart));

    RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "Block header doesn't fit");
    cSeqsSize = ZSTD_entropyCompressSeqStore(seqStore,
                &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
                &zc->appliedParams,
                op + ZSTD_blockHeaderSize, dstCapacity - ZSTD_blockHeaderSize,
                srcSize,
                zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */,
                zc->bmi2);
    FORWARD_IF_ERROR(cSeqsSize, "ZSTD_entropyCompressSeqStore failed!");

    if (!zc->isFirstBlock &&
        cSeqsSize < rleMaxLength &&
        ZSTD_isRLE((BYTE const*)src, srcSize)) {
        /* We don't want to emit our first block as a RLE even if it qualifies because
        * doing so will cause the decoder (cli only) to throw a "should consume all input error."
        * This is only an issue for zstd <= v1.4.3
        */
        cSeqsSize = 1;
    }

    /* Sequence collection not supported when block splitting */
    if (zc->seqCollector.collectSequences) {
        FORWARD_IF_ERROR(ZSTD_copyBlockSequences(&zc->seqCollector, seqStore, dRepOriginal.rep), "copyBlockSequences failed");
        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
        return 0;
    }

    if (cSeqsSize == 0) {
        cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, srcSize, lastBlock);
        FORWARD_IF_ERROR(cSize, "Nocompress block failed");
        DEBUGLOG(5, "Writing out nocompress block, size: %zu", cSize);
        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
    } else if (cSeqsSize == 1) {
        cSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, srcSize, lastBlock);
        FORWARD_IF_ERROR(cSize, "RLE compress block failed");
        DEBUGLOG(5, "Writing out RLE block, size: %zu", cSize);
        *dRep = dRepOriginal; /* reset simulated decompression repcode history */
    } else {
        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
        writeBlockHeader(op, cSeqsSize, srcSize, lastBlock);
        cSize = ZSTD_blockHeaderSize + cSeqsSize;
        DEBUGLOG(5, "Writing out compressed block, size: %zu", cSize);
    }

    if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
        zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;

    return cSize;
}

/* Struct to keep track of where we are in our recursive calls. */
typedef struct {
    U32* splitLocations;    /* Array of split indices */
    size_t idx;             /* The current index within splitLocations being worked on */
} seqStoreSplits;

#define MIN_SEQUENCES_BLOCK_SPLITTING 300

/* Helper function to perform the recursive search for block splits.
 * Estimates the cost of seqStore prior to split, and estimates the cost of splitting the sequences in half.
 * If advantageous to split, then we recurse down the two sub-blocks.
 * If not, or if an error occurred in estimation, then we do not recurse.
 *
 * Note: The recursion depth is capped by a heuristic minimum number of sequences,
 * defined by MIN_SEQUENCES_BLOCK_SPLITTING.
 * In theory, this means the absolute largest recursion depth is 10 == log2(maxNbSeqInBlock/MIN_SEQUENCES_BLOCK_SPLITTING).
 * In practice, recursion depth usually doesn't go beyond 4.
 *
 * Furthermore, the number of splits is capped by ZSTD_MAX_NB_BLOCK_SPLITS.
 * At ZSTD_MAX_NB_BLOCK_SPLITS == 196 with the current existing blockSize
 * maximum of 128 KB, this value is actually impossible to reach.
 */
static void
ZSTD_deriveBlockSplitsHelper(seqStoreSplits* splits, size_t startIdx, size_t endIdx,
                             ZSTD_CCtx* zc, const SeqStore_t* origSeqStore)
{
    SeqStore_t* const fullSeqStoreChunk = &zc->blockSplitCtx.fullSeqStoreChunk;
    SeqStore_t* const firstHalfSeqStore = &zc->blockSplitCtx.firstHalfSeqStore;
    SeqStore_t* const secondHalfSeqStore = &zc->blockSplitCtx.secondHalfSeqStore;
    size_t estimatedOriginalSize;
    size_t estimatedFirstHalfSize;
    size_t estimatedSecondHalfSize;
    size_t midIdx = (startIdx + endIdx)/2;

    DEBUGLOG(5, "ZSTD_deriveBlockSplitsHelper: startIdx=%zu endIdx=%zu", startIdx, endIdx);
    assert(endIdx >= startIdx);
    if (endIdx - startIdx < MIN_SEQUENCES_BLOCK_SPLITTING || splits->idx >= ZSTD_MAX_NB_BLOCK_SPLITS) {
        DEBUGLOG(6, "ZSTD_deriveBlockSplitsHelper: Too few sequences (%zu)", endIdx - startIdx);
        return;
    }
    ZSTD_deriveSeqStoreChunk(fullSeqStoreChunk, origSeqStore, startIdx, endIdx);
    ZSTD_deriveSeqStoreChunk(firstHalfSeqStore, origSeqStore, startIdx, midIdx);
    ZSTD_deriveSeqStoreChunk(secondHalfSeqStore, origSeqStore, midIdx, endIdx);
    estimatedOriginalSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(fullSeqStoreChunk, zc);
    estimatedFirstHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(firstHalfSeqStore, zc);
    estimatedSecondHalfSize = ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(secondHalfSeqStore, zc);
    DEBUGLOG(5, "Estimated original block size: %zu -- First half split: %zu -- Second half split: %zu",
             estimatedOriginalSize, estimatedFirstHalfSize, estimatedSecondHalfSize);
    if (ZSTD_isError(estimatedOriginalSize) || ZSTD_isError(estimatedFirstHalfSize) || ZSTD_isError(estimatedSecondHalfSize)) {
        return;
    }
    if (estimatedFirstHalfSize + estimatedSecondHalfSize < estimatedOriginalSize) {
        DEBUGLOG(5, "split decided at seqNb:%zu", midIdx);
        ZSTD_deriveBlockSplitsHelper(splits, startIdx, midIdx, zc, origSeqStore);
        splits->splitLocations[splits->idx] = (U32)midIdx;
        splits->idx++;
        ZSTD_deriveBlockSplitsHelper(splits, midIdx, endIdx, zc, origSeqStore);
    }
}

/* Base recursive function.
 * Populates a table with intra-block partition indices that can improve compression ratio.
 *
 * @return: number of splits made (which equals the size of the partition table - 1).
 */
static size_t ZSTD_deriveBlockSplits(ZSTD_CCtx* zc, U32 partitions[], U32 nbSeq)
{
    seqStoreSplits splits;
    splits.splitLocations = partitions;
    splits.idx = 0;
    if (nbSeq <= 4) {
        DEBUGLOG(5, "ZSTD_deriveBlockSplits: Too few sequences to split (%u <= 4)", nbSeq);
        /* Refuse to try and split anything with less than 4 sequences */
        return 0;
    }
    ZSTD_deriveBlockSplitsHelper(&splits, 0, nbSeq, zc, &zc->seqStore);
    splits.splitLocations[splits.idx] = nbSeq;
    DEBUGLOG(5, "ZSTD_deriveBlockSplits: final nb partitions: %zu", splits.idx+1);
    return splits.idx;
}

/* ZSTD_compressBlock_splitBlock():
 * Attempts to split a given block into multiple blocks to improve compression ratio.
 *
 * Returns combined size of all blocks (which includes headers), or a ZSTD error code.
 */
static size_t
ZSTD_compressBlock_splitBlock_internal(ZSTD_CCtx* zc,
                                    void* dst, size_t dstCapacity,
                              const void* src, size_t blockSize,
                                    U32 lastBlock, U32 nbSeq)
{
    size_t cSize = 0;
    const BYTE* ip = (const BYTE*)src;
    BYTE* op = (BYTE*)dst;
    size_t i = 0;
    size_t srcBytesTotal = 0;
    U32* const partitions = zc->blockSplitCtx.partitions; /* size == ZSTD_MAX_NB_BLOCK_SPLITS */
    SeqStore_t* const nextSeqStore = &zc->blockSplitCtx.nextSeqStore;
    SeqStore_t* const currSeqStore = &zc->blockSplitCtx.currSeqStore;
    size_t const numSplits = ZSTD_deriveBlockSplits(zc, partitions, nbSeq);

    /* If a block is split and some partitions are emitted as RLE/uncompressed, then repcode history
     * may become invalid. In order to reconcile potentially invalid repcodes, we keep track of two
     * separate repcode histories that simulate repcode history on compression and decompression side,
     * and use the histories to determine whether we must replace a particular repcode with its raw offset.
     *
     * 1) cRep gets updated for each partition, regardless of whether the block was emitted as uncompressed
     *    or RLE. This allows us to retrieve the offset value that an invalid repcode references within
     *    a nocompress/RLE block.
     * 2) dRep gets updated only for compressed partitions, and when a repcode gets replaced, will use
     *    the replacement offset value rather than the original repcode to update the repcode history.
     *    dRep also will be the final repcode history sent to the next block.
     *
     * See ZSTD_seqStore_resolveOffCodes() for more details.
     */
    Repcodes_t dRep;
    Repcodes_t cRep;
    ZSTD_memcpy(dRep.rep, zc->blockState.prevCBlock->rep, sizeof(Repcodes_t));
    ZSTD_memcpy(cRep.rep, zc->blockState.prevCBlock->rep, sizeof(Repcodes_t));
    ZSTD_memset(nextSeqStore, 0, sizeof(SeqStore_t));

    DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
                (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit,
                (unsigned)zc->blockState.matchState.nextToUpdate);

    if (numSplits == 0) {
        size_t cSizeSingleBlock =
            ZSTD_compressSeqStore_singleBlock(zc, &zc->seqStore,
                                            &dRep, &cRep,
                                            op, dstCapacity,
                                            ip, blockSize,
                                            lastBlock, 0 /* isPartition */);
        FORWARD_IF_ERROR(cSizeSingleBlock, "Compressing single block from splitBlock_internal() failed!");
        DEBUGLOG(5, "ZSTD_compressBlock_splitBlock_internal: No splits");
        assert(zc->blockSizeMax <= ZSTD_BLOCKSIZE_MAX);
        assert(cSizeSingleBlock <= zc->blockSizeMax + ZSTD_blockHeaderSize);
        return cSizeSingleBlock;
    }

    ZSTD_deriveSeqStoreChunk(currSeqStore, &zc->seqStore, 0, partitions[0]);
    for (i = 0; i <= numSplits; ++i) {
        size_t cSizeChunk;
        U32 const lastPartition = (i == numSplits);
        U32 lastBlockEntireSrc = 0;

        size_t srcBytes = ZSTD_countSeqStoreLiteralsBytes(currSeqStore) + ZSTD_countSeqStoreMatchBytes(currSeqStore);
        srcBytesTotal += srcBytes;
        if (lastPartition) {
            /* This is the final partition, need to account for possible last literals */
            srcBytes += blockSize - srcBytesTotal;
            lastBlockEntireSrc = lastBlock;
        } else {
            ZSTD_deriveSeqStoreChunk(nextSeqStore, &zc->seqStore, partitions[i], partitions[i+1]);
        }

        cSizeChunk = ZSTD_compressSeqStore_singleBlock(zc, currSeqStore,
                                                      &dRep, &cRep,
                                                       op, dstCapacity,
                                                       ip, srcBytes,
                                                       lastBlockEntireSrc, 1 /* isPartition */);
        DEBUGLOG(5, "Estimated size: %zu vs %zu : actual size",
                    ZSTD_buildEntropyStatisticsAndEstimateSubBlockSize(currSeqStore, zc), cSizeChunk);
        FORWARD_IF_ERROR(cSizeChunk, "Compressing chunk failed!");

        ip += srcBytes;
        op += cSizeChunk;
        dstCapacity -= cSizeChunk;
        cSize += cSizeChunk;
        *currSeqStore = *nextSeqStore;
        assert(cSizeChunk <= zc->blockSizeMax + ZSTD_blockHeaderSize);
    }
    /* cRep and dRep may have diverged during the compression.
     * If so, we use the dRep repcodes for the next block.
     */
    ZSTD_memcpy(zc->blockState.prevCBlock->rep, dRep.rep, sizeof(Repcodes_t));
    return cSize;
}

static size_t
ZSTD_compressBlock_splitBlock(ZSTD_CCtx* zc,
                              void* dst, size_t dstCapacity,
                              const void* src, size_t srcSize, U32 lastBlock)
{
    U32 nbSeq;
    size_t cSize;
    DEBUGLOG(5, "ZSTD_compressBlock_splitBlock");
    assert(zc->appliedParams.postBlockSplitter == ZSTD_ps_enable);

    {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
        FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
        if (bss == ZSTDbss_noCompress) {
            if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
                zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;
            RETURN_ERROR_IF(zc->seqCollector.collectSequences, sequenceProducer_failed, "Uncompressible block");
            cSize = ZSTD_noCompressBlock(dst, dstCapacity, src, srcSize, lastBlock);
            FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
            DEBUGLOG(5, "ZSTD_compressBlock_splitBlock: Nocompress block");
            return cSize;
        }
        nbSeq = (U32)(zc->seqStore.sequences - zc->seqStore.sequencesStart);
    }

    cSize = ZSTD_compressBlock_splitBlock_internal(zc, dst, dstCapacity, src, srcSize, lastBlock, nbSeq);
    FORWARD_IF_ERROR(cSize, "Splitting blocks failed!");
    return cSize;
}

static size_t
ZSTD_compressBlock_internal(ZSTD_CCtx* zc,
                            void* dst, size_t dstCapacity,
                            const void* src, size_t srcSize, U32 frame)
{
    /* This is an estimated upper bound for the length of an rle block.
     * This isn't the actual upper bound.
     * Finding the real threshold needs further investigation.
     */
    const U32 rleMaxLength = 25;
    size_t cSize;
    const BYTE* ip = (const BYTE*)src;
    BYTE* op = (BYTE*)dst;
    DEBUGLOG(5, "ZSTD_compressBlock_internal (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u)",
                (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit,
                (unsigned)zc->blockState.matchState.nextToUpdate);

    {   const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
        FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");
        if (bss == ZSTDbss_noCompress) {
            RETURN_ERROR_IF(zc->seqCollector.collectSequences, sequenceProducer_failed, "Uncompressible block");
            cSize = 0;
            goto out;
        }
    }

    if (zc->seqCollector.collectSequences) {
        FORWARD_IF_ERROR(ZSTD_copyBlockSequences(&zc->seqCollector, ZSTD_getSeqStore(zc), zc->blockState.prevCBlock->rep), "copyBlockSequences failed");
        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
        return 0;
    }

    /* encode sequences and literals */
    cSize = ZSTD_entropyCompressSeqStore(&zc->seqStore,
            &zc->blockState.prevCBlock->entropy, &zc->blockState.nextCBlock->entropy,
            &zc->appliedParams,
            dst, dstCapacity,
            srcSize,
            zc->tmpWorkspace, zc->tmpWkspSize /* statically allocated in resetCCtx */,
            zc->bmi2);

    if (frame &&
        /* We don't want to emit our first block as a RLE even if it qualifies because
         * doing so will cause the decoder (cli only) to throw a "should consume all input error."
         * This is only an issue for zstd <= v1.4.3
         */
        !zc->isFirstBlock &&
        cSize < rleMaxLength &&
        ZSTD_isRLE(ip, srcSize))
    {
        cSize = 1;
        op[0] = ip[0];
    }

out:
    if (!ZSTD_isError(cSize) && cSize > 1) {
        ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
    }
    /* We check that dictionaries have offset codes available for the first
     * block. After the first block, the offcode table might not have large
     * enough codes to represent the offsets in the data.
     */
    if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
        zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;

    return cSize;
}

static size_t ZSTD_compressBlock_targetCBlockSize_body(ZSTD_CCtx* zc,
                               void* dst, size_t dstCapacity,
                               const void* src, size_t srcSize,
                               const size_t bss, U32 lastBlock)
{
    DEBUGLOG(6, "Attempting ZSTD_compressSuperBlock()");
    if (bss == ZSTDbss_compress) {
        if (/* We don't want to emit our first block as a RLE even if it qualifies because
            * doing so will cause the decoder (cli only) to throw a "should consume all input error."
            * This is only an issue for zstd <= v1.4.3
            */
            !zc->isFirstBlock &&
            ZSTD_maybeRLE(&zc->seqStore) &&
            ZSTD_isRLE((BYTE const*)src, srcSize))
        {
            return ZSTD_rleCompressBlock(dst, dstCapacity, *(BYTE const*)src, srcSize, lastBlock);
        }
        /* Attempt superblock compression.
         *
         * Note that compressed size of ZSTD_compressSuperBlock() is not bound by the
         * standard ZSTD_compressBound(). This is a problem, because even if we have
         * space now, taking an extra byte now could cause us to run out of space later
         * and violate ZSTD_compressBound().
         *
         * Define blockBound(blockSize) = blockSize + ZSTD_blockHeaderSize.
         *
         * In order to respect ZSTD_compressBound() we must attempt to emit a raw
         * uncompressed block in these cases:
         *   * cSize == 0: Return code for an uncompressed block.
         *   * cSize == dstSize_tooSmall: We may have expanded beyond blockBound(srcSize).
         *     ZSTD_noCompressBlock() will return dstSize_tooSmall if we are really out of
         *     output space.
         *   * cSize >= blockBound(srcSize): We have expanded the block too much so
         *     emit an uncompressed block.
         */
        {   size_t const cSize =
                ZSTD_compressSuperBlock(zc, dst, dstCapacity, src, srcSize, lastBlock);
            if (cSize != ERROR(dstSize_tooSmall)) {
                size_t const maxCSize =
                    srcSize - ZSTD_minGain(srcSize, zc->appliedParams.cParams.strategy);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressSuperBlock failed");
                if (cSize != 0 && cSize < maxCSize + ZSTD_blockHeaderSize) {
                    ZSTD_blockState_confirmRepcodesAndEntropyTables(&zc->blockState);
                    return cSize;
                }
            }
        }
    } /* if (bss == ZSTDbss_compress)*/

    DEBUGLOG(6, "Resorting to ZSTD_noCompressBlock()");
    /* Superblock compression failed, attempt to emit a single no compress block.
     * The decoder will be able to stream this block since it is uncompressed.
     */
    return ZSTD_noCompressBlock(dst, dstCapacity, src, srcSize, lastBlock);
}

static size_t ZSTD_compressBlock_targetCBlockSize(ZSTD_CCtx* zc,
                               void* dst, size_t dstCapacity,
                               const void* src, size_t srcSize,
                               U32 lastBlock)
{
    size_t cSize = 0;
    const size_t bss = ZSTD_buildSeqStore(zc, src, srcSize);
    DEBUGLOG(5, "ZSTD_compressBlock_targetCBlockSize (dstCapacity=%u, dictLimit=%u, nextToUpdate=%u, srcSize=%zu)",
                (unsigned)dstCapacity, (unsigned)zc->blockState.matchState.window.dictLimit, (unsigned)zc->blockState.matchState.nextToUpdate, srcSize);
    FORWARD_IF_ERROR(bss, "ZSTD_buildSeqStore failed");

    cSize = ZSTD_compressBlock_targetCBlockSize_body(zc, dst, dstCapacity, src, srcSize, bss, lastBlock);
    FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_targetCBlockSize_body failed");

    if (zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
        zc->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;

    return cSize;
}

static void ZSTD_overflowCorrectIfNeeded(ZSTD_MatchState_t* ms,
                                         ZSTD_cwksp* ws,
                                         ZSTD_CCtx_params const* params,
                                         void const* ip,
                                         void const* iend)
{
    U32 const cycleLog = ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy);
    U32 const maxDist = (U32)1 << params->cParams.windowLog;
    if (ZSTD_window_needOverflowCorrection(ms->window, cycleLog, maxDist, ms->loadedDictEnd, ip, iend)) {
        U32 const correction = ZSTD_window_correctOverflow(&ms->window, cycleLog, maxDist, ip);
        ZSTD_STATIC_ASSERT(ZSTD_CHAINLOG_MAX <= 30);
        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX_32 <= 30);
        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
        ZSTD_cwksp_mark_tables_dirty(ws);
        ZSTD_reduceIndex(ms, params, correction);
        ZSTD_cwksp_mark_tables_clean(ws);
        if (ms->nextToUpdate < correction) ms->nextToUpdate = 0;
        else ms->nextToUpdate -= correction;
        /* invalidate dictionaries on overflow correction */
        ms->loadedDictEnd = 0;
        ms->dictMatchState = NULL;
    }
}

/**** skipping file: zstd_preSplit.h ****/

static size_t ZSTD_optimalBlockSize(ZSTD_CCtx* cctx, const void* src, size_t srcSize, size_t blockSizeMax, int splitLevel, ZSTD_strategy strat, S64 savings)
{
    /* split level based on compression strategy, from `fast` to `btultra2` */
    static const int splitLevels[] = { 0, 0, 1, 2, 2, 3, 3, 4, 4, 4 };
    /* note: conservatively only split full blocks (128 KB) currently.
     * While it's possible to go lower, let's keep it simple for a first implementation.
     * Besides, benefits of splitting are reduced when blocks are already small.
     */
    if (srcSize < 128 KB || blockSizeMax < 128 KB)
        return MIN(srcSize, blockSizeMax);
    /* do not split incompressible data though:
     * require verified savings to allow pre-splitting.
     * Note: as a consequence, the first full block is not split.
     */
    if (savings < 3) {
        DEBUGLOG(6, "don't attempt splitting: savings (%i) too low", (int)savings);
        return 128 KB;
    }
    /* apply @splitLevel, or use default value (which depends on @strat).
     * note that splitting heuristic is still conditioned by @savings >= 3,
     * so the first block will not reach this code path */
    if (splitLevel == 1) return 128 KB;
    if (splitLevel == 0) {
        assert(ZSTD_fast <= strat && strat <= ZSTD_btultra2);
        splitLevel = splitLevels[strat];
    } else {
        assert(2 <= splitLevel && splitLevel <= 6);
        splitLevel -= 2;
    }
    return ZSTD_splitBlock(src, blockSizeMax, splitLevel, cctx->tmpWorkspace, cctx->tmpWkspSize);
}

/*! ZSTD_compress_frameChunk() :
*   Compress a chunk of data into one or multiple blocks.
*   All blocks will be terminated, all input will be consumed.
*   Function will issue an error if there is not enough `dstCapacity` to hold the compressed content.
*   Frame is supposed already started (header already produced)
*  @return : compressed size, or an error code
*/
static size_t ZSTD_compress_frameChunk(ZSTD_CCtx* cctx,
                                     void* dst, size_t dstCapacity,
                               const void* src, size_t srcSize,
                                     U32 lastFrameChunk)
{
    size_t blockSizeMax = cctx->blockSizeMax;
    size_t remaining = srcSize;
    const BYTE* ip = (const BYTE*)src;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* op = ostart;
    U32 const maxDist = (U32)1 << cctx->appliedParams.cParams.windowLog;
    S64 savings = (S64)cctx->consumedSrcSize - (S64)cctx->producedCSize;

    assert(cctx->appliedParams.cParams.windowLog <= ZSTD_WINDOWLOG_MAX);

    DEBUGLOG(5, "ZSTD_compress_frameChunk (srcSize=%u, blockSizeMax=%u)", (unsigned)srcSize, (unsigned)blockSizeMax);
    if (cctx->appliedParams.fParams.checksumFlag && srcSize)
        XXH64_update(&cctx->xxhState, src, srcSize);

    while (remaining) {
        ZSTD_MatchState_t* const ms = &cctx->blockState.matchState;
        size_t const blockSize = ZSTD_optimalBlockSize(cctx,
                                ip, remaining,
                                blockSizeMax,
                                cctx->appliedParams.preBlockSplitter_level,
                                cctx->appliedParams.cParams.strategy,
                                savings);
        U32 const lastBlock = lastFrameChunk & (blockSize == remaining);
        assert(blockSize <= remaining);

        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
         * additional 1. We need to revisit and change this logic to be more consistent */
        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize + MIN_CBLOCK_SIZE + 1,
                        dstSize_tooSmall,
                        "not enough space to store compressed block");

        ZSTD_overflowCorrectIfNeeded(
            ms, &cctx->workspace, &cctx->appliedParams, ip, ip + blockSize);
        ZSTD_checkDictValidity(&ms->window, ip + blockSize, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);
        ZSTD_window_enforceMaxDist(&ms->window, ip, maxDist, &ms->loadedDictEnd, &ms->dictMatchState);

        /* Ensure hash/chain table insertion resumes no sooner than lowlimit */
        if (ms->nextToUpdate < ms->window.lowLimit) ms->nextToUpdate = ms->window.lowLimit;

        {   size_t cSize;
            if (ZSTD_useTargetCBlockSize(&cctx->appliedParams)) {
                cSize = ZSTD_compressBlock_targetCBlockSize(cctx, op, dstCapacity, ip, blockSize, lastBlock);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_targetCBlockSize failed");
                assert(cSize > 0);
                assert(cSize <= blockSize + ZSTD_blockHeaderSize);
            } else if (ZSTD_blockSplitterEnabled(&cctx->appliedParams)) {
                cSize = ZSTD_compressBlock_splitBlock(cctx, op, dstCapacity, ip, blockSize, lastBlock);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_splitBlock failed");
                assert(cSize > 0 || cctx->seqCollector.collectSequences == 1);
            } else {
                cSize = ZSTD_compressBlock_internal(cctx,
                                        op+ZSTD_blockHeaderSize, dstCapacity-ZSTD_blockHeaderSize,
                                        ip, blockSize, 1 /* frame */);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressBlock_internal failed");

                if (cSize == 0) {  /* block is not compressible */
                    cSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
                    FORWARD_IF_ERROR(cSize, "ZSTD_noCompressBlock failed");
                } else {
                    U32 const cBlockHeader = cSize == 1 ?
                        lastBlock + (((U32)bt_rle)<<1) + (U32)(blockSize << 3) :
                        lastBlock + (((U32)bt_compressed)<<1) + (U32)(cSize << 3);
                    MEM_writeLE24(op, cBlockHeader);
                    cSize += ZSTD_blockHeaderSize;
                }
            }  /* if (ZSTD_useTargetCBlockSize(&cctx->appliedParams))*/

            /* @savings is employed to ensure that splitting doesn't worsen expansion of incompressible data.
             * Without splitting, the maximum expansion is 3 bytes per full block.
             * An adversarial input could attempt to fudge the split detector,
             * and make it split incompressible data, resulting in more block headers.
             * Note that, since ZSTD_COMPRESSBOUND() assumes a worst case scenario of 1KB per block,
             * and the splitter never creates blocks that small (current lower limit is 8 KB),
             * there is already no risk to expand beyond ZSTD_COMPRESSBOUND() limit.
             * But if the goal is to not expand by more than 3-bytes per 128 KB full block,
             * then yes, it becomes possible to make the block splitter oversplit incompressible data.
             * Using @savings, we enforce an even more conservative condition,
             * requiring the presence of enough savings (at least 3 bytes) to authorize splitting,
             * otherwise only full blocks are used.
             * But being conservative is fine,
             * since splitting barely compressible blocks is not fruitful anyway */
            savings += (S64)blockSize - (S64)cSize;

            ip += blockSize;
            assert(remaining >= blockSize);
            remaining -= blockSize;
            op += cSize;
            assert(dstCapacity >= cSize);
            dstCapacity -= cSize;
            cctx->isFirstBlock = 0;
            DEBUGLOG(5, "ZSTD_compress_frameChunk: adding a block of size %u",
                        (unsigned)cSize);
    }   }

    if (lastFrameChunk && (op>ostart)) cctx->stage = ZSTDcs_ending;
    return (size_t)(op-ostart);
}


static size_t ZSTD_writeFrameHeader(void* dst, size_t dstCapacity,
                                    const ZSTD_CCtx_params* params,
                                    U64 pledgedSrcSize, U32 dictID)
{
    BYTE* const op = (BYTE*)dst;
    U32   const dictIDSizeCodeLength = (dictID>0) + (dictID>=256) + (dictID>=65536);   /* 0-3 */
    U32   const dictIDSizeCode = params->fParams.noDictIDFlag ? 0 : dictIDSizeCodeLength;   /* 0-3 */
    U32   const checksumFlag = params->fParams.checksumFlag>0;
    U32   const windowSize = (U32)1 << params->cParams.windowLog;
    U32   const singleSegment = params->fParams.contentSizeFlag && (windowSize >= pledgedSrcSize);
    BYTE  const windowLogByte = (BYTE)((params->cParams.windowLog - ZSTD_WINDOWLOG_ABSOLUTEMIN) << 3);
    U32   const fcsCode = params->fParams.contentSizeFlag ?
                     (pledgedSrcSize>=256) + (pledgedSrcSize>=65536+256) + (pledgedSrcSize>=0xFFFFFFFFU) : 0;  /* 0-3 */
    BYTE  const frameHeaderDescriptionByte = (BYTE)(dictIDSizeCode + (checksumFlag<<2) + (singleSegment<<5) + (fcsCode<<6) );
    size_t pos=0;

    assert(!(params->fParams.contentSizeFlag && pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN));
    RETURN_ERROR_IF(dstCapacity < ZSTD_FRAMEHEADERSIZE_MAX, dstSize_tooSmall,
                    "dst buf is too small to fit worst-case frame header size.");
    DEBUGLOG(4, "ZSTD_writeFrameHeader : dictIDFlag : %u ; dictID : %u ; dictIDSizeCode : %u",
                !params->fParams.noDictIDFlag, (unsigned)dictID, (unsigned)dictIDSizeCode);
    if (params->format == ZSTD_f_zstd1) {
        MEM_writeLE32(dst, ZSTD_MAGICNUMBER);
        pos = 4;
    }
    op[pos++] = frameHeaderDescriptionByte;
    if (!singleSegment) op[pos++] = windowLogByte;
    switch(dictIDSizeCode)
    {
        default:
            assert(0); /* impossible */
            ZSTD_FALLTHROUGH;
        case 0 : break;
        case 1 : op[pos] = (BYTE)(dictID); pos++; break;
        case 2 : MEM_writeLE16(op+pos, (U16)dictID); pos+=2; break;
        case 3 : MEM_writeLE32(op+pos, dictID); pos+=4; break;
    }
    switch(fcsCode)
    {
        default:
            assert(0); /* impossible */
            ZSTD_FALLTHROUGH;
        case 0 : if (singleSegment) op[pos++] = (BYTE)(pledgedSrcSize); break;
        case 1 : MEM_writeLE16(op+pos, (U16)(pledgedSrcSize-256)); pos+=2; break;
        case 2 : MEM_writeLE32(op+pos, (U32)(pledgedSrcSize)); pos+=4; break;
        case 3 : MEM_writeLE64(op+pos, (U64)(pledgedSrcSize)); pos+=8; break;
    }
    return pos;
}

/* ZSTD_writeSkippableFrame_advanced() :
 * Writes out a skippable frame with the specified magic number variant (16 are supported),
 * from ZSTD_MAGIC_SKIPPABLE_START to ZSTD_MAGIC_SKIPPABLE_START+15, and the desired source data.
 *
 * Returns the total number of bytes written, or a ZSTD error code.
 */
size_t ZSTD_writeSkippableFrame(void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize, unsigned magicVariant) {
    BYTE* op = (BYTE*)dst;
    RETURN_ERROR_IF(dstCapacity < srcSize + ZSTD_SKIPPABLEHEADERSIZE /* Skippable frame overhead */,
                    dstSize_tooSmall, "Not enough room for skippable frame");
    RETURN_ERROR_IF(srcSize > (unsigned)0xFFFFFFFF, srcSize_wrong, "Src size too large for skippable frame");
    RETURN_ERROR_IF(magicVariant > 15, parameter_outOfBound, "Skippable frame magic number variant not supported");

    MEM_writeLE32(op, (U32)(ZSTD_MAGIC_SKIPPABLE_START + magicVariant));
    MEM_writeLE32(op+4, (U32)srcSize);
    ZSTD_memcpy(op+8, src, srcSize);
    return srcSize + ZSTD_SKIPPABLEHEADERSIZE;
}

/* ZSTD_writeLastEmptyBlock() :
 * output an empty Block with end-of-frame mark to complete a frame
 * @return : size of data written into `dst` (== ZSTD_blockHeaderSize (defined in zstd_internal.h))
 *           or an error code if `dstCapacity` is too small (<ZSTD_blockHeaderSize)
 */
size_t ZSTD_writeLastEmptyBlock(void* dst, size_t dstCapacity)
{
    RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall,
                    "dst buf is too small to write frame trailer empty block.");
    {   U32 const cBlockHeader24 = 1 /*lastBlock*/ + (((U32)bt_raw)<<1);  /* 0 size */
        MEM_writeLE24(dst, cBlockHeader24);
        return ZSTD_blockHeaderSize;
    }
}

void ZSTD_referenceExternalSequences(ZSTD_CCtx* cctx, rawSeq* seq, size_t nbSeq)
{
    assert(cctx->stage == ZSTDcs_init);
    assert(nbSeq == 0 || cctx->appliedParams.ldmParams.enableLdm != ZSTD_ps_enable);
    cctx->externSeqStore.seq = seq;
    cctx->externSeqStore.size = nbSeq;
    cctx->externSeqStore.capacity = nbSeq;
    cctx->externSeqStore.pos = 0;
    cctx->externSeqStore.posInSequence = 0;
}


static size_t ZSTD_compressContinue_internal (ZSTD_CCtx* cctx,
                              void* dst, size_t dstCapacity,
                        const void* src, size_t srcSize,
                               U32 frame, U32 lastFrameChunk)
{
    ZSTD_MatchState_t* const ms = &cctx->blockState.matchState;
    size_t fhSize = 0;

    DEBUGLOG(5, "ZSTD_compressContinue_internal, stage: %u, srcSize: %u",
                cctx->stage, (unsigned)srcSize);
    RETURN_ERROR_IF(cctx->stage==ZSTDcs_created, stage_wrong,
                    "missing init (ZSTD_compressBegin)");

    if (frame && (cctx->stage==ZSTDcs_init)) {
        fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, &cctx->appliedParams,
                                       cctx->pledgedSrcSizePlusOne-1, cctx->dictID);
        FORWARD_IF_ERROR(fhSize, "ZSTD_writeFrameHeader failed");
        assert(fhSize <= dstCapacity);
        dstCapacity -= fhSize;
        dst = (char*)dst + fhSize;
        cctx->stage = ZSTDcs_ongoing;
    }

    if (!srcSize) return fhSize;  /* do not generate an empty block if no input */

    if (!ZSTD_window_update(&ms->window, src, srcSize, ms->forceNonContiguous)) {
        ms->forceNonContiguous = 0;
        ms->nextToUpdate = ms->window.dictLimit;
    }
    if (cctx->appliedParams.ldmParams.enableLdm == ZSTD_ps_enable) {
        ZSTD_window_update(&cctx->ldmState.window, src, srcSize, /* forceNonContiguous */ 0);
    }

    if (!frame) {
        /* overflow check and correction for block mode */
        ZSTD_overflowCorrectIfNeeded(
            ms, &cctx->workspace, &cctx->appliedParams,
            src, (BYTE const*)src + srcSize);
    }

    DEBUGLOG(5, "ZSTD_compressContinue_internal (blockSize=%u)", (unsigned)cctx->blockSizeMax);
    {   size_t const cSize = frame ?
                             ZSTD_compress_frameChunk (cctx, dst, dstCapacity, src, srcSize, lastFrameChunk) :
                             ZSTD_compressBlock_internal (cctx, dst, dstCapacity, src, srcSize, 0 /* frame */);
        FORWARD_IF_ERROR(cSize, "%s", frame ? "ZSTD_compress_frameChunk failed" : "ZSTD_compressBlock_internal failed");
        cctx->consumedSrcSize += srcSize;
        cctx->producedCSize += (cSize + fhSize);
        assert(!(cctx->appliedParams.fParams.contentSizeFlag && cctx->pledgedSrcSizePlusOne == 0));
        if (cctx->pledgedSrcSizePlusOne != 0) {  /* control src size */
            ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN == (unsigned long long)-1);
            RETURN_ERROR_IF(
                cctx->consumedSrcSize+1 > cctx->pledgedSrcSizePlusOne,
                srcSize_wrong,
                "error : pledgedSrcSize = %u, while realSrcSize >= %u",
                (unsigned)cctx->pledgedSrcSizePlusOne-1,
                (unsigned)cctx->consumedSrcSize);
        }
        return cSize + fhSize;
    }
}

size_t ZSTD_compressContinue_public(ZSTD_CCtx* cctx,
                                        void* dst, size_t dstCapacity,
                                  const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_compressContinue (srcSize=%u)", (unsigned)srcSize);
    return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 1 /* frame mode */, 0 /* last chunk */);
}

/* NOTE: Must just wrap ZSTD_compressContinue_public() */
size_t ZSTD_compressContinue(ZSTD_CCtx* cctx,
                             void* dst, size_t dstCapacity,
                       const void* src, size_t srcSize)
{
    return ZSTD_compressContinue_public(cctx, dst, dstCapacity, src, srcSize);
}

static size_t ZSTD_getBlockSize_deprecated(const ZSTD_CCtx* cctx)
{
    ZSTD_compressionParameters const cParams = cctx->appliedParams.cParams;
    assert(!ZSTD_checkCParams(cParams));
    return MIN(cctx->appliedParams.maxBlockSize, (size_t)1 << cParams.windowLog);
}

/* NOTE: Must just wrap ZSTD_getBlockSize_deprecated() */
size_t ZSTD_getBlockSize(const ZSTD_CCtx* cctx)
{
    return ZSTD_getBlockSize_deprecated(cctx);
}

/* NOTE: Must just wrap ZSTD_compressBlock_deprecated() */
size_t ZSTD_compressBlock_deprecated(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_compressBlock: srcSize = %u", (unsigned)srcSize);
    { size_t const blockSizeMax = ZSTD_getBlockSize_deprecated(cctx);
      RETURN_ERROR_IF(srcSize > blockSizeMax, srcSize_wrong, "input is larger than a block"); }

    return ZSTD_compressContinue_internal(cctx, dst, dstCapacity, src, srcSize, 0 /* frame mode */, 0 /* last chunk */);
}

/* NOTE: Must just wrap ZSTD_compressBlock_deprecated() */
size_t ZSTD_compressBlock(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    return ZSTD_compressBlock_deprecated(cctx, dst, dstCapacity, src, srcSize);
}

/*! ZSTD_loadDictionaryContent() :
 *  @return : 0, or an error code
 */
static size_t
ZSTD_loadDictionaryContent(ZSTD_MatchState_t* ms,
                        ldmState_t* ls,
                        ZSTD_cwksp* ws,
                        ZSTD_CCtx_params const* params,
                        const void* src, size_t srcSize,
                        ZSTD_dictTableLoadMethod_e dtlm,
                        ZSTD_tableFillPurpose_e tfp)
{
    const BYTE* ip = (const BYTE*) src;
    const BYTE* const iend = ip + srcSize;
    int const loadLdmDict = params->ldmParams.enableLdm == ZSTD_ps_enable && ls != NULL;

    /* Assert that the ms params match the params we're being given */
    ZSTD_assertEqualCParams(params->cParams, ms->cParams);

    {   /* Ensure large dictionaries can't cause index overflow */

        /* Allow the dictionary to set indices up to exactly ZSTD_CURRENT_MAX.
         * Dictionaries right at the edge will immediately trigger overflow
         * correction, but I don't want to insert extra constraints here.
         */
        U32 maxDictSize = ZSTD_CURRENT_MAX - ZSTD_WINDOW_START_INDEX;

        int const CDictTaggedIndices = ZSTD_CDictIndicesAreTagged(&params->cParams);
        if (CDictTaggedIndices && tfp == ZSTD_tfp_forCDict) {
            /* Some dictionary matchfinders in zstd use "short cache",
             * which treats the lower ZSTD_SHORT_CACHE_TAG_BITS of each
             * CDict hashtable entry as a tag rather than as part of an index.
             * When short cache is used, we need to truncate the dictionary
             * so that its indices don't overlap with the tag. */
            U32 const shortCacheMaxDictSize = (1u << (32 - ZSTD_SHORT_CACHE_TAG_BITS)) - ZSTD_WINDOW_START_INDEX;
            maxDictSize = MIN(maxDictSize, shortCacheMaxDictSize);
            assert(!loadLdmDict);
        }

        /* If the dictionary is too large, only load the suffix of the dictionary. */
        if (srcSize > maxDictSize) {
            ip = iend - maxDictSize;
            src = ip;
            srcSize = maxDictSize;
        }
    }

    if (srcSize > ZSTD_CHUNKSIZE_MAX) {
        /* We must have cleared our windows when our source is this large. */
        assert(ZSTD_window_isEmpty(ms->window));
        if (loadLdmDict) assert(ZSTD_window_isEmpty(ls->window));
    }
    ZSTD_window_update(&ms->window, src, srcSize, /* forceNonContiguous */ 0);

    DEBUGLOG(4, "ZSTD_loadDictionaryContent: useRowMatchFinder=%d", (int)params->useRowMatchFinder);

    if (loadLdmDict) { /* Load the entire dict into LDM matchfinders. */
        DEBUGLOG(4, "ZSTD_loadDictionaryContent: Trigger loadLdmDict");
        ZSTD_window_update(&ls->window, src, srcSize, /* forceNonContiguous */ 0);
        ls->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ls->window.base);
        ZSTD_ldm_fillHashTable(ls, ip, iend, &params->ldmParams);
        DEBUGLOG(4, "ZSTD_loadDictionaryContent: ZSTD_ldm_fillHashTable completes");
    }

    /* If the dict is larger than we can reasonably index in our tables, only load the suffix. */
    {   U32 maxDictSize = 1U << MIN(MAX(params->cParams.hashLog + 3, params->cParams.chainLog + 1), 31);
        if (srcSize > maxDictSize) {
            ip = iend - maxDictSize;
            src = ip;
            srcSize = maxDictSize;
        }
    }

    ms->nextToUpdate = (U32)(ip - ms->window.base);
    ms->loadedDictEnd = params->forceWindow ? 0 : (U32)(iend - ms->window.base);
    ms->forceNonContiguous = params->deterministicRefPrefix;

    if (srcSize <= HASH_READ_SIZE) return 0;

    ZSTD_overflowCorrectIfNeeded(ms, ws, params, ip, iend);

    switch(params->cParams.strategy)
    {
    case ZSTD_fast:
        ZSTD_fillHashTable(ms, iend, dtlm, tfp);
        break;
    case ZSTD_dfast:
#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
        ZSTD_fillDoubleHashTable(ms, iend, dtlm, tfp);
#else
        assert(0); /* shouldn't be called: cparams should've been adjusted. */
#endif
        break;

    case ZSTD_greedy:
    case ZSTD_lazy:
    case ZSTD_lazy2:
#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR)
        assert(srcSize >= HASH_READ_SIZE);
        if (ms->dedicatedDictSearch) {
            assert(ms->chainTable != NULL);
            ZSTD_dedicatedDictSearch_lazy_loadDictionary(ms, iend-HASH_READ_SIZE);
        } else {
            assert(params->useRowMatchFinder != ZSTD_ps_auto);
            if (params->useRowMatchFinder == ZSTD_ps_enable) {
                size_t const tagTableSize = ((size_t)1 << params->cParams.hashLog);
                ZSTD_memset(ms->tagTable, 0, tagTableSize);
                ZSTD_row_update(ms, iend-HASH_READ_SIZE);
                DEBUGLOG(4, "Using row-based hash table for lazy dict");
            } else {
                ZSTD_insertAndFindFirstIndex(ms, iend-HASH_READ_SIZE);
                DEBUGLOG(4, "Using chain-based hash table for lazy dict");
            }
        }
#else
        assert(0); /* shouldn't be called: cparams should've been adjusted. */
#endif
        break;

    case ZSTD_btlazy2:   /* we want the dictionary table fully sorted */
    case ZSTD_btopt:
    case ZSTD_btultra:
    case ZSTD_btultra2:
#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)
        assert(srcSize >= HASH_READ_SIZE);
        DEBUGLOG(4, "Fill %u bytes into the Binary Tree", (unsigned)srcSize);
        ZSTD_updateTree(ms, iend-HASH_READ_SIZE, iend);
#else
        assert(0); /* shouldn't be called: cparams should've been adjusted. */
#endif
        break;

    default:
        assert(0);  /* not possible : not a valid strategy id */
    }

    ms->nextToUpdate = (U32)(iend - ms->window.base);
    return 0;
}


/* Dictionaries that assign zero probability to symbols that show up causes problems
 * when FSE encoding. Mark dictionaries with zero probability symbols as FSE_repeat_check
 * and only dictionaries with 100% valid symbols can be assumed valid.
 */
static FSE_repeat ZSTD_dictNCountRepeat(short* normalizedCounter, unsigned dictMaxSymbolValue, unsigned maxSymbolValue)
{
    U32 s;
    if (dictMaxSymbolValue < maxSymbolValue) {
        return FSE_repeat_check;
    }
    for (s = 0; s <= maxSymbolValue; ++s) {
        if (normalizedCounter[s] == 0) {
            return FSE_repeat_check;
        }
    }
    return FSE_repeat_valid;
}

size_t ZSTD_loadCEntropy(ZSTD_compressedBlockState_t* bs, void* workspace,
                         const void* const dict, size_t dictSize)
{
    short offcodeNCount[MaxOff+1];
    unsigned offcodeMaxValue = MaxOff;
    const BYTE* dictPtr = (const BYTE*)dict;    /* skip magic num and dict ID */
    const BYTE* const dictEnd = dictPtr + dictSize;
    dictPtr += 8;
    bs->entropy.huf.repeatMode = HUF_repeat_check;

    {   unsigned maxSymbolValue = 255;
        unsigned hasZeroWeights = 1;
        size_t const hufHeaderSize = HUF_readCTable((HUF_CElt*)bs->entropy.huf.CTable, &maxSymbolValue, dictPtr,
            (size_t)(dictEnd-dictPtr), &hasZeroWeights);

        /* We only set the loaded table as valid if it contains all non-zero
         * weights. Otherwise, we set it to check */
        if (!hasZeroWeights && maxSymbolValue == 255)
            bs->entropy.huf.repeatMode = HUF_repeat_valid;

        RETURN_ERROR_IF(HUF_isError(hufHeaderSize), dictionary_corrupted, "");
        dictPtr += hufHeaderSize;
    }

    {   unsigned offcodeLog;
        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, "");
        /* fill all offset symbols to avoid garbage at end of table */
        RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                bs->entropy.fse.offcodeCTable,
                offcodeNCount, MaxOff, offcodeLog,
                workspace, HUF_WORKSPACE_SIZE)),
            dictionary_corrupted, "");
        /* Defer checking offcodeMaxValue because we need to know the size of the dictionary content */
        dictPtr += offcodeHeaderSize;
    }

    {   short matchlengthNCount[MaxML+1];
        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, "");
        RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                bs->entropy.fse.matchlengthCTable,
                matchlengthNCount, matchlengthMaxValue, matchlengthLog,
                workspace, HUF_WORKSPACE_SIZE)),
            dictionary_corrupted, "");
        bs->entropy.fse.matchlength_repeatMode = ZSTD_dictNCountRepeat(matchlengthNCount, matchlengthMaxValue, MaxML);
        dictPtr += matchlengthHeaderSize;
    }

    {   short litlengthNCount[MaxLL+1];
        unsigned litlengthMaxValue = MaxLL, litlengthLog;
        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, "");
        RETURN_ERROR_IF(FSE_isError(FSE_buildCTable_wksp(
                bs->entropy.fse.litlengthCTable,
                litlengthNCount, litlengthMaxValue, litlengthLog,
                workspace, HUF_WORKSPACE_SIZE)),
            dictionary_corrupted, "");
        bs->entropy.fse.litlength_repeatMode = ZSTD_dictNCountRepeat(litlengthNCount, litlengthMaxValue, MaxLL);
        dictPtr += litlengthHeaderSize;
    }

    RETURN_ERROR_IF(dictPtr+12 > dictEnd, dictionary_corrupted, "");
    bs->rep[0] = MEM_readLE32(dictPtr+0);
    bs->rep[1] = MEM_readLE32(dictPtr+4);
    bs->rep[2] = MEM_readLE32(dictPtr+8);
    dictPtr += 12;

    {   size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
        U32 offcodeMax = MaxOff;
        if (dictContentSize <= ((U32)-1) - 128 KB) {
            U32 const maxOffset = (U32)dictContentSize + 128 KB; /* The maximum offset that must be supported */
            offcodeMax = ZSTD_highbit32(maxOffset); /* Calculate minimum offset code required to represent maxOffset */
        }
        /* All offset values <= dictContentSize + 128 KB must be representable for a valid table */
        bs->entropy.fse.offcode_repeatMode = ZSTD_dictNCountRepeat(offcodeNCount, offcodeMaxValue, MIN(offcodeMax, MaxOff));

        /* All repCodes must be <= dictContentSize and != 0 */
        {   U32 u;
            for (u=0; u<3; u++) {
                RETURN_ERROR_IF(bs->rep[u] == 0, dictionary_corrupted, "");
                RETURN_ERROR_IF(bs->rep[u] > dictContentSize, dictionary_corrupted, "");
    }   }   }

    return (size_t)(dictPtr - (const BYTE*)dict);
}

/* Dictionary format :
 * See :
 * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#dictionary-format
 */
/*! ZSTD_loadZstdDictionary() :
 * @return : dictID, or an error code
 *  assumptions : magic number supposed already checked
 *                dictSize supposed >= 8
 */
static size_t ZSTD_loadZstdDictionary(ZSTD_compressedBlockState_t* bs,
                                      ZSTD_MatchState_t* ms,
                                      ZSTD_cwksp* ws,
                                      ZSTD_CCtx_params const* params,
                                      const void* dict, size_t dictSize,
                                      ZSTD_dictTableLoadMethod_e dtlm,
                                      ZSTD_tableFillPurpose_e tfp,
                                      void* workspace)
{
    const BYTE* dictPtr = (const BYTE*)dict;
    const BYTE* const dictEnd = dictPtr + dictSize;
    size_t dictID;
    size_t eSize;
    ZSTD_STATIC_ASSERT(HUF_WORKSPACE_SIZE >= (1<<MAX(MLFSELog,LLFSELog)));
    assert(dictSize >= 8);
    assert(MEM_readLE32(dictPtr) == ZSTD_MAGIC_DICTIONARY);

    dictID = params->fParams.noDictIDFlag ? 0 :  MEM_readLE32(dictPtr + 4 /* skip magic number */ );
    eSize = ZSTD_loadCEntropy(bs, workspace, dict, dictSize);
    FORWARD_IF_ERROR(eSize, "ZSTD_loadCEntropy failed");
    dictPtr += eSize;

    {
        size_t const dictContentSize = (size_t)(dictEnd - dictPtr);
        FORWARD_IF_ERROR(ZSTD_loadDictionaryContent(
            ms, NULL, ws, params, dictPtr, dictContentSize, dtlm, tfp), "");
    }
    return dictID;
}

/** ZSTD_compress_insertDictionary() :
*   @return : dictID, or an error code */
static size_t
ZSTD_compress_insertDictionary(ZSTD_compressedBlockState_t* bs,
                               ZSTD_MatchState_t* ms,
                               ldmState_t* ls,
                               ZSTD_cwksp* ws,
                         const ZSTD_CCtx_params* params,
                         const void* dict, size_t dictSize,
                               ZSTD_dictContentType_e dictContentType,
                               ZSTD_dictTableLoadMethod_e dtlm,
                               ZSTD_tableFillPurpose_e tfp,
                               void* workspace)
{
    DEBUGLOG(4, "ZSTD_compress_insertDictionary (dictSize=%u)", (U32)dictSize);
    if ((dict==NULL) || (dictSize<8)) {
        RETURN_ERROR_IF(dictContentType == ZSTD_dct_fullDict, dictionary_wrong, "");
        return 0;
    }

    ZSTD_reset_compressedBlockState(bs);

    /* dict restricted modes */
    if (dictContentType == ZSTD_dct_rawContent)
        return ZSTD_loadDictionaryContent(ms, ls, ws, params, dict, dictSize, dtlm, tfp);

    if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) {
        if (dictContentType == ZSTD_dct_auto) {
            DEBUGLOG(4, "raw content dictionary detected");
            return ZSTD_loadDictionaryContent(
                ms, ls, ws, params, dict, dictSize, dtlm, tfp);
        }
        RETURN_ERROR_IF(dictContentType == ZSTD_dct_fullDict, dictionary_wrong, "");
        assert(0);   /* impossible */
    }

    /* dict as full zstd dictionary */
    return ZSTD_loadZstdDictionary(
        bs, ms, ws, params, dict, dictSize, dtlm, tfp, workspace);
}

#define ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF (128 KB)
#define ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER (6ULL)

/*! ZSTD_compressBegin_internal() :
 * Assumption : either @dict OR @cdict (or none) is non-NULL, never both
 * @return : 0, or an error code */
static size_t ZSTD_compressBegin_internal(ZSTD_CCtx* cctx,
                                    const void* dict, size_t dictSize,
                                    ZSTD_dictContentType_e dictContentType,
                                    ZSTD_dictTableLoadMethod_e dtlm,
                                    const ZSTD_CDict* cdict,
                                    const ZSTD_CCtx_params* params, U64 pledgedSrcSize,
                                    ZSTD_buffered_policy_e zbuff)
{
    size_t const dictContentSize = cdict ? cdict->dictContentSize : dictSize;
#if ZSTD_TRACE
    cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
#endif
    DEBUGLOG(4, "ZSTD_compressBegin_internal: wlog=%u", params->cParams.windowLog);
    /* params are supposed to be fully validated at this point */
    assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
    if ( (cdict)
      && (cdict->dictContentSize > 0)
      && ( pledgedSrcSize < ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF
        || pledgedSrcSize < cdict->dictContentSize * ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER
        || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
        || cdict->compressionLevel == 0)
      && (params->attachDictPref != ZSTD_dictForceLoad) ) {
        return ZSTD_resetCCtx_usingCDict(cctx, cdict, params, pledgedSrcSize, zbuff);
    }

    FORWARD_IF_ERROR( ZSTD_resetCCtx_internal(cctx, params, pledgedSrcSize,
                                     dictContentSize,
                                     ZSTDcrp_makeClean, zbuff) , "");
    {   size_t const dictID = cdict ?
                ZSTD_compress_insertDictionary(
                        cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                        &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, cdict->dictContent,
                        cdict->dictContentSize, cdict->dictContentType, dtlm,
                        ZSTD_tfp_forCCtx, cctx->tmpWorkspace)
              : ZSTD_compress_insertDictionary(
                        cctx->blockState.prevCBlock, &cctx->blockState.matchState,
                        &cctx->ldmState, &cctx->workspace, &cctx->appliedParams, dict, dictSize,
                        dictContentType, dtlm, ZSTD_tfp_forCCtx, cctx->tmpWorkspace);
        FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
        assert(dictID <= UINT_MAX);
        cctx->dictID = (U32)dictID;
        cctx->dictContentSize = dictContentSize;
    }
    return 0;
}

size_t ZSTD_compressBegin_advanced_internal(ZSTD_CCtx* cctx,
                                    const void* dict, size_t dictSize,
                                    ZSTD_dictContentType_e dictContentType,
                                    ZSTD_dictTableLoadMethod_e dtlm,
                                    const ZSTD_CDict* cdict,
                                    const ZSTD_CCtx_params* params,
                                    unsigned long long pledgedSrcSize)
{
    DEBUGLOG(4, "ZSTD_compressBegin_advanced_internal: wlog=%u", params->cParams.windowLog);
    /* compression parameters verification and optimization */
    FORWARD_IF_ERROR( ZSTD_checkCParams(params->cParams) , "");
    return ZSTD_compressBegin_internal(cctx,
                                       dict, dictSize, dictContentType, dtlm,
                                       cdict,
                                       params, pledgedSrcSize,
                                       ZSTDb_not_buffered);
}

/*! ZSTD_compressBegin_advanced() :
*   @return : 0, or an error code */
size_t ZSTD_compressBegin_advanced(ZSTD_CCtx* cctx,
                             const void* dict, size_t dictSize,
                                   ZSTD_parameters params, unsigned long long pledgedSrcSize)
{
    ZSTD_CCtx_params cctxParams;
    ZSTD_CCtxParams_init_internal(&cctxParams, &params, ZSTD_NO_CLEVEL);
    return ZSTD_compressBegin_advanced_internal(cctx,
                                            dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast,
                                            NULL /*cdict*/,
                                            &cctxParams, pledgedSrcSize);
}

static size_t
ZSTD_compressBegin_usingDict_deprecated(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
{
    ZSTD_CCtx_params cctxParams;
    {   ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_noAttachDict);
        ZSTD_CCtxParams_init_internal(&cctxParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel);
    }
    DEBUGLOG(4, "ZSTD_compressBegin_usingDict (dictSize=%u)", (unsigned)dictSize);
    return ZSTD_compressBegin_internal(cctx, dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL,
                                       &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, ZSTDb_not_buffered);
}

size_t
ZSTD_compressBegin_usingDict(ZSTD_CCtx* cctx, const void* dict, size_t dictSize, int compressionLevel)
{
    return ZSTD_compressBegin_usingDict_deprecated(cctx, dict, dictSize, compressionLevel);
}

size_t ZSTD_compressBegin(ZSTD_CCtx* cctx, int compressionLevel)
{
    return ZSTD_compressBegin_usingDict_deprecated(cctx, NULL, 0, compressionLevel);
}


/*! ZSTD_writeEpilogue() :
*   Ends a frame.
*   @return : nb of bytes written into dst (or an error code) */
static size_t ZSTD_writeEpilogue(ZSTD_CCtx* cctx, void* dst, size_t dstCapacity)
{
    BYTE* const ostart = (BYTE*)dst;
    BYTE* op = ostart;

    DEBUGLOG(4, "ZSTD_writeEpilogue");
    RETURN_ERROR_IF(cctx->stage == ZSTDcs_created, stage_wrong, "init missing");

    /* special case : empty frame */
    if (cctx->stage == ZSTDcs_init) {
        size_t fhSize = ZSTD_writeFrameHeader(dst, dstCapacity, &cctx->appliedParams, 0, 0);
        FORWARD_IF_ERROR(fhSize, "ZSTD_writeFrameHeader failed");
        dstCapacity -= fhSize;
        op += fhSize;
        cctx->stage = ZSTDcs_ongoing;
    }

    if (cctx->stage != ZSTDcs_ending) {
        /* write one last empty block, make it the "last" block */
        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1) + 0;
        ZSTD_STATIC_ASSERT(ZSTD_BLOCKHEADERSIZE == 3);
        RETURN_ERROR_IF(dstCapacity<3, dstSize_tooSmall, "no room for epilogue");
        MEM_writeLE24(op, cBlockHeader24);
        op += ZSTD_blockHeaderSize;
        dstCapacity -= ZSTD_blockHeaderSize;
    }

    if (cctx->appliedParams.fParams.checksumFlag) {
        U32 const checksum = (U32) XXH64_digest(&cctx->xxhState);
        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for checksum");
        DEBUGLOG(4, "ZSTD_writeEpilogue: write checksum : %08X", (unsigned)checksum);
        MEM_writeLE32(op, checksum);
        op += 4;
    }

    cctx->stage = ZSTDcs_created;  /* return to "created but no init" status */
    return (size_t)(op-ostart);
}

void ZSTD_CCtx_trace(ZSTD_CCtx* cctx, size_t extraCSize)
{
#if ZSTD_TRACE
    if (cctx->traceCtx && ZSTD_trace_compress_end != NULL) {
        int const streaming = cctx->inBuffSize > 0 || cctx->outBuffSize > 0 || cctx->appliedParams.nbWorkers > 0;
        ZSTD_Trace trace;
        ZSTD_memset(&trace, 0, sizeof(trace));
        trace.version = ZSTD_VERSION_NUMBER;
        trace.streaming = streaming;
        trace.dictionaryID = cctx->dictID;
        trace.dictionarySize = cctx->dictContentSize;
        trace.uncompressedSize = cctx->consumedSrcSize;
        trace.compressedSize = cctx->producedCSize + extraCSize;
        trace.params = &cctx->appliedParams;
        trace.cctx = cctx;
        ZSTD_trace_compress_end(cctx->traceCtx, &trace);
    }
    cctx->traceCtx = 0;
#else
    (void)cctx;
    (void)extraCSize;
#endif
}

size_t ZSTD_compressEnd_public(ZSTD_CCtx* cctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize)
{
    size_t endResult;
    size_t const cSize = ZSTD_compressContinue_internal(cctx,
                                dst, dstCapacity, src, srcSize,
                                1 /* frame mode */, 1 /* last chunk */);
    FORWARD_IF_ERROR(cSize, "ZSTD_compressContinue_internal failed");
    endResult = ZSTD_writeEpilogue(cctx, (char*)dst + cSize, dstCapacity-cSize);
    FORWARD_IF_ERROR(endResult, "ZSTD_writeEpilogue failed");
    assert(!(cctx->appliedParams.fParams.contentSizeFlag && cctx->pledgedSrcSizePlusOne == 0));
    if (cctx->pledgedSrcSizePlusOne != 0) {  /* control src size */
        ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_UNKNOWN == (unsigned long long)-1);
        DEBUGLOG(4, "end of frame : controlling src size");
        RETURN_ERROR_IF(
            cctx->pledgedSrcSizePlusOne != cctx->consumedSrcSize+1,
            srcSize_wrong,
             "error : pledgedSrcSize = %u, while realSrcSize = %u",
            (unsigned)cctx->pledgedSrcSizePlusOne-1,
            (unsigned)cctx->consumedSrcSize);
    }
    ZSTD_CCtx_trace(cctx, endResult);
    return cSize + endResult;
}

/* NOTE: Must just wrap ZSTD_compressEnd_public() */
size_t ZSTD_compressEnd(ZSTD_CCtx* cctx,
                        void* dst, size_t dstCapacity,
                  const void* src, size_t srcSize)
{
    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
}

size_t ZSTD_compress_advanced (ZSTD_CCtx* cctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                         const void* dict,size_t dictSize,
                               ZSTD_parameters params)
{
    DEBUGLOG(4, "ZSTD_compress_advanced");
    FORWARD_IF_ERROR(ZSTD_checkCParams(params.cParams), "");
    ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, ZSTD_NO_CLEVEL);
    return ZSTD_compress_advanced_internal(cctx,
                                           dst, dstCapacity,
                                           src, srcSize,
                                           dict, dictSize,
                                           &cctx->simpleApiParams);
}

/* Internal */
size_t ZSTD_compress_advanced_internal(
        ZSTD_CCtx* cctx,
        void* dst, size_t dstCapacity,
        const void* src, size_t srcSize,
        const void* dict,size_t dictSize,
        const ZSTD_CCtx_params* params)
{
    DEBUGLOG(4, "ZSTD_compress_advanced_internal (srcSize:%u)", (unsigned)srcSize);
    FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx,
                         dict, dictSize, ZSTD_dct_auto, ZSTD_dtlm_fast, NULL,
                         params, srcSize, ZSTDb_not_buffered) , "");
    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
}

size_t ZSTD_compress_usingDict(ZSTD_CCtx* cctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize,
                         const void* dict, size_t dictSize,
                               int compressionLevel)
{
    {
        ZSTD_parameters const params = ZSTD_getParams_internal(compressionLevel, srcSize, dict ? dictSize : 0, ZSTD_cpm_noAttachDict);
        assert(params.fParams.contentSizeFlag == 1);
        ZSTD_CCtxParams_init_internal(&cctx->simpleApiParams, &params, (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT: compressionLevel);
    }
    DEBUGLOG(4, "ZSTD_compress_usingDict (srcSize=%u)", (unsigned)srcSize);
    return ZSTD_compress_advanced_internal(cctx, dst, dstCapacity, src, srcSize, dict, dictSize, &cctx->simpleApiParams);
}

size_t ZSTD_compressCCtx(ZSTD_CCtx* cctx,
                         void* dst, size_t dstCapacity,
                   const void* src, size_t srcSize,
                         int compressionLevel)
{
    DEBUGLOG(4, "ZSTD_compressCCtx (srcSize=%u)", (unsigned)srcSize);
    assert(cctx != NULL);
    return ZSTD_compress_usingDict(cctx, dst, dstCapacity, src, srcSize, NULL, 0, compressionLevel);
}

size_t ZSTD_compress(void* dst, size_t dstCapacity,
               const void* src, size_t srcSize,
                     int compressionLevel)
{
    size_t result;
#if ZSTD_COMPRESS_HEAPMODE
    ZSTD_CCtx* cctx = ZSTD_createCCtx();
    RETURN_ERROR_IF(!cctx, memory_allocation, "ZSTD_createCCtx failed");
    result = ZSTD_compressCCtx(cctx, dst, dstCapacity, src, srcSize, compressionLevel);
    ZSTD_freeCCtx(cctx);
#else
    ZSTD_CCtx ctxBody;
    ZSTD_initCCtx(&ctxBody, ZSTD_defaultCMem);
    result = ZSTD_compressCCtx(&ctxBody, dst, dstCapacity, src, srcSize, compressionLevel);
    ZSTD_freeCCtxContent(&ctxBody);   /* can't free ctxBody itself, as it's on stack; free only heap content */
#endif
    return result;
}


/* =====  Dictionary API  ===== */

/*! ZSTD_estimateCDictSize_advanced() :
 *  Estimate amount of memory that will be needed to create a dictionary with following arguments */
size_t ZSTD_estimateCDictSize_advanced(
        size_t dictSize, ZSTD_compressionParameters cParams,
        ZSTD_dictLoadMethod_e dictLoadMethod)
{
    DEBUGLOG(5, "sizeof(ZSTD_CDict) : %u", (unsigned)sizeof(ZSTD_CDict));
    return ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
         + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE)
         /* enableDedicatedDictSearch == 1 ensures that CDict estimation will not be too small
          * in case we are using DDS with row-hash. */
         + ZSTD_sizeof_matchState(&cParams, ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams),
                                  /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0)
         + (dictLoadMethod == ZSTD_dlm_byRef ? 0
            : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void *))));
}

size_t ZSTD_estimateCDictSize(size_t dictSize, int compressionLevel)
{
    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
    return ZSTD_estimateCDictSize_advanced(dictSize, cParams, ZSTD_dlm_byCopy);
}

size_t ZSTD_sizeof_CDict(const ZSTD_CDict* cdict)
{
    if (cdict==NULL) return 0;   /* support sizeof on NULL */
    DEBUGLOG(5, "sizeof(*cdict) : %u", (unsigned)sizeof(*cdict));
    /* cdict may be in the workspace */
    return (cdict->workspace.workspace == cdict ? 0 : sizeof(*cdict))
        + ZSTD_cwksp_sizeof(&cdict->workspace);
}

static size_t ZSTD_initCDict_internal(
                    ZSTD_CDict* cdict,
              const void* dictBuffer, size_t dictSize,
                    ZSTD_dictLoadMethod_e dictLoadMethod,
                    ZSTD_dictContentType_e dictContentType,
                    ZSTD_CCtx_params params)
{
    DEBUGLOG(3, "ZSTD_initCDict_internal (dictContentType:%u)", (unsigned)dictContentType);
    assert(!ZSTD_checkCParams(params.cParams));
    cdict->matchState.cParams = params.cParams;
    cdict->matchState.dedicatedDictSearch = params.enableDedicatedDictSearch;
    if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dictBuffer) || (!dictSize)) {
        cdict->dictContent = dictBuffer;
    } else {
         void *internalBuffer = ZSTD_cwksp_reserve_object(&cdict->workspace, ZSTD_cwksp_align(dictSize, sizeof(void*)));
        RETURN_ERROR_IF(!internalBuffer, memory_allocation, "NULL pointer!");
        cdict->dictContent = internalBuffer;
        ZSTD_memcpy(internalBuffer, dictBuffer, dictSize);
    }
    cdict->dictContentSize = dictSize;
    cdict->dictContentType = dictContentType;

    cdict->entropyWorkspace = (U32*)ZSTD_cwksp_reserve_object(&cdict->workspace, HUF_WORKSPACE_SIZE);


    /* Reset the state to no dictionary */
    ZSTD_reset_compressedBlockState(&cdict->cBlockState);
    FORWARD_IF_ERROR(ZSTD_reset_matchState(
        &cdict->matchState,
        &cdict->workspace,
        &params.cParams,
        params.useRowMatchFinder,
        ZSTDcrp_makeClean,
        ZSTDirp_reset,
        ZSTD_resetTarget_CDict), "");
    /* (Maybe) load the dictionary
     * Skips loading the dictionary if it is < 8 bytes.
     */
    {   params.compressionLevel = ZSTD_CLEVEL_DEFAULT;
        params.fParams.contentSizeFlag = 1;
        {   size_t const dictID = ZSTD_compress_insertDictionary(
                    &cdict->cBlockState, &cdict->matchState, NULL, &cdict->workspace,
                    &params, cdict->dictContent, cdict->dictContentSize,
                    dictContentType, ZSTD_dtlm_full, ZSTD_tfp_forCDict, cdict->entropyWorkspace);
            FORWARD_IF_ERROR(dictID, "ZSTD_compress_insertDictionary failed");
            assert(dictID <= (size_t)(U32)-1);
            cdict->dictID = (U32)dictID;
        }
    }

    return 0;
}

static ZSTD_CDict*
ZSTD_createCDict_advanced_internal(size_t dictSize,
                                ZSTD_dictLoadMethod_e dictLoadMethod,
                                ZSTD_compressionParameters cParams,
                                ZSTD_ParamSwitch_e useRowMatchFinder,
                                int enableDedicatedDictSearch,
                                ZSTD_customMem customMem)
{
    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;
    DEBUGLOG(3, "ZSTD_createCDict_advanced_internal (dictSize=%u)", (unsigned)dictSize);

    {   size_t const workspaceSize =
            ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict)) +
            ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE) +
            ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, enableDedicatedDictSearch, /* forCCtx */ 0) +
            (dictLoadMethod == ZSTD_dlm_byRef ? 0
             : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))));
        void* const workspace = ZSTD_customMalloc(workspaceSize, customMem);
        ZSTD_cwksp ws;
        ZSTD_CDict* cdict;

        if (!workspace) {
            ZSTD_customFree(workspace, customMem);
            return NULL;
        }

        ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_dynamic_alloc);

        cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict));
        assert(cdict != NULL);
        ZSTD_cwksp_move(&cdict->workspace, &ws);
        cdict->customMem = customMem;
        cdict->compressionLevel = ZSTD_NO_CLEVEL; /* signals advanced API usage */
        cdict->useRowMatchFinder = useRowMatchFinder;
        return cdict;
    }
}

ZSTD_CDict* ZSTD_createCDict_advanced(const void* dictBuffer, size_t dictSize,
                                      ZSTD_dictLoadMethod_e dictLoadMethod,
                                      ZSTD_dictContentType_e dictContentType,
                                      ZSTD_compressionParameters cParams,
                                      ZSTD_customMem customMem)
{
    ZSTD_CCtx_params cctxParams;
    ZSTD_memset(&cctxParams, 0, sizeof(cctxParams));
    DEBUGLOG(3, "ZSTD_createCDict_advanced, dictSize=%u, mode=%u", (unsigned)dictSize, (unsigned)dictContentType);
    ZSTD_CCtxParams_init(&cctxParams, 0);
    cctxParams.cParams = cParams;
    cctxParams.customMem = customMem;
    return ZSTD_createCDict_advanced2(
        dictBuffer, dictSize,
        dictLoadMethod, dictContentType,
        &cctxParams, customMem);
}

ZSTD_CDict* ZSTD_createCDict_advanced2(
        const void* dict, size_t dictSize,
        ZSTD_dictLoadMethod_e dictLoadMethod,
        ZSTD_dictContentType_e dictContentType,
        const ZSTD_CCtx_params* originalCctxParams,
        ZSTD_customMem customMem)
{
    ZSTD_CCtx_params cctxParams = *originalCctxParams;
    ZSTD_compressionParameters cParams;
    ZSTD_CDict* cdict;

    DEBUGLOG(3, "ZSTD_createCDict_advanced2, dictSize=%u, mode=%u", (unsigned)dictSize, (unsigned)dictContentType);
    if (!customMem.customAlloc ^ !customMem.customFree) return NULL;

    if (cctxParams.enableDedicatedDictSearch) {
        cParams = ZSTD_dedicatedDictSearch_getCParams(
            cctxParams.compressionLevel, dictSize);
        ZSTD_overrideCParams(&cParams, &cctxParams.cParams);
    } else {
        cParams = ZSTD_getCParamsFromCCtxParams(
            &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
    }

    if (!ZSTD_dedicatedDictSearch_isSupported(&cParams)) {
        /* Fall back to non-DDSS params */
        cctxParams.enableDedicatedDictSearch = 0;
        cParams = ZSTD_getCParamsFromCCtxParams(
            &cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
    }

    DEBUGLOG(3, "ZSTD_createCDict_advanced2: DedicatedDictSearch=%u", cctxParams.enableDedicatedDictSearch);
    cctxParams.cParams = cParams;
    cctxParams.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(cctxParams.useRowMatchFinder, &cParams);

    cdict = ZSTD_createCDict_advanced_internal(dictSize,
                        dictLoadMethod, cctxParams.cParams,
                        cctxParams.useRowMatchFinder, cctxParams.enableDedicatedDictSearch,
                        customMem);

    if (!cdict || ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                    dict, dictSize,
                                    dictLoadMethod, dictContentType,
                                    cctxParams) )) {
        ZSTD_freeCDict(cdict);
        return NULL;
    }

    return cdict;
}

ZSTD_CDict* ZSTD_createCDict(const void* dict, size_t dictSize, int compressionLevel)
{
    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
    ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize,
                                                  ZSTD_dlm_byCopy, ZSTD_dct_auto,
                                                  cParams, ZSTD_defaultCMem);
    if (cdict)
        cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel;
    return cdict;
}

ZSTD_CDict* ZSTD_createCDict_byReference(const void* dict, size_t dictSize, int compressionLevel)
{
    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, ZSTD_CONTENTSIZE_UNKNOWN, dictSize, ZSTD_cpm_createCDict);
    ZSTD_CDict* const cdict = ZSTD_createCDict_advanced(dict, dictSize,
                                     ZSTD_dlm_byRef, ZSTD_dct_auto,
                                     cParams, ZSTD_defaultCMem);
    if (cdict)
        cdict->compressionLevel = (compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : compressionLevel;
    return cdict;
}

size_t ZSTD_freeCDict(ZSTD_CDict* cdict)
{
    if (cdict==NULL) return 0;   /* support free on NULL */
    {   ZSTD_customMem const cMem = cdict->customMem;
        int cdictInWorkspace = ZSTD_cwksp_owns_buffer(&cdict->workspace, cdict);
        ZSTD_cwksp_free(&cdict->workspace, cMem);
        if (!cdictInWorkspace) {
            ZSTD_customFree(cdict, cMem);
        }
        return 0;
    }
}

/*! ZSTD_initStaticCDict_advanced() :
 *  Generate a digested dictionary in provided memory area.
 *  workspace: The memory area to emplace the dictionary into.
 *             Provided pointer must 8-bytes aligned.
 *             It must outlive dictionary usage.
 *  workspaceSize: Use ZSTD_estimateCDictSize()
 *                 to determine how large workspace must be.
 *  cParams : use ZSTD_getCParams() to transform a compression level
 *            into its relevant cParams.
 * @return : pointer to ZSTD_CDict*, or NULL if error (size too small)
 *  Note : there is no corresponding "free" function.
 *         Since workspace was allocated externally, it must be freed externally.
 */
const ZSTD_CDict* ZSTD_initStaticCDict(
                                 void* workspace, size_t workspaceSize,
                           const void* dict, size_t dictSize,
                                 ZSTD_dictLoadMethod_e dictLoadMethod,
                                 ZSTD_dictContentType_e dictContentType,
                                 ZSTD_compressionParameters cParams)
{
    ZSTD_ParamSwitch_e const useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(ZSTD_ps_auto, &cParams);
    /* enableDedicatedDictSearch == 1 ensures matchstate is not too small in case this CDict will be used for DDS + row hash */
    size_t const matchStateSize = ZSTD_sizeof_matchState(&cParams, useRowMatchFinder, /* enableDedicatedDictSearch */ 1, /* forCCtx */ 0);
    size_t const neededSize = ZSTD_cwksp_alloc_size(sizeof(ZSTD_CDict))
                            + (dictLoadMethod == ZSTD_dlm_byRef ? 0
                               : ZSTD_cwksp_alloc_size(ZSTD_cwksp_align(dictSize, sizeof(void*))))
                            + ZSTD_cwksp_alloc_size(HUF_WORKSPACE_SIZE)
                            + matchStateSize;
    ZSTD_CDict* cdict;
    ZSTD_CCtx_params params;

    DEBUGLOG(4, "ZSTD_initStaticCDict (dictSize==%u)", (unsigned)dictSize);
    if ((size_t)workspace & 7) return NULL;  /* 8-aligned */

    {
        ZSTD_cwksp ws;
        ZSTD_cwksp_init(&ws, workspace, workspaceSize, ZSTD_cwksp_static_alloc);
        cdict = (ZSTD_CDict*)ZSTD_cwksp_reserve_object(&ws, sizeof(ZSTD_CDict));
        if (cdict == NULL) return NULL;
        ZSTD_cwksp_move(&cdict->workspace, &ws);
    }

    if (workspaceSize < neededSize) return NULL;

    ZSTD_CCtxParams_init(&params, 0);
    params.cParams = cParams;
    params.useRowMatchFinder = useRowMatchFinder;
    cdict->useRowMatchFinder = useRowMatchFinder;
    cdict->compressionLevel = ZSTD_NO_CLEVEL;

    if (ZSTD_isError( ZSTD_initCDict_internal(cdict,
                                              dict, dictSize,
                                              dictLoadMethod, dictContentType,
                                              params) ))
        return NULL;

    return cdict;
}

ZSTD_compressionParameters ZSTD_getCParamsFromCDict(const ZSTD_CDict* cdict)
{
    assert(cdict != NULL);
    return cdict->matchState.cParams;
}

/*! ZSTD_getDictID_fromCDict() :
 *  Provides the dictID of the dictionary loaded into `cdict`.
 *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
 *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
unsigned ZSTD_getDictID_fromCDict(const ZSTD_CDict* cdict)
{
    if (cdict==NULL) return 0;
    return cdict->dictID;
}

/* ZSTD_compressBegin_usingCDict_internal() :
 * Implementation of various ZSTD_compressBegin_usingCDict* functions.
 */
static size_t ZSTD_compressBegin_usingCDict_internal(
    ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
    ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
{
    ZSTD_CCtx_params cctxParams;
    DEBUGLOG(4, "ZSTD_compressBegin_usingCDict_internal");
    RETURN_ERROR_IF(cdict==NULL, dictionary_wrong, "NULL pointer!");
    /* Initialize the cctxParams from the cdict */
    {
        ZSTD_parameters params;
        params.fParams = fParams;
        params.cParams = ( pledgedSrcSize < ZSTD_USE_CDICT_PARAMS_SRCSIZE_CUTOFF
                        || pledgedSrcSize < cdict->dictContentSize * ZSTD_USE_CDICT_PARAMS_DICTSIZE_MULTIPLIER
                        || pledgedSrcSize == ZSTD_CONTENTSIZE_UNKNOWN
                        || cdict->compressionLevel == 0 ) ?
                ZSTD_getCParamsFromCDict(cdict)
              : ZSTD_getCParams(cdict->compressionLevel,
                                pledgedSrcSize,
                                cdict->dictContentSize);
        ZSTD_CCtxParams_init_internal(&cctxParams, &params, cdict->compressionLevel);
    }
    /* Increase window log to fit the entire dictionary and source if the
     * source size is known. Limit the increase to 19, which is the
     * window log for compression level 1 with the largest source size.
     */
    if (pledgedSrcSize != ZSTD_CONTENTSIZE_UNKNOWN) {
        U32 const limitedSrcSize = (U32)MIN(pledgedSrcSize, 1U << 19);
        U32 const limitedSrcLog = limitedSrcSize > 1 ? ZSTD_highbit32(limitedSrcSize - 1) + 1 : 1;
        cctxParams.cParams.windowLog = MAX(cctxParams.cParams.windowLog, limitedSrcLog);
    }
    return ZSTD_compressBegin_internal(cctx,
                                        NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast,
                                        cdict,
                                        &cctxParams, pledgedSrcSize,
                                        ZSTDb_not_buffered);
}


/* ZSTD_compressBegin_usingCDict_advanced() :
 * This function is DEPRECATED.
 * cdict must be != NULL */
size_t ZSTD_compressBegin_usingCDict_advanced(
    ZSTD_CCtx* const cctx, const ZSTD_CDict* const cdict,
    ZSTD_frameParameters const fParams, unsigned long long const pledgedSrcSize)
{
    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, pledgedSrcSize);
}

/* ZSTD_compressBegin_usingCDict() :
 * cdict must be != NULL */
size_t ZSTD_compressBegin_usingCDict_deprecated(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
{
    ZSTD_frameParameters const fParams = { 0 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
    return ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, ZSTD_CONTENTSIZE_UNKNOWN);
}

size_t ZSTD_compressBegin_usingCDict(ZSTD_CCtx* cctx, const ZSTD_CDict* cdict)
{
    return ZSTD_compressBegin_usingCDict_deprecated(cctx, cdict);
}

/*! ZSTD_compress_usingCDict_internal():
 * Implementation of various ZSTD_compress_usingCDict* functions.
 */
static size_t ZSTD_compress_usingCDict_internal(ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
{
    FORWARD_IF_ERROR(ZSTD_compressBegin_usingCDict_internal(cctx, cdict, fParams, srcSize), ""); /* will check if cdict != NULL */
    return ZSTD_compressEnd_public(cctx, dst, dstCapacity, src, srcSize);
}

/*! ZSTD_compress_usingCDict_advanced():
 * This function is DEPRECATED.
 */
size_t ZSTD_compress_usingCDict_advanced(ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                const ZSTD_CDict* cdict, ZSTD_frameParameters fParams)
{
    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
}

/*! ZSTD_compress_usingCDict() :
 *  Compression using a digested Dictionary.
 *  Faster startup than ZSTD_compress_usingDict(), recommended when same dictionary is used multiple times.
 *  Note that compression parameters are decided at CDict creation time
 *  while frame parameters are hardcoded */
size_t ZSTD_compress_usingCDict(ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                                const void* src, size_t srcSize,
                                const ZSTD_CDict* cdict)
{
    ZSTD_frameParameters const fParams = { 1 /*content*/, 0 /*checksum*/, 0 /*noDictID*/ };
    return ZSTD_compress_usingCDict_internal(cctx, dst, dstCapacity, src, srcSize, cdict, fParams);
}


/* ******************************************************************
*  Streaming
********************************************************************/

ZSTD_CStream* ZSTD_createCStream(void)
{
    DEBUGLOG(3, "ZSTD_createCStream");
    return ZSTD_createCStream_advanced(ZSTD_defaultCMem);
}

ZSTD_CStream* ZSTD_initStaticCStream(void *workspace, size_t workspaceSize)
{
    return ZSTD_initStaticCCtx(workspace, workspaceSize);
}

ZSTD_CStream* ZSTD_createCStream_advanced(ZSTD_customMem customMem)
{   /* CStream and CCtx are now same object */
    return ZSTD_createCCtx_advanced(customMem);
}

size_t ZSTD_freeCStream(ZSTD_CStream* zcs)
{
    return ZSTD_freeCCtx(zcs);   /* same object */
}


/*======   Initialization   ======*/

size_t ZSTD_CStreamInSize(void)  { return ZSTD_BLOCKSIZE_MAX; }

size_t ZSTD_CStreamOutSize(void)
{
    return ZSTD_compressBound(ZSTD_BLOCKSIZE_MAX) + ZSTD_blockHeaderSize + 4 /* 32-bits hash */ ;
}

static ZSTD_CParamMode_e ZSTD_getCParamMode(ZSTD_CDict const* cdict, ZSTD_CCtx_params const* params, U64 pledgedSrcSize)
{
    if (cdict != NULL && ZSTD_shouldAttachDict(cdict, params, pledgedSrcSize))
        return ZSTD_cpm_attachDict;
    else
        return ZSTD_cpm_noAttachDict;
}

/* ZSTD_resetCStream():
 * pledgedSrcSize == 0 means "unknown" */
size_t ZSTD_resetCStream(ZSTD_CStream* zcs, unsigned long long pss)
{
    /* temporary : 0 interpreted as "unknown" during transition period.
     * Users willing to specify "unknown" **must** use ZSTD_CONTENTSIZE_UNKNOWN.
     * 0 will be interpreted as "empty" in the future.
     */
    U64 const pledgedSrcSize = (pss==0) ? ZSTD_CONTENTSIZE_UNKNOWN : pss;
    DEBUGLOG(4, "ZSTD_resetCStream: pledgedSrcSize = %u", (unsigned)pledgedSrcSize);
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
    return 0;
}

/*! ZSTD_initCStream_internal() :
 *  Note : for lib/compress only. Used by zstdmt_compress.c.
 *  Assumption 1 : params are valid
 *  Assumption 2 : either dict, or cdict, is defined, not both */
size_t ZSTD_initCStream_internal(ZSTD_CStream* zcs,
                    const void* dict, size_t dictSize, const ZSTD_CDict* cdict,
                    const ZSTD_CCtx_params* params,
                    unsigned long long pledgedSrcSize)
{
    DEBUGLOG(4, "ZSTD_initCStream_internal");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
    assert(!ZSTD_isError(ZSTD_checkCParams(params->cParams)));
    zcs->requestedParams = *params;
    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */
    if (dict) {
        FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , "");
    } else {
        /* Dictionary is cleared if !cdict */
        FORWARD_IF_ERROR( ZSTD_CCtx_refCDict(zcs, cdict) , "");
    }
    return 0;
}

/* ZSTD_initCStream_usingCDict_advanced() :
 * same as ZSTD_initCStream_usingCDict(), with control over frame parameters */
size_t ZSTD_initCStream_usingCDict_advanced(ZSTD_CStream* zcs,
                                            const ZSTD_CDict* cdict,
                                            ZSTD_frameParameters fParams,
                                            unsigned long long pledgedSrcSize)
{
    DEBUGLOG(4, "ZSTD_initCStream_usingCDict_advanced");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
    zcs->requestedParams.fParams = fParams;
    FORWARD_IF_ERROR( ZSTD_CCtx_refCDict(zcs, cdict) , "");
    return 0;
}

/* note : cdict must outlive compression session */
size_t ZSTD_initCStream_usingCDict(ZSTD_CStream* zcs, const ZSTD_CDict* cdict)
{
    DEBUGLOG(4, "ZSTD_initCStream_usingCDict");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_refCDict(zcs, cdict) , "");
    return 0;
}


/* ZSTD_initCStream_advanced() :
 * pledgedSrcSize must be exact.
 * if srcSize is not known at init time, use value ZSTD_CONTENTSIZE_UNKNOWN.
 * dict is loaded with default parameters ZSTD_dct_auto and ZSTD_dlm_byCopy. */
size_t ZSTD_initCStream_advanced(ZSTD_CStream* zcs,
                                 const void* dict, size_t dictSize,
                                 ZSTD_parameters params, unsigned long long pss)
{
    /* for compatibility with older programs relying on this behavior.
     * Users should now specify ZSTD_CONTENTSIZE_UNKNOWN.
     * This line will be removed in the future.
     */
    U64 const pledgedSrcSize = (pss==0 && params.fParams.contentSizeFlag==0) ? ZSTD_CONTENTSIZE_UNKNOWN : pss;
    DEBUGLOG(4, "ZSTD_initCStream_advanced");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
    FORWARD_IF_ERROR( ZSTD_checkCParams(params.cParams) , "");
    ZSTD_CCtxParams_setZstdParams(&zcs->requestedParams, &params);
    FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , "");
    return 0;
}

size_t ZSTD_initCStream_usingDict(ZSTD_CStream* zcs, const void* dict, size_t dictSize, int compressionLevel)
{
    DEBUGLOG(4, "ZSTD_initCStream_usingDict");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_loadDictionary(zcs, dict, dictSize) , "");
    return 0;
}

size_t ZSTD_initCStream_srcSize(ZSTD_CStream* zcs, int compressionLevel, unsigned long long pss)
{
    /* temporary : 0 interpreted as "unknown" during transition period.
     * Users willing to specify "unknown" **must** use ZSTD_CONTENTSIZE_UNKNOWN.
     * 0 will be interpreted as "empty" in the future.
     */
    U64 const pledgedSrcSize = (pss==0) ? ZSTD_CONTENTSIZE_UNKNOWN : pss;
    DEBUGLOG(4, "ZSTD_initCStream_srcSize");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_refCDict(zcs, NULL) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setPledgedSrcSize(zcs, pledgedSrcSize) , "");
    return 0;
}

size_t ZSTD_initCStream(ZSTD_CStream* zcs, int compressionLevel)
{
    DEBUGLOG(4, "ZSTD_initCStream");
    FORWARD_IF_ERROR( ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_refCDict(zcs, NULL) , "");
    FORWARD_IF_ERROR( ZSTD_CCtx_setParameter(zcs, ZSTD_c_compressionLevel, compressionLevel) , "");
    return 0;
}

/*======   Compression   ======*/

static size_t ZSTD_nextInputSizeHint(const ZSTD_CCtx* cctx)
{
    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
        return cctx->blockSizeMax - cctx->stableIn_notConsumed;
    }
    assert(cctx->appliedParams.inBufferMode == ZSTD_bm_buffered);
    {   size_t hintInSize = cctx->inBuffTarget - cctx->inBuffPos;
        if (hintInSize==0) hintInSize = cctx->blockSizeMax;
        return hintInSize;
    }
}

/** ZSTD_compressStream_generic():
 *  internal function for all *compressStream*() variants
 * @return : hint size for next input to complete ongoing block */
static size_t ZSTD_compressStream_generic(ZSTD_CStream* zcs,
                                          ZSTD_outBuffer* output,
                                          ZSTD_inBuffer* input,
                                          ZSTD_EndDirective const flushMode)
{
    const char* const istart = (assert(input != NULL), (const char*)input->src);
    const char* const iend = (istart != NULL) ? istart + input->size : istart;
    const char* ip = (istart != NULL) ? istart + input->pos : istart;
    char* const ostart = (assert(output != NULL), (char*)output->dst);
    char* const oend = (ostart != NULL) ? ostart + output->size : ostart;
    char* op = (ostart != NULL) ? ostart + output->pos : ostart;
    U32 someMoreWork = 1;

    /* check expectations */
    DEBUGLOG(5, "ZSTD_compressStream_generic, flush=%i, srcSize = %zu", (int)flushMode, input->size - input->pos);
    assert(zcs != NULL);
    if (zcs->appliedParams.inBufferMode == ZSTD_bm_stable) {
        assert(input->pos >= zcs->stableIn_notConsumed);
        input->pos -= zcs->stableIn_notConsumed;
        if (ip) ip -= zcs->stableIn_notConsumed;
        zcs->stableIn_notConsumed = 0;
    }
    if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) {
        assert(zcs->inBuff != NULL);
        assert(zcs->inBuffSize > 0);
    }
    if (zcs->appliedParams.outBufferMode == ZSTD_bm_buffered) {
        assert(zcs->outBuff !=  NULL);
        assert(zcs->outBuffSize > 0);
    }
    if (input->src == NULL) assert(input->size == 0);
    assert(input->pos <= input->size);
    if (output->dst == NULL) assert(output->size == 0);
    assert(output->pos <= output->size);
    assert((U32)flushMode <= (U32)ZSTD_e_end);

    while (someMoreWork) {
        switch(zcs->streamStage)
        {
        case zcss_init:
            RETURN_ERROR(init_missing, "call ZSTD_initCStream() first!");

        case zcss_load:
            if ( (flushMode == ZSTD_e_end)
              && ( (size_t)(oend-op) >= ZSTD_compressBound((size_t)(iend-ip))     /* Enough output space */
                || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)  /* OR we are allowed to return dstSizeTooSmall */
              && (zcs->inBuffPos == 0) ) {
                /* shortcut to compression pass directly into output buffer */
                size_t const cSize = ZSTD_compressEnd_public(zcs,
                                                op, (size_t)(oend-op),
                                                ip, (size_t)(iend-ip));
                DEBUGLOG(4, "ZSTD_compressEnd : cSize=%u", (unsigned)cSize);
                FORWARD_IF_ERROR(cSize, "ZSTD_compressEnd failed");
                ip = iend;
                op += cSize;
                zcs->frameEnded = 1;
                ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
                someMoreWork = 0; break;
            }
            /* complete loading into inBuffer in buffered mode */
            if (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered) {
                size_t const toLoad = zcs->inBuffTarget - zcs->inBuffPos;
                size_t const loaded = ZSTD_limitCopy(
                                        zcs->inBuff + zcs->inBuffPos, toLoad,
                                        ip, (size_t)(iend-ip));
                zcs->inBuffPos += loaded;
                if (ip) ip += loaded;
                if ( (flushMode == ZSTD_e_continue)
                  && (zcs->inBuffPos < zcs->inBuffTarget) ) {
                    /* not enough input to fill full block : stop here */
                    someMoreWork = 0; break;
                }
                if ( (flushMode == ZSTD_e_flush)
                  && (zcs->inBuffPos == zcs->inToCompress) ) {
                    /* empty */
                    someMoreWork = 0; break;
                }
            } else {
                assert(zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
                if ( (flushMode == ZSTD_e_continue)
                  && ( (size_t)(iend - ip) < zcs->blockSizeMax) ) {
                    /* can't compress a full block : stop here */
                    zcs->stableIn_notConsumed = (size_t)(iend - ip);
                    ip = iend;  /* pretend to have consumed input */
                    someMoreWork = 0; break;
                }
                if ( (flushMode == ZSTD_e_flush)
                  && (ip == iend) ) {
                    /* empty */
                    someMoreWork = 0; break;
                }
            }
            /* compress current block (note : this stage cannot be stopped in the middle) */
            DEBUGLOG(5, "stream compression stage (flushMode==%u)", flushMode);
            {   int const inputBuffered = (zcs->appliedParams.inBufferMode == ZSTD_bm_buffered);
                void* cDst;
                size_t cSize;
                size_t oSize = (size_t)(oend-op);
                size_t const iSize = inputBuffered ? zcs->inBuffPos - zcs->inToCompress
                                                   : MIN((size_t)(iend - ip), zcs->blockSizeMax);
                if (oSize >= ZSTD_compressBound(iSize) || zcs->appliedParams.outBufferMode == ZSTD_bm_stable)
                    cDst = op;   /* compress into output buffer, to skip flush stage */
                else
                    cDst = zcs->outBuff, oSize = zcs->outBuffSize;
                if (inputBuffered) {
                    unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip==iend);
                    cSize = lastBlock ?
                            ZSTD_compressEnd_public(zcs, cDst, oSize,
                                        zcs->inBuff + zcs->inToCompress, iSize) :
                            ZSTD_compressContinue_public(zcs, cDst, oSize,
                                        zcs->inBuff + zcs->inToCompress, iSize);
                    FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
                    zcs->frameEnded = lastBlock;
                    /* prepare next block */
                    zcs->inBuffTarget = zcs->inBuffPos + zcs->blockSizeMax;
                    if (zcs->inBuffTarget > zcs->inBuffSize)
                        zcs->inBuffPos = 0, zcs->inBuffTarget = zcs->blockSizeMax;
                    DEBUGLOG(5, "inBuffTarget:%u / inBuffSize:%u",
                            (unsigned)zcs->inBuffTarget, (unsigned)zcs->inBuffSize);
                    if (!lastBlock)
                        assert(zcs->inBuffTarget <= zcs->inBuffSize);
                    zcs->inToCompress = zcs->inBuffPos;
                } else { /* !inputBuffered, hence ZSTD_bm_stable */
                    unsigned const lastBlock = (flushMode == ZSTD_e_end) && (ip + iSize == iend);
                    cSize = lastBlock ?
                            ZSTD_compressEnd_public(zcs, cDst, oSize, ip, iSize) :
                            ZSTD_compressContinue_public(zcs, cDst, oSize, ip, iSize);
                    /* Consume the input prior to error checking to mirror buffered mode. */
                    if (ip) ip += iSize;
                    FORWARD_IF_ERROR(cSize, "%s", lastBlock ? "ZSTD_compressEnd failed" : "ZSTD_compressContinue failed");
                    zcs->frameEnded = lastBlock;
                    if (lastBlock) assert(ip == iend);
                }
                if (cDst == op) {  /* no need to flush */
                    op += cSize;
                    if (zcs->frameEnded) {
                        DEBUGLOG(5, "Frame completed directly in outBuffer");
                        someMoreWork = 0;
                        ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
                    }
                    break;
                }
                zcs->outBuffContentSize = cSize;
                zcs->outBuffFlushedSize = 0;
                zcs->streamStage = zcss_flush; /* pass-through to flush stage */
            }
	    ZSTD_FALLTHROUGH;
        case zcss_flush:
            DEBUGLOG(5, "flush stage");
            assert(zcs->appliedParams.outBufferMode == ZSTD_bm_buffered);
            {   size_t const toFlush = zcs->outBuffContentSize - zcs->outBuffFlushedSize;
                size_t const flushed = ZSTD_limitCopy(op, (size_t)(oend-op),
                            zcs->outBuff + zcs->outBuffFlushedSize, toFlush);
                DEBUGLOG(5, "toFlush: %u into %u ==> flushed: %u",
                            (unsigned)toFlush, (unsigned)(oend-op), (unsigned)flushed);
                if (flushed)
                    op += flushed;
                zcs->outBuffFlushedSize += flushed;
                if (toFlush!=flushed) {
                    /* flush not fully completed, presumably because dst is too small */
                    assert(op==oend);
                    someMoreWork = 0;
                    break;
                }
                zcs->outBuffContentSize = zcs->outBuffFlushedSize = 0;
                if (zcs->frameEnded) {
                    DEBUGLOG(5, "Frame completed on flush");
                    someMoreWork = 0;
                    ZSTD_CCtx_reset(zcs, ZSTD_reset_session_only);
                    break;
                }
                zcs->streamStage = zcss_load;
                break;
            }

        default: /* impossible */
            assert(0);
        }
    }

    input->pos = (size_t)(ip - istart);
    output->pos = (size_t)(op - ostart);
    if (zcs->frameEnded) return 0;
    return ZSTD_nextInputSizeHint(zcs);
}

static size_t ZSTD_nextInputSizeHint_MTorST(const ZSTD_CCtx* cctx)
{
#ifdef ZSTD_MULTITHREAD
    if (cctx->appliedParams.nbWorkers >= 1) {
        assert(cctx->mtctx != NULL);
        return ZSTDMT_nextInputSizeHint(cctx->mtctx);
    }
#endif
    return ZSTD_nextInputSizeHint(cctx);

}

size_t ZSTD_compressStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
{
    FORWARD_IF_ERROR( ZSTD_compressStream2(zcs, output, input, ZSTD_e_continue) , "");
    return ZSTD_nextInputSizeHint_MTorST(zcs);
}

/* After a compression call set the expected input/output buffer.
 * This is validated at the start of the next compression call.
 */
static void
ZSTD_setBufferExpectations(ZSTD_CCtx* cctx, const ZSTD_outBuffer* output, const ZSTD_inBuffer* input)
{
    DEBUGLOG(5, "ZSTD_setBufferExpectations (for advanced stable in/out modes)");
    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
        cctx->expectedInBuffer = *input;
    }
    if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) {
        cctx->expectedOutBufferSize = output->size - output->pos;
    }
}

/* Validate that the input/output buffers match the expectations set by
 * ZSTD_setBufferExpectations.
 */
static size_t ZSTD_checkBufferStability(ZSTD_CCtx const* cctx,
                                        ZSTD_outBuffer const* output,
                                        ZSTD_inBuffer const* input,
                                        ZSTD_EndDirective endOp)
{
    if (cctx->appliedParams.inBufferMode == ZSTD_bm_stable) {
        ZSTD_inBuffer const expect = cctx->expectedInBuffer;
        if (expect.src != input->src || expect.pos != input->pos)
            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableInBuffer enabled but input differs!");
    }
    (void)endOp;
    if (cctx->appliedParams.outBufferMode == ZSTD_bm_stable) {
        size_t const outBufferSize = output->size - output->pos;
        if (cctx->expectedOutBufferSize != outBufferSize)
            RETURN_ERROR(stabilityCondition_notRespected, "ZSTD_c_stableOutBuffer enabled but output size differs!");
    }
    return 0;
}

/*
 * If @endOp == ZSTD_e_end, @inSize becomes pledgedSrcSize.
 * Otherwise, it's ignored.
 * @return: 0 on success, or a ZSTD_error code otherwise.
 */
static size_t ZSTD_CCtx_init_compressStream2(ZSTD_CCtx* cctx,
                                             ZSTD_EndDirective endOp,
                                             size_t inSize)
{
    ZSTD_CCtx_params params = cctx->requestedParams;
    ZSTD_prefixDict const prefixDict = cctx->prefixDict;
    FORWARD_IF_ERROR( ZSTD_initLocalDict(cctx) , ""); /* Init the local dict if present. */
    ZSTD_memset(&cctx->prefixDict, 0, sizeof(cctx->prefixDict));   /* single usage */
    assert(prefixDict.dict==NULL || cctx->cdict==NULL);    /* only one can be set */
    if (cctx->cdict && !cctx->localDict.cdict) {
        /* Let the cdict's compression level take priority over the requested params.
         * But do not take the cdict's compression level if the "cdict" is actually a localDict
         * generated from ZSTD_initLocalDict().
         */
        params.compressionLevel = cctx->cdict->compressionLevel;
    }
    DEBUGLOG(4, "ZSTD_CCtx_init_compressStream2 : transparent init stage");
    if (endOp == ZSTD_e_end) cctx->pledgedSrcSizePlusOne = inSize + 1;  /* auto-determine pledgedSrcSize */

    {   size_t const dictSize = prefixDict.dict
                ? prefixDict.dictSize
                : (cctx->cdict ? cctx->cdict->dictContentSize : 0);
        ZSTD_CParamMode_e const mode = ZSTD_getCParamMode(cctx->cdict, &params, cctx->pledgedSrcSizePlusOne - 1);
        params.cParams = ZSTD_getCParamsFromCCtxParams(
                &params, cctx->pledgedSrcSizePlusOne-1,
                dictSize, mode);
    }

    params.postBlockSplitter = ZSTD_resolveBlockSplitterMode(params.postBlockSplitter, &params.cParams);
    params.ldmParams.enableLdm = ZSTD_resolveEnableLdm(params.ldmParams.enableLdm, &params.cParams);
    params.useRowMatchFinder = ZSTD_resolveRowMatchFinderMode(params.useRowMatchFinder, &params.cParams);
    params.validateSequences = ZSTD_resolveExternalSequenceValidation(params.validateSequences);
    params.maxBlockSize = ZSTD_resolveMaxBlockSize(params.maxBlockSize);
    params.searchForExternalRepcodes = ZSTD_resolveExternalRepcodeSearch(params.searchForExternalRepcodes, params.compressionLevel);

#ifdef ZSTD_MULTITHREAD
    /* If external matchfinder is enabled, make sure to fail before checking job size (for consistency) */
    RETURN_ERROR_IF(
        ZSTD_hasExtSeqProd(&params) && params.nbWorkers >= 1,
        parameter_combination_unsupported,
        "External sequence producer isn't supported with nbWorkers >= 1"
    );

    if ((cctx->pledgedSrcSizePlusOne-1) <= ZSTDMT_JOBSIZE_MIN) {
        params.nbWorkers = 0; /* do not invoke multi-threading when src size is too small */
    }
    if (params.nbWorkers > 0) {
# if ZSTD_TRACE
        cctx->traceCtx = (ZSTD_trace_compress_begin != NULL) ? ZSTD_trace_compress_begin(cctx) : 0;
# endif
        /* mt context creation */
        if (cctx->mtctx == NULL) {
            DEBUGLOG(4, "ZSTD_compressStream2: creating new mtctx for nbWorkers=%u",
                        params.nbWorkers);
            cctx->mtctx = ZSTDMT_createCCtx_advanced((U32)params.nbWorkers, cctx->customMem, cctx->pool);
            RETURN_ERROR_IF(cctx->mtctx == NULL, memory_allocation, "NULL pointer!");
        }
        /* mt compression */
        DEBUGLOG(4, "call ZSTDMT_initCStream_internal as nbWorkers=%u", params.nbWorkers);
        FORWARD_IF_ERROR( ZSTDMT_initCStream_internal(
                    cctx->mtctx,
                    prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType,
                    cctx->cdict, params, cctx->pledgedSrcSizePlusOne-1) , "");
        cctx->dictID = cctx->cdict ? cctx->cdict->dictID : 0;
        cctx->dictContentSize = cctx->cdict ? cctx->cdict->dictContentSize : prefixDict.dictSize;
        cctx->consumedSrcSize = 0;
        cctx->producedCSize = 0;
        cctx->streamStage = zcss_load;
        cctx->appliedParams = params;
    } else
#endif  /* ZSTD_MULTITHREAD */
    {   U64 const pledgedSrcSize = cctx->pledgedSrcSizePlusOne - 1;
        assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
        FORWARD_IF_ERROR( ZSTD_compressBegin_internal(cctx,
                prefixDict.dict, prefixDict.dictSize, prefixDict.dictContentType, ZSTD_dtlm_fast,
                cctx->cdict,
                &params, pledgedSrcSize,
                ZSTDb_buffered) , "");
        assert(cctx->appliedParams.nbWorkers == 0);
        cctx->inToCompress = 0;
        cctx->inBuffPos = 0;
        if (cctx->appliedParams.inBufferMode == ZSTD_bm_buffered) {
            /* for small input: avoid automatic flush on reaching end of block, since
            * it would require to add a 3-bytes null block to end frame
            */
            cctx->inBuffTarget = cctx->blockSizeMax + (cctx->blockSizeMax == pledgedSrcSize);
        } else {
            cctx->inBuffTarget = 0;
        }
        cctx->outBuffContentSize = cctx->outBuffFlushedSize = 0;
        cctx->streamStage = zcss_load;
        cctx->frameEnded = 0;
    }
    return 0;
}

/* @return provides a minimum amount of data remaining to be flushed from internal buffers
 */
size_t ZSTD_compressStream2( ZSTD_CCtx* cctx,
                             ZSTD_outBuffer* output,
                             ZSTD_inBuffer* input,
                             ZSTD_EndDirective endOp)
{
    DEBUGLOG(5, "ZSTD_compressStream2, endOp=%u ", (unsigned)endOp);
    /* check conditions */
    RETURN_ERROR_IF(output->pos > output->size, dstSize_tooSmall, "invalid output buffer");
    RETURN_ERROR_IF(input->pos  > input->size, srcSize_wrong, "invalid input buffer");
    RETURN_ERROR_IF((U32)endOp > (U32)ZSTD_e_end, parameter_outOfBound, "invalid endDirective");
    assert(cctx != NULL);

    /* transparent initialization stage */
    if (cctx->streamStage == zcss_init) {
        size_t const inputSize = input->size - input->pos;  /* no obligation to start from pos==0 */
        size_t const totalInputSize = inputSize + cctx->stableIn_notConsumed;
        if ( (cctx->requestedParams.inBufferMode == ZSTD_bm_stable) /* input is presumed stable, across invocations */
          && (endOp == ZSTD_e_continue)                             /* no flush requested, more input to come */
          && (totalInputSize < ZSTD_BLOCKSIZE_MAX) ) {              /* not even reached one block yet */
            if (cctx->stableIn_notConsumed) {  /* not the first time */
                /* check stable source guarantees */
                RETURN_ERROR_IF(input->src != cctx->expectedInBuffer.src, stabilityCondition_notRespected, "stableInBuffer condition not respected: wrong src pointer");
                RETURN_ERROR_IF(input->pos != cctx->expectedInBuffer.size, stabilityCondition_notRespected, "stableInBuffer condition not respected: externally modified pos");
            }
            /* pretend input was consumed, to give a sense forward progress */
            input->pos = input->size;
            /* save stable inBuffer, for later control, and flush/end */
            cctx->expectedInBuffer = *input;
            /* but actually input wasn't consumed, so keep track of position from where compression shall resume */
            cctx->stableIn_notConsumed += inputSize;
            /* don't initialize yet, wait for the first block of flush() order, for better parameters adaptation */
            return ZSTD_FRAMEHEADERSIZE_MIN(cctx->requestedParams.format);  /* at least some header to produce */
        }
        FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, endOp, totalInputSize), "compressStream2 initialization failed");
        ZSTD_setBufferExpectations(cctx, output, input);   /* Set initial buffer expectations now that we've initialized */
    }
    /* end of transparent initialization stage */

    FORWARD_IF_ERROR(ZSTD_checkBufferStability(cctx, output, input, endOp), "invalid buffers");
    /* compression stage */
#ifdef ZSTD_MULTITHREAD
    if (cctx->appliedParams.nbWorkers > 0) {
        size_t flushMin;
        if (cctx->cParamsChanged) {
            ZSTDMT_updateCParams_whileCompressing(cctx->mtctx, &cctx->requestedParams);
            cctx->cParamsChanged = 0;
        }
        if (cctx->stableIn_notConsumed) {
            assert(cctx->appliedParams.inBufferMode == ZSTD_bm_stable);
            /* some early data was skipped - make it available for consumption */
            assert(input->pos >= cctx->stableIn_notConsumed);
            input->pos -= cctx->stableIn_notConsumed;
            cctx->stableIn_notConsumed = 0;
        }
        for (;;) {
            size_t const ipos = input->pos;
            size_t const opos = output->pos;
            flushMin = ZSTDMT_compressStream_generic(cctx->mtctx, output, input, endOp);
            cctx->consumedSrcSize += (U64)(input->pos - ipos);
            cctx->producedCSize += (U64)(output->pos - opos);
            if ( ZSTD_isError(flushMin)
              || (endOp == ZSTD_e_end && flushMin == 0) ) { /* compression completed */
                if (flushMin == 0)
                    ZSTD_CCtx_trace(cctx, 0);
                ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only);
            }
            FORWARD_IF_ERROR(flushMin, "ZSTDMT_compressStream_generic failed");

            if (endOp == ZSTD_e_continue) {
                /* We only require some progress with ZSTD_e_continue, not maximal progress.
                 * We're done if we've consumed or produced any bytes, or either buffer is
                 * full.
                 */
                if (input->pos != ipos || output->pos != opos || input->pos == input->size || output->pos == output->size)
                    break;
            } else {
                assert(endOp == ZSTD_e_flush || endOp == ZSTD_e_end);
                /* We require maximal progress. We're done when the flush is complete or the
                 * output buffer is full.
                 */
                if (flushMin == 0 || output->pos == output->size)
                    break;
            }
        }
        DEBUGLOG(5, "completed ZSTD_compressStream2 delegating to ZSTDMT_compressStream_generic");
        /* Either we don't require maximum forward progress, we've finished the
         * flush, or we are out of output space.
         */
        assert(endOp == ZSTD_e_continue || flushMin == 0 || output->pos == output->size);
        ZSTD_setBufferExpectations(cctx, output, input);
        return flushMin;
    }
#endif /* ZSTD_MULTITHREAD */
    FORWARD_IF_ERROR( ZSTD_compressStream_generic(cctx, output, input, endOp) , "");
    DEBUGLOG(5, "completed ZSTD_compressStream2");
    ZSTD_setBufferExpectations(cctx, output, input);
    return cctx->outBuffContentSize - cctx->outBuffFlushedSize; /* remaining to flush */
}

size_t ZSTD_compressStream2_simpleArgs (
                            ZSTD_CCtx* cctx,
                            void* dst, size_t dstCapacity, size_t* dstPos,
                      const void* src, size_t srcSize, size_t* srcPos,
                            ZSTD_EndDirective endOp)
{
    ZSTD_outBuffer output;
    ZSTD_inBuffer  input;
    output.dst = dst;
    output.size = dstCapacity;
    output.pos = *dstPos;
    input.src = src;
    input.size = srcSize;
    input.pos = *srcPos;
    /* ZSTD_compressStream2() will check validity of dstPos and srcPos */
    {   size_t const cErr = ZSTD_compressStream2(cctx, &output, &input, endOp);
        *dstPos = output.pos;
        *srcPos = input.pos;
        return cErr;
    }
}

size_t ZSTD_compress2(ZSTD_CCtx* cctx,
                      void* dst, size_t dstCapacity,
                      const void* src, size_t srcSize)
{
    ZSTD_bufferMode_e const originalInBufferMode = cctx->requestedParams.inBufferMode;
    ZSTD_bufferMode_e const originalOutBufferMode = cctx->requestedParams.outBufferMode;
    DEBUGLOG(4, "ZSTD_compress2 (srcSize=%u)", (unsigned)srcSize);
    ZSTD_CCtx_reset(cctx, ZSTD_reset_session_only);
    /* Enable stable input/output buffers. */
    cctx->requestedParams.inBufferMode = ZSTD_bm_stable;
    cctx->requestedParams.outBufferMode = ZSTD_bm_stable;
    {   size_t oPos = 0;
        size_t iPos = 0;
        size_t const result = ZSTD_compressStream2_simpleArgs(cctx,
                                        dst, dstCapacity, &oPos,
                                        src, srcSize, &iPos,
                                        ZSTD_e_end);
        /* Reset to the original values. */
        cctx->requestedParams.inBufferMode = originalInBufferMode;
        cctx->requestedParams.outBufferMode = originalOutBufferMode;

        FORWARD_IF_ERROR(result, "ZSTD_compressStream2_simpleArgs failed");
        if (result != 0) {  /* compression not completed, due to lack of output space */
            assert(oPos == dstCapacity);
            RETURN_ERROR(dstSize_tooSmall, "");
        }
        assert(iPos == srcSize);   /* all input is expected consumed */
        return oPos;
    }
}

/* ZSTD_validateSequence() :
 * @offBase : must use the format required by ZSTD_storeSeq()
 * @returns a ZSTD error code if sequence is not valid
 */
static size_t
ZSTD_validateSequence(U32 offBase, U32 matchLength, U32 minMatch,
                      size_t posInSrc, U32 windowLog, size_t dictSize, int useSequenceProducer)
{
    U32 const windowSize = 1u << windowLog;
    /* posInSrc represents the amount of data the decoder would decode up to this point.
     * As long as the amount of data decoded is less than or equal to window size, offsets may be
     * larger than the total length of output decoded in order to reference the dict, even larger than
     * window size. After output surpasses windowSize, we're limited to windowSize offsets again.
     */
    size_t const offsetBound = posInSrc > windowSize ? (size_t)windowSize : posInSrc + (size_t)dictSize;
    size_t const matchLenLowerBound = (minMatch == 3 || useSequenceProducer) ? 3 : 4;
    RETURN_ERROR_IF(offBase > OFFSET_TO_OFFBASE(offsetBound), externalSequences_invalid, "Offset too large!");
    /* Validate maxNbSeq is large enough for the given matchLength and minMatch */
    RETURN_ERROR_IF(matchLength < matchLenLowerBound, externalSequences_invalid, "Matchlength too small for the minMatch");
    return 0;
}

/* Returns an offset code, given a sequence's raw offset, the ongoing repcode array, and whether litLength == 0 */
static U32 ZSTD_finalizeOffBase(U32 rawOffset, const U32 rep[ZSTD_REP_NUM], U32 ll0)
{
    U32 offBase = OFFSET_TO_OFFBASE(rawOffset);

    if (!ll0 && rawOffset == rep[0]) {
        offBase = REPCODE1_TO_OFFBASE;
    } else if (rawOffset == rep[1]) {
        offBase = REPCODE_TO_OFFBASE(2 - ll0);
    } else if (rawOffset == rep[2]) {
        offBase = REPCODE_TO_OFFBASE(3 - ll0);
    } else if (ll0 && rawOffset == rep[0] - 1) {
        offBase = REPCODE3_TO_OFFBASE;
    }
    return offBase;
}

/* This function scans through an array of ZSTD_Sequence,
 * storing the sequences it reads, until it reaches a block delimiter.
 * Note that the block delimiter includes the last literals of the block.
 * @blockSize must be == sum(sequence_lengths).
 * @returns @blockSize on success, and a ZSTD_error otherwise.
 */
static size_t
ZSTD_transferSequences_wBlockDelim(ZSTD_CCtx* cctx,
                                   ZSTD_SequencePosition* seqPos,
                             const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
                             const void* src, size_t blockSize,
                                   ZSTD_ParamSwitch_e externalRepSearch)
{
    U32 idx = seqPos->idx;
    U32 const startIdx = idx;
    BYTE const* ip = (BYTE const*)(src);
    const BYTE* const iend = ip + blockSize;
    Repcodes_t updatedRepcodes;
    U32 dictSize;

    DEBUGLOG(5, "ZSTD_transferSequences_wBlockDelim (blockSize = %zu)", blockSize);

    if (cctx->cdict) {
        dictSize = (U32)cctx->cdict->dictContentSize;
    } else if (cctx->prefixDict.dict) {
        dictSize = (U32)cctx->prefixDict.dictSize;
    } else {
        dictSize = 0;
    }
    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));
    for (; idx < inSeqsSize && (inSeqs[idx].matchLength != 0 || inSeqs[idx].offset != 0); ++idx) {
        U32 const litLength = inSeqs[idx].litLength;
        U32 const matchLength = inSeqs[idx].matchLength;
        U32 offBase;

        if (externalRepSearch == ZSTD_ps_disable) {
            offBase = OFFSET_TO_OFFBASE(inSeqs[idx].offset);
        } else {
            U32 const ll0 = (litLength == 0);
            offBase = ZSTD_finalizeOffBase(inSeqs[idx].offset, updatedRepcodes.rep, ll0);
            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
        }

        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
        if (cctx->appliedParams.validateSequences) {
            seqPos->posInSrc += litLength + matchLength;
            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch,
                                                seqPos->posInSrc,
                                                cctx->appliedParams.cParams.windowLog, dictSize,
                                                ZSTD_hasExtSeqProd(&cctx->appliedParams)),
                                                "Sequence validation failed");
        }
        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
                        "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
        ip += matchLength + litLength;
    }
    RETURN_ERROR_IF(idx == inSeqsSize, externalSequences_invalid, "Block delimiter not found.");

    /* If we skipped repcode search while parsing, we need to update repcodes now */
    assert(externalRepSearch != ZSTD_ps_auto);
    assert(idx >= startIdx);
    if (externalRepSearch == ZSTD_ps_disable && idx != startIdx) {
        U32* const rep = updatedRepcodes.rep;
        U32 lastSeqIdx = idx - 1; /* index of last non-block-delimiter sequence */

        if (lastSeqIdx >= startIdx + 2) {
            rep[2] = inSeqs[lastSeqIdx - 2].offset;
            rep[1] = inSeqs[lastSeqIdx - 1].offset;
            rep[0] = inSeqs[lastSeqIdx].offset;
        } else if (lastSeqIdx == startIdx + 1) {
            rep[2] = rep[0];
            rep[1] = inSeqs[lastSeqIdx - 1].offset;
            rep[0] = inSeqs[lastSeqIdx].offset;
        } else {
            assert(lastSeqIdx == startIdx);
            rep[2] = rep[1];
            rep[1] = rep[0];
            rep[0] = inSeqs[lastSeqIdx].offset;
        }
    }

    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));

    if (inSeqs[idx].litLength) {
        DEBUGLOG(6, "Storing last literals of size: %u", inSeqs[idx].litLength);
        ZSTD_storeLastLiterals(&cctx->seqStore, ip, inSeqs[idx].litLength);
        ip += inSeqs[idx].litLength;
        seqPos->posInSrc += inSeqs[idx].litLength;
    }
    RETURN_ERROR_IF(ip != iend, externalSequences_invalid, "Blocksize doesn't agree with block delimiter!");
    seqPos->idx = idx+1;
    return blockSize;
}

/*
 * This function attempts to scan through @blockSize bytes in @src
 * represented by the sequences in @inSeqs,
 * storing any (partial) sequences.
 *
 * Occasionally, we may want to reduce the actual number of bytes consumed from @src
 * to avoid splitting a match, notably if it would produce a match smaller than MINMATCH.
 *
 * @returns the number of bytes consumed from @src, necessarily <= @blockSize.
 * Otherwise, it may return a ZSTD error if something went wrong.
 */
static size_t
ZSTD_transferSequences_noDelim(ZSTD_CCtx* cctx,
                               ZSTD_SequencePosition* seqPos,
                         const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
                         const void* src, size_t blockSize,
                               ZSTD_ParamSwitch_e externalRepSearch)
{
    U32 idx = seqPos->idx;
    U32 startPosInSequence = seqPos->posInSequence;
    U32 endPosInSequence = seqPos->posInSequence + (U32)blockSize;
    size_t dictSize;
    const BYTE* const istart = (const BYTE*)(src);
    const BYTE* ip = istart;
    const BYTE* iend = istart + blockSize;  /* May be adjusted if we decide to process fewer than blockSize bytes */
    Repcodes_t updatedRepcodes;
    U32 bytesAdjustment = 0;
    U32 finalMatchSplit = 0;

    /* TODO(embg) support fast parsing mode in noBlockDelim mode */
    (void)externalRepSearch;

    if (cctx->cdict) {
        dictSize = cctx->cdict->dictContentSize;
    } else if (cctx->prefixDict.dict) {
        dictSize = cctx->prefixDict.dictSize;
    } else {
        dictSize = 0;
    }
    DEBUGLOG(5, "ZSTD_transferSequences_noDelim: idx: %u PIS: %u blockSize: %zu", idx, startPosInSequence, blockSize);
    DEBUGLOG(5, "Start seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));
    while (endPosInSequence && idx < inSeqsSize && !finalMatchSplit) {
        const ZSTD_Sequence currSeq = inSeqs[idx];
        U32 litLength = currSeq.litLength;
        U32 matchLength = currSeq.matchLength;
        U32 const rawOffset = currSeq.offset;
        U32 offBase;

        /* Modify the sequence depending on where endPosInSequence lies */
        if (endPosInSequence >= currSeq.litLength + currSeq.matchLength) {
            if (startPosInSequence >= litLength) {
                startPosInSequence -= litLength;
                litLength = 0;
                matchLength -= startPosInSequence;
            } else {
                litLength -= startPosInSequence;
            }
            /* Move to the next sequence */
            endPosInSequence -= currSeq.litLength + currSeq.matchLength;
            startPosInSequence = 0;
        } else {
            /* This is the final (partial) sequence we're adding from inSeqs, and endPosInSequence
               does not reach the end of the match. So, we have to split the sequence */
            DEBUGLOG(6, "Require a split: diff: %u, idx: %u PIS: %u",
                     currSeq.litLength + currSeq.matchLength - endPosInSequence, idx, endPosInSequence);
            if (endPosInSequence > litLength) {
                U32 firstHalfMatchLength;
                litLength = startPosInSequence >= litLength ? 0 : litLength - startPosInSequence;
                firstHalfMatchLength = endPosInSequence - startPosInSequence - litLength;
                if (matchLength > blockSize && firstHalfMatchLength >= cctx->appliedParams.cParams.minMatch) {
                    /* Only ever split the match if it is larger than the block size */
                    U32 secondHalfMatchLength = currSeq.matchLength + currSeq.litLength - endPosInSequence;
                    if (secondHalfMatchLength < cctx->appliedParams.cParams.minMatch) {
                        /* Move the endPosInSequence backward so that it creates match of minMatch length */
                        endPosInSequence -= cctx->appliedParams.cParams.minMatch - secondHalfMatchLength;
                        bytesAdjustment = cctx->appliedParams.cParams.minMatch - secondHalfMatchLength;
                        firstHalfMatchLength -= bytesAdjustment;
                    }
                    matchLength = firstHalfMatchLength;
                    /* Flag that we split the last match - after storing the sequence, exit the loop,
                       but keep the value of endPosInSequence */
                    finalMatchSplit = 1;
                } else {
                    /* Move the position in sequence backwards so that we don't split match, and break to store
                     * the last literals. We use the original currSeq.litLength as a marker for where endPosInSequence
                     * should go. We prefer to do this whenever it is not necessary to split the match, or if doing so
                     * would cause the first half of the match to be too small
                     */
                    bytesAdjustment = endPosInSequence - currSeq.litLength;
                    endPosInSequence = currSeq.litLength;
                    break;
                }
            } else {
                /* This sequence ends inside the literals, break to store the last literals */
                break;
            }
        }
        /* Check if this offset can be represented with a repcode */
        {   U32 const ll0 = (litLength == 0);
            offBase = ZSTD_finalizeOffBase(rawOffset, updatedRepcodes.rep, ll0);
            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
        }

        if (cctx->appliedParams.validateSequences) {
            seqPos->posInSrc += litLength + matchLength;
            FORWARD_IF_ERROR(ZSTD_validateSequence(offBase, matchLength, cctx->appliedParams.cParams.minMatch, seqPos->posInSrc,
                                                   cctx->appliedParams.cParams.windowLog, dictSize, ZSTD_hasExtSeqProd(&cctx->appliedParams)),
                                                   "Sequence validation failed");
        }
        DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
        RETURN_ERROR_IF(idx - seqPos->idx >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
                        "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");
        ZSTD_storeSeq(&cctx->seqStore, litLength, ip, iend, offBase, matchLength);
        ip += matchLength + litLength;
        if (!finalMatchSplit)
            idx++; /* Next Sequence */
    }
    DEBUGLOG(5, "Ending seq: idx: %u (of: %u ml: %u ll: %u)", idx, inSeqs[idx].offset, inSeqs[idx].matchLength, inSeqs[idx].litLength);
    assert(idx == inSeqsSize || endPosInSequence <= inSeqs[idx].litLength + inSeqs[idx].matchLength);
    seqPos->idx = idx;
    seqPos->posInSequence = endPosInSequence;
    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));

    iend -= bytesAdjustment;
    if (ip != iend) {
        /* Store any last literals */
        U32 const lastLLSize = (U32)(iend - ip);
        assert(ip <= iend);
        DEBUGLOG(6, "Storing last literals of size: %u", lastLLSize);
        ZSTD_storeLastLiterals(&cctx->seqStore, ip, lastLLSize);
        seqPos->posInSrc += lastLLSize;
    }

    return (size_t)(iend-istart);
}

/* @seqPos represents a position within @inSeqs,
 * it is read and updated by this function,
 * once the goal to produce a block of size @blockSize is reached.
 * @return: nb of bytes consumed from @src, necessarily <= @blockSize.
 */
typedef size_t (*ZSTD_SequenceCopier_f)(ZSTD_CCtx* cctx,
                                        ZSTD_SequencePosition* seqPos,
                                  const ZSTD_Sequence* const inSeqs, size_t inSeqsSize,
                                  const void* src, size_t blockSize,
                                        ZSTD_ParamSwitch_e externalRepSearch);

static ZSTD_SequenceCopier_f ZSTD_selectSequenceCopier(ZSTD_SequenceFormat_e mode)
{
    assert(ZSTD_cParam_withinBounds(ZSTD_c_blockDelimiters, (int)mode));
    if (mode == ZSTD_sf_explicitBlockDelimiters) {
        return ZSTD_transferSequences_wBlockDelim;
    }
    assert(mode == ZSTD_sf_noBlockDelimiters);
    return ZSTD_transferSequences_noDelim;
}

/* Discover the size of next block by searching for the delimiter.
 * Note that a block delimiter **must** exist in this mode,
 * otherwise it's an input error.
 * The block size retrieved will be later compared to ensure it remains within bounds */
static size_t
blockSize_explicitDelimiter(const ZSTD_Sequence* inSeqs, size_t inSeqsSize, ZSTD_SequencePosition seqPos)
{
    int end = 0;
    size_t blockSize = 0;
    size_t spos = seqPos.idx;
    DEBUGLOG(6, "blockSize_explicitDelimiter : seq %zu / %zu", spos, inSeqsSize);
    assert(spos <= inSeqsSize);
    while (spos < inSeqsSize) {
        end = (inSeqs[spos].offset == 0);
        blockSize += inSeqs[spos].litLength + inSeqs[spos].matchLength;
        if (end) {
            if (inSeqs[spos].matchLength != 0)
                RETURN_ERROR(externalSequences_invalid, "delimiter format error : both matchlength and offset must be == 0");
            break;
        }
        spos++;
    }
    if (!end)
        RETURN_ERROR(externalSequences_invalid, "Reached end of sequences without finding a block delimiter");
    return blockSize;
}

static size_t determine_blockSize(ZSTD_SequenceFormat_e mode,
                           size_t blockSize, size_t remaining,
                     const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                           ZSTD_SequencePosition seqPos)
{
    DEBUGLOG(6, "determine_blockSize : remainingSize = %zu", remaining);
    if (mode == ZSTD_sf_noBlockDelimiters) {
        /* Note: more a "target" block size */
        return MIN(remaining, blockSize);
    }
    assert(mode == ZSTD_sf_explicitBlockDelimiters);
    {   size_t const explicitBlockSize = blockSize_explicitDelimiter(inSeqs, inSeqsSize, seqPos);
        FORWARD_IF_ERROR(explicitBlockSize, "Error while determining block size with explicit delimiters");
        if (explicitBlockSize > blockSize)
            RETURN_ERROR(externalSequences_invalid, "sequences incorrectly define a too large block");
        if (explicitBlockSize > remaining)
            RETURN_ERROR(externalSequences_invalid, "sequences define a frame longer than source");
        return explicitBlockSize;
    }
}

/* Compress all provided sequences, block-by-block.
 *
 * Returns the cumulative size of all compressed blocks (including their headers),
 * otherwise a ZSTD error.
 */
static size_t
ZSTD_compressSequences_internal(ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                          const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                          const void* src, size_t srcSize)
{
    size_t cSize = 0;
    size_t remaining = srcSize;
    ZSTD_SequencePosition seqPos = {0, 0, 0};

    const BYTE* ip = (BYTE const*)src;
    BYTE* op = (BYTE*)dst;
    ZSTD_SequenceCopier_f const sequenceCopier = ZSTD_selectSequenceCopier(cctx->appliedParams.blockDelimiters);

    DEBUGLOG(4, "ZSTD_compressSequences_internal srcSize: %zu, inSeqsSize: %zu", srcSize, inSeqsSize);
    /* Special case: empty frame */
    if (remaining == 0) {
        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1);
        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "No room for empty frame block header");
        MEM_writeLE32(op, cBlockHeader24);
        op += ZSTD_blockHeaderSize;
        dstCapacity -= ZSTD_blockHeaderSize;
        cSize += ZSTD_blockHeaderSize;
    }

    while (remaining) {
        size_t compressedSeqsSize;
        size_t cBlockSize;
        size_t blockSize = determine_blockSize(cctx->appliedParams.blockDelimiters,
                                        cctx->blockSizeMax, remaining,
                                        inSeqs, inSeqsSize, seqPos);
        U32 const lastBlock = (blockSize == remaining);
        FORWARD_IF_ERROR(blockSize, "Error while trying to determine block size");
        assert(blockSize <= remaining);
        ZSTD_resetSeqStore(&cctx->seqStore);

        blockSize = sequenceCopier(cctx,
                                   &seqPos, inSeqs, inSeqsSize,
                                   ip, blockSize,
                                   cctx->appliedParams.searchForExternalRepcodes);
        FORWARD_IF_ERROR(blockSize, "Bad sequence copy");

        /* If blocks are too small, emit as a nocompress block */
        /* TODO: See 3090. We reduced MIN_CBLOCK_SIZE from 3 to 2 so to compensate we are adding
         * additional 1. We need to revisit and change this logic to be more consistent */
        if (blockSize < MIN_CBLOCK_SIZE+ZSTD_blockHeaderSize+1+1) {
            cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
            FORWARD_IF_ERROR(cBlockSize, "Nocompress block failed");
            DEBUGLOG(5, "Block too small (%zu): data remains uncompressed: cSize=%zu", blockSize, cBlockSize);
            cSize += cBlockSize;
            ip += blockSize;
            op += cBlockSize;
            remaining -= blockSize;
            dstCapacity -= cBlockSize;
            continue;
        }

        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "not enough dstCapacity to write a new compressed block");
        compressedSeqsSize = ZSTD_entropyCompressSeqStore(&cctx->seqStore,
                                &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
                                &cctx->appliedParams,
                                op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
                                blockSize,
                                cctx->tmpWorkspace, cctx->tmpWkspSize /* statically allocated in resetCCtx */,
                                cctx->bmi2);
        FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed");
        DEBUGLOG(5, "Compressed sequences size: %zu", compressedSeqsSize);

        if (!cctx->isFirstBlock &&
            ZSTD_maybeRLE(&cctx->seqStore) &&
            ZSTD_isRLE(ip, blockSize)) {
            /* Note: don't emit the first block as RLE even if it qualifies because
             * doing so will cause the decoder (cli <= v1.4.3 only) to throw an (invalid) error
             * "should consume all input error."
             */
            compressedSeqsSize = 1;
        }

        if (compressedSeqsSize == 0) {
            /* ZSTD_noCompressBlock writes the block header as well */
            cBlockSize = ZSTD_noCompressBlock(op, dstCapacity, ip, blockSize, lastBlock);
            FORWARD_IF_ERROR(cBlockSize, "ZSTD_noCompressBlock failed");
            DEBUGLOG(5, "Writing out nocompress block, size: %zu", cBlockSize);
        } else if (compressedSeqsSize == 1) {
            cBlockSize = ZSTD_rleCompressBlock(op, dstCapacity, *ip, blockSize, lastBlock);
            FORWARD_IF_ERROR(cBlockSize, "ZSTD_rleCompressBlock failed");
            DEBUGLOG(5, "Writing out RLE block, size: %zu", cBlockSize);
        } else {
            U32 cBlockHeader;
            /* Error checking and repcodes update */
            ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState);
            if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
                cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;

            /* Write block header into beginning of block*/
            cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3);
            MEM_writeLE24(op, cBlockHeader);
            cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize;
            DEBUGLOG(5, "Writing out compressed block, size: %zu", cBlockSize);
        }

        cSize += cBlockSize;

        if (lastBlock) {
            break;
        } else {
            ip += blockSize;
            op += cBlockSize;
            remaining -= blockSize;
            dstCapacity -= cBlockSize;
            cctx->isFirstBlock = 0;
        }
        DEBUGLOG(5, "cSize running total: %zu (remaining dstCapacity=%zu)", cSize, dstCapacity);
    }

    DEBUGLOG(4, "cSize final total: %zu", cSize);
    return cSize;
}

size_t ZSTD_compressSequences(ZSTD_CCtx* cctx,
                              void* dst, size_t dstCapacity,
                              const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                              const void* src, size_t srcSize)
{
    BYTE* op = (BYTE*)dst;
    size_t cSize = 0;

    /* Transparent initialization stage, same as compressStream2() */
    DEBUGLOG(4, "ZSTD_compressSequences (nbSeqs=%zu,dstCapacity=%zu)", inSeqsSize, dstCapacity);
    assert(cctx != NULL);
    FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, srcSize), "CCtx initialization failed");

    /* Begin writing output, starting with frame header */
    {   size_t const frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity,
                    &cctx->appliedParams, srcSize, cctx->dictID);
        op += frameHeaderSize;
        assert(frameHeaderSize <= dstCapacity);
        dstCapacity -= frameHeaderSize;
        cSize += frameHeaderSize;
    }
    if (cctx->appliedParams.fParams.checksumFlag && srcSize) {
        XXH64_update(&cctx->xxhState, src, srcSize);
    }

    /* Now generate compressed blocks */
    {   size_t const cBlocksSize = ZSTD_compressSequences_internal(cctx,
                                                           op, dstCapacity,
                                                           inSeqs, inSeqsSize,
                                                           src, srcSize);
        FORWARD_IF_ERROR(cBlocksSize, "Compressing blocks failed!");
        cSize += cBlocksSize;
        assert(cBlocksSize <= dstCapacity);
        dstCapacity -= cBlocksSize;
    }

    /* Complete with frame checksum, if needed */
    if (cctx->appliedParams.fParams.checksumFlag) {
        U32 const checksum = (U32) XXH64_digest(&cctx->xxhState);
        RETURN_ERROR_IF(dstCapacity<4, dstSize_tooSmall, "no room for checksum");
        DEBUGLOG(4, "Write checksum : %08X", (unsigned)checksum);
        MEM_writeLE32((char*)dst + cSize, checksum);
        cSize += 4;
    }

    DEBUGLOG(4, "Final compressed size: %zu", cSize);
    return cSize;
}


#if defined(__AVX2__)

#include <immintrin.h>  /* AVX2 intrinsics */

/*
 * Convert 2 sequences per iteration, using AVX2 intrinsics:
 *   - offset -> offBase = offset + 2
 *   - litLength -> (U16) litLength
 *   - matchLength -> (U16)(matchLength - 3)
 *   - rep is ignored
 * Store only 8 bytes per SeqDef (offBase[4], litLength[2], mlBase[2]).
 *
 * At the end, instead of extracting two __m128i,
 * we use _mm256_permute4x64_epi64(..., 0xE8) to move lane2 into lane1,
 * then store the lower 16 bytes in one go.
 *
 * @returns 0 on succes, with no long length detected
 * @returns > 0 if there is one long length (> 65535),
 * indicating the position, and type.
 */
static size_t convertSequences_noRepcodes(
    SeqDef* dstSeqs,
    const ZSTD_Sequence* inSeqs,
    size_t nbSequences)
{
    /*
     * addition:
     *   For each 128-bit half: (offset+2, litLength+0, matchLength-3, rep+0)
     */
    const __m256i addition = _mm256_setr_epi32(
        ZSTD_REP_NUM, 0, -MINMATCH, 0,    /* for sequence i */
        ZSTD_REP_NUM, 0, -MINMATCH, 0     /* for sequence i+1 */
    );

    /* limit: check if there is a long length */
    const __m256i limit = _mm256_set1_epi32(65535);

    /*
     * shuffle mask for byte-level rearrangement in each 128-bit half:
     *
     * Input layout (after addition) per 128-bit half:
     *   [ offset+2 (4 bytes) | litLength (4 bytes) | matchLength (4 bytes) | rep (4 bytes) ]
     * We only need:
     *   offBase (4 bytes) = offset+2
     *   litLength (2 bytes) = low 2 bytes of litLength
     *   mlBase (2 bytes) = low 2 bytes of (matchLength)
     * => Bytes [0..3, 4..5, 8..9], zero the rest.
     */
    const __m256i mask = _mm256_setr_epi8(
        /* For the lower 128 bits => sequence i */
         0, 1, 2, 3,       /* offset+2 */
         4, 5,             /* litLength (16 bits) */
         8, 9,             /* matchLength (16 bits) */
         (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,
         (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,

        /* For the upper 128 bits => sequence i+1 */
        16,17,18,19,       /* offset+2 */
        20,21,             /* litLength */
        24,25,             /* matchLength */
        (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80,
        (BYTE)0x80, (BYTE)0x80, (BYTE)0x80, (BYTE)0x80
    );

    /*
     * Next, we'll use _mm256_permute4x64_epi64(vshf, 0xE8).
     * Explanation of 0xE8 = 11101000b => [lane0, lane2, lane2, lane3].
     * So the lower 128 bits become [lane0, lane2] => combining seq0 and seq1.
     */
#define PERM_LANE_0X_E8 0xE8  /* [0,2,2,3] in lane indices */

    size_t longLen = 0, i = 0;

    /* AVX permutation depends on the specific definition of target structures */
    ZSTD_STATIC_ASSERT(sizeof(ZSTD_Sequence) == 16);
    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, offset) == 0);
    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, litLength) == 4);
    ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, matchLength) == 8);
    ZSTD_STATIC_ASSERT(sizeof(SeqDef) == 8);
    ZSTD_STATIC_ASSERT(offsetof(SeqDef, offBase) == 0);
    ZSTD_STATIC_ASSERT(offsetof(SeqDef, litLength) == 4);
    ZSTD_STATIC_ASSERT(offsetof(SeqDef, mlBase) == 6);

    /* Process 2 sequences per loop iteration */
    for (; i + 1 < nbSequences; i += 2) {
        /* Load 2 ZSTD_Sequence (32 bytes) */
        __m256i vin  = _mm256_loadu_si256((const __m256i*)(const void*)&inSeqs[i]);

        /* Add {2, 0, -3, 0} in each 128-bit half */
        __m256i vadd = _mm256_add_epi32(vin, addition);

        /* Check for long length */
        __m256i ll_cmp  = _mm256_cmpgt_epi32(vadd, limit);  /* 0xFFFFFFFF for element > 65535 */
        int ll_res  = _mm256_movemask_epi8(ll_cmp);

        /* Shuffle bytes so each half gives us the 8 bytes we need */
        __m256i vshf = _mm256_shuffle_epi8(vadd, mask);
        /*
         * Now:
         *   Lane0 = seq0's 8 bytes
         *   Lane1 = 0
         *   Lane2 = seq1's 8 bytes
         *   Lane3 = 0
         */

        /* Permute 64-bit lanes => move Lane2 down into Lane1. */
        __m256i vperm = _mm256_permute4x64_epi64(vshf, PERM_LANE_0X_E8);
        /*
         * Now the lower 16 bytes (Lane0+Lane1) = [seq0, seq1].
         * The upper 16 bytes are [Lane2, Lane3] = [seq1, 0], but we won't use them.
         */

        /* Store only the lower 16 bytes => 2 SeqDef (8 bytes each) */
        _mm_storeu_si128((__m128i *)(void*)&dstSeqs[i], _mm256_castsi256_si128(vperm));
        /*
         * This writes out 16 bytes total:
         *   - offset 0..7  => seq0 (offBase, litLength, mlBase)
         *   - offset 8..15 => seq1 (offBase, litLength, mlBase)
         */

        /* check (unlikely) long lengths > 65535
         * indices for lengths correspond to bits [4..7], [8..11], [20..23], [24..27]
         * => combined mask = 0x0FF00FF0
         */
        if (UNLIKELY((ll_res & 0x0FF00FF0) != 0)) {
            /* long length detected: let's figure out which one*/
            if (inSeqs[i].matchLength > 65535+MINMATCH) {
                assert(longLen == 0);
                longLen = i + 1;
            }
            if (inSeqs[i].litLength > 65535) {
                assert(longLen == 0);
                longLen = i + nbSequences + 1;
            }
            if (inSeqs[i+1].matchLength > 65535+MINMATCH) {
                assert(longLen == 0);
                longLen = i + 1 + 1;
            }
            if (inSeqs[i+1].litLength > 65535) {
                assert(longLen == 0);
                longLen = i + 1 + nbSequences + 1;
            }
        }
    }

    /* Handle leftover if @nbSequences is odd */
    if (i < nbSequences) {
        /* process last sequence */
        assert(i == nbSequences - 1);
        dstSeqs[i].offBase = OFFSET_TO_OFFBASE(inSeqs[i].offset);
        dstSeqs[i].litLength = (U16)inSeqs[i].litLength;
        dstSeqs[i].mlBase = (U16)(inSeqs[i].matchLength - MINMATCH);
        /* check (unlikely) long lengths > 65535 */
        if (UNLIKELY(inSeqs[i].matchLength > 65535+MINMATCH)) {
            assert(longLen == 0);
            longLen = i + 1;
        }
        if (UNLIKELY(inSeqs[i].litLength > 65535)) {
            assert(longLen == 0);
            longLen = i + nbSequences + 1;
        }
    }

    return longLen;
}

/* the vector implementation could also be ported to SSSE3,
 * but since this implementation is targeting modern systems (>= Sapphire Rapid),
 * it's not useful to develop and maintain code for older pre-AVX2 platforms */

#else /* no AVX2 */

static size_t convertSequences_noRepcodes(
    SeqDef* dstSeqs,
    const ZSTD_Sequence* inSeqs,
    size_t nbSequences)
{
    size_t longLen = 0;
    size_t n;
    for (n=0; n<nbSequences; n++) {
        dstSeqs[n].offBase = OFFSET_TO_OFFBASE(inSeqs[n].offset);
        dstSeqs[n].litLength = (U16)inSeqs[n].litLength;
        dstSeqs[n].mlBase = (U16)(inSeqs[n].matchLength - MINMATCH);
        /* check for long length > 65535 */
        if (UNLIKELY(inSeqs[n].matchLength > 65535+MINMATCH)) {
            assert(longLen == 0);
            longLen = n + 1;
        }
        if (UNLIKELY(inSeqs[n].litLength > 65535)) {
            assert(longLen == 0);
            longLen = n + nbSequences + 1;
        }
    }
    return longLen;
}

#endif

/*
 * Precondition: Sequences must end on an explicit Block Delimiter
 * @return: 0 on success, or an error code.
 * Note: Sequence validation functionality has been disabled (removed).
 * This is helpful to generate a lean main pipeline, improving performance.
 * It may be re-inserted later.
 */
size_t ZSTD_convertBlockSequences(ZSTD_CCtx* cctx,
                const ZSTD_Sequence* const inSeqs, size_t nbSequences,
                int repcodeResolution)
{
    Repcodes_t updatedRepcodes;
    size_t seqNb = 0;

    DEBUGLOG(5, "ZSTD_convertBlockSequences (nbSequences = %zu)", nbSequences);

    RETURN_ERROR_IF(nbSequences >= cctx->seqStore.maxNbSeq, externalSequences_invalid,
                    "Not enough memory allocated. Try adjusting ZSTD_c_minMatch.");

    ZSTD_memcpy(updatedRepcodes.rep, cctx->blockState.prevCBlock->rep, sizeof(Repcodes_t));

    /* check end condition */
    assert(nbSequences >= 1);
    assert(inSeqs[nbSequences-1].matchLength == 0);
    assert(inSeqs[nbSequences-1].offset == 0);

    /* Convert Sequences from public format to internal format */
    if (!repcodeResolution) {
        size_t const longl = convertSequences_noRepcodes(cctx->seqStore.sequencesStart, inSeqs, nbSequences-1);
        cctx->seqStore.sequences = cctx->seqStore.sequencesStart + nbSequences-1;
        if (longl) {
            DEBUGLOG(5, "long length");
            assert(cctx->seqStore.longLengthType == ZSTD_llt_none);
            if (longl <= nbSequences-1) {
                DEBUGLOG(5, "long match length detected at pos %zu", longl-1);
                cctx->seqStore.longLengthType = ZSTD_llt_matchLength;
                cctx->seqStore.longLengthPos = (U32)(longl-1);
            } else {
                DEBUGLOG(5, "long literals length detected at pos %zu", longl-nbSequences);
                assert(longl <= 2* (nbSequences-1));
                cctx->seqStore.longLengthType = ZSTD_llt_literalLength;
                cctx->seqStore.longLengthPos = (U32)(longl-(nbSequences-1)-1);
            }
        }
    } else {
        for (seqNb = 0; seqNb < nbSequences - 1 ; seqNb++) {
            U32 const litLength = inSeqs[seqNb].litLength;
            U32 const matchLength = inSeqs[seqNb].matchLength;
            U32 const ll0 = (litLength == 0);
            U32 const offBase = ZSTD_finalizeOffBase(inSeqs[seqNb].offset, updatedRepcodes.rep, ll0);

            DEBUGLOG(6, "Storing sequence: (of: %u, ml: %u, ll: %u)", offBase, matchLength, litLength);
            ZSTD_storeSeqOnly(&cctx->seqStore, litLength, offBase, matchLength);
            ZSTD_updateRep(updatedRepcodes.rep, offBase, ll0);
        }
    }

    /* If we skipped repcode search while parsing, we need to update repcodes now */
    if (!repcodeResolution && nbSequences > 1) {
        U32* const rep = updatedRepcodes.rep;

        if (nbSequences >= 4) {
            U32 lastSeqIdx = (U32)nbSequences - 2; /* index of last full sequence */
            rep[2] = inSeqs[lastSeqIdx - 2].offset;
            rep[1] = inSeqs[lastSeqIdx - 1].offset;
            rep[0] = inSeqs[lastSeqIdx].offset;
        } else if (nbSequences == 3) {
            rep[2] = rep[0];
            rep[1] = inSeqs[0].offset;
            rep[0] = inSeqs[1].offset;
        } else {
            assert(nbSequences == 2);
            rep[2] = rep[1];
            rep[1] = rep[0];
            rep[0] = inSeqs[0].offset;
        }
    }

    ZSTD_memcpy(cctx->blockState.nextCBlock->rep, updatedRepcodes.rep, sizeof(Repcodes_t));

    return 0;
}

#if defined(ZSTD_ARCH_X86_AVX2)

BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
{
    size_t i;
    __m256i const zeroVec = _mm256_setzero_si256();
    __m256i sumVec = zeroVec;  /* accumulates match+lit in 32-bit lanes */
    ZSTD_ALIGNED(32) U32 tmp[8];      /* temporary buffer for reduction */
    size_t mSum = 0, lSum = 0;
    ZSTD_STATIC_ASSERT(sizeof(ZSTD_Sequence) == 16);

    /* Process 2 structs (32 bytes) at a time */
    for (i = 0; i + 2 <= nbSeqs; i += 2) {
        /* Load two consecutive ZSTD_Sequence (8×4 = 32 bytes) */
        __m256i data     = _mm256_loadu_si256((const __m256i*)(const void*)&seqs[i]);
        /* check end of block signal */
        __m256i cmp      = _mm256_cmpeq_epi32(data, zeroVec);
        int cmp_res      = _mm256_movemask_epi8(cmp);
        /* indices for match lengths correspond to bits [8..11], [24..27]
         * => combined mask = 0x0F000F00 */
        ZSTD_STATIC_ASSERT(offsetof(ZSTD_Sequence, matchLength) == 8);
        if (cmp_res & 0x0F000F00) break;
        /* Accumulate in sumVec */
        sumVec           = _mm256_add_epi32(sumVec, data);
    }

    /* Horizontal reduction */
    _mm256_store_si256((__m256i*)tmp, sumVec);
    lSum = tmp[1] + tmp[5];
    mSum = tmp[2] + tmp[6];

    /* Handle the leftover */
    for (; i < nbSeqs; i++) {
        lSum += seqs[i].litLength;
        mSum += seqs[i].matchLength;
        if (seqs[i].matchLength == 0) break; /* end of block */
    }

    if (i==nbSeqs) {
        /* reaching end of sequences: end of block signal was not present */
        BlockSummary bs;
        bs.nbSequences = ERROR(externalSequences_invalid);
        return bs;
    }
    {   BlockSummary bs;
        bs.nbSequences = i+1;
        bs.blockSize = lSum + mSum;
        bs.litSize = lSum;
        return bs;
    }
}

#else

BlockSummary ZSTD_get1BlockSummary(const ZSTD_Sequence* seqs, size_t nbSeqs)
{
    size_t totalMatchSize = 0;
    size_t litSize = 0;
    size_t n;
    assert(seqs);
    for (n=0; n<nbSeqs; n++) {
        totalMatchSize += seqs[n].matchLength;
        litSize += seqs[n].litLength;
        if (seqs[n].matchLength == 0) {
            assert(seqs[n].offset == 0);
            break;
        }
    }
    if (n==nbSeqs) {
        BlockSummary bs;
        bs.nbSequences = ERROR(externalSequences_invalid);
        return bs;
    }
    {   BlockSummary bs;
        bs.nbSequences = n+1;
        bs.blockSize = litSize + totalMatchSize;
        bs.litSize = litSize;
        return bs;
    }
}
#endif


static size_t
ZSTD_compressSequencesAndLiterals_internal(ZSTD_CCtx* cctx,
                                void* dst, size_t dstCapacity,
                          const ZSTD_Sequence* inSeqs, size_t nbSequences,
                          const void* literals, size_t litSize, size_t srcSize)
{
    size_t remaining = srcSize;
    size_t cSize = 0;
    BYTE* op = (BYTE*)dst;
    int const repcodeResolution = (cctx->appliedParams.searchForExternalRepcodes == ZSTD_ps_enable);
    assert(cctx->appliedParams.searchForExternalRepcodes != ZSTD_ps_auto);

    DEBUGLOG(4, "ZSTD_compressSequencesAndLiterals_internal: nbSeqs=%zu, litSize=%zu", nbSequences, litSize);
    RETURN_ERROR_IF(nbSequences == 0, externalSequences_invalid, "Requires at least 1 end-of-block");

    /* Special case: empty frame */
    if ((nbSequences == 1) && (inSeqs[0].litLength == 0)) {
        U32 const cBlockHeader24 = 1 /* last block */ + (((U32)bt_raw)<<1);
        RETURN_ERROR_IF(dstCapacity<3, dstSize_tooSmall, "No room for empty frame block header");
        MEM_writeLE24(op, cBlockHeader24);
        op += ZSTD_blockHeaderSize;
        dstCapacity -= ZSTD_blockHeaderSize;
        cSize += ZSTD_blockHeaderSize;
    }

    while (nbSequences) {
        size_t compressedSeqsSize, cBlockSize, conversionStatus;
        BlockSummary const block = ZSTD_get1BlockSummary(inSeqs, nbSequences);
        U32 const lastBlock = (block.nbSequences == nbSequences);
        FORWARD_IF_ERROR(block.nbSequences, "Error while trying to determine nb of sequences for a block");
        assert(block.nbSequences <= nbSequences);
        RETURN_ERROR_IF(block.litSize > litSize, externalSequences_invalid, "discrepancy: Sequences require more literals than present in buffer");
        ZSTD_resetSeqStore(&cctx->seqStore);

        conversionStatus = ZSTD_convertBlockSequences(cctx,
                            inSeqs, block.nbSequences,
                            repcodeResolution);
        FORWARD_IF_ERROR(conversionStatus, "Bad sequence conversion");
        inSeqs += block.nbSequences;
        nbSequences -= block.nbSequences;
        remaining -= block.blockSize;

        /* Note: when blockSize is very small, other variant send it uncompressed.
         * Here, we still send the sequences, because we don't have the original source to send it uncompressed.
         * One could imagine in theory reproducing the source from the sequences,
         * but that's complex and costly memory intensive, and goes against the objectives of this variant. */

        RETURN_ERROR_IF(dstCapacity < ZSTD_blockHeaderSize, dstSize_tooSmall, "not enough dstCapacity to write a new compressed block");

        compressedSeqsSize = ZSTD_entropyCompressSeqStore_internal(
                                op + ZSTD_blockHeaderSize /* Leave space for block header */, dstCapacity - ZSTD_blockHeaderSize,
                                literals, block.litSize,
                                &cctx->seqStore,
                                &cctx->blockState.prevCBlock->entropy, &cctx->blockState.nextCBlock->entropy,
                                &cctx->appliedParams,
                                cctx->tmpWorkspace, cctx->tmpWkspSize /* statically allocated in resetCCtx */,
                                cctx->bmi2);
        FORWARD_IF_ERROR(compressedSeqsSize, "Compressing sequences of block failed");
        /* note: the spec forbids for any compressed block to be larger than maximum block size */
        if (compressedSeqsSize > cctx->blockSizeMax) compressedSeqsSize = 0;
        DEBUGLOG(5, "Compressed sequences size: %zu", compressedSeqsSize);
        litSize -= block.litSize;
        literals = (const char*)literals + block.litSize;

        /* Note: difficult to check source for RLE block when only Literals are provided,
         * but it could be considered from analyzing the sequence directly */

        if (compressedSeqsSize == 0) {
            /* Sending uncompressed blocks is out of reach, because the source is not provided.
             * In theory, one could use the sequences to regenerate the source, like a decompressor,
             * but it's complex, and memory hungry, killing the purpose of this variant.
             * Current outcome: generate an error code.
             */
            RETURN_ERROR(cannotProduce_uncompressedBlock, "ZSTD_compressSequencesAndLiterals cannot generate an uncompressed block");
        } else {
            U32 cBlockHeader;
            assert(compressedSeqsSize > 1); /* no RLE */
            /* Error checking and repcodes update */
            ZSTD_blockState_confirmRepcodesAndEntropyTables(&cctx->blockState);
            if (cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode == FSE_repeat_valid)
                cctx->blockState.prevCBlock->entropy.fse.offcode_repeatMode = FSE_repeat_check;

            /* Write block header into beginning of block*/
            cBlockHeader = lastBlock + (((U32)bt_compressed)<<1) + (U32)(compressedSeqsSize << 3);
            MEM_writeLE24(op, cBlockHeader);
            cBlockSize = ZSTD_blockHeaderSize + compressedSeqsSize;
            DEBUGLOG(5, "Writing out compressed block, size: %zu", cBlockSize);
        }

        cSize += cBlockSize;
        op += cBlockSize;
        dstCapacity -= cBlockSize;
        cctx->isFirstBlock = 0;
        DEBUGLOG(5, "cSize running total: %zu (remaining dstCapacity=%zu)", cSize, dstCapacity);

        if (lastBlock) {
            assert(nbSequences == 0);
            break;
        }
    }

    RETURN_ERROR_IF(litSize != 0, externalSequences_invalid, "literals must be entirely and exactly consumed");
    RETURN_ERROR_IF(remaining != 0, externalSequences_invalid, "Sequences must represent a total of exactly srcSize=%zu", srcSize);
    DEBUGLOG(4, "cSize final total: %zu", cSize);
    return cSize;
}

size_t
ZSTD_compressSequencesAndLiterals(ZSTD_CCtx* cctx,
                    void* dst, size_t dstCapacity,
                    const ZSTD_Sequence* inSeqs, size_t inSeqsSize,
                    const void* literals, size_t litSize, size_t litCapacity,
                    size_t decompressedSize)
{
    BYTE* op = (BYTE*)dst;
    size_t cSize = 0;

    /* Transparent initialization stage, same as compressStream2() */
    DEBUGLOG(4, "ZSTD_compressSequencesAndLiterals (dstCapacity=%zu)", dstCapacity);
    assert(cctx != NULL);
    if (litCapacity < litSize) {
        RETURN_ERROR(workSpace_tooSmall, "literals buffer is not large enough: must be at least 8 bytes larger than litSize (risk of read out-of-bound)");
    }
    FORWARD_IF_ERROR(ZSTD_CCtx_init_compressStream2(cctx, ZSTD_e_end, decompressedSize), "CCtx initialization failed");

    if (cctx->appliedParams.blockDelimiters == ZSTD_sf_noBlockDelimiters) {
        RETURN_ERROR(frameParameter_unsupported, "This mode is only compatible with explicit delimiters");
    }
    if (cctx->appliedParams.validateSequences) {
        RETURN_ERROR(parameter_unsupported, "This mode is not compatible with Sequence validation");
    }
    if (cctx->appliedParams.fParams.checksumFlag) {
        RETURN_ERROR(frameParameter_unsupported, "this mode is not compatible with frame checksum");
    }

    /* Begin writing output, starting with frame header */
    {   size_t const frameHeaderSize = ZSTD_writeFrameHeader(op, dstCapacity,
                    &cctx->appliedParams, decompressedSize, cctx->dictID);
        op += frameHeaderSize;
        assert(frameHeaderSize <= dstCapacity);
        dstCapacity -= frameHeaderSize;
        cSize += frameHeaderSize;
    }

    /* Now generate compressed blocks */
    {   size_t const cBlocksSize = ZSTD_compressSequencesAndLiterals_internal(cctx,
                                            op, dstCapacity,
                                            inSeqs, inSeqsSize,
                                            literals, litSize, decompressedSize);
        FORWARD_IF_ERROR(cBlocksSize, "Compressing blocks failed!");
        cSize += cBlocksSize;
        assert(cBlocksSize <= dstCapacity);
        dstCapacity -= cBlocksSize;
    }

    DEBUGLOG(4, "Final compressed size: %zu", cSize);
    return cSize;
}

/*======   Finalize   ======*/

static ZSTD_inBuffer inBuffer_forEndFlush(const ZSTD_CStream* zcs)
{
    const ZSTD_inBuffer nullInput = { NULL, 0, 0 };
    const int stableInput = (zcs->appliedParams.inBufferMode == ZSTD_bm_stable);
    return stableInput ? zcs->expectedInBuffer : nullInput;
}

/*! ZSTD_flushStream() :
 * @return : amount of data remaining to flush */
size_t ZSTD_flushStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
{
    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
    input.size = input.pos; /* do not ingest more input during flush */
    return ZSTD_compressStream2(zcs, output, &input, ZSTD_e_flush);
}

size_t ZSTD_endStream(ZSTD_CStream* zcs, ZSTD_outBuffer* output)
{
    ZSTD_inBuffer input = inBuffer_forEndFlush(zcs);
    size_t const remainingToFlush = ZSTD_compressStream2(zcs, output, &input, ZSTD_e_end);
    FORWARD_IF_ERROR(remainingToFlush , "ZSTD_compressStream2(,,ZSTD_e_end) failed");
    if (zcs->appliedParams.nbWorkers > 0) return remainingToFlush;   /* minimal estimation */
    /* single thread mode : attempt to calculate remaining to flush more precisely */
    {   size_t const lastBlockSize = zcs->frameEnded ? 0 : ZSTD_BLOCKHEADERSIZE;
        size_t const checksumSize = (size_t)(zcs->frameEnded ? 0 : zcs->appliedParams.fParams.checksumFlag * 4);
        size_t const toFlush = remainingToFlush + lastBlockSize + checksumSize;
        DEBUGLOG(4, "ZSTD_endStream : remaining to flush : %u", (unsigned)toFlush);
        return toFlush;
    }
}


/*-=====  Pre-defined compression levels  =====-*/
/**** start inlining clevels.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_CLEVELS_H
#define ZSTD_CLEVELS_H

#define ZSTD_STATIC_LINKING_ONLY  /* ZSTD_compressionParameters  */
/**** skipping file: ../zstd.h ****/

/*-=====  Pre-defined compression levels  =====-*/

#define ZSTD_MAX_CLEVEL     22

#ifdef __GNUC__
__attribute__((__unused__))
#endif

static const ZSTD_compressionParameters ZSTD_defaultCParameters[4][ZSTD_MAX_CLEVEL+1] = {
{   /* "default" - for any srcSize > 256 KB */
    /* W,  C,  H,  S,  L, TL, strat */
    { 19, 12, 13,  1,  6,  1, ZSTD_fast    },  /* base for negative levels */
    { 19, 13, 14,  1,  7,  0, ZSTD_fast    },  /* level  1 */
    { 20, 15, 16,  1,  6,  0, ZSTD_fast    },  /* level  2 */
    { 21, 16, 17,  1,  5,  0, ZSTD_dfast   },  /* level  3 */
    { 21, 18, 18,  1,  5,  0, ZSTD_dfast   },  /* level  4 */
    { 21, 18, 19,  3,  5,  2, ZSTD_greedy  },  /* level  5 */
    { 21, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6 */
    { 21, 19, 20,  4,  5,  8, ZSTD_lazy    },  /* level  7 */
    { 21, 19, 20,  4,  5, 16, ZSTD_lazy2   },  /* level  8 */
    { 22, 20, 21,  4,  5, 16, ZSTD_lazy2   },  /* level  9 */
    { 22, 21, 22,  5,  5, 16, ZSTD_lazy2   },  /* level 10 */
    { 22, 21, 22,  6,  5, 16, ZSTD_lazy2   },  /* level 11 */
    { 22, 22, 23,  6,  5, 32, ZSTD_lazy2   },  /* level 12 */
    { 22, 22, 22,  4,  5, 32, ZSTD_btlazy2 },  /* level 13 */
    { 22, 22, 23,  5,  5, 32, ZSTD_btlazy2 },  /* level 14 */
    { 22, 23, 23,  6,  5, 32, ZSTD_btlazy2 },  /* level 15 */
    { 22, 22, 22,  5,  5, 48, ZSTD_btopt   },  /* level 16 */
    { 23, 23, 22,  5,  4, 64, ZSTD_btopt   },  /* level 17 */
    { 23, 23, 22,  6,  3, 64, ZSTD_btultra },  /* level 18 */
    { 23, 24, 22,  7,  3,256, ZSTD_btultra2},  /* level 19 */
    { 25, 25, 23,  7,  3,256, ZSTD_btultra2},  /* level 20 */
    { 26, 26, 24,  7,  3,512, ZSTD_btultra2},  /* level 21 */
    { 27, 27, 25,  9,  3,999, ZSTD_btultra2},  /* level 22 */
},
{   /* for srcSize <= 256 KB */
    /* W,  C,  H,  S,  L,  T, strat */
    { 18, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
    { 18, 13, 14,  1,  6,  0, ZSTD_fast    },  /* level  1 */
    { 18, 14, 14,  1,  5,  0, ZSTD_dfast   },  /* level  2 */
    { 18, 16, 16,  1,  4,  0, ZSTD_dfast   },  /* level  3 */
    { 18, 16, 17,  3,  5,  2, ZSTD_greedy  },  /* level  4.*/
    { 18, 17, 18,  5,  5,  2, ZSTD_greedy  },  /* level  5.*/
    { 18, 18, 19,  3,  5,  4, ZSTD_lazy    },  /* level  6.*/
    { 18, 18, 19,  4,  4,  4, ZSTD_lazy    },  /* level  7 */
    { 18, 18, 19,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
    { 18, 18, 19,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
    { 18, 18, 19,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
    { 18, 18, 19,  5,  4, 12, ZSTD_btlazy2 },  /* level 11.*/
    { 18, 19, 19,  7,  4, 12, ZSTD_btlazy2 },  /* level 12.*/
    { 18, 18, 19,  4,  4, 16, ZSTD_btopt   },  /* level 13 */
    { 18, 18, 19,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
    { 18, 18, 19,  6,  3,128, ZSTD_btopt   },  /* level 15.*/
    { 18, 19, 19,  6,  3,128, ZSTD_btultra },  /* level 16.*/
    { 18, 19, 19,  8,  3,256, ZSTD_btultra },  /* level 17.*/
    { 18, 19, 19,  6,  3,128, ZSTD_btultra2},  /* level 18.*/
    { 18, 19, 19,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
    { 18, 19, 19, 10,  3,512, ZSTD_btultra2},  /* level 20.*/
    { 18, 19, 19, 12,  3,512, ZSTD_btultra2},  /* level 21.*/
    { 18, 19, 19, 13,  3,999, ZSTD_btultra2},  /* level 22.*/
},
{   /* for srcSize <= 128 KB */
    /* W,  C,  H,  S,  L,  T, strat */
    { 17, 12, 12,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
    { 17, 12, 13,  1,  6,  0, ZSTD_fast    },  /* level  1 */
    { 17, 13, 15,  1,  5,  0, ZSTD_fast    },  /* level  2 */
    { 17, 15, 16,  2,  5,  0, ZSTD_dfast   },  /* level  3 */
    { 17, 17, 17,  2,  4,  0, ZSTD_dfast   },  /* level  4 */
    { 17, 16, 17,  3,  4,  2, ZSTD_greedy  },  /* level  5 */
    { 17, 16, 17,  3,  4,  4, ZSTD_lazy    },  /* level  6 */
    { 17, 16, 17,  3,  4,  8, ZSTD_lazy2   },  /* level  7 */
    { 17, 16, 17,  4,  4,  8, ZSTD_lazy2   },  /* level  8 */
    { 17, 16, 17,  5,  4,  8, ZSTD_lazy2   },  /* level  9 */
    { 17, 16, 17,  6,  4,  8, ZSTD_lazy2   },  /* level 10 */
    { 17, 17, 17,  5,  4,  8, ZSTD_btlazy2 },  /* level 11 */
    { 17, 18, 17,  7,  4, 12, ZSTD_btlazy2 },  /* level 12 */
    { 17, 18, 17,  3,  4, 12, ZSTD_btopt   },  /* level 13.*/
    { 17, 18, 17,  4,  3, 32, ZSTD_btopt   },  /* level 14.*/
    { 17, 18, 17,  6,  3,256, ZSTD_btopt   },  /* level 15.*/
    { 17, 18, 17,  6,  3,128, ZSTD_btultra },  /* level 16.*/
    { 17, 18, 17,  8,  3,256, ZSTD_btultra },  /* level 17.*/
    { 17, 18, 17, 10,  3,512, ZSTD_btultra },  /* level 18.*/
    { 17, 18, 17,  5,  3,256, ZSTD_btultra2},  /* level 19.*/
    { 17, 18, 17,  7,  3,512, ZSTD_btultra2},  /* level 20.*/
    { 17, 18, 17,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
    { 17, 18, 17, 11,  3,999, ZSTD_btultra2},  /* level 22.*/
},
{   /* for srcSize <= 16 KB */
    /* W,  C,  H,  S,  L,  T, strat */
    { 14, 12, 13,  1,  5,  1, ZSTD_fast    },  /* base for negative levels */
    { 14, 14, 15,  1,  5,  0, ZSTD_fast    },  /* level  1 */
    { 14, 14, 15,  1,  4,  0, ZSTD_fast    },  /* level  2 */
    { 14, 14, 15,  2,  4,  0, ZSTD_dfast   },  /* level  3 */
    { 14, 14, 14,  4,  4,  2, ZSTD_greedy  },  /* level  4 */
    { 14, 14, 14,  3,  4,  4, ZSTD_lazy    },  /* level  5.*/
    { 14, 14, 14,  4,  4,  8, ZSTD_lazy2   },  /* level  6 */
    { 14, 14, 14,  6,  4,  8, ZSTD_lazy2   },  /* level  7 */
    { 14, 14, 14,  8,  4,  8, ZSTD_lazy2   },  /* level  8.*/
    { 14, 15, 14,  5,  4,  8, ZSTD_btlazy2 },  /* level  9.*/
    { 14, 15, 14,  9,  4,  8, ZSTD_btlazy2 },  /* level 10.*/
    { 14, 15, 14,  3,  4, 12, ZSTD_btopt   },  /* level 11.*/
    { 14, 15, 14,  4,  3, 24, ZSTD_btopt   },  /* level 12.*/
    { 14, 15, 14,  5,  3, 32, ZSTD_btultra },  /* level 13.*/
    { 14, 15, 15,  6,  3, 64, ZSTD_btultra },  /* level 14.*/
    { 14, 15, 15,  7,  3,256, ZSTD_btultra },  /* level 15.*/
    { 14, 15, 15,  5,  3, 48, ZSTD_btultra2},  /* level 16.*/
    { 14, 15, 15,  6,  3,128, ZSTD_btultra2},  /* level 17.*/
    { 14, 15, 15,  7,  3,256, ZSTD_btultra2},  /* level 18.*/
    { 14, 15, 15,  8,  3,256, ZSTD_btultra2},  /* level 19.*/
    { 14, 15, 15,  8,  3,512, ZSTD_btultra2},  /* level 20.*/
    { 14, 15, 15,  9,  3,512, ZSTD_btultra2},  /* level 21.*/
    { 14, 15, 15, 10,  3,999, ZSTD_btultra2},  /* level 22.*/
},
};


#endif  /* ZSTD_CLEVELS_H */
/**** ended inlining clevels.h ****/

int ZSTD_maxCLevel(void) { return ZSTD_MAX_CLEVEL; }
int ZSTD_minCLevel(void) { return (int)-ZSTD_TARGETLENGTH_MAX; }
int ZSTD_defaultCLevel(void) { return ZSTD_CLEVEL_DEFAULT; }

static ZSTD_compressionParameters ZSTD_dedicatedDictSearch_getCParams(int const compressionLevel, size_t const dictSize)
{
    ZSTD_compressionParameters cParams = ZSTD_getCParams_internal(compressionLevel, 0, dictSize, ZSTD_cpm_createCDict);
    switch (cParams.strategy) {
        case ZSTD_fast:
        case ZSTD_dfast:
            break;
        case ZSTD_greedy:
        case ZSTD_lazy:
        case ZSTD_lazy2:
            cParams.hashLog += ZSTD_LAZY_DDSS_BUCKET_LOG;
            break;
        case ZSTD_btlazy2:
        case ZSTD_btopt:
        case ZSTD_btultra:
        case ZSTD_btultra2:
            break;
    }
    return cParams;
}

static int ZSTD_dedicatedDictSearch_isSupported(
        ZSTD_compressionParameters const* cParams)
{
    return (cParams->strategy >= ZSTD_greedy)
        && (cParams->strategy <= ZSTD_lazy2)
        && (cParams->hashLog > cParams->chainLog)
        && (cParams->chainLog <= 24);
}

/**
 * Reverses the adjustment applied to cparams when enabling dedicated dict
 * search. This is used to recover the params set to be used in the working
 * context. (Otherwise, those tables would also grow.)
 */
static void ZSTD_dedicatedDictSearch_revertCParams(
        ZSTD_compressionParameters* cParams) {
    switch (cParams->strategy) {
        case ZSTD_fast:
        case ZSTD_dfast:
            break;
        case ZSTD_greedy:
        case ZSTD_lazy:
        case ZSTD_lazy2:
            cParams->hashLog -= ZSTD_LAZY_DDSS_BUCKET_LOG;
            if (cParams->hashLog < ZSTD_HASHLOG_MIN) {
                cParams->hashLog = ZSTD_HASHLOG_MIN;
            }
            break;
        case ZSTD_btlazy2:
        case ZSTD_btopt:
        case ZSTD_btultra:
        case ZSTD_btultra2:
            break;
    }
}

static U64 ZSTD_getCParamRowSize(U64 srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
{
    switch (mode) {
    case ZSTD_cpm_unknown:
    case ZSTD_cpm_noAttachDict:
    case ZSTD_cpm_createCDict:
        break;
    case ZSTD_cpm_attachDict:
        dictSize = 0;
        break;
    default:
        assert(0);
        break;
    }
    {   int const unknown = srcSizeHint == ZSTD_CONTENTSIZE_UNKNOWN;
        size_t const addedSize = unknown && dictSize > 0 ? 500 : 0;
        return unknown && dictSize == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : srcSizeHint+dictSize+addedSize;
    }
}

/*! ZSTD_getCParams_internal() :
 * @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize.
 *  Note: srcSizeHint 0 means 0, use ZSTD_CONTENTSIZE_UNKNOWN for unknown.
 *        Use dictSize == 0 for unknown or unused.
 *  Note: `mode` controls how we treat the `dictSize`. See docs for `ZSTD_CParamMode_e`. */
static ZSTD_compressionParameters ZSTD_getCParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
{
    U64 const rSize = ZSTD_getCParamRowSize(srcSizeHint, dictSize, mode);
    U32 const tableID = (rSize <= 256 KB) + (rSize <= 128 KB) + (rSize <= 16 KB);
    int row;
    DEBUGLOG(5, "ZSTD_getCParams_internal (cLevel=%i)", compressionLevel);

    /* row */
    if (compressionLevel == 0) row = ZSTD_CLEVEL_DEFAULT;   /* 0 == default */
    else if (compressionLevel < 0) row = 0;   /* entry 0 is baseline for fast mode */
    else if (compressionLevel > ZSTD_MAX_CLEVEL) row = ZSTD_MAX_CLEVEL;
    else row = compressionLevel;

    {   ZSTD_compressionParameters cp = ZSTD_defaultCParameters[tableID][row];
        DEBUGLOG(5, "ZSTD_getCParams_internal selected tableID: %u row: %u strat: %u", tableID, row, (U32)cp.strategy);
        /* acceleration factor */
        if (compressionLevel < 0) {
            int const clampedCompressionLevel = MAX(ZSTD_minCLevel(), compressionLevel);
            cp.targetLength = (unsigned)(-clampedCompressionLevel);
        }
        /* refine parameters based on srcSize & dictSize */
        return ZSTD_adjustCParams_internal(cp, srcSizeHint, dictSize, mode, ZSTD_ps_auto);
    }
}

/*! ZSTD_getCParams() :
 * @return ZSTD_compressionParameters structure for a selected compression level, srcSize and dictSize.
 *  Size values are optional, provide 0 if not known or unused */
ZSTD_compressionParameters ZSTD_getCParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
{
    if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
    return ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
}

/*! ZSTD_getParams() :
 *  same idea as ZSTD_getCParams()
 * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`).
 *  Fields of `ZSTD_frameParameters` are set to default values */
static ZSTD_parameters
ZSTD_getParams_internal(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize, ZSTD_CParamMode_e mode)
{
    ZSTD_parameters params;
    ZSTD_compressionParameters const cParams = ZSTD_getCParams_internal(compressionLevel, srcSizeHint, dictSize, mode);
    DEBUGLOG(5, "ZSTD_getParams (cLevel=%i)", compressionLevel);
    ZSTD_memset(&params, 0, sizeof(params));
    params.cParams = cParams;
    params.fParams.contentSizeFlag = 1;
    return params;
}

/*! ZSTD_getParams() :
 *  same idea as ZSTD_getCParams()
 * @return a `ZSTD_parameters` structure (instead of `ZSTD_compressionParameters`).
 *  Fields of `ZSTD_frameParameters` are set to default values */
ZSTD_parameters ZSTD_getParams(int compressionLevel, unsigned long long srcSizeHint, size_t dictSize)
{
    if (srcSizeHint == 0) srcSizeHint = ZSTD_CONTENTSIZE_UNKNOWN;
    return ZSTD_getParams_internal(compressionLevel, srcSizeHint, dictSize, ZSTD_cpm_unknown);
}

void ZSTD_registerSequenceProducer(
    ZSTD_CCtx* zc,
    void* extSeqProdState,
    ZSTD_sequenceProducer_F extSeqProdFunc)
{
    assert(zc != NULL);
    ZSTD_CCtxParams_registerSequenceProducer(
        &zc->requestedParams, extSeqProdState, extSeqProdFunc
    );
}

void ZSTD_CCtxParams_registerSequenceProducer(
  ZSTD_CCtx_params* params,
  void* extSeqProdState,
  ZSTD_sequenceProducer_F extSeqProdFunc)
{
    assert(params != NULL);
    if (extSeqProdFunc != NULL) {
        params->extSeqProdFunc = extSeqProdFunc;
        params->extSeqProdState = extSeqProdState;
    } else {
        params->extSeqProdFunc = NULL;
        params->extSeqProdState = NULL;
    }
}
/**** ended inlining compress/zstd_compress.c ****/
/**** start inlining compress/zstd_double_fast.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_double_fast.h ****/

#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_fillDoubleHashTableForCDict(ZSTD_MatchState_t* ms,
                              void const* end, ZSTD_dictTableLoadMethod_e dtlm)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashLarge = ms->hashTable;
    U32  const hBitsL = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
    U32  const mls = cParams->minMatch;
    U32* const hashSmall = ms->chainTable;
    U32  const hBitsS = cParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
    const BYTE* const base = ms->window.base;
    const BYTE* ip = base + ms->nextToUpdate;
    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
    const U32 fastHashFillStep = 3;

    /* Always insert every fastHashFillStep position into the hash tables.
     * Insert the other positions into the large hash table if their entry
     * is empty.
     */
    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
        U32 const curr = (U32)(ip - base);
        U32 i;
        for (i = 0; i < fastHashFillStep; ++i) {
            size_t const smHashAndTag = ZSTD_hashPtr(ip + i, hBitsS, mls);
            size_t const lgHashAndTag = ZSTD_hashPtr(ip + i, hBitsL, 8);
            if (i == 0) {
                ZSTD_writeTaggedIndex(hashSmall, smHashAndTag, curr + i);
            }
            if (i == 0 || hashLarge[lgHashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {
                ZSTD_writeTaggedIndex(hashLarge, lgHashAndTag, curr + i);
            }
            /* Only load extra positions for ZSTD_dtlm_full */
            if (dtlm == ZSTD_dtlm_fast)
                break;
    }   }
}

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_fillDoubleHashTableForCCtx(ZSTD_MatchState_t* ms,
                              void const* end, ZSTD_dictTableLoadMethod_e dtlm)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashLarge = ms->hashTable;
    U32  const hBitsL = cParams->hashLog;
    U32  const mls = cParams->minMatch;
    U32* const hashSmall = ms->chainTable;
    U32  const hBitsS = cParams->chainLog;
    const BYTE* const base = ms->window.base;
    const BYTE* ip = base + ms->nextToUpdate;
    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
    const U32 fastHashFillStep = 3;

    /* Always insert every fastHashFillStep position into the hash tables.
     * Insert the other positions into the large hash table if their entry
     * is empty.
     */
    for (; ip + fastHashFillStep - 1 <= iend; ip += fastHashFillStep) {
        U32 const curr = (U32)(ip - base);
        U32 i;
        for (i = 0; i < fastHashFillStep; ++i) {
            size_t const smHash = ZSTD_hashPtr(ip + i, hBitsS, mls);
            size_t const lgHash = ZSTD_hashPtr(ip + i, hBitsL, 8);
            if (i == 0)
                hashSmall[smHash] = curr + i;
            if (i == 0 || hashLarge[lgHash] == 0)
                hashLarge[lgHash] = curr + i;
            /* Only load extra positions for ZSTD_dtlm_full */
            if (dtlm == ZSTD_dtlm_fast)
                break;
        }   }
}

void ZSTD_fillDoubleHashTable(ZSTD_MatchState_t* ms,
                        const void* const end,
                        ZSTD_dictTableLoadMethod_e dtlm,
                        ZSTD_tableFillPurpose_e tfp)
{
    if (tfp == ZSTD_tfp_forCDict) {
        ZSTD_fillDoubleHashTableForCDict(ms, end, dtlm);
    } else {
        ZSTD_fillDoubleHashTableForCCtx(ms, end, dtlm);
    }
}


FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_doubleFast_noDict_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize, U32 const mls /* template */)
{
    ZSTD_compressionParameters const* cParams = &ms->cParams;
    U32* const hashLong = ms->hashTable;
    const U32 hBitsL = cParams->hashLog;
    U32* const hashSmall = ms->chainTable;
    const U32 hBitsS = cParams->chainLog;
    const BYTE* const base = ms->window.base;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* anchor = istart;
    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
    /* presumes that, if there is a dictionary, it must be using Attach mode */
    const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
    const BYTE* const prefixLowest = base + prefixLowestIndex;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - HASH_READ_SIZE;
    U32 offset_1=rep[0], offset_2=rep[1];
    U32 offsetSaved1 = 0, offsetSaved2 = 0;

    size_t mLength;
    U32 offset;
    U32 curr;

    /* how many positions to search before increasing step size */
    const size_t kStepIncr = 1 << kSearchStrength;
    /* the position at which to increment the step size if no match is found */
    const BYTE* nextStep;
    size_t step; /* the current step size */

    size_t hl0; /* the long hash at ip */
    size_t hl1; /* the long hash at ip1 */

    U32 idxl0; /* the long match index for ip */
    U32 idxl1; /* the long match index for ip1 */

    const BYTE* matchl0; /* the long match for ip */
    const BYTE* matchs0; /* the short match for ip */
    const BYTE* matchl1; /* the long match for ip1 */
    const BYTE* matchs0_safe; /* matchs0 or safe address */

    const BYTE* ip = istart; /* the current position */
    const BYTE* ip1; /* the next position */
    /* Array of ~random data, should have low probability of matching data
     * we load from here instead of from tables, if matchl0/matchl1 are
     * invalid indices. Used to avoid unpredictable branches. */
    const BYTE dummy[] = {0x12,0x34,0x56,0x78,0x9a,0xbc,0xde,0xf0,0xe2,0xb4};

    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_noDict_generic");

    /* init */
    ip += ((ip - prefixLowest) == 0);
    {
        U32 const current = (U32)(ip - base);
        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, current, cParams->windowLog);
        U32 const maxRep = current - windowLow;
        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
    }

    /* Outer Loop: one iteration per match found and stored */
    while (1) {
        step = 1;
        nextStep = ip + kStepIncr;
        ip1 = ip + step;

        if (ip1 > ilimit) {
            goto _cleanup;
        }

        hl0 = ZSTD_hashPtr(ip, hBitsL, 8);
        idxl0 = hashLong[hl0];
        matchl0 = base + idxl0;

        /* Inner Loop: one iteration per search / position */
        do {
            const size_t hs0 = ZSTD_hashPtr(ip, hBitsS, mls);
            const U32 idxs0 = hashSmall[hs0];
            curr = (U32)(ip-base);
            matchs0 = base + idxs0;

            hashLong[hl0] = hashSmall[hs0] = curr;   /* update hash tables */

            /* check noDict repcode */
            if ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1))) {
                mLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
                ip++;
                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
                goto _match_stored;
            }

            hl1 = ZSTD_hashPtr(ip1, hBitsL, 8);

            /* idxl0 > prefixLowestIndex is a (somewhat) unpredictable branch.
             * However expression below complies into conditional move. Since
             * match is unlikely and we only *branch* on idxl0 > prefixLowestIndex
             * if there is a match, all branches become predictable. */
            {   const BYTE*  const matchl0_safe = ZSTD_selectAddr(idxl0, prefixLowestIndex, matchl0, &dummy[0]);

                /* check prefix long match */
                if (MEM_read64(matchl0_safe) == MEM_read64(ip) && matchl0_safe == matchl0) {
                    mLength = ZSTD_count(ip+8, matchl0+8, iend) + 8;
                    offset = (U32)(ip-matchl0);
                    while (((ip>anchor) & (matchl0>prefixLowest)) && (ip[-1] == matchl0[-1])) { ip--; matchl0--; mLength++; } /* catch up */
                    goto _match_found;
            }   }

            idxl1 = hashLong[hl1];
            matchl1 = base + idxl1;

            /* Same optimization as matchl0 above */
            matchs0_safe = ZSTD_selectAddr(idxs0, prefixLowestIndex, matchs0, &dummy[0]);

            /* check prefix short match */
            if(MEM_read32(matchs0_safe) == MEM_read32(ip) && matchs0_safe == matchs0) {
                  goto _search_next_long;
            }

            if (ip1 >= nextStep) {
                PREFETCH_L1(ip1 + 64);
                PREFETCH_L1(ip1 + 128);
                step++;
                nextStep += kStepIncr;
            }
            ip = ip1;
            ip1 += step;

            hl0 = hl1;
            idxl0 = idxl1;
            matchl0 = matchl1;
    #if defined(__aarch64__)
            PREFETCH_L1(ip+256);
    #endif
        } while (ip1 <= ilimit);

_cleanup:
        /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
         * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
        offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;

        /* save reps for next block */
        rep[0] = offset_1 ? offset_1 : offsetSaved1;
        rep[1] = offset_2 ? offset_2 : offsetSaved2;

        /* Return the last literals size */
        return (size_t)(iend - anchor);

_search_next_long:

        /* short match found: let's check for a longer one */
        mLength = ZSTD_count(ip+4, matchs0+4, iend) + 4;
        offset = (U32)(ip - matchs0);

        /* check long match at +1 position */
        if ((idxl1 > prefixLowestIndex) && (MEM_read64(matchl1) == MEM_read64(ip1))) {
            size_t const l1len = ZSTD_count(ip1+8, matchl1+8, iend) + 8;
            if (l1len > mLength) {
                /* use the long match instead */
                ip = ip1;
                mLength = l1len;
                offset = (U32)(ip-matchl1);
                matchs0 = matchl1;
            }
        }

        while (((ip>anchor) & (matchs0>prefixLowest)) && (ip[-1] == matchs0[-1])) { ip--; matchs0--; mLength++; } /* complete backward */

        /* fall-through */

_match_found: /* requires ip, offset, mLength */
        offset_2 = offset_1;
        offset_1 = offset;

        if (step < 4) {
            /* It is unsafe to write this value back to the hashtable when ip1 is
             * greater than or equal to the new ip we will have after we're done
             * processing this match. Rather than perform that test directly
             * (ip1 >= ip + mLength), which costs speed in practice, we do a simpler
             * more predictable test. The minmatch even if we take a short match is
             * 4 bytes, so as long as step, the distance between ip and ip1
             * (initially) is less than 4, we know ip1 < new ip. */
            hashLong[hl1] = (U32)(ip1 - base);
        }

        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);

_match_stored:
        /* match found */
        ip += mLength;
        anchor = ip;

        if (ip <= ilimit) {
            /* Complementary insertion */
            /* done after iLimit test, as candidates could be > iend-8 */
            {   U32 const indexToInsert = curr+2;
                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
            }

            /* check immediate repcode */
            while ( (ip <= ilimit)
                 && ( (offset_2>0)
                    & (MEM_read32(ip) == MEM_read32(ip - offset_2)) )) {
                /* store sequence */
                size_t const rLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
                U32 const tmpOff = offset_2; offset_2 = offset_1; offset_1 = tmpOff;  /* swap offset_2 <=> offset_1 */
                hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = (U32)(ip-base);
                hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = (U32)(ip-base);
                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                ip += rLength;
                anchor = ip;
                continue;   /* faster when present ... (?) */
            }
        }
    }
}


FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_doubleFast_dictMatchState_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize,
        U32 const mls /* template */)
{
    ZSTD_compressionParameters const* cParams = &ms->cParams;
    U32* const hashLong = ms->hashTable;
    const U32 hBitsL = cParams->hashLog;
    U32* const hashSmall = ms->chainTable;
    const U32 hBitsS = cParams->chainLog;
    const BYTE* const base = ms->window.base;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip = istart;
    const BYTE* anchor = istart;
    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
    /* presumes that, if there is a dictionary, it must be using Attach mode */
    const U32 prefixLowestIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
    const BYTE* const prefixLowest = base + prefixLowestIndex;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - HASH_READ_SIZE;
    U32 offset_1=rep[0], offset_2=rep[1];

    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
    const ZSTD_compressionParameters* const dictCParams = &dms->cParams;
    const U32* const dictHashLong  = dms->hashTable;
    const U32* const dictHashSmall = dms->chainTable;
    const U32 dictStartIndex       = dms->window.dictLimit;
    const BYTE* const dictBase     = dms->window.base;
    const BYTE* const dictStart    = dictBase + dictStartIndex;
    const BYTE* const dictEnd      = dms->window.nextSrc;
    const U32 dictIndexDelta       = prefixLowestIndex - (U32)(dictEnd - dictBase);
    const U32 dictHBitsL           = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
    const U32 dictHBitsS           = dictCParams->chainLog + ZSTD_SHORT_CACHE_TAG_BITS;
    const U32 dictAndPrefixLength  = (U32)((ip - prefixLowest) + (dictEnd - dictStart));

    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_dictMatchState_generic");

    /* if a dictionary is attached, it must be within window range */
    assert(ms->window.dictLimit + (1U << cParams->windowLog) >= endIndex);

    if (ms->prefetchCDictTables) {
        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
        size_t const chainTableBytes = (((size_t)1) << dictCParams->chainLog) * sizeof(U32);
        PREFETCH_AREA(dictHashLong, hashTableBytes);
        PREFETCH_AREA(dictHashSmall, chainTableBytes);
    }

    /* init */
    ip += (dictAndPrefixLength == 0);

    /* dictMatchState repCode checks don't currently handle repCode == 0
     * disabling. */
    assert(offset_1 <= dictAndPrefixLength);
    assert(offset_2 <= dictAndPrefixLength);

    /* Main Search Loop */
    while (ip < ilimit) {   /* < instead of <=, because repcode check at (ip+1) */
        size_t mLength;
        U32 offset;
        size_t const h2 = ZSTD_hashPtr(ip, hBitsL, 8);
        size_t const h = ZSTD_hashPtr(ip, hBitsS, mls);
        size_t const dictHashAndTagL = ZSTD_hashPtr(ip, dictHBitsL, 8);
        size_t const dictHashAndTagS = ZSTD_hashPtr(ip, dictHBitsS, mls);
        U32 const dictMatchIndexAndTagL = dictHashLong[dictHashAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS];
        U32 const dictMatchIndexAndTagS = dictHashSmall[dictHashAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS];
        int const dictTagsMatchL = ZSTD_comparePackedTags(dictMatchIndexAndTagL, dictHashAndTagL);
        int const dictTagsMatchS = ZSTD_comparePackedTags(dictMatchIndexAndTagS, dictHashAndTagS);
        U32 const curr = (U32)(ip-base);
        U32 const matchIndexL = hashLong[h2];
        U32 matchIndexS = hashSmall[h];
        const BYTE* matchLong = base + matchIndexL;
        const BYTE* match = base + matchIndexS;
        const U32 repIndex = curr + 1 - offset_1;
        const BYTE* repMatch = (repIndex < prefixLowestIndex) ?
                               dictBase + (repIndex - dictIndexDelta) :
                               base + repIndex;
        hashLong[h2] = hashSmall[h] = curr;   /* update hash tables */

        /* check repcode */
        if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
            && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
            const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
            ip++;
            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
            goto _match_stored;
        }

        if ((matchIndexL >= prefixLowestIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
            /* check prefix long match */
            mLength = ZSTD_count(ip+8, matchLong+8, iend) + 8;
            offset = (U32)(ip-matchLong);
            while (((ip>anchor) & (matchLong>prefixLowest)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; } /* catch up */
            goto _match_found;
        } else if (dictTagsMatchL) {
            /* check dictMatchState long match */
            U32 const dictMatchIndexL = dictMatchIndexAndTagL >> ZSTD_SHORT_CACHE_TAG_BITS;
            const BYTE* dictMatchL = dictBase + dictMatchIndexL;
            assert(dictMatchL < dictEnd);

            if (dictMatchL > dictStart && MEM_read64(dictMatchL) == MEM_read64(ip)) {
                mLength = ZSTD_count_2segments(ip+8, dictMatchL+8, iend, dictEnd, prefixLowest) + 8;
                offset = (U32)(curr - dictMatchIndexL - dictIndexDelta);
                while (((ip>anchor) & (dictMatchL>dictStart)) && (ip[-1] == dictMatchL[-1])) { ip--; dictMatchL--; mLength++; } /* catch up */
                goto _match_found;
        }   }

        if (matchIndexS > prefixLowestIndex) {
            /* short match  candidate */
            if (MEM_read32(match) == MEM_read32(ip)) {
                goto _search_next_long;
            }
        } else if (dictTagsMatchS) {
            /* check dictMatchState short match */
            U32 const dictMatchIndexS = dictMatchIndexAndTagS >> ZSTD_SHORT_CACHE_TAG_BITS;
            match = dictBase + dictMatchIndexS;
            matchIndexS = dictMatchIndexS + dictIndexDelta;

            if (match > dictStart && MEM_read32(match) == MEM_read32(ip)) {
                goto _search_next_long;
        }   }

        ip += ((ip-anchor) >> kSearchStrength) + 1;
#if defined(__aarch64__)
        PREFETCH_L1(ip+256);
#endif
        continue;

_search_next_long:
        {   size_t const hl3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
            size_t const dictHashAndTagL3 = ZSTD_hashPtr(ip+1, dictHBitsL, 8);
            U32 const matchIndexL3 = hashLong[hl3];
            U32 const dictMatchIndexAndTagL3 = dictHashLong[dictHashAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS];
            int const dictTagsMatchL3 = ZSTD_comparePackedTags(dictMatchIndexAndTagL3, dictHashAndTagL3);
            const BYTE* matchL3 = base + matchIndexL3;
            hashLong[hl3] = curr + 1;

            /* check prefix long +1 match */
            if ((matchIndexL3 >= prefixLowestIndex) && (MEM_read64(matchL3) == MEM_read64(ip+1))) {
                mLength = ZSTD_count(ip+9, matchL3+8, iend) + 8;
                ip++;
                offset = (U32)(ip-matchL3);
                while (((ip>anchor) & (matchL3>prefixLowest)) && (ip[-1] == matchL3[-1])) { ip--; matchL3--; mLength++; } /* catch up */
                goto _match_found;
            } else if (dictTagsMatchL3) {
                /* check dict long +1 match */
                U32 const dictMatchIndexL3 = dictMatchIndexAndTagL3 >> ZSTD_SHORT_CACHE_TAG_BITS;
                const BYTE* dictMatchL3 = dictBase + dictMatchIndexL3;
                assert(dictMatchL3 < dictEnd);
                if (dictMatchL3 > dictStart && MEM_read64(dictMatchL3) == MEM_read64(ip+1)) {
                    mLength = ZSTD_count_2segments(ip+1+8, dictMatchL3+8, iend, dictEnd, prefixLowest) + 8;
                    ip++;
                    offset = (U32)(curr + 1 - dictMatchIndexL3 - dictIndexDelta);
                    while (((ip>anchor) & (dictMatchL3>dictStart)) && (ip[-1] == dictMatchL3[-1])) { ip--; dictMatchL3--; mLength++; } /* catch up */
                    goto _match_found;
        }   }   }

        /* if no long +1 match, explore the short match we found */
        if (matchIndexS < prefixLowestIndex) {
            mLength = ZSTD_count_2segments(ip+4, match+4, iend, dictEnd, prefixLowest) + 4;
            offset = (U32)(curr - matchIndexS);
            while (((ip>anchor) & (match>dictStart)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
        } else {
            mLength = ZSTD_count(ip+4, match+4, iend) + 4;
            offset = (U32)(ip - match);
            while (((ip>anchor) & (match>prefixLowest)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
        }

_match_found:
        offset_2 = offset_1;
        offset_1 = offset;

        ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);

_match_stored:
        /* match found */
        ip += mLength;
        anchor = ip;

        if (ip <= ilimit) {
            /* Complementary insertion */
            /* done after iLimit test, as candidates could be > iend-8 */
            {   U32 const indexToInsert = curr+2;
                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
            }

            /* check immediate repcode */
            while (ip <= ilimit) {
                U32 const current2 = (U32)(ip-base);
                U32 const repIndex2 = current2 - offset_2;
                const BYTE* repMatch2 = repIndex2 < prefixLowestIndex ?
                        dictBase + repIndex2 - dictIndexDelta :
                        base + repIndex2;
                if ( (ZSTD_index_overlap_check(prefixLowestIndex, repIndex2))
                   && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                    const BYTE* const repEnd2 = repIndex2 < prefixLowestIndex ? dictEnd : iend;
                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixLowest) + 4;
                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                    ip += repLength2;
                    anchor = ip;
                    continue;
                }
                break;
            }
        }
    }   /* while (ip < ilimit) */

    /* save reps for next block */
    rep[0] = offset_1;
    rep[1] = offset_2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}

#define ZSTD_GEN_DFAST_FN(dictMode, mls)                                                                 \
    static size_t ZSTD_compressBlock_doubleFast_##dictMode##_##mls(                                      \
            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                          \
            void const* src, size_t srcSize)                                                             \
    {                                                                                                    \
        return ZSTD_compressBlock_doubleFast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mls); \
    }

ZSTD_GEN_DFAST_FN(noDict, 4)
ZSTD_GEN_DFAST_FN(noDict, 5)
ZSTD_GEN_DFAST_FN(noDict, 6)
ZSTD_GEN_DFAST_FN(noDict, 7)

ZSTD_GEN_DFAST_FN(dictMatchState, 4)
ZSTD_GEN_DFAST_FN(dictMatchState, 5)
ZSTD_GEN_DFAST_FN(dictMatchState, 6)
ZSTD_GEN_DFAST_FN(dictMatchState, 7)


size_t ZSTD_compressBlock_doubleFast(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    const U32 mls = ms->cParams.minMatch;
    switch(mls)
    {
    default: /* includes case 3 */
    case 4 :
        return ZSTD_compressBlock_doubleFast_noDict_4(ms, seqStore, rep, src, srcSize);
    case 5 :
        return ZSTD_compressBlock_doubleFast_noDict_5(ms, seqStore, rep, src, srcSize);
    case 6 :
        return ZSTD_compressBlock_doubleFast_noDict_6(ms, seqStore, rep, src, srcSize);
    case 7 :
        return ZSTD_compressBlock_doubleFast_noDict_7(ms, seqStore, rep, src, srcSize);
    }
}


size_t ZSTD_compressBlock_doubleFast_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    const U32 mls = ms->cParams.minMatch;
    switch(mls)
    {
    default: /* includes case 3 */
    case 4 :
        return ZSTD_compressBlock_doubleFast_dictMatchState_4(ms, seqStore, rep, src, srcSize);
    case 5 :
        return ZSTD_compressBlock_doubleFast_dictMatchState_5(ms, seqStore, rep, src, srcSize);
    case 6 :
        return ZSTD_compressBlock_doubleFast_dictMatchState_6(ms, seqStore, rep, src, srcSize);
    case 7 :
        return ZSTD_compressBlock_doubleFast_dictMatchState_7(ms, seqStore, rep, src, srcSize);
    }
}


static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_doubleFast_extDict_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize,
        U32 const mls /* template */)
{
    ZSTD_compressionParameters const* cParams = &ms->cParams;
    U32* const hashLong = ms->hashTable;
    U32  const hBitsL = cParams->hashLog;
    U32* const hashSmall = ms->chainTable;
    U32  const hBitsS = cParams->chainLog;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip = istart;
    const BYTE* anchor = istart;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - 8;
    const BYTE* const base = ms->window.base;
    const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
    const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
    const U32   dictStartIndex = lowLimit;
    const U32   dictLimit = ms->window.dictLimit;
    const U32   prefixStartIndex = (dictLimit > lowLimit) ? dictLimit : lowLimit;
    const BYTE* const prefixStart = base + prefixStartIndex;
    const BYTE* const dictBase = ms->window.dictBase;
    const BYTE* const dictStart = dictBase + dictStartIndex;
    const BYTE* const dictEnd = dictBase + prefixStartIndex;
    U32 offset_1=rep[0], offset_2=rep[1];

    DEBUGLOG(5, "ZSTD_compressBlock_doubleFast_extDict_generic (srcSize=%zu)", srcSize);

    /* if extDict is invalidated due to maxDistance, switch to "regular" variant */
    if (prefixStartIndex == dictStartIndex)
        return ZSTD_compressBlock_doubleFast(ms, seqStore, rep, src, srcSize);

    /* Search Loop */
    while (ip < ilimit) {  /* < instead of <=, because (ip+1) */
        const size_t hSmall = ZSTD_hashPtr(ip, hBitsS, mls);
        const U32 matchIndex = hashSmall[hSmall];
        const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
        const BYTE* match = matchBase + matchIndex;

        const size_t hLong = ZSTD_hashPtr(ip, hBitsL, 8);
        const U32 matchLongIndex = hashLong[hLong];
        const BYTE* const matchLongBase = matchLongIndex < prefixStartIndex ? dictBase : base;
        const BYTE* matchLong = matchLongBase + matchLongIndex;

        const U32 curr = (U32)(ip-base);
        const U32 repIndex = curr + 1 - offset_1;   /* offset_1 expected <= curr +1 */
        const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
        const BYTE* const repMatch = repBase + repIndex;
        size_t mLength;
        hashSmall[hSmall] = hashLong[hLong] = curr;   /* update hash table */

        if (((ZSTD_index_overlap_check(prefixStartIndex, repIndex))
            & (offset_1 <= curr+1 - dictStartIndex)) /* note: we are searching at curr+1 */
          && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
            const BYTE* repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
            mLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixStart) + 4;
            ip++;
            ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
        } else {
            if ((matchLongIndex > dictStartIndex) && (MEM_read64(matchLong) == MEM_read64(ip))) {
                const BYTE* const matchEnd = matchLongIndex < prefixStartIndex ? dictEnd : iend;
                const BYTE* const lowMatchPtr = matchLongIndex < prefixStartIndex ? dictStart : prefixStart;
                U32 offset;
                mLength = ZSTD_count_2segments(ip+8, matchLong+8, iend, matchEnd, prefixStart) + 8;
                offset = curr - matchLongIndex;
                while (((ip>anchor) & (matchLong>lowMatchPtr)) && (ip[-1] == matchLong[-1])) { ip--; matchLong--; mLength++; }   /* catch up */
                offset_2 = offset_1;
                offset_1 = offset;
                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);

            } else if ((matchIndex > dictStartIndex) && (MEM_read32(match) == MEM_read32(ip))) {
                size_t const h3 = ZSTD_hashPtr(ip+1, hBitsL, 8);
                U32 const matchIndex3 = hashLong[h3];
                const BYTE* const match3Base = matchIndex3 < prefixStartIndex ? dictBase : base;
                const BYTE* match3 = match3Base + matchIndex3;
                U32 offset;
                hashLong[h3] = curr + 1;
                if ( (matchIndex3 > dictStartIndex) && (MEM_read64(match3) == MEM_read64(ip+1)) ) {
                    const BYTE* const matchEnd = matchIndex3 < prefixStartIndex ? dictEnd : iend;
                    const BYTE* const lowMatchPtr = matchIndex3 < prefixStartIndex ? dictStart : prefixStart;
                    mLength = ZSTD_count_2segments(ip+9, match3+8, iend, matchEnd, prefixStart) + 8;
                    ip++;
                    offset = curr+1 - matchIndex3;
                    while (((ip>anchor) & (match3>lowMatchPtr)) && (ip[-1] == match3[-1])) { ip--; match3--; mLength++; } /* catch up */
                } else {
                    const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
                    const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
                    mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
                    offset = curr - matchIndex;
                    while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; }   /* catch up */
                }
                offset_2 = offset_1;
                offset_1 = offset;
                ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);

            } else {
                ip += ((ip-anchor) >> kSearchStrength) + 1;
                continue;
        }   }

        /* move to next sequence start */
        ip += mLength;
        anchor = ip;

        if (ip <= ilimit) {
            /* Complementary insertion */
            /* done after iLimit test, as candidates could be > iend-8 */
            {   U32 const indexToInsert = curr+2;
                hashLong[ZSTD_hashPtr(base+indexToInsert, hBitsL, 8)] = indexToInsert;
                hashLong[ZSTD_hashPtr(ip-2, hBitsL, 8)] = (U32)(ip-2-base);
                hashSmall[ZSTD_hashPtr(base+indexToInsert, hBitsS, mls)] = indexToInsert;
                hashSmall[ZSTD_hashPtr(ip-1, hBitsS, mls)] = (U32)(ip-1-base);
            }

            /* check immediate repcode */
            while (ip <= ilimit) {
                U32 const current2 = (U32)(ip-base);
                U32 const repIndex2 = current2 - offset_2;
                const BYTE* repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
                if ( ((ZSTD_index_overlap_check(prefixStartIndex, repIndex2))
                    & (offset_2 <= current2 - dictStartIndex))
                  && (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                    size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                    U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                    hashSmall[ZSTD_hashPtr(ip, hBitsS, mls)] = current2;
                    hashLong[ZSTD_hashPtr(ip, hBitsL, 8)] = current2;
                    ip += repLength2;
                    anchor = ip;
                    continue;
                }
                break;
    }   }   }

    /* save reps for next block */
    rep[0] = offset_1;
    rep[1] = offset_2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}

ZSTD_GEN_DFAST_FN(extDict, 4)
ZSTD_GEN_DFAST_FN(extDict, 5)
ZSTD_GEN_DFAST_FN(extDict, 6)
ZSTD_GEN_DFAST_FN(extDict, 7)

size_t ZSTD_compressBlock_doubleFast_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    U32 const mls = ms->cParams.minMatch;
    switch(mls)
    {
    default: /* includes case 3 */
    case 4 :
        return ZSTD_compressBlock_doubleFast_extDict_4(ms, seqStore, rep, src, srcSize);
    case 5 :
        return ZSTD_compressBlock_doubleFast_extDict_5(ms, seqStore, rep, src, srcSize);
    case 6 :
        return ZSTD_compressBlock_doubleFast_extDict_6(ms, seqStore, rep, src, srcSize);
    case 7 :
        return ZSTD_compressBlock_doubleFast_extDict_7(ms, seqStore, rep, src, srcSize);
    }
}

#endif /* ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR */
/**** ended inlining compress/zstd_double_fast.c ****/
/**** start inlining compress/zstd_fast.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_fast.h ****/

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_fillHashTableForCDict(ZSTD_MatchState_t* ms,
                        const void* const end,
                        ZSTD_dictTableLoadMethod_e dtlm)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32  const hBits = cParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;
    U32  const mls = cParams->minMatch;
    const BYTE* const base = ms->window.base;
    const BYTE* ip = base + ms->nextToUpdate;
    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
    const U32 fastHashFillStep = 3;

    /* Currently, we always use ZSTD_dtlm_full for filling CDict tables.
     * Feel free to remove this assert if there's a good reason! */
    assert(dtlm == ZSTD_dtlm_full);

    /* Always insert every fastHashFillStep position into the hash table.
     * Insert the other positions if their hash entry is empty.
     */
    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
        U32 const curr = (U32)(ip - base);
        {   size_t const hashAndTag = ZSTD_hashPtr(ip, hBits, mls);
            ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr);   }

        if (dtlm == ZSTD_dtlm_fast) continue;
        /* Only load extra positions for ZSTD_dtlm_full */
        {   U32 p;
            for (p = 1; p < fastHashFillStep; ++p) {
                size_t const hashAndTag = ZSTD_hashPtr(ip + p, hBits, mls);
                if (hashTable[hashAndTag >> ZSTD_SHORT_CACHE_TAG_BITS] == 0) {  /* not yet filled */
                    ZSTD_writeTaggedIndex(hashTable, hashAndTag, curr + p);
    }   }   }   }
}

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_fillHashTableForCCtx(ZSTD_MatchState_t* ms,
                        const void* const end,
                        ZSTD_dictTableLoadMethod_e dtlm)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32  const hBits = cParams->hashLog;
    U32  const mls = cParams->minMatch;
    const BYTE* const base = ms->window.base;
    const BYTE* ip = base + ms->nextToUpdate;
    const BYTE* const iend = ((const BYTE*)end) - HASH_READ_SIZE;
    const U32 fastHashFillStep = 3;

    /* Currently, we always use ZSTD_dtlm_fast for filling CCtx tables.
     * Feel free to remove this assert if there's a good reason! */
    assert(dtlm == ZSTD_dtlm_fast);

    /* Always insert every fastHashFillStep position into the hash table.
     * Insert the other positions if their hash entry is empty.
     */
    for ( ; ip + fastHashFillStep < iend + 2; ip += fastHashFillStep) {
        U32 const curr = (U32)(ip - base);
        size_t const hash0 = ZSTD_hashPtr(ip, hBits, mls);
        hashTable[hash0] = curr;
        if (dtlm == ZSTD_dtlm_fast) continue;
        /* Only load extra positions for ZSTD_dtlm_full */
        {   U32 p;
            for (p = 1; p < fastHashFillStep; ++p) {
                size_t const hash = ZSTD_hashPtr(ip + p, hBits, mls);
                if (hashTable[hash] == 0) {  /* not yet filled */
                    hashTable[hash] = curr + p;
    }   }   }   }
}

void ZSTD_fillHashTable(ZSTD_MatchState_t* ms,
                        const void* const end,
                        ZSTD_dictTableLoadMethod_e dtlm,
                        ZSTD_tableFillPurpose_e tfp)
{
    if (tfp == ZSTD_tfp_forCDict) {
        ZSTD_fillHashTableForCDict(ms, end, dtlm);
    } else {
        ZSTD_fillHashTableForCCtx(ms, end, dtlm);
    }
}


typedef int (*ZSTD_match4Found) (const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit);

static int
ZSTD_match4Found_cmov(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit)
{
    /* Array of ~random data, should have low probability of matching data.
     * Load from here if the index is invalid.
     * Used to avoid unpredictable branches. */
    static const BYTE dummy[] = {0x12,0x34,0x56,0x78};

    /* currentIdx >= lowLimit is a (somewhat) unpredictable branch.
     * However expression below compiles into conditional move.
     */
    const BYTE* mvalAddr = ZSTD_selectAddr(matchIdx, idxLowLimit, matchAddress, dummy);
    /* Note: this used to be written as : return test1 && test2;
     * Unfortunately, once inlined, these tests become branches,
     * in which case it becomes critical that they are executed in the right order (test1 then test2).
     * So we have to write these tests in a specific manner to ensure their ordering.
     */
    if (MEM_read32(currentPtr) != MEM_read32(mvalAddr)) return 0;
    /* force ordering of these tests, which matters once the function is inlined, as they become branches */
#if defined(__GNUC__)
    __asm__("");
#endif
    return matchIdx >= idxLowLimit;
}

static int
ZSTD_match4Found_branch(const BYTE* currentPtr, const BYTE* matchAddress, U32 matchIdx, U32 idxLowLimit)
{
    /* using a branch instead of a cmov,
     * because it's faster in scenarios where matchIdx >= idxLowLimit is generally true,
     * aka almost all candidates are within range */
    U32 mval;
    if (matchIdx >= idxLowLimit) {
        mval = MEM_read32(matchAddress);
    } else {
        mval = MEM_read32(currentPtr) ^ 1; /* guaranteed to not match. */
    }

    return (MEM_read32(currentPtr) == mval);
}


/**
 * If you squint hard enough (and ignore repcodes), the search operation at any
 * given position is broken into 4 stages:
 *
 * 1. Hash   (map position to hash value via input read)
 * 2. Lookup (map hash val to index via hashtable read)
 * 3. Load   (map index to value at that position via input read)
 * 4. Compare
 *
 * Each of these steps involves a memory read at an address which is computed
 * from the previous step. This means these steps must be sequenced and their
 * latencies are cumulative.
 *
 * Rather than do 1->2->3->4 sequentially for a single position before moving
 * onto the next, this implementation interleaves these operations across the
 * next few positions:
 *
 * R = Repcode Read & Compare
 * H = Hash
 * T = Table Lookup
 * M = Match Read & Compare
 *
 * Pos | Time -->
 * ----+-------------------
 * N   | ... M
 * N+1 | ...   TM
 * N+2 |    R H   T M
 * N+3 |         H    TM
 * N+4 |           R H   T M
 * N+5 |                H   ...
 * N+6 |                  R ...
 *
 * This is very much analogous to the pipelining of execution in a CPU. And just
 * like a CPU, we have to dump the pipeline when we find a match (i.e., take a
 * branch).
 *
 * When this happens, we throw away our current state, and do the following prep
 * to re-enter the loop:
 *
 * Pos | Time -->
 * ----+-------------------
 * N   | H T
 * N+1 |  H
 *
 * This is also the work we do at the beginning to enter the loop initially.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_fast_noDict_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize,
        U32 const mls, int useCmov)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32 const hlog = cParams->hashLog;
    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1; /* min 2 */
    const BYTE* const base = ms->window.base;
    const BYTE* const istart = (const BYTE*)src;
    const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
    const U32   prefixStartIndex = ZSTD_getLowestPrefixIndex(ms, endIndex, cParams->windowLog);
    const BYTE* const prefixStart = base + prefixStartIndex;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - HASH_READ_SIZE;

    const BYTE* anchor = istart;
    const BYTE* ip0 = istart;
    const BYTE* ip1;
    const BYTE* ip2;
    const BYTE* ip3;
    U32 current0;

    U32 rep_offset1 = rep[0];
    U32 rep_offset2 = rep[1];
    U32 offsetSaved1 = 0, offsetSaved2 = 0;

    size_t hash0; /* hash for ip0 */
    size_t hash1; /* hash for ip1 */
    U32 matchIdx; /* match idx for ip0 */

    U32 offcode;
    const BYTE* match0;
    size_t mLength;

    /* ip0 and ip1 are always adjacent. The targetLength skipping and
     * uncompressibility acceleration is applied to every other position,
     * matching the behavior of #1562. step therefore represents the gap
     * between pairs of positions, from ip0 to ip2 or ip1 to ip3. */
    size_t step;
    const BYTE* nextStep;
    const size_t kStepIncr = (1 << (kSearchStrength - 1));
    const ZSTD_match4Found matchFound = useCmov ? ZSTD_match4Found_cmov : ZSTD_match4Found_branch;

    DEBUGLOG(5, "ZSTD_compressBlock_fast_generic");
    ip0 += (ip0 == prefixStart);
    {   U32 const curr = (U32)(ip0 - base);
        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, cParams->windowLog);
        U32 const maxRep = curr - windowLow;
        if (rep_offset2 > maxRep) offsetSaved2 = rep_offset2, rep_offset2 = 0;
        if (rep_offset1 > maxRep) offsetSaved1 = rep_offset1, rep_offset1 = 0;
    }

    /* start each op */
_start: /* Requires: ip0 */

    step = stepSize;
    nextStep = ip0 + kStepIncr;

    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
    ip1 = ip0 + 1;
    ip2 = ip0 + step;
    ip3 = ip2 + 1;

    if (ip3 >= ilimit) {
        goto _cleanup;
    }

    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
    hash1 = ZSTD_hashPtr(ip1, hlog, mls);

    matchIdx = hashTable[hash0];

    do {
        /* load repcode match for ip[2]*/
        const U32 rval = MEM_read32(ip2 - rep_offset1);

        /* write back hash table entry */
        current0 = (U32)(ip0 - base);
        hashTable[hash0] = current0;

        /* check repcode at ip[2] */
        if ((MEM_read32(ip2) == rval) & (rep_offset1 > 0)) {
            ip0 = ip2;
            match0 = ip0 - rep_offset1;
            mLength = ip0[-1] == match0[-1];
            ip0 -= mLength;
            match0 -= mLength;
            offcode = REPCODE1_TO_OFFBASE;
            mLength += 4;

            /* Write next hash table entry: it's already calculated.
             * This write is known to be safe because ip1 is before the
             * repcode (ip2). */
            hashTable[hash1] = (U32)(ip1 - base);

            goto _match;
        }

         if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) {
            /* Write next hash table entry (it's already calculated).
            * This write is known to be safe because the ip1 == ip0 + 1,
            * so searching will resume after ip1 */
            hashTable[hash1] = (U32)(ip1 - base);

            goto _offset;
        }

        /* lookup ip[1] */
        matchIdx = hashTable[hash1];

        /* hash ip[2] */
        hash0 = hash1;
        hash1 = ZSTD_hashPtr(ip2, hlog, mls);

        /* advance to next positions */
        ip0 = ip1;
        ip1 = ip2;
        ip2 = ip3;

        /* write back hash table entry */
        current0 = (U32)(ip0 - base);
        hashTable[hash0] = current0;

         if (matchFound(ip0, base + matchIdx, matchIdx, prefixStartIndex)) {
            /* Write next hash table entry, since it's already calculated */
            if (step <= 4) {
                /* Avoid writing an index if it's >= position where search will resume.
                * The minimum possible match has length 4, so search can resume at ip0 + 4.
                */
                hashTable[hash1] = (U32)(ip1 - base);
            }
            goto _offset;
        }

        /* lookup ip[1] */
        matchIdx = hashTable[hash1];

        /* hash ip[2] */
        hash0 = hash1;
        hash1 = ZSTD_hashPtr(ip2, hlog, mls);

        /* advance to next positions */
        ip0 = ip1;
        ip1 = ip2;
        ip2 = ip0 + step;
        ip3 = ip1 + step;

        /* calculate step */
        if (ip2 >= nextStep) {
            step++;
            PREFETCH_L1(ip1 + 64);
            PREFETCH_L1(ip1 + 128);
            nextStep += kStepIncr;
        }
    } while (ip3 < ilimit);

_cleanup:
    /* Note that there are probably still a couple positions one could search.
     * However, it seems to be a meaningful performance hit to try to search
     * them. So let's not. */

    /* When the repcodes are outside of the prefix, we set them to zero before the loop.
     * When the offsets are still zero, we need to restore them after the block to have a correct
     * repcode history. If only one offset was invalid, it is easy. The tricky case is when both
     * offsets were invalid. We need to figure out which offset to refill with.
     *     - If both offsets are zero they are in the same order.
     *     - If both offsets are non-zero, we won't restore the offsets from `offsetSaved[12]`.
     *     - If only one is zero, we need to decide which offset to restore.
     *         - If rep_offset1 is non-zero, then rep_offset2 must be offsetSaved1.
     *         - It is impossible for rep_offset2 to be non-zero.
     *
     * So if rep_offset1 started invalid (offsetSaved1 != 0) and became valid (rep_offset1 != 0), then
     * set rep[0] = rep_offset1 and rep[1] = offsetSaved1.
     */
    offsetSaved2 = ((offsetSaved1 != 0) && (rep_offset1 != 0)) ? offsetSaved1 : offsetSaved2;

    /* save reps for next block */
    rep[0] = rep_offset1 ? rep_offset1 : offsetSaved1;
    rep[1] = rep_offset2 ? rep_offset2 : offsetSaved2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);

_offset: /* Requires: ip0, idx */

    /* Compute the offset code. */
    match0 = base + matchIdx;
    rep_offset2 = rep_offset1;
    rep_offset1 = (U32)(ip0-match0);
    offcode = OFFSET_TO_OFFBASE(rep_offset1);
    mLength = 4;

    /* Count the backwards match length. */
    while (((ip0>anchor) & (match0>prefixStart)) && (ip0[-1] == match0[-1])) {
        ip0--;
        match0--;
        mLength++;
    }

_match: /* Requires: ip0, match0, offcode */

    /* Count the forward length. */
    mLength += ZSTD_count(ip0 + mLength, match0 + mLength, iend);

    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);

    ip0 += mLength;
    anchor = ip0;

    /* Fill table and check for immediate repcode. */
    if (ip0 <= ilimit) {
        /* Fill Table */
        assert(base+current0+2 > istart);  /* check base overflow */
        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);

        if (rep_offset2 > 0) { /* rep_offset2==0 means rep_offset2 is invalidated */
            while ( (ip0 <= ilimit) && (MEM_read32(ip0) == MEM_read32(ip0 - rep_offset2)) ) {
                /* store sequence */
                size_t const rLength = ZSTD_count(ip0+4, ip0+4-rep_offset2, iend) + 4;
                { U32 const tmpOff = rep_offset2; rep_offset2 = rep_offset1; rep_offset1 = tmpOff; } /* swap rep_offset2 <=> rep_offset1 */
                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
                ip0 += rLength;
                ZSTD_storeSeq(seqStore, 0 /*litLen*/, anchor, iend, REPCODE1_TO_OFFBASE, rLength);
                anchor = ip0;
                continue;   /* faster when present (confirmed on gcc-8) ... (?) */
    }   }   }

    goto _start;
}

#define ZSTD_GEN_FAST_FN(dictMode, mml, cmov)                                                       \
    static size_t ZSTD_compressBlock_fast_##dictMode##_##mml##_##cmov(                              \
            ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],                    \
            void const* src, size_t srcSize)                                                       \
    {                                                                                              \
        return ZSTD_compressBlock_fast_##dictMode##_generic(ms, seqStore, rep, src, srcSize, mml, cmov); \
    }

ZSTD_GEN_FAST_FN(noDict, 4, 1)
ZSTD_GEN_FAST_FN(noDict, 5, 1)
ZSTD_GEN_FAST_FN(noDict, 6, 1)
ZSTD_GEN_FAST_FN(noDict, 7, 1)

ZSTD_GEN_FAST_FN(noDict, 4, 0)
ZSTD_GEN_FAST_FN(noDict, 5, 0)
ZSTD_GEN_FAST_FN(noDict, 6, 0)
ZSTD_GEN_FAST_FN(noDict, 7, 0)

size_t ZSTD_compressBlock_fast(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    U32 const mml = ms->cParams.minMatch;
    /* use cmov when "candidate in range" branch is likely unpredictable */
    int const useCmov = ms->cParams.windowLog < 19;
    assert(ms->dictMatchState == NULL);
    if (useCmov) {
        switch(mml)
        {
        default: /* includes case 3 */
        case 4 :
            return ZSTD_compressBlock_fast_noDict_4_1(ms, seqStore, rep, src, srcSize);
        case 5 :
            return ZSTD_compressBlock_fast_noDict_5_1(ms, seqStore, rep, src, srcSize);
        case 6 :
            return ZSTD_compressBlock_fast_noDict_6_1(ms, seqStore, rep, src, srcSize);
        case 7 :
            return ZSTD_compressBlock_fast_noDict_7_1(ms, seqStore, rep, src, srcSize);
        }
    } else {
        /* use a branch instead */
        switch(mml)
        {
        default: /* includes case 3 */
        case 4 :
            return ZSTD_compressBlock_fast_noDict_4_0(ms, seqStore, rep, src, srcSize);
        case 5 :
            return ZSTD_compressBlock_fast_noDict_5_0(ms, seqStore, rep, src, srcSize);
        case 6 :
            return ZSTD_compressBlock_fast_noDict_6_0(ms, seqStore, rep, src, srcSize);
        case 7 :
            return ZSTD_compressBlock_fast_noDict_7_0(ms, seqStore, rep, src, srcSize);
        }
    }
}

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_fast_dictMatchState_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32 const hlog = cParams->hashLog;
    /* support stepSize of 0 */
    U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
    const BYTE* const base = ms->window.base;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip0 = istart;
    const BYTE* ip1 = ip0 + stepSize; /* we assert below that stepSize >= 1 */
    const BYTE* anchor = istart;
    const U32   prefixStartIndex = ms->window.dictLimit;
    const BYTE* const prefixStart = base + prefixStartIndex;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - HASH_READ_SIZE;
    U32 offset_1=rep[0], offset_2=rep[1];

    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
    const ZSTD_compressionParameters* const dictCParams = &dms->cParams ;
    const U32* const dictHashTable = dms->hashTable;
    const U32 dictStartIndex       = dms->window.dictLimit;
    const BYTE* const dictBase     = dms->window.base;
    const BYTE* const dictStart    = dictBase + dictStartIndex;
    const BYTE* const dictEnd      = dms->window.nextSrc;
    const U32 dictIndexDelta       = prefixStartIndex - (U32)(dictEnd - dictBase);
    const U32 dictAndPrefixLength  = (U32)(istart - prefixStart + dictEnd - dictStart);
    const U32 dictHBits            = dictCParams->hashLog + ZSTD_SHORT_CACHE_TAG_BITS;

    /* if a dictionary is still attached, it necessarily means that
     * it is within window size. So we just check it. */
    const U32 maxDistance = 1U << cParams->windowLog;
    const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
    assert(endIndex - prefixStartIndex <= maxDistance);
    (void)maxDistance; (void)endIndex;   /* these variables are not used when assert() is disabled */

    (void)hasStep; /* not currently specialized on whether it's accelerated */

    /* ensure there will be no underflow
     * when translating a dict index into a local index */
    assert(prefixStartIndex >= (U32)(dictEnd - dictBase));

    if (ms->prefetchCDictTables) {
        size_t const hashTableBytes = (((size_t)1) << dictCParams->hashLog) * sizeof(U32);
        PREFETCH_AREA(dictHashTable, hashTableBytes);
    }

    /* init */
    DEBUGLOG(5, "ZSTD_compressBlock_fast_dictMatchState_generic");
    ip0 += (dictAndPrefixLength == 0);
    /* dictMatchState repCode checks don't currently handle repCode == 0
     * disabling. */
    assert(offset_1 <= dictAndPrefixLength);
    assert(offset_2 <= dictAndPrefixLength);

    /* Outer search loop */
    assert(stepSize >= 1);
    while (ip1 <= ilimit) {   /* repcode check at (ip0 + 1) is safe because ip0 < ip1 */
        size_t mLength;
        size_t hash0 = ZSTD_hashPtr(ip0, hlog, mls);

        size_t const dictHashAndTag0 = ZSTD_hashPtr(ip0, dictHBits, mls);
        U32 dictMatchIndexAndTag = dictHashTable[dictHashAndTag0 >> ZSTD_SHORT_CACHE_TAG_BITS];
        int dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag0);

        U32 matchIndex = hashTable[hash0];
        U32 curr = (U32)(ip0 - base);
        size_t step = stepSize;
        const size_t kStepIncr = 1 << kSearchStrength;
        const BYTE* nextStep = ip0 + kStepIncr;

        /* Inner search loop */
        while (1) {
            const BYTE* match = base + matchIndex;
            const U32 repIndex = curr + 1 - offset_1;
            const BYTE* repMatch = (repIndex < prefixStartIndex) ?
                                   dictBase + (repIndex - dictIndexDelta) :
                                   base + repIndex;
            const size_t hash1 = ZSTD_hashPtr(ip1, hlog, mls);
            size_t const dictHashAndTag1 = ZSTD_hashPtr(ip1, dictHBits, mls);
            hashTable[hash0] = curr;   /* update hash table */

            if ((ZSTD_index_overlap_check(prefixStartIndex, repIndex))
                && (MEM_read32(repMatch) == MEM_read32(ip0 + 1))) {
                const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
                mLength = ZSTD_count_2segments(ip0 + 1 + 4, repMatch + 4, iend, repMatchEnd, prefixStart) + 4;
                ip0++;
                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, REPCODE1_TO_OFFBASE, mLength);
                break;
            }

            if (dictTagsMatch) {
                /* Found a possible dict match */
                const U32 dictMatchIndex = dictMatchIndexAndTag >> ZSTD_SHORT_CACHE_TAG_BITS;
                const BYTE* dictMatch = dictBase + dictMatchIndex;
                if (dictMatchIndex > dictStartIndex &&
                    MEM_read32(dictMatch) == MEM_read32(ip0)) {
                    /* To replicate extDict parse behavior, we only use dict matches when the normal matchIndex is invalid */
                    if (matchIndex <= prefixStartIndex) {
                        U32 const offset = (U32) (curr - dictMatchIndex - dictIndexDelta);
                        mLength = ZSTD_count_2segments(ip0 + 4, dictMatch + 4, iend, dictEnd, prefixStart) + 4;
                        while (((ip0 > anchor) & (dictMatch > dictStart))
                            && (ip0[-1] == dictMatch[-1])) {
                            ip0--;
                            dictMatch--;
                            mLength++;
                        } /* catch up */
                        offset_2 = offset_1;
                        offset_1 = offset;
                        ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
                        break;
                    }
                }
            }

            if (ZSTD_match4Found_cmov(ip0, match, matchIndex, prefixStartIndex)) {
                /* found a regular match of size >= 4 */
                U32 const offset = (U32) (ip0 - match);
                mLength = ZSTD_count(ip0 + 4, match + 4, iend) + 4;
                while (((ip0 > anchor) & (match > prefixStart))
                       && (ip0[-1] == match[-1])) {
                    ip0--;
                    match--;
                    mLength++;
                } /* catch up */
                offset_2 = offset_1;
                offset_1 = offset;
                ZSTD_storeSeq(seqStore, (size_t) (ip0 - anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
                break;
            }

            /* Prepare for next iteration */
            dictMatchIndexAndTag = dictHashTable[dictHashAndTag1 >> ZSTD_SHORT_CACHE_TAG_BITS];
            dictTagsMatch = ZSTD_comparePackedTags(dictMatchIndexAndTag, dictHashAndTag1);
            matchIndex = hashTable[hash1];

            if (ip1 >= nextStep) {
                step++;
                nextStep += kStepIncr;
            }
            ip0 = ip1;
            ip1 = ip1 + step;
            if (ip1 > ilimit) goto _cleanup;

            curr = (U32)(ip0 - base);
            hash0 = hash1;
        }   /* end inner search loop */

        /* match found */
        assert(mLength);
        ip0 += mLength;
        anchor = ip0;

        if (ip0 <= ilimit) {
            /* Fill Table */
            assert(base+curr+2 > istart);  /* check base overflow */
            hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;  /* here because curr+2 could be > iend-8 */
            hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);

            /* check immediate repcode */
            while (ip0 <= ilimit) {
                U32 const current2 = (U32)(ip0-base);
                U32 const repIndex2 = current2 - offset_2;
                const BYTE* repMatch2 = repIndex2 < prefixStartIndex ?
                        dictBase - dictIndexDelta + repIndex2 :
                        base + repIndex2;
                if ( (ZSTD_index_overlap_check(prefixStartIndex, repIndex2))
                   && (MEM_read32(repMatch2) == MEM_read32(ip0))) {
                    const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                    size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                    U32 tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset;   /* swap offset_2 <=> offset_1 */
                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                    hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = current2;
                    ip0 += repLength2;
                    anchor = ip0;
                    continue;
                }
                break;
            }
        }

        /* Prepare for next iteration */
        assert(ip0 == anchor);
        ip1 = ip0 + stepSize;
    }

_cleanup:
    /* save reps for next block */
    rep[0] = offset_1;
    rep[1] = offset_2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}


ZSTD_GEN_FAST_FN(dictMatchState, 4, 0)
ZSTD_GEN_FAST_FN(dictMatchState, 5, 0)
ZSTD_GEN_FAST_FN(dictMatchState, 6, 0)
ZSTD_GEN_FAST_FN(dictMatchState, 7, 0)

size_t ZSTD_compressBlock_fast_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    U32 const mls = ms->cParams.minMatch;
    assert(ms->dictMatchState != NULL);
    switch(mls)
    {
    default: /* includes case 3 */
    case 4 :
        return ZSTD_compressBlock_fast_dictMatchState_4_0(ms, seqStore, rep, src, srcSize);
    case 5 :
        return ZSTD_compressBlock_fast_dictMatchState_5_0(ms, seqStore, rep, src, srcSize);
    case 6 :
        return ZSTD_compressBlock_fast_dictMatchState_6_0(ms, seqStore, rep, src, srcSize);
    case 7 :
        return ZSTD_compressBlock_fast_dictMatchState_7_0(ms, seqStore, rep, src, srcSize);
    }
}


static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_fast_extDict_generic(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize, U32 const mls, U32 const hasStep)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32 const hlog = cParams->hashLog;
    /* support stepSize of 0 */
    size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
    const BYTE* const base = ms->window.base;
    const BYTE* const dictBase = ms->window.dictBase;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* anchor = istart;
    const U32   endIndex = (U32)((size_t)(istart - base) + srcSize);
    const U32   lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
    const U32   dictStartIndex = lowLimit;
    const BYTE* const dictStart = dictBase + dictStartIndex;
    const U32   dictLimit = ms->window.dictLimit;
    const U32   prefixStartIndex = dictLimit < lowLimit ? lowLimit : dictLimit;
    const BYTE* const prefixStart = base + prefixStartIndex;
    const BYTE* const dictEnd = dictBase + prefixStartIndex;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - 8;
    U32 offset_1=rep[0], offset_2=rep[1];
    U32 offsetSaved1 = 0, offsetSaved2 = 0;

    const BYTE* ip0 = istart;
    const BYTE* ip1;
    const BYTE* ip2;
    const BYTE* ip3;
    U32 current0;


    size_t hash0; /* hash for ip0 */
    size_t hash1; /* hash for ip1 */
    U32 idx; /* match idx for ip0 */
    const BYTE* idxBase; /* base pointer for idx */

    U32 offcode;
    const BYTE* match0;
    size_t mLength;
    const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */

    size_t step;
    const BYTE* nextStep;
    const size_t kStepIncr = (1 << (kSearchStrength - 1));

    (void)hasStep; /* not currently specialized on whether it's accelerated */

    DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);

    /* switch to "regular" variant if extDict is invalidated due to maxDistance */
    if (prefixStartIndex == dictStartIndex)
        return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);

    {   U32 const curr = (U32)(ip0 - base);
        U32 const maxRep = curr - dictStartIndex;
        if (offset_2 >= maxRep) offsetSaved2 = offset_2, offset_2 = 0;
        if (offset_1 >= maxRep) offsetSaved1 = offset_1, offset_1 = 0;
    }

    /* start each op */
_start: /* Requires: ip0 */

    step = stepSize;
    nextStep = ip0 + kStepIncr;

    /* calculate positions, ip0 - anchor == 0, so we skip step calc */
    ip1 = ip0 + 1;
    ip2 = ip0 + step;
    ip3 = ip2 + 1;

    if (ip3 >= ilimit) {
        goto _cleanup;
    }

    hash0 = ZSTD_hashPtr(ip0, hlog, mls);
    hash1 = ZSTD_hashPtr(ip1, hlog, mls);

    idx = hashTable[hash0];
    idxBase = idx < prefixStartIndex ? dictBase : base;

    do {
        {   /* load repcode match for ip[2] */
            U32 const current2 = (U32)(ip2 - base);
            U32 const repIndex = current2 - offset_1;
            const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
            U32 rval;
            if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
                 & (offset_1 > 0) ) {
                rval = MEM_read32(repBase + repIndex);
            } else {
                rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
            }

            /* write back hash table entry */
            current0 = (U32)(ip0 - base);
            hashTable[hash0] = current0;

            /* check repcode at ip[2] */
            if (MEM_read32(ip2) == rval) {
                ip0 = ip2;
                match0 = repBase + repIndex;
                matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
                assert((match0 != prefixStart) & (match0 != dictStart));
                mLength = ip0[-1] == match0[-1];
                ip0 -= mLength;
                match0 -= mLength;
                offcode = REPCODE1_TO_OFFBASE;
                mLength += 4;
                goto _match;
        }   }

        {   /* load match for ip[0] */
            U32 const mval = idx >= dictStartIndex ?
                    MEM_read32(idxBase + idx) :
                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */

            /* check match at ip[0] */
            if (MEM_read32(ip0) == mval) {
                /* found a match! */
                goto _offset;
        }   }

        /* lookup ip[1] */
        idx = hashTable[hash1];
        idxBase = idx < prefixStartIndex ? dictBase : base;

        /* hash ip[2] */
        hash0 = hash1;
        hash1 = ZSTD_hashPtr(ip2, hlog, mls);

        /* advance to next positions */
        ip0 = ip1;
        ip1 = ip2;
        ip2 = ip3;

        /* write back hash table entry */
        current0 = (U32)(ip0 - base);
        hashTable[hash0] = current0;

        {   /* load match for ip[0] */
            U32 const mval = idx >= dictStartIndex ?
                    MEM_read32(idxBase + idx) :
                    MEM_read32(ip0) ^ 1; /* guaranteed not to match */

            /* check match at ip[0] */
            if (MEM_read32(ip0) == mval) {
                /* found a match! */
                goto _offset;
        }   }

        /* lookup ip[1] */
        idx = hashTable[hash1];
        idxBase = idx < prefixStartIndex ? dictBase : base;

        /* hash ip[2] */
        hash0 = hash1;
        hash1 = ZSTD_hashPtr(ip2, hlog, mls);

        /* advance to next positions */
        ip0 = ip1;
        ip1 = ip2;
        ip2 = ip0 + step;
        ip3 = ip1 + step;

        /* calculate step */
        if (ip2 >= nextStep) {
            step++;
            PREFETCH_L1(ip1 + 64);
            PREFETCH_L1(ip1 + 128);
            nextStep += kStepIncr;
        }
    } while (ip3 < ilimit);

_cleanup:
    /* Note that there are probably still a couple positions we could search.
     * However, it seems to be a meaningful performance hit to try to search
     * them. So let's not. */

    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;

    /* save reps for next block */
    rep[0] = offset_1 ? offset_1 : offsetSaved1;
    rep[1] = offset_2 ? offset_2 : offsetSaved2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);

_offset: /* Requires: ip0, idx, idxBase */

    /* Compute the offset code. */
    {   U32 const offset = current0 - idx;
        const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
        matchEnd = idx < prefixStartIndex ? dictEnd : iend;
        match0 = idxBase + idx;
        offset_2 = offset_1;
        offset_1 = offset;
        offcode = OFFSET_TO_OFFBASE(offset);
        mLength = 4;

        /* Count the backwards match length. */
        while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
            ip0--;
            match0--;
            mLength++;
    }   }

_match: /* Requires: ip0, match0, offcode, matchEnd */

    /* Count the forward length. */
    assert(matchEnd != 0);
    mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);

    ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);

    ip0 += mLength;
    anchor = ip0;

    /* write next hash table entry */
    if (ip1 < ip0) {
        hashTable[hash1] = (U32)(ip1 - base);
    }

    /* Fill table and check for immediate repcode. */
    if (ip0 <= ilimit) {
        /* Fill Table */
        assert(base+current0+2 > istart);  /* check base overflow */
        hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2;  /* here because current+2 could be > iend-8 */
        hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);

        while (ip0 <= ilimit) {
            U32 const repIndex2 = (U32)(ip0-base) - offset_2;
            const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
            if ( ((ZSTD_index_overlap_check(prefixStartIndex, repIndex2)) & (offset_2 > 0))
                 && (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
                const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
                size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
                { U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; }  /* swap offset_2 <=> offset_1 */
                ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
                hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
                ip0 += repLength2;
                anchor = ip0;
                continue;
            }
            break;
    }   }

    goto _start;
}

ZSTD_GEN_FAST_FN(extDict, 4, 0)
ZSTD_GEN_FAST_FN(extDict, 5, 0)
ZSTD_GEN_FAST_FN(extDict, 6, 0)
ZSTD_GEN_FAST_FN(extDict, 7, 0)

size_t ZSTD_compressBlock_fast_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    U32 const mls = ms->cParams.minMatch;
    assert(ms->dictMatchState == NULL);
    switch(mls)
    {
    default: /* includes case 3 */
    case 4 :
        return ZSTD_compressBlock_fast_extDict_4_0(ms, seqStore, rep, src, srcSize);
    case 5 :
        return ZSTD_compressBlock_fast_extDict_5_0(ms, seqStore, rep, src, srcSize);
    case 6 :
        return ZSTD_compressBlock_fast_extDict_6_0(ms, seqStore, rep, src, srcSize);
    case 7 :
        return ZSTD_compressBlock_fast_extDict_7_0(ms, seqStore, rep, src, srcSize);
    }
}
/**** ended inlining compress/zstd_fast.c ****/
/**** start inlining compress/zstd_lazy.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_lazy.h ****/
/**** skipping file: ../common/bits.h ****/

#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)

#define kLazySkippingStep 8


/*-*************************************
*  Binary Tree search
***************************************/

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_updateDUBT(ZSTD_MatchState_t* ms,
                const BYTE* ip, const BYTE* iend,
                U32 mls)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const hashTable = ms->hashTable;
    U32  const hashLog = cParams->hashLog;

    U32* const bt = ms->chainTable;
    U32  const btLog  = cParams->chainLog - 1;
    U32  const btMask = (1 << btLog) - 1;

    const BYTE* const base = ms->window.base;
    U32 const target = (U32)(ip - base);
    U32 idx = ms->nextToUpdate;

    if (idx != target)
        DEBUGLOG(7, "ZSTD_updateDUBT, from %u to %u (dictLimit:%u)",
                    idx, target, ms->window.dictLimit);
    assert(ip + 8 <= iend);   /* condition for ZSTD_hashPtr */
    (void)iend;

    assert(idx >= ms->window.dictLimit);   /* condition for valid base+idx */
    for ( ; idx < target ; idx++) {
        size_t const h  = ZSTD_hashPtr(base + idx, hashLog, mls);   /* assumption : ip + 8 <= iend */
        U32    const matchIndex = hashTable[h];

        U32*   const nextCandidatePtr = bt + 2*(idx&btMask);
        U32*   const sortMarkPtr  = nextCandidatePtr + 1;

        DEBUGLOG(8, "ZSTD_updateDUBT: insert %u", idx);
        hashTable[h] = idx;   /* Update Hash Table */
        *nextCandidatePtr = matchIndex;   /* update BT like a chain */
        *sortMarkPtr = ZSTD_DUBT_UNSORTED_MARK;
    }
    ms->nextToUpdate = target;
}


/** ZSTD_insertDUBT1() :
 *  sort one already inserted but unsorted position
 *  assumption : curr >= btlow == (curr - btmask)
 *  doesn't fail */
static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_insertDUBT1(const ZSTD_MatchState_t* ms,
                 U32 curr, const BYTE* inputEnd,
                 U32 nbCompares, U32 btLow,
                 const ZSTD_dictMode_e dictMode)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const bt = ms->chainTable;
    U32  const btLog  = cParams->chainLog - 1;
    U32  const btMask = (1 << btLog) - 1;
    size_t commonLengthSmaller=0, commonLengthLarger=0;
    const BYTE* const base = ms->window.base;
    const BYTE* const dictBase = ms->window.dictBase;
    const U32 dictLimit = ms->window.dictLimit;
    const BYTE* const ip = (curr>=dictLimit) ? base + curr : dictBase + curr;
    const BYTE* const iend = (curr>=dictLimit) ? inputEnd : dictBase + dictLimit;
    const BYTE* const dictEnd = dictBase + dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    const BYTE* match;
    U32* smallerPtr = bt + 2*(curr&btMask);
    U32* largerPtr  = smallerPtr + 1;
    U32 matchIndex = *smallerPtr;   /* this candidate is unsorted : next sorted candidate is reached through *smallerPtr, while *largerPtr contains previous unsorted candidate (which is already saved and can be overwritten) */
    U32 dummy32;   /* to be nullified at the end */
    U32 const windowValid = ms->window.lowLimit;
    U32 const maxDistance = 1U << cParams->windowLog;
    U32 const windowLow = (curr - windowValid > maxDistance) ? curr - maxDistance : windowValid;


    DEBUGLOG(8, "ZSTD_insertDUBT1(%u) (dictLimit=%u, lowLimit=%u)",
                curr, dictLimit, windowLow);
    assert(curr >= btLow);
    assert(ip < iend);   /* condition for ZSTD_count */

    for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
        U32* const nextPtr = bt + 2*(matchIndex & btMask);
        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
        assert(matchIndex < curr);
        /* note : all candidates are now supposed sorted,
         * but it's still possible to have nextPtr[1] == ZSTD_DUBT_UNSORTED_MARK
         * when a real index has the same value as ZSTD_DUBT_UNSORTED_MARK */

        if ( (dictMode != ZSTD_extDict)
          || (matchIndex+matchLength >= dictLimit)  /* both in current segment*/
          || (curr < dictLimit) /* both in extDict */) {
            const BYTE* const mBase = ( (dictMode != ZSTD_extDict)
                                     || (matchIndex+matchLength >= dictLimit)) ?
                                        base : dictBase;
            assert( (matchIndex+matchLength >= dictLimit)   /* might be wrong if extDict is incorrectly set to 0 */
                 || (curr < dictLimit) );
            match = mBase + matchIndex;
            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
        } else {
            match = dictBase + matchIndex;
            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
            if (matchIndex+matchLength >= dictLimit)
                match = base + matchIndex;   /* preparation for next read of match[matchLength] */
        }

        DEBUGLOG(8, "ZSTD_insertDUBT1: comparing %u with %u : found %u common bytes ",
                    curr, matchIndex, (U32)matchLength);

        if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
            break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
        }

        if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
            /* match is smaller than current */
            *smallerPtr = matchIndex;             /* update smaller idx */
            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
            DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is smaller : next => %u",
                        matchIndex, btLow, nextPtr[1]);
            smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
        } else {
            /* match is larger than current */
            *largerPtr = matchIndex;
            commonLengthLarger = matchLength;
            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
            DEBUGLOG(8, "ZSTD_insertDUBT1: %u (>btLow=%u) is larger => %u",
                        matchIndex, btLow, nextPtr[0]);
            largerPtr = nextPtr;
            matchIndex = nextPtr[0];
    }   }

    *smallerPtr = *largerPtr = 0;
}


static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_DUBT_findBetterDictMatch (
        const ZSTD_MatchState_t* ms,
        const BYTE* const ip, const BYTE* const iend,
        size_t* offsetPtr,
        size_t bestLength,
        U32 nbCompares,
        U32 const mls,
        const ZSTD_dictMode_e dictMode)
{
    const ZSTD_MatchState_t * const dms = ms->dictMatchState;
    const ZSTD_compressionParameters* const dmsCParams = &dms->cParams;
    const U32 * const dictHashTable = dms->hashTable;
    U32         const hashLog = dmsCParams->hashLog;
    size_t      const h  = ZSTD_hashPtr(ip, hashLog, mls);
    U32               dictMatchIndex = dictHashTable[h];

    const BYTE* const base = ms->window.base;
    const BYTE* const prefixStart = base + ms->window.dictLimit;
    U32         const curr = (U32)(ip-base);
    const BYTE* const dictBase = dms->window.base;
    const BYTE* const dictEnd = dms->window.nextSrc;
    U32         const dictHighLimit = (U32)(dms->window.nextSrc - dms->window.base);
    U32         const dictLowLimit = dms->window.lowLimit;
    U32         const dictIndexDelta = ms->window.lowLimit - dictHighLimit;

    U32*        const dictBt = dms->chainTable;
    U32         const btLog  = dmsCParams->chainLog - 1;
    U32         const btMask = (1 << btLog) - 1;
    U32         const btLow = (btMask >= dictHighLimit - dictLowLimit) ? dictLowLimit : dictHighLimit - btMask;

    size_t commonLengthSmaller=0, commonLengthLarger=0;

    (void)dictMode;
    assert(dictMode == ZSTD_dictMatchState);

    for (; nbCompares && (dictMatchIndex > dictLowLimit); --nbCompares) {
        U32* const nextPtr = dictBt + 2*(dictMatchIndex & btMask);
        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
        const BYTE* match = dictBase + dictMatchIndex;
        matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
        if (dictMatchIndex+matchLength >= dictHighLimit)
            match = base + dictMatchIndex + dictIndexDelta;   /* to prepare for next usage of match[matchLength] */

        if (matchLength > bestLength) {
            U32 matchIndex = dictMatchIndex + dictIndexDelta;
            if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr-matchIndex+1) - ZSTD_highbit32((U32)offsetPtr[0]+1)) ) {
                DEBUGLOG(9, "ZSTD_DUBT_findBetterDictMatch(%u) : found better match length %u -> %u and offsetCode %u -> %u (dictMatchIndex %u, matchIndex %u)",
                    curr, (U32)bestLength, (U32)matchLength, (U32)*offsetPtr, OFFSET_TO_OFFBASE(curr - matchIndex), dictMatchIndex, matchIndex);
                bestLength = matchLength, *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
            }
            if (ip+matchLength == iend) {   /* reached end of input : ip[matchLength] is not valid, no way to know if it's larger or smaller than match */
                break;   /* drop, to guarantee consistency (miss a little bit of compression) */
            }
        }

        if (match[matchLength] < ip[matchLength]) {
            if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
            dictMatchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
        } else {
            /* match is larger than current */
            if (dictMatchIndex <= btLow) { break; }   /* beyond tree size, stop the search */
            commonLengthLarger = matchLength;
            dictMatchIndex = nextPtr[0];
        }
    }

    if (bestLength >= MINMATCH) {
        U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offsetPtr); (void)mIndex;
        DEBUGLOG(8, "ZSTD_DUBT_findBetterDictMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
                    curr, (U32)bestLength, (U32)*offsetPtr, mIndex);
    }
    return bestLength;

}


static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_DUBT_findBestMatch(ZSTD_MatchState_t* ms,
                        const BYTE* const ip, const BYTE* const iend,
                        size_t* offBasePtr,
                        U32 const mls,
                        const ZSTD_dictMode_e dictMode)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32*   const hashTable = ms->hashTable;
    U32    const hashLog = cParams->hashLog;
    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
    U32          matchIndex  = hashTable[h];

    const BYTE* const base = ms->window.base;
    U32    const curr = (U32)(ip-base);
    U32    const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);

    U32*   const bt = ms->chainTable;
    U32    const btLog  = cParams->chainLog - 1;
    U32    const btMask = (1 << btLog) - 1;
    U32    const btLow = (btMask >= curr) ? 0 : curr - btMask;
    U32    const unsortLimit = MAX(btLow, windowLow);

    U32*         nextCandidate = bt + 2*(matchIndex&btMask);
    U32*         unsortedMark = bt + 2*(matchIndex&btMask) + 1;
    U32          nbCompares = 1U << cParams->searchLog;
    U32          nbCandidates = nbCompares;
    U32          previousCandidate = 0;

    DEBUGLOG(7, "ZSTD_DUBT_findBestMatch (%u) ", curr);
    assert(ip <= iend-8);   /* required for h calculation */
    assert(dictMode != ZSTD_dedicatedDictSearch);

    /* reach end of unsorted candidates list */
    while ( (matchIndex > unsortLimit)
         && (*unsortedMark == ZSTD_DUBT_UNSORTED_MARK)
         && (nbCandidates > 1) ) {
        DEBUGLOG(8, "ZSTD_DUBT_findBestMatch: candidate %u is unsorted",
                    matchIndex);
        *unsortedMark = previousCandidate;  /* the unsortedMark becomes a reversed chain, to move up back to original position */
        previousCandidate = matchIndex;
        matchIndex = *nextCandidate;
        nextCandidate = bt + 2*(matchIndex&btMask);
        unsortedMark = bt + 2*(matchIndex&btMask) + 1;
        nbCandidates --;
    }

    /* nullify last candidate if it's still unsorted
     * simplification, detrimental to compression ratio, beneficial for speed */
    if ( (matchIndex > unsortLimit)
      && (*unsortedMark==ZSTD_DUBT_UNSORTED_MARK) ) {
        DEBUGLOG(7, "ZSTD_DUBT_findBestMatch: nullify last unsorted candidate %u",
                    matchIndex);
        *nextCandidate = *unsortedMark = 0;
    }

    /* batch sort stacked candidates */
    matchIndex = previousCandidate;
    while (matchIndex) {  /* will end on matchIndex == 0 */
        U32* const nextCandidateIdxPtr = bt + 2*(matchIndex&btMask) + 1;
        U32 const nextCandidateIdx = *nextCandidateIdxPtr;
        ZSTD_insertDUBT1(ms, matchIndex, iend,
                         nbCandidates, unsortLimit, dictMode);
        matchIndex = nextCandidateIdx;
        nbCandidates++;
    }

    /* find longest match */
    {   size_t commonLengthSmaller = 0, commonLengthLarger = 0;
        const BYTE* const dictBase = ms->window.dictBase;
        const U32 dictLimit = ms->window.dictLimit;
        const BYTE* const dictEnd = dictBase + dictLimit;
        const BYTE* const prefixStart = base + dictLimit;
        U32* smallerPtr = bt + 2*(curr&btMask);
        U32* largerPtr  = bt + 2*(curr&btMask) + 1;
        U32 matchEndIdx = curr + 8 + 1;
        U32 dummy32;   /* to be nullified at the end */
        size_t bestLength = 0;

        matchIndex  = hashTable[h];
        hashTable[h] = curr;   /* Update Hash Table */

        for (; nbCompares && (matchIndex > windowLow); --nbCompares) {
            U32* const nextPtr = bt + 2*(matchIndex & btMask);
            size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
            const BYTE* match;

            if ((dictMode != ZSTD_extDict) || (matchIndex+matchLength >= dictLimit)) {
                match = base + matchIndex;
                matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
            } else {
                match = dictBase + matchIndex;
                matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
                if (matchIndex+matchLength >= dictLimit)
                    match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
            }

            if (matchLength > bestLength) {
                if (matchLength > matchEndIdx - matchIndex)
                    matchEndIdx = matchIndex + (U32)matchLength;
                if ( (4*(int)(matchLength-bestLength)) > (int)(ZSTD_highbit32(curr - matchIndex + 1) - ZSTD_highbit32((U32)*offBasePtr)) )
                    bestLength = matchLength, *offBasePtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
                    if (dictMode == ZSTD_dictMatchState) {
                        nbCompares = 0; /* in addition to avoiding checking any
                                         * further in this loop, make sure we
                                         * skip checking in the dictionary. */
                    }
                    break;   /* drop, to guarantee consistency (miss a little bit of compression) */
                }
            }

            if (match[matchLength] < ip[matchLength]) {
                /* match is smaller than current */
                *smallerPtr = matchIndex;             /* update smaller idx */
                commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
                if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
                smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
                matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
            } else {
                /* match is larger than current */
                *largerPtr = matchIndex;
                commonLengthLarger = matchLength;
                if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
                largerPtr = nextPtr;
                matchIndex = nextPtr[0];
        }   }

        *smallerPtr = *largerPtr = 0;

        assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
        if (dictMode == ZSTD_dictMatchState && nbCompares) {
            bestLength = ZSTD_DUBT_findBetterDictMatch(
                    ms, ip, iend,
                    offBasePtr, bestLength, nbCompares,
                    mls, dictMode);
        }

        assert(matchEndIdx > curr+8); /* ensure nextToUpdate is increased */
        ms->nextToUpdate = matchEndIdx - 8;   /* skip repetitive patterns */
        if (bestLength >= MINMATCH) {
            U32 const mIndex = curr - (U32)OFFBASE_TO_OFFSET(*offBasePtr); (void)mIndex;
            DEBUGLOG(8, "ZSTD_DUBT_findBestMatch(%u) : found match of length %u and offsetCode %u (pos %u)",
                        curr, (U32)bestLength, (U32)*offBasePtr, mIndex);
        }
        return bestLength;
    }
}


/** ZSTD_BtFindBestMatch() : Tree updater, providing best match */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_BtFindBestMatch( ZSTD_MatchState_t* ms,
                const BYTE* const ip, const BYTE* const iLimit,
                      size_t* offBasePtr,
                const U32 mls /* template */,
                const ZSTD_dictMode_e dictMode)
{
    DEBUGLOG(7, "ZSTD_BtFindBestMatch");
    if (ip < ms->window.base + ms->nextToUpdate) return 0;   /* skipped area */
    ZSTD_updateDUBT(ms, ip, iLimit, mls);
    return ZSTD_DUBT_findBestMatch(ms, ip, iLimit, offBasePtr, mls, dictMode);
}

/***********************************
* Dedicated dict search
***********************************/

void ZSTD_dedicatedDictSearch_lazy_loadDictionary(ZSTD_MatchState_t* ms, const BYTE* const ip)
{
    const BYTE* const base = ms->window.base;
    U32 const target = (U32)(ip - base);
    U32* const hashTable = ms->hashTable;
    U32* const chainTable = ms->chainTable;
    U32 const chainSize = 1 << ms->cParams.chainLog;
    U32 idx = ms->nextToUpdate;
    U32 const minChain = chainSize < target - idx ? target - chainSize : idx;
    U32 const bucketSize = 1 << ZSTD_LAZY_DDSS_BUCKET_LOG;
    U32 const cacheSize = bucketSize - 1;
    U32 const chainAttempts = (1 << ms->cParams.searchLog) - cacheSize;
    U32 const chainLimit = chainAttempts > 255 ? 255 : chainAttempts;

    /* We know the hashtable is oversized by a factor of `bucketSize`.
     * We are going to temporarily pretend `bucketSize == 1`, keeping only a
     * single entry. We will use the rest of the space to construct a temporary
     * chaintable.
     */
    U32 const hashLog = ms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
    U32* const tmpHashTable = hashTable;
    U32* const tmpChainTable = hashTable + ((size_t)1 << hashLog);
    U32 const tmpChainSize = (U32)((1 << ZSTD_LAZY_DDSS_BUCKET_LOG) - 1) << hashLog;
    U32 const tmpMinChain = tmpChainSize < target ? target - tmpChainSize : idx;
    U32 hashIdx;

    assert(ms->cParams.chainLog <= 24);
    assert(ms->cParams.hashLog > ms->cParams.chainLog);
    assert(idx != 0);
    assert(tmpMinChain <= minChain);

    /* fill conventional hash table and conventional chain table */
    for ( ; idx < target; idx++) {
        U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch);
        if (idx >= tmpMinChain) {
            tmpChainTable[idx - tmpMinChain] = hashTable[h];
        }
        tmpHashTable[h] = idx;
    }

    /* sort chains into ddss chain table */
    {
        U32 chainPos = 0;
        for (hashIdx = 0; hashIdx < (1U << hashLog); hashIdx++) {
            U32 count;
            U32 countBeyondMinChain = 0;
            U32 i = tmpHashTable[hashIdx];
            for (count = 0; i >= tmpMinChain && count < cacheSize; count++) {
                /* skip through the chain to the first position that won't be
                 * in the hash cache bucket */
                if (i < minChain) {
                    countBeyondMinChain++;
                }
                i = tmpChainTable[i - tmpMinChain];
            }
            if (count == cacheSize) {
                for (count = 0; count < chainLimit;) {
                    if (i < minChain) {
                        if (!i || ++countBeyondMinChain > cacheSize) {
                            /* only allow pulling `cacheSize` number of entries
                             * into the cache or chainTable beyond `minChain`,
                             * to replace the entries pulled out of the
                             * chainTable into the cache. This lets us reach
                             * back further without increasing the total number
                             * of entries in the chainTable, guaranteeing the
                             * DDSS chain table will fit into the space
                             * allocated for the regular one. */
                            break;
                        }
                    }
                    chainTable[chainPos++] = i;
                    count++;
                    if (i < tmpMinChain) {
                        break;
                    }
                    i = tmpChainTable[i - tmpMinChain];
                }
            } else {
                count = 0;
            }
            if (count) {
                tmpHashTable[hashIdx] = ((chainPos - count) << 8) + count;
            } else {
                tmpHashTable[hashIdx] = 0;
            }
        }
        assert(chainPos <= chainSize); /* I believe this is guaranteed... */
    }

    /* move chain pointers into the last entry of each hash bucket */
    for (hashIdx = (1 << hashLog); hashIdx; ) {
        U32 const bucketIdx = --hashIdx << ZSTD_LAZY_DDSS_BUCKET_LOG;
        U32 const chainPackedPointer = tmpHashTable[hashIdx];
        U32 i;
        for (i = 0; i < cacheSize; i++) {
            hashTable[bucketIdx + i] = 0;
        }
        hashTable[bucketIdx + bucketSize - 1] = chainPackedPointer;
    }

    /* fill the buckets of the hash table */
    for (idx = ms->nextToUpdate; idx < target; idx++) {
        U32 const h = (U32)ZSTD_hashPtr(base + idx, hashLog, ms->cParams.minMatch)
                   << ZSTD_LAZY_DDSS_BUCKET_LOG;
        U32 i;
        /* Shift hash cache down 1. */
        for (i = cacheSize - 1; i; i--)
            hashTable[h + i] = hashTable[h + i - 1];
        hashTable[h] = idx;
    }

    ms->nextToUpdate = target;
}

/* Returns the longest match length found in the dedicated dict search structure.
 * If none are longer than the argument ml, then ml will be returned.
 */
FORCE_INLINE_TEMPLATE
size_t ZSTD_dedicatedDictSearch_lazy_search(size_t* offsetPtr, size_t ml, U32 nbAttempts,
                                            const ZSTD_MatchState_t* const dms,
                                            const BYTE* const ip, const BYTE* const iLimit,
                                            const BYTE* const prefixStart, const U32 curr,
                                            const U32 dictLimit, const size_t ddsIdx) {
    const U32 ddsLowestIndex  = dms->window.dictLimit;
    const BYTE* const ddsBase = dms->window.base;
    const BYTE* const ddsEnd  = dms->window.nextSrc;
    const U32 ddsSize         = (U32)(ddsEnd - ddsBase);
    const U32 ddsIndexDelta   = dictLimit - ddsSize;
    const U32 bucketSize      = (1 << ZSTD_LAZY_DDSS_BUCKET_LOG);
    const U32 bucketLimit     = nbAttempts < bucketSize - 1 ? nbAttempts : bucketSize - 1;
    U32 ddsAttempt;
    U32 matchIndex;

    for (ddsAttempt = 0; ddsAttempt < bucketSize - 1; ddsAttempt++) {
        PREFETCH_L1(ddsBase + dms->hashTable[ddsIdx + ddsAttempt]);
    }

    {
        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
        U32 const chainIndex = chainPackedPointer >> 8;

        PREFETCH_L1(&dms->chainTable[chainIndex]);
    }

    for (ddsAttempt = 0; ddsAttempt < bucketLimit; ddsAttempt++) {
        size_t currentMl=0;
        const BYTE* match;
        matchIndex = dms->hashTable[ddsIdx + ddsAttempt];
        match = ddsBase + matchIndex;

        if (!matchIndex) {
            return ml;
        }

        /* guaranteed by table construction */
        (void)ddsLowestIndex;
        assert(matchIndex >= ddsLowestIndex);
        assert(match+4 <= ddsEnd);
        if (MEM_read32(match) == MEM_read32(ip)) {
            /* assumption : matchIndex <= dictLimit-4 (by table construction) */
            currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
        }

        /* save best solution */
        if (currentMl > ml) {
            ml = currentMl;
            *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
            if (ip+currentMl == iLimit) {
                /* best possible, avoids read overflow on next attempt */
                return ml;
            }
        }
    }

    {
        U32 const chainPackedPointer = dms->hashTable[ddsIdx + bucketSize - 1];
        U32 chainIndex = chainPackedPointer >> 8;
        U32 const chainLength = chainPackedPointer & 0xFF;
        U32 const chainAttempts = nbAttempts - ddsAttempt;
        U32 const chainLimit = chainAttempts > chainLength ? chainLength : chainAttempts;
        U32 chainAttempt;

        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++) {
            PREFETCH_L1(ddsBase + dms->chainTable[chainIndex + chainAttempt]);
        }

        for (chainAttempt = 0 ; chainAttempt < chainLimit; chainAttempt++, chainIndex++) {
            size_t currentMl=0;
            const BYTE* match;
            matchIndex = dms->chainTable[chainIndex];
            match = ddsBase + matchIndex;

            /* guaranteed by table construction */
            assert(matchIndex >= ddsLowestIndex);
            assert(match+4 <= ddsEnd);
            if (MEM_read32(match) == MEM_read32(ip)) {
                /* assumption : matchIndex <= dictLimit-4 (by table construction) */
                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, ddsEnd, prefixStart) + 4;
            }

            /* save best solution */
            if (currentMl > ml) {
                ml = currentMl;
                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + ddsIndexDelta));
                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
            }
        }
    }
    return ml;
}


/* *********************************
*  Hash Chain
***********************************/
#define NEXT_IN_CHAIN(d, mask)   chainTable[(d) & (mask)]

/* Update chains up to ip (excluded)
   Assumption : always within prefix (i.e. not within extDict) */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_insertAndFindFirstIndex_internal(
                        ZSTD_MatchState_t* ms,
                        const ZSTD_compressionParameters* const cParams,
                        const BYTE* ip, U32 const mls, U32 const lazySkipping)
{
    U32* const hashTable  = ms->hashTable;
    const U32 hashLog = cParams->hashLog;
    U32* const chainTable = ms->chainTable;
    const U32 chainMask = (1 << cParams->chainLog) - 1;
    const BYTE* const base = ms->window.base;
    const U32 target = (U32)(ip - base);
    U32 idx = ms->nextToUpdate;

    while(idx < target) { /* catch up */
        size_t const h = ZSTD_hashPtr(base+idx, hashLog, mls);
        NEXT_IN_CHAIN(idx, chainMask) = hashTable[h];
        hashTable[h] = idx;
        idx++;
        /* Stop inserting every position when in the lazy skipping mode. */
        if (lazySkipping)
            break;
    }

    ms->nextToUpdate = target;
    return hashTable[ZSTD_hashPtr(ip, hashLog, mls)];
}

U32 ZSTD_insertAndFindFirstIndex(ZSTD_MatchState_t* ms, const BYTE* ip) {
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    return ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, ms->cParams.minMatch, /* lazySkipping*/ 0);
}

/* inlining is important to hardwire a hot branch (template emulation) */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_HcFindBestMatch(
                        ZSTD_MatchState_t* ms,
                        const BYTE* const ip, const BYTE* const iLimit,
                        size_t* offsetPtr,
                        const U32 mls, const ZSTD_dictMode_e dictMode)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32* const chainTable = ms->chainTable;
    const U32 chainSize = (1 << cParams->chainLog);
    const U32 chainMask = chainSize-1;
    const BYTE* const base = ms->window.base;
    const BYTE* const dictBase = ms->window.dictBase;
    const U32 dictLimit = ms->window.dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    const BYTE* const dictEnd = dictBase + dictLimit;
    const U32 curr = (U32)(ip-base);
    const U32 maxDistance = 1U << cParams->windowLog;
    const U32 lowestValid = ms->window.lowLimit;
    const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
    const U32 isDictionary = (ms->loadedDictEnd != 0);
    const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
    const U32 minChain = curr > chainSize ? curr - chainSize : 0;
    U32 nbAttempts = 1U << cParams->searchLog;
    size_t ml=4-1;

    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
    const U32 ddsHashLog = dictMode == ZSTD_dedicatedDictSearch
                         ? dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG : 0;
    const size_t ddsIdx = dictMode == ZSTD_dedicatedDictSearch
                        ? ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG : 0;

    U32 matchIndex;

    if (dictMode == ZSTD_dedicatedDictSearch) {
        const U32* entry = &dms->hashTable[ddsIdx];
        PREFETCH_L1(entry);
    }

    /* HC4 match finder */
    matchIndex = ZSTD_insertAndFindFirstIndex_internal(ms, cParams, ip, mls, ms->lazySkipping);

    for ( ; (matchIndex>=lowLimit) & (nbAttempts>0) ; nbAttempts--) {
        size_t currentMl=0;
        if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
            const BYTE* const match = base + matchIndex;
            assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
            /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
            if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                currentMl = ZSTD_count(ip, match, iLimit);
        } else {
            const BYTE* const match = dictBase + matchIndex;
            assert(match+4 <= dictEnd);
            if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
        }

        /* save best solution */
        if (currentMl > ml) {
            ml = currentMl;
            *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
            if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
        }

        if (matchIndex <= minChain) break;
        matchIndex = NEXT_IN_CHAIN(matchIndex, chainMask);
    }

    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
    if (dictMode == ZSTD_dedicatedDictSearch) {
        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts, dms,
                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
    } else if (dictMode == ZSTD_dictMatchState) {
        const U32* const dmsChainTable = dms->chainTable;
        const U32 dmsChainSize         = (1 << dms->cParams.chainLog);
        const U32 dmsChainMask         = dmsChainSize - 1;
        const U32 dmsLowestIndex       = dms->window.dictLimit;
        const BYTE* const dmsBase      = dms->window.base;
        const BYTE* const dmsEnd       = dms->window.nextSrc;
        const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
        const U32 dmsIndexDelta        = dictLimit - dmsSize;
        const U32 dmsMinChain = dmsSize > dmsChainSize ? dmsSize - dmsChainSize : 0;

        matchIndex = dms->hashTable[ZSTD_hashPtr(ip, dms->cParams.hashLog, mls)];

        for ( ; (matchIndex>=dmsLowestIndex) & (nbAttempts>0) ; nbAttempts--) {
            size_t currentMl=0;
            const BYTE* const match = dmsBase + matchIndex;
            assert(match+4 <= dmsEnd);
            if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
                currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;

            /* save best solution */
            if (currentMl > ml) {
                ml = currentMl;
                assert(curr > matchIndex + dmsIndexDelta);
                *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
            }

            if (matchIndex <= dmsMinChain) break;

            matchIndex = dmsChainTable[matchIndex & dmsChainMask];
        }
    }

    return ml;
}

/* *********************************
* (SIMD) Row-based matchfinder
***********************************/
/* Constants for row-based hash */
#define ZSTD_ROW_HASH_TAG_MASK ((1u << ZSTD_ROW_HASH_TAG_BITS) - 1)
#define ZSTD_ROW_HASH_MAX_ENTRIES 64    /* absolute maximum number of entries per row, for all configurations */

#define ZSTD_ROW_HASH_CACHE_MASK (ZSTD_ROW_HASH_CACHE_SIZE - 1)

typedef U64 ZSTD_VecMask;   /* Clarifies when we are interacting with a U64 representing a mask of matches */

/* ZSTD_VecMask_next():
 * Starting from the LSB, returns the idx of the next non-zero bit.
 * Basically counting the nb of trailing zeroes.
 */
MEM_STATIC U32 ZSTD_VecMask_next(ZSTD_VecMask val) {
    return ZSTD_countTrailingZeros64(val);
}

/* ZSTD_row_nextIndex():
 * Returns the next index to insert at within a tagTable row, and updates the "head"
 * value to reflect the update. Essentially cycles backwards from [1, {entries per row})
 */
FORCE_INLINE_TEMPLATE U32 ZSTD_row_nextIndex(BYTE* const tagRow, U32 const rowMask) {
    U32 next = (*tagRow-1) & rowMask;
    next += (next == 0) ? rowMask : 0; /* skip first position */
    *tagRow = (BYTE)next;
    return next;
}

/* ZSTD_isAligned():
 * Checks that a pointer is aligned to "align" bytes which must be a power of 2.
 */
MEM_STATIC int ZSTD_isAligned(void const* ptr, size_t align) {
    assert((align & (align - 1)) == 0);
    return (((size_t)ptr) & (align - 1)) == 0;
}

/* ZSTD_row_prefetch():
 * Performs prefetching for the hashTable and tagTable at a given row.
 */
FORCE_INLINE_TEMPLATE void ZSTD_row_prefetch(U32 const* hashTable, BYTE const* tagTable, U32 const relRow, U32 const rowLog) {
    PREFETCH_L1(hashTable + relRow);
    if (rowLog >= 5) {
        PREFETCH_L1(hashTable + relRow + 16);
        /* Note: prefetching more of the hash table does not appear to be beneficial for 128-entry rows */
    }
    PREFETCH_L1(tagTable + relRow);
    if (rowLog == 6) {
        PREFETCH_L1(tagTable + relRow + 32);
    }
    assert(rowLog == 4 || rowLog == 5 || rowLog == 6);
    assert(ZSTD_isAligned(hashTable + relRow, 64));                 /* prefetched hash row always 64-byte aligned */
    assert(ZSTD_isAligned(tagTable + relRow, (size_t)1 << rowLog)); /* prefetched tagRow sits on correct multiple of bytes (32,64,128) */
}

/* ZSTD_row_fillHashCache():
 * Fill up the hash cache starting at idx, prefetching up to ZSTD_ROW_HASH_CACHE_SIZE entries,
 * but not beyond iLimit.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_row_fillHashCache(ZSTD_MatchState_t* ms, const BYTE* base,
                                   U32 const rowLog, U32 const mls,
                                   U32 idx, const BYTE* const iLimit)
{
    U32 const* const hashTable = ms->hashTable;
    BYTE const* const tagTable = ms->tagTable;
    U32 const hashLog = ms->rowHashLog;
    U32 const maxElemsToPrefetch = (base + idx) > iLimit ? 0 : (U32)(iLimit - (base + idx) + 1);
    U32 const lim = idx + MIN(ZSTD_ROW_HASH_CACHE_SIZE, maxElemsToPrefetch);

    for (; idx < lim; ++idx) {
        U32 const hash = (U32)ZSTD_hashPtrSalted(base + idx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
        U32 const row = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
        ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
        ms->hashCache[idx & ZSTD_ROW_HASH_CACHE_MASK] = hash;
    }

    DEBUGLOG(6, "ZSTD_row_fillHashCache(): [%u %u %u %u %u %u %u %u]", ms->hashCache[0], ms->hashCache[1],
                                                     ms->hashCache[2], ms->hashCache[3], ms->hashCache[4],
                                                     ms->hashCache[5], ms->hashCache[6], ms->hashCache[7]);
}

/* ZSTD_row_nextCachedHash():
 * Returns the hash of base + idx, and replaces the hash in the hash cache with the byte at
 * base + idx + ZSTD_ROW_HASH_CACHE_SIZE. Also prefetches the appropriate rows from hashTable and tagTable.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_row_nextCachedHash(U32* cache, U32 const* hashTable,
                                                  BYTE const* tagTable, BYTE const* base,
                                                  U32 idx, U32 const hashLog,
                                                  U32 const rowLog, U32 const mls,
                                                  U64 const hashSalt)
{
    U32 const newHash = (U32)ZSTD_hashPtrSalted(base+idx+ZSTD_ROW_HASH_CACHE_SIZE, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
    U32 const row = (newHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
    ZSTD_row_prefetch(hashTable, tagTable, row, rowLog);
    {   U32 const hash = cache[idx & ZSTD_ROW_HASH_CACHE_MASK];
        cache[idx & ZSTD_ROW_HASH_CACHE_MASK] = newHash;
        return hash;
    }
}

/* ZSTD_row_update_internalImpl():
 * Updates the hash table with positions starting from updateStartIdx until updateEndIdx.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_row_update_internalImpl(ZSTD_MatchState_t* ms,
                                  U32 updateStartIdx, U32 const updateEndIdx,
                                  U32 const mls, U32 const rowLog,
                                  U32 const rowMask, U32 const useCache)
{
    U32* const hashTable = ms->hashTable;
    BYTE* const tagTable = ms->tagTable;
    U32 const hashLog = ms->rowHashLog;
    const BYTE* const base = ms->window.base;

    DEBUGLOG(6, "ZSTD_row_update_internalImpl(): updateStartIdx=%u, updateEndIdx=%u", updateStartIdx, updateEndIdx);
    for (; updateStartIdx < updateEndIdx; ++updateStartIdx) {
        U32 const hash = useCache ? ZSTD_row_nextCachedHash(ms->hashCache, hashTable, tagTable, base, updateStartIdx, hashLog, rowLog, mls, ms->hashSalt)
                                  : (U32)ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt);
        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
        U32* const row = hashTable + relRow;
        BYTE* tagRow = tagTable + relRow;
        U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);

        assert(hash == ZSTD_hashPtrSalted(base + updateStartIdx, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, ms->hashSalt));
        tagRow[pos] = hash & ZSTD_ROW_HASH_TAG_MASK;
        row[pos] = updateStartIdx;
    }
}

/* ZSTD_row_update_internal():
 * Inserts the byte at ip into the appropriate position in the hash table, and updates ms->nextToUpdate.
 * Skips sections of long matches as is necessary.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_row_update_internal(ZSTD_MatchState_t* ms, const BYTE* ip,
                              U32 const mls, U32 const rowLog,
                              U32 const rowMask, U32 const useCache)
{
    U32 idx = ms->nextToUpdate;
    const BYTE* const base = ms->window.base;
    const U32 target = (U32)(ip - base);
    const U32 kSkipThreshold = 384;
    const U32 kMaxMatchStartPositionsToUpdate = 96;
    const U32 kMaxMatchEndPositionsToUpdate = 32;

    if (useCache) {
        /* Only skip positions when using hash cache, i.e.
         * if we are loading a dict, don't skip anything.
         * If we decide to skip, then we only update a set number
         * of positions at the beginning and end of the match.
         */
        if (UNLIKELY(target - idx > kSkipThreshold)) {
            U32 const bound = idx + kMaxMatchStartPositionsToUpdate;
            ZSTD_row_update_internalImpl(ms, idx, bound, mls, rowLog, rowMask, useCache);
            idx = target - kMaxMatchEndPositionsToUpdate;
            ZSTD_row_fillHashCache(ms, base, rowLog, mls, idx, ip+1);
        }
    }
    assert(target >= idx);
    ZSTD_row_update_internalImpl(ms, idx, target, mls, rowLog, rowMask, useCache);
    ms->nextToUpdate = target;
}

/* ZSTD_row_update():
 * External wrapper for ZSTD_row_update_internal(). Used for filling the hashtable during dictionary
 * processing.
 */
void ZSTD_row_update(ZSTD_MatchState_t* const ms, const BYTE* ip) {
    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);
    const U32 rowMask = (1u << rowLog) - 1;
    const U32 mls = MIN(ms->cParams.minMatch, 6 /* mls caps out at 6 */);

    DEBUGLOG(5, "ZSTD_row_update(), rowLog=%u", rowLog);
    ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 0 /* don't use cache */);
}

/* Returns the mask width of bits group of which will be set to 1. Given not all
 * architectures have easy movemask instruction, this helps to iterate over
 * groups of bits easier and faster.
 */
FORCE_INLINE_TEMPLATE U32
ZSTD_row_matchMaskGroupWidth(const U32 rowEntries)
{
    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
    (void)rowEntries;
#if defined(ZSTD_ARCH_ARM_NEON)
    /* NEON path only works for little endian */
    if (!MEM_isLittleEndian()) {
        return 1;
    }
    if (rowEntries == 16) {
        return 4;
    }
    if (rowEntries == 32) {
        return 2;
    }
    if (rowEntries == 64) {
        return 1;
    }
#endif
    return 1;
}

#if defined(ZSTD_ARCH_X86_SSE2)
FORCE_INLINE_TEMPLATE ZSTD_VecMask
ZSTD_row_getSSEMask(int nbChunks, const BYTE* const src, const BYTE tag, const U32 head)
{
    const __m128i comparisonMask = _mm_set1_epi8((char)tag);
    int matches[4] = {0};
    int i;
    assert(nbChunks == 1 || nbChunks == 2 || nbChunks == 4);
    for (i=0; i<nbChunks; i++) {
        const __m128i chunk = _mm_loadu_si128((const __m128i*)(const void*)(src + 16*i));
        const __m128i equalMask = _mm_cmpeq_epi8(chunk, comparisonMask);
        matches[i] = _mm_movemask_epi8(equalMask);
    }
    if (nbChunks == 1) return ZSTD_rotateRight_U16((U16)matches[0], head);
    if (nbChunks == 2) return ZSTD_rotateRight_U32((U32)matches[1] << 16 | (U32)matches[0], head);
    assert(nbChunks == 4);
    return ZSTD_rotateRight_U64((U64)matches[3] << 48 | (U64)matches[2] << 32 | (U64)matches[1] << 16 | (U64)matches[0], head);
}
#endif

#if defined(ZSTD_ARCH_ARM_NEON)
FORCE_INLINE_TEMPLATE ZSTD_VecMask
ZSTD_row_getNEONMask(const U32 rowEntries, const BYTE* const src, const BYTE tag, const U32 headGrouped)
{
    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
    if (rowEntries == 16) {
        /* vshrn_n_u16 shifts by 4 every u16 and narrows to 8 lower bits.
         * After that groups of 4 bits represent the equalMask. We lower
         * all bits except the highest in these groups by doing AND with
         * 0x88 = 0b10001000.
         */
        const uint8x16_t chunk = vld1q_u8(src);
        const uint16x8_t equalMask = vreinterpretq_u16_u8(vceqq_u8(chunk, vdupq_n_u8(tag)));
        const uint8x8_t res = vshrn_n_u16(equalMask, 4);
        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0);
        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x8888888888888888ull;
    } else if (rowEntries == 32) {
        /* Same idea as with rowEntries == 16 but doing AND with
         * 0x55 = 0b01010101.
         */
        const uint16x8x2_t chunk = vld2q_u16((const uint16_t*)(const void*)src);
        const uint8x16_t chunk0 = vreinterpretq_u8_u16(chunk.val[0]);
        const uint8x16_t chunk1 = vreinterpretq_u8_u16(chunk.val[1]);
        const uint8x16_t dup = vdupq_n_u8(tag);
        const uint8x8_t t0 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk0, dup)), 6);
        const uint8x8_t t1 = vshrn_n_u16(vreinterpretq_u16_u8(vceqq_u8(chunk1, dup)), 6);
        const uint8x8_t res = vsli_n_u8(t0, t1, 4);
        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(res), 0) ;
        return ZSTD_rotateRight_U64(matches, headGrouped) & 0x5555555555555555ull;
    } else { /* rowEntries == 64 */
        const uint8x16x4_t chunk = vld4q_u8(src);
        const uint8x16_t dup = vdupq_n_u8(tag);
        const uint8x16_t cmp0 = vceqq_u8(chunk.val[0], dup);
        const uint8x16_t cmp1 = vceqq_u8(chunk.val[1], dup);
        const uint8x16_t cmp2 = vceqq_u8(chunk.val[2], dup);
        const uint8x16_t cmp3 = vceqq_u8(chunk.val[3], dup);

        const uint8x16_t t0 = vsriq_n_u8(cmp1, cmp0, 1);
        const uint8x16_t t1 = vsriq_n_u8(cmp3, cmp2, 1);
        const uint8x16_t t2 = vsriq_n_u8(t1, t0, 2);
        const uint8x16_t t3 = vsriq_n_u8(t2, t2, 4);
        const uint8x8_t t4 = vshrn_n_u16(vreinterpretq_u16_u8(t3), 4);
        const U64 matches = vget_lane_u64(vreinterpret_u64_u8(t4), 0);
        return ZSTD_rotateRight_U64(matches, headGrouped);
    }
}
#endif

/* Returns a ZSTD_VecMask (U64) that has the nth group (determined by
 * ZSTD_row_matchMaskGroupWidth) of bits set to 1 if the newly-computed "tag"
 * matches the hash at the nth position in a row of the tagTable.
 * Each row is a circular buffer beginning at the value of "headGrouped". So we
 * must rotate the "matches" bitfield to match up with the actual layout of the
 * entries within the hashTable */
FORCE_INLINE_TEMPLATE ZSTD_VecMask
ZSTD_row_getMatchMask(const BYTE* const tagRow, const BYTE tag, const U32 headGrouped, const U32 rowEntries)
{
    const BYTE* const src = tagRow;
    assert((rowEntries == 16) || (rowEntries == 32) || rowEntries == 64);
    assert(rowEntries <= ZSTD_ROW_HASH_MAX_ENTRIES);
    assert(ZSTD_row_matchMaskGroupWidth(rowEntries) * rowEntries <= sizeof(ZSTD_VecMask) * 8);

#if defined(ZSTD_ARCH_X86_SSE2)

    return ZSTD_row_getSSEMask(rowEntries / 16, src, tag, headGrouped);

#else /* SW or NEON-LE */

# if defined(ZSTD_ARCH_ARM_NEON)
  /* This NEON path only works for little endian - otherwise use SWAR below */
    if (MEM_isLittleEndian()) {
        return ZSTD_row_getNEONMask(rowEntries, src, tag, headGrouped);
    }
# endif /* ZSTD_ARCH_ARM_NEON */
    /* SWAR */
    {   const int chunkSize = sizeof(size_t);
        const size_t shiftAmount = ((chunkSize * 8) - chunkSize);
        const size_t xFF = ~((size_t)0);
        const size_t x01 = xFF / 0xFF;
        const size_t x80 = x01 << 7;
        const size_t splatChar = tag * x01;
        ZSTD_VecMask matches = 0;
        int i = rowEntries - chunkSize;
        assert((sizeof(size_t) == 4) || (sizeof(size_t) == 8));
        if (MEM_isLittleEndian()) { /* runtime check so have two loops */
            const size_t extractMagic = (xFF / 0x7F) >> chunkSize;
            do {
                size_t chunk = MEM_readST(&src[i]);
                chunk ^= splatChar;
                chunk = (((chunk | x80) - x01) | chunk) & x80;
                matches <<= chunkSize;
                matches |= (chunk * extractMagic) >> shiftAmount;
                i -= chunkSize;
            } while (i >= 0);
        } else { /* big endian: reverse bits during extraction */
            const size_t msb = xFF ^ (xFF >> 1);
            const size_t extractMagic = (msb / 0x1FF) | msb;
            do {
                size_t chunk = MEM_readST(&src[i]);
                chunk ^= splatChar;
                chunk = (((chunk | x80) - x01) | chunk) & x80;
                matches <<= chunkSize;
                matches |= ((chunk >> 7) * extractMagic) >> shiftAmount;
                i -= chunkSize;
            } while (i >= 0);
        }
        matches = ~matches;
        if (rowEntries == 16) {
            return ZSTD_rotateRight_U16((U16)matches, headGrouped);
        } else if (rowEntries == 32) {
            return ZSTD_rotateRight_U32((U32)matches, headGrouped);
        } else {
            return ZSTD_rotateRight_U64((U64)matches, headGrouped);
        }
    }
#endif
}

/* The high-level approach of the SIMD row based match finder is as follows:
 * - Figure out where to insert the new entry:
 *      - Generate a hash for current input position and split it into a one byte of tag and `rowHashLog` bits of index.
 *           - The hash is salted by a value that changes on every context reset, so when the same table is used
 *             we will avoid collisions that would otherwise slow us down by introducing phantom matches.
 *      - The hashTable is effectively split into groups or "rows" of 15 or 31 entries of U32, and the index determines
 *        which row to insert into.
 *      - Determine the correct position within the row to insert the entry into. Each row of 15 or 31 can
 *        be considered as a circular buffer with a "head" index that resides in the tagTable (overall 16 or 32 bytes
 *        per row).
 * - Use SIMD to efficiently compare the tags in the tagTable to the 1-byte tag calculated for the position and
 *   generate a bitfield that we can cycle through to check the collisions in the hash table.
 * - Pick the longest match.
 * - Insert the tag into the equivalent row and position in the tagTable.
 */
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_RowFindBestMatch(
                        ZSTD_MatchState_t* ms,
                        const BYTE* const ip, const BYTE* const iLimit,
                        size_t* offsetPtr,
                        const U32 mls, const ZSTD_dictMode_e dictMode,
                        const U32 rowLog)
{
    U32* const hashTable = ms->hashTable;
    BYTE* const tagTable = ms->tagTable;
    U32* const hashCache = ms->hashCache;
    const U32 hashLog = ms->rowHashLog;
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    const BYTE* const base = ms->window.base;
    const BYTE* const dictBase = ms->window.dictBase;
    const U32 dictLimit = ms->window.dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    const BYTE* const dictEnd = dictBase + dictLimit;
    const U32 curr = (U32)(ip-base);
    const U32 maxDistance = 1U << cParams->windowLog;
    const U32 lowestValid = ms->window.lowLimit;
    const U32 withinMaxDistance = (curr - lowestValid > maxDistance) ? curr - maxDistance : lowestValid;
    const U32 isDictionary = (ms->loadedDictEnd != 0);
    const U32 lowLimit = isDictionary ? lowestValid : withinMaxDistance;
    const U32 rowEntries = (1U << rowLog);
    const U32 rowMask = rowEntries - 1;
    const U32 cappedSearchLog = MIN(cParams->searchLog, rowLog); /* nb of searches is capped at nb entries per row */
    const U32 groupWidth = ZSTD_row_matchMaskGroupWidth(rowEntries);
    const U64 hashSalt = ms->hashSalt;
    U32 nbAttempts = 1U << cappedSearchLog;
    size_t ml=4-1;
    U32 hash;

    /* DMS/DDS variables that may be referenced laster */
    const ZSTD_MatchState_t* const dms = ms->dictMatchState;

    /* Initialize the following variables to satisfy static analyzer */
    size_t ddsIdx = 0;
    U32 ddsExtraAttempts = 0; /* cctx hash tables are limited in searches, but allow extra searches into DDS */
    U32 dmsTag = 0;
    U32* dmsRow = NULL;
    BYTE* dmsTagRow = NULL;

    if (dictMode == ZSTD_dedicatedDictSearch) {
        const U32 ddsHashLog = dms->cParams.hashLog - ZSTD_LAZY_DDSS_BUCKET_LOG;
        {   /* Prefetch DDS hashtable entry */
            ddsIdx = ZSTD_hashPtr(ip, ddsHashLog, mls) << ZSTD_LAZY_DDSS_BUCKET_LOG;
            PREFETCH_L1(&dms->hashTable[ddsIdx]);
        }
        ddsExtraAttempts = cParams->searchLog > rowLog ? 1U << (cParams->searchLog - rowLog) : 0;
    }

    if (dictMode == ZSTD_dictMatchState) {
        /* Prefetch DMS rows */
        U32* const dmsHashTable = dms->hashTable;
        BYTE* const dmsTagTable = dms->tagTable;
        U32 const dmsHash = (U32)ZSTD_hashPtr(ip, dms->rowHashLog + ZSTD_ROW_HASH_TAG_BITS, mls);
        U32 const dmsRelRow = (dmsHash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
        dmsTag = dmsHash & ZSTD_ROW_HASH_TAG_MASK;
        dmsTagRow = (BYTE*)(dmsTagTable + dmsRelRow);
        dmsRow = dmsHashTable + dmsRelRow;
        ZSTD_row_prefetch(dmsHashTable, dmsTagTable, dmsRelRow, rowLog);
    }

    /* Update the hashTable and tagTable up to (but not including) ip */
    if (!ms->lazySkipping) {
        ZSTD_row_update_internal(ms, ip, mls, rowLog, rowMask, 1 /* useCache */);
        hash = ZSTD_row_nextCachedHash(hashCache, hashTable, tagTable, base, curr, hashLog, rowLog, mls, hashSalt);
    } else {
        /* Stop inserting every position when in the lazy skipping mode.
         * The hash cache is also not kept up to date in this mode.
         */
        hash = (U32)ZSTD_hashPtrSalted(ip, hashLog + ZSTD_ROW_HASH_TAG_BITS, mls, hashSalt);
        ms->nextToUpdate = curr;
    }
    ms->hashSaltEntropy += hash; /* collect salt entropy */

    {   /* Get the hash for ip, compute the appropriate row */
        U32 const relRow = (hash >> ZSTD_ROW_HASH_TAG_BITS) << rowLog;
        U32 const tag = hash & ZSTD_ROW_HASH_TAG_MASK;
        U32* const row = hashTable + relRow;
        BYTE* tagRow = (BYTE*)(tagTable + relRow);
        U32 const headGrouped = (*tagRow & rowMask) * groupWidth;
        U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
        size_t numMatches = 0;
        size_t currMatch = 0;
        ZSTD_VecMask matches = ZSTD_row_getMatchMask(tagRow, (BYTE)tag, headGrouped, rowEntries);

        /* Cycle through the matches and prefetch */
        for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
            U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
            U32 const matchIndex = row[matchPos];
            if(matchPos == 0) continue;
            assert(numMatches < rowEntries);
            if (matchIndex < lowLimit)
                break;
            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
                PREFETCH_L1(base + matchIndex);
            } else {
                PREFETCH_L1(dictBase + matchIndex);
            }
            matchBuffer[numMatches++] = matchIndex;
            --nbAttempts;
        }

        /* Speed opt: insert current byte into hashtable too. This allows us to avoid one iteration of the loop
           in ZSTD_row_update_internal() at the next search. */
        {
            U32 const pos = ZSTD_row_nextIndex(tagRow, rowMask);
            tagRow[pos] = (BYTE)tag;
            row[pos] = ms->nextToUpdate++;
        }

        /* Return the longest match */
        for (; currMatch < numMatches; ++currMatch) {
            U32 const matchIndex = matchBuffer[currMatch];
            size_t currentMl=0;
            assert(matchIndex < curr);
            assert(matchIndex >= lowLimit);

            if ((dictMode != ZSTD_extDict) || matchIndex >= dictLimit) {
                const BYTE* const match = base + matchIndex;
                assert(matchIndex >= dictLimit);   /* ensures this is true if dictMode != ZSTD_extDict */
                /* read 4B starting from (match + ml + 1 - sizeof(U32)) */
                if (MEM_read32(match + ml - 3) == MEM_read32(ip + ml - 3))   /* potentially better */
                    currentMl = ZSTD_count(ip, match, iLimit);
            } else {
                const BYTE* const match = dictBase + matchIndex;
                assert(match+4 <= dictEnd);
                if (MEM_read32(match) == MEM_read32(ip))   /* assumption : matchIndex <= dictLimit-4 (by table construction) */
                    currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dictEnd, prefixStart) + 4;
            }

            /* Save best solution */
            if (currentMl > ml) {
                ml = currentMl;
                *offsetPtr = OFFSET_TO_OFFBASE(curr - matchIndex);
                if (ip+currentMl == iLimit) break; /* best possible, avoids read overflow on next attempt */
            }
        }
    }

    assert(nbAttempts <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
    if (dictMode == ZSTD_dedicatedDictSearch) {
        ml = ZSTD_dedicatedDictSearch_lazy_search(offsetPtr, ml, nbAttempts + ddsExtraAttempts, dms,
                                                  ip, iLimit, prefixStart, curr, dictLimit, ddsIdx);
    } else if (dictMode == ZSTD_dictMatchState) {
        /* TODO: Measure and potentially add prefetching to DMS */
        const U32 dmsLowestIndex       = dms->window.dictLimit;
        const BYTE* const dmsBase      = dms->window.base;
        const BYTE* const dmsEnd       = dms->window.nextSrc;
        const U32 dmsSize              = (U32)(dmsEnd - dmsBase);
        const U32 dmsIndexDelta        = dictLimit - dmsSize;

        {   U32 const headGrouped = (*dmsTagRow & rowMask) * groupWidth;
            U32 matchBuffer[ZSTD_ROW_HASH_MAX_ENTRIES];
            size_t numMatches = 0;
            size_t currMatch = 0;
            ZSTD_VecMask matches = ZSTD_row_getMatchMask(dmsTagRow, (BYTE)dmsTag, headGrouped, rowEntries);

            for (; (matches > 0) && (nbAttempts > 0); matches &= (matches - 1)) {
                U32 const matchPos = ((headGrouped + ZSTD_VecMask_next(matches)) / groupWidth) & rowMask;
                U32 const matchIndex = dmsRow[matchPos];
                if(matchPos == 0) continue;
                if (matchIndex < dmsLowestIndex)
                    break;
                PREFETCH_L1(dmsBase + matchIndex);
                matchBuffer[numMatches++] = matchIndex;
                --nbAttempts;
            }

            /* Return the longest match */
            for (; currMatch < numMatches; ++currMatch) {
                U32 const matchIndex = matchBuffer[currMatch];
                size_t currentMl=0;
                assert(matchIndex >= dmsLowestIndex);
                assert(matchIndex < curr);

                {   const BYTE* const match = dmsBase + matchIndex;
                    assert(match+4 <= dmsEnd);
                    if (MEM_read32(match) == MEM_read32(ip))
                        currentMl = ZSTD_count_2segments(ip+4, match+4, iLimit, dmsEnd, prefixStart) + 4;
                }

                if (currentMl > ml) {
                    ml = currentMl;
                    assert(curr > matchIndex + dmsIndexDelta);
                    *offsetPtr = OFFSET_TO_OFFBASE(curr - (matchIndex + dmsIndexDelta));
                    if (ip+currentMl == iLimit) break;
                }
            }
        }
    }
    return ml;
}


/**
 * Generate search functions templated on (dictMode, mls, rowLog).
 * These functions are outlined for code size & compilation time.
 * ZSTD_searchMax() dispatches to the correct implementation function.
 *
 * TODO: The start of the search function involves loading and calculating a
 * bunch of constants from the ZSTD_MatchState_t. These computations could be
 * done in an initialization function, and saved somewhere in the match state.
 * Then we could pass a pointer to the saved state instead of the match state,
 * and avoid duplicate computations.
 *
 * TODO: Move the match re-winding into searchMax. This improves compression
 * ratio, and unlocks further simplifications with the next TODO.
 *
 * TODO: Try moving the repcode search into searchMax. After the re-winding
 * and repcode search are in searchMax, there is no more logic in the match
 * finder loop that requires knowledge about the dictMode. So we should be
 * able to avoid force inlining it, and we can join the extDict loop with
 * the single segment loop. It should go in searchMax instead of its own
 * function to avoid having multiple virtual function calls per search.
 */

#define ZSTD_BT_SEARCH_FN(dictMode, mls) ZSTD_BtFindBestMatch_##dictMode##_##mls
#define ZSTD_HC_SEARCH_FN(dictMode, mls) ZSTD_HcFindBestMatch_##dictMode##_##mls
#define ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog) ZSTD_RowFindBestMatch_##dictMode##_##mls##_##rowLog

#define ZSTD_SEARCH_FN_ATTRS FORCE_NOINLINE

#define GEN_ZSTD_BT_SEARCH_FN(dictMode, mls)                                           \
    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_BT_SEARCH_FN(dictMode, mls)(                      \
            ZSTD_MatchState_t* ms,                                                     \
            const BYTE* ip, const BYTE* const iLimit,                                  \
            size_t* offBasePtr)                                                        \
    {                                                                                  \
        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                           \
        return ZSTD_BtFindBestMatch(ms, ip, iLimit, offBasePtr, mls, ZSTD_##dictMode); \
    }                                                                                  \

#define GEN_ZSTD_HC_SEARCH_FN(dictMode, mls)                                          \
    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_HC_SEARCH_FN(dictMode, mls)(                     \
            ZSTD_MatchState_t* ms,                                                    \
            const BYTE* ip, const BYTE* const iLimit,                                 \
            size_t* offsetPtr)                                                        \
    {                                                                                 \
        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                          \
        return ZSTD_HcFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode); \
    }                                                                                 \

#define GEN_ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)                                          \
    ZSTD_SEARCH_FN_ATTRS size_t ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(                     \
            ZSTD_MatchState_t* ms,                                                             \
            const BYTE* ip, const BYTE* const iLimit,                                          \
            size_t* offsetPtr)                                                                 \
    {                                                                                          \
        assert(MAX(4, MIN(6, ms->cParams.minMatch)) == mls);                                   \
        assert(MAX(4, MIN(6, ms->cParams.searchLog)) == rowLog);                               \
        return ZSTD_RowFindBestMatch(ms, ip, iLimit, offsetPtr, mls, ZSTD_##dictMode, rowLog); \
    }                                                                                          \

#define ZSTD_FOR_EACH_ROWLOG(X, dictMode, mls) \
    X(dictMode, mls, 4)                        \
    X(dictMode, mls, 5)                        \
    X(dictMode, mls, 6)

#define ZSTD_FOR_EACH_MLS_ROWLOG(X, dictMode) \
    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 4)      \
    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 5)      \
    ZSTD_FOR_EACH_ROWLOG(X, dictMode, 6)

#define ZSTD_FOR_EACH_MLS(X, dictMode) \
    X(dictMode, 4)                     \
    X(dictMode, 5)                     \
    X(dictMode, 6)

#define ZSTD_FOR_EACH_DICT_MODE(X, ...) \
    X(__VA_ARGS__, noDict)              \
    X(__VA_ARGS__, extDict)             \
    X(__VA_ARGS__, dictMatchState)      \
    X(__VA_ARGS__, dedicatedDictSearch)

/* Generate row search fns for each combination of (dictMode, mls, rowLog) */
ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS_ROWLOG, GEN_ZSTD_ROW_SEARCH_FN)
/* Generate binary Tree search fns for each combination of (dictMode, mls) */
ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_BT_SEARCH_FN)
/* Generate hash chain search fns for each combination of (dictMode, mls) */
ZSTD_FOR_EACH_DICT_MODE(ZSTD_FOR_EACH_MLS, GEN_ZSTD_HC_SEARCH_FN)

typedef enum { search_hashChain=0, search_binaryTree=1, search_rowHash=2 } searchMethod_e;

#define GEN_ZSTD_CALL_BT_SEARCH_FN(dictMode, mls)                         \
    case mls:                                                             \
        return ZSTD_BT_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
#define GEN_ZSTD_CALL_HC_SEARCH_FN(dictMode, mls)                         \
    case mls:                                                             \
        return ZSTD_HC_SEARCH_FN(dictMode, mls)(ms, ip, iend, offsetPtr);
#define GEN_ZSTD_CALL_ROW_SEARCH_FN(dictMode, mls, rowLog)                         \
    case rowLog:                                                                   \
        return ZSTD_ROW_SEARCH_FN(dictMode, mls, rowLog)(ms, ip, iend, offsetPtr);

#define ZSTD_SWITCH_MLS(X, dictMode)   \
    switch (mls) {                     \
        ZSTD_FOR_EACH_MLS(X, dictMode) \
    }

#define ZSTD_SWITCH_ROWLOG(dictMode, mls)                                    \
    case mls:                                                                \
        switch (rowLog) {                                                    \
            ZSTD_FOR_EACH_ROWLOG(GEN_ZSTD_CALL_ROW_SEARCH_FN, dictMode, mls) \
        }                                                                    \
        ZSTD_UNREACHABLE;                                                    \
        break;

#define ZSTD_SWITCH_SEARCH_METHOD(dictMode)                       \
    switch (searchMethod) {                                       \
        case search_hashChain:                                    \
            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_HC_SEARCH_FN, dictMode) \
            break;                                                \
        case search_binaryTree:                                   \
            ZSTD_SWITCH_MLS(GEN_ZSTD_CALL_BT_SEARCH_FN, dictMode) \
            break;                                                \
        case search_rowHash:                                      \
            ZSTD_SWITCH_MLS(ZSTD_SWITCH_ROWLOG, dictMode)         \
            break;                                                \
    }                                                             \
    ZSTD_UNREACHABLE;

/**
 * Searches for the longest match at @p ip.
 * Dispatches to the correct implementation function based on the
 * (searchMethod, dictMode, mls, rowLog). We use switch statements
 * here instead of using an indirect function call through a function
 * pointer because after Spectre and Meltdown mitigations, indirect
 * function calls can be very costly, especially in the kernel.
 *
 * NOTE: dictMode and searchMethod should be templated, so those switch
 * statements should be optimized out. Only the mls & rowLog switches
 * should be left.
 *
 * @param ms The match state.
 * @param ip The position to search at.
 * @param iend The end of the input data.
 * @param[out] offsetPtr Stores the match offset into this pointer.
 * @param mls The minimum search length, in the range [4, 6].
 * @param rowLog The row log (if applicable), in the range [4, 6].
 * @param searchMethod The search method to use (templated).
 * @param dictMode The dictMode (templated).
 *
 * @returns The length of the longest match found, or < mls if no match is found.
 * If a match is found its offset is stored in @p offsetPtr.
 */
FORCE_INLINE_TEMPLATE size_t ZSTD_searchMax(
    ZSTD_MatchState_t* ms,
    const BYTE* ip,
    const BYTE* iend,
    size_t* offsetPtr,
    U32 const mls,
    U32 const rowLog,
    searchMethod_e const searchMethod,
    ZSTD_dictMode_e const dictMode)
{
    if (dictMode == ZSTD_noDict) {
        ZSTD_SWITCH_SEARCH_METHOD(noDict)
    } else if (dictMode == ZSTD_extDict) {
        ZSTD_SWITCH_SEARCH_METHOD(extDict)
    } else if (dictMode == ZSTD_dictMatchState) {
        ZSTD_SWITCH_SEARCH_METHOD(dictMatchState)
    } else if (dictMode == ZSTD_dedicatedDictSearch) {
        ZSTD_SWITCH_SEARCH_METHOD(dedicatedDictSearch)
    }
    ZSTD_UNREACHABLE;
    return 0;
}

/* *******************************
*  Common parser - lazy strategy
*********************************/

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_lazy_generic(
                        ZSTD_MatchState_t* ms, SeqStore_t* seqStore,
                        U32 rep[ZSTD_REP_NUM],
                        const void* src, size_t srcSize,
                        const searchMethod_e searchMethod, const U32 depth,
                        ZSTD_dictMode_e const dictMode)
{
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip = istart;
    const BYTE* anchor = istart;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = (searchMethod == search_rowHash) ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
    const BYTE* const base = ms->window.base;
    const U32 prefixLowestIndex = ms->window.dictLimit;
    const BYTE* const prefixLowest = base + prefixLowestIndex;
    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);

    U32 offset_1 = rep[0], offset_2 = rep[1];
    U32 offsetSaved1 = 0, offsetSaved2 = 0;

    const int isDMS = dictMode == ZSTD_dictMatchState;
    const int isDDS = dictMode == ZSTD_dedicatedDictSearch;
    const int isDxS = isDMS || isDDS;
    const ZSTD_MatchState_t* const dms = ms->dictMatchState;
    const U32 dictLowestIndex      = isDxS ? dms->window.dictLimit : 0;
    const BYTE* const dictBase     = isDxS ? dms->window.base : NULL;
    const BYTE* const dictLowest   = isDxS ? dictBase + dictLowestIndex : NULL;
    const BYTE* const dictEnd      = isDxS ? dms->window.nextSrc : NULL;
    const U32 dictIndexDelta       = isDxS ?
                                     prefixLowestIndex - (U32)(dictEnd - dictBase) :
                                     0;
    const U32 dictAndPrefixLength = (U32)((ip - prefixLowest) + (dictEnd - dictLowest));

    DEBUGLOG(5, "ZSTD_compressBlock_lazy_generic (dictMode=%u) (searchFunc=%u)", (U32)dictMode, (U32)searchMethod);
    ip += (dictAndPrefixLength == 0);
    if (dictMode == ZSTD_noDict) {
        U32 const curr = (U32)(ip - base);
        U32 const windowLow = ZSTD_getLowestPrefixIndex(ms, curr, ms->cParams.windowLog);
        U32 const maxRep = curr - windowLow;
        if (offset_2 > maxRep) offsetSaved2 = offset_2, offset_2 = 0;
        if (offset_1 > maxRep) offsetSaved1 = offset_1, offset_1 = 0;
    }
    if (isDxS) {
        /* dictMatchState repCode checks don't currently handle repCode == 0
         * disabling. */
        assert(offset_1 <= dictAndPrefixLength);
        assert(offset_2 <= dictAndPrefixLength);
    }

    /* Reset the lazy skipping state */
    ms->lazySkipping = 0;

    if (searchMethod == search_rowHash) {
        ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
    }

    /* Match Loop */
#if defined(__GNUC__) && defined(__x86_64__)
    /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
     * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
     */
    __asm__(".p2align 5");
#endif
    while (ip < ilimit) {
        size_t matchLength=0;
        size_t offBase = REPCODE1_TO_OFFBASE;
        const BYTE* start=ip+1;
        DEBUGLOG(7, "search baseline (depth 0)");

        /* check repCode */
        if (isDxS) {
            const U32 repIndex = (U32)(ip - base) + 1 - offset_1;
            const BYTE* repMatch = ((dictMode == ZSTD_dictMatchState || dictMode == ZSTD_dedicatedDictSearch)
                                && repIndex < prefixLowestIndex) ?
                                   dictBase + (repIndex - dictIndexDelta) :
                                   base + repIndex;
            if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                && (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
                const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                if (depth==0) goto _storeSequence;
            }
        }
        if ( dictMode == ZSTD_noDict
          && ((offset_1 > 0) & (MEM_read32(ip+1-offset_1) == MEM_read32(ip+1)))) {
            matchLength = ZSTD_count(ip+1+4, ip+1+4-offset_1, iend) + 4;
            if (depth==0) goto _storeSequence;
        }

        /* first search (depth 0) */
        {   size_t offbaseFound = 999999999;
            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &offbaseFound, mls, rowLog, searchMethod, dictMode);
            if (ml2 > matchLength)
                matchLength = ml2, start = ip, offBase = offbaseFound;
        }

        if (matchLength < 4) {
            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength) + 1;   /* jump faster over incompressible sections */;
            ip += step;
            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
             * In this mode we stop inserting every position into our tables, and only insert
             * positions that we search, which is one in step positions.
             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
             * triggered once we've gone 2KB without finding any matches.
             */
            ms->lazySkipping = step > kLazySkippingStep;
            continue;
        }

        /* let's try to find a better solution */
        if (depth>=1)
        while (ip<ilimit) {
            DEBUGLOG(7, "search depth 1");
            ip ++;
            if ( (dictMode == ZSTD_noDict)
              && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                int const gain2 = (int)(mlRep * 3);
                int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                if ((mlRep >= 4) && (gain2 > gain1))
                    matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
            }
            if (isDxS) {
                const U32 repIndex = (U32)(ip - base) - offset_1;
                const BYTE* repMatch = repIndex < prefixLowestIndex ?
                               dictBase + (repIndex - dictIndexDelta) :
                               base + repIndex;
                if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                    && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                    const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                    size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                    int const gain2 = (int)(mlRep * 3);
                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                    if ((mlRep >= 4) && (gain2 > gain1))
                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                }
            }
            {   size_t ofbCandidate=999999999;
                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                if ((ml2 >= 4) && (gain2 > gain1)) {
                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                    continue;   /* search a better one */
            }   }

            /* let's find an even better one */
            if ((depth==2) && (ip<ilimit)) {
                DEBUGLOG(7, "search depth 2");
                ip ++;
                if ( (dictMode == ZSTD_noDict)
                  && (offBase) && ((offset_1>0) & (MEM_read32(ip) == MEM_read32(ip - offset_1)))) {
                    size_t const mlRep = ZSTD_count(ip+4, ip+4-offset_1, iend) + 4;
                    int const gain2 = (int)(mlRep * 4);
                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                    if ((mlRep >= 4) && (gain2 > gain1))
                        matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                }
                if (isDxS) {
                    const U32 repIndex = (U32)(ip - base) - offset_1;
                    const BYTE* repMatch = repIndex < prefixLowestIndex ?
                                   dictBase + (repIndex - dictIndexDelta) :
                                   base + repIndex;
                    if ((ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                        && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                        const BYTE* repMatchEnd = repIndex < prefixLowestIndex ? dictEnd : iend;
                        size_t const mlRep = ZSTD_count_2segments(ip+4, repMatch+4, iend, repMatchEnd, prefixLowest) + 4;
                        int const gain2 = (int)(mlRep * 4);
                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                        if ((mlRep >= 4) && (gain2 > gain1))
                            matchLength = mlRep, offBase = REPCODE1_TO_OFFBASE, start = ip;
                    }
                }
                {   size_t ofbCandidate=999999999;
                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, dictMode);
                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                    if ((ml2 >= 4) && (gain2 > gain1)) {
                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                        continue;
            }   }   }
            break;  /* nothing found : store previous solution */
        }

        /* NOTE:
         * Pay attention that `start[-value]` can lead to strange undefined behavior
         * notably if `value` is unsigned, resulting in a large positive `-value`.
         */
        /* catch up */
        if (OFFBASE_IS_OFFSET(offBase)) {
            if (dictMode == ZSTD_noDict) {
                while ( ((start > anchor) & (start - OFFBASE_TO_OFFSET(offBase) > prefixLowest))
                     && (start[-1] == (start-OFFBASE_TO_OFFSET(offBase))[-1]) )  /* only search for offset within prefix */
                    { start--; matchLength++; }
            }
            if (isDxS) {
                U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
                const BYTE* match = (matchIndex < prefixLowestIndex) ? dictBase + matchIndex - dictIndexDelta : base + matchIndex;
                const BYTE* const mStart = (matchIndex < prefixLowestIndex) ? dictLowest : prefixLowest;
                while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
            }
            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
        }
        /* store sequence */
_storeSequence:
        {   size_t const litLength = (size_t)(start - anchor);
            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
            anchor = ip = start + matchLength;
        }
        if (ms->lazySkipping) {
            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
            if (searchMethod == search_rowHash) {
                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
            }
            ms->lazySkipping = 0;
        }

        /* check immediate repcode */
        if (isDxS) {
            while (ip <= ilimit) {
                U32 const current2 = (U32)(ip-base);
                U32 const repIndex = current2 - offset_2;
                const BYTE* repMatch = repIndex < prefixLowestIndex ?
                        dictBase - dictIndexDelta + repIndex :
                        base + repIndex;
                if ( (ZSTD_index_overlap_check(prefixLowestIndex, repIndex))
                   && (MEM_read32(repMatch) == MEM_read32(ip)) ) {
                    const BYTE* const repEnd2 = repIndex < prefixLowestIndex ? dictEnd : iend;
                    matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd2, prefixLowest) + 4;
                    offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset_2 <=> offset_1 */
                    ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                    ip += matchLength;
                    anchor = ip;
                    continue;
                }
                break;
            }
        }

        if (dictMode == ZSTD_noDict) {
            while ( ((ip <= ilimit) & (offset_2>0))
                 && (MEM_read32(ip) == MEM_read32(ip - offset_2)) ) {
                /* store sequence */
                matchLength = ZSTD_count(ip+4, ip+4-offset_2, iend) + 4;
                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase; /* swap repcodes */
                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                ip += matchLength;
                anchor = ip;
                continue;   /* faster when present ... (?) */
    }   }   }

    /* If offset_1 started invalid (offsetSaved1 != 0) and became valid (offset_1 != 0),
     * rotate saved offsets. See comment in ZSTD_compressBlock_fast_noDict for more context. */
    offsetSaved2 = ((offsetSaved1 != 0) && (offset_1 != 0)) ? offsetSaved1 : offsetSaved2;

    /* save reps for next block */
    rep[0] = offset_1 ? offset_1 : offsetSaved1;
    rep[1] = offset_2 ? offset_2 : offsetSaved2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}
#endif /* build exclusions */


#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_greedy(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_noDict);
}

size_t ZSTD_compressBlock_greedy_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_greedy_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0, ZSTD_dedicatedDictSearch);
}

size_t ZSTD_compressBlock_greedy_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_noDict);
}

size_t ZSTD_compressBlock_greedy_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_greedy_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0, ZSTD_dedicatedDictSearch);
}
#endif

#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_noDict);
}

size_t ZSTD_compressBlock_lazy_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_lazy_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1, ZSTD_dedicatedDictSearch);
}

size_t ZSTD_compressBlock_lazy_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_noDict);
}

size_t ZSTD_compressBlock_lazy_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_lazy_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1, ZSTD_dedicatedDictSearch);
}
#endif

#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_noDict);
}

size_t ZSTD_compressBlock_lazy2_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2, ZSTD_dedicatedDictSearch);
}

size_t ZSTD_compressBlock_lazy2_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_noDict);
}

size_t ZSTD_compressBlock_lazy2_dictMatchState_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_lazy2_dedicatedDictSearch_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2, ZSTD_dedicatedDictSearch);
}
#endif

#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btlazy2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_noDict);
}

size_t ZSTD_compressBlock_btlazy2_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2, ZSTD_dictMatchState);
}
#endif

#if !defined(ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR)
FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_compressBlock_lazy_extDict_generic(
                        ZSTD_MatchState_t* ms, SeqStore_t* seqStore,
                        U32 rep[ZSTD_REP_NUM],
                        const void* src, size_t srcSize,
                        const searchMethod_e searchMethod, const U32 depth)
{
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip = istart;
    const BYTE* anchor = istart;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = searchMethod == search_rowHash ? iend - 8 - ZSTD_ROW_HASH_CACHE_SIZE : iend - 8;
    const BYTE* const base = ms->window.base;
    const U32 dictLimit = ms->window.dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    const BYTE* const dictBase = ms->window.dictBase;
    const BYTE* const dictEnd  = dictBase + dictLimit;
    const BYTE* const dictStart  = dictBase + ms->window.lowLimit;
    const U32 windowLog = ms->cParams.windowLog;
    const U32 mls = BOUNDED(4, ms->cParams.minMatch, 6);
    const U32 rowLog = BOUNDED(4, ms->cParams.searchLog, 6);

    U32 offset_1 = rep[0], offset_2 = rep[1];

    DEBUGLOG(5, "ZSTD_compressBlock_lazy_extDict_generic (searchFunc=%u)", (U32)searchMethod);

    /* Reset the lazy skipping state */
    ms->lazySkipping = 0;

    /* init */
    ip += (ip == prefixStart);
    if (searchMethod == search_rowHash) {
        ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
    }

    /* Match Loop */
#if defined(__GNUC__) && defined(__x86_64__)
    /* I've measured random a 5% speed loss on levels 5 & 6 (greedy) when the
     * code alignment is perturbed. To fix the instability align the loop on 32-bytes.
     */
    __asm__(".p2align 5");
#endif
    while (ip < ilimit) {
        size_t matchLength=0;
        size_t offBase = REPCODE1_TO_OFFBASE;
        const BYTE* start=ip+1;
        U32 curr = (U32)(ip-base);

        /* check repCode */
        {   const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr+1, windowLog);
            const U32 repIndex = (U32)(curr+1 - offset_1);
            const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
            const BYTE* const repMatch = repBase + repIndex;
            if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
               & (offset_1 <= curr+1 - windowLow) ) /* note: we are searching at curr+1 */
            if (MEM_read32(ip+1) == MEM_read32(repMatch)) {
                /* repcode detected we should take it */
                const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                matchLength = ZSTD_count_2segments(ip+1+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                if (depth==0) goto _storeSequence;
        }   }

        /* first search (depth 0) */
        {   size_t ofbCandidate = 999999999;
            size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
            if (ml2 > matchLength)
                matchLength = ml2, start = ip, offBase = ofbCandidate;
        }

        if (matchLength < 4) {
            size_t const step = ((size_t)(ip-anchor) >> kSearchStrength);
            ip += step + 1;   /* jump faster over incompressible sections */
            /* Enter the lazy skipping mode once we are skipping more than 8 bytes at a time.
             * In this mode we stop inserting every position into our tables, and only insert
             * positions that we search, which is one in step positions.
             * The exact cutoff is flexible, I've just chosen a number that is reasonably high,
             * so we minimize the compression ratio loss in "normal" scenarios. This mode gets
             * triggered once we've gone 2KB without finding any matches.
             */
            ms->lazySkipping = step > kLazySkippingStep;
            continue;
        }

        /* let's try to find a better solution */
        if (depth>=1)
        while (ip<ilimit) {
            ip ++;
            curr++;
            /* check repCode */
            if (offBase) {
                const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                const U32 repIndex = (U32)(curr - offset_1);
                const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                const BYTE* const repMatch = repBase + repIndex;
                if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                   & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                if (MEM_read32(ip) == MEM_read32(repMatch)) {
                    /* repcode detected */
                    const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                    size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                    int const gain2 = (int)(repLength * 3);
                    int const gain1 = (int)(matchLength*3 - ZSTD_highbit32((U32)offBase) + 1);
                    if ((repLength >= 4) && (gain2 > gain1))
                        matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
            }   }

            /* search match, depth 1 */
            {   size_t ofbCandidate = 999999999;
                size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
                int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
                int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 4);
                if ((ml2 >= 4) && (gain2 > gain1)) {
                    matchLength = ml2, offBase = ofbCandidate, start = ip;
                    continue;   /* search a better one */
            }   }

            /* let's find an even better one */
            if ((depth==2) && (ip<ilimit)) {
                ip ++;
                curr++;
                /* check repCode */
                if (offBase) {
                    const U32 windowLow = ZSTD_getLowestMatchIndex(ms, curr, windowLog);
                    const U32 repIndex = (U32)(curr - offset_1);
                    const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
                    const BYTE* const repMatch = repBase + repIndex;
                    if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
                       & (offset_1 <= curr - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
                    if (MEM_read32(ip) == MEM_read32(repMatch)) {
                        /* repcode detected */
                        const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                        size_t const repLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                        int const gain2 = (int)(repLength * 4);
                        int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 1);
                        if ((repLength >= 4) && (gain2 > gain1))
                            matchLength = repLength, offBase = REPCODE1_TO_OFFBASE, start = ip;
                }   }

                /* search match, depth 2 */
                {   size_t ofbCandidate = 999999999;
                    size_t const ml2 = ZSTD_searchMax(ms, ip, iend, &ofbCandidate, mls, rowLog, searchMethod, ZSTD_extDict);
                    int const gain2 = (int)(ml2*4 - ZSTD_highbit32((U32)ofbCandidate));   /* raw approx */
                    int const gain1 = (int)(matchLength*4 - ZSTD_highbit32((U32)offBase) + 7);
                    if ((ml2 >= 4) && (gain2 > gain1)) {
                        matchLength = ml2, offBase = ofbCandidate, start = ip;
                        continue;
            }   }   }
            break;  /* nothing found : store previous solution */
        }

        /* catch up */
        if (OFFBASE_IS_OFFSET(offBase)) {
            U32 const matchIndex = (U32)((size_t)(start-base) - OFFBASE_TO_OFFSET(offBase));
            const BYTE* match = (matchIndex < dictLimit) ? dictBase + matchIndex : base + matchIndex;
            const BYTE* const mStart = (matchIndex < dictLimit) ? dictStart : prefixStart;
            while ((start>anchor) && (match>mStart) && (start[-1] == match[-1])) { start--; match--; matchLength++; }  /* catch up */
            offset_2 = offset_1; offset_1 = (U32)OFFBASE_TO_OFFSET(offBase);
        }

        /* store sequence */
_storeSequence:
        {   size_t const litLength = (size_t)(start - anchor);
            ZSTD_storeSeq(seqStore, litLength, anchor, iend, (U32)offBase, matchLength);
            anchor = ip = start + matchLength;
        }
        if (ms->lazySkipping) {
            /* We've found a match, disable lazy skipping mode, and refill the hash cache. */
            if (searchMethod == search_rowHash) {
                ZSTD_row_fillHashCache(ms, base, rowLog, mls, ms->nextToUpdate, ilimit);
            }
            ms->lazySkipping = 0;
        }

        /* check immediate repcode */
        while (ip <= ilimit) {
            const U32 repCurrent = (U32)(ip-base);
            const U32 windowLow = ZSTD_getLowestMatchIndex(ms, repCurrent, windowLog);
            const U32 repIndex = repCurrent - offset_2;
            const BYTE* const repBase = repIndex < dictLimit ? dictBase : base;
            const BYTE* const repMatch = repBase + repIndex;
            if ( (ZSTD_index_overlap_check(dictLimit, repIndex))
               & (offset_2 <= repCurrent - windowLow) ) /* equivalent to `curr > repIndex >= windowLow` */
            if (MEM_read32(ip) == MEM_read32(repMatch)) {
                /* repcode detected we should take it */
                const BYTE* const repEnd = repIndex < dictLimit ? dictEnd : iend;
                matchLength = ZSTD_count_2segments(ip+4, repMatch+4, iend, repEnd, prefixStart) + 4;
                offBase = offset_2; offset_2 = offset_1; offset_1 = (U32)offBase;   /* swap offset history */
                ZSTD_storeSeq(seqStore, 0, anchor, iend, REPCODE1_TO_OFFBASE, matchLength);
                ip += matchLength;
                anchor = ip;
                continue;   /* faster when present ... (?) */
            }
            break;
    }   }

    /* Save reps for next block */
    rep[0] = offset_1;
    rep[1] = offset_2;

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}
#endif /* build exclusions */

#ifndef ZSTD_EXCLUDE_GREEDY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_greedy_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 0);
}

size_t ZSTD_compressBlock_greedy_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 0);
}
#endif

#ifndef ZSTD_EXCLUDE_LAZY_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)

{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 1);
}

size_t ZSTD_compressBlock_lazy_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)

{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 1);
}
#endif

#ifndef ZSTD_EXCLUDE_LAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_lazy2_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)

{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_hashChain, 2);
}

size_t ZSTD_compressBlock_lazy2_extDict_row(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)
{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_rowHash, 2);
}
#endif

#ifndef ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btlazy2_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        void const* src, size_t srcSize)

{
    return ZSTD_compressBlock_lazy_extDict_generic(ms, seqStore, rep, src, srcSize, search_binaryTree, 2);
}
#endif
/**** ended inlining compress/zstd_lazy.c ****/
/**** start inlining compress/zstd_ldm.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: zstd_ldm.h ****/

/**** skipping file: ../common/debug.h ****/
/**** skipping file: ../common/xxhash.h ****/
/**** skipping file: zstd_fast.h ****/
/**** skipping file: zstd_double_fast.h ****/
/**** start inlining zstd_ldm_geartab.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_LDM_GEARTAB_H
#define ZSTD_LDM_GEARTAB_H

/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/mem.h ****/

static UNUSED_ATTR const U64 ZSTD_ldm_gearTab[256] = {
    0xf5b8f72c5f77775c, 0x84935f266b7ac412, 0xb647ada9ca730ccc,
    0xb065bb4b114fb1de, 0x34584e7e8c3a9fd0, 0x4e97e17c6ae26b05,
    0x3a03d743bc99a604, 0xcecd042422c4044f, 0x76de76c58524259e,
    0x9c8528f65badeaca, 0x86563706e2097529, 0x2902475fa375d889,
    0xafb32a9739a5ebe6, 0xce2714da3883e639, 0x21eaf821722e69e,
    0x37b628620b628,    0x49a8d455d88caf5,  0x8556d711e6958140,
    0x4f7ae74fc605c1f,  0x829f0c3468bd3a20, 0x4ffdc885c625179e,
    0x8473de048a3daf1b, 0x51008822b05646b2, 0x69d75d12b2d1cc5f,
    0x8c9d4a19159154bc, 0xc3cc10f4abbd4003, 0xd06ddc1cecb97391,
    0xbe48e6e7ed80302e, 0x3481db31cee03547, 0xacc3f67cdaa1d210,
    0x65cb771d8c7f96cc, 0x8eb27177055723dd, 0xc789950d44cd94be,
    0x934feadc3700b12b, 0x5e485f11edbdf182, 0x1e2e2a46fd64767a,
    0x2969ca71d82efa7c, 0x9d46e9935ebbba2e, 0xe056b67e05e6822b,
    0x94d73f55739d03a0, 0xcd7010bdb69b5a03, 0x455ef9fcd79b82f4,
    0x869cb54a8749c161, 0x38d1a4fa6185d225, 0xb475166f94bbe9bb,
    0xa4143548720959f1, 0x7aed4780ba6b26ba, 0xd0ce264439e02312,
    0x84366d746078d508, 0xa8ce973c72ed17be, 0x21c323a29a430b01,
    0x9962d617e3af80ee, 0xab0ce91d9c8cf75b, 0x530e8ee6d19a4dbc,
    0x2ef68c0cf53f5d72, 0xc03a681640a85506, 0x496e4e9f9c310967,
    0x78580472b59b14a0, 0x273824c23b388577, 0x66bf923ad45cb553,
    0x47ae1a5a2492ba86, 0x35e304569e229659, 0x4765182a46870b6f,
    0x6cbab625e9099412, 0xddac9a2e598522c1, 0x7172086e666624f2,
    0xdf5003ca503b7837, 0x88c0c1db78563d09, 0x58d51865acfc289d,
    0x177671aec65224f1, 0xfb79d8a241e967d7, 0x2be1e101cad9a49a,
    0x6625682f6e29186b, 0x399553457ac06e50, 0x35dffb4c23abb74,
    0x429db2591f54aade, 0xc52802a8037d1009, 0x6acb27381f0b25f3,
    0xf45e2551ee4f823b, 0x8b0ea2d99580c2f7, 0x3bed519cbcb4e1e1,
    0xff452823dbb010a,  0x9d42ed614f3dd267, 0x5b9313c06257c57b,
    0xa114b8008b5e1442, 0xc1fe311c11c13d4b, 0x66e8763ea34c5568,
    0x8b982af1c262f05d, 0xee8876faaa75fbb7, 0x8a62a4d0d172bb2a,
    0xc13d94a3b7449a97, 0x6dbbba9dc15d037c, 0xc786101f1d92e0f1,
    0xd78681a907a0b79b, 0xf61aaf2962c9abb9, 0x2cfd16fcd3cb7ad9,
    0x868c5b6744624d21, 0x25e650899c74ddd7, 0xba042af4a7c37463,
    0x4eb1a539465a3eca, 0xbe09dbf03b05d5ca, 0x774e5a362b5472ba,
    0x47a1221229d183cd, 0x504b0ca18ef5a2df, 0xdffbdfbde2456eb9,
    0x46cd2b2fbee34634, 0xf2aef8fe819d98c3, 0x357f5276d4599d61,
    0x24a5483879c453e3, 0x88026889192b4b9,  0x28da96671782dbec,
    0x4ef37c40588e9aaa, 0x8837b90651bc9fb3, 0xc164f741d3f0e5d6,
    0xbc135a0a704b70ba, 0x69cd868f7622ada,  0xbc37ba89e0b9c0ab,
    0x47c14a01323552f6, 0x4f00794bacee98bb, 0x7107de7d637a69d5,
    0x88af793bb6f2255e, 0xf3c6466b8799b598, 0xc288c616aa7f3b59,
    0x81ca63cf42fca3fd, 0x88d85ace36a2674b, 0xd056bd3792389e7,
    0xe55c396c4e9dd32d, 0xbefb504571e6c0a6, 0x96ab32115e91e8cc,
    0xbf8acb18de8f38d1, 0x66dae58801672606, 0x833b6017872317fb,
    0xb87c16f2d1c92864, 0xdb766a74e58b669c, 0x89659f85c61417be,
    0xc8daad856011ea0c, 0x76a4b565b6fe7eae, 0xa469d085f6237312,
    0xaaf0365683a3e96c, 0x4dbb746f8424f7b8, 0x638755af4e4acc1,
    0x3d7807f5bde64486, 0x17be6d8f5bbb7639, 0x903f0cd44dc35dc,
    0x67b672eafdf1196c, 0xa676ff93ed4c82f1, 0x521d1004c5053d9d,
    0x37ba9ad09ccc9202, 0x84e54d297aacfb51, 0xa0b4b776a143445,
    0x820d471e20b348e,  0x1874383cb83d46dc, 0x97edeec7a1efe11c,
    0xb330e50b1bdc42aa, 0x1dd91955ce70e032, 0xa514cdb88f2939d5,
    0x2791233fd90db9d3, 0x7b670a4cc50f7a9b, 0x77c07d2a05c6dfa5,
    0xe3778b6646d0a6fa, 0xb39c8eda47b56749, 0x933ed448addbef28,
    0xaf846af6ab7d0bf4, 0xe5af208eb666e49,  0x5e6622f73534cd6a,
    0x297daeca42ef5b6e, 0x862daef3d35539a6, 0xe68722498f8e1ea9,
    0x981c53093dc0d572, 0xfa09b0bfbf86fbf5, 0x30b1e96166219f15,
    0x70e7d466bdc4fb83, 0x5a66736e35f2a8e9, 0xcddb59d2b7c1baef,
    0xd6c7d247d26d8996, 0xea4e39eac8de1ba3, 0x539c8bb19fa3aff2,
    0x9f90e4c5fd508d8,  0xa34e5956fbaf3385, 0x2e2f8e151d3ef375,
    0x173691e9b83faec1, 0xb85a8d56bf016379, 0x8382381267408ae3,
    0xb90f901bbdc0096d, 0x7c6ad32933bcec65, 0x76bb5e2f2c8ad595,
    0x390f851a6cf46d28, 0xc3e6064da1c2da72, 0xc52a0c101cfa5389,
    0xd78eaf84a3fbc530, 0x3781b9e2288b997e, 0x73c2f6dea83d05c4,
    0x4228e364c5b5ed7,  0x9d7a3edf0da43911, 0x8edcfeda24686756,
    0x5e7667a7b7a9b3a1, 0x4c4f389fa143791d, 0xb08bc1023da7cddc,
    0x7ab4be3ae529b1cc, 0x754e6132dbe74ff9, 0x71635442a839df45,
    0x2f6fb1643fbe52de, 0x961e0a42cf7a8177, 0xf3b45d83d89ef2ea,
    0xee3de4cf4a6e3e9b, 0xcd6848542c3295e7, 0xe4cee1664c78662f,
    0x9947548b474c68c4, 0x25d73777a5ed8b0b, 0xc915b1d636b7fc,
    0x21c2ba75d9b0d2da, 0x5f6b5dcf608a64a1, 0xdcf333255ff9570c,
    0x633b922418ced4ee, 0xc136dde0b004b34a, 0x58cc83b05d4b2f5a,
    0x5eb424dda28e42d2, 0x62df47369739cd98, 0xb4e0b42485e4ce17,
    0x16e1f0c1f9a8d1e7, 0x8ec3916707560ebf, 0x62ba6e2df2cc9db3,
    0xcbf9f4ff77d83a16, 0x78d9d7d07d2bbcc4, 0xef554ce1e02c41f4,
    0x8d7581127eccf94d, 0xa9b53336cb3c8a05, 0x38c42c0bf45c4f91,
    0x640893cdf4488863, 0x80ec34bc575ea568, 0x39f324f5b48eaa40,
    0xe9d9ed1f8eff527f, 0x9224fc058cc5a214, 0xbaba00b04cfe7741,
    0x309a9f120fcf52af, 0xa558f3ec65626212, 0x424bec8b7adabe2f,
    0x41622513a6aea433, 0xb88da2d5324ca798, 0xd287733b245528a4,
    0x9a44697e6d68aec3, 0x7b1093be2f49bb28, 0x50bbec632e3d8aad,
    0x6cd90723e1ea8283, 0x897b9e7431b02bf3, 0x219efdcb338a7047,
    0x3b0311f0a27c0656, 0xdb17bf91c0db96e7, 0x8cd4fd6b4e85a5b2,
    0xfab071054ba6409d, 0x40d6fe831fa9dfd9, 0xaf358debad7d791e,
    0xeb8d0e25a65e3e58, 0xbbcbd3df14e08580, 0xcf751f27ecdab2b,
    0x2b4da14f2613d8f4
};

#endif /* ZSTD_LDM_GEARTAB_H */
/**** ended inlining zstd_ldm_geartab.h ****/

#define LDM_BUCKET_SIZE_LOG 4
#define LDM_MIN_MATCH_LENGTH 64
#define LDM_HASH_RLOG 7

typedef struct {
    U64 rolling;
    U64 stopMask;
} ldmRollingHashState_t;

/** ZSTD_ldm_gear_init():
 *
 * Initializes the rolling hash state such that it will honor the
 * settings in params. */
static void ZSTD_ldm_gear_init(ldmRollingHashState_t* state, ldmParams_t const* params)
{
    unsigned maxBitsInMask = MIN(params->minMatchLength, 64);
    unsigned hashRateLog = params->hashRateLog;

    state->rolling = ~(U32)0;

    /* The choice of the splitting criterion is subject to two conditions:
     *   1. it has to trigger on average every 2^(hashRateLog) bytes;
     *   2. ideally, it has to depend on a window of minMatchLength bytes.
     *
     * In the gear hash algorithm, bit n depends on the last n bytes;
     * so in order to obtain a good quality splitting criterion it is
     * preferable to use bits with high weight.
     *
     * To match condition 1 we use a mask with hashRateLog bits set
     * and, because of the previous remark, we make sure these bits
     * have the highest possible weight while still respecting
     * condition 2.
     */
    if (hashRateLog > 0 && hashRateLog <= maxBitsInMask) {
        state->stopMask = (((U64)1 << hashRateLog) - 1) << (maxBitsInMask - hashRateLog);
    } else {
        /* In this degenerate case we simply honor the hash rate. */
        state->stopMask = ((U64)1 << hashRateLog) - 1;
    }
}

/** ZSTD_ldm_gear_reset()
 * Feeds [data, data + minMatchLength) into the hash without registering any
 * splits. This effectively resets the hash state. This is used when skipping
 * over data, either at the beginning of a block, or skipping sections.
 */
static void ZSTD_ldm_gear_reset(ldmRollingHashState_t* state,
                                BYTE const* data, size_t minMatchLength)
{
    U64 hash = state->rolling;
    size_t n = 0;

#define GEAR_ITER_ONCE() do {                                  \
        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
        n += 1;                                                \
    } while (0)
    while (n + 3 < minMatchLength) {
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
    }
    while (n < minMatchLength) {
        GEAR_ITER_ONCE();
    }
#undef GEAR_ITER_ONCE
}

/** ZSTD_ldm_gear_feed():
 *
 * Registers in the splits array all the split points found in the first
 * size bytes following the data pointer. This function terminates when
 * either all the data has been processed or LDM_BATCH_SIZE splits are
 * present in the splits array.
 *
 * Precondition: The splits array must not be full.
 * Returns: The number of bytes processed. */
static size_t ZSTD_ldm_gear_feed(ldmRollingHashState_t* state,
                                 BYTE const* data, size_t size,
                                 size_t* splits, unsigned* numSplits)
{
    size_t n;
    U64 hash, mask;

    hash = state->rolling;
    mask = state->stopMask;
    n = 0;

#define GEAR_ITER_ONCE() do { \
        hash = (hash << 1) + ZSTD_ldm_gearTab[data[n] & 0xff]; \
        n += 1; \
        if (UNLIKELY((hash & mask) == 0)) { \
            splits[*numSplits] = n; \
            *numSplits += 1; \
            if (*numSplits == LDM_BATCH_SIZE) \
                goto done; \
        } \
    } while (0)

    while (n + 3 < size) {
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
        GEAR_ITER_ONCE();
    }
    while (n < size) {
        GEAR_ITER_ONCE();
    }

#undef GEAR_ITER_ONCE

done:
    state->rolling = hash;
    return n;
}

void ZSTD_ldm_adjustParameters(ldmParams_t* params,
                        const ZSTD_compressionParameters* cParams)
{
    params->windowLog = cParams->windowLog;
    ZSTD_STATIC_ASSERT(LDM_BUCKET_SIZE_LOG <= ZSTD_LDM_BUCKETSIZELOG_MAX);
    DEBUGLOG(4, "ZSTD_ldm_adjustParameters");
    if (params->hashRateLog == 0) {
        if (params->hashLog > 0) {
            /* if params->hashLog is set, derive hashRateLog from it */
            assert(params->hashLog <= ZSTD_HASHLOG_MAX);
            if (params->windowLog > params->hashLog) {
                params->hashRateLog = params->windowLog - params->hashLog;
            }
        } else {
            assert(1 <= (int)cParams->strategy && (int)cParams->strategy <= 9);
            /* mapping from [fast, rate7] to [btultra2, rate4] */
            params->hashRateLog = 7 - (cParams->strategy/3);
        }
    }
    if (params->hashLog == 0) {
        params->hashLog = BOUNDED(ZSTD_HASHLOG_MIN, params->windowLog - params->hashRateLog, ZSTD_HASHLOG_MAX);
    }
    if (params->minMatchLength == 0) {
        params->minMatchLength = LDM_MIN_MATCH_LENGTH;
        if (cParams->strategy >= ZSTD_btultra)
            params->minMatchLength /= 2;
    }
    if (params->bucketSizeLog==0) {
        assert(1 <= (int)cParams->strategy && (int)cParams->strategy <= 9);
        params->bucketSizeLog = BOUNDED(LDM_BUCKET_SIZE_LOG, (U32)cParams->strategy, ZSTD_LDM_BUCKETSIZELOG_MAX);
    }
    params->bucketSizeLog = MIN(params->bucketSizeLog, params->hashLog);
}

size_t ZSTD_ldm_getTableSize(ldmParams_t params)
{
    size_t const ldmHSize = ((size_t)1) << params.hashLog;
    size_t const ldmBucketSizeLog = MIN(params.bucketSizeLog, params.hashLog);
    size_t const ldmBucketSize = ((size_t)1) << (params.hashLog - ldmBucketSizeLog);
    size_t const totalSize = ZSTD_cwksp_alloc_size(ldmBucketSize)
                           + ZSTD_cwksp_alloc_size(ldmHSize * sizeof(ldmEntry_t));
    return params.enableLdm == ZSTD_ps_enable ? totalSize : 0;
}

size_t ZSTD_ldm_getMaxNbSeq(ldmParams_t params, size_t maxChunkSize)
{
    return params.enableLdm == ZSTD_ps_enable ? (maxChunkSize / params.minMatchLength) : 0;
}

/** ZSTD_ldm_getBucket() :
 *  Returns a pointer to the start of the bucket associated with hash. */
static ldmEntry_t* ZSTD_ldm_getBucket(
        const ldmState_t* ldmState, size_t hash, U32 const bucketSizeLog)
{
    return ldmState->hashTable + (hash << bucketSizeLog);
}

/** ZSTD_ldm_insertEntry() :
 *  Insert the entry with corresponding hash into the hash table */
static void ZSTD_ldm_insertEntry(ldmState_t* ldmState,
                                 size_t const hash, const ldmEntry_t entry,
                                 U32 const bucketSizeLog)
{
    BYTE* const pOffset = ldmState->bucketOffsets + hash;
    unsigned const offset = *pOffset;

    *(ZSTD_ldm_getBucket(ldmState, hash, bucketSizeLog) + offset) = entry;
    *pOffset = (BYTE)((offset + 1) & ((1u << bucketSizeLog) - 1));

}

/** ZSTD_ldm_countBackwardsMatch() :
 *  Returns the number of bytes that match backwards before pIn and pMatch.
 *
 *  We count only bytes where pMatch >= pBase and pIn >= pAnchor. */
static size_t ZSTD_ldm_countBackwardsMatch(
            const BYTE* pIn, const BYTE* pAnchor,
            const BYTE* pMatch, const BYTE* pMatchBase)
{
    size_t matchLength = 0;
    while (pIn > pAnchor && pMatch > pMatchBase && pIn[-1] == pMatch[-1]) {
        pIn--;
        pMatch--;
        matchLength++;
    }
    return matchLength;
}

/** ZSTD_ldm_countBackwardsMatch_2segments() :
 *  Returns the number of bytes that match backwards from pMatch,
 *  even with the backwards match spanning 2 different segments.
 *
 *  On reaching `pMatchBase`, start counting from mEnd */
static size_t ZSTD_ldm_countBackwardsMatch_2segments(
                    const BYTE* pIn, const BYTE* pAnchor,
                    const BYTE* pMatch, const BYTE* pMatchBase,
                    const BYTE* pExtDictStart, const BYTE* pExtDictEnd)
{
    size_t matchLength = ZSTD_ldm_countBackwardsMatch(pIn, pAnchor, pMatch, pMatchBase);
    if (pMatch - matchLength != pMatchBase || pMatchBase == pExtDictStart) {
        /* If backwards match is entirely in the extDict or prefix, immediately return */
        return matchLength;
    }
    DEBUGLOG(7, "ZSTD_ldm_countBackwardsMatch_2segments: found 2-parts backwards match (length in prefix==%zu)", matchLength);
    matchLength += ZSTD_ldm_countBackwardsMatch(pIn - matchLength, pAnchor, pExtDictEnd, pExtDictStart);
    DEBUGLOG(7, "final backwards match length = %zu", matchLength);
    return matchLength;
}

/** ZSTD_ldm_fillFastTables() :
 *
 *  Fills the relevant tables for the ZSTD_fast and ZSTD_dfast strategies.
 *  This is similar to ZSTD_loadDictionaryContent.
 *
 *  The tables for the other strategies are filled within their
 *  block compressors. */
static size_t ZSTD_ldm_fillFastTables(ZSTD_MatchState_t* ms,
                                      void const* end)
{
    const BYTE* const iend = (const BYTE*)end;

    switch(ms->cParams.strategy)
    {
    case ZSTD_fast:
        ZSTD_fillHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
        break;

    case ZSTD_dfast:
#ifndef ZSTD_EXCLUDE_DFAST_BLOCK_COMPRESSOR
        ZSTD_fillDoubleHashTable(ms, iend, ZSTD_dtlm_fast, ZSTD_tfp_forCCtx);
#else
        assert(0); /* shouldn't be called: cparams should've been adjusted. */
#endif
        break;

    case ZSTD_greedy:
    case ZSTD_lazy:
    case ZSTD_lazy2:
    case ZSTD_btlazy2:
    case ZSTD_btopt:
    case ZSTD_btultra:
    case ZSTD_btultra2:
        break;
    default:
        assert(0);  /* not possible : not a valid strategy id */
    }

    return 0;
}

void ZSTD_ldm_fillHashTable(
            ldmState_t* ldmState, const BYTE* ip,
            const BYTE* iend, ldmParams_t const* params)
{
    U32 const minMatchLength = params->minMatchLength;
    U32 const bucketSizeLog = params->bucketSizeLog;
    U32 const hBits = params->hashLog - bucketSizeLog;
    BYTE const* const base = ldmState->window.base;
    BYTE const* const istart = ip;
    ldmRollingHashState_t hashState;
    size_t* const splits = ldmState->splitIndices;
    unsigned numSplits;

    DEBUGLOG(5, "ZSTD_ldm_fillHashTable");

    ZSTD_ldm_gear_init(&hashState, params);
    while (ip < iend) {
        size_t hashed;
        unsigned n;

        numSplits = 0;
        hashed = ZSTD_ldm_gear_feed(&hashState, ip, (size_t)(iend - ip), splits, &numSplits);

        for (n = 0; n < numSplits; n++) {
            if (ip + splits[n] >= istart + minMatchLength) {
                BYTE const* const split = ip + splits[n] - minMatchLength;
                U64 const xxhash = XXH64(split, minMatchLength, 0);
                U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));
                ldmEntry_t entry;

                entry.offset = (U32)(split - base);
                entry.checksum = (U32)(xxhash >> 32);
                ZSTD_ldm_insertEntry(ldmState, hash, entry, params->bucketSizeLog);
            }
        }

        ip += hashed;
    }
}


/** ZSTD_ldm_limitTableUpdate() :
 *
 *  Sets cctx->nextToUpdate to a position corresponding closer to anchor
 *  if it is far way
 *  (after a long match, only update tables a limited amount). */
static void ZSTD_ldm_limitTableUpdate(ZSTD_MatchState_t* ms, const BYTE* anchor)
{
    U32 const curr = (U32)(anchor - ms->window.base);
    if (curr > ms->nextToUpdate + 1024) {
        ms->nextToUpdate =
            curr - MIN(512, curr - ms->nextToUpdate - 1024);
    }
}

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_ldm_generateSequences_internal(
        ldmState_t* ldmState, RawSeqStore_t* rawSeqStore,
        ldmParams_t const* params, void const* src, size_t srcSize)
{
    /* LDM parameters */
    int const extDict = ZSTD_window_hasExtDict(ldmState->window);
    U32 const minMatchLength = params->minMatchLength;
    U32 const entsPerBucket = 1U << params->bucketSizeLog;
    U32 const hBits = params->hashLog - params->bucketSizeLog;
    /* Prefix and extDict parameters */
    U32 const dictLimit = ldmState->window.dictLimit;
    U32 const lowestIndex = extDict ? ldmState->window.lowLimit : dictLimit;
    BYTE const* const base = ldmState->window.base;
    BYTE const* const dictBase = extDict ? ldmState->window.dictBase : NULL;
    BYTE const* const dictStart = extDict ? dictBase + lowestIndex : NULL;
    BYTE const* const dictEnd = extDict ? dictBase + dictLimit : NULL;
    BYTE const* const lowPrefixPtr = base + dictLimit;
    /* Input bounds */
    BYTE const* const istart = (BYTE const*)src;
    BYTE const* const iend = istart + srcSize;
    BYTE const* const ilimit = iend - HASH_READ_SIZE;
    /* Input positions */
    BYTE const* anchor = istart;
    BYTE const* ip = istart;
    /* Rolling hash state */
    ldmRollingHashState_t hashState;
    /* Arrays for staged-processing */
    size_t* const splits = ldmState->splitIndices;
    ldmMatchCandidate_t* const candidates = ldmState->matchCandidates;
    unsigned numSplits;

    if (srcSize < minMatchLength)
        return iend - anchor;

    /* Initialize the rolling hash state with the first minMatchLength bytes */
    ZSTD_ldm_gear_init(&hashState, params);
    ZSTD_ldm_gear_reset(&hashState, ip, minMatchLength);
    ip += minMatchLength;

    while (ip < ilimit) {
        size_t hashed;
        unsigned n;

        numSplits = 0;
        hashed = ZSTD_ldm_gear_feed(&hashState, ip, ilimit - ip,
                                    splits, &numSplits);

        for (n = 0; n < numSplits; n++) {
            BYTE const* const split = ip + splits[n] - minMatchLength;
            U64 const xxhash = XXH64(split, minMatchLength, 0);
            U32 const hash = (U32)(xxhash & (((U32)1 << hBits) - 1));

            candidates[n].split = split;
            candidates[n].hash = hash;
            candidates[n].checksum = (U32)(xxhash >> 32);
            candidates[n].bucket = ZSTD_ldm_getBucket(ldmState, hash, params->bucketSizeLog);
            PREFETCH_L1(candidates[n].bucket);
        }

        for (n = 0; n < numSplits; n++) {
            size_t forwardMatchLength = 0, backwardMatchLength = 0,
                   bestMatchLength = 0, mLength;
            U32 offset;
            BYTE const* const split = candidates[n].split;
            U32 const checksum = candidates[n].checksum;
            U32 const hash = candidates[n].hash;
            ldmEntry_t* const bucket = candidates[n].bucket;
            ldmEntry_t const* cur;
            ldmEntry_t const* bestEntry = NULL;
            ldmEntry_t newEntry;

            newEntry.offset = (U32)(split - base);
            newEntry.checksum = checksum;

            /* If a split point would generate a sequence overlapping with
             * the previous one, we merely register it in the hash table and
             * move on */
            if (split < anchor) {
                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);
                continue;
            }

            for (cur = bucket; cur < bucket + entsPerBucket; cur++) {
                size_t curForwardMatchLength, curBackwardMatchLength,
                       curTotalMatchLength;
                if (cur->checksum != checksum || cur->offset <= lowestIndex) {
                    continue;
                }
                if (extDict) {
                    BYTE const* const curMatchBase =
                        cur->offset < dictLimit ? dictBase : base;
                    BYTE const* const pMatch = curMatchBase + cur->offset;
                    BYTE const* const matchEnd =
                        cur->offset < dictLimit ? dictEnd : iend;
                    BYTE const* const lowMatchPtr =
                        cur->offset < dictLimit ? dictStart : lowPrefixPtr;
                    curForwardMatchLength =
                        ZSTD_count_2segments(split, pMatch, iend, matchEnd, lowPrefixPtr);
                    if (curForwardMatchLength < minMatchLength) {
                        continue;
                    }
                    curBackwardMatchLength = ZSTD_ldm_countBackwardsMatch_2segments(
                            split, anchor, pMatch, lowMatchPtr, dictStart, dictEnd);
                } else { /* !extDict */
                    BYTE const* const pMatch = base + cur->offset;
                    curForwardMatchLength = ZSTD_count(split, pMatch, iend);
                    if (curForwardMatchLength < minMatchLength) {
                        continue;
                    }
                    curBackwardMatchLength =
                        ZSTD_ldm_countBackwardsMatch(split, anchor, pMatch, lowPrefixPtr);
                }
                curTotalMatchLength = curForwardMatchLength + curBackwardMatchLength;

                if (curTotalMatchLength > bestMatchLength) {
                    bestMatchLength = curTotalMatchLength;
                    forwardMatchLength = curForwardMatchLength;
                    backwardMatchLength = curBackwardMatchLength;
                    bestEntry = cur;
                }
            }

            /* No match found -- insert an entry into the hash table
             * and process the next candidate match */
            if (bestEntry == NULL) {
                ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);
                continue;
            }

            /* Match found */
            offset = (U32)(split - base) - bestEntry->offset;
            mLength = forwardMatchLength + backwardMatchLength;
            {
                rawSeq* const seq = rawSeqStore->seq + rawSeqStore->size;

                /* Out of sequence storage */
                if (rawSeqStore->size == rawSeqStore->capacity)
                    return ERROR(dstSize_tooSmall);
                seq->litLength = (U32)(split - backwardMatchLength - anchor);
                seq->matchLength = (U32)mLength;
                seq->offset = offset;
                rawSeqStore->size++;
            }

            /* Insert the current entry into the hash table --- it must be
             * done after the previous block to avoid clobbering bestEntry */
            ZSTD_ldm_insertEntry(ldmState, hash, newEntry, params->bucketSizeLog);

            anchor = split + forwardMatchLength;

            /* If we find a match that ends after the data that we've hashed
             * then we have a repeating, overlapping, pattern. E.g. all zeros.
             * If one repetition of the pattern matches our `stopMask` then all
             * repetitions will. We don't need to insert them all into out table,
             * only the first one. So skip over overlapping matches.
             * This is a major speed boost (20x) for compressing a single byte
             * repeated, when that byte ends up in the table.
             */
            if (anchor > ip + hashed) {
                ZSTD_ldm_gear_reset(&hashState, anchor - minMatchLength, minMatchLength);
                /* Continue the outer loop at anchor (ip + hashed == anchor). */
                ip = anchor - hashed;
                break;
            }
        }

        ip += hashed;
    }

    return iend - anchor;
}

/*! ZSTD_ldm_reduceTable() :
 *  reduce table indexes by `reducerValue` */
static void ZSTD_ldm_reduceTable(ldmEntry_t* const table, U32 const size,
                                 U32 const reducerValue)
{
    U32 u;
    for (u = 0; u < size; u++) {
        if (table[u].offset < reducerValue) table[u].offset = 0;
        else table[u].offset -= reducerValue;
    }
}

size_t ZSTD_ldm_generateSequences(
        ldmState_t* ldmState, RawSeqStore_t* sequences,
        ldmParams_t const* params, void const* src, size_t srcSize)
{
    U32 const maxDist = 1U << params->windowLog;
    BYTE const* const istart = (BYTE const*)src;
    BYTE const* const iend = istart + srcSize;
    size_t const kMaxChunkSize = 1 << 20;
    size_t const nbChunks = (srcSize / kMaxChunkSize) + ((srcSize % kMaxChunkSize) != 0);
    size_t chunk;
    size_t leftoverSize = 0;

    assert(ZSTD_CHUNKSIZE_MAX >= kMaxChunkSize);
    /* Check that ZSTD_window_update() has been called for this chunk prior
     * to passing it to this function.
     */
    assert(ldmState->window.nextSrc >= (BYTE const*)src + srcSize);
    /* The input could be very large (in zstdmt), so it must be broken up into
     * chunks to enforce the maximum distance and handle overflow correction.
     */
    assert(sequences->pos <= sequences->size);
    assert(sequences->size <= sequences->capacity);
    for (chunk = 0; chunk < nbChunks && sequences->size < sequences->capacity; ++chunk) {
        BYTE const* const chunkStart = istart + chunk * kMaxChunkSize;
        size_t const remaining = (size_t)(iend - chunkStart);
        BYTE const *const chunkEnd =
            (remaining < kMaxChunkSize) ? iend : chunkStart + kMaxChunkSize;
        size_t const chunkSize = chunkEnd - chunkStart;
        size_t newLeftoverSize;
        size_t const prevSize = sequences->size;

        assert(chunkStart < iend);
        /* 1. Perform overflow correction if necessary. */
        if (ZSTD_window_needOverflowCorrection(ldmState->window, 0, maxDist, ldmState->loadedDictEnd, chunkStart, chunkEnd)) {
            U32 const ldmHSize = 1U << params->hashLog;
            U32 const correction = ZSTD_window_correctOverflow(
                &ldmState->window, /* cycleLog */ 0, maxDist, chunkStart);
            ZSTD_ldm_reduceTable(ldmState->hashTable, ldmHSize, correction);
            /* invalidate dictionaries on overflow correction */
            ldmState->loadedDictEnd = 0;
        }
        /* 2. We enforce the maximum offset allowed.
         *
         * kMaxChunkSize should be small enough that we don't lose too much of
         * the window through early invalidation.
         * TODO: * Test the chunk size.
         *       * Try invalidation after the sequence generation and test the
         *         offset against maxDist directly.
         *
         * NOTE: Because of dictionaries + sequence splitting we MUST make sure
         * that any offset used is valid at the END of the sequence, since it may
         * be split into two sequences. This condition holds when using
         * ZSTD_window_enforceMaxDist(), but if we move to checking offsets
         * against maxDist directly, we'll have to carefully handle that case.
         */
        ZSTD_window_enforceMaxDist(&ldmState->window, chunkEnd, maxDist, &ldmState->loadedDictEnd, NULL);
        /* 3. Generate the sequences for the chunk, and get newLeftoverSize. */
        newLeftoverSize = ZSTD_ldm_generateSequences_internal(
            ldmState, sequences, params, chunkStart, chunkSize);
        if (ZSTD_isError(newLeftoverSize))
            return newLeftoverSize;
        /* 4. We add the leftover literals from previous iterations to the first
         *    newly generated sequence, or add the `newLeftoverSize` if none are
         *    generated.
         */
        /* Prepend the leftover literals from the last call */
        if (prevSize < sequences->size) {
            sequences->seq[prevSize].litLength += (U32)leftoverSize;
            leftoverSize = newLeftoverSize;
        } else {
            assert(newLeftoverSize == chunkSize);
            leftoverSize += chunkSize;
        }
    }
    return 0;
}

void
ZSTD_ldm_skipSequences(RawSeqStore_t* rawSeqStore, size_t srcSize, U32 const minMatch)
{
    while (srcSize > 0 && rawSeqStore->pos < rawSeqStore->size) {
        rawSeq* seq = rawSeqStore->seq + rawSeqStore->pos;
        if (srcSize <= seq->litLength) {
            /* Skip past srcSize literals */
            seq->litLength -= (U32)srcSize;
            return;
        }
        srcSize -= seq->litLength;
        seq->litLength = 0;
        if (srcSize < seq->matchLength) {
            /* Skip past the first srcSize of the match */
            seq->matchLength -= (U32)srcSize;
            if (seq->matchLength < minMatch) {
                /* The match is too short, omit it */
                if (rawSeqStore->pos + 1 < rawSeqStore->size) {
                    seq[1].litLength += seq[0].matchLength;
                }
                rawSeqStore->pos++;
            }
            return;
        }
        srcSize -= seq->matchLength;
        seq->matchLength = 0;
        rawSeqStore->pos++;
    }
}

/**
 * If the sequence length is longer than remaining then the sequence is split
 * between this block and the next.
 *
 * Returns the current sequence to handle, or if the rest of the block should
 * be literals, it returns a sequence with offset == 0.
 */
static rawSeq maybeSplitSequence(RawSeqStore_t* rawSeqStore,
                                 U32 const remaining, U32 const minMatch)
{
    rawSeq sequence = rawSeqStore->seq[rawSeqStore->pos];
    assert(sequence.offset > 0);
    /* Likely: No partial sequence */
    if (remaining >= sequence.litLength + sequence.matchLength) {
        rawSeqStore->pos++;
        return sequence;
    }
    /* Cut the sequence short (offset == 0 ==> rest is literals). */
    if (remaining <= sequence.litLength) {
        sequence.offset = 0;
    } else if (remaining < sequence.litLength + sequence.matchLength) {
        sequence.matchLength = remaining - sequence.litLength;
        if (sequence.matchLength < minMatch) {
            sequence.offset = 0;
        }
    }
    /* Skip past `remaining` bytes for the future sequences. */
    ZSTD_ldm_skipSequences(rawSeqStore, remaining, minMatch);
    return sequence;
}

void ZSTD_ldm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes) {
    U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
    while (currPos && rawSeqStore->pos < rawSeqStore->size) {
        rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
        if (currPos >= currSeq.litLength + currSeq.matchLength) {
            currPos -= currSeq.litLength + currSeq.matchLength;
            rawSeqStore->pos++;
        } else {
            rawSeqStore->posInSequence = currPos;
            break;
        }
    }
    if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
        rawSeqStore->posInSequence = 0;
    }
}

size_t ZSTD_ldm_blockCompress(RawSeqStore_t* rawSeqStore,
    ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
    ZSTD_ParamSwitch_e useRowMatchFinder,
    void const* src, size_t srcSize)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    unsigned const minMatch = cParams->minMatch;
    ZSTD_BlockCompressor_f const blockCompressor =
        ZSTD_selectBlockCompressor(cParams->strategy, useRowMatchFinder, ZSTD_matchState_dictMode(ms));
    /* Input bounds */
    BYTE const* const istart = (BYTE const*)src;
    BYTE const* const iend = istart + srcSize;
    /* Input positions */
    BYTE const* ip = istart;

    DEBUGLOG(5, "ZSTD_ldm_blockCompress: srcSize=%zu", srcSize);
    /* If using opt parser, use LDMs only as candidates rather than always accepting them */
    if (cParams->strategy >= ZSTD_btopt) {
        size_t lastLLSize;
        ms->ldmSeqStore = rawSeqStore;
        lastLLSize = blockCompressor(ms, seqStore, rep, src, srcSize);
        ZSTD_ldm_skipRawSeqStoreBytes(rawSeqStore, srcSize);
        return lastLLSize;
    }

    assert(rawSeqStore->pos <= rawSeqStore->size);
    assert(rawSeqStore->size <= rawSeqStore->capacity);
    /* Loop through each sequence and apply the block compressor to the literals */
    while (rawSeqStore->pos < rawSeqStore->size && ip < iend) {
        /* maybeSplitSequence updates rawSeqStore->pos */
        rawSeq const sequence = maybeSplitSequence(rawSeqStore,
                                                   (U32)(iend - ip), minMatch);
        /* End signal */
        if (sequence.offset == 0)
            break;

        assert(ip + sequence.litLength + sequence.matchLength <= iend);

        /* Fill tables for block compressor */
        ZSTD_ldm_limitTableUpdate(ms, ip);
        ZSTD_ldm_fillFastTables(ms, ip);
        /* Run the block compressor */
        DEBUGLOG(5, "pos %u : calling block compressor on segment of size %u", (unsigned)(ip-istart), sequence.litLength);
        {
            int i;
            size_t const newLitLength =
                blockCompressor(ms, seqStore, rep, ip, sequence.litLength);
            ip += sequence.litLength;
            /* Update the repcodes */
            for (i = ZSTD_REP_NUM - 1; i > 0; i--)
                rep[i] = rep[i-1];
            rep[0] = sequence.offset;
            /* Store the sequence */
            ZSTD_storeSeq(seqStore, newLitLength, ip - newLitLength, iend,
                          OFFSET_TO_OFFBASE(sequence.offset),
                          sequence.matchLength);
            ip += sequence.matchLength;
        }
    }
    /* Fill the tables for the block compressor */
    ZSTD_ldm_limitTableUpdate(ms, ip);
    ZSTD_ldm_fillFastTables(ms, ip);
    /* Compress the last literals */
    return blockCompressor(ms, seqStore, rep, ip, iend - ip);
}
/**** ended inlining compress/zstd_ldm.c ****/
/**** start inlining compress/zstd_opt.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: hist.h ****/
/**** skipping file: zstd_opt.h ****/

#if !defined(ZSTD_EXCLUDE_BTLAZY2_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR) \
 || !defined(ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR)

#define ZSTD_LITFREQ_ADD    2   /* scaling factor for litFreq, so that frequencies adapt faster to new stats */
#define ZSTD_MAX_PRICE     (1<<30)

#define ZSTD_PREDEF_THRESHOLD 8   /* if srcSize < ZSTD_PREDEF_THRESHOLD, symbols' cost is assumed static, directly determined by pre-defined distributions */


/*-*************************************
*  Price functions for optimal parser
***************************************/

#if 0    /* approximation at bit level (for tests) */
#  define BITCOST_ACCURACY 0
#  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
#  define WEIGHT(stat, opt) ((void)(opt), ZSTD_bitWeight(stat))
#elif 0  /* fractional bit accuracy (for tests) */
#  define BITCOST_ACCURACY 8
#  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
#  define WEIGHT(stat,opt) ((void)(opt), ZSTD_fracWeight(stat))
#else    /* opt==approx, ultra==accurate */
#  define BITCOST_ACCURACY 8
#  define BITCOST_MULTIPLIER (1 << BITCOST_ACCURACY)
#  define WEIGHT(stat,opt) ((opt) ? ZSTD_fracWeight(stat) : ZSTD_bitWeight(stat))
#endif

/* ZSTD_bitWeight() :
 * provide estimated "cost" of a stat in full bits only */
MEM_STATIC U32 ZSTD_bitWeight(U32 stat)
{
    return (ZSTD_highbit32(stat+1) * BITCOST_MULTIPLIER);
}

/* ZSTD_fracWeight() :
 * provide fractional-bit "cost" of a stat,
 * using linear interpolation approximation */
MEM_STATIC U32 ZSTD_fracWeight(U32 rawStat)
{
    U32 const stat = rawStat + 1;
    U32 const hb = ZSTD_highbit32(stat);
    U32 const BWeight = hb * BITCOST_MULTIPLIER;
    /* Fweight was meant for "Fractional weight"
     * but it's effectively a value between 1 and 2
     * using fixed point arithmetic */
    U32 const FWeight = (stat << BITCOST_ACCURACY) >> hb;
    U32 const weight = BWeight + FWeight;
    assert(hb + BITCOST_ACCURACY < 31);
    return weight;
}

#if (DEBUGLEVEL>=2)
/* debugging function,
 * @return price in bytes as fractional value
 * for debug messages only */
MEM_STATIC double ZSTD_fCost(int price)
{
    return (double)price / (BITCOST_MULTIPLIER*8);
}
#endif

static int ZSTD_compressedLiterals(optState_t const* const optPtr)
{
    return optPtr->literalCompressionMode != ZSTD_ps_disable;
}

static void ZSTD_setBasePrices(optState_t* optPtr, int optLevel)
{
    if (ZSTD_compressedLiterals(optPtr))
        optPtr->litSumBasePrice = WEIGHT(optPtr->litSum, optLevel);
    optPtr->litLengthSumBasePrice = WEIGHT(optPtr->litLengthSum, optLevel);
    optPtr->matchLengthSumBasePrice = WEIGHT(optPtr->matchLengthSum, optLevel);
    optPtr->offCodeSumBasePrice = WEIGHT(optPtr->offCodeSum, optLevel);
}


static U32 sum_u32(const unsigned table[], size_t nbElts)
{
    size_t n;
    U32 total = 0;
    for (n=0; n<nbElts; n++) {
        total += table[n];
    }
    return total;
}

typedef enum { base_0possible=0, base_1guaranteed=1 } base_directive_e;

static U32
ZSTD_downscaleStats(unsigned* table, U32 lastEltIndex, U32 shift, base_directive_e base1)
{
    U32 s, sum=0;
    DEBUGLOG(5, "ZSTD_downscaleStats (nbElts=%u, shift=%u)",
            (unsigned)lastEltIndex+1, (unsigned)shift );
    assert(shift < 30);
    for (s=0; s<lastEltIndex+1; s++) {
        unsigned const base = base1 ? 1 : (table[s]>0);
        unsigned const newStat = base + (table[s] >> shift);
        sum += newStat;
        table[s] = newStat;
    }
    return sum;
}

/* ZSTD_scaleStats() :
 * reduce all elt frequencies in table if sum too large
 * return the resulting sum of elements */
static U32 ZSTD_scaleStats(unsigned* table, U32 lastEltIndex, U32 logTarget)
{
    U32 const prevsum = sum_u32(table, lastEltIndex+1);
    U32 const factor = prevsum >> logTarget;
    DEBUGLOG(5, "ZSTD_scaleStats (nbElts=%u, target=%u)", (unsigned)lastEltIndex+1, (unsigned)logTarget);
    assert(logTarget < 30);
    if (factor <= 1) return prevsum;
    return ZSTD_downscaleStats(table, lastEltIndex, ZSTD_highbit32(factor), base_1guaranteed);
}

/* ZSTD_rescaleFreqs() :
 * if first block (detected by optPtr->litLengthSum == 0) : init statistics
 *    take hints from dictionary if there is one
 *    and init from zero if there is none,
 *    using src for literals stats, and baseline stats for sequence symbols
 * otherwise downscale existing stats, to be used as seed for next block.
 */
static void
ZSTD_rescaleFreqs(optState_t* const optPtr,
            const BYTE* const src, size_t const srcSize,
                  int const optLevel)
{
    int const compressedLiterals = ZSTD_compressedLiterals(optPtr);
    DEBUGLOG(5, "ZSTD_rescaleFreqs (srcSize=%u)", (unsigned)srcSize);
    optPtr->priceType = zop_dynamic;

    if (optPtr->litLengthSum == 0) {  /* no literals stats collected -> first block assumed -> init */

        /* heuristic: use pre-defined stats for too small inputs */
        if (srcSize <= ZSTD_PREDEF_THRESHOLD) {
            DEBUGLOG(5, "srcSize <= %i : use predefined stats", ZSTD_PREDEF_THRESHOLD);
            optPtr->priceType = zop_predef;
        }

        assert(optPtr->symbolCosts != NULL);
        if (optPtr->symbolCosts->huf.repeatMode == HUF_repeat_valid) {

            /* huffman stats covering the full value set : table presumed generated by dictionary */
            optPtr->priceType = zop_dynamic;

            if (compressedLiterals) {
                /* generate literals statistics from huffman table */
                unsigned lit;
                assert(optPtr->litFreq != NULL);
                optPtr->litSum = 0;
                for (lit=0; lit<=MaxLit; lit++) {
                    U32 const scaleLog = 11;   /* scale to 2K */
                    U32 const bitCost = HUF_getNbBitsFromCTable(optPtr->symbolCosts->huf.CTable, lit);
                    assert(bitCost <= scaleLog);
                    optPtr->litFreq[lit] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                    optPtr->litSum += optPtr->litFreq[lit];
            }   }

            {   unsigned ll;
                FSE_CState_t llstate;
                FSE_initCState(&llstate, optPtr->symbolCosts->fse.litlengthCTable);
                optPtr->litLengthSum = 0;
                for (ll=0; ll<=MaxLL; ll++) {
                    U32 const scaleLog = 10;   /* scale to 1K */
                    U32 const bitCost = FSE_getMaxNbBits(llstate.symbolTT, ll);
                    assert(bitCost < scaleLog);
                    optPtr->litLengthFreq[ll] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                    optPtr->litLengthSum += optPtr->litLengthFreq[ll];
            }   }

            {   unsigned ml;
                FSE_CState_t mlstate;
                FSE_initCState(&mlstate, optPtr->symbolCosts->fse.matchlengthCTable);
                optPtr->matchLengthSum = 0;
                for (ml=0; ml<=MaxML; ml++) {
                    U32 const scaleLog = 10;
                    U32 const bitCost = FSE_getMaxNbBits(mlstate.symbolTT, ml);
                    assert(bitCost < scaleLog);
                    optPtr->matchLengthFreq[ml] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                    optPtr->matchLengthSum += optPtr->matchLengthFreq[ml];
            }   }

            {   unsigned of;
                FSE_CState_t ofstate;
                FSE_initCState(&ofstate, optPtr->symbolCosts->fse.offcodeCTable);
                optPtr->offCodeSum = 0;
                for (of=0; of<=MaxOff; of++) {
                    U32 const scaleLog = 10;
                    U32 const bitCost = FSE_getMaxNbBits(ofstate.symbolTT, of);
                    assert(bitCost < scaleLog);
                    optPtr->offCodeFreq[of] = bitCost ? 1 << (scaleLog-bitCost) : 1 /*minimum to calculate cost*/;
                    optPtr->offCodeSum += optPtr->offCodeFreq[of];
            }   }

        } else {  /* first block, no dictionary */

            assert(optPtr->litFreq != NULL);
            if (compressedLiterals) {
                /* base initial cost of literals on direct frequency within src */
                unsigned lit = MaxLit;
                HIST_count_simple(optPtr->litFreq, &lit, src, srcSize);   /* use raw first block to init statistics */
                optPtr->litSum = ZSTD_downscaleStats(optPtr->litFreq, MaxLit, 8, base_0possible);
            }

            {   unsigned const baseLLfreqs[MaxLL+1] = {
                    4, 2, 1, 1, 1, 1, 1, 1,
                    1, 1, 1, 1, 1, 1, 1, 1,
                    1, 1, 1, 1, 1, 1, 1, 1,
                    1, 1, 1, 1, 1, 1, 1, 1,
                    1, 1, 1, 1
                };
                ZSTD_memcpy(optPtr->litLengthFreq, baseLLfreqs, sizeof(baseLLfreqs));
                optPtr->litLengthSum = sum_u32(baseLLfreqs, MaxLL+1);
            }

            {   unsigned ml;
                for (ml=0; ml<=MaxML; ml++)
                    optPtr->matchLengthFreq[ml] = 1;
            }
            optPtr->matchLengthSum = MaxML+1;

            {   unsigned const baseOFCfreqs[MaxOff+1] = {
                    6, 2, 1, 1, 2, 3, 4, 4,
                    4, 3, 2, 1, 1, 1, 1, 1,
                    1, 1, 1, 1, 1, 1, 1, 1,
                    1, 1, 1, 1, 1, 1, 1, 1
                };
                ZSTD_memcpy(optPtr->offCodeFreq, baseOFCfreqs, sizeof(baseOFCfreqs));
                optPtr->offCodeSum = sum_u32(baseOFCfreqs, MaxOff+1);
            }

        }

    } else {   /* new block : scale down accumulated statistics */

        if (compressedLiterals)
            optPtr->litSum = ZSTD_scaleStats(optPtr->litFreq, MaxLit, 12);
        optPtr->litLengthSum = ZSTD_scaleStats(optPtr->litLengthFreq, MaxLL, 11);
        optPtr->matchLengthSum = ZSTD_scaleStats(optPtr->matchLengthFreq, MaxML, 11);
        optPtr->offCodeSum = ZSTD_scaleStats(optPtr->offCodeFreq, MaxOff, 11);
    }

    ZSTD_setBasePrices(optPtr, optLevel);
}

/* ZSTD_rawLiteralsCost() :
 * price of literals (only) in specified segment (which length can be 0).
 * does not include price of literalLength symbol */
static U32 ZSTD_rawLiteralsCost(const BYTE* const literals, U32 const litLength,
                                const optState_t* const optPtr,
                                int optLevel)
{
    DEBUGLOG(8, "ZSTD_rawLiteralsCost (%u literals)", litLength);
    if (litLength == 0) return 0;

    if (!ZSTD_compressedLiterals(optPtr))
        return (litLength << 3) * BITCOST_MULTIPLIER;  /* Uncompressed - 8 bytes per literal. */

    if (optPtr->priceType == zop_predef)
        return (litLength*6) * BITCOST_MULTIPLIER;  /* 6 bit per literal - no statistic used */

    /* dynamic statistics */
    {   U32 price = optPtr->litSumBasePrice * litLength;
        U32 const litPriceMax = optPtr->litSumBasePrice - BITCOST_MULTIPLIER;
        U32 u;
        assert(optPtr->litSumBasePrice >= BITCOST_MULTIPLIER);
        for (u=0; u < litLength; u++) {
            U32 litPrice = WEIGHT(optPtr->litFreq[literals[u]], optLevel);
            if (UNLIKELY(litPrice > litPriceMax)) litPrice = litPriceMax;
            price -= litPrice;
        }
        return price;
    }
}

/* ZSTD_litLengthPrice() :
 * cost of literalLength symbol */
static U32 ZSTD_litLengthPrice(U32 const litLength, const optState_t* const optPtr, int optLevel)
{
    assert(litLength <= ZSTD_BLOCKSIZE_MAX);
    if (optPtr->priceType == zop_predef)
        return WEIGHT(litLength, optLevel);

    /* ZSTD_LLcode() can't compute litLength price for sizes >= ZSTD_BLOCKSIZE_MAX
     * because it isn't representable in the zstd format.
     * So instead just pretend it would cost 1 bit more than ZSTD_BLOCKSIZE_MAX - 1.
     * In such a case, the block would be all literals.
     */
    if (litLength == ZSTD_BLOCKSIZE_MAX)
        return BITCOST_MULTIPLIER + ZSTD_litLengthPrice(ZSTD_BLOCKSIZE_MAX - 1, optPtr, optLevel);

    /* dynamic statistics */
    {   U32 const llCode = ZSTD_LLcode(litLength);
        return (LL_bits[llCode] * BITCOST_MULTIPLIER)
             + optPtr->litLengthSumBasePrice
             - WEIGHT(optPtr->litLengthFreq[llCode], optLevel);
    }
}

/* ZSTD_getMatchPrice() :
 * Provides the cost of the match part (offset + matchLength) of a sequence.
 * Must be combined with ZSTD_fullLiteralsCost() to get the full cost of a sequence.
 * @offBase : sumtype, representing an offset or a repcode, and using numeric representation of ZSTD_storeSeq()
 * @optLevel: when <2, favors small offset for decompression speed (improved cache efficiency)
 */
FORCE_INLINE_TEMPLATE U32
ZSTD_getMatchPrice(U32 const offBase,
                   U32 const matchLength,
             const optState_t* const optPtr,
                   int const optLevel)
{
    U32 price;
    U32 const offCode = ZSTD_highbit32(offBase);
    U32 const mlBase = matchLength - MINMATCH;
    assert(matchLength >= MINMATCH);

    if (optPtr->priceType == zop_predef)  /* fixed scheme, does not use statistics */
        return WEIGHT(mlBase, optLevel)
             + ((16 + offCode) * BITCOST_MULTIPLIER); /* emulated offset cost */

    /* dynamic statistics */
    price = (offCode * BITCOST_MULTIPLIER) + (optPtr->offCodeSumBasePrice - WEIGHT(optPtr->offCodeFreq[offCode], optLevel));
    if ((optLevel<2) /*static*/ && offCode >= 20)
        price += (offCode-19)*2 * BITCOST_MULTIPLIER; /* handicap for long distance offsets, favor decompression speed */

    /* match Length */
    {   U32 const mlCode = ZSTD_MLcode(mlBase);
        price += (ML_bits[mlCode] * BITCOST_MULTIPLIER) + (optPtr->matchLengthSumBasePrice - WEIGHT(optPtr->matchLengthFreq[mlCode], optLevel));
    }

    price += BITCOST_MULTIPLIER / 5;   /* heuristic : make matches a bit more costly to favor less sequences -> faster decompression speed */

    DEBUGLOG(8, "ZSTD_getMatchPrice(ml:%u) = %u", matchLength, price);
    return price;
}

/* ZSTD_updateStats() :
 * assumption : literals + litLength <= iend */
static void ZSTD_updateStats(optState_t* const optPtr,
                             U32 litLength, const BYTE* literals,
                             U32 offBase, U32 matchLength)
{
    /* literals */
    if (ZSTD_compressedLiterals(optPtr)) {
        U32 u;
        for (u=0; u < litLength; u++)
            optPtr->litFreq[literals[u]] += ZSTD_LITFREQ_ADD;
        optPtr->litSum += litLength*ZSTD_LITFREQ_ADD;
    }

    /* literal Length */
    {   U32 const llCode = ZSTD_LLcode(litLength);
        optPtr->litLengthFreq[llCode]++;
        optPtr->litLengthSum++;
    }

    /* offset code : follows storeSeq() numeric representation */
    {   U32 const offCode = ZSTD_highbit32(offBase);
        assert(offCode <= MaxOff);
        optPtr->offCodeFreq[offCode]++;
        optPtr->offCodeSum++;
    }

    /* match Length */
    {   U32 const mlBase = matchLength - MINMATCH;
        U32 const mlCode = ZSTD_MLcode(mlBase);
        optPtr->matchLengthFreq[mlCode]++;
        optPtr->matchLengthSum++;
    }
}


/* ZSTD_readMINMATCH() :
 * function safe only for comparisons
 * assumption : memPtr must be at least 4 bytes before end of buffer */
MEM_STATIC U32 ZSTD_readMINMATCH(const void* memPtr, U32 length)
{
    switch (length)
    {
    default :
    case 4 : return MEM_read32(memPtr);
    case 3 : if (MEM_isLittleEndian())
                return MEM_read32(memPtr)<<8;
             else
                return MEM_read32(memPtr)>>8;
    }
}


/* Update hashTable3 up to ip (excluded)
   Assumption : always within prefix (i.e. not within extDict) */
static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_insertAndFindFirstIndexHash3 (const ZSTD_MatchState_t* ms,
                                       U32* nextToUpdate3,
                                       const BYTE* const ip)
{
    U32* const hashTable3 = ms->hashTable3;
    U32 const hashLog3 = ms->hashLog3;
    const BYTE* const base = ms->window.base;
    U32 idx = *nextToUpdate3;
    U32 const target = (U32)(ip - base);
    size_t const hash3 = ZSTD_hash3Ptr(ip, hashLog3);
    assert(hashLog3 > 0);

    while(idx < target) {
        hashTable3[ZSTD_hash3Ptr(base+idx, hashLog3)] = idx;
        idx++;
    }

    *nextToUpdate3 = target;
    return hashTable3[hash3];
}


/*-*************************************
*  Binary Tree search
***************************************/
/** ZSTD_insertBt1() : add one or multiple positions to tree.
 * @param ip assumed <= iend-8 .
 * @param target The target of ZSTD_updateTree_internal() - we are filling to this position
 * @return : nb of positions added */
static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_insertBt1(
                const ZSTD_MatchState_t* ms,
                const BYTE* const ip, const BYTE* const iend,
                U32 const target,
                U32 const mls, const int extDict)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32*   const hashTable = ms->hashTable;
    U32    const hashLog = cParams->hashLog;
    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
    U32*   const bt = ms->chainTable;
    U32    const btLog  = cParams->chainLog - 1;
    U32    const btMask = (1 << btLog) - 1;
    U32 matchIndex = hashTable[h];
    size_t commonLengthSmaller=0, commonLengthLarger=0;
    const BYTE* const base = ms->window.base;
    const BYTE* const dictBase = ms->window.dictBase;
    const U32 dictLimit = ms->window.dictLimit;
    const BYTE* const dictEnd = dictBase + dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    const BYTE* match;
    const U32 curr = (U32)(ip-base);
    const U32 btLow = btMask >= curr ? 0 : curr - btMask;
    U32* smallerPtr = bt + 2*(curr&btMask);
    U32* largerPtr  = smallerPtr + 1;
    U32 dummy32;   /* to be nullified at the end */
    /* windowLow is based on target because
     * we only need positions that will be in the window at the end of the tree update.
     */
    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, target, cParams->windowLog);
    U32 matchEndIdx = curr+8+1;
    size_t bestLength = 8;
    U32 nbCompares = 1U << cParams->searchLog;
#ifdef ZSTD_C_PREDICT
    U32 predictedSmall = *(bt + 2*((curr-1)&btMask) + 0);
    U32 predictedLarge = *(bt + 2*((curr-1)&btMask) + 1);
    predictedSmall += (predictedSmall>0);
    predictedLarge += (predictedLarge>0);
#endif /* ZSTD_C_PREDICT */

    DEBUGLOG(8, "ZSTD_insertBt1 (%u)", curr);

    assert(curr <= target);
    assert(ip <= iend-8);   /* required for h calculation */
    hashTable[h] = curr;   /* Update Hash Table */

    assert(windowLow > 0);
    for (; nbCompares && (matchIndex >= windowLow); --nbCompares) {
        U32* const nextPtr = bt + 2*(matchIndex & btMask);
        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
        assert(matchIndex < curr);

#ifdef ZSTD_C_PREDICT   /* note : can create issues when hlog small <= 11 */
        const U32* predictPtr = bt + 2*((matchIndex-1) & btMask);   /* written this way, as bt is a roll buffer */
        if (matchIndex == predictedSmall) {
            /* no need to check length, result known */
            *smallerPtr = matchIndex;
            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
            smallerPtr = nextPtr+1;               /* new "smaller" => larger of match */
            matchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
            predictedSmall = predictPtr[1] + (predictPtr[1]>0);
            continue;
        }
        if (matchIndex == predictedLarge) {
            *largerPtr = matchIndex;
            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
            largerPtr = nextPtr;
            matchIndex = nextPtr[0];
            predictedLarge = predictPtr[0] + (predictPtr[0]>0);
            continue;
        }
#endif

        if (!extDict || (matchIndex+matchLength >= dictLimit)) {
            assert(matchIndex+matchLength >= dictLimit);   /* might be wrong if actually extDict */
            match = base + matchIndex;
            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
        } else {
            match = dictBase + matchIndex;
            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
            if (matchIndex+matchLength >= dictLimit)
                match = base + matchIndex;   /* to prepare for next usage of match[matchLength] */
        }

        if (matchLength > bestLength) {
            bestLength = matchLength;
            if (matchLength > matchEndIdx - matchIndex)
                matchEndIdx = matchIndex + (U32)matchLength;
        }

        if (ip+matchLength == iend) {   /* equal : no way to know if inf or sup */
            break;   /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
        }

        if (match[matchLength] < ip[matchLength]) {  /* necessarily within buffer */
            /* match is smaller than current */
            *smallerPtr = matchIndex;             /* update smaller idx */
            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
            smallerPtr = nextPtr+1;               /* new "candidate" => larger than match, which was smaller than target */
            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous and closer to current */
        } else {
            /* match is larger than current */
            *largerPtr = matchIndex;
            commonLengthLarger = matchLength;
            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop searching */
            largerPtr = nextPtr;
            matchIndex = nextPtr[0];
    }   }

    *smallerPtr = *largerPtr = 0;
    {   U32 positions = 0;
        if (bestLength > 384) positions = MIN(192, (U32)(bestLength - 384));   /* speed optimization */
        assert(matchEndIdx > curr + 8);
        return MAX(positions, matchEndIdx - (curr + 8));
    }
}

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_updateTree_internal(
                ZSTD_MatchState_t* ms,
                const BYTE* const ip, const BYTE* const iend,
                const U32 mls, const ZSTD_dictMode_e dictMode)
{
    const BYTE* const base = ms->window.base;
    U32 const target = (U32)(ip - base);
    U32 idx = ms->nextToUpdate;
    DEBUGLOG(7, "ZSTD_updateTree_internal, from %u to %u  (dictMode:%u)",
                idx, target, dictMode);

    while(idx < target) {
        U32 const forward = ZSTD_insertBt1(ms, base+idx, iend, target, mls, dictMode == ZSTD_extDict);
        assert(idx < (U32)(idx + forward));
        idx += forward;
    }
    assert((size_t)(ip - base) <= (size_t)(U32)(-1));
    assert((size_t)(iend - base) <= (size_t)(U32)(-1));
    ms->nextToUpdate = target;
}

void ZSTD_updateTree(ZSTD_MatchState_t* ms, const BYTE* ip, const BYTE* iend) {
    ZSTD_updateTree_internal(ms, ip, iend, ms->cParams.minMatch, ZSTD_noDict);
}

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32
ZSTD_insertBtAndGetAllMatches (
                ZSTD_match_t* matches,  /* store result (found matches) in this table (presumed large enough) */
                ZSTD_MatchState_t* ms,
                U32* nextToUpdate3,
                const BYTE* const ip, const BYTE* const iLimit,
                const ZSTD_dictMode_e dictMode,
                const U32 rep[ZSTD_REP_NUM],
                const U32 ll0,  /* tells if associated literal length is 0 or not. This value must be 0 or 1 */
                const U32 lengthToBeat,
                const U32 mls /* template */)
{
    const ZSTD_compressionParameters* const cParams = &ms->cParams;
    U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
    const BYTE* const base = ms->window.base;
    U32 const curr = (U32)(ip-base);
    U32 const hashLog = cParams->hashLog;
    U32 const minMatch = (mls==3) ? 3 : 4;
    U32* const hashTable = ms->hashTable;
    size_t const h  = ZSTD_hashPtr(ip, hashLog, mls);
    U32 matchIndex  = hashTable[h];
    U32* const bt   = ms->chainTable;
    U32 const btLog = cParams->chainLog - 1;
    U32 const btMask= (1U << btLog) - 1;
    size_t commonLengthSmaller=0, commonLengthLarger=0;
    const BYTE* const dictBase = ms->window.dictBase;
    U32 const dictLimit = ms->window.dictLimit;
    const BYTE* const dictEnd = dictBase + dictLimit;
    const BYTE* const prefixStart = base + dictLimit;
    U32 const btLow = (btMask >= curr) ? 0 : curr - btMask;
    U32 const windowLow = ZSTD_getLowestMatchIndex(ms, curr, cParams->windowLog);
    U32 const matchLow = windowLow ? windowLow : 1;
    U32* smallerPtr = bt + 2*(curr&btMask);
    U32* largerPtr  = bt + 2*(curr&btMask) + 1;
    U32 matchEndIdx = curr+8+1;   /* farthest referenced position of any match => detects repetitive patterns */
    U32 dummy32;   /* to be nullified at the end */
    U32 mnum = 0;
    U32 nbCompares = 1U << cParams->searchLog;

    const ZSTD_MatchState_t* dms    = dictMode == ZSTD_dictMatchState ? ms->dictMatchState : NULL;
    const ZSTD_compressionParameters* const dmsCParams =
                                      dictMode == ZSTD_dictMatchState ? &dms->cParams : NULL;
    const BYTE* const dmsBase       = dictMode == ZSTD_dictMatchState ? dms->window.base : NULL;
    const BYTE* const dmsEnd        = dictMode == ZSTD_dictMatchState ? dms->window.nextSrc : NULL;
    U32         const dmsHighLimit  = dictMode == ZSTD_dictMatchState ? (U32)(dmsEnd - dmsBase) : 0;
    U32         const dmsLowLimit   = dictMode == ZSTD_dictMatchState ? dms->window.lowLimit : 0;
    U32         const dmsIndexDelta = dictMode == ZSTD_dictMatchState ? windowLow - dmsHighLimit : 0;
    U32         const dmsHashLog    = dictMode == ZSTD_dictMatchState ? dmsCParams->hashLog : hashLog;
    U32         const dmsBtLog      = dictMode == ZSTD_dictMatchState ? dmsCParams->chainLog - 1 : btLog;
    U32         const dmsBtMask     = dictMode == ZSTD_dictMatchState ? (1U << dmsBtLog) - 1 : 0;
    U32         const dmsBtLow      = dictMode == ZSTD_dictMatchState && dmsBtMask < dmsHighLimit - dmsLowLimit ? dmsHighLimit - dmsBtMask : dmsLowLimit;

    size_t bestLength = lengthToBeat-1;
    DEBUGLOG(8, "ZSTD_insertBtAndGetAllMatches: current=%u", curr);

    /* check repCode */
    assert(ll0 <= 1);   /* necessarily 1 or 0 */
    {   U32 const lastR = ZSTD_REP_NUM + ll0;
        U32 repCode;
        for (repCode = ll0; repCode < lastR; repCode++) {
            U32 const repOffset = (repCode==ZSTD_REP_NUM) ? (rep[0] - 1) : rep[repCode];
            U32 const repIndex = curr - repOffset;
            U32 repLen = 0;
            assert(curr >= dictLimit);
            if (repOffset-1 /* intentional overflow, discards 0 and -1 */ < curr-dictLimit) {  /* equivalent to `curr > repIndex >= dictLimit` */
                /* We must validate the repcode offset because when we're using a dictionary the
                 * valid offset range shrinks when the dictionary goes out of bounds.
                 */
                if ((repIndex >= windowLow) & (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(ip - repOffset, minMatch))) {
                    repLen = (U32)ZSTD_count(ip+minMatch, ip+minMatch-repOffset, iLimit) + minMatch;
                }
            } else {  /* repIndex < dictLimit || repIndex >= curr */
                const BYTE* const repMatch = dictMode == ZSTD_dictMatchState ?
                                             dmsBase + repIndex - dmsIndexDelta :
                                             dictBase + repIndex;
                assert(curr >= windowLow);
                if ( dictMode == ZSTD_extDict
                  && ( ((repOffset-1) /*intentional overflow*/ < curr - windowLow)  /* equivalent to `curr > repIndex >= windowLow` */
                     & (ZSTD_index_overlap_check(dictLimit, repIndex)) )
                  && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                    repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dictEnd, prefixStart) + minMatch;
                }
                if (dictMode == ZSTD_dictMatchState
                  && ( ((repOffset-1) /*intentional overflow*/ < curr - (dmsLowLimit + dmsIndexDelta))  /* equivalent to `curr > repIndex >= dmsLowLimit` */
                     & (ZSTD_index_overlap_check(dictLimit, repIndex)) )
                  && (ZSTD_readMINMATCH(ip, minMatch) == ZSTD_readMINMATCH(repMatch, minMatch)) ) {
                    repLen = (U32)ZSTD_count_2segments(ip+minMatch, repMatch+minMatch, iLimit, dmsEnd, prefixStart) + minMatch;
            }   }
            /* save longer solution */
            if (repLen > bestLength) {
                DEBUGLOG(8, "found repCode %u (ll0:%u, offset:%u) of length %u",
                            repCode, ll0, repOffset, repLen);
                bestLength = repLen;
                matches[mnum].off = REPCODE_TO_OFFBASE(repCode - ll0 + 1);  /* expect value between 1 and 3 */
                matches[mnum].len = (U32)repLen;
                mnum++;
                if ( (repLen > sufficient_len)
                   | (ip+repLen == iLimit) ) {  /* best possible */
                    return mnum;
    }   }   }   }

    /* HC3 match finder */
    if ((mls == 3) /*static*/ && (bestLength < mls)) {
        U32 const matchIndex3 = ZSTD_insertAndFindFirstIndexHash3(ms, nextToUpdate3, ip);
        if ((matchIndex3 >= matchLow)
          & (curr - matchIndex3 < (1<<18)) /*heuristic : longer distance likely too expensive*/ ) {
            size_t mlen;
            if ((dictMode == ZSTD_noDict) /*static*/ || (dictMode == ZSTD_dictMatchState) /*static*/ || (matchIndex3 >= dictLimit)) {
                const BYTE* const match = base + matchIndex3;
                mlen = ZSTD_count(ip, match, iLimit);
            } else {
                const BYTE* const match = dictBase + matchIndex3;
                mlen = ZSTD_count_2segments(ip, match, iLimit, dictEnd, prefixStart);
            }

            /* save best solution */
            if (mlen >= mls /* == 3 > bestLength */) {
                DEBUGLOG(8, "found small match with hlog3, of length %u",
                            (U32)mlen);
                bestLength = mlen;
                assert(curr > matchIndex3);
                assert(mnum==0);  /* no prior solution */
                matches[0].off = OFFSET_TO_OFFBASE(curr - matchIndex3);
                matches[0].len = (U32)mlen;
                mnum = 1;
                if ( (mlen > sufficient_len) |
                     (ip+mlen == iLimit) ) {  /* best possible length */
                    ms->nextToUpdate = curr+1;  /* skip insertion */
                    return 1;
        }   }   }
        /* no dictMatchState lookup: dicts don't have a populated HC3 table */
    }  /* if (mls == 3) */

    hashTable[h] = curr;   /* Update Hash Table */

    for (; nbCompares && (matchIndex >= matchLow); --nbCompares) {
        U32* const nextPtr = bt + 2*(matchIndex & btMask);
        const BYTE* match;
        size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
        assert(curr > matchIndex);

        if ((dictMode == ZSTD_noDict) || (dictMode == ZSTD_dictMatchState) || (matchIndex+matchLength >= dictLimit)) {
            assert(matchIndex+matchLength >= dictLimit);  /* ensure the condition is correct when !extDict */
            match = base + matchIndex;
            if (matchIndex >= dictLimit) assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
            matchLength += ZSTD_count(ip+matchLength, match+matchLength, iLimit);
        } else {
            match = dictBase + matchIndex;
            assert(memcmp(match, ip, matchLength) == 0);  /* ensure early section of match is equal as expected */
            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dictEnd, prefixStart);
            if (matchIndex+matchLength >= dictLimit)
                match = base + matchIndex;   /* prepare for match[matchLength] read */
        }

        if (matchLength > bestLength) {
            DEBUGLOG(8, "found match of length %u at distance %u (offBase=%u)",
                    (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
            assert(matchEndIdx > matchIndex);
            if (matchLength > matchEndIdx - matchIndex)
                matchEndIdx = matchIndex + (U32)matchLength;
            bestLength = matchLength;
            matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
            matches[mnum].len = (U32)matchLength;
            mnum++;
            if ( (matchLength > ZSTD_OPT_NUM)
               | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                if (dictMode == ZSTD_dictMatchState) nbCompares = 0; /* break should also skip searching dms */
                break; /* drop, to preserve bt consistency (miss a little bit of compression) */
        }   }

        if (match[matchLength] < ip[matchLength]) {
            /* match smaller than current */
            *smallerPtr = matchIndex;             /* update smaller idx */
            commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
            if (matchIndex <= btLow) { smallerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
            smallerPtr = nextPtr+1;               /* new candidate => larger than match, which was smaller than current */
            matchIndex = nextPtr[1];              /* new matchIndex, larger than previous, closer to current */
        } else {
            *largerPtr = matchIndex;
            commonLengthLarger = matchLength;
            if (matchIndex <= btLow) { largerPtr=&dummy32; break; }   /* beyond tree size, stop the search */
            largerPtr = nextPtr;
            matchIndex = nextPtr[0];
    }   }

    *smallerPtr = *largerPtr = 0;

    assert(nbCompares <= (1U << ZSTD_SEARCHLOG_MAX)); /* Check we haven't underflowed. */
    if (dictMode == ZSTD_dictMatchState && nbCompares) {
        size_t const dmsH = ZSTD_hashPtr(ip, dmsHashLog, mls);
        U32 dictMatchIndex = dms->hashTable[dmsH];
        const U32* const dmsBt = dms->chainTable;
        commonLengthSmaller = commonLengthLarger = 0;
        for (; nbCompares && (dictMatchIndex > dmsLowLimit); --nbCompares) {
            const U32* const nextPtr = dmsBt + 2*(dictMatchIndex & dmsBtMask);
            size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger);   /* guaranteed minimum nb of common bytes */
            const BYTE* match = dmsBase + dictMatchIndex;
            matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iLimit, dmsEnd, prefixStart);
            if (dictMatchIndex+matchLength >= dmsHighLimit)
                match = base + dictMatchIndex + dmsIndexDelta;   /* to prepare for next usage of match[matchLength] */

            if (matchLength > bestLength) {
                matchIndex = dictMatchIndex + dmsIndexDelta;
                DEBUGLOG(8, "found dms match of length %u at distance %u (offBase=%u)",
                        (U32)matchLength, curr - matchIndex, OFFSET_TO_OFFBASE(curr - matchIndex));
                if (matchLength > matchEndIdx - matchIndex)
                    matchEndIdx = matchIndex + (U32)matchLength;
                bestLength = matchLength;
                matches[mnum].off = OFFSET_TO_OFFBASE(curr - matchIndex);
                matches[mnum].len = (U32)matchLength;
                mnum++;
                if ( (matchLength > ZSTD_OPT_NUM)
                   | (ip+matchLength == iLimit) /* equal : no way to know if inf or sup */) {
                    break;   /* drop, to guarantee consistency (miss a little bit of compression) */
            }   }

            if (dictMatchIndex <= dmsBtLow) { break; }   /* beyond tree size, stop the search */
            if (match[matchLength] < ip[matchLength]) {
                commonLengthSmaller = matchLength;    /* all smaller will now have at least this guaranteed common length */
                dictMatchIndex = nextPtr[1];              /* new matchIndex larger than previous (closer to current) */
            } else {
                /* match is larger than current */
                commonLengthLarger = matchLength;
                dictMatchIndex = nextPtr[0];
    }   }   }  /* if (dictMode == ZSTD_dictMatchState) */

    assert(matchEndIdx > curr+8);
    ms->nextToUpdate = matchEndIdx - 8;  /* skip repetitive patterns */
    return mnum;
}

typedef U32 (*ZSTD_getAllMatchesFn)(
    ZSTD_match_t*,
    ZSTD_MatchState_t*,
    U32*,
    const BYTE*,
    const BYTE*,
    const U32 rep[ZSTD_REP_NUM],
    U32 const ll0,
    U32 const lengthToBeat);

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
U32 ZSTD_btGetAllMatches_internal(
        ZSTD_match_t* matches,
        ZSTD_MatchState_t* ms,
        U32* nextToUpdate3,
        const BYTE* ip,
        const BYTE* const iHighLimit,
        const U32 rep[ZSTD_REP_NUM],
        U32 const ll0,
        U32 const lengthToBeat,
        const ZSTD_dictMode_e dictMode,
        const U32 mls)
{
    assert(BOUNDED(3, ms->cParams.minMatch, 6) == mls);
    DEBUGLOG(8, "ZSTD_BtGetAllMatches(dictMode=%d, mls=%u)", (int)dictMode, mls);
    if (ip < ms->window.base + ms->nextToUpdate)
        return 0;   /* skipped area */
    ZSTD_updateTree_internal(ms, ip, iHighLimit, mls, dictMode);
    return ZSTD_insertBtAndGetAllMatches(matches, ms, nextToUpdate3, ip, iHighLimit, dictMode, rep, ll0, lengthToBeat, mls);
}

#define ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls) ZSTD_btGetAllMatches_##dictMode##_##mls

#define GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, mls)            \
    static U32 ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, mls)(      \
            ZSTD_match_t* matches,                             \
            ZSTD_MatchState_t* ms,                             \
            U32* nextToUpdate3,                                \
            const BYTE* ip,                                    \
            const BYTE* const iHighLimit,                      \
            const U32 rep[ZSTD_REP_NUM],                       \
            U32 const ll0,                                     \
            U32 const lengthToBeat)                            \
    {                                                          \
        return ZSTD_btGetAllMatches_internal(                  \
                matches, ms, nextToUpdate3, ip, iHighLimit,    \
                rep, ll0, lengthToBeat, ZSTD_##dictMode, mls); \
    }

#define GEN_ZSTD_BT_GET_ALL_MATCHES(dictMode)  \
    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 3)  \
    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 4)  \
    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 5)  \
    GEN_ZSTD_BT_GET_ALL_MATCHES_(dictMode, 6)

GEN_ZSTD_BT_GET_ALL_MATCHES(noDict)
GEN_ZSTD_BT_GET_ALL_MATCHES(extDict)
GEN_ZSTD_BT_GET_ALL_MATCHES(dictMatchState)

#define ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMode)  \
    {                                            \
        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 3), \
        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 4), \
        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 5), \
        ZSTD_BT_GET_ALL_MATCHES_FN(dictMode, 6)  \
    }

static ZSTD_getAllMatchesFn
ZSTD_selectBtGetAllMatches(ZSTD_MatchState_t const* ms, ZSTD_dictMode_e const dictMode)
{
    ZSTD_getAllMatchesFn const getAllMatchesFns[3][4] = {
        ZSTD_BT_GET_ALL_MATCHES_ARRAY(noDict),
        ZSTD_BT_GET_ALL_MATCHES_ARRAY(extDict),
        ZSTD_BT_GET_ALL_MATCHES_ARRAY(dictMatchState)
    };
    U32 const mls = BOUNDED(3, ms->cParams.minMatch, 6);
    assert((U32)dictMode < 3);
    assert(mls - 3 < 4);
    return getAllMatchesFns[(int)dictMode][mls - 3];
}

/*************************
*  LDM helper functions  *
*************************/

/* Struct containing info needed to make decision about ldm inclusion */
typedef struct {
    RawSeqStore_t seqStore;   /* External match candidates store for this block */
    U32 startPosInBlock;      /* Start position of the current match candidate */
    U32 endPosInBlock;        /* End position of the current match candidate */
    U32 offset;               /* Offset of the match candidate */
} ZSTD_optLdm_t;

/* ZSTD_optLdm_skipRawSeqStoreBytes():
 * Moves forward in @rawSeqStore by @nbBytes,
 * which will update the fields 'pos' and 'posInSequence'.
 */
static void ZSTD_optLdm_skipRawSeqStoreBytes(RawSeqStore_t* rawSeqStore, size_t nbBytes)
{
    U32 currPos = (U32)(rawSeqStore->posInSequence + nbBytes);
    while (currPos && rawSeqStore->pos < rawSeqStore->size) {
        rawSeq currSeq = rawSeqStore->seq[rawSeqStore->pos];
        if (currPos >= currSeq.litLength + currSeq.matchLength) {
            currPos -= currSeq.litLength + currSeq.matchLength;
            rawSeqStore->pos++;
        } else {
            rawSeqStore->posInSequence = currPos;
            break;
        }
    }
    if (currPos == 0 || rawSeqStore->pos == rawSeqStore->size) {
        rawSeqStore->posInSequence = 0;
    }
}

/* ZSTD_opt_getNextMatchAndUpdateSeqStore():
 * Calculates the beginning and end of the next match in the current block.
 * Updates 'pos' and 'posInSequence' of the ldmSeqStore.
 */
static void
ZSTD_opt_getNextMatchAndUpdateSeqStore(ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
                                       U32 blockBytesRemaining)
{
    rawSeq currSeq;
    U32 currBlockEndPos;
    U32 literalsBytesRemaining;
    U32 matchBytesRemaining;

    /* Setting match end position to MAX to ensure we never use an LDM during this block */
    if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
        optLdm->startPosInBlock = UINT_MAX;
        optLdm->endPosInBlock = UINT_MAX;
        return;
    }
    /* Calculate appropriate bytes left in matchLength and litLength
     * after adjusting based on ldmSeqStore->posInSequence */
    currSeq = optLdm->seqStore.seq[optLdm->seqStore.pos];
    assert(optLdm->seqStore.posInSequence <= currSeq.litLength + currSeq.matchLength);
    currBlockEndPos = currPosInBlock + blockBytesRemaining;
    literalsBytesRemaining = (optLdm->seqStore.posInSequence < currSeq.litLength) ?
            currSeq.litLength - (U32)optLdm->seqStore.posInSequence :
            0;
    matchBytesRemaining = (literalsBytesRemaining == 0) ?
            currSeq.matchLength - ((U32)optLdm->seqStore.posInSequence - currSeq.litLength) :
            currSeq.matchLength;

    /* If there are more literal bytes than bytes remaining in block, no ldm is possible */
    if (literalsBytesRemaining >= blockBytesRemaining) {
        optLdm->startPosInBlock = UINT_MAX;
        optLdm->endPosInBlock = UINT_MAX;
        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, blockBytesRemaining);
        return;
    }

    /* Matches may be < minMatch by this process. In that case, we will reject them
       when we are deciding whether or not to add the ldm */
    optLdm->startPosInBlock = currPosInBlock + literalsBytesRemaining;
    optLdm->endPosInBlock = optLdm->startPosInBlock + matchBytesRemaining;
    optLdm->offset = currSeq.offset;

    if (optLdm->endPosInBlock > currBlockEndPos) {
        /* Match ends after the block ends, we can't use the whole match */
        optLdm->endPosInBlock = currBlockEndPos;
        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, currBlockEndPos - currPosInBlock);
    } else {
        /* Consume nb of bytes equal to size of sequence left */
        ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, literalsBytesRemaining + matchBytesRemaining);
    }
}

/* ZSTD_optLdm_maybeAddMatch():
 * Adds a match if it's long enough,
 * based on it's 'matchStartPosInBlock' and 'matchEndPosInBlock',
 * into 'matches'. Maintains the correct ordering of 'matches'.
 */
static void ZSTD_optLdm_maybeAddMatch(ZSTD_match_t* matches, U32* nbMatches,
                                      const ZSTD_optLdm_t* optLdm, U32 currPosInBlock,
                                      U32 minMatch)
{
    U32 const posDiff = currPosInBlock - optLdm->startPosInBlock;
    /* Note: ZSTD_match_t actually contains offBase and matchLength (before subtracting MINMATCH) */
    U32 const candidateMatchLength = optLdm->endPosInBlock - optLdm->startPosInBlock - posDiff;

    /* Ensure that current block position is not outside of the match */
    if (currPosInBlock < optLdm->startPosInBlock
      || currPosInBlock >= optLdm->endPosInBlock
      || candidateMatchLength < minMatch) {
        return;
    }

    if (*nbMatches == 0 || ((candidateMatchLength > matches[*nbMatches-1].len) && *nbMatches < ZSTD_OPT_NUM)) {
        U32 const candidateOffBase = OFFSET_TO_OFFBASE(optLdm->offset);
        DEBUGLOG(6, "ZSTD_optLdm_maybeAddMatch(): Adding ldm candidate match (offBase: %u matchLength %u) at block position=%u",
                 candidateOffBase, candidateMatchLength, currPosInBlock);
        matches[*nbMatches].len = candidateMatchLength;
        matches[*nbMatches].off = candidateOffBase;
        (*nbMatches)++;
    }
}

/* ZSTD_optLdm_processMatchCandidate():
 * Wrapper function to update ldm seq store and call ldm functions as necessary.
 */
static void
ZSTD_optLdm_processMatchCandidate(ZSTD_optLdm_t* optLdm,
                                  ZSTD_match_t* matches, U32* nbMatches,
                                  U32 currPosInBlock, U32 remainingBytes,
                                  U32 minMatch)
{
    if (optLdm->seqStore.size == 0 || optLdm->seqStore.pos >= optLdm->seqStore.size) {
        return;
    }

    if (currPosInBlock >= optLdm->endPosInBlock) {
        if (currPosInBlock > optLdm->endPosInBlock) {
            /* The position at which ZSTD_optLdm_processMatchCandidate() is called is not necessarily
             * at the end of a match from the ldm seq store, and will often be some bytes
             * over beyond matchEndPosInBlock. As such, we need to correct for these "overshoots"
             */
            U32 const posOvershoot = currPosInBlock - optLdm->endPosInBlock;
            ZSTD_optLdm_skipRawSeqStoreBytes(&optLdm->seqStore, posOvershoot);
        }
        ZSTD_opt_getNextMatchAndUpdateSeqStore(optLdm, currPosInBlock, remainingBytes);
    }
    ZSTD_optLdm_maybeAddMatch(matches, nbMatches, optLdm, currPosInBlock, minMatch);
}


/*-*******************************
*  Optimal parser
*********************************/

#if 0 /* debug */

static void
listStats(const U32* table, int lastEltID)
{
    int const nbElts = lastEltID + 1;
    int enb;
    for (enb=0; enb < nbElts; enb++) {
        (void)table;
        /* RAWLOG(2, "%3i:%3i,  ", enb, table[enb]); */
        RAWLOG(2, "%4i,", table[enb]);
    }
    RAWLOG(2, " \n");
}

#endif

#define LIT_PRICE(_p) (int)ZSTD_rawLiteralsCost(_p, 1, optStatePtr, optLevel)
#define LL_PRICE(_l) (int)ZSTD_litLengthPrice(_l, optStatePtr, optLevel)
#define LL_INCPRICE(_l) (LL_PRICE(_l) - LL_PRICE(_l-1))

FORCE_INLINE_TEMPLATE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t
ZSTD_compressBlock_opt_generic(ZSTD_MatchState_t* ms,
                               SeqStore_t* seqStore,
                               U32 rep[ZSTD_REP_NUM],
                         const void* src, size_t srcSize,
                         const int optLevel,
                         const ZSTD_dictMode_e dictMode)
{
    optState_t* const optStatePtr = &ms->opt;
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* ip = istart;
    const BYTE* anchor = istart;
    const BYTE* const iend = istart + srcSize;
    const BYTE* const ilimit = iend - 8;
    const BYTE* const base = ms->window.base;
    const BYTE* const prefixStart = base + ms->window.dictLimit;
    const ZSTD_compressionParameters* const cParams = &ms->cParams;

    ZSTD_getAllMatchesFn getAllMatches = ZSTD_selectBtGetAllMatches(ms, dictMode);

    U32 const sufficient_len = MIN(cParams->targetLength, ZSTD_OPT_NUM -1);
    U32 const minMatch = (cParams->minMatch == 3) ? 3 : 4;
    U32 nextToUpdate3 = ms->nextToUpdate;

    ZSTD_optimal_t* const opt = optStatePtr->priceTable;
    ZSTD_match_t* const matches = optStatePtr->matchTable;
    ZSTD_optimal_t lastStretch;
    ZSTD_optLdm_t optLdm;

    ZSTD_memset(&lastStretch, 0, sizeof(ZSTD_optimal_t));

    optLdm.seqStore = ms->ldmSeqStore ? *ms->ldmSeqStore : kNullRawSeqStore;
    optLdm.endPosInBlock = optLdm.startPosInBlock = optLdm.offset = 0;
    ZSTD_opt_getNextMatchAndUpdateSeqStore(&optLdm, (U32)(ip-istart), (U32)(iend-ip));

    /* init */
    DEBUGLOG(5, "ZSTD_compressBlock_opt_generic: current=%u, prefix=%u, nextToUpdate=%u",
                (U32)(ip - base), ms->window.dictLimit, ms->nextToUpdate);
    assert(optLevel <= 2);
    ZSTD_rescaleFreqs(optStatePtr, (const BYTE*)src, srcSize, optLevel);
    ip += (ip==prefixStart);

    /* Match Loop */
    while (ip < ilimit) {
        U32 cur, last_pos = 0;

        /* find first match */
        {   U32 const litlen = (U32)(ip - anchor);
            U32 const ll0 = !litlen;
            U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, ip, iend, rep, ll0, minMatch);
            ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
                                              (U32)(ip-istart), (U32)(iend-ip),
                                              minMatch);
            if (!nbMatches) {
                DEBUGLOG(8, "no match found at cPos %u", (unsigned)(ip-istart));
                ip++;
                continue;
            }

            /* Match found: let's store this solution, and eventually find more candidates.
             * During this forward pass, @opt is used to store stretches,
             * defined as "a match followed by N literals".
             * Note how this is different from a Sequence, which is "N literals followed by a match".
             * Storing stretches allows us to store different match predecessors
             * for each literal position part of a literals run. */

            /* initialize opt[0] */
            opt[0].mlen = 0;  /* there are only literals so far */
            opt[0].litlen = litlen;
            /* No need to include the actual price of the literals before the first match
             * because it is static for the duration of the forward pass, and is included
             * in every subsequent price. But, we include the literal length because
             * the cost variation of litlen depends on the value of litlen.
             */
            opt[0].price = LL_PRICE(litlen);
            ZSTD_STATIC_ASSERT(sizeof(opt[0].rep[0]) == sizeof(rep[0]));
            ZSTD_memcpy(&opt[0].rep, rep, sizeof(opt[0].rep));

            /* large match -> immediate encoding */
            {   U32 const maxML = matches[nbMatches-1].len;
                U32 const maxOffBase = matches[nbMatches-1].off;
                DEBUGLOG(6, "found %u matches of maxLength=%u and maxOffBase=%u at cPos=%u => start new series",
                            nbMatches, maxML, maxOffBase, (U32)(ip-prefixStart));

                if (maxML > sufficient_len) {
                    lastStretch.litlen = 0;
                    lastStretch.mlen = maxML;
                    lastStretch.off = maxOffBase;
                    DEBUGLOG(6, "large match (%u>%u) => immediate encoding",
                                maxML, sufficient_len);
                    cur = 0;
                    last_pos = maxML;
                    goto _shortestPath;
            }   }

            /* set prices for first matches starting position == 0 */
            assert(opt[0].price >= 0);
            {   U32 pos;
                U32 matchNb;
                for (pos = 1; pos < minMatch; pos++) {
                    opt[pos].price = ZSTD_MAX_PRICE;
                    opt[pos].mlen = 0;
                    opt[pos].litlen = litlen + pos;
                }
                for (matchNb = 0; matchNb < nbMatches; matchNb++) {
                    U32 const offBase = matches[matchNb].off;
                    U32 const end = matches[matchNb].len;
                    for ( ; pos <= end ; pos++ ) {
                        int const matchPrice = (int)ZSTD_getMatchPrice(offBase, pos, optStatePtr, optLevel);
                        int const sequencePrice = opt[0].price + matchPrice;
                        DEBUGLOG(7, "rPos:%u => set initial price : %.2f",
                                    pos, ZSTD_fCost(sequencePrice));
                        opt[pos].mlen = pos;
                        opt[pos].off = offBase;
                        opt[pos].litlen = 0; /* end of match */
                        opt[pos].price = sequencePrice + LL_PRICE(0);
                    }
                }
                last_pos = pos-1;
                opt[pos].price = ZSTD_MAX_PRICE;
            }
        }

        /* check further positions */
        for (cur = 1; cur <= last_pos; cur++) {
            const BYTE* const inr = ip + cur;
            assert(cur <= ZSTD_OPT_NUM);
            DEBUGLOG(7, "cPos:%i==rPos:%u", (int)(inr-istart), cur);

            /* Fix current position with one literal if cheaper */
            {   U32 const litlen = opt[cur-1].litlen + 1;
                int const price = opt[cur-1].price
                                + LIT_PRICE(ip+cur-1)
                                + LL_INCPRICE(litlen);
                assert(price < 1000000000); /* overflow check */
                if (price <= opt[cur].price) {
                    ZSTD_optimal_t const prevMatch = opt[cur];
                    DEBUGLOG(7, "cPos:%i==rPos:%u : better price (%.2f<=%.2f) using literal (ll==%u) (hist:%u,%u,%u)",
                                (int)(inr-istart), cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price), litlen,
                                opt[cur-1].rep[0], opt[cur-1].rep[1], opt[cur-1].rep[2]);
                    opt[cur] = opt[cur-1];
                    opt[cur].litlen = litlen;
                    opt[cur].price = price;
                    if ( (optLevel >= 1) /* additional check only for higher modes */
                      && (prevMatch.litlen == 0) /* replace a match */
                      && (LL_INCPRICE(1) < 0) /* ll1 is cheaper than ll0 */
                      && LIKELY(ip + cur < iend)
                    ) {
                        /* check next position, in case it would be cheaper */
                        int with1literal = prevMatch.price + LIT_PRICE(ip+cur) + LL_INCPRICE(1);
                        int withMoreLiterals = price + LIT_PRICE(ip+cur) + LL_INCPRICE(litlen+1);
                        DEBUGLOG(7, "then at next rPos %u : match+1lit %.2f vs %ulits %.2f",
                                cur+1, ZSTD_fCost(with1literal), litlen+1, ZSTD_fCost(withMoreLiterals));
                        if ( (with1literal < withMoreLiterals)
                          && (with1literal < opt[cur+1].price) ) {
                            /* update offset history - before it disappears */
                            U32 const prev = cur - prevMatch.mlen;
                            Repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, prevMatch.off, opt[prev].litlen==0);
                            assert(cur >= prevMatch.mlen);
                            DEBUGLOG(7, "==> match+1lit is cheaper (%.2f < %.2f) (hist:%u,%u,%u) !",
                                        ZSTD_fCost(with1literal), ZSTD_fCost(withMoreLiterals),
                                        newReps.rep[0], newReps.rep[1], newReps.rep[2] );
                            opt[cur+1] = prevMatch;  /* mlen & offbase */
                            ZSTD_memcpy(opt[cur+1].rep, &newReps, sizeof(Repcodes_t));
                            opt[cur+1].litlen = 1;
                            opt[cur+1].price = with1literal;
                            if (last_pos < cur+1) last_pos = cur+1;
                        }
                    }
                } else {
                    DEBUGLOG(7, "cPos:%i==rPos:%u : literal would cost more (%.2f>%.2f)",
                                (int)(inr-istart), cur, ZSTD_fCost(price), ZSTD_fCost(opt[cur].price));
                }
            }

            /* Offset history is not updated during match comparison.
             * Do it here, now that the match is selected and confirmed.
             */
            ZSTD_STATIC_ASSERT(sizeof(opt[cur].rep) == sizeof(Repcodes_t));
            assert(cur >= opt[cur].mlen);
            if (opt[cur].litlen == 0) {
                /* just finished a match => alter offset history */
                U32 const prev = cur - opt[cur].mlen;
                Repcodes_t const newReps = ZSTD_newRep(opt[prev].rep, opt[cur].off, opt[prev].litlen==0);
                ZSTD_memcpy(opt[cur].rep, &newReps, sizeof(Repcodes_t));
            }

            /* last match must start at a minimum distance of 8 from oend */
            if (inr > ilimit) continue;

            if (cur == last_pos) break;

            if ( (optLevel==0) /*static_test*/
              && (opt[cur+1].price <= opt[cur].price + (BITCOST_MULTIPLIER/2)) ) {
                DEBUGLOG(7, "skip current position : next rPos(%u) price is cheaper", cur+1);
                continue;  /* skip unpromising positions; about ~+6% speed, -0.01 ratio */
            }

            assert(opt[cur].price >= 0);
            {   U32 const ll0 = (opt[cur].litlen == 0);
                int const previousPrice = opt[cur].price;
                int const basePrice = previousPrice + LL_PRICE(0);
                U32 nbMatches = getAllMatches(matches, ms, &nextToUpdate3, inr, iend, opt[cur].rep, ll0, minMatch);
                U32 matchNb;

                ZSTD_optLdm_processMatchCandidate(&optLdm, matches, &nbMatches,
                                                  (U32)(inr-istart), (U32)(iend-inr),
                                                  minMatch);

                if (!nbMatches) {
                    DEBUGLOG(7, "rPos:%u : no match found", cur);
                    continue;
                }

                {   U32 const longestML = matches[nbMatches-1].len;
                    DEBUGLOG(7, "cPos:%i==rPos:%u, found %u matches, of longest ML=%u",
                                (int)(inr-istart), cur, nbMatches, longestML);

                    if ( (longestML > sufficient_len)
                      || (cur + longestML >= ZSTD_OPT_NUM)
                      || (ip + cur + longestML >= iend) ) {
                        lastStretch.mlen = longestML;
                        lastStretch.off = matches[nbMatches-1].off;
                        lastStretch.litlen = 0;
                        last_pos = cur + longestML;
                        goto _shortestPath;
                }   }

                /* set prices using matches found at position == cur */
                for (matchNb = 0; matchNb < nbMatches; matchNb++) {
                    U32 const offset = matches[matchNb].off;
                    U32 const lastML = matches[matchNb].len;
                    U32 const startML = (matchNb>0) ? matches[matchNb-1].len+1 : minMatch;
                    U32 mlen;

                    DEBUGLOG(7, "testing match %u => offBase=%4u, mlen=%2u, llen=%2u",
                                matchNb, matches[matchNb].off, lastML, opt[cur].litlen);

                    for (mlen = lastML; mlen >= startML; mlen--) {  /* scan downward */
                        U32 const pos = cur + mlen;
                        int const price = basePrice + (int)ZSTD_getMatchPrice(offset, mlen, optStatePtr, optLevel);

                        if ((pos > last_pos) || (price < opt[pos].price)) {
                            DEBUGLOG(7, "rPos:%u (ml=%2u) => new better price (%.2f<%.2f)",
                                        pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
                            while (last_pos < pos) {
                                /* fill empty positions, for future comparisons */
                                last_pos++;
                                opt[last_pos].price = ZSTD_MAX_PRICE;
                                opt[last_pos].litlen = !0;  /* just needs to be != 0, to mean "not an end of match" */
                            }
                            opt[pos].mlen = mlen;
                            opt[pos].off = offset;
                            opt[pos].litlen = 0;
                            opt[pos].price = price;
                        } else {
                            DEBUGLOG(7, "rPos:%u (ml=%2u) => new price is worse (%.2f>=%.2f)",
                                        pos, mlen, ZSTD_fCost(price), ZSTD_fCost(opt[pos].price));
                            if (optLevel==0) break;  /* early update abort; gets ~+10% speed for about -0.01 ratio loss */
                        }
            }   }   }
            opt[last_pos+1].price = ZSTD_MAX_PRICE;
        }  /* for (cur = 1; cur <= last_pos; cur++) */

        lastStretch = opt[last_pos];
        assert(cur >= lastStretch.mlen);
        cur = last_pos - lastStretch.mlen;

_shortestPath:   /* cur, last_pos, best_mlen, best_off have to be set */
        assert(opt[0].mlen == 0);
        assert(last_pos >= lastStretch.mlen);
        assert(cur == last_pos - lastStretch.mlen);

        if (lastStretch.mlen==0) {
            /* no solution : all matches have been converted into literals */
            assert(lastStretch.litlen == (ip - anchor) + last_pos);
            ip += last_pos;
            continue;
        }
        assert(lastStretch.off > 0);

        /* Update offset history */
        if (lastStretch.litlen == 0) {
            /* finishing on a match : update offset history */
            Repcodes_t const reps = ZSTD_newRep(opt[cur].rep, lastStretch.off, opt[cur].litlen==0);
            ZSTD_memcpy(rep, &reps, sizeof(Repcodes_t));
        } else {
            ZSTD_memcpy(rep, lastStretch.rep, sizeof(Repcodes_t));
            assert(cur >= lastStretch.litlen);
            cur -= lastStretch.litlen;
        }

        /* Let's write the shortest path solution.
         * It is stored in @opt in reverse order,
         * starting from @storeEnd (==cur+2),
         * effectively partially @opt overwriting.
         * Content is changed too:
         * - So far, @opt stored stretches, aka a match followed by literals
         * - Now, it will store sequences, aka literals followed by a match
         */
        {   U32 const storeEnd = cur + 2;
            U32 storeStart = storeEnd;
            U32 stretchPos = cur;

            DEBUGLOG(6, "start reverse traversal (last_pos:%u, cur:%u)",
                        last_pos, cur); (void)last_pos;
            assert(storeEnd < ZSTD_OPT_SIZE);
            DEBUGLOG(6, "last stretch copied into pos=%u (llen=%u,mlen=%u,ofc=%u)",
                        storeEnd, lastStretch.litlen, lastStretch.mlen, lastStretch.off);
            if (lastStretch.litlen > 0) {
                /* last "sequence" is unfinished: just a bunch of literals */
                opt[storeEnd].litlen = lastStretch.litlen;
                opt[storeEnd].mlen = 0;
                storeStart = storeEnd-1;
                opt[storeStart] = lastStretch;
            } {
                opt[storeEnd] = lastStretch;  /* note: litlen will be fixed */
                storeStart = storeEnd;
            }
            while (1) {
                ZSTD_optimal_t nextStretch = opt[stretchPos];
                opt[storeStart].litlen = nextStretch.litlen;
                DEBUGLOG(6, "selected sequence (llen=%u,mlen=%u,ofc=%u)",
                            opt[storeStart].litlen, opt[storeStart].mlen, opt[storeStart].off);
                if (nextStretch.mlen == 0) {
                    /* reaching beginning of segment */
                    break;
                }
                storeStart--;
                opt[storeStart] = nextStretch; /* note: litlen will be fixed */
                assert(nextStretch.litlen + nextStretch.mlen <= stretchPos);
                stretchPos -= nextStretch.litlen + nextStretch.mlen;
            }

            /* save sequences */
            DEBUGLOG(6, "sending selected sequences into seqStore");
            {   U32 storePos;
                for (storePos=storeStart; storePos <= storeEnd; storePos++) {
                    U32 const llen = opt[storePos].litlen;
                    U32 const mlen = opt[storePos].mlen;
                    U32 const offBase = opt[storePos].off;
                    U32 const advance = llen + mlen;
                    DEBUGLOG(6, "considering seq starting at %i, llen=%u, mlen=%u",
                                (int)(anchor - istart), (unsigned)llen, (unsigned)mlen);

                    if (mlen==0) {  /* only literals => must be last "sequence", actually starting a new stream of sequences */
                        assert(storePos == storeEnd);   /* must be last sequence */
                        ip = anchor + llen;     /* last "sequence" is a bunch of literals => don't progress anchor */
                        continue;   /* will finish */
                    }

                    assert(anchor + llen <= iend);
                    ZSTD_updateStats(optStatePtr, llen, anchor, offBase, mlen);
                    ZSTD_storeSeq(seqStore, llen, anchor, iend, offBase, mlen);
                    anchor += advance;
                    ip = anchor;
            }   }
            DEBUGLOG(7, "new offset history : %u, %u, %u", rep[0], rep[1], rep[2]);

            /* update all costs */
            ZSTD_setBasePrices(optStatePtr, optLevel);
        }
    }   /* while (ip < ilimit) */

    /* Return the last literals size */
    return (size_t)(iend - anchor);
}
#endif /* build exclusions */

#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
static size_t ZSTD_compressBlock_opt0(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
{
    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 0 /* optLevel */, dictMode);
}
#endif

#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
static size_t ZSTD_compressBlock_opt2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize, const ZSTD_dictMode_e dictMode)
{
    return ZSTD_compressBlock_opt_generic(ms, seqStore, rep, src, srcSize, 2 /* optLevel */, dictMode);
}
#endif

#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btopt(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_compressBlock_btopt");
    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
}
#endif


#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
/* ZSTD_initStats_ultra():
 * make a first compression pass, just to seed stats with more accurate starting values.
 * only works on first block, with no dictionary and no ldm.
 * this function cannot error out, its narrow contract must be respected.
 */
static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_initStats_ultra(ZSTD_MatchState_t* ms,
                          SeqStore_t* seqStore,
                          U32 rep[ZSTD_REP_NUM],
                    const void* src, size_t srcSize)
{
    U32 tmpRep[ZSTD_REP_NUM];  /* updated rep codes will sink here */
    ZSTD_memcpy(tmpRep, rep, sizeof(tmpRep));

    DEBUGLOG(4, "ZSTD_initStats_ultra (srcSize=%zu)", srcSize);
    assert(ms->opt.litLengthSum == 0);    /* first block */
    assert(seqStore->sequences == seqStore->sequencesStart);   /* no ldm */
    assert(ms->window.dictLimit == ms->window.lowLimit);   /* no dictionary */
    assert(ms->window.dictLimit - ms->nextToUpdate <= 1);  /* no prefix (note: intentional overflow, defined as 2-complement) */

    ZSTD_compressBlock_opt2(ms, seqStore, tmpRep, src, srcSize, ZSTD_noDict);   /* generate stats into ms->opt*/

    /* invalidate first scan from history, only keep entropy stats */
    ZSTD_resetSeqStore(seqStore);
    ms->window.base -= srcSize;
    ms->window.dictLimit += (U32)srcSize;
    ms->window.lowLimit = ms->window.dictLimit;
    ms->nextToUpdate = ms->window.dictLimit;

}

size_t ZSTD_compressBlock_btultra(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_compressBlock_btultra (srcSize=%zu)", srcSize);
    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
}

size_t ZSTD_compressBlock_btultra2(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    U32 const curr = (U32)((const BYTE*)src - ms->window.base);
    DEBUGLOG(5, "ZSTD_compressBlock_btultra2 (srcSize=%zu)", srcSize);

    /* 2-passes strategy:
     * this strategy makes a first pass over first block to collect statistics
     * in order to seed next round's statistics with it.
     * After 1st pass, function forgets history, and starts a new block.
     * Consequently, this can only work if no data has been previously loaded in tables,
     * aka, no dictionary, no prefix, no ldm preprocessing.
     * The compression ratio gain is generally small (~0.5% on first block),
     * the cost is 2x cpu time on first block. */
    assert(srcSize <= ZSTD_BLOCKSIZE_MAX);
    if ( (ms->opt.litLengthSum==0)   /* first block */
      && (seqStore->sequences == seqStore->sequencesStart)  /* no ldm */
      && (ms->window.dictLimit == ms->window.lowLimit)   /* no dictionary */
      && (curr == ms->window.dictLimit)    /* start of frame, nothing already loaded nor skipped */
      && (srcSize > ZSTD_PREDEF_THRESHOLD) /* input large enough to not employ default stats */
      ) {
        ZSTD_initStats_ultra(ms, seqStore, rep, src, srcSize);
    }

    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_noDict);
}
#endif

#ifndef ZSTD_EXCLUDE_BTOPT_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btopt_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_btopt_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    return ZSTD_compressBlock_opt0(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
}
#endif

#ifndef ZSTD_EXCLUDE_BTULTRA_BLOCK_COMPRESSOR
size_t ZSTD_compressBlock_btultra_dictMatchState(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_dictMatchState);
}

size_t ZSTD_compressBlock_btultra_extDict(
        ZSTD_MatchState_t* ms, SeqStore_t* seqStore, U32 rep[ZSTD_REP_NUM],
        const void* src, size_t srcSize)
{
    return ZSTD_compressBlock_opt2(ms, seqStore, rep, src, srcSize, ZSTD_extDict);
}
#endif

/* note : no btultra2 variant for extDict nor dictMatchState,
 * because btultra2 is not meant to work with dictionaries
 * and is only specific for the first block (no prefix) */
/**** ended inlining compress/zstd_opt.c ****/
#ifdef ZSTD_MULTITHREAD
/**** start inlining compress/zstdmt_compress.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/* ======   Compiler specifics   ====== */
#if defined(_MSC_VER)
#  pragma warning(disable : 4204)   /* disable: C4204: non-constant aggregate initializer */
#endif


/* ======   Dependencies   ====== */
/**** skipping file: ../common/allocations.h ****/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/pool.h ****/
/**** skipping file: ../common/threading.h ****/
/**** skipping file: zstd_compress_internal.h ****/
/**** skipping file: zstd_ldm.h ****/
/**** skipping file: zstdmt_compress.h ****/

/* Guards code to support resizing the SeqPool.
 * We will want to resize the SeqPool to save memory in the future.
 * Until then, comment the code out since it is unused.
 */
#define ZSTD_RESIZE_SEQPOOL 0

/* ======   Debug   ====== */
#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=2) \
    && !defined(_MSC_VER) \
    && !defined(__MINGW32__)

#  include <stdio.h>
#  include <unistd.h>
#  include <sys/times.h>

#  define DEBUG_PRINTHEX(l,p,n)                                       \
    do {                                                              \
        unsigned debug_u;                                             \
        for (debug_u=0; debug_u<(n); debug_u++)                       \
            RAWLOG(l, "%02X ", ((const unsigned char*)(p))[debug_u]); \
        RAWLOG(l, " \n");                                             \
    } while (0)

static unsigned long long GetCurrentClockTimeMicroseconds(void)
{
   static clock_t _ticksPerSecond = 0;
   if (_ticksPerSecond <= 0) _ticksPerSecond = sysconf(_SC_CLK_TCK);

   {   struct tms junk; clock_t newTicks = (clock_t) times(&junk);
       return ((((unsigned long long)newTicks)*(1000000))/_ticksPerSecond);
}  }

#define MUTEX_WAIT_TIME_DLEVEL 6
#define ZSTD_PTHREAD_MUTEX_LOCK(mutex)                                                  \
    do {                                                                                \
        if (DEBUGLEVEL >= MUTEX_WAIT_TIME_DLEVEL) {                                     \
            unsigned long long const beforeTime = GetCurrentClockTimeMicroseconds();    \
            ZSTD_pthread_mutex_lock(mutex);                                             \
            {   unsigned long long const afterTime = GetCurrentClockTimeMicroseconds(); \
                unsigned long long const elapsedTime = (afterTime-beforeTime);          \
                if (elapsedTime > 1000) {                                               \
                    /* or whatever threshold you like; I'm using 1 millisecond here */  \
                    DEBUGLOG(MUTEX_WAIT_TIME_DLEVEL,                                    \
                        "Thread took %llu microseconds to acquire mutex %s \n",         \
                        elapsedTime, #mutex);                                           \
            }   }                                                                       \
        } else {                                                                        \
            ZSTD_pthread_mutex_lock(mutex);                                             \
        }                                                                               \
    } while (0)

#else

#  define ZSTD_PTHREAD_MUTEX_LOCK(m) ZSTD_pthread_mutex_lock(m)
#  define DEBUG_PRINTHEX(l,p,n) do { } while (0)

#endif


/* =====   Buffer Pool   ===== */
/* a single Buffer Pool can be invoked from multiple threads in parallel */

typedef struct buffer_s {
    void* start;
    size_t capacity;
} Buffer;

static const Buffer g_nullBuffer = { NULL, 0 };

typedef struct ZSTDMT_bufferPool_s {
    ZSTD_pthread_mutex_t poolMutex;
    size_t bufferSize;
    unsigned totalBuffers;
    unsigned nbBuffers;
    ZSTD_customMem cMem;
    Buffer* buffers;
} ZSTDMT_bufferPool;

static void ZSTDMT_freeBufferPool(ZSTDMT_bufferPool* bufPool)
{
    DEBUGLOG(3, "ZSTDMT_freeBufferPool (address:%08X)", (U32)(size_t)bufPool);
    if (!bufPool) return;   /* compatibility with free on NULL */
    if (bufPool->buffers) {
        unsigned u;
        for (u=0; u<bufPool->totalBuffers; u++) {
            DEBUGLOG(4, "free buffer %2u (address:%08X)", u, (U32)(size_t)bufPool->buffers[u].start);
            ZSTD_customFree(bufPool->buffers[u].start, bufPool->cMem);
        }
        ZSTD_customFree(bufPool->buffers, bufPool->cMem);
    }
    ZSTD_pthread_mutex_destroy(&bufPool->poolMutex);
    ZSTD_customFree(bufPool, bufPool->cMem);
}

static ZSTDMT_bufferPool* ZSTDMT_createBufferPool(unsigned maxNbBuffers, ZSTD_customMem cMem)
{
    ZSTDMT_bufferPool* const bufPool =
        (ZSTDMT_bufferPool*)ZSTD_customCalloc(sizeof(ZSTDMT_bufferPool), cMem);
    if (bufPool==NULL) return NULL;
    if (ZSTD_pthread_mutex_init(&bufPool->poolMutex, NULL)) {
        ZSTD_customFree(bufPool, cMem);
        return NULL;
    }
    bufPool->buffers = (Buffer*)ZSTD_customCalloc(maxNbBuffers * sizeof(Buffer), cMem);
    if (bufPool->buffers==NULL) {
        ZSTDMT_freeBufferPool(bufPool);
        return NULL;
    }
    bufPool->bufferSize = 64 KB;
    bufPool->totalBuffers = maxNbBuffers;
    bufPool->nbBuffers = 0;
    bufPool->cMem = cMem;
    return bufPool;
}

/* only works at initialization, not during compression */
static size_t ZSTDMT_sizeof_bufferPool(ZSTDMT_bufferPool* bufPool)
{
    size_t const poolSize = sizeof(*bufPool);
    size_t const arraySize = bufPool->totalBuffers * sizeof(Buffer);
    unsigned u;
    size_t totalBufferSize = 0;
    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
    for (u=0; u<bufPool->totalBuffers; u++)
        totalBufferSize += bufPool->buffers[u].capacity;
    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);

    return poolSize + arraySize + totalBufferSize;
}

/* ZSTDMT_setBufferSize() :
 * all future buffers provided by this buffer pool will have _at least_ this size
 * note : it's better for all buffers to have same size,
 * as they become freely interchangeable, reducing malloc/free usages and memory fragmentation */
static void ZSTDMT_setBufferSize(ZSTDMT_bufferPool* const bufPool, size_t const bSize)
{
    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
    DEBUGLOG(4, "ZSTDMT_setBufferSize: bSize = %u", (U32)bSize);
    bufPool->bufferSize = bSize;
    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
}


static ZSTDMT_bufferPool* ZSTDMT_expandBufferPool(ZSTDMT_bufferPool* srcBufPool, unsigned maxNbBuffers)
{
    if (srcBufPool==NULL) return NULL;
    if (srcBufPool->totalBuffers >= maxNbBuffers) /* good enough */
        return srcBufPool;
    /* need a larger buffer pool */
    {   ZSTD_customMem const cMem = srcBufPool->cMem;
        size_t const bSize = srcBufPool->bufferSize;   /* forward parameters */
        ZSTDMT_bufferPool* newBufPool;
        ZSTDMT_freeBufferPool(srcBufPool);
        newBufPool = ZSTDMT_createBufferPool(maxNbBuffers, cMem);
        if (newBufPool==NULL) return newBufPool;
        ZSTDMT_setBufferSize(newBufPool, bSize);
        return newBufPool;
    }
}

/** ZSTDMT_getBuffer() :
 *  assumption : bufPool must be valid
 * @return : a buffer, with start pointer and size
 *  note: allocation may fail, in this case, start==NULL and size==0 */
static Buffer ZSTDMT_getBuffer(ZSTDMT_bufferPool* bufPool)
{
    size_t const bSize = bufPool->bufferSize;
    DEBUGLOG(5, "ZSTDMT_getBuffer: bSize = %u", (U32)bufPool->bufferSize);
    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
    if (bufPool->nbBuffers) {   /* try to use an existing buffer */
        Buffer const buf = bufPool->buffers[--(bufPool->nbBuffers)];
        size_t const availBufferSize = buf.capacity;
        bufPool->buffers[bufPool->nbBuffers] = g_nullBuffer;
        if ((availBufferSize >= bSize) & ((availBufferSize>>3) <= bSize)) {
            /* large enough, but not too much */
            DEBUGLOG(5, "ZSTDMT_getBuffer: provide buffer %u of size %u",
                        bufPool->nbBuffers, (U32)buf.capacity);
            ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
            return buf;
        }
        /* size conditions not respected : scratch this buffer, create new one */
        DEBUGLOG(5, "ZSTDMT_getBuffer: existing buffer does not meet size conditions => freeing");
        ZSTD_customFree(buf.start, bufPool->cMem);
    }
    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
    /* create new buffer */
    DEBUGLOG(5, "ZSTDMT_getBuffer: create a new buffer");
    {   Buffer buffer;
        void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
        buffer.start = start;   /* note : start can be NULL if malloc fails ! */
        buffer.capacity = (start==NULL) ? 0 : bSize;
        if (start==NULL) {
            DEBUGLOG(5, "ZSTDMT_getBuffer: buffer allocation failure !!");
        } else {
            DEBUGLOG(5, "ZSTDMT_getBuffer: created buffer of size %u", (U32)bSize);
        }
        return buffer;
    }
}

#if ZSTD_RESIZE_SEQPOOL
/** ZSTDMT_resizeBuffer() :
 * assumption : bufPool must be valid
 * @return : a buffer that is at least the buffer pool buffer size.
 *           If a reallocation happens, the data in the input buffer is copied.
 */
static Buffer ZSTDMT_resizeBuffer(ZSTDMT_bufferPool* bufPool, Buffer buffer)
{
    size_t const bSize = bufPool->bufferSize;
    if (buffer.capacity < bSize) {
        void* const start = ZSTD_customMalloc(bSize, bufPool->cMem);
        Buffer newBuffer;
        newBuffer.start = start;
        newBuffer.capacity = start == NULL ? 0 : bSize;
        if (start != NULL) {
            assert(newBuffer.capacity >= buffer.capacity);
            ZSTD_memcpy(newBuffer.start, buffer.start, buffer.capacity);
            DEBUGLOG(5, "ZSTDMT_resizeBuffer: created buffer of size %u", (U32)bSize);
            return newBuffer;
        }
        DEBUGLOG(5, "ZSTDMT_resizeBuffer: buffer allocation failure !!");
    }
    return buffer;
}
#endif

/* store buffer for later re-use, up to pool capacity */
static void ZSTDMT_releaseBuffer(ZSTDMT_bufferPool* bufPool, Buffer buf)
{
    DEBUGLOG(5, "ZSTDMT_releaseBuffer");
    if (buf.start == NULL) return;   /* compatible with release on NULL */
    ZSTD_pthread_mutex_lock(&bufPool->poolMutex);
    if (bufPool->nbBuffers < bufPool->totalBuffers) {
        bufPool->buffers[bufPool->nbBuffers++] = buf;  /* stored for later use */
        DEBUGLOG(5, "ZSTDMT_releaseBuffer: stored buffer of size %u in slot %u",
                    (U32)buf.capacity, (U32)(bufPool->nbBuffers-1));
        ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
        return;
    }
    ZSTD_pthread_mutex_unlock(&bufPool->poolMutex);
    /* Reached bufferPool capacity (note: should not happen) */
    DEBUGLOG(5, "ZSTDMT_releaseBuffer: pool capacity reached => freeing ");
    ZSTD_customFree(buf.start, bufPool->cMem);
}

/* We need 2 output buffers per worker since each dstBuff must be flushed after it is released.
 * The 3 additional buffers are as follows:
 *   1 buffer for input loading
 *   1 buffer for "next input" when submitting current one
 *   1 buffer stuck in queue */
#define BUF_POOL_MAX_NB_BUFFERS(nbWorkers) (2*(nbWorkers) + 3)

/* After a worker releases its rawSeqStore, it is immediately ready for reuse.
 * So we only need one seq buffer per worker. */
#define SEQ_POOL_MAX_NB_BUFFERS(nbWorkers) (nbWorkers)

/* =====   Seq Pool Wrapper   ====== */

typedef ZSTDMT_bufferPool ZSTDMT_seqPool;

static size_t ZSTDMT_sizeof_seqPool(ZSTDMT_seqPool* seqPool)
{
    return ZSTDMT_sizeof_bufferPool(seqPool);
}

static RawSeqStore_t bufferToSeq(Buffer buffer)
{
    RawSeqStore_t seq = kNullRawSeqStore;
    seq.seq = (rawSeq*)buffer.start;
    seq.capacity = buffer.capacity / sizeof(rawSeq);
    return seq;
}

static Buffer seqToBuffer(RawSeqStore_t seq)
{
    Buffer buffer;
    buffer.start = seq.seq;
    buffer.capacity = seq.capacity * sizeof(rawSeq);
    return buffer;
}

static RawSeqStore_t ZSTDMT_getSeq(ZSTDMT_seqPool* seqPool)
{
    if (seqPool->bufferSize == 0) {
        return kNullRawSeqStore;
    }
    return bufferToSeq(ZSTDMT_getBuffer(seqPool));
}

#if ZSTD_RESIZE_SEQPOOL
static RawSeqStore_t ZSTDMT_resizeSeq(ZSTDMT_seqPool* seqPool, RawSeqStore_t seq)
{
  return bufferToSeq(ZSTDMT_resizeBuffer(seqPool, seqToBuffer(seq)));
}
#endif

static void ZSTDMT_releaseSeq(ZSTDMT_seqPool* seqPool, RawSeqStore_t seq)
{
  ZSTDMT_releaseBuffer(seqPool, seqToBuffer(seq));
}

static void ZSTDMT_setNbSeq(ZSTDMT_seqPool* const seqPool, size_t const nbSeq)
{
  ZSTDMT_setBufferSize(seqPool, nbSeq * sizeof(rawSeq));
}

static ZSTDMT_seqPool* ZSTDMT_createSeqPool(unsigned nbWorkers, ZSTD_customMem cMem)
{
    ZSTDMT_seqPool* const seqPool = ZSTDMT_createBufferPool(SEQ_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
    if (seqPool == NULL) return NULL;
    ZSTDMT_setNbSeq(seqPool, 0);
    return seqPool;
}

static void ZSTDMT_freeSeqPool(ZSTDMT_seqPool* seqPool)
{
    ZSTDMT_freeBufferPool(seqPool);
}

static ZSTDMT_seqPool* ZSTDMT_expandSeqPool(ZSTDMT_seqPool* pool, U32 nbWorkers)
{
    return ZSTDMT_expandBufferPool(pool, SEQ_POOL_MAX_NB_BUFFERS(nbWorkers));
}


/* =====   CCtx Pool   ===== */
/* a single CCtx Pool can be invoked from multiple threads in parallel */

typedef struct {
    ZSTD_pthread_mutex_t poolMutex;
    int totalCCtx;
    int availCCtx;
    ZSTD_customMem cMem;
    ZSTD_CCtx** cctxs;
} ZSTDMT_CCtxPool;

/* note : all CCtx borrowed from the pool must be reverted back to the pool _before_ freeing the pool */
static void ZSTDMT_freeCCtxPool(ZSTDMT_CCtxPool* pool)
{
    if (!pool) return;
    ZSTD_pthread_mutex_destroy(&pool->poolMutex);
    if (pool->cctxs) {
        int cid;
        for (cid=0; cid<pool->totalCCtx; cid++)
            ZSTD_freeCCtx(pool->cctxs[cid]);  /* free compatible with NULL */
        ZSTD_customFree(pool->cctxs, pool->cMem);
    }
    ZSTD_customFree(pool, pool->cMem);
}

/* ZSTDMT_createCCtxPool() :
 * implies nbWorkers >= 1 , checked by caller ZSTDMT_createCCtx() */
static ZSTDMT_CCtxPool* ZSTDMT_createCCtxPool(int nbWorkers,
                                              ZSTD_customMem cMem)
{
    ZSTDMT_CCtxPool* const cctxPool =
        (ZSTDMT_CCtxPool*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtxPool), cMem);
    assert(nbWorkers > 0);
    if (!cctxPool) return NULL;
    if (ZSTD_pthread_mutex_init(&cctxPool->poolMutex, NULL)) {
        ZSTD_customFree(cctxPool, cMem);
        return NULL;
    }
    cctxPool->totalCCtx = nbWorkers;
    cctxPool->cctxs = (ZSTD_CCtx**)ZSTD_customCalloc(nbWorkers * sizeof(ZSTD_CCtx*), cMem);
    if (!cctxPool->cctxs) {
        ZSTDMT_freeCCtxPool(cctxPool);
        return NULL;
    }
    cctxPool->cMem = cMem;
    cctxPool->cctxs[0] = ZSTD_createCCtx_advanced(cMem);
    if (!cctxPool->cctxs[0]) { ZSTDMT_freeCCtxPool(cctxPool); return NULL; }
    cctxPool->availCCtx = 1;   /* at least one cctx for single-thread mode */
    DEBUGLOG(3, "cctxPool created, with %u workers", nbWorkers);
    return cctxPool;
}

static ZSTDMT_CCtxPool* ZSTDMT_expandCCtxPool(ZSTDMT_CCtxPool* srcPool,
                                              int nbWorkers)
{
    if (srcPool==NULL) return NULL;
    if (nbWorkers <= srcPool->totalCCtx) return srcPool;   /* good enough */
    /* need a larger cctx pool */
    {   ZSTD_customMem const cMem = srcPool->cMem;
        ZSTDMT_freeCCtxPool(srcPool);
        return ZSTDMT_createCCtxPool(nbWorkers, cMem);
    }
}

/* only works during initialization phase, not during compression */
static size_t ZSTDMT_sizeof_CCtxPool(ZSTDMT_CCtxPool* cctxPool)
{
    ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
    {   unsigned const nbWorkers = cctxPool->totalCCtx;
        size_t const poolSize = sizeof(*cctxPool);
        size_t const arraySize = cctxPool->totalCCtx * sizeof(ZSTD_CCtx*);
        size_t totalCCtxSize = 0;
        unsigned u;
        for (u=0; u<nbWorkers; u++) {
            totalCCtxSize += ZSTD_sizeof_CCtx(cctxPool->cctxs[u]);
        }
        ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
        assert(nbWorkers > 0);
        return poolSize + arraySize + totalCCtxSize;
    }
}

static ZSTD_CCtx* ZSTDMT_getCCtx(ZSTDMT_CCtxPool* cctxPool)
{
    DEBUGLOG(5, "ZSTDMT_getCCtx");
    ZSTD_pthread_mutex_lock(&cctxPool->poolMutex);
    if (cctxPool->availCCtx) {
        cctxPool->availCCtx--;
        {   ZSTD_CCtx* const cctx = cctxPool->cctxs[cctxPool->availCCtx];
            ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
            return cctx;
    }   }
    ZSTD_pthread_mutex_unlock(&cctxPool->poolMutex);
    DEBUGLOG(5, "create one more CCtx");
    return ZSTD_createCCtx_advanced(cctxPool->cMem);   /* note : can be NULL, when creation fails ! */
}

static void ZSTDMT_releaseCCtx(ZSTDMT_CCtxPool* pool, ZSTD_CCtx* cctx)
{
    if (cctx==NULL) return;   /* compatibility with release on NULL */
    ZSTD_pthread_mutex_lock(&pool->poolMutex);
    if (pool->availCCtx < pool->totalCCtx)
        pool->cctxs[pool->availCCtx++] = cctx;
    else {
        /* pool overflow : should not happen, since totalCCtx==nbWorkers */
        DEBUGLOG(4, "CCtx pool overflow : free cctx");
        ZSTD_freeCCtx(cctx);
    }
    ZSTD_pthread_mutex_unlock(&pool->poolMutex);
}

/* ====   Serial State   ==== */

typedef struct {
    void const* start;
    size_t size;
} Range;

typedef struct {
    /* All variables in the struct are protected by mutex. */
    ZSTD_pthread_mutex_t mutex;
    ZSTD_pthread_cond_t cond;
    ZSTD_CCtx_params params;
    ldmState_t ldmState;
    XXH64_state_t xxhState;
    unsigned nextJobID;
    /* Protects ldmWindow.
     * Must be acquired after the main mutex when acquiring both.
     */
    ZSTD_pthread_mutex_t ldmWindowMutex;
    ZSTD_pthread_cond_t ldmWindowCond;  /* Signaled when ldmWindow is updated */
    ZSTD_window_t ldmWindow;  /* A thread-safe copy of ldmState.window */
} SerialState;

static int
ZSTDMT_serialState_reset(SerialState* serialState,
                         ZSTDMT_seqPool* seqPool,
                         ZSTD_CCtx_params params,
                         size_t jobSize,
                         const void* dict, size_t const dictSize,
                         ZSTD_dictContentType_e dictContentType)
{
    /* Adjust parameters */
    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
        DEBUGLOG(4, "LDM window size = %u KB", (1U << params.cParams.windowLog) >> 10);
        ZSTD_ldm_adjustParameters(&params.ldmParams, &params.cParams);
        assert(params.ldmParams.hashLog >= params.ldmParams.bucketSizeLog);
        assert(params.ldmParams.hashRateLog < 32);
    } else {
        ZSTD_memset(&params.ldmParams, 0, sizeof(params.ldmParams));
    }
    serialState->nextJobID = 0;
    if (params.fParams.checksumFlag)
        XXH64_reset(&serialState->xxhState, 0);
    if (params.ldmParams.enableLdm == ZSTD_ps_enable) {
        ZSTD_customMem cMem = params.customMem;
        unsigned const hashLog = params.ldmParams.hashLog;
        size_t const hashSize = ((size_t)1 << hashLog) * sizeof(ldmEntry_t);
        unsigned const bucketLog =
            params.ldmParams.hashLog - params.ldmParams.bucketSizeLog;
        unsigned const prevBucketLog =
            serialState->params.ldmParams.hashLog -
            serialState->params.ldmParams.bucketSizeLog;
        size_t const numBuckets = (size_t)1 << bucketLog;
        /* Size the seq pool tables */
        ZSTDMT_setNbSeq(seqPool, ZSTD_ldm_getMaxNbSeq(params.ldmParams, jobSize));
        /* Reset the window */
        ZSTD_window_init(&serialState->ldmState.window);
        /* Resize tables and output space if necessary. */
        if (serialState->ldmState.hashTable == NULL || serialState->params.ldmParams.hashLog < hashLog) {
            ZSTD_customFree(serialState->ldmState.hashTable, cMem);
            serialState->ldmState.hashTable = (ldmEntry_t*)ZSTD_customMalloc(hashSize, cMem);
        }
        if (serialState->ldmState.bucketOffsets == NULL || prevBucketLog < bucketLog) {
            ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
            serialState->ldmState.bucketOffsets = (BYTE*)ZSTD_customMalloc(numBuckets, cMem);
        }
        if (!serialState->ldmState.hashTable || !serialState->ldmState.bucketOffsets)
            return 1;
        /* Zero the tables */
        ZSTD_memset(serialState->ldmState.hashTable, 0, hashSize);
        ZSTD_memset(serialState->ldmState.bucketOffsets, 0, numBuckets);

        /* Update window state and fill hash table with dict */
        serialState->ldmState.loadedDictEnd = 0;
        if (dictSize > 0) {
            if (dictContentType == ZSTD_dct_rawContent) {
                BYTE const* const dictEnd = (const BYTE*)dict + dictSize;
                ZSTD_window_update(&serialState->ldmState.window, dict, dictSize, /* forceNonContiguous */ 0);
                ZSTD_ldm_fillHashTable(&serialState->ldmState, (const BYTE*)dict, dictEnd, &params.ldmParams);
                serialState->ldmState.loadedDictEnd = params.forceWindow ? 0 : (U32)(dictEnd - serialState->ldmState.window.base);
            } else {
                /* don't even load anything */
            }
        }

        /* Initialize serialState's copy of ldmWindow. */
        serialState->ldmWindow = serialState->ldmState.window;
    }

    serialState->params = params;
    serialState->params.jobSize = (U32)jobSize;
    return 0;
}

static int ZSTDMT_serialState_init(SerialState* serialState)
{
    int initError = 0;
    ZSTD_memset(serialState, 0, sizeof(*serialState));
    initError |= ZSTD_pthread_mutex_init(&serialState->mutex, NULL);
    initError |= ZSTD_pthread_cond_init(&serialState->cond, NULL);
    initError |= ZSTD_pthread_mutex_init(&serialState->ldmWindowMutex, NULL);
    initError |= ZSTD_pthread_cond_init(&serialState->ldmWindowCond, NULL);
    return initError;
}

static void ZSTDMT_serialState_free(SerialState* serialState)
{
    ZSTD_customMem cMem = serialState->params.customMem;
    ZSTD_pthread_mutex_destroy(&serialState->mutex);
    ZSTD_pthread_cond_destroy(&serialState->cond);
    ZSTD_pthread_mutex_destroy(&serialState->ldmWindowMutex);
    ZSTD_pthread_cond_destroy(&serialState->ldmWindowCond);
    ZSTD_customFree(serialState->ldmState.hashTable, cMem);
    ZSTD_customFree(serialState->ldmState.bucketOffsets, cMem);
}

static void
ZSTDMT_serialState_genSequences(SerialState* serialState,
                                RawSeqStore_t* seqStore,
                                Range src, unsigned jobID)
{
    /* Wait for our turn */
    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
    while (serialState->nextJobID < jobID) {
        DEBUGLOG(5, "wait for serialState->cond");
        ZSTD_pthread_cond_wait(&serialState->cond, &serialState->mutex);
    }
    /* A future job may error and skip our job */
    if (serialState->nextJobID == jobID) {
        /* It is now our turn, do any processing necessary */
        if (serialState->params.ldmParams.enableLdm == ZSTD_ps_enable) {
            size_t error;
            DEBUGLOG(6, "ZSTDMT_serialState_genSequences: LDM update");
            assert(seqStore->seq != NULL && seqStore->pos == 0 &&
                   seqStore->size == 0 && seqStore->capacity > 0);
            assert(src.size <= serialState->params.jobSize);
            ZSTD_window_update(&serialState->ldmState.window, src.start, src.size, /* forceNonContiguous */ 0);
            error = ZSTD_ldm_generateSequences(
                &serialState->ldmState, seqStore,
                &serialState->params.ldmParams, src.start, src.size);
            /* We provide a large enough buffer to never fail. */
            assert(!ZSTD_isError(error)); (void)error;
            /* Update ldmWindow to match the ldmState.window and signal the main
             * thread if it is waiting for a buffer.
             */
            ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
            serialState->ldmWindow = serialState->ldmState.window;
            ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
            ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
        }
        if (serialState->params.fParams.checksumFlag && src.size > 0)
            XXH64_update(&serialState->xxhState, src.start, src.size);
    }
    /* Now it is the next jobs turn */
    serialState->nextJobID++;
    ZSTD_pthread_cond_broadcast(&serialState->cond);
    ZSTD_pthread_mutex_unlock(&serialState->mutex);
}

static void
ZSTDMT_serialState_applySequences(const SerialState* serialState, /* just for an assert() check */
                                  ZSTD_CCtx* jobCCtx,
                                  const RawSeqStore_t* seqStore)
{
    if (seqStore->size > 0) {
        DEBUGLOG(5, "ZSTDMT_serialState_applySequences: uploading %u external sequences", (unsigned)seqStore->size);
        assert(serialState->params.ldmParams.enableLdm == ZSTD_ps_enable); (void)serialState;
        assert(jobCCtx);
        ZSTD_referenceExternalSequences(jobCCtx, seqStore->seq, seqStore->size);
    }
}

static void ZSTDMT_serialState_ensureFinished(SerialState* serialState,
                                              unsigned jobID, size_t cSize)
{
    ZSTD_PTHREAD_MUTEX_LOCK(&serialState->mutex);
    if (serialState->nextJobID <= jobID) {
        assert(ZSTD_isError(cSize)); (void)cSize;
        DEBUGLOG(5, "Skipping past job %u because of error", jobID);
        serialState->nextJobID = jobID + 1;
        ZSTD_pthread_cond_broadcast(&serialState->cond);

        ZSTD_PTHREAD_MUTEX_LOCK(&serialState->ldmWindowMutex);
        ZSTD_window_clear(&serialState->ldmWindow);
        ZSTD_pthread_cond_signal(&serialState->ldmWindowCond);
        ZSTD_pthread_mutex_unlock(&serialState->ldmWindowMutex);
    }
    ZSTD_pthread_mutex_unlock(&serialState->mutex);

}


/* ------------------------------------------ */
/* =====          Worker thread         ===== */
/* ------------------------------------------ */

static const Range kNullRange = { NULL, 0 };

typedef struct {
    size_t   consumed;                 /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx */
    size_t   cSize;                    /* SHARED - set0 by mtctx, then modified by worker AND read by mtctx, then set0 by mtctx */
    ZSTD_pthread_mutex_t job_mutex;    /* Thread-safe - used by mtctx and worker */
    ZSTD_pthread_cond_t job_cond;      /* Thread-safe - used by mtctx and worker */
    ZSTDMT_CCtxPool* cctxPool;         /* Thread-safe - used by mtctx and (all) workers */
    ZSTDMT_bufferPool* bufPool;        /* Thread-safe - used by mtctx and (all) workers */
    ZSTDMT_seqPool* seqPool;           /* Thread-safe - used by mtctx and (all) workers */
    SerialState* serial;               /* Thread-safe - used by mtctx and (all) workers */
    Buffer dstBuff;                    /* set by worker (or mtctx), then read by worker & mtctx, then modified by mtctx => no barrier */
    Range prefix;                      /* set by mtctx, then read by worker & mtctx => no barrier */
    Range src;                         /* set by mtctx, then read by worker & mtctx => no barrier */
    unsigned jobID;                    /* set by mtctx, then read by worker => no barrier */
    unsigned firstJob;                 /* set by mtctx, then read by worker => no barrier */
    unsigned lastJob;                  /* set by mtctx, then read by worker => no barrier */
    ZSTD_CCtx_params params;           /* set by mtctx, then read by worker => no barrier */
    const ZSTD_CDict* cdict;           /* set by mtctx, then read by worker => no barrier */
    unsigned long long fullFrameSize;  /* set by mtctx, then read by worker => no barrier */
    size_t   dstFlushed;               /* used only by mtctx */
    unsigned frameChecksumNeeded;      /* used only by mtctx */
} ZSTDMT_jobDescription;

#define JOB_ERROR(e)                                \
    do {                                            \
        ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);   \
        job->cSize = e;                             \
        ZSTD_pthread_mutex_unlock(&job->job_mutex); \
        goto _endJob;                               \
    } while (0)

/* ZSTDMT_compressionJob() is a POOL_function type */
static void ZSTDMT_compressionJob(void* jobDescription)
{
    ZSTDMT_jobDescription* const job = (ZSTDMT_jobDescription*)jobDescription;
    ZSTD_CCtx_params jobParams = job->params;   /* do not modify job->params ! copy it, modify the copy */
    ZSTD_CCtx* const cctx = ZSTDMT_getCCtx(job->cctxPool);
    RawSeqStore_t rawSeqStore = ZSTDMT_getSeq(job->seqPool);
    Buffer dstBuff = job->dstBuff;
    size_t lastCBlockSize = 0;

    DEBUGLOG(5, "ZSTDMT_compressionJob: job %u", job->jobID);
    /* resources */
    if (cctx==NULL) JOB_ERROR(ERROR(memory_allocation));
    if (dstBuff.start == NULL) {   /* streaming job : doesn't provide a dstBuffer */
        dstBuff = ZSTDMT_getBuffer(job->bufPool);
        if (dstBuff.start==NULL) JOB_ERROR(ERROR(memory_allocation));
        job->dstBuff = dstBuff;   /* this value can be read in ZSTDMT_flush, when it copies the whole job */
    }
    if (jobParams.ldmParams.enableLdm == ZSTD_ps_enable && rawSeqStore.seq == NULL)
        JOB_ERROR(ERROR(memory_allocation));

    /* Don't compute the checksum for chunks, since we compute it externally,
     * but write it in the header.
     */
    if (job->jobID != 0) jobParams.fParams.checksumFlag = 0;
    /* Don't run LDM for the chunks, since we handle it externally */
    jobParams.ldmParams.enableLdm = ZSTD_ps_disable;
    /* Correct nbWorkers to 0. */
    jobParams.nbWorkers = 0;


    /* init */

    /* Perform serial step as early as possible */
    ZSTDMT_serialState_genSequences(job->serial, &rawSeqStore, job->src, job->jobID);

    if (job->cdict) {
        size_t const initError = ZSTD_compressBegin_advanced_internal(cctx, NULL, 0, ZSTD_dct_auto, ZSTD_dtlm_fast, job->cdict, &jobParams, job->fullFrameSize);
        assert(job->firstJob);  /* only allowed for first job */
        if (ZSTD_isError(initError)) JOB_ERROR(initError);
    } else {
        U64 const pledgedSrcSize = job->firstJob ? job->fullFrameSize : job->src.size;
        {   size_t const forceWindowError = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_forceMaxWindow, !job->firstJob);
            if (ZSTD_isError(forceWindowError)) JOB_ERROR(forceWindowError);
        }
        if (!job->firstJob) {
            size_t const err = ZSTD_CCtxParams_setParameter(&jobParams, ZSTD_c_deterministicRefPrefix, 0);
            if (ZSTD_isError(err)) JOB_ERROR(err);
        }
        DEBUGLOG(6, "ZSTDMT_compressionJob: job %u: loading prefix of size %zu", job->jobID, job->prefix.size);
        {   size_t const initError = ZSTD_compressBegin_advanced_internal(cctx,
                                        job->prefix.start, job->prefix.size, ZSTD_dct_rawContent,
                                        ZSTD_dtlm_fast,
                                        NULL, /*cdict*/
                                        &jobParams, pledgedSrcSize);
            if (ZSTD_isError(initError)) JOB_ERROR(initError);
    }   }

    /* External Sequences can only be applied after CCtx initialization */
    ZSTDMT_serialState_applySequences(job->serial, cctx, &rawSeqStore);

    if (!job->firstJob) {  /* flush and overwrite frame header when it's not first job */
        size_t const hSize = ZSTD_compressContinue_public(cctx, dstBuff.start, dstBuff.capacity, job->src.start, 0);
        if (ZSTD_isError(hSize)) JOB_ERROR(hSize);
        DEBUGLOG(5, "ZSTDMT_compressionJob: flush and overwrite %u bytes of frame header (not first job)", (U32)hSize);
        ZSTD_invalidateRepCodes(cctx);
    }

    /* compress the entire job by smaller chunks, for better granularity */
    {   size_t const chunkSize = 4*ZSTD_BLOCKSIZE_MAX;
        int const nbChunks = (int)((job->src.size + (chunkSize-1)) / chunkSize);
        const BYTE* ip = (const BYTE*) job->src.start;
        BYTE* const ostart = (BYTE*)dstBuff.start;
        BYTE* op = ostart;
        BYTE* oend = op + dstBuff.capacity;
        int chunkNb;
        if (sizeof(size_t) > sizeof(int)) assert(job->src.size < ((size_t)INT_MAX) * chunkSize);   /* check overflow */
        DEBUGLOG(5, "ZSTDMT_compressionJob: compress %u bytes in %i blocks", (U32)job->src.size, nbChunks);
        assert(job->cSize == 0);
        for (chunkNb = 1; chunkNb < nbChunks; chunkNb++) {
            size_t const cSize = ZSTD_compressContinue_public(cctx, op, oend-op, ip, chunkSize);
            if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
            ip += chunkSize;
            op += cSize; assert(op < oend);
            /* stats */
            ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
            job->cSize += cSize;
            job->consumed = chunkSize * chunkNb;
            DEBUGLOG(5, "ZSTDMT_compressionJob: compress new block : cSize==%u bytes (total: %u)",
                        (U32)cSize, (U32)job->cSize);
            ZSTD_pthread_cond_signal(&job->job_cond);   /* warns some more data is ready to be flushed */
            ZSTD_pthread_mutex_unlock(&job->job_mutex);
        }
        /* last block */
        assert(chunkSize > 0);
        assert((chunkSize & (chunkSize - 1)) == 0);  /* chunkSize must be power of 2 for mask==(chunkSize-1) to work */
        if ((nbChunks > 0) | job->lastJob /*must output a "last block" flag*/ ) {
            size_t const lastBlockSize1 = job->src.size & (chunkSize-1);
            size_t const lastBlockSize = ((lastBlockSize1==0) & (job->src.size>=chunkSize)) ? chunkSize : lastBlockSize1;
            size_t const cSize = (job->lastJob) ?
                 ZSTD_compressEnd_public(cctx, op, oend-op, ip, lastBlockSize) :
                 ZSTD_compressContinue_public(cctx, op, oend-op, ip, lastBlockSize);
            if (ZSTD_isError(cSize)) JOB_ERROR(cSize);
            lastCBlockSize = cSize;
    }   }
    if (!job->firstJob) {
        /* Double check that we don't have an ext-dict, because then our
         * repcode invalidation doesn't work.
         */
        assert(!ZSTD_window_hasExtDict(cctx->blockState.matchState.window));
    }
    ZSTD_CCtx_trace(cctx, 0);

_endJob:
    ZSTDMT_serialState_ensureFinished(job->serial, job->jobID, job->cSize);
    if (job->prefix.size > 0)
        DEBUGLOG(5, "Finished with prefix: %zx", (size_t)job->prefix.start);
    DEBUGLOG(5, "Finished with source: %zx", (size_t)job->src.start);
    /* release resources */
    ZSTDMT_releaseSeq(job->seqPool, rawSeqStore);
    ZSTDMT_releaseCCtx(job->cctxPool, cctx);
    /* report */
    ZSTD_PTHREAD_MUTEX_LOCK(&job->job_mutex);
    if (ZSTD_isError(job->cSize)) assert(lastCBlockSize == 0);
    job->cSize += lastCBlockSize;
    job->consumed = job->src.size;  /* when job->consumed == job->src.size , compression job is presumed completed */
    ZSTD_pthread_cond_signal(&job->job_cond);
    ZSTD_pthread_mutex_unlock(&job->job_mutex);
}


/* ------------------------------------------ */
/* =====   Multi-threaded compression   ===== */
/* ------------------------------------------ */

typedef struct {
    Range prefix;         /* read-only non-owned prefix buffer */
    Buffer buffer;
    size_t filled;
} InBuff_t;

typedef struct {
  BYTE* buffer;     /* The round input buffer. All jobs get references
                     * to pieces of the buffer. ZSTDMT_tryGetInputRange()
                     * handles handing out job input buffers, and makes
                     * sure it doesn't overlap with any pieces still in use.
                     */
  size_t capacity;  /* The capacity of buffer. */
  size_t pos;       /* The position of the current inBuff in the round
                     * buffer. Updated past the end if the inBuff once
                     * the inBuff is sent to the worker thread.
                     * pos <= capacity.
                     */
} RoundBuff_t;

static const RoundBuff_t kNullRoundBuff = {NULL, 0, 0};

#define RSYNC_LENGTH 32
/* Don't create chunks smaller than the zstd block size.
 * This stops us from regressing compression ratio too much,
 * and ensures our output fits in ZSTD_compressBound().
 *
 * If this is shrunk < ZSTD_BLOCKSIZELOG_MIN then
 * ZSTD_COMPRESSBOUND() will need to be updated.
 */
#define RSYNC_MIN_BLOCK_LOG ZSTD_BLOCKSIZELOG_MAX
#define RSYNC_MIN_BLOCK_SIZE (1<<RSYNC_MIN_BLOCK_LOG)

typedef struct {
  U64 hash;
  U64 hitMask;
  U64 primePower;
} RSyncState_t;

struct ZSTDMT_CCtx_s {
    POOL_ctx* factory;
    ZSTDMT_jobDescription* jobs;
    ZSTDMT_bufferPool* bufPool;
    ZSTDMT_CCtxPool* cctxPool;
    ZSTDMT_seqPool* seqPool;
    ZSTD_CCtx_params params;
    size_t targetSectionSize;
    size_t targetPrefixSize;
    int jobReady;        /* 1 => one job is already prepared, but pool has shortage of workers. Don't create a new job. */
    InBuff_t inBuff;
    RoundBuff_t roundBuff;
    SerialState serial;
    RSyncState_t rsync;
    unsigned jobIDMask;
    unsigned doneJobID;
    unsigned nextJobID;
    unsigned frameEnded;
    unsigned allJobsCompleted;
    unsigned long long frameContentSize;
    unsigned long long consumed;
    unsigned long long produced;
    ZSTD_customMem cMem;
    ZSTD_CDict* cdictLocal;
    const ZSTD_CDict* cdict;
    unsigned providedFactory: 1;
};

static void ZSTDMT_freeJobsTable(ZSTDMT_jobDescription* jobTable, U32 nbJobs, ZSTD_customMem cMem)
{
    U32 jobNb;
    if (jobTable == NULL) return;
    for (jobNb=0; jobNb<nbJobs; jobNb++) {
        ZSTD_pthread_mutex_destroy(&jobTable[jobNb].job_mutex);
        ZSTD_pthread_cond_destroy(&jobTable[jobNb].job_cond);
    }
    ZSTD_customFree(jobTable, cMem);
}

/* ZSTDMT_allocJobsTable()
 * allocate and init a job table.
 * update *nbJobsPtr to next power of 2 value, as size of table */
static ZSTDMT_jobDescription* ZSTDMT_createJobsTable(U32* nbJobsPtr, ZSTD_customMem cMem)
{
    U32 const nbJobsLog2 = ZSTD_highbit32(*nbJobsPtr) + 1;
    U32 const nbJobs = 1 << nbJobsLog2;
    U32 jobNb;
    ZSTDMT_jobDescription* const jobTable = (ZSTDMT_jobDescription*)
                ZSTD_customCalloc(nbJobs * sizeof(ZSTDMT_jobDescription), cMem);
    int initError = 0;
    if (jobTable==NULL) return NULL;
    *nbJobsPtr = nbJobs;
    for (jobNb=0; jobNb<nbJobs; jobNb++) {
        initError |= ZSTD_pthread_mutex_init(&jobTable[jobNb].job_mutex, NULL);
        initError |= ZSTD_pthread_cond_init(&jobTable[jobNb].job_cond, NULL);
    }
    if (initError != 0) {
        ZSTDMT_freeJobsTable(jobTable, nbJobs, cMem);
        return NULL;
    }
    return jobTable;
}

static size_t ZSTDMT_expandJobsTable (ZSTDMT_CCtx* mtctx, U32 nbWorkers) {
    U32 nbJobs = nbWorkers + 2;
    if (nbJobs > mtctx->jobIDMask+1) {  /* need more job capacity */
        ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
        mtctx->jobIDMask = 0;
        mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, mtctx->cMem);
        if (mtctx->jobs==NULL) return ERROR(memory_allocation);
        assert((nbJobs != 0) && ((nbJobs & (nbJobs - 1)) == 0));  /* ensure nbJobs is a power of 2 */
        mtctx->jobIDMask = nbJobs - 1;
    }
    return 0;
}


/* ZSTDMT_CCtxParam_setNbWorkers():
 * Internal use only */
static size_t ZSTDMT_CCtxParam_setNbWorkers(ZSTD_CCtx_params* params, unsigned nbWorkers)
{
    return ZSTD_CCtxParams_setParameter(params, ZSTD_c_nbWorkers, (int)nbWorkers);
}

MEM_STATIC ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced_internal(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
{
    ZSTDMT_CCtx* mtctx;
    U32 nbJobs = nbWorkers + 2;
    int initError;
    DEBUGLOG(3, "ZSTDMT_createCCtx_advanced (nbWorkers = %u)", nbWorkers);

    if (nbWorkers < 1) return NULL;
    nbWorkers = MIN(nbWorkers , ZSTDMT_NBWORKERS_MAX);
    if ((cMem.customAlloc!=NULL) ^ (cMem.customFree!=NULL))
        /* invalid custom allocator */
        return NULL;

    mtctx = (ZSTDMT_CCtx*) ZSTD_customCalloc(sizeof(ZSTDMT_CCtx), cMem);
    if (!mtctx) return NULL;
    ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
    mtctx->cMem = cMem;
    mtctx->allJobsCompleted = 1;
    if (pool != NULL) {
      mtctx->factory = pool;
      mtctx->providedFactory = 1;
    }
    else {
      mtctx->factory = POOL_create_advanced(nbWorkers, 0, cMem);
      mtctx->providedFactory = 0;
    }
    mtctx->jobs = ZSTDMT_createJobsTable(&nbJobs, cMem);
    assert(nbJobs > 0); assert((nbJobs & (nbJobs - 1)) == 0);  /* ensure nbJobs is a power of 2 */
    mtctx->jobIDMask = nbJobs - 1;
    mtctx->bufPool = ZSTDMT_createBufferPool(BUF_POOL_MAX_NB_BUFFERS(nbWorkers), cMem);
    mtctx->cctxPool = ZSTDMT_createCCtxPool(nbWorkers, cMem);
    mtctx->seqPool = ZSTDMT_createSeqPool(nbWorkers, cMem);
    initError = ZSTDMT_serialState_init(&mtctx->serial);
    mtctx->roundBuff = kNullRoundBuff;
    if (!mtctx->factory | !mtctx->jobs | !mtctx->bufPool | !mtctx->cctxPool | !mtctx->seqPool | initError) {
        ZSTDMT_freeCCtx(mtctx);
        return NULL;
    }
    DEBUGLOG(3, "mt_cctx created, for %u threads", nbWorkers);
    return mtctx;
}

ZSTDMT_CCtx* ZSTDMT_createCCtx_advanced(unsigned nbWorkers, ZSTD_customMem cMem, ZSTD_threadPool* pool)
{
#ifdef ZSTD_MULTITHREAD
    return ZSTDMT_createCCtx_advanced_internal(nbWorkers, cMem, pool);
#else
    (void)nbWorkers;
    (void)cMem;
    (void)pool;
    return NULL;
#endif
}


/* ZSTDMT_releaseAllJobResources() :
 * note : ensure all workers are killed first ! */
static void ZSTDMT_releaseAllJobResources(ZSTDMT_CCtx* mtctx)
{
    unsigned jobID;
    DEBUGLOG(3, "ZSTDMT_releaseAllJobResources");
    for (jobID=0; jobID <= mtctx->jobIDMask; jobID++) {
        /* Copy the mutex/cond out */
        ZSTD_pthread_mutex_t const mutex = mtctx->jobs[jobID].job_mutex;
        ZSTD_pthread_cond_t const cond = mtctx->jobs[jobID].job_cond;

        DEBUGLOG(4, "job%02u: release dst address %08X", jobID, (U32)(size_t)mtctx->jobs[jobID].dstBuff.start);
        ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[jobID].dstBuff);

        /* Clear the job description, but keep the mutex/cond */
        ZSTD_memset(&mtctx->jobs[jobID], 0, sizeof(mtctx->jobs[jobID]));
        mtctx->jobs[jobID].job_mutex = mutex;
        mtctx->jobs[jobID].job_cond = cond;
    }
    mtctx->inBuff.buffer = g_nullBuffer;
    mtctx->inBuff.filled = 0;
    mtctx->allJobsCompleted = 1;
}

static void ZSTDMT_waitForAllJobsCompleted(ZSTDMT_CCtx* mtctx)
{
    DEBUGLOG(4, "ZSTDMT_waitForAllJobsCompleted");
    while (mtctx->doneJobID < mtctx->nextJobID) {
        unsigned const jobID = mtctx->doneJobID & mtctx->jobIDMask;
        ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[jobID].job_mutex);
        while (mtctx->jobs[jobID].consumed < mtctx->jobs[jobID].src.size) {
            DEBUGLOG(4, "waiting for jobCompleted signal from job %u", mtctx->doneJobID);   /* we want to block when waiting for data to flush */
            ZSTD_pthread_cond_wait(&mtctx->jobs[jobID].job_cond, &mtctx->jobs[jobID].job_mutex);
        }
        ZSTD_pthread_mutex_unlock(&mtctx->jobs[jobID].job_mutex);
        mtctx->doneJobID++;
    }
}

size_t ZSTDMT_freeCCtx(ZSTDMT_CCtx* mtctx)
{
    if (mtctx==NULL) return 0;   /* compatible with free on NULL */
    if (!mtctx->providedFactory)
        POOL_free(mtctx->factory);   /* stop and free worker threads */
    ZSTDMT_releaseAllJobResources(mtctx);  /* release job resources into pools first */
    ZSTDMT_freeJobsTable(mtctx->jobs, mtctx->jobIDMask+1, mtctx->cMem);
    ZSTDMT_freeBufferPool(mtctx->bufPool);
    ZSTDMT_freeCCtxPool(mtctx->cctxPool);
    ZSTDMT_freeSeqPool(mtctx->seqPool);
    ZSTDMT_serialState_free(&mtctx->serial);
    ZSTD_freeCDict(mtctx->cdictLocal);
    if (mtctx->roundBuff.buffer)
        ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
    ZSTD_customFree(mtctx, mtctx->cMem);
    return 0;
}

size_t ZSTDMT_sizeof_CCtx(ZSTDMT_CCtx* mtctx)
{
    if (mtctx == NULL) return 0;   /* supports sizeof NULL */
    return sizeof(*mtctx)
            + POOL_sizeof(mtctx->factory)
            + ZSTDMT_sizeof_bufferPool(mtctx->bufPool)
            + (mtctx->jobIDMask+1) * sizeof(ZSTDMT_jobDescription)
            + ZSTDMT_sizeof_CCtxPool(mtctx->cctxPool)
            + ZSTDMT_sizeof_seqPool(mtctx->seqPool)
            + ZSTD_sizeof_CDict(mtctx->cdictLocal)
            + mtctx->roundBuff.capacity;
}


/* ZSTDMT_resize() :
 * @return : error code if fails, 0 on success */
static size_t ZSTDMT_resize(ZSTDMT_CCtx* mtctx, unsigned nbWorkers)
{
    if (POOL_resize(mtctx->factory, nbWorkers)) return ERROR(memory_allocation);
    FORWARD_IF_ERROR( ZSTDMT_expandJobsTable(mtctx, nbWorkers) , "");
    mtctx->bufPool = ZSTDMT_expandBufferPool(mtctx->bufPool, BUF_POOL_MAX_NB_BUFFERS(nbWorkers));
    if (mtctx->bufPool == NULL) return ERROR(memory_allocation);
    mtctx->cctxPool = ZSTDMT_expandCCtxPool(mtctx->cctxPool, nbWorkers);
    if (mtctx->cctxPool == NULL) return ERROR(memory_allocation);
    mtctx->seqPool = ZSTDMT_expandSeqPool(mtctx->seqPool, nbWorkers);
    if (mtctx->seqPool == NULL) return ERROR(memory_allocation);
    ZSTDMT_CCtxParam_setNbWorkers(&mtctx->params, nbWorkers);
    return 0;
}


/*! ZSTDMT_updateCParams_whileCompressing() :
 *  Updates a selected set of compression parameters, remaining compatible with currently active frame.
 *  New parameters will be applied to next compression job. */
void ZSTDMT_updateCParams_whileCompressing(ZSTDMT_CCtx* mtctx, const ZSTD_CCtx_params* cctxParams)
{
    U32 const saved_wlog = mtctx->params.cParams.windowLog;   /* Do not modify windowLog while compressing */
    int const compressionLevel = cctxParams->compressionLevel;
    DEBUGLOG(5, "ZSTDMT_updateCParams_whileCompressing (level:%i)",
                compressionLevel);
    mtctx->params.compressionLevel = compressionLevel;
    {   ZSTD_compressionParameters cParams = ZSTD_getCParamsFromCCtxParams(cctxParams, ZSTD_CONTENTSIZE_UNKNOWN, 0, ZSTD_cpm_noAttachDict);
        cParams.windowLog = saved_wlog;
        mtctx->params.cParams = cParams;
    }
}

/* ZSTDMT_getFrameProgression():
 * tells how much data has been consumed (input) and produced (output) for current frame.
 * able to count progression inside worker threads.
 * Note : mutex will be acquired during statistics collection inside workers. */
ZSTD_frameProgression ZSTDMT_getFrameProgression(ZSTDMT_CCtx* mtctx)
{
    ZSTD_frameProgression fps;
    DEBUGLOG(5, "ZSTDMT_getFrameProgression");
    fps.ingested = mtctx->consumed + mtctx->inBuff.filled;
    fps.consumed = mtctx->consumed;
    fps.produced = fps.flushed = mtctx->produced;
    fps.currentJobID = mtctx->nextJobID;
    fps.nbActiveWorkers = 0;
    {   unsigned jobNb;
        unsigned lastJobNb = mtctx->nextJobID + mtctx->jobReady; assert(mtctx->jobReady <= 1);
        DEBUGLOG(6, "ZSTDMT_getFrameProgression: jobs: from %u to <%u (jobReady:%u)",
                    mtctx->doneJobID, lastJobNb, mtctx->jobReady);
        for (jobNb = mtctx->doneJobID ; jobNb < lastJobNb ; jobNb++) {
            unsigned const wJobID = jobNb & mtctx->jobIDMask;
            ZSTDMT_jobDescription* jobPtr = &mtctx->jobs[wJobID];
            ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
            {   size_t const cResult = jobPtr->cSize;
                size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
                size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
                assert(flushed <= produced);
                fps.ingested += jobPtr->src.size;
                fps.consumed += jobPtr->consumed;
                fps.produced += produced;
                fps.flushed  += flushed;
                fps.nbActiveWorkers += (jobPtr->consumed < jobPtr->src.size);
            }
            ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
        }
    }
    return fps;
}


size_t ZSTDMT_toFlushNow(ZSTDMT_CCtx* mtctx)
{
    size_t toFlush;
    unsigned const jobID = mtctx->doneJobID;
    assert(jobID <= mtctx->nextJobID);
    if (jobID == mtctx->nextJobID) return 0;   /* no active job => nothing to flush */

    /* look into oldest non-fully-flushed job */
    {   unsigned const wJobID = jobID & mtctx->jobIDMask;
        ZSTDMT_jobDescription* const jobPtr = &mtctx->jobs[wJobID];
        ZSTD_pthread_mutex_lock(&jobPtr->job_mutex);
        {   size_t const cResult = jobPtr->cSize;
            size_t const produced = ZSTD_isError(cResult) ? 0 : cResult;
            size_t const flushed = ZSTD_isError(cResult) ? 0 : jobPtr->dstFlushed;
            assert(flushed <= produced);
            assert(jobPtr->consumed <= jobPtr->src.size);
            toFlush = produced - flushed;
            /* if toFlush==0, nothing is available to flush.
             * However, jobID is expected to still be active:
             * if jobID was already completed and fully flushed,
             * ZSTDMT_flushProduced() should have already moved onto next job.
             * Therefore, some input has not yet been consumed. */
            if (toFlush==0) {
                assert(jobPtr->consumed < jobPtr->src.size);
            }
        }
        ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
    }

    return toFlush;
}


/* ------------------------------------------ */
/* =====   Multi-threaded compression   ===== */
/* ------------------------------------------ */

static unsigned ZSTDMT_computeTargetJobLog(const ZSTD_CCtx_params* params)
{
    unsigned jobLog;
    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
        /* In Long Range Mode, the windowLog is typically oversized.
         * In which case, it's preferable to determine the jobSize
         * based on cycleLog instead. */
        jobLog = MAX(21, ZSTD_cycleLog(params->cParams.chainLog, params->cParams.strategy) + 3);
    } else {
        jobLog = MAX(20, params->cParams.windowLog + 2);
    }
    return MIN(jobLog, (unsigned)ZSTDMT_JOBLOG_MAX);
}

static int ZSTDMT_overlapLog_default(ZSTD_strategy strat)
{
    switch(strat)
    {
        case ZSTD_btultra2:
            return 9;
        case ZSTD_btultra:
        case ZSTD_btopt:
            return 8;
        case ZSTD_btlazy2:
        case ZSTD_lazy2:
            return 7;
        case ZSTD_lazy:
        case ZSTD_greedy:
        case ZSTD_dfast:
        case ZSTD_fast:
        default:;
    }
    return 6;
}

static int ZSTDMT_overlapLog(int ovlog, ZSTD_strategy strat)
{
    assert(0 <= ovlog && ovlog <= 9);
    if (ovlog == 0) return ZSTDMT_overlapLog_default(strat);
    return ovlog;
}

static size_t ZSTDMT_computeOverlapSize(const ZSTD_CCtx_params* params)
{
    int const overlapRLog = 9 - ZSTDMT_overlapLog(params->overlapLog, params->cParams.strategy);
    int ovLog = (overlapRLog >= 8) ? 0 : (params->cParams.windowLog - overlapRLog);
    assert(0 <= overlapRLog && overlapRLog <= 8);
    if (params->ldmParams.enableLdm == ZSTD_ps_enable) {
        /* In Long Range Mode, the windowLog is typically oversized.
         * In which case, it's preferable to determine the jobSize
         * based on chainLog instead.
         * Then, ovLog becomes a fraction of the jobSize, rather than windowSize */
        ovLog = MIN(params->cParams.windowLog, ZSTDMT_computeTargetJobLog(params) - 2)
                - overlapRLog;
    }
    assert(0 <= ovLog && ovLog <= ZSTD_WINDOWLOG_MAX);
    DEBUGLOG(4, "overlapLog : %i", params->overlapLog);
    DEBUGLOG(4, "overlap size : %i", 1 << ovLog);
    return (ovLog==0) ? 0 : (size_t)1 << ovLog;
}

/* ====================================== */
/* =======      Streaming API     ======= */
/* ====================================== */

size_t ZSTDMT_initCStream_internal(
        ZSTDMT_CCtx* mtctx,
        const void* dict, size_t dictSize, ZSTD_dictContentType_e dictContentType,
        const ZSTD_CDict* cdict, ZSTD_CCtx_params params,
        unsigned long long pledgedSrcSize)
{
    DEBUGLOG(4, "ZSTDMT_initCStream_internal (pledgedSrcSize=%u, nbWorkers=%u, cctxPool=%u)",
                (U32)pledgedSrcSize, params.nbWorkers, mtctx->cctxPool->totalCCtx);

    /* params supposed partially fully validated at this point */
    assert(!ZSTD_isError(ZSTD_checkCParams(params.cParams)));
    assert(!((dict) && (cdict)));  /* either dict or cdict, not both */

    /* init */
    if (params.nbWorkers != mtctx->params.nbWorkers)
        FORWARD_IF_ERROR( ZSTDMT_resize(mtctx, (unsigned)params.nbWorkers) , "");

    if (params.jobSize != 0 && params.jobSize < ZSTDMT_JOBSIZE_MIN) params.jobSize = ZSTDMT_JOBSIZE_MIN;
    if (params.jobSize > (size_t)ZSTDMT_JOBSIZE_MAX) params.jobSize = (size_t)ZSTDMT_JOBSIZE_MAX;

    if (mtctx->allJobsCompleted == 0) {   /* previous compression not correctly finished */
        ZSTDMT_waitForAllJobsCompleted(mtctx);
        ZSTDMT_releaseAllJobResources(mtctx);
        mtctx->allJobsCompleted = 1;
    }

    mtctx->params = params;
    mtctx->frameContentSize = pledgedSrcSize;
    ZSTD_freeCDict(mtctx->cdictLocal);
    if (dict) {
        mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
                                                    ZSTD_dlm_byCopy, dictContentType, /* note : a loadPrefix becomes an internal CDict */
                                                    params.cParams, mtctx->cMem);
        mtctx->cdict = mtctx->cdictLocal;
        if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
    } else {
        mtctx->cdictLocal = NULL;
        mtctx->cdict = cdict;
    }

    mtctx->targetPrefixSize = ZSTDMT_computeOverlapSize(&params);
    DEBUGLOG(4, "overlapLog=%i => %u KB", params.overlapLog, (U32)(mtctx->targetPrefixSize>>10));
    mtctx->targetSectionSize = params.jobSize;
    if (mtctx->targetSectionSize == 0) {
        mtctx->targetSectionSize = 1ULL << ZSTDMT_computeTargetJobLog(&params);
    }
    assert(mtctx->targetSectionSize <= (size_t)ZSTDMT_JOBSIZE_MAX);

    if (params.rsyncable) {
        /* Aim for the targetsectionSize as the average job size. */
        U32 const jobSizeKB = (U32)(mtctx->targetSectionSize >> 10);
        U32 const rsyncBits = (assert(jobSizeKB >= 1), ZSTD_highbit32(jobSizeKB) + 10);
        /* We refuse to create jobs < RSYNC_MIN_BLOCK_SIZE bytes, so make sure our
         * expected job size is at least 4x larger. */
        assert(rsyncBits >= RSYNC_MIN_BLOCK_LOG + 2);
        DEBUGLOG(4, "rsyncLog = %u", rsyncBits);
        mtctx->rsync.hash = 0;
        mtctx->rsync.hitMask = (1ULL << rsyncBits) - 1;
        mtctx->rsync.primePower = ZSTD_rollingHash_primePower(RSYNC_LENGTH);
    }
    if (mtctx->targetSectionSize < mtctx->targetPrefixSize) mtctx->targetSectionSize = mtctx->targetPrefixSize;  /* job size must be >= overlap size */
    DEBUGLOG(4, "Job Size : %u KB (note : set to %u)", (U32)(mtctx->targetSectionSize>>10), (U32)params.jobSize);
    DEBUGLOG(4, "inBuff Size : %u KB", (U32)(mtctx->targetSectionSize>>10));
    ZSTDMT_setBufferSize(mtctx->bufPool, ZSTD_compressBound(mtctx->targetSectionSize));
    {
        /* If ldm is enabled we need windowSize space. */
        size_t const windowSize = mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable ? (1U << mtctx->params.cParams.windowLog) : 0;
        /* Two buffers of slack, plus extra space for the overlap
         * This is the minimum slack that LDM works with. One extra because
         * flush might waste up to targetSectionSize-1 bytes. Another extra
         * for the overlap (if > 0), then one to fill which doesn't overlap
         * with the LDM window.
         */
        size_t const nbSlackBuffers = 2 + (mtctx->targetPrefixSize > 0);
        size_t const slackSize = mtctx->targetSectionSize * nbSlackBuffers;
        /* Compute the total size, and always have enough slack */
        size_t const nbWorkers = MAX(mtctx->params.nbWorkers, 1);
        size_t const sectionsSize = mtctx->targetSectionSize * nbWorkers;
        size_t const capacity = MAX(windowSize, sectionsSize) + slackSize;
        if (mtctx->roundBuff.capacity < capacity) {
            if (mtctx->roundBuff.buffer)
                ZSTD_customFree(mtctx->roundBuff.buffer, mtctx->cMem);
            mtctx->roundBuff.buffer = (BYTE*)ZSTD_customMalloc(capacity, mtctx->cMem);
            if (mtctx->roundBuff.buffer == NULL) {
                mtctx->roundBuff.capacity = 0;
                return ERROR(memory_allocation);
            }
            mtctx->roundBuff.capacity = capacity;
        }
    }
    DEBUGLOG(4, "roundBuff capacity : %u KB", (U32)(mtctx->roundBuff.capacity>>10));
    mtctx->roundBuff.pos = 0;
    mtctx->inBuff.buffer = g_nullBuffer;
    mtctx->inBuff.filled = 0;
    mtctx->inBuff.prefix = kNullRange;
    mtctx->doneJobID = 0;
    mtctx->nextJobID = 0;
    mtctx->frameEnded = 0;
    mtctx->allJobsCompleted = 0;
    mtctx->consumed = 0;
    mtctx->produced = 0;

    /* update dictionary */
    ZSTD_freeCDict(mtctx->cdictLocal);
    mtctx->cdictLocal = NULL;
    mtctx->cdict = NULL;
    if (dict) {
        if (dictContentType == ZSTD_dct_rawContent) {
            mtctx->inBuff.prefix.start = (const BYTE*)dict;
            mtctx->inBuff.prefix.size = dictSize;
        } else {
            /* note : a loadPrefix becomes an internal CDict */
            mtctx->cdictLocal = ZSTD_createCDict_advanced(dict, dictSize,
                                                        ZSTD_dlm_byRef, dictContentType,
                                                        params.cParams, mtctx->cMem);
            mtctx->cdict = mtctx->cdictLocal;
            if (mtctx->cdictLocal == NULL) return ERROR(memory_allocation);
        }
    } else {
        mtctx->cdict = cdict;
    }

    if (ZSTDMT_serialState_reset(&mtctx->serial, mtctx->seqPool, params, mtctx->targetSectionSize,
                                 dict, dictSize, dictContentType))
        return ERROR(memory_allocation);


    return 0;
}


/* ZSTDMT_writeLastEmptyBlock()
 * Write a single empty block with an end-of-frame to finish a frame.
 * Job must be created from streaming variant.
 * This function is always successful if expected conditions are fulfilled.
 */
static void ZSTDMT_writeLastEmptyBlock(ZSTDMT_jobDescription* job)
{
    assert(job->lastJob == 1);
    assert(job->src.size == 0);   /* last job is empty -> will be simplified into a last empty block */
    assert(job->firstJob == 0);   /* cannot be first job, as it also needs to create frame header */
    assert(job->dstBuff.start == NULL);   /* invoked from streaming variant only (otherwise, dstBuff might be user's output) */
    job->dstBuff = ZSTDMT_getBuffer(job->bufPool);
    if (job->dstBuff.start == NULL) {
      job->cSize = ERROR(memory_allocation);
      return;
    }
    assert(job->dstBuff.capacity >= ZSTD_blockHeaderSize);   /* no buffer should ever be that small */
    job->src = kNullRange;
    job->cSize = ZSTD_writeLastEmptyBlock(job->dstBuff.start, job->dstBuff.capacity);
    assert(!ZSTD_isError(job->cSize));
    assert(job->consumed == 0);
}

static size_t ZSTDMT_createCompressionJob(ZSTDMT_CCtx* mtctx, size_t srcSize, ZSTD_EndDirective endOp)
{
    unsigned const jobID = mtctx->nextJobID & mtctx->jobIDMask;
    int const endFrame = (endOp == ZSTD_e_end);

    if (mtctx->nextJobID > mtctx->doneJobID + mtctx->jobIDMask) {
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: will not create new job : table is full");
        assert((mtctx->nextJobID & mtctx->jobIDMask) == (mtctx->doneJobID & mtctx->jobIDMask));
        return 0;
    }

    if (!mtctx->jobReady) {
        BYTE const* src = (BYTE const*)mtctx->inBuff.buffer.start;
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: preparing job %u to compress %u bytes with %u preload ",
                    mtctx->nextJobID, (U32)srcSize, (U32)mtctx->inBuff.prefix.size);
        mtctx->jobs[jobID].src.start = src;
        mtctx->jobs[jobID].src.size = srcSize;
        assert(mtctx->inBuff.filled >= srcSize);
        mtctx->jobs[jobID].prefix = mtctx->inBuff.prefix;
        mtctx->jobs[jobID].consumed = 0;
        mtctx->jobs[jobID].cSize = 0;
        mtctx->jobs[jobID].params = mtctx->params;
        mtctx->jobs[jobID].cdict = mtctx->nextJobID==0 ? mtctx->cdict : NULL;
        mtctx->jobs[jobID].fullFrameSize = mtctx->frameContentSize;
        mtctx->jobs[jobID].dstBuff = g_nullBuffer;
        mtctx->jobs[jobID].cctxPool = mtctx->cctxPool;
        mtctx->jobs[jobID].bufPool = mtctx->bufPool;
        mtctx->jobs[jobID].seqPool = mtctx->seqPool;
        mtctx->jobs[jobID].serial = &mtctx->serial;
        mtctx->jobs[jobID].jobID = mtctx->nextJobID;
        mtctx->jobs[jobID].firstJob = (mtctx->nextJobID==0);
        mtctx->jobs[jobID].lastJob = endFrame;
        mtctx->jobs[jobID].frameChecksumNeeded = mtctx->params.fParams.checksumFlag && endFrame && (mtctx->nextJobID>0);
        mtctx->jobs[jobID].dstFlushed = 0;

        /* Update the round buffer pos and clear the input buffer to be reset */
        mtctx->roundBuff.pos += srcSize;
        mtctx->inBuff.buffer = g_nullBuffer;
        mtctx->inBuff.filled = 0;
        /* Set the prefix for next job */
        if (!endFrame) {
            size_t const newPrefixSize = MIN(srcSize, mtctx->targetPrefixSize);
            mtctx->inBuff.prefix.start = src + srcSize - newPrefixSize;
            mtctx->inBuff.prefix.size = newPrefixSize;
        } else {   /* endFrame==1 => no need for another input buffer */
            mtctx->inBuff.prefix = kNullRange;
            mtctx->frameEnded = endFrame;
            if (mtctx->nextJobID == 0) {
                /* single job exception : checksum is already calculated directly within worker thread */
                mtctx->params.fParams.checksumFlag = 0;
        }   }

        if ( (srcSize == 0)
          && (mtctx->nextJobID>0)/*single job must also write frame header*/ ) {
            DEBUGLOG(5, "ZSTDMT_createCompressionJob: creating a last empty block to end frame");
            assert(endOp == ZSTD_e_end);  /* only possible case : need to end the frame with an empty last block */
            ZSTDMT_writeLastEmptyBlock(mtctx->jobs + jobID);
            mtctx->nextJobID++;
            return 0;
        }
    }

    DEBUGLOG(5, "ZSTDMT_createCompressionJob: posting job %u : %u bytes  (end:%u, jobNb == %u (mod:%u))",
                mtctx->nextJobID,
                (U32)mtctx->jobs[jobID].src.size,
                mtctx->jobs[jobID].lastJob,
                mtctx->nextJobID,
                jobID);
    if (POOL_tryAdd(mtctx->factory, ZSTDMT_compressionJob, &mtctx->jobs[jobID])) {
        mtctx->nextJobID++;
        mtctx->jobReady = 0;
    } else {
        DEBUGLOG(5, "ZSTDMT_createCompressionJob: no worker available for job %u", mtctx->nextJobID);
        mtctx->jobReady = 1;
    }
    return 0;
}


/*! ZSTDMT_flushProduced() :
 *  flush whatever data has been produced but not yet flushed in current job.
 *  move to next job if current one is fully flushed.
 * `output` : `pos` will be updated with amount of data flushed .
 * `blockToFlush` : if >0, the function will block and wait if there is no data available to flush .
 * @return : amount of data remaining within internal buffer, 0 if no more, 1 if unknown but > 0, or an error code */
static size_t ZSTDMT_flushProduced(ZSTDMT_CCtx* mtctx, ZSTD_outBuffer* output, unsigned blockToFlush, ZSTD_EndDirective end)
{
    unsigned const wJobID = mtctx->doneJobID & mtctx->jobIDMask;
    DEBUGLOG(5, "ZSTDMT_flushProduced (blocking:%u , job %u <= %u)",
                blockToFlush, mtctx->doneJobID, mtctx->nextJobID);
    assert(output->size >= output->pos);

    ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
    if (  blockToFlush
      && (mtctx->doneJobID < mtctx->nextJobID) ) {
        assert(mtctx->jobs[wJobID].dstFlushed <= mtctx->jobs[wJobID].cSize);
        while (mtctx->jobs[wJobID].dstFlushed == mtctx->jobs[wJobID].cSize) {  /* nothing to flush */
            if (mtctx->jobs[wJobID].consumed == mtctx->jobs[wJobID].src.size) {
                DEBUGLOG(5, "job %u is completely consumed (%u == %u) => don't wait for cond, there will be none",
                            mtctx->doneJobID, (U32)mtctx->jobs[wJobID].consumed, (U32)mtctx->jobs[wJobID].src.size);
                break;
            }
            DEBUGLOG(5, "waiting for something to flush from job %u (currently flushed: %u bytes)",
                        mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
            ZSTD_pthread_cond_wait(&mtctx->jobs[wJobID].job_cond, &mtctx->jobs[wJobID].job_mutex);  /* block when nothing to flush but some to come */
    }   }

    /* try to flush something */
    {   size_t cSize = mtctx->jobs[wJobID].cSize;                  /* shared */
        size_t const srcConsumed = mtctx->jobs[wJobID].consumed;   /* shared */
        size_t const srcSize = mtctx->jobs[wJobID].src.size;       /* read-only, could be done after mutex lock, but no-declaration-after-statement */
        ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);
        if (ZSTD_isError(cSize)) {
            DEBUGLOG(5, "ZSTDMT_flushProduced: job %u : compression error detected : %s",
                        mtctx->doneJobID, ZSTD_getErrorName(cSize));
            ZSTDMT_waitForAllJobsCompleted(mtctx);
            ZSTDMT_releaseAllJobResources(mtctx);
            return cSize;
        }
        /* add frame checksum if necessary (can only happen once) */
        assert(srcConsumed <= srcSize);
        if ( (srcConsumed == srcSize)   /* job completed -> worker no longer active */
          && mtctx->jobs[wJobID].frameChecksumNeeded ) {
            U32 const checksum = (U32)XXH64_digest(&mtctx->serial.xxhState);
            DEBUGLOG(4, "ZSTDMT_flushProduced: writing checksum : %08X \n", checksum);
            MEM_writeLE32((char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].cSize, checksum);
            cSize += 4;
            mtctx->jobs[wJobID].cSize += 4;  /* can write this shared value, as worker is no longer active */
            mtctx->jobs[wJobID].frameChecksumNeeded = 0;
        }

        if (cSize > 0) {   /* compression is ongoing or completed */
            size_t const toFlush = MIN(cSize - mtctx->jobs[wJobID].dstFlushed, output->size - output->pos);
            DEBUGLOG(5, "ZSTDMT_flushProduced: Flushing %u bytes from job %u (completion:%u/%u, generated:%u)",
                        (U32)toFlush, mtctx->doneJobID, (U32)srcConsumed, (U32)srcSize, (U32)cSize);
            assert(mtctx->doneJobID < mtctx->nextJobID);
            assert(cSize >= mtctx->jobs[wJobID].dstFlushed);
            assert(mtctx->jobs[wJobID].dstBuff.start != NULL);
            if (toFlush > 0) {
                ZSTD_memcpy((char*)output->dst + output->pos,
                    (const char*)mtctx->jobs[wJobID].dstBuff.start + mtctx->jobs[wJobID].dstFlushed,
                    toFlush);
            }
            output->pos += toFlush;
            mtctx->jobs[wJobID].dstFlushed += toFlush;  /* can write : this value is only used by mtctx */

            if ( (srcConsumed == srcSize)    /* job is completed */
              && (mtctx->jobs[wJobID].dstFlushed == cSize) ) {   /* output buffer fully flushed => free this job position */
                DEBUGLOG(5, "Job %u completed (%u bytes), moving to next one",
                        mtctx->doneJobID, (U32)mtctx->jobs[wJobID].dstFlushed);
                ZSTDMT_releaseBuffer(mtctx->bufPool, mtctx->jobs[wJobID].dstBuff);
                DEBUGLOG(5, "dstBuffer released");
                mtctx->jobs[wJobID].dstBuff = g_nullBuffer;
                mtctx->jobs[wJobID].cSize = 0;   /* ensure this job slot is considered "not started" in future check */
                mtctx->consumed += srcSize;
                mtctx->produced += cSize;
                mtctx->doneJobID++;
        }   }

        /* return value : how many bytes left in buffer ; fake it to 1 when unknown but >0 */
        if (cSize > mtctx->jobs[wJobID].dstFlushed) return (cSize - mtctx->jobs[wJobID].dstFlushed);
        if (srcSize > srcConsumed) return 1;   /* current job not completely compressed */
    }
    if (mtctx->doneJobID < mtctx->nextJobID) return 1;   /* some more jobs ongoing */
    if (mtctx->jobReady) return 1;      /* one job is ready to push, just not yet in the list */
    if (mtctx->inBuff.filled > 0) return 1;   /* input is not empty, and still needs to be converted into a job */
    mtctx->allJobsCompleted = mtctx->frameEnded;   /* all jobs are entirely flushed => if this one is last one, frame is completed */
    if (end == ZSTD_e_end) return !mtctx->frameEnded;  /* for ZSTD_e_end, question becomes : is frame completed ? instead of : are internal buffers fully flushed ? */
    return 0;   /* internal buffers fully flushed */
}

/**
 * Returns the range of data used by the earliest job that is not yet complete.
 * If the data of the first job is broken up into two segments, we cover both
 * sections.
 */
static Range ZSTDMT_getInputDataInUse(ZSTDMT_CCtx* mtctx)
{
    unsigned const firstJobID = mtctx->doneJobID;
    unsigned const lastJobID = mtctx->nextJobID;
    unsigned jobID;

    /* no need to check during first round */
    size_t roundBuffCapacity = mtctx->roundBuff.capacity;
    size_t nbJobs1stRoundMin = roundBuffCapacity / mtctx->targetSectionSize;
    if (lastJobID < nbJobs1stRoundMin) return kNullRange;

    for (jobID = firstJobID; jobID < lastJobID; ++jobID) {
        unsigned const wJobID = jobID & mtctx->jobIDMask;
        size_t consumed;

        ZSTD_PTHREAD_MUTEX_LOCK(&mtctx->jobs[wJobID].job_mutex);
        consumed = mtctx->jobs[wJobID].consumed;
        ZSTD_pthread_mutex_unlock(&mtctx->jobs[wJobID].job_mutex);

        if (consumed < mtctx->jobs[wJobID].src.size) {
            Range range = mtctx->jobs[wJobID].prefix;
            if (range.size == 0) {
                /* Empty prefix */
                range = mtctx->jobs[wJobID].src;
            }
            /* Job source in multiple segments not supported yet */
            assert(range.start <= mtctx->jobs[wJobID].src.start);
            return range;
        }
    }
    return kNullRange;
}

/**
 * Returns non-zero iff buffer and range overlap.
 */
static int ZSTDMT_isOverlapped(Buffer buffer, Range range)
{
    BYTE const* const bufferStart = (BYTE const*)buffer.start;
    BYTE const* const rangeStart = (BYTE const*)range.start;

    if (rangeStart == NULL || bufferStart == NULL)
        return 0;

    {
        BYTE const* const bufferEnd = bufferStart + buffer.capacity;
        BYTE const* const rangeEnd = rangeStart + range.size;

        /* Empty ranges cannot overlap */
        if (bufferStart == bufferEnd || rangeStart == rangeEnd)
            return 0;

        return bufferStart < rangeEnd && rangeStart < bufferEnd;
    }
}

static int ZSTDMT_doesOverlapWindow(Buffer buffer, ZSTD_window_t window)
{
    Range extDict;
    Range prefix;

    DEBUGLOG(5, "ZSTDMT_doesOverlapWindow");
    extDict.start = window.dictBase + window.lowLimit;
    extDict.size = window.dictLimit - window.lowLimit;

    prefix.start = window.base + window.dictLimit;
    prefix.size = window.nextSrc - (window.base + window.dictLimit);
    DEBUGLOG(5, "extDict [0x%zx, 0x%zx)",
                (size_t)extDict.start,
                (size_t)extDict.start + extDict.size);
    DEBUGLOG(5, "prefix  [0x%zx, 0x%zx)",
                (size_t)prefix.start,
                (size_t)prefix.start + prefix.size);

    return ZSTDMT_isOverlapped(buffer, extDict)
        || ZSTDMT_isOverlapped(buffer, prefix);
}

static void ZSTDMT_waitForLdmComplete(ZSTDMT_CCtx* mtctx, Buffer buffer)
{
    if (mtctx->params.ldmParams.enableLdm == ZSTD_ps_enable) {
        ZSTD_pthread_mutex_t* mutex = &mtctx->serial.ldmWindowMutex;
        DEBUGLOG(5, "ZSTDMT_waitForLdmComplete");
        DEBUGLOG(5, "source  [0x%zx, 0x%zx)",
                    (size_t)buffer.start,
                    (size_t)buffer.start + buffer.capacity);
        ZSTD_PTHREAD_MUTEX_LOCK(mutex);
        while (ZSTDMT_doesOverlapWindow(buffer, mtctx->serial.ldmWindow)) {
            DEBUGLOG(5, "Waiting for LDM to finish...");
            ZSTD_pthread_cond_wait(&mtctx->serial.ldmWindowCond, mutex);
        }
        DEBUGLOG(6, "Done waiting for LDM to finish");
        ZSTD_pthread_mutex_unlock(mutex);
    }
}

/**
 * Attempts to set the inBuff to the next section to fill.
 * If any part of the new section is still in use we give up.
 * Returns non-zero if the buffer is filled.
 */
static int ZSTDMT_tryGetInputRange(ZSTDMT_CCtx* mtctx)
{
    Range const inUse = ZSTDMT_getInputDataInUse(mtctx);
    size_t const spaceLeft = mtctx->roundBuff.capacity - mtctx->roundBuff.pos;
    size_t const spaceNeeded = mtctx->targetSectionSize;
    Buffer buffer;

    DEBUGLOG(5, "ZSTDMT_tryGetInputRange");
    assert(mtctx->inBuff.buffer.start == NULL);
    assert(mtctx->roundBuff.capacity >= spaceNeeded);

    if (spaceLeft < spaceNeeded) {
        /* ZSTD_invalidateRepCodes() doesn't work for extDict variants.
         * Simply copy the prefix to the beginning in that case.
         */
        BYTE* const start = (BYTE*)mtctx->roundBuff.buffer;
        size_t const prefixSize = mtctx->inBuff.prefix.size;

        buffer.start = start;
        buffer.capacity = prefixSize;
        if (ZSTDMT_isOverlapped(buffer, inUse)) {
            DEBUGLOG(5, "Waiting for buffer...");
            return 0;
        }
        ZSTDMT_waitForLdmComplete(mtctx, buffer);
        ZSTD_memmove(start, mtctx->inBuff.prefix.start, prefixSize);
        mtctx->inBuff.prefix.start = start;
        mtctx->roundBuff.pos = prefixSize;
    }
    buffer.start = mtctx->roundBuff.buffer + mtctx->roundBuff.pos;
    buffer.capacity = spaceNeeded;

    if (ZSTDMT_isOverlapped(buffer, inUse)) {
        DEBUGLOG(5, "Waiting for buffer...");
        return 0;
    }
    assert(!ZSTDMT_isOverlapped(buffer, mtctx->inBuff.prefix));

    ZSTDMT_waitForLdmComplete(mtctx, buffer);

    DEBUGLOG(5, "Using prefix range [%zx, %zx)",
                (size_t)mtctx->inBuff.prefix.start,
                (size_t)mtctx->inBuff.prefix.start + mtctx->inBuff.prefix.size);
    DEBUGLOG(5, "Using source range [%zx, %zx)",
                (size_t)buffer.start,
                (size_t)buffer.start + buffer.capacity);


    mtctx->inBuff.buffer = buffer;
    mtctx->inBuff.filled = 0;
    assert(mtctx->roundBuff.pos + buffer.capacity <= mtctx->roundBuff.capacity);
    return 1;
}

typedef struct {
  size_t toLoad;  /* The number of bytes to load from the input. */
  int flush;      /* Boolean declaring if we must flush because we found a synchronization point. */
} SyncPoint;

/**
 * Searches through the input for a synchronization point. If one is found, we
 * will instruct the caller to flush, and return the number of bytes to load.
 * Otherwise, we will load as many bytes as possible and instruct the caller
 * to continue as normal.
 */
static SyncPoint
findSynchronizationPoint(ZSTDMT_CCtx const* mtctx, ZSTD_inBuffer const input)
{
    BYTE const* const istart = (BYTE const*)input.src + input.pos;
    U64 const primePower = mtctx->rsync.primePower;
    U64 const hitMask = mtctx->rsync.hitMask;

    SyncPoint syncPoint;
    U64 hash;
    BYTE const* prev;
    size_t pos;

    syncPoint.toLoad = MIN(input.size - input.pos, mtctx->targetSectionSize - mtctx->inBuff.filled);
    syncPoint.flush = 0;
    if (!mtctx->params.rsyncable)
        /* Rsync is disabled. */
        return syncPoint;
    if (mtctx->inBuff.filled + input.size - input.pos < RSYNC_MIN_BLOCK_SIZE)
        /* We don't emit synchronization points if it would produce too small blocks.
         * We don't have enough input to find a synchronization point, so don't look.
         */
        return syncPoint;
    if (mtctx->inBuff.filled + syncPoint.toLoad < RSYNC_LENGTH)
        /* Not enough to compute the hash.
         * We will miss any synchronization points in this RSYNC_LENGTH byte
         * window. However, since it depends only in the internal buffers, if the
         * state is already synchronized, we will remain synchronized.
         * Additionally, the probability that we miss a synchronization point is
         * low: RSYNC_LENGTH / targetSectionSize.
         */
        return syncPoint;
    /* Initialize the loop variables. */
    if (mtctx->inBuff.filled < RSYNC_MIN_BLOCK_SIZE) {
        /* We don't need to scan the first RSYNC_MIN_BLOCK_SIZE positions
         * because they can't possibly be a sync point. So we can start
         * part way through the input buffer.
         */
        pos = RSYNC_MIN_BLOCK_SIZE - mtctx->inBuff.filled;
        if (pos >= RSYNC_LENGTH) {
            prev = istart + pos - RSYNC_LENGTH;
            hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
        } else {
            assert(mtctx->inBuff.filled >= RSYNC_LENGTH);
            prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
            hash = ZSTD_rollingHash_compute(prev + pos, (RSYNC_LENGTH - pos));
            hash = ZSTD_rollingHash_append(hash, istart, pos);
        }
    } else {
        /* We have enough bytes buffered to initialize the hash,
         * and have processed enough bytes to find a sync point.
         * Start scanning at the beginning of the input.
         */
        assert(mtctx->inBuff.filled >= RSYNC_MIN_BLOCK_SIZE);
        assert(RSYNC_MIN_BLOCK_SIZE >= RSYNC_LENGTH);
        pos = 0;
        prev = (BYTE const*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled - RSYNC_LENGTH;
        hash = ZSTD_rollingHash_compute(prev, RSYNC_LENGTH);
        if ((hash & hitMask) == hitMask) {
            /* We're already at a sync point so don't load any more until
             * we're able to flush this sync point.
             * This likely happened because the job table was full so we
             * couldn't add our job.
             */
            syncPoint.toLoad = 0;
            syncPoint.flush = 1;
            return syncPoint;
        }
    }
    /* Starting with the hash of the previous RSYNC_LENGTH bytes, roll
     * through the input. If we hit a synchronization point, then cut the
     * job off, and tell the compressor to flush the job. Otherwise, load
     * all the bytes and continue as normal.
     * If we go too long without a synchronization point (targetSectionSize)
     * then a block will be emitted anyways, but this is okay, since if we
     * are already synchronized we will remain synchronized.
     */
    assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
    for (; pos < syncPoint.toLoad; ++pos) {
        BYTE const toRemove = pos < RSYNC_LENGTH ? prev[pos] : istart[pos - RSYNC_LENGTH];
        /* This assert is very expensive, and Debian compiles with asserts enabled.
         * So disable it for now. We can get similar coverage by checking it at the
         * beginning & end of the loop.
         * assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
         */
        hash = ZSTD_rollingHash_rotate(hash, toRemove, istart[pos], primePower);
        assert(mtctx->inBuff.filled + pos >= RSYNC_MIN_BLOCK_SIZE);
        if ((hash & hitMask) == hitMask) {
            syncPoint.toLoad = pos + 1;
            syncPoint.flush = 1;
            ++pos; /* for assert */
            break;
        }
    }
    assert(pos < RSYNC_LENGTH || ZSTD_rollingHash_compute(istart + pos - RSYNC_LENGTH, RSYNC_LENGTH) == hash);
    return syncPoint;
}

size_t ZSTDMT_nextInputSizeHint(const ZSTDMT_CCtx* mtctx)
{
    size_t hintInSize = mtctx->targetSectionSize - mtctx->inBuff.filled;
    if (hintInSize==0) hintInSize = mtctx->targetSectionSize;
    return hintInSize;
}

/** ZSTDMT_compressStream_generic() :
 *  internal use only - exposed to be invoked from zstd_compress.c
 *  assumption : output and input are valid (pos <= size)
 * @return : minimum amount of data remaining to flush, 0 if none */
size_t ZSTDMT_compressStream_generic(ZSTDMT_CCtx* mtctx,
                                     ZSTD_outBuffer* output,
                                     ZSTD_inBuffer* input,
                                     ZSTD_EndDirective endOp)
{
    unsigned forwardInputProgress = 0;
    DEBUGLOG(5, "ZSTDMT_compressStream_generic (endOp=%u, srcSize=%u)",
                (U32)endOp, (U32)(input->size - input->pos));
    assert(output->pos <= output->size);
    assert(input->pos  <= input->size);

    if ((mtctx->frameEnded) && (endOp==ZSTD_e_continue)) {
        /* current frame being ended. Only flush/end are allowed */
        return ERROR(stage_wrong);
    }

    /* fill input buffer */
    if ( (!mtctx->jobReady)
      && (input->size > input->pos) ) {   /* support NULL input */
        if (mtctx->inBuff.buffer.start == NULL) {
            assert(mtctx->inBuff.filled == 0); /* Can't fill an empty buffer */
            if (!ZSTDMT_tryGetInputRange(mtctx)) {
                /* It is only possible for this operation to fail if there are
                 * still compression jobs ongoing.
                 */
                DEBUGLOG(5, "ZSTDMT_tryGetInputRange failed");
                assert(mtctx->doneJobID != mtctx->nextJobID);
            } else
                DEBUGLOG(5, "ZSTDMT_tryGetInputRange completed successfully : mtctx->inBuff.buffer.start = %p", mtctx->inBuff.buffer.start);
        }
        if (mtctx->inBuff.buffer.start != NULL) {
            SyncPoint const syncPoint = findSynchronizationPoint(mtctx, *input);
            if (syncPoint.flush && endOp == ZSTD_e_continue) {
                endOp = ZSTD_e_flush;
            }
            assert(mtctx->inBuff.buffer.capacity >= mtctx->targetSectionSize);
            DEBUGLOG(5, "ZSTDMT_compressStream_generic: adding %u bytes on top of %u to buffer of size %u",
                        (U32)syncPoint.toLoad, (U32)mtctx->inBuff.filled, (U32)mtctx->targetSectionSize);
            ZSTD_memcpy((char*)mtctx->inBuff.buffer.start + mtctx->inBuff.filled, (const char*)input->src + input->pos, syncPoint.toLoad);
            input->pos += syncPoint.toLoad;
            mtctx->inBuff.filled += syncPoint.toLoad;
            forwardInputProgress = syncPoint.toLoad>0;
        }
    }
    if ((input->pos < input->size) && (endOp == ZSTD_e_end)) {
        /* Can't end yet because the input is not fully consumed.
            * We are in one of these cases:
            * - mtctx->inBuff is NULL & empty: we couldn't get an input buffer so don't create a new job.
            * - We filled the input buffer: flush this job but don't end the frame.
            * - We hit a synchronization point: flush this job but don't end the frame.
            */
        assert(mtctx->inBuff.filled == 0 || mtctx->inBuff.filled == mtctx->targetSectionSize || mtctx->params.rsyncable);
        endOp = ZSTD_e_flush;
    }

    if ( (mtctx->jobReady)
      || (mtctx->inBuff.filled >= mtctx->targetSectionSize)  /* filled enough : let's compress */
      || ((endOp != ZSTD_e_continue) && (mtctx->inBuff.filled > 0))  /* something to flush : let's go */
      || ((endOp == ZSTD_e_end) && (!mtctx->frameEnded)) ) {   /* must finish the frame with a zero-size block */
        size_t const jobSize = mtctx->inBuff.filled;
        assert(mtctx->inBuff.filled <= mtctx->targetSectionSize);
        FORWARD_IF_ERROR( ZSTDMT_createCompressionJob(mtctx, jobSize, endOp) , "");
    }

    /* check for potential compressed data ready to be flushed */
    {   size_t const remainingToFlush = ZSTDMT_flushProduced(mtctx, output, !forwardInputProgress, endOp); /* block if there was no forward input progress */
        if (input->pos < input->size) return MAX(remainingToFlush, 1);  /* input not consumed : do not end flush yet */
        DEBUGLOG(5, "end of ZSTDMT_compressStream_generic: remainingToFlush = %u", (U32)remainingToFlush);
        return remainingToFlush;
    }
}
/**** ended inlining compress/zstdmt_compress.c ****/
#endif

/**** start inlining decompress/huf_decompress.c ****/
/* ******************************************************************
 * huff0 huffman decoder,
 * part of Finite State Entropy library
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 *
 *  You can contact the author at :
 *  - FSE+HUF source repository : https://github.com/Cyan4973/FiniteStateEntropy
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
****************************************************************** */

/* **************************************************************
*  Dependencies
****************************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/bitstream.h ****/
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: ../common/error_private.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../common/bits.h ****/

/* **************************************************************
*  Constants
****************************************************************/

#define HUF_DECODER_FAST_TABLELOG 11

/* **************************************************************
*  Macros
****************************************************************/

#ifdef HUF_DISABLE_FAST_DECODE
# define HUF_ENABLE_FAST_DECODE 0
#else
# define HUF_ENABLE_FAST_DECODE 1
#endif

/* These two optional macros force the use one way or another of the two
 * Huffman decompression implementations. You can't force in both directions
 * at the same time.
 */
#if defined(HUF_FORCE_DECOMPRESS_X1) && \
    defined(HUF_FORCE_DECOMPRESS_X2)
#error "Cannot force the use of the X1 and X2 decoders at the same time!"
#endif

/* When DYNAMIC_BMI2 is enabled, fast decoders are only called when bmi2 is
 * supported at runtime, so we can add the BMI2 target attribute.
 * When it is disabled, we will still get BMI2 if it is enabled statically.
 */
#if DYNAMIC_BMI2
# define HUF_FAST_BMI2_ATTRS BMI2_TARGET_ATTRIBUTE
#else
# define HUF_FAST_BMI2_ATTRS
#endif

#ifdef __cplusplus
# define HUF_EXTERN_C extern "C"
#else
# define HUF_EXTERN_C
#endif
#define HUF_ASM_DECL HUF_EXTERN_C

#if DYNAMIC_BMI2
# define HUF_NEED_BMI2_FUNCTION 1
#else
# define HUF_NEED_BMI2_FUNCTION 0
#endif

/* **************************************************************
*  Error Management
****************************************************************/
#define HUF_isError ERR_isError


/* **************************************************************
*  Byte alignment for workSpace management
****************************************************************/
#define HUF_ALIGN(x, a)         HUF_ALIGN_MASK((x), (a) - 1)
#define HUF_ALIGN_MASK(x, mask) (((x) + (mask)) & ~(mask))


/* **************************************************************
*  BMI2 Variant Wrappers
****************************************************************/
typedef size_t (*HUF_DecompressUsingDTableFn)(void *dst, size_t dstSize,
                                              const void *cSrc,
                                              size_t cSrcSize,
                                              const HUF_DTable *DTable);

#if DYNAMIC_BMI2

#define HUF_DGEN(fn)                                                        \
                                                                            \
    static size_t fn##_default(                                             \
                  void* dst,  size_t dstSize,                               \
            const void* cSrc, size_t cSrcSize,                              \
            const HUF_DTable* DTable)                                       \
    {                                                                       \
        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
    }                                                                       \
                                                                            \
    static BMI2_TARGET_ATTRIBUTE size_t fn##_bmi2(                          \
                  void* dst,  size_t dstSize,                               \
            const void* cSrc, size_t cSrcSize,                              \
            const HUF_DTable* DTable)                                       \
    {                                                                       \
        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
    }                                                                       \
                                                                            \
    static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
    {                                                                       \
        if (flags & HUF_flags_bmi2) {                                       \
            return fn##_bmi2(dst, dstSize, cSrc, cSrcSize, DTable);         \
        }                                                                   \
        return fn##_default(dst, dstSize, cSrc, cSrcSize, DTable);          \
    }

#else

#define HUF_DGEN(fn)                                                        \
    static size_t fn(void* dst, size_t dstSize, void const* cSrc,           \
                     size_t cSrcSize, HUF_DTable const* DTable, int flags)  \
    {                                                                       \
        (void)flags;                                                        \
        return fn##_body(dst, dstSize, cSrc, cSrcSize, DTable);             \
    }

#endif


/*-***************************/
/*  generic DTableDesc       */
/*-***************************/
typedef struct { BYTE maxTableLog; BYTE tableType; BYTE tableLog; BYTE reserved; } DTableDesc;

static DTableDesc HUF_getDTableDesc(const HUF_DTable* table)
{
    DTableDesc dtd;
    ZSTD_memcpy(&dtd, table, sizeof(dtd));
    return dtd;
}

static size_t HUF_initFastDStream(BYTE const* ip) {
    BYTE const lastByte = ip[7];
    size_t const bitsConsumed = lastByte ? 8 - ZSTD_highbit32(lastByte) : 0;
    size_t const value = MEM_readLEST(ip) | 1;
    assert(bitsConsumed <= 8);
    assert(sizeof(size_t) == 8);
    return value << bitsConsumed;
}


/**
 * The input/output arguments to the Huffman fast decoding loop:
 *
 * ip [in/out] - The input pointers, must be updated to reflect what is consumed.
 * op [in/out] - The output pointers, must be updated to reflect what is written.
 * bits [in/out] - The bitstream containers, must be updated to reflect the current state.
 * dt [in] - The decoding table.
 * ilowest [in] - The beginning of the valid range of the input. Decoders may read
 *                down to this pointer. It may be below iend[0].
 * oend [in] - The end of the output stream. op[3] must not cross oend.
 * iend [in] - The end of each input stream. ip[i] may cross iend[i],
 *             as long as it is above ilowest, but that indicates corruption.
 */
typedef struct {
    BYTE const* ip[4];
    BYTE* op[4];
    U64 bits[4];
    void const* dt;
    BYTE const* ilowest;
    BYTE* oend;
    BYTE const* iend[4];
} HUF_DecompressFastArgs;

typedef void (*HUF_DecompressFastLoopFn)(HUF_DecompressFastArgs*);

/**
 * Initializes args for the fast decoding loop.
 * @returns 1 on success
 *          0 if the fallback implementation should be used.
 *          Or an error code on failure.
 */
static size_t HUF_DecompressFastArgs_init(HUF_DecompressFastArgs* args, void* dst, size_t dstSize, void const* src, size_t srcSize, const HUF_DTable* DTable)
{
    void const* dt = DTable + 1;
    U32 const dtLog = HUF_getDTableDesc(DTable).tableLog;

    const BYTE* const istart = (const BYTE*)src;

    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);

    /* The fast decoding loop assumes 64-bit little-endian.
     * This condition is false on x32.
     */
    if (!MEM_isLittleEndian() || MEM_32bits())
        return 0;

    /* Avoid nullptr addition */
    if (dstSize == 0)
        return 0;
    assert(dst != NULL);

    /* strict minimum : jump table + 1 byte per stream */
    if (srcSize < 10)
        return ERROR(corruption_detected);

    /* Must have at least 8 bytes per stream because we don't handle initializing smaller bit containers.
     * If table log is not correct at this point, fallback to the old decoder.
     * On small inputs we don't have enough data to trigger the fast loop, so use the old decoder.
     */
    if (dtLog != HUF_DECODER_FAST_TABLELOG)
        return 0;

    /* Read the jump table. */
    {
        size_t const length1 = MEM_readLE16(istart);
        size_t const length2 = MEM_readLE16(istart+2);
        size_t const length3 = MEM_readLE16(istart+4);
        size_t const length4 = srcSize - (length1 + length2 + length3 + 6);
        args->iend[0] = istart + 6;  /* jumpTable */
        args->iend[1] = args->iend[0] + length1;
        args->iend[2] = args->iend[1] + length2;
        args->iend[3] = args->iend[2] + length3;

        /* HUF_initFastDStream() requires this, and this small of an input
         * won't benefit from the ASM loop anyways.
         */
        if (length1 < 8 || length2 < 8 || length3 < 8 || length4 < 8)
            return 0;
        if (length4 > srcSize) return ERROR(corruption_detected);   /* overflow */
    }
    /* ip[] contains the position that is currently loaded into bits[]. */
    args->ip[0] = args->iend[1] - sizeof(U64);
    args->ip[1] = args->iend[2] - sizeof(U64);
    args->ip[2] = args->iend[3] - sizeof(U64);
    args->ip[3] = (BYTE const*)src + srcSize - sizeof(U64);

    /* op[] contains the output pointers. */
    args->op[0] = (BYTE*)dst;
    args->op[1] = args->op[0] + (dstSize+3)/4;
    args->op[2] = args->op[1] + (dstSize+3)/4;
    args->op[3] = args->op[2] + (dstSize+3)/4;

    /* No point to call the ASM loop for tiny outputs. */
    if (args->op[3] >= oend)
        return 0;

    /* bits[] is the bit container.
        * It is read from the MSB down to the LSB.
        * It is shifted left as it is read, and zeros are
        * shifted in. After the lowest valid bit a 1 is
        * set, so that CountTrailingZeros(bits[]) can be used
        * to count how many bits we've consumed.
        */
    args->bits[0] = HUF_initFastDStream(args->ip[0]);
    args->bits[1] = HUF_initFastDStream(args->ip[1]);
    args->bits[2] = HUF_initFastDStream(args->ip[2]);
    args->bits[3] = HUF_initFastDStream(args->ip[3]);

    /* The decoders must be sure to never read beyond ilowest.
     * This is lower than iend[0], but allowing decoders to read
     * down to ilowest can allow an extra iteration or two in the
     * fast loop.
     */
    args->ilowest = istart;

    args->oend = oend;
    args->dt = dt;

    return 1;
}

static size_t HUF_initRemainingDStream(BIT_DStream_t* bit, HUF_DecompressFastArgs const* args, int stream, BYTE* segmentEnd)
{
    /* Validate that we haven't overwritten. */
    if (args->op[stream] > segmentEnd)
        return ERROR(corruption_detected);
    /* Validate that we haven't read beyond iend[].
        * Note that ip[] may be < iend[] because the MSB is
        * the next bit to read, and we may have consumed 100%
        * of the stream, so down to iend[i] - 8 is valid.
        */
    if (args->ip[stream] < args->iend[stream] - 8)
        return ERROR(corruption_detected);

    /* Construct the BIT_DStream_t. */
    assert(sizeof(size_t) == 8);
    bit->bitContainer = MEM_readLEST(args->ip[stream]);
    bit->bitsConsumed = ZSTD_countTrailingZeros64(args->bits[stream]);
    bit->start = (const char*)args->ilowest;
    bit->limitPtr = bit->start + sizeof(size_t);
    bit->ptr = (const char*)args->ip[stream];

    return 0;
}

/* Calls X(N) for each stream 0, 1, 2, 3. */
#define HUF_4X_FOR_EACH_STREAM(X) \
    do {                          \
        X(0);                     \
        X(1);                     \
        X(2);                     \
        X(3);                     \
    } while (0)

/* Calls X(N, var) for each stream 0, 1, 2, 3. */
#define HUF_4X_FOR_EACH_STREAM_WITH_VAR(X, var) \
    do {                                        \
        X(0, (var));                            \
        X(1, (var));                            \
        X(2, (var));                            \
        X(3, (var));                            \
    } while (0)


#ifndef HUF_FORCE_DECOMPRESS_X2

/*-***************************/
/*  single-symbol decoding   */
/*-***************************/
typedef struct { BYTE nbBits; BYTE byte; } HUF_DEltX1;   /* single-symbol decoding */

/**
 * Packs 4 HUF_DEltX1 structs into a U64. This is used to lay down 4 entries at
 * a time.
 */
static U64 HUF_DEltX1_set4(BYTE symbol, BYTE nbBits) {
    U64 D4;
    if (MEM_isLittleEndian()) {
        D4 = (U64)((symbol << 8) + nbBits);
    } else {
        D4 = (U64)(symbol + (nbBits << 8));
    }
    assert(D4 < (1U << 16));
    D4 *= 0x0001000100010001ULL;
    return D4;
}

/**
 * Increase the tableLog to targetTableLog and rescales the stats.
 * If tableLog > targetTableLog this is a no-op.
 * @returns New tableLog
 */
static U32 HUF_rescaleStats(BYTE* huffWeight, U32* rankVal, U32 nbSymbols, U32 tableLog, U32 targetTableLog)
{
    if (tableLog > targetTableLog)
        return tableLog;
    if (tableLog < targetTableLog) {
        U32 const scale = targetTableLog - tableLog;
        U32 s;
        /* Increase the weight for all non-zero probability symbols by scale. */
        for (s = 0; s < nbSymbols; ++s) {
            huffWeight[s] += (BYTE)((huffWeight[s] == 0) ? 0 : scale);
        }
        /* Update rankVal to reflect the new weights.
         * All weights except 0 get moved to weight + scale.
         * Weights [1, scale] are empty.
         */
        for (s = targetTableLog; s > scale; --s) {
            rankVal[s] = rankVal[s - scale];
        }
        for (s = scale; s > 0; --s) {
            rankVal[s] = 0;
        }
    }
    return targetTableLog;
}

typedef struct {
        U32 rankVal[HUF_TABLELOG_ABSOLUTEMAX + 1];
        U32 rankStart[HUF_TABLELOG_ABSOLUTEMAX + 1];
        U32 statsWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
        BYTE symbols[HUF_SYMBOLVALUE_MAX + 1];
        BYTE huffWeight[HUF_SYMBOLVALUE_MAX + 1];
} HUF_ReadDTableX1_Workspace;

size_t HUF_readDTableX1_wksp(HUF_DTable* DTable, const void* src, size_t srcSize, void* workSpace, size_t wkspSize, int flags)
{
    U32 tableLog = 0;
    U32 nbSymbols = 0;
    size_t iSize;
    void* const dtPtr = DTable + 1;
    HUF_DEltX1* const dt = (HUF_DEltX1*)dtPtr;
    HUF_ReadDTableX1_Workspace* wksp = (HUF_ReadDTableX1_Workspace*)workSpace;

    DEBUG_STATIC_ASSERT(HUF_DECOMPRESS_WORKSPACE_SIZE >= sizeof(*wksp));
    if (sizeof(*wksp) > wkspSize) return ERROR(tableLog_tooLarge);

    DEBUG_STATIC_ASSERT(sizeof(DTableDesc) == sizeof(HUF_DTable));
    /* ZSTD_memset(huffWeight, 0, sizeof(huffWeight)); */   /* is not necessary, even though some analyzer complain ... */

    iSize = HUF_readStats_wksp(wksp->huffWeight, HUF_SYMBOLVALUE_MAX + 1, wksp->rankVal, &nbSymbols, &tableLog, src, srcSize, wksp->statsWksp, sizeof(wksp->statsWksp), flags);
    if (HUF_isError(iSize)) return iSize;


    /* Table header */
    {   DTableDesc dtd = HUF_getDTableDesc(DTable);
        U32 const maxTableLog = dtd.maxTableLog + 1;
        U32 const targetTableLog = MIN(maxTableLog, HUF_DECODER_FAST_TABLELOG);
        tableLog = HUF_rescaleStats(wksp->huffWeight, wksp->rankVal, nbSymbols, tableLog, targetTableLog);
        if (tableLog > (U32)(dtd.maxTableLog+1)) return ERROR(tableLog_tooLarge);   /* DTable too small, Huffman tree cannot fit in */
        dtd.tableType = 0;
        dtd.tableLog = (BYTE)tableLog;
        ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
    }

    /* Compute symbols and rankStart given rankVal:
     *
     * rankVal already contains the number of values of each weight.
     *
     * symbols contains the symbols ordered by weight. First are the rankVal[0]
     * weight 0 symbols, followed by the rankVal[1] weight 1 symbols, and so on.
     * symbols[0] is filled (but unused) to avoid a branch.
     *
     * rankStart contains the offset where each rank belongs in the DTable.
     * rankStart[0] is not filled because there are no entries in the table for
     * weight 0.
     */
    {   int n;
        U32 nextRankStart = 0;
        int const unroll = 4;
        int const nLimit = (int)nbSymbols - unroll + 1;
        for (n=0; n<(int)tableLog+1; n++) {
            U32 const curr = nextRankStart;
            nextRankStart += wksp->rankVal[n];
            wksp->rankStart[n] = curr;
        }
        for (n=0; n < nLimit; n += unroll) {
            int u;
            for (u=0; u < unroll; ++u) {
                size_t const w = wksp->huffWeight[n+u];
                wksp->symbols[wksp->rankStart[w]++] = (BYTE)(n+u);
            }
        }
        for (; n < (int)nbSymbols; ++n) {
            size_t const w = wksp->huffWeight[n];
            wksp->symbols[wksp->rankStart[w]++] = (BYTE)n;
        }
    }

    /* fill DTable
     * We fill all entries of each weight in order.
     * That way length is a constant for each iteration of the outer loop.
     * We can switch based on the length to a different inner loop which is
     * optimized for that particular case.
     */
    {   U32 w;
        int symbol = wksp->rankVal[0];
        int rankStart = 0;
        for (w=1; w<tableLog+1; ++w) {
            int const symbolCount = wksp->rankVal[w];
            int const length = (1 << w) >> 1;
            int uStart = rankStart;
            BYTE const nbBits = (BYTE)(tableLog + 1 - w);
            int s;
            int u;
            switch (length) {
            case 1:
                for (s=0; s<symbolCount; ++s) {
                    HUF_DEltX1 D;
                    D.byte = wksp->symbols[symbol + s];
                    D.nbBits = nbBits;
                    dt[uStart] = D;
                    uStart += 1;
                }
                break;
            case 2:
                for (s=0; s<symbolCount; ++s) {
                    HUF_DEltX1 D;
                    D.byte = wksp->symbols[symbol + s];
                    D.nbBits = nbBits;
                    dt[uStart+0] = D;
                    dt[uStart+1] = D;
                    uStart += 2;
                }
                break;
            case 4:
                for (s=0; s<symbolCount; ++s) {
                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
                    MEM_write64(dt + uStart, D4);
                    uStart += 4;
                }
                break;
            case 8:
                for (s=0; s<symbolCount; ++s) {
                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
                    MEM_write64(dt + uStart, D4);
                    MEM_write64(dt + uStart + 4, D4);
                    uStart += 8;
                }
                break;
            default:
                for (s=0; s<symbolCount; ++s) {
                    U64 const D4 = HUF_DEltX1_set4(wksp->symbols[symbol + s], nbBits);
                    for (u=0; u < length; u += 16) {
                        MEM_write64(dt + uStart + u + 0, D4);
                        MEM_write64(dt + uStart + u + 4, D4);
                        MEM_write64(dt + uStart + u + 8, D4);
                        MEM_write64(dt + uStart + u + 12, D4);
                    }
                    assert(u == length);
                    uStart += length;
                }
                break;
            }
            symbol += symbolCount;
            rankStart += symbolCount * length;
        }
    }
    return iSize;
}

FORCE_INLINE_TEMPLATE BYTE
HUF_decodeSymbolX1(BIT_DStream_t* Dstream, const HUF_DEltX1* dt, const U32 dtLog)
{
    size_t const val = BIT_lookBitsFast(Dstream, dtLog); /* note : dtLog >= 1 */
    BYTE const c = dt[val].byte;
    BIT_skipBits(Dstream, dt[val].nbBits);
    return c;
}

#define HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr) \
    do { *ptr++ = HUF_decodeSymbolX1(DStreamPtr, dt, dtLog); } while (0)

#define HUF_DECODE_SYMBOLX1_1(ptr, DStreamPtr)      \
    do {                                            \
        if (MEM_64bits() || (HUF_TABLELOG_MAX<=12)) \
            HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr); \
    } while (0)

#define HUF_DECODE_SYMBOLX1_2(ptr, DStreamPtr)      \
    do {                                            \
        if (MEM_64bits())                           \
            HUF_DECODE_SYMBOLX1_0(ptr, DStreamPtr); \
    } while (0)

HINT_INLINE size_t
HUF_decodeStreamX1(BYTE* p, BIT_DStream_t* const bitDPtr, BYTE* const pEnd, const HUF_DEltX1* const dt, const U32 dtLog)
{
    BYTE* const pStart = p;

    /* up to 4 symbols at a time */
    if ((pEnd - p) > 3) {
        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-3)) {
            HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
            HUF_DECODE_SYMBOLX1_1(p, bitDPtr);
            HUF_DECODE_SYMBOLX1_2(p, bitDPtr);
            HUF_DECODE_SYMBOLX1_0(p, bitDPtr);
        }
    } else {
        BIT_reloadDStream(bitDPtr);
    }

    /* [0-3] symbols remaining */
    if (MEM_32bits())
        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd))
            HUF_DECODE_SYMBOLX1_0(p, bitDPtr);

    /* no more data to retrieve from bitstream, no need to reload */
    while (p < pEnd)
        HUF_DECODE_SYMBOLX1_0(p, bitDPtr);

    return (size_t)(pEnd-pStart);
}

FORCE_INLINE_TEMPLATE size_t
HUF_decompress1X1_usingDTable_internal_body(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable)
{
    BYTE* op = (BYTE*)dst;
    BYTE* const oend = ZSTD_maybeNullPtrAdd(op, dstSize);
    const void* dtPtr = DTable + 1;
    const HUF_DEltX1* const dt = (const HUF_DEltX1*)dtPtr;
    BIT_DStream_t bitD;
    DTableDesc const dtd = HUF_getDTableDesc(DTable);
    U32 const dtLog = dtd.tableLog;

    CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );

    HUF_decodeStreamX1(op, &bitD, oend, dt, dtLog);

    if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);

    return dstSize;
}

/* HUF_decompress4X1_usingDTable_internal_body():
 * Conditions :
 * @dstSize >= 6
 */
FORCE_INLINE_TEMPLATE size_t
HUF_decompress4X1_usingDTable_internal_body(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable)
{
    /* Check */
    if (cSrcSize < 10) return ERROR(corruption_detected);  /* strict minimum : jump table + 1 byte per stream */
    if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */

    {   const BYTE* const istart = (const BYTE*) cSrc;
        BYTE* const ostart = (BYTE*) dst;
        BYTE* const oend = ostart + dstSize;
        BYTE* const olimit = oend - 3;
        const void* const dtPtr = DTable + 1;
        const HUF_DEltX1* const dt = (const HUF_DEltX1*)dtPtr;

        /* Init */
        BIT_DStream_t bitD1;
        BIT_DStream_t bitD2;
        BIT_DStream_t bitD3;
        BIT_DStream_t bitD4;
        size_t const length1 = MEM_readLE16(istart);
        size_t const length2 = MEM_readLE16(istart+2);
        size_t const length3 = MEM_readLE16(istart+4);
        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
        const BYTE* const istart1 = istart + 6;  /* jumpTable */
        const BYTE* const istart2 = istart1 + length1;
        const BYTE* const istart3 = istart2 + length2;
        const BYTE* const istart4 = istart3 + length3;
        const size_t segmentSize = (dstSize+3) / 4;
        BYTE* const opStart2 = ostart + segmentSize;
        BYTE* const opStart3 = opStart2 + segmentSize;
        BYTE* const opStart4 = opStart3 + segmentSize;
        BYTE* op1 = ostart;
        BYTE* op2 = opStart2;
        BYTE* op3 = opStart3;
        BYTE* op4 = opStart4;
        DTableDesc const dtd = HUF_getDTableDesc(DTable);
        U32 const dtLog = dtd.tableLog;
        U32 endSignal = 1;

        if (length4 > cSrcSize) return ERROR(corruption_detected);   /* overflow */
        if (opStart4 > oend) return ERROR(corruption_detected);      /* overflow */
        assert(dstSize >= 6); /* validated above */
        CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
        CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
        CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
        CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );

        /* up to 16 symbols per loop (4 symbols per stream) in 64-bit mode */
        if ((size_t)(oend - op4) >= sizeof(size_t)) {
            for ( ; (endSignal) & (op4 < olimit) ; ) {
                HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX1_1(op1, &bitD1);
                HUF_DECODE_SYMBOLX1_1(op2, &bitD2);
                HUF_DECODE_SYMBOLX1_1(op3, &bitD3);
                HUF_DECODE_SYMBOLX1_1(op4, &bitD4);
                HUF_DECODE_SYMBOLX1_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX1_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX1_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX1_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX1_0(op1, &bitD1);
                HUF_DECODE_SYMBOLX1_0(op2, &bitD2);
                HUF_DECODE_SYMBOLX1_0(op3, &bitD3);
                HUF_DECODE_SYMBOLX1_0(op4, &bitD4);
                endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
                endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
                endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
                endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
            }
        }

        /* check corruption */
        /* note : should not be necessary : op# advance in lock step, and we control op4.
         *        but curiously, binary generated by gcc 7.2 & 7.3 with -mbmi2 runs faster when >=1 test is present */
        if (op1 > opStart2) return ERROR(corruption_detected);
        if (op2 > opStart3) return ERROR(corruption_detected);
        if (op3 > opStart4) return ERROR(corruption_detected);
        /* note : op4 supposed already verified within main loop */

        /* finish bitStreams one by one */
        HUF_decodeStreamX1(op1, &bitD1, opStart2, dt, dtLog);
        HUF_decodeStreamX1(op2, &bitD2, opStart3, dt, dtLog);
        HUF_decodeStreamX1(op3, &bitD3, opStart4, dt, dtLog);
        HUF_decodeStreamX1(op4, &bitD4, oend,     dt, dtLog);

        /* check */
        { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
          if (!endCheck) return ERROR(corruption_detected); }

        /* decoded size */
        return dstSize;
    }
}

#if HUF_NEED_BMI2_FUNCTION
static BMI2_TARGET_ATTRIBUTE
size_t HUF_decompress4X1_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable) {
    return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
}
#endif

static
size_t HUF_decompress4X1_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable) {
    return HUF_decompress4X1_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
}

#if ZSTD_ENABLE_ASM_X86_64_BMI2

HUF_ASM_DECL void HUF_decompress4X1_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;

#endif

static HUF_FAST_BMI2_ATTRS
void HUF_decompress4X1_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
{
    U64 bits[4];
    BYTE const* ip[4];
    BYTE* op[4];
    U16 const* const dtable = (U16 const*)args->dt;
    BYTE* const oend = args->oend;
    BYTE const* const ilowest = args->ilowest;

    /* Copy the arguments to local variables */
    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
    ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
    ZSTD_memcpy(&op, &args->op, sizeof(op));

    assert(MEM_isLittleEndian());
    assert(!MEM_32bits());

    for (;;) {
        BYTE* olimit;
        int stream;

        /* Assert loop preconditions */
#ifndef NDEBUG
        for (stream = 0; stream < 4; ++stream) {
            assert(op[stream] <= (stream == 3 ? oend : op[stream + 1]));
            assert(ip[stream] >= ilowest);
        }
#endif
        /* Compute olimit */
        {
            /* Each iteration produces 5 output symbols per stream */
            size_t const oiters = (size_t)(oend - op[3]) / 5;
            /* Each iteration consumes up to 11 bits * 5 = 55 bits < 7 bytes
             * per stream.
             */
            size_t const iiters = (size_t)(ip[0] - ilowest) / 7;
            /* We can safely run iters iterations before running bounds checks */
            size_t const iters = MIN(oiters, iiters);
            size_t const symbols = iters * 5;

            /* We can simply check that op[3] < olimit, instead of checking all
             * of our bounds, since we can't hit the other bounds until we've run
             * iters iterations, which only happens when op[3] == olimit.
             */
            olimit = op[3] + symbols;

            /* Exit fast decoding loop once we reach the end. */
            if (op[3] == olimit)
                break;

            /* Exit the decoding loop if any input pointer has crossed the
             * previous one. This indicates corruption, and a precondition
             * to our loop is that ip[i] >= ip[0].
             */
            for (stream = 1; stream < 4; ++stream) {
                if (ip[stream] < ip[stream - 1])
                    goto _out;
            }
        }

#ifndef NDEBUG
        for (stream = 1; stream < 4; ++stream) {
            assert(ip[stream] >= ip[stream - 1]);
        }
#endif

#define HUF_4X1_DECODE_SYMBOL(_stream, _symbol)                 \
    do {                                                        \
        int const index = (int)(bits[(_stream)] >> 53);         \
        int const entry = (int)dtable[index];                   \
        bits[(_stream)] <<= (entry & 0x3F);                     \
        op[(_stream)][(_symbol)] = (BYTE)((entry >> 8) & 0xFF); \
    } while (0)

#define HUF_4X1_RELOAD_STREAM(_stream)                              \
    do {                                                            \
        int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
        int const nbBits = ctz & 7;                                 \
        int const nbBytes = ctz >> 3;                               \
        op[(_stream)] += 5;                                         \
        ip[(_stream)] -= nbBytes;                                   \
        bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1;            \
        bits[(_stream)] <<= nbBits;                                 \
    } while (0)

        /* Manually unroll the loop because compilers don't consistently
         * unroll the inner loops, which destroys performance.
         */
        do {
            /* Decode 5 symbols in each of the 4 streams */
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 0);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 1);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 2);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 3);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X1_DECODE_SYMBOL, 4);

            /* Reload each of the 4 the bitstreams */
            HUF_4X_FOR_EACH_STREAM(HUF_4X1_RELOAD_STREAM);
        } while (op[3] < olimit);

#undef HUF_4X1_DECODE_SYMBOL
#undef HUF_4X1_RELOAD_STREAM
    }

_out:

    /* Save the final values of each of the state variables back to args. */
    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
    ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
    ZSTD_memcpy(&args->op, &op, sizeof(op));
}

/**
 * @returns @p dstSize on success (>= 6)
 *          0 if the fallback implementation should be used
 *          An error if an error occurred
 */
static HUF_FAST_BMI2_ATTRS
size_t
HUF_decompress4X1_usingDTable_internal_fast(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable,
    HUF_DecompressFastLoopFn loopFn)
{
    void const* dt = DTable + 1;
    BYTE const* const ilowest = (BYTE const*)cSrc;
    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);
    HUF_DecompressFastArgs args;
    {   size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
        FORWARD_IF_ERROR(ret, "Failed to init fast loop args");
        if (ret == 0)
            return 0;
    }

    assert(args.ip[0] >= args.ilowest);
    loopFn(&args);

    /* Our loop guarantees that ip[] >= ilowest and that we haven't
    * overwritten any op[].
    */
    assert(args.ip[0] >= ilowest);
    assert(args.ip[0] >= ilowest);
    assert(args.ip[1] >= ilowest);
    assert(args.ip[2] >= ilowest);
    assert(args.ip[3] >= ilowest);
    assert(args.op[3] <= oend);

    assert(ilowest == args.ilowest);
    assert(ilowest + 6 == args.iend[0]);
    (void)ilowest;

    /* finish bit streams one by one. */
    {   size_t const segmentSize = (dstSize+3) / 4;
        BYTE* segmentEnd = (BYTE*)dst;
        int i;
        for (i = 0; i < 4; ++i) {
            BIT_DStream_t bit;
            if (segmentSize <= (size_t)(oend - segmentEnd))
                segmentEnd += segmentSize;
            else
                segmentEnd = oend;
            FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
            /* Decompress and validate that we've produced exactly the expected length. */
            args.op[i] += HUF_decodeStreamX1(args.op[i], &bit, segmentEnd, (HUF_DEltX1 const*)dt, HUF_DECODER_FAST_TABLELOG);
            if (args.op[i] != segmentEnd) return ERROR(corruption_detected);
        }
    }

    /* decoded size */
    assert(dstSize != 0);
    return dstSize;
}

HUF_DGEN(HUF_decompress1X1_usingDTable_internal)

static size_t HUF_decompress4X1_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
{
    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X1_usingDTable_internal_default;
    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X1_usingDTable_internal_fast_c_loop;

#if DYNAMIC_BMI2
    if (flags & HUF_flags_bmi2) {
        fallbackFn = HUF_decompress4X1_usingDTable_internal_bmi2;
# if ZSTD_ENABLE_ASM_X86_64_BMI2
        if (!(flags & HUF_flags_disableAsm)) {
            loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
        }
# endif
    } else {
        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
    }
#endif

#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
    if (!(flags & HUF_flags_disableAsm)) {
        loopFn = HUF_decompress4X1_usingDTable_internal_fast_asm_loop;
    }
#endif

    if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
        size_t const ret = HUF_decompress4X1_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
        if (ret != 0)
            return ret;
    }
    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
}

static size_t HUF_decompress4X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                   const void* cSrc, size_t cSrcSize,
                                   void* workSpace, size_t wkspSize, int flags)
{
    const BYTE* ip = (const BYTE*) cSrc;

    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
    if (HUF_isError(hSize)) return hSize;
    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
    ip += hSize; cSrcSize -= hSize;

    return HUF_decompress4X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
}

#endif /* HUF_FORCE_DECOMPRESS_X2 */


#ifndef HUF_FORCE_DECOMPRESS_X1

/* *************************/
/* double-symbols decoding */
/* *************************/

typedef struct { U16 sequence; BYTE nbBits; BYTE length; } HUF_DEltX2;  /* double-symbols decoding */
typedef struct { BYTE symbol; } sortedSymbol_t;
typedef U32 rankValCol_t[HUF_TABLELOG_MAX + 1];
typedef rankValCol_t rankVal_t[HUF_TABLELOG_MAX];

/**
 * Constructs a HUF_DEltX2 in a U32.
 */
static U32 HUF_buildDEltX2U32(U32 symbol, U32 nbBits, U32 baseSeq, int level)
{
    U32 seq;
    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, sequence) == 0);
    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, nbBits) == 2);
    DEBUG_STATIC_ASSERT(offsetof(HUF_DEltX2, length) == 3);
    DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(U32));
    if (MEM_isLittleEndian()) {
        seq = level == 1 ? symbol : (baseSeq + (symbol << 8));
        return seq + (nbBits << 16) + ((U32)level << 24);
    } else {
        seq = level == 1 ? (symbol << 8) : ((baseSeq << 8) + symbol);
        return (seq << 16) + (nbBits << 8) + (U32)level;
    }
}

/**
 * Constructs a HUF_DEltX2.
 */
static HUF_DEltX2 HUF_buildDEltX2(U32 symbol, U32 nbBits, U32 baseSeq, int level)
{
    HUF_DEltX2 DElt;
    U32 const val = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
    DEBUG_STATIC_ASSERT(sizeof(DElt) == sizeof(val));
    ZSTD_memcpy(&DElt, &val, sizeof(val));
    return DElt;
}

/**
 * Constructs 2 HUF_DEltX2s and packs them into a U64.
 */
static U64 HUF_buildDEltX2U64(U32 symbol, U32 nbBits, U16 baseSeq, int level)
{
    U32 DElt = HUF_buildDEltX2U32(symbol, nbBits, baseSeq, level);
    return (U64)DElt + ((U64)DElt << 32);
}

/**
 * Fills the DTable rank with all the symbols from [begin, end) that are each
 * nbBits long.
 *
 * @param DTableRank The start of the rank in the DTable.
 * @param begin The first symbol to fill (inclusive).
 * @param end The last symbol to fill (exclusive).
 * @param nbBits Each symbol is nbBits long.
 * @param tableLog The table log.
 * @param baseSeq If level == 1 { 0 } else { the first level symbol }
 * @param level The level in the table. Must be 1 or 2.
 */
static void HUF_fillDTableX2ForWeight(
    HUF_DEltX2* DTableRank,
    sortedSymbol_t const* begin, sortedSymbol_t const* end,
    U32 nbBits, U32 tableLog,
    U16 baseSeq, int const level)
{
    U32 const length = 1U << ((tableLog - nbBits) & 0x1F /* quiet static-analyzer */);
    const sortedSymbol_t* ptr;
    assert(level >= 1 && level <= 2);
    switch (length) {
    case 1:
        for (ptr = begin; ptr != end; ++ptr) {
            HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
            *DTableRank++ = DElt;
        }
        break;
    case 2:
        for (ptr = begin; ptr != end; ++ptr) {
            HUF_DEltX2 const DElt = HUF_buildDEltX2(ptr->symbol, nbBits, baseSeq, level);
            DTableRank[0] = DElt;
            DTableRank[1] = DElt;
            DTableRank += 2;
        }
        break;
    case 4:
        for (ptr = begin; ptr != end; ++ptr) {
            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
            ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
            ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
            DTableRank += 4;
        }
        break;
    case 8:
        for (ptr = begin; ptr != end; ++ptr) {
            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
            ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
            ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
            ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
            ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
            DTableRank += 8;
        }
        break;
    default:
        for (ptr = begin; ptr != end; ++ptr) {
            U64 const DEltX2 = HUF_buildDEltX2U64(ptr->symbol, nbBits, baseSeq, level);
            HUF_DEltX2* const DTableRankEnd = DTableRank + length;
            for (; DTableRank != DTableRankEnd; DTableRank += 8) {
                ZSTD_memcpy(DTableRank + 0, &DEltX2, sizeof(DEltX2));
                ZSTD_memcpy(DTableRank + 2, &DEltX2, sizeof(DEltX2));
                ZSTD_memcpy(DTableRank + 4, &DEltX2, sizeof(DEltX2));
                ZSTD_memcpy(DTableRank + 6, &DEltX2, sizeof(DEltX2));
            }
        }
        break;
    }
}

/* HUF_fillDTableX2Level2() :
 * `rankValOrigin` must be a table of at least (HUF_TABLELOG_MAX + 1) U32 */
static void HUF_fillDTableX2Level2(HUF_DEltX2* DTable, U32 targetLog, const U32 consumedBits,
                           const U32* rankVal, const int minWeight, const int maxWeight1,
                           const sortedSymbol_t* sortedSymbols, U32 const* rankStart,
                           U32 nbBitsBaseline, U16 baseSeq)
{
    /* Fill skipped values (all positions up to rankVal[minWeight]).
     * These are positions only get a single symbol because the combined weight
     * is too large.
     */
    if (minWeight>1) {
        U32 const length = 1U << ((targetLog - consumedBits) & 0x1F /* quiet static-analyzer */);
        U64 const DEltX2 = HUF_buildDEltX2U64(baseSeq, consumedBits, /* baseSeq */ 0, /* level */ 1);
        int const skipSize = rankVal[minWeight];
        assert(length > 1);
        assert((U32)skipSize < length);
        switch (length) {
        case 2:
            assert(skipSize == 1);
            ZSTD_memcpy(DTable, &DEltX2, sizeof(DEltX2));
            break;
        case 4:
            assert(skipSize <= 4);
            ZSTD_memcpy(DTable + 0, &DEltX2, sizeof(DEltX2));
            ZSTD_memcpy(DTable + 2, &DEltX2, sizeof(DEltX2));
            break;
        default:
            {
                int i;
                for (i = 0; i < skipSize; i += 8) {
                    ZSTD_memcpy(DTable + i + 0, &DEltX2, sizeof(DEltX2));
                    ZSTD_memcpy(DTable + i + 2, &DEltX2, sizeof(DEltX2));
                    ZSTD_memcpy(DTable + i + 4, &DEltX2, sizeof(DEltX2));
                    ZSTD_memcpy(DTable + i + 6, &DEltX2, sizeof(DEltX2));
                }
            }
        }
    }

    /* Fill each of the second level symbols by weight. */
    {
        int w;
        for (w = minWeight; w < maxWeight1; ++w) {
            int const begin = rankStart[w];
            int const end = rankStart[w+1];
            U32 const nbBits = nbBitsBaseline - w;
            U32 const totalBits = nbBits + consumedBits;
            HUF_fillDTableX2ForWeight(
                DTable + rankVal[w],
                sortedSymbols + begin, sortedSymbols + end,
                totalBits, targetLog,
                baseSeq, /* level */ 2);
        }
    }
}

static void HUF_fillDTableX2(HUF_DEltX2* DTable, const U32 targetLog,
                           const sortedSymbol_t* sortedList,
                           const U32* rankStart, rankValCol_t* rankValOrigin, const U32 maxWeight,
                           const U32 nbBitsBaseline)
{
    U32* const rankVal = rankValOrigin[0];
    const int scaleLog = nbBitsBaseline - targetLog;   /* note : targetLog >= srcLog, hence scaleLog <= 1 */
    const U32 minBits  = nbBitsBaseline - maxWeight;
    int w;
    int const wEnd = (int)maxWeight + 1;

    /* Fill DTable in order of weight. */
    for (w = 1; w < wEnd; ++w) {
        int const begin = (int)rankStart[w];
        int const end = (int)rankStart[w+1];
        U32 const nbBits = nbBitsBaseline - w;

        if (targetLog-nbBits >= minBits) {
            /* Enough room for a second symbol. */
            int start = rankVal[w];
            U32 const length = 1U << ((targetLog - nbBits) & 0x1F /* quiet static-analyzer */);
            int minWeight = nbBits + scaleLog;
            int s;
            if (minWeight < 1) minWeight = 1;
            /* Fill the DTable for every symbol of weight w.
             * These symbols get at least 1 second symbol.
             */
            for (s = begin; s != end; ++s) {
                HUF_fillDTableX2Level2(
                    DTable + start, targetLog, nbBits,
                    rankValOrigin[nbBits], minWeight, wEnd,
                    sortedList, rankStart,
                    nbBitsBaseline, sortedList[s].symbol);
                start += length;
            }
        } else {
            /* Only a single symbol. */
            HUF_fillDTableX2ForWeight(
                DTable + rankVal[w],
                sortedList + begin, sortedList + end,
                nbBits, targetLog,
                /* baseSeq */ 0, /* level */ 1);
        }
    }
}

typedef struct {
    rankValCol_t rankVal[HUF_TABLELOG_MAX];
    U32 rankStats[HUF_TABLELOG_MAX + 1];
    U32 rankStart0[HUF_TABLELOG_MAX + 3];
    sortedSymbol_t sortedSymbol[HUF_SYMBOLVALUE_MAX + 1];
    BYTE weightList[HUF_SYMBOLVALUE_MAX + 1];
    U32 calleeWksp[HUF_READ_STATS_WORKSPACE_SIZE_U32];
} HUF_ReadDTableX2_Workspace;

size_t HUF_readDTableX2_wksp(HUF_DTable* DTable,
                       const void* src, size_t srcSize,
                             void* workSpace, size_t wkspSize, int flags)
{
    U32 tableLog, maxW, nbSymbols;
    DTableDesc dtd = HUF_getDTableDesc(DTable);
    U32 maxTableLog = dtd.maxTableLog;
    size_t iSize;
    void* dtPtr = DTable+1;   /* force compiler to avoid strict-aliasing */
    HUF_DEltX2* const dt = (HUF_DEltX2*)dtPtr;
    U32 *rankStart;

    HUF_ReadDTableX2_Workspace* const wksp = (HUF_ReadDTableX2_Workspace*)workSpace;

    if (sizeof(*wksp) > wkspSize) return ERROR(GENERIC);

    rankStart = wksp->rankStart0 + 1;
    ZSTD_memset(wksp->rankStats, 0, sizeof(wksp->rankStats));
    ZSTD_memset(wksp->rankStart0, 0, sizeof(wksp->rankStart0));

    DEBUG_STATIC_ASSERT(sizeof(HUF_DEltX2) == sizeof(HUF_DTable));   /* if compiler fails here, assertion is wrong */
    if (maxTableLog > HUF_TABLELOG_MAX) return ERROR(tableLog_tooLarge);
    /* ZSTD_memset(weightList, 0, sizeof(weightList)); */  /* is not necessary, even though some analyzer complain ... */

    iSize = HUF_readStats_wksp(wksp->weightList, HUF_SYMBOLVALUE_MAX + 1, wksp->rankStats, &nbSymbols, &tableLog, src, srcSize, wksp->calleeWksp, sizeof(wksp->calleeWksp), flags);
    if (HUF_isError(iSize)) return iSize;

    /* check result */
    if (tableLog > maxTableLog) return ERROR(tableLog_tooLarge);   /* DTable can't fit code depth */
    if (tableLog <= HUF_DECODER_FAST_TABLELOG && maxTableLog > HUF_DECODER_FAST_TABLELOG) maxTableLog = HUF_DECODER_FAST_TABLELOG;

    /* find maxWeight */
    for (maxW = tableLog; wksp->rankStats[maxW]==0; maxW--) {}  /* necessarily finds a solution before 0 */

    /* Get start index of each weight */
    {   U32 w, nextRankStart = 0;
        for (w=1; w<maxW+1; w++) {
            U32 curr = nextRankStart;
            nextRankStart += wksp->rankStats[w];
            rankStart[w] = curr;
        }
        rankStart[0] = nextRankStart;   /* put all 0w symbols at the end of sorted list*/
        rankStart[maxW+1] = nextRankStart;
    }

    /* sort symbols by weight */
    {   U32 s;
        for (s=0; s<nbSymbols; s++) {
            U32 const w = wksp->weightList[s];
            U32 const r = rankStart[w]++;
            wksp->sortedSymbol[r].symbol = (BYTE)s;
        }
        rankStart[0] = 0;   /* forget 0w symbols; this is beginning of weight(1) */
    }

    /* Build rankVal */
    {   U32* const rankVal0 = wksp->rankVal[0];
        {   int const rescale = (maxTableLog-tableLog) - 1;   /* tableLog <= maxTableLog */
            U32 nextRankVal = 0;
            U32 w;
            for (w=1; w<maxW+1; w++) {
                U32 curr = nextRankVal;
                nextRankVal += wksp->rankStats[w] << (w+rescale);
                rankVal0[w] = curr;
        }   }
        {   U32 const minBits = tableLog+1 - maxW;
            U32 consumed;
            for (consumed = minBits; consumed < maxTableLog - minBits + 1; consumed++) {
                U32* const rankValPtr = wksp->rankVal[consumed];
                U32 w;
                for (w = 1; w < maxW+1; w++) {
                    rankValPtr[w] = rankVal0[w] >> consumed;
    }   }   }   }

    HUF_fillDTableX2(dt, maxTableLog,
                   wksp->sortedSymbol,
                   wksp->rankStart0, wksp->rankVal, maxW,
                   tableLog+1);

    dtd.tableLog = (BYTE)maxTableLog;
    dtd.tableType = 1;
    ZSTD_memcpy(DTable, &dtd, sizeof(dtd));
    return iSize;
}


FORCE_INLINE_TEMPLATE U32
HUF_decodeSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
{
    size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
    ZSTD_memcpy(op, &dt[val].sequence, 2);
    BIT_skipBits(DStream, dt[val].nbBits);
    return dt[val].length;
}

FORCE_INLINE_TEMPLATE U32
HUF_decodeLastSymbolX2(void* op, BIT_DStream_t* DStream, const HUF_DEltX2* dt, const U32 dtLog)
{
    size_t const val = BIT_lookBitsFast(DStream, dtLog);   /* note : dtLog >= 1 */
    ZSTD_memcpy(op, &dt[val].sequence, 1);
    if (dt[val].length==1) {
        BIT_skipBits(DStream, dt[val].nbBits);
    } else {
        if (DStream->bitsConsumed < (sizeof(DStream->bitContainer)*8)) {
            BIT_skipBits(DStream, dt[val].nbBits);
            if (DStream->bitsConsumed > (sizeof(DStream->bitContainer)*8))
                /* ugly hack; works only because it's the last symbol. Note : can't easily extract nbBits from just this symbol */
                DStream->bitsConsumed = (sizeof(DStream->bitContainer)*8);
        }
    }
    return 1;
}

#define HUF_DECODE_SYMBOLX2_0(ptr, DStreamPtr) \
    do { ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); } while (0)

#define HUF_DECODE_SYMBOLX2_1(ptr, DStreamPtr)                     \
    do {                                                           \
        if (MEM_64bits() || (HUF_TABLELOG_MAX<=12))                \
            ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); \
    } while (0)

#define HUF_DECODE_SYMBOLX2_2(ptr, DStreamPtr)                     \
    do {                                                           \
        if (MEM_64bits())                                          \
            ptr += HUF_decodeSymbolX2(ptr, DStreamPtr, dt, dtLog); \
    } while (0)

HINT_INLINE size_t
HUF_decodeStreamX2(BYTE* p, BIT_DStream_t* bitDPtr, BYTE* const pEnd,
                const HUF_DEltX2* const dt, const U32 dtLog)
{
    BYTE* const pStart = p;

    /* up to 8 symbols at a time */
    if ((size_t)(pEnd - p) >= sizeof(bitDPtr->bitContainer)) {
        if (dtLog <= 11 && MEM_64bits()) {
            /* up to 10 symbols at a time */
            while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-9)) {
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
            }
        } else {
            /* up to 8 symbols at a time */
            while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p < pEnd-(sizeof(bitDPtr->bitContainer)-1))) {
                HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_1(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_2(p, bitDPtr);
                HUF_DECODE_SYMBOLX2_0(p, bitDPtr);
            }
        }
    } else {
        BIT_reloadDStream(bitDPtr);
    }

    /* closer to end : up to 2 symbols at a time */
    if ((size_t)(pEnd - p) >= 2) {
        while ((BIT_reloadDStream(bitDPtr) == BIT_DStream_unfinished) & (p <= pEnd-2))
            HUF_DECODE_SYMBOLX2_0(p, bitDPtr);

        while (p <= pEnd-2)
            HUF_DECODE_SYMBOLX2_0(p, bitDPtr);   /* no need to reload : reached the end of DStream */
    }

    if (p < pEnd)
        p += HUF_decodeLastSymbolX2(p, bitDPtr, dt, dtLog);

    return p-pStart;
}

FORCE_INLINE_TEMPLATE size_t
HUF_decompress1X2_usingDTable_internal_body(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable)
{
    BIT_DStream_t bitD;

    /* Init */
    CHECK_F( BIT_initDStream(&bitD, cSrc, cSrcSize) );

    /* decode */
    {   BYTE* const ostart = (BYTE*) dst;
        BYTE* const oend = ZSTD_maybeNullPtrAdd(ostart, dstSize);
        const void* const dtPtr = DTable+1;   /* force compiler to not use strict-aliasing */
        const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;
        DTableDesc const dtd = HUF_getDTableDesc(DTable);
        HUF_decodeStreamX2(ostart, &bitD, oend, dt, dtd.tableLog);
    }

    /* check */
    if (!BIT_endOfDStream(&bitD)) return ERROR(corruption_detected);

    /* decoded size */
    return dstSize;
}

/* HUF_decompress4X2_usingDTable_internal_body():
 * Conditions:
 * @dstSize >= 6
 */
FORCE_INLINE_TEMPLATE size_t
HUF_decompress4X2_usingDTable_internal_body(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable)
{
    if (cSrcSize < 10) return ERROR(corruption_detected);   /* strict minimum : jump table + 1 byte per stream */
    if (dstSize < 6) return ERROR(corruption_detected);         /* stream 4-split doesn't work */

    {   const BYTE* const istart = (const BYTE*) cSrc;
        BYTE* const ostart = (BYTE*) dst;
        BYTE* const oend = ostart + dstSize;
        BYTE* const olimit = oend - (sizeof(size_t)-1);
        const void* const dtPtr = DTable+1;
        const HUF_DEltX2* const dt = (const HUF_DEltX2*)dtPtr;

        /* Init */
        BIT_DStream_t bitD1;
        BIT_DStream_t bitD2;
        BIT_DStream_t bitD3;
        BIT_DStream_t bitD4;
        size_t const length1 = MEM_readLE16(istart);
        size_t const length2 = MEM_readLE16(istart+2);
        size_t const length3 = MEM_readLE16(istart+4);
        size_t const length4 = cSrcSize - (length1 + length2 + length3 + 6);
        const BYTE* const istart1 = istart + 6;  /* jumpTable */
        const BYTE* const istart2 = istart1 + length1;
        const BYTE* const istart3 = istart2 + length2;
        const BYTE* const istart4 = istart3 + length3;
        size_t const segmentSize = (dstSize+3) / 4;
        BYTE* const opStart2 = ostart + segmentSize;
        BYTE* const opStart3 = opStart2 + segmentSize;
        BYTE* const opStart4 = opStart3 + segmentSize;
        BYTE* op1 = ostart;
        BYTE* op2 = opStart2;
        BYTE* op3 = opStart3;
        BYTE* op4 = opStart4;
        U32 endSignal = 1;
        DTableDesc const dtd = HUF_getDTableDesc(DTable);
        U32 const dtLog = dtd.tableLog;

        if (length4 > cSrcSize) return ERROR(corruption_detected);  /* overflow */
        if (opStart4 > oend) return ERROR(corruption_detected);     /* overflow */
        assert(dstSize >= 6 /* validated above */);
        CHECK_F( BIT_initDStream(&bitD1, istart1, length1) );
        CHECK_F( BIT_initDStream(&bitD2, istart2, length2) );
        CHECK_F( BIT_initDStream(&bitD3, istart3, length3) );
        CHECK_F( BIT_initDStream(&bitD4, istart4, length4) );

        /* 16-32 symbols per loop (4-8 symbols per stream) */
        if ((size_t)(oend - op4) >= sizeof(size_t)) {
            for ( ; (endSignal) & (op4 < olimit); ) {
#if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
                endSignal &= BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished;
                endSignal &= BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished;
                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
                endSignal &= BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished;
                endSignal &= BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished;
#else
                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_1(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_1(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_1(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_1(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_2(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_2(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_2(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_2(op4, &bitD4);
                HUF_DECODE_SYMBOLX2_0(op1, &bitD1);
                HUF_DECODE_SYMBOLX2_0(op2, &bitD2);
                HUF_DECODE_SYMBOLX2_0(op3, &bitD3);
                HUF_DECODE_SYMBOLX2_0(op4, &bitD4);
                endSignal = (U32)LIKELY((U32)
                            (BIT_reloadDStreamFast(&bitD1) == BIT_DStream_unfinished)
                        & (BIT_reloadDStreamFast(&bitD2) == BIT_DStream_unfinished)
                        & (BIT_reloadDStreamFast(&bitD3) == BIT_DStream_unfinished)
                        & (BIT_reloadDStreamFast(&bitD4) == BIT_DStream_unfinished));
#endif
            }
        }

        /* check corruption */
        if (op1 > opStart2) return ERROR(corruption_detected);
        if (op2 > opStart3) return ERROR(corruption_detected);
        if (op3 > opStart4) return ERROR(corruption_detected);
        /* note : op4 already verified within main loop */

        /* finish bitStreams one by one */
        HUF_decodeStreamX2(op1, &bitD1, opStart2, dt, dtLog);
        HUF_decodeStreamX2(op2, &bitD2, opStart3, dt, dtLog);
        HUF_decodeStreamX2(op3, &bitD3, opStart4, dt, dtLog);
        HUF_decodeStreamX2(op4, &bitD4, oend,     dt, dtLog);

        /* check */
        { U32 const endCheck = BIT_endOfDStream(&bitD1) & BIT_endOfDStream(&bitD2) & BIT_endOfDStream(&bitD3) & BIT_endOfDStream(&bitD4);
          if (!endCheck) return ERROR(corruption_detected); }

        /* decoded size */
        return dstSize;
    }
}

#if HUF_NEED_BMI2_FUNCTION
static BMI2_TARGET_ATTRIBUTE
size_t HUF_decompress4X2_usingDTable_internal_bmi2(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable) {
    return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
}
#endif

static
size_t HUF_decompress4X2_usingDTable_internal_default(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable) {
    return HUF_decompress4X2_usingDTable_internal_body(dst, dstSize, cSrc, cSrcSize, DTable);
}

#if ZSTD_ENABLE_ASM_X86_64_BMI2

HUF_ASM_DECL void HUF_decompress4X2_usingDTable_internal_fast_asm_loop(HUF_DecompressFastArgs* args) ZSTDLIB_HIDDEN;

#endif

static HUF_FAST_BMI2_ATTRS
void HUF_decompress4X2_usingDTable_internal_fast_c_loop(HUF_DecompressFastArgs* args)
{
    U64 bits[4];
    BYTE const* ip[4];
    BYTE* op[4];
    BYTE* oend[4];
    HUF_DEltX2 const* const dtable = (HUF_DEltX2 const*)args->dt;
    BYTE const* const ilowest = args->ilowest;

    /* Copy the arguments to local registers. */
    ZSTD_memcpy(&bits, &args->bits, sizeof(bits));
    ZSTD_memcpy((void*)(&ip), &args->ip, sizeof(ip));
    ZSTD_memcpy(&op, &args->op, sizeof(op));

    oend[0] = op[1];
    oend[1] = op[2];
    oend[2] = op[3];
    oend[3] = args->oend;

    assert(MEM_isLittleEndian());
    assert(!MEM_32bits());

    for (;;) {
        BYTE* olimit;
        int stream;

        /* Assert loop preconditions */
#ifndef NDEBUG
        for (stream = 0; stream < 4; ++stream) {
            assert(op[stream] <= oend[stream]);
            assert(ip[stream] >= ilowest);
        }
#endif
        /* Compute olimit */
        {
            /* Each loop does 5 table lookups for each of the 4 streams.
             * Each table lookup consumes up to 11 bits of input, and produces
             * up to 2 bytes of output.
             */
            /* We can consume up to 7 bytes of input per iteration per stream.
             * We also know that each input pointer is >= ip[0]. So we can run
             * iters loops before running out of input.
             */
            size_t iters = (size_t)(ip[0] - ilowest) / 7;
            /* Each iteration can produce up to 10 bytes of output per stream.
             * Each output stream my advance at different rates. So take the
             * minimum number of safe iterations among all the output streams.
             */
            for (stream = 0; stream < 4; ++stream) {
                size_t const oiters = (size_t)(oend[stream] - op[stream]) / 10;
                iters = MIN(iters, oiters);
            }

            /* Each iteration produces at least 5 output symbols. So until
             * op[3] crosses olimit, we know we haven't executed iters
             * iterations yet. This saves us maintaining an iters counter,
             * at the expense of computing the remaining # of iterations
             * more frequently.
             */
            olimit = op[3] + (iters * 5);

            /* Exit the fast decoding loop once we reach the end. */
            if (op[3] == olimit)
                break;

            /* Exit the decoding loop if any input pointer has crossed the
             * previous one. This indicates corruption, and a precondition
             * to our loop is that ip[i] >= ip[0].
             */
            for (stream = 1; stream < 4; ++stream) {
                if (ip[stream] < ip[stream - 1])
                    goto _out;
            }
        }

#ifndef NDEBUG
        for (stream = 1; stream < 4; ++stream) {
            assert(ip[stream] >= ip[stream - 1]);
        }
#endif

#define HUF_4X2_DECODE_SYMBOL(_stream, _decode3)                      \
    do {                                                              \
        if ((_decode3) || (_stream) != 3) {                           \
            int const index = (int)(bits[(_stream)] >> 53);           \
            HUF_DEltX2 const entry = dtable[index];                   \
            MEM_write16(op[(_stream)], entry.sequence); \
            bits[(_stream)] <<= (entry.nbBits) & 0x3F;                \
            op[(_stream)] += (entry.length);                          \
        }                                                             \
    } while (0)

#define HUF_4X2_RELOAD_STREAM(_stream)                                  \
    do {                                                                \
        HUF_4X2_DECODE_SYMBOL(3, 1);                                    \
        {                                                               \
            int const ctz = ZSTD_countTrailingZeros64(bits[(_stream)]); \
            int const nbBits = ctz & 7;                                 \
            int const nbBytes = ctz >> 3;                               \
            ip[(_stream)] -= nbBytes;                                   \
            bits[(_stream)] = MEM_read64(ip[(_stream)]) | 1;            \
            bits[(_stream)] <<= nbBits;                                 \
        }                                                               \
    } while (0)

        /* Manually unroll the loop because compilers don't consistently
         * unroll the inner loops, which destroys performance.
         */
        do {
            /* Decode 5 symbols from each of the first 3 streams.
             * The final stream will be decoded during the reload phase
             * to reduce register pressure.
             */
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);
            HUF_4X_FOR_EACH_STREAM_WITH_VAR(HUF_4X2_DECODE_SYMBOL, 0);

            /* Decode one symbol from the final stream */
            HUF_4X2_DECODE_SYMBOL(3, 1);

            /* Decode 4 symbols from the final stream & reload bitstreams.
             * The final stream is reloaded last, meaning that all 5 symbols
             * are decoded from the final stream before it is reloaded.
             */
            HUF_4X_FOR_EACH_STREAM(HUF_4X2_RELOAD_STREAM);
        } while (op[3] < olimit);
    }

#undef HUF_4X2_DECODE_SYMBOL
#undef HUF_4X2_RELOAD_STREAM

_out:

    /* Save the final values of each of the state variables back to args. */
    ZSTD_memcpy(&args->bits, &bits, sizeof(bits));
    ZSTD_memcpy((void*)(&args->ip), &ip, sizeof(ip));
    ZSTD_memcpy(&args->op, &op, sizeof(op));
}


static HUF_FAST_BMI2_ATTRS size_t
HUF_decompress4X2_usingDTable_internal_fast(
          void* dst,  size_t dstSize,
    const void* cSrc, size_t cSrcSize,
    const HUF_DTable* DTable,
    HUF_DecompressFastLoopFn loopFn) {
    void const* dt = DTable + 1;
    const BYTE* const ilowest = (const BYTE*)cSrc;
    BYTE* const oend = ZSTD_maybeNullPtrAdd((BYTE*)dst, dstSize);
    HUF_DecompressFastArgs args;
    {
        size_t const ret = HUF_DecompressFastArgs_init(&args, dst, dstSize, cSrc, cSrcSize, DTable);
        FORWARD_IF_ERROR(ret, "Failed to init asm args");
        if (ret == 0)
            return 0;
    }

    assert(args.ip[0] >= args.ilowest);
    loopFn(&args);

    /* note : op4 already verified within main loop */
    assert(args.ip[0] >= ilowest);
    assert(args.ip[1] >= ilowest);
    assert(args.ip[2] >= ilowest);
    assert(args.ip[3] >= ilowest);
    assert(args.op[3] <= oend);

    assert(ilowest == args.ilowest);
    assert(ilowest + 6 == args.iend[0]);
    (void)ilowest;

    /* finish bitStreams one by one */
    {
        size_t const segmentSize = (dstSize+3) / 4;
        BYTE* segmentEnd = (BYTE*)dst;
        int i;
        for (i = 0; i < 4; ++i) {
            BIT_DStream_t bit;
            if (segmentSize <= (size_t)(oend - segmentEnd))
                segmentEnd += segmentSize;
            else
                segmentEnd = oend;
            FORWARD_IF_ERROR(HUF_initRemainingDStream(&bit, &args, i, segmentEnd), "corruption");
            args.op[i] += HUF_decodeStreamX2(args.op[i], &bit, segmentEnd, (HUF_DEltX2 const*)dt, HUF_DECODER_FAST_TABLELOG);
            if (args.op[i] != segmentEnd)
                return ERROR(corruption_detected);
        }
    }

    /* decoded size */
    return dstSize;
}

static size_t HUF_decompress4X2_usingDTable_internal(void* dst, size_t dstSize, void const* cSrc,
                    size_t cSrcSize, HUF_DTable const* DTable, int flags)
{
    HUF_DecompressUsingDTableFn fallbackFn = HUF_decompress4X2_usingDTable_internal_default;
    HUF_DecompressFastLoopFn loopFn = HUF_decompress4X2_usingDTable_internal_fast_c_loop;

#if DYNAMIC_BMI2
    if (flags & HUF_flags_bmi2) {
        fallbackFn = HUF_decompress4X2_usingDTable_internal_bmi2;
# if ZSTD_ENABLE_ASM_X86_64_BMI2
        if (!(flags & HUF_flags_disableAsm)) {
            loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
        }
# endif
    } else {
        return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
    }
#endif

#if ZSTD_ENABLE_ASM_X86_64_BMI2 && defined(__BMI2__)
    if (!(flags & HUF_flags_disableAsm)) {
        loopFn = HUF_decompress4X2_usingDTable_internal_fast_asm_loop;
    }
#endif

    if (HUF_ENABLE_FAST_DECODE && !(flags & HUF_flags_disableFast)) {
        size_t const ret = HUF_decompress4X2_usingDTable_internal_fast(dst, dstSize, cSrc, cSrcSize, DTable, loopFn);
        if (ret != 0)
            return ret;
    }
    return fallbackFn(dst, dstSize, cSrc, cSrcSize, DTable);
}

HUF_DGEN(HUF_decompress1X2_usingDTable_internal)

size_t HUF_decompress1X2_DCtx_wksp(HUF_DTable* DCtx, void* dst, size_t dstSize,
                                   const void* cSrc, size_t cSrcSize,
                                   void* workSpace, size_t wkspSize, int flags)
{
    const BYTE* ip = (const BYTE*) cSrc;

    size_t const hSize = HUF_readDTableX2_wksp(DCtx, cSrc, cSrcSize,
                                               workSpace, wkspSize, flags);
    if (HUF_isError(hSize)) return hSize;
    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
    ip += hSize; cSrcSize -= hSize;

    return HUF_decompress1X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, DCtx, flags);
}

static size_t HUF_decompress4X2_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                   const void* cSrc, size_t cSrcSize,
                                   void* workSpace, size_t wkspSize, int flags)
{
    const BYTE* ip = (const BYTE*) cSrc;

    size_t hSize = HUF_readDTableX2_wksp(dctx, cSrc, cSrcSize,
                                         workSpace, wkspSize, flags);
    if (HUF_isError(hSize)) return hSize;
    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
    ip += hSize; cSrcSize -= hSize;

    return HUF_decompress4X2_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
}

#endif /* HUF_FORCE_DECOMPRESS_X1 */


/* ***********************************/
/* Universal decompression selectors */
/* ***********************************/


#if !defined(HUF_FORCE_DECOMPRESS_X1) && !defined(HUF_FORCE_DECOMPRESS_X2)
typedef struct { U32 tableTime; U32 decode256Time; } algo_time_t;
static const algo_time_t algoTime[16 /* Quantization */][2 /* single, double */] =
{
    /* single, double, quad */
    {{0,0}, {1,1}},  /* Q==0 : impossible */
    {{0,0}, {1,1}},  /* Q==1 : impossible */
    {{ 150,216}, { 381,119}},   /* Q == 2 : 12-18% */
    {{ 170,205}, { 514,112}},   /* Q == 3 : 18-25% */
    {{ 177,199}, { 539,110}},   /* Q == 4 : 25-32% */
    {{ 197,194}, { 644,107}},   /* Q == 5 : 32-38% */
    {{ 221,192}, { 735,107}},   /* Q == 6 : 38-44% */
    {{ 256,189}, { 881,106}},   /* Q == 7 : 44-50% */
    {{ 359,188}, {1167,109}},   /* Q == 8 : 50-56% */
    {{ 582,187}, {1570,114}},   /* Q == 9 : 56-62% */
    {{ 688,187}, {1712,122}},   /* Q ==10 : 62-69% */
    {{ 825,186}, {1965,136}},   /* Q ==11 : 69-75% */
    {{ 976,185}, {2131,150}},   /* Q ==12 : 75-81% */
    {{1180,186}, {2070,175}},   /* Q ==13 : 81-87% */
    {{1377,185}, {1731,202}},   /* Q ==14 : 87-93% */
    {{1412,185}, {1695,202}},   /* Q ==15 : 93-99% */
};
#endif

/** HUF_selectDecoder() :
 *  Tells which decoder is likely to decode faster,
 *  based on a set of pre-computed metrics.
 * @return : 0==HUF_decompress4X1, 1==HUF_decompress4X2 .
 *  Assumption : 0 < dstSize <= 128 KB */
U32 HUF_selectDecoder (size_t dstSize, size_t cSrcSize)
{
    assert(dstSize > 0);
    assert(dstSize <= 128*1024);
#if defined(HUF_FORCE_DECOMPRESS_X1)
    (void)dstSize;
    (void)cSrcSize;
    return 0;
#elif defined(HUF_FORCE_DECOMPRESS_X2)
    (void)dstSize;
    (void)cSrcSize;
    return 1;
#else
    /* decoder timing evaluation */
    {   U32 const Q = (cSrcSize >= dstSize) ? 15 : (U32)(cSrcSize * 16 / dstSize);   /* Q < 16 */
        U32 const D256 = (U32)(dstSize >> 8);
        U32 const DTime0 = algoTime[Q][0].tableTime + (algoTime[Q][0].decode256Time * D256);
        U32 DTime1 = algoTime[Q][1].tableTime + (algoTime[Q][1].decode256Time * D256);
        DTime1 += DTime1 >> 5;  /* small advantage to algorithm using less memory, to reduce cache eviction */
        return DTime1 < DTime0;
    }
#endif
}

size_t HUF_decompress1X_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize,
                                  const void* cSrc, size_t cSrcSize,
                                  void* workSpace, size_t wkspSize, int flags)
{
    /* validation checks */
    if (dstSize == 0) return ERROR(dstSize_tooSmall);
    if (cSrcSize > dstSize) return ERROR(corruption_detected);   /* invalid */
    if (cSrcSize == dstSize) { ZSTD_memcpy(dst, cSrc, dstSize); return dstSize; }   /* not compressed */
    if (cSrcSize == 1) { ZSTD_memset(dst, *(const BYTE*)cSrc, dstSize); return dstSize; }   /* RLE */

    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
#if defined(HUF_FORCE_DECOMPRESS_X1)
        (void)algoNb;
        assert(algoNb == 0);
        return HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
                                cSrcSize, workSpace, wkspSize, flags);
#elif defined(HUF_FORCE_DECOMPRESS_X2)
        (void)algoNb;
        assert(algoNb == 1);
        return HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
                                cSrcSize, workSpace, wkspSize, flags);
#else
        return algoNb ? HUF_decompress1X2_DCtx_wksp(dctx, dst, dstSize, cSrc,
                                cSrcSize, workSpace, wkspSize, flags):
                        HUF_decompress1X1_DCtx_wksp(dctx, dst, dstSize, cSrc,
                                cSrcSize, workSpace, wkspSize, flags);
#endif
    }
}


size_t HUF_decompress1X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
{
    DTableDesc const dtd = HUF_getDTableDesc(DTable);
#if defined(HUF_FORCE_DECOMPRESS_X1)
    (void)dtd;
    assert(dtd.tableType == 0);
    return HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#elif defined(HUF_FORCE_DECOMPRESS_X2)
    (void)dtd;
    assert(dtd.tableType == 1);
    return HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#else
    return dtd.tableType ? HUF_decompress1X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
                           HUF_decompress1X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#endif
}

#ifndef HUF_FORCE_DECOMPRESS_X2
size_t HUF_decompress1X1_DCtx_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
{
    const BYTE* ip = (const BYTE*) cSrc;

    size_t const hSize = HUF_readDTableX1_wksp(dctx, cSrc, cSrcSize, workSpace, wkspSize, flags);
    if (HUF_isError(hSize)) return hSize;
    if (hSize >= cSrcSize) return ERROR(srcSize_wrong);
    ip += hSize; cSrcSize -= hSize;

    return HUF_decompress1X1_usingDTable_internal(dst, dstSize, ip, cSrcSize, dctx, flags);
}
#endif

size_t HUF_decompress4X_usingDTable(void* dst, size_t maxDstSize, const void* cSrc, size_t cSrcSize, const HUF_DTable* DTable, int flags)
{
    DTableDesc const dtd = HUF_getDTableDesc(DTable);
#if defined(HUF_FORCE_DECOMPRESS_X1)
    (void)dtd;
    assert(dtd.tableType == 0);
    return HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#elif defined(HUF_FORCE_DECOMPRESS_X2)
    (void)dtd;
    assert(dtd.tableType == 1);
    return HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#else
    return dtd.tableType ? HUF_decompress4X2_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags) :
                           HUF_decompress4X1_usingDTable_internal(dst, maxDstSize, cSrc, cSrcSize, DTable, flags);
#endif
}

size_t HUF_decompress4X_hufOnly_wksp(HUF_DTable* dctx, void* dst, size_t dstSize, const void* cSrc, size_t cSrcSize, void* workSpace, size_t wkspSize, int flags)
{
    /* validation checks */
    if (dstSize == 0) return ERROR(dstSize_tooSmall);
    if (cSrcSize == 0) return ERROR(corruption_detected);

    {   U32 const algoNb = HUF_selectDecoder(dstSize, cSrcSize);
#if defined(HUF_FORCE_DECOMPRESS_X1)
        (void)algoNb;
        assert(algoNb == 0);
        return HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
#elif defined(HUF_FORCE_DECOMPRESS_X2)
        (void)algoNb;
        assert(algoNb == 1);
        return HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
#else
        return algoNb ? HUF_decompress4X2_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags) :
                        HUF_decompress4X1_DCtx_wksp(dctx, dst, dstSize, cSrc, cSrcSize, workSpace, wkspSize, flags);
#endif
    }
}
/**** ended inlining decompress/huf_decompress.c ****/
/**** start inlining decompress/zstd_ddict.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* zstd_ddict.c :
 * concentrates all logic that needs to know the internals of ZSTD_DDict object */

/*-*******************************************************
*  Dependencies
*********************************************************/
/**** skipping file: ../common/allocations.h ****/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/cpu.h ****/
/**** skipping file: ../common/mem.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** start inlining zstd_decompress_internal.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/* zstd_decompress_internal:
 * objects and definitions shared within lib/decompress modules */

 #ifndef ZSTD_DECOMPRESS_INTERNAL_H
 #define ZSTD_DECOMPRESS_INTERNAL_H


/*-*******************************************************
 *  Dependencies
 *********************************************************/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/zstd_internal.h ****/


/*-*******************************************************
 *  Constants
 *********************************************************/
static UNUSED_ATTR const U32 LL_base[MaxLL+1] = {
                 0,    1,    2,     3,     4,     5,     6,      7,
                 8,    9,   10,    11,    12,    13,    14,     15,
                16,   18,   20,    22,    24,    28,    32,     40,
                48,   64, 0x80, 0x100, 0x200, 0x400, 0x800, 0x1000,
                0x2000, 0x4000, 0x8000, 0x10000 };

static UNUSED_ATTR const U32 OF_base[MaxOff+1] = {
                 0,        1,       1,       5,     0xD,     0x1D,     0x3D,     0x7D,
                 0xFD,   0x1FD,   0x3FD,   0x7FD,   0xFFD,   0x1FFD,   0x3FFD,   0x7FFD,
                 0xFFFD, 0x1FFFD, 0x3FFFD, 0x7FFFD, 0xFFFFD, 0x1FFFFD, 0x3FFFFD, 0x7FFFFD,
                 0xFFFFFD, 0x1FFFFFD, 0x3FFFFFD, 0x7FFFFFD, 0xFFFFFFD, 0x1FFFFFFD, 0x3FFFFFFD, 0x7FFFFFFD };

static UNUSED_ATTR const U8 OF_bits[MaxOff+1] = {
                     0,  1,  2,  3,  4,  5,  6,  7,
                     8,  9, 10, 11, 12, 13, 14, 15,
                    16, 17, 18, 19, 20, 21, 22, 23,
                    24, 25, 26, 27, 28, 29, 30, 31 };

static UNUSED_ATTR const U32 ML_base[MaxML+1] = {
                     3,  4,  5,    6,     7,     8,     9,    10,
                    11, 12, 13,   14,    15,    16,    17,    18,
                    19, 20, 21,   22,    23,    24,    25,    26,
                    27, 28, 29,   30,    31,    32,    33,    34,
                    35, 37, 39,   41,    43,    47,    51,    59,
                    67, 83, 99, 0x83, 0x103, 0x203, 0x403, 0x803,
                    0x1003, 0x2003, 0x4003, 0x8003, 0x10003 };


/*-*******************************************************
 *  Decompression types
 *********************************************************/
 typedef struct {
     U32 fastMode;
     U32 tableLog;
 } ZSTD_seqSymbol_header;

 typedef struct {
     U16  nextState;
     BYTE nbAdditionalBits;
     BYTE nbBits;
     U32  baseValue;
 } ZSTD_seqSymbol;

 #define SEQSYMBOL_TABLE_SIZE(log)   (1 + (1 << (log)))

#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE (sizeof(S16) * (MaxSeq + 1) + (1u << MaxFSELog) + sizeof(U64))
#define ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32 ((ZSTD_BUILD_FSE_TABLE_WKSP_SIZE + sizeof(U32) - 1) / sizeof(U32))
#define ZSTD_HUFFDTABLE_CAPACITY_LOG 12

typedef struct {
    ZSTD_seqSymbol LLTable[SEQSYMBOL_TABLE_SIZE(LLFSELog)];    /* Note : Space reserved for FSE Tables */
    ZSTD_seqSymbol OFTable[SEQSYMBOL_TABLE_SIZE(OffFSELog)];   /* is also used as temporary workspace while building hufTable during DDict creation */
    ZSTD_seqSymbol MLTable[SEQSYMBOL_TABLE_SIZE(MLFSELog)];    /* and therefore must be at least HUF_DECOMPRESS_WORKSPACE_SIZE large */
    HUF_DTable hufTable[HUF_DTABLE_SIZE(ZSTD_HUFFDTABLE_CAPACITY_LOG)];  /* can accommodate HUF_decompress4X */
    U32 rep[ZSTD_REP_NUM];
    U32 workspace[ZSTD_BUILD_FSE_TABLE_WKSP_SIZE_U32];
} ZSTD_entropyDTables_t;

typedef enum { ZSTDds_getFrameHeaderSize, ZSTDds_decodeFrameHeader,
               ZSTDds_decodeBlockHeader, ZSTDds_decompressBlock,
               ZSTDds_decompressLastBlock, ZSTDds_checkChecksum,
               ZSTDds_decodeSkippableHeader, ZSTDds_skipFrame } ZSTD_dStage;

typedef enum { zdss_init=0, zdss_loadHeader,
               zdss_read, zdss_load, zdss_flush } ZSTD_dStreamStage;

typedef enum {
    ZSTD_use_indefinitely = -1,  /* Use the dictionary indefinitely */
    ZSTD_dont_use = 0,           /* Do not use the dictionary (if one exists free it) */
    ZSTD_use_once = 1            /* Use the dictionary once and set to ZSTD_dont_use */
} ZSTD_dictUses_e;

/* Hashset for storing references to multiple ZSTD_DDict within ZSTD_DCtx */
typedef struct {
    const ZSTD_DDict** ddictPtrTable;
    size_t ddictPtrTableSize;
    size_t ddictPtrCount;
} ZSTD_DDictHashSet;

#ifndef ZSTD_DECODER_INTERNAL_BUFFER
#  define ZSTD_DECODER_INTERNAL_BUFFER  (1 << 16)
#endif

#define ZSTD_LBMIN 64
#define ZSTD_LBMAX (128 << 10)

/* extra buffer, compensates when dst is not large enough to store litBuffer */
#define ZSTD_LITBUFFEREXTRASIZE  BOUNDED(ZSTD_LBMIN, ZSTD_DECODER_INTERNAL_BUFFER, ZSTD_LBMAX)

typedef enum {
    ZSTD_not_in_dst = 0,  /* Stored entirely within litExtraBuffer */
    ZSTD_in_dst = 1,           /* Stored entirely within dst (in memory after current output write) */
    ZSTD_split = 2            /* Split between litExtraBuffer and dst */
} ZSTD_litLocation_e;

struct ZSTD_DCtx_s
{
    const ZSTD_seqSymbol* LLTptr;
    const ZSTD_seqSymbol* MLTptr;
    const ZSTD_seqSymbol* OFTptr;
    const HUF_DTable* HUFptr;
    ZSTD_entropyDTables_t entropy;
    U32 workspace[HUF_DECOMPRESS_WORKSPACE_SIZE_U32];   /* space needed when building huffman tables */
    const void* previousDstEnd;   /* detect continuity */
    const void* prefixStart;      /* start of current segment */
    const void* virtualStart;     /* virtual start of previous segment if it was just before current one */
    const void* dictEnd;          /* end of previous segment */
    size_t expected;
    ZSTD_FrameHeader fParams;
    U64 processedCSize;
    U64 decodedSize;
    blockType_e bType;            /* used in ZSTD_decompressContinue(), store blockType between block header decoding and block decompression stages */
    ZSTD_dStage stage;
    U32 litEntropy;
    U32 fseEntropy;
    XXH64_state_t xxhState;
    size_t headerSize;
    ZSTD_format_e format;
    ZSTD_forceIgnoreChecksum_e forceIgnoreChecksum;   /* User specified: if == 1, will ignore checksums in compressed frame. Default == 0 */
    U32 validateChecksum;         /* if == 1, will validate checksum. Is == 1 if (fParams.checksumFlag == 1) and (forceIgnoreChecksum == 0). */
    const BYTE* litPtr;
    ZSTD_customMem customMem;
    size_t litSize;
    size_t rleSize;
    size_t staticSize;
    int isFrameDecompression;
#if DYNAMIC_BMI2
    int bmi2;                     /* == 1 if the CPU supports BMI2 and 0 otherwise. CPU support is determined dynamically once per context lifetime. */
#endif

    /* dictionary */
    ZSTD_DDict* ddictLocal;
    const ZSTD_DDict* ddict;     /* set by ZSTD_initDStream_usingDDict(), or ZSTD_DCtx_refDDict() */
    U32 dictID;
    int ddictIsCold;             /* if == 1 : dictionary is "new" for working context, and presumed "cold" (not in cpu cache) */
    ZSTD_dictUses_e dictUses;
    ZSTD_DDictHashSet* ddictSet;                    /* Hash set for multiple ddicts */
    ZSTD_refMultipleDDicts_e refMultipleDDicts;     /* User specified: if == 1, will allow references to multiple DDicts. Default == 0 (disabled) */
    int disableHufAsm;
    int maxBlockSizeParam;

    /* streaming */
    ZSTD_dStreamStage streamStage;
    char*  inBuff;
    size_t inBuffSize;
    size_t inPos;
    size_t maxWindowSize;
    char*  outBuff;
    size_t outBuffSize;
    size_t outStart;
    size_t outEnd;
    size_t lhSize;
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
    void* legacyContext;
    U32 previousLegacyVersion;
    U32 legacyVersion;
#endif
    U32 hostageByte;
    int noForwardProgress;
    ZSTD_bufferMode_e outBufferMode;
    ZSTD_outBuffer expectedOutBuffer;

    /* workspace */
    BYTE* litBuffer;
    const BYTE* litBufferEnd;
    ZSTD_litLocation_e litBufferLocation;
    BYTE litExtraBuffer[ZSTD_LITBUFFEREXTRASIZE + WILDCOPY_OVERLENGTH]; /* literal buffer can be split between storage within dst and within this scratch buffer */
    BYTE headerBuffer[ZSTD_FRAMEHEADERSIZE_MAX];

    size_t oversizedDuration;

#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    void const* dictContentBeginForFuzzing;
    void const* dictContentEndForFuzzing;
#endif

    /* Tracing */
#if ZSTD_TRACE
    ZSTD_TraceCtx traceCtx;
#endif
};  /* typedef'd to ZSTD_DCtx within "zstd.h" */

MEM_STATIC int ZSTD_DCtx_get_bmi2(const struct ZSTD_DCtx_s *dctx) {
#if DYNAMIC_BMI2
    return dctx->bmi2;
#else
    (void)dctx;
    return 0;
#endif
}

/*-*******************************************************
 *  Shared internal functions
 *********************************************************/

/*! ZSTD_loadDEntropy() :
 *  dict : must point at beginning of a valid zstd dictionary.
 * @return : size of dictionary header (size of magic number + dict ID + entropy tables) */
size_t ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
                   const void* const dict, size_t const dictSize);

/*! ZSTD_checkContinuity() :
 *  check if next `dst` follows previous position, where decompression ended.
 *  If yes, do nothing (continue on current segment).
 *  If not, classify previous segment as "external dictionary", and start a new segment.
 *  This function cannot fail. */
void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize);


#endif /* ZSTD_DECOMPRESS_INTERNAL_H */
/**** ended inlining zstd_decompress_internal.h ****/
/**** start inlining zstd_ddict.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


#ifndef ZSTD_DDICT_H
#define ZSTD_DDICT_H

/*-*******************************************************
 *  Dependencies
 *********************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../zstd.h ****/


/*-*******************************************************
 *  Interface
 *********************************************************/

/* note: several prototypes are already published in `zstd.h` :
 * ZSTD_createDDict()
 * ZSTD_createDDict_byReference()
 * ZSTD_createDDict_advanced()
 * ZSTD_freeDDict()
 * ZSTD_initStaticDDict()
 * ZSTD_sizeof_DDict()
 * ZSTD_estimateDDictSize()
 * ZSTD_getDictID_fromDict()
 */

const void* ZSTD_DDict_dictContent(const ZSTD_DDict* ddict);
size_t ZSTD_DDict_dictSize(const ZSTD_DDict* ddict);

void ZSTD_copyDDictParameters(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict);


#endif /* ZSTD_DDICT_H */
/**** ended inlining zstd_ddict.h ****/

#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
#error Using excluded file: ../legacy/zstd_legacy.h (re-amalgamate source to fix)
#endif


/*-*******************************************************
*  Types
*********************************************************/
struct ZSTD_DDict_s {
    void* dictBuffer;
    const void* dictContent;
    size_t dictSize;
    ZSTD_entropyDTables_t entropy;
    U32 dictID;
    U32 entropyPresent;
    ZSTD_customMem cMem;
};  /* typedef'd to ZSTD_DDict within "zstd.h" */

const void* ZSTD_DDict_dictContent(const ZSTD_DDict* ddict)
{
    assert(ddict != NULL);
    return ddict->dictContent;
}

size_t ZSTD_DDict_dictSize(const ZSTD_DDict* ddict)
{
    assert(ddict != NULL);
    return ddict->dictSize;
}

void ZSTD_copyDDictParameters(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
{
    DEBUGLOG(4, "ZSTD_copyDDictParameters");
    assert(dctx != NULL);
    assert(ddict != NULL);
    dctx->dictID = ddict->dictID;
    dctx->prefixStart = ddict->dictContent;
    dctx->virtualStart = ddict->dictContent;
    dctx->dictEnd = (const BYTE*)ddict->dictContent + ddict->dictSize;
    dctx->previousDstEnd = dctx->dictEnd;
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    dctx->dictContentBeginForFuzzing = dctx->prefixStart;
    dctx->dictContentEndForFuzzing = dctx->previousDstEnd;
#endif
    if (ddict->entropyPresent) {
        dctx->litEntropy = 1;
        dctx->fseEntropy = 1;
        dctx->LLTptr = ddict->entropy.LLTable;
        dctx->MLTptr = ddict->entropy.MLTable;
        dctx->OFTptr = ddict->entropy.OFTable;
        dctx->HUFptr = ddict->entropy.hufTable;
        dctx->entropy.rep[0] = ddict->entropy.rep[0];
        dctx->entropy.rep[1] = ddict->entropy.rep[1];
        dctx->entropy.rep[2] = ddict->entropy.rep[2];
    } else {
        dctx->litEntropy = 0;
        dctx->fseEntropy = 0;
    }
}


static size_t
ZSTD_loadEntropy_intoDDict(ZSTD_DDict* ddict,
                           ZSTD_dictContentType_e dictContentType)
{
    ddict->dictID = 0;
    ddict->entropyPresent = 0;
    if (dictContentType == ZSTD_dct_rawContent) return 0;

    if (ddict->dictSize < 8) {
        if (dictContentType == ZSTD_dct_fullDict)
            return ERROR(dictionary_corrupted);   /* only accept specified dictionaries */
        return 0;   /* pure content mode */
    }
    {   U32 const magic = MEM_readLE32(ddict->dictContent);
        if (magic != ZSTD_MAGIC_DICTIONARY) {
            if (dictContentType == ZSTD_dct_fullDict)
                return ERROR(dictionary_corrupted);   /* only accept specified dictionaries */
            return 0;   /* pure content mode */
        }
    }
    ddict->dictID = MEM_readLE32((const char*)ddict->dictContent + ZSTD_FRAMEIDSIZE);

    /* load entropy tables */
    RETURN_ERROR_IF(ZSTD_isError(ZSTD_loadDEntropy(
            &ddict->entropy, ddict->dictContent, ddict->dictSize)),
        dictionary_corrupted, "");
    ddict->entropyPresent = 1;
    return 0;
}


static size_t ZSTD_initDDict_internal(ZSTD_DDict* ddict,
                                      const void* dict, size_t dictSize,
                                      ZSTD_dictLoadMethod_e dictLoadMethod,
                                      ZSTD_dictContentType_e dictContentType)
{
    if ((dictLoadMethod == ZSTD_dlm_byRef) || (!dict) || (!dictSize)) {
        ddict->dictBuffer = NULL;
        ddict->dictContent = dict;
        if (!dict) dictSize = 0;
    } else {
        void* const internalBuffer = ZSTD_customMalloc(dictSize, ddict->cMem);
        ddict->dictBuffer = internalBuffer;
        ddict->dictContent = internalBuffer;
        if (!internalBuffer) return ERROR(memory_allocation);
        ZSTD_memcpy(internalBuffer, dict, dictSize);
    }
    ddict->dictSize = dictSize;
    ddict->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */

    /* parse dictionary content */
    FORWARD_IF_ERROR( ZSTD_loadEntropy_intoDDict(ddict, dictContentType) , "");

    return 0;
}

ZSTD_DDict* ZSTD_createDDict_advanced(const void* dict, size_t dictSize,
                                      ZSTD_dictLoadMethod_e dictLoadMethod,
                                      ZSTD_dictContentType_e dictContentType,
                                      ZSTD_customMem customMem)
{
    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;

    {   ZSTD_DDict* const ddict = (ZSTD_DDict*) ZSTD_customMalloc(sizeof(ZSTD_DDict), customMem);
        if (ddict == NULL) return NULL;
        ddict->cMem = customMem;
        {   size_t const initResult = ZSTD_initDDict_internal(ddict,
                                            dict, dictSize,
                                            dictLoadMethod, dictContentType);
            if (ZSTD_isError(initResult)) {
                ZSTD_freeDDict(ddict);
                return NULL;
        }   }
        return ddict;
    }
}

/*! ZSTD_createDDict() :
*   Create a digested dictionary, to start decompression without startup delay.
*   `dict` content is copied inside DDict.
*   Consequently, `dict` can be released after `ZSTD_DDict` creation */
ZSTD_DDict* ZSTD_createDDict(const void* dict, size_t dictSize)
{
    ZSTD_customMem const allocator = { NULL, NULL, NULL };
    return ZSTD_createDDict_advanced(dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto, allocator);
}

/*! ZSTD_createDDict_byReference() :
 *  Create a digested dictionary, to start decompression without startup delay.
 *  Dictionary content is simply referenced, it will be accessed during decompression.
 *  Warning : dictBuffer must outlive DDict (DDict must be freed before dictBuffer) */
ZSTD_DDict* ZSTD_createDDict_byReference(const void* dictBuffer, size_t dictSize)
{
    ZSTD_customMem const allocator = { NULL, NULL, NULL };
    return ZSTD_createDDict_advanced(dictBuffer, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto, allocator);
}


const ZSTD_DDict* ZSTD_initStaticDDict(
                                void* sBuffer, size_t sBufferSize,
                                const void* dict, size_t dictSize,
                                ZSTD_dictLoadMethod_e dictLoadMethod,
                                ZSTD_dictContentType_e dictContentType)
{
    size_t const neededSpace = sizeof(ZSTD_DDict)
                             + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize);
    ZSTD_DDict* const ddict = (ZSTD_DDict*)sBuffer;
    assert(sBuffer != NULL);
    assert(dict != NULL);
    if ((size_t)sBuffer & 7) return NULL;   /* 8-aligned */
    if (sBufferSize < neededSpace) return NULL;
    if (dictLoadMethod == ZSTD_dlm_byCopy) {
        ZSTD_memcpy(ddict+1, dict, dictSize);  /* local copy */
        dict = ddict+1;
    }
    if (ZSTD_isError( ZSTD_initDDict_internal(ddict,
                                              dict, dictSize,
                                              ZSTD_dlm_byRef, dictContentType) ))
        return NULL;
    return ddict;
}


size_t ZSTD_freeDDict(ZSTD_DDict* ddict)
{
    if (ddict==NULL) return 0;   /* support free on NULL */
    {   ZSTD_customMem const cMem = ddict->cMem;
        ZSTD_customFree(ddict->dictBuffer, cMem);
        ZSTD_customFree(ddict, cMem);
        return 0;
    }
}

/*! ZSTD_estimateDDictSize() :
 *  Estimate amount of memory that will be needed to create a dictionary for decompression.
 *  Note : dictionary created by reference using ZSTD_dlm_byRef are smaller */
size_t ZSTD_estimateDDictSize(size_t dictSize, ZSTD_dictLoadMethod_e dictLoadMethod)
{
    return sizeof(ZSTD_DDict) + (dictLoadMethod == ZSTD_dlm_byRef ? 0 : dictSize);
}

size_t ZSTD_sizeof_DDict(const ZSTD_DDict* ddict)
{
    if (ddict==NULL) return 0;   /* support sizeof on NULL */
    return sizeof(*ddict) + (ddict->dictBuffer ? ddict->dictSize : 0) ;
}

/*! ZSTD_getDictID_fromDDict() :
 *  Provides the dictID of the dictionary loaded into `ddict`.
 *  If @return == 0, the dictionary is not conformant to Zstandard specification, or empty.
 *  Non-conformant dictionaries can still be loaded, but as content-only dictionaries. */
unsigned ZSTD_getDictID_fromDDict(const ZSTD_DDict* ddict)
{
    if (ddict==NULL) return 0;
    return ddict->dictID;
}
/**** ended inlining decompress/zstd_ddict.c ****/
/**** start inlining decompress/zstd_decompress.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/* ***************************************************************
*  Tuning parameters
*****************************************************************/
/*!
 * HEAPMODE :
 * Select how default decompression function ZSTD_decompress() allocates its context,
 * on stack (0), or into heap (1, default; requires malloc()).
 * Note that functions with explicit context such as ZSTD_decompressDCtx() are unaffected.
 */
#ifndef ZSTD_HEAPMODE
#  define ZSTD_HEAPMODE 1
#endif

/*!
*  LEGACY_SUPPORT :
*  if set to 1+, ZSTD_decompress() can decode older formats (v0.1+)
*/
#ifndef ZSTD_LEGACY_SUPPORT
#  define ZSTD_LEGACY_SUPPORT 0
#endif

/*!
 *  MAXWINDOWSIZE_DEFAULT :
 *  maximum window size accepted by DStream __by default__.
 *  Frames requiring more memory will be rejected.
 *  It's possible to set a different limit using ZSTD_DCtx_setMaxWindowSize().
 */
#ifndef ZSTD_MAXWINDOWSIZE_DEFAULT
#  define ZSTD_MAXWINDOWSIZE_DEFAULT (((U32)1 << ZSTD_WINDOWLOG_LIMIT_DEFAULT) + 1)
#endif

/*!
 *  NO_FORWARD_PROGRESS_MAX :
 *  maximum allowed nb of calls to ZSTD_decompressStream()
 *  without any forward progress
 *  (defined as: no byte read from input, and no byte flushed to output)
 *  before triggering an error.
 */
#ifndef ZSTD_NO_FORWARD_PROGRESS_MAX
#  define ZSTD_NO_FORWARD_PROGRESS_MAX 16
#endif


/*-*******************************************************
*  Dependencies
*********************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/allocations.h ****/
/**** skipping file: ../common/error_private.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/bits.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: ../common/xxhash.h ****/
/**** skipping file: zstd_decompress_internal.h ****/
/**** skipping file: zstd_ddict.h ****/
/**** start inlining zstd_decompress_block.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


#ifndef ZSTD_DEC_BLOCK_H
#define ZSTD_DEC_BLOCK_H

/*-*******************************************************
 *  Dependencies
 *********************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../zstd.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: zstd_decompress_internal.h ****/


/* ===   Prototypes   === */

/* note: prototypes already published within `zstd.h` :
 * ZSTD_decompressBlock()
 */

/* note: prototypes already published within `zstd_internal.h` :
 * ZSTD_getcBlockSize()
 * ZSTD_decodeSeqHeaders()
 */


 /* Streaming state is used to inform allocation of the literal buffer */
typedef enum {
    not_streaming = 0,
    is_streaming = 1
} streaming_operation;

/* ZSTD_decompressBlock_internal() :
 * decompress block, starting at `src`,
 * into destination buffer `dst`.
 * @return : decompressed block size,
 *           or an error code (which can be tested using ZSTD_isError())
 */
size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                               void* dst, size_t dstCapacity,
                         const void* src, size_t srcSize, const streaming_operation streaming);

/* ZSTD_buildFSETable() :
 * generate FSE decoding table for one symbol (ll, ml or off)
 * this function must be called with valid parameters only
 * (dt is large enough, normalizedCounter distribution total is a power of 2, max is within range, etc.)
 * in which case it cannot fail.
 * The workspace must be 4-byte aligned and at least ZSTD_BUILD_FSE_TABLE_WKSP_SIZE bytes, which is
 * defined in zstd_decompress_internal.h.
 * Internal use only.
 */
void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
             const short* normalizedCounter, unsigned maxSymbolValue,
             const U32* baseValue, const U8* nbAdditionalBits,
                   unsigned tableLog, void* wksp, size_t wkspSize,
                   int bmi2);

/* Internal definition of ZSTD_decompressBlock() to avoid deprecation warnings. */
size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
                            void* dst, size_t dstCapacity,
                      const void* src, size_t srcSize);


#endif /* ZSTD_DEC_BLOCK_H */
/**** ended inlining zstd_decompress_block.h ****/

#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
#error Using excluded file: ../legacy/zstd_legacy.h (re-amalgamate source to fix)
#endif


/*************************************
 * Multiple DDicts Hashset internals *
 *************************************/

#define DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT 4
#define DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT 3  /* These two constants represent SIZE_MULT/COUNT_MULT load factor without using a float.
                                                    * Currently, that means a 0.75 load factor.
                                                    * So, if count * COUNT_MULT / size * SIZE_MULT != 0, then we've exceeded
                                                    * the load factor of the ddict hash set.
                                                    */

#define DDICT_HASHSET_TABLE_BASE_SIZE 64
#define DDICT_HASHSET_RESIZE_FACTOR 2

/* Hash function to determine starting position of dict insertion within the table
 * Returns an index between [0, hashSet->ddictPtrTableSize]
 */
static size_t ZSTD_DDictHashSet_getIndex(const ZSTD_DDictHashSet* hashSet, U32 dictID) {
    const U64 hash = XXH64(&dictID, sizeof(U32), 0);
    /* DDict ptr table size is a multiple of 2, use size - 1 as mask to get index within [0, hashSet->ddictPtrTableSize) */
    return hash & (hashSet->ddictPtrTableSize - 1);
}

/* Adds DDict to a hashset without resizing it.
 * If inserting a DDict with a dictID that already exists in the set, replaces the one in the set.
 * Returns 0 if successful, or a zstd error code if something went wrong.
 */
static size_t ZSTD_DDictHashSet_emplaceDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict) {
    const U32 dictID = ZSTD_getDictID_fromDDict(ddict);
    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
    RETURN_ERROR_IF(hashSet->ddictPtrCount == hashSet->ddictPtrTableSize, GENERIC, "Hash set is full!");
    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
    while (hashSet->ddictPtrTable[idx] != NULL) {
        /* Replace existing ddict if inserting ddict with same dictID */
        if (ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]) == dictID) {
            DEBUGLOG(4, "DictID already exists, replacing rather than adding");
            hashSet->ddictPtrTable[idx] = ddict;
            return 0;
        }
        idx &= idxRangeMask;
        idx++;
    }
    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
    hashSet->ddictPtrTable[idx] = ddict;
    hashSet->ddictPtrCount++;
    return 0;
}

/* Expands hash table by factor of DDICT_HASHSET_RESIZE_FACTOR and
 * rehashes all values, allocates new table, frees old table.
 * Returns 0 on success, otherwise a zstd error code.
 */
static size_t ZSTD_DDictHashSet_expand(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
    size_t newTableSize = hashSet->ddictPtrTableSize * DDICT_HASHSET_RESIZE_FACTOR;
    const ZSTD_DDict** newTable = (const ZSTD_DDict**)ZSTD_customCalloc(sizeof(ZSTD_DDict*) * newTableSize, customMem);
    const ZSTD_DDict** oldTable = hashSet->ddictPtrTable;
    size_t oldTableSize = hashSet->ddictPtrTableSize;
    size_t i;

    DEBUGLOG(4, "Expanding DDict hash table! Old size: %zu new size: %zu", oldTableSize, newTableSize);
    RETURN_ERROR_IF(!newTable, memory_allocation, "Expanded hashset allocation failed!");
    hashSet->ddictPtrTable = newTable;
    hashSet->ddictPtrTableSize = newTableSize;
    hashSet->ddictPtrCount = 0;
    for (i = 0; i < oldTableSize; ++i) {
        if (oldTable[i] != NULL) {
            FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, oldTable[i]), "");
        }
    }
    ZSTD_customFree((void*)oldTable, customMem);
    DEBUGLOG(4, "Finished re-hash");
    return 0;
}

/* Fetches a DDict with the given dictID
 * Returns the ZSTD_DDict* with the requested dictID. If it doesn't exist, then returns NULL.
 */
static const ZSTD_DDict* ZSTD_DDictHashSet_getDDict(ZSTD_DDictHashSet* hashSet, U32 dictID) {
    size_t idx = ZSTD_DDictHashSet_getIndex(hashSet, dictID);
    const size_t idxRangeMask = hashSet->ddictPtrTableSize - 1;
    DEBUGLOG(4, "Hashed index: for dictID: %u is %zu", dictID, idx);
    for (;;) {
        size_t currDictID = ZSTD_getDictID_fromDDict(hashSet->ddictPtrTable[idx]);
        if (currDictID == dictID || currDictID == 0) {
            /* currDictID == 0 implies a NULL ddict entry */
            break;
        } else {
            idx &= idxRangeMask;    /* Goes to start of table when we reach the end */
            idx++;
        }
    }
    DEBUGLOG(4, "Final idx after probing for dictID %u is: %zu", dictID, idx);
    return hashSet->ddictPtrTable[idx];
}

/* Allocates space for and returns a ddict hash set
 * The hash set's ZSTD_DDict* table has all values automatically set to NULL to begin with.
 * Returns NULL if allocation failed.
 */
static ZSTD_DDictHashSet* ZSTD_createDDictHashSet(ZSTD_customMem customMem) {
    ZSTD_DDictHashSet* ret = (ZSTD_DDictHashSet*)ZSTD_customMalloc(sizeof(ZSTD_DDictHashSet), customMem);
    DEBUGLOG(4, "Allocating new hash set");
    if (!ret)
        return NULL;
    ret->ddictPtrTable = (const ZSTD_DDict**)ZSTD_customCalloc(DDICT_HASHSET_TABLE_BASE_SIZE * sizeof(ZSTD_DDict*), customMem);
    if (!ret->ddictPtrTable) {
        ZSTD_customFree(ret, customMem);
        return NULL;
    }
    ret->ddictPtrTableSize = DDICT_HASHSET_TABLE_BASE_SIZE;
    ret->ddictPtrCount = 0;
    return ret;
}

/* Frees the table of ZSTD_DDict* within a hashset, then frees the hashset itself.
 * Note: The ZSTD_DDict* within the table are NOT freed.
 */
static void ZSTD_freeDDictHashSet(ZSTD_DDictHashSet* hashSet, ZSTD_customMem customMem) {
    DEBUGLOG(4, "Freeing ddict hash set");
    if (hashSet && hashSet->ddictPtrTable) {
        ZSTD_customFree((void*)hashSet->ddictPtrTable, customMem);
    }
    if (hashSet) {
        ZSTD_customFree(hashSet, customMem);
    }
}

/* Public function: Adds a DDict into the ZSTD_DDictHashSet, possibly triggering a resize of the hash set.
 * Returns 0 on success, or a ZSTD error.
 */
static size_t ZSTD_DDictHashSet_addDDict(ZSTD_DDictHashSet* hashSet, const ZSTD_DDict* ddict, ZSTD_customMem customMem) {
    DEBUGLOG(4, "Adding dict ID: %u to hashset with - Count: %zu Tablesize: %zu", ZSTD_getDictID_fromDDict(ddict), hashSet->ddictPtrCount, hashSet->ddictPtrTableSize);
    if (hashSet->ddictPtrCount * DDICT_HASHSET_MAX_LOAD_FACTOR_COUNT_MULT / hashSet->ddictPtrTableSize * DDICT_HASHSET_MAX_LOAD_FACTOR_SIZE_MULT != 0) {
        FORWARD_IF_ERROR(ZSTD_DDictHashSet_expand(hashSet, customMem), "");
    }
    FORWARD_IF_ERROR(ZSTD_DDictHashSet_emplaceDDict(hashSet, ddict), "");
    return 0;
}

/*-*************************************************************
*   Context management
***************************************************************/
size_t ZSTD_sizeof_DCtx (const ZSTD_DCtx* dctx)
{
    if (dctx==NULL) return 0;   /* support sizeof NULL */
    return sizeof(*dctx)
           + ZSTD_sizeof_DDict(dctx->ddictLocal)
           + dctx->inBuffSize + dctx->outBuffSize;
}

size_t ZSTD_estimateDCtxSize(void) { return sizeof(ZSTD_DCtx); }


static size_t ZSTD_startingInputLength(ZSTD_format_e format)
{
    size_t const startingInputLength = ZSTD_FRAMEHEADERSIZE_PREFIX(format);
    /* only supports formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless */
    assert( (format == ZSTD_f_zstd1) || (format == ZSTD_f_zstd1_magicless) );
    return startingInputLength;
}

static void ZSTD_DCtx_resetParameters(ZSTD_DCtx* dctx)
{
    assert(dctx->streamStage == zdss_init);
    dctx->format = ZSTD_f_zstd1;
    dctx->maxWindowSize = ZSTD_MAXWINDOWSIZE_DEFAULT;
    dctx->outBufferMode = ZSTD_bm_buffered;
    dctx->forceIgnoreChecksum = ZSTD_d_validateChecksum;
    dctx->refMultipleDDicts = ZSTD_rmd_refSingleDDict;
    dctx->disableHufAsm = 0;
    dctx->maxBlockSizeParam = 0;
}

static void ZSTD_initDCtx_internal(ZSTD_DCtx* dctx)
{
    dctx->staticSize  = 0;
    dctx->ddict       = NULL;
    dctx->ddictLocal  = NULL;
    dctx->dictEnd     = NULL;
    dctx->ddictIsCold = 0;
    dctx->dictUses = ZSTD_dont_use;
    dctx->inBuff      = NULL;
    dctx->inBuffSize  = 0;
    dctx->outBuffSize = 0;
    dctx->streamStage = zdss_init;
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
    dctx->legacyContext = NULL;
    dctx->previousLegacyVersion = 0;
#endif
    dctx->noForwardProgress = 0;
    dctx->oversizedDuration = 0;
    dctx->isFrameDecompression = 1;
#if DYNAMIC_BMI2
    dctx->bmi2 = ZSTD_cpuSupportsBmi2();
#endif
    dctx->ddictSet = NULL;
    ZSTD_DCtx_resetParameters(dctx);
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    dctx->dictContentEndForFuzzing = NULL;
#endif
}

ZSTD_DCtx* ZSTD_initStaticDCtx(void *workspace, size_t workspaceSize)
{
    ZSTD_DCtx* const dctx = (ZSTD_DCtx*) workspace;

    if ((size_t)workspace & 7) return NULL;  /* 8-aligned */
    if (workspaceSize < sizeof(ZSTD_DCtx)) return NULL;  /* minimum size */

    ZSTD_initDCtx_internal(dctx);
    dctx->staticSize = workspaceSize;
    dctx->inBuff = (char*)(dctx+1);
    return dctx;
}

static ZSTD_DCtx* ZSTD_createDCtx_internal(ZSTD_customMem customMem) {
    if ((!customMem.customAlloc) ^ (!customMem.customFree)) return NULL;

    {   ZSTD_DCtx* const dctx = (ZSTD_DCtx*)ZSTD_customMalloc(sizeof(*dctx), customMem);
        if (!dctx) return NULL;
        dctx->customMem = customMem;
        ZSTD_initDCtx_internal(dctx);
        return dctx;
    }
}

ZSTD_DCtx* ZSTD_createDCtx_advanced(ZSTD_customMem customMem)
{
    return ZSTD_createDCtx_internal(customMem);
}

ZSTD_DCtx* ZSTD_createDCtx(void)
{
    DEBUGLOG(3, "ZSTD_createDCtx");
    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
}

static void ZSTD_clearDict(ZSTD_DCtx* dctx)
{
    ZSTD_freeDDict(dctx->ddictLocal);
    dctx->ddictLocal = NULL;
    dctx->ddict = NULL;
    dctx->dictUses = ZSTD_dont_use;
}

size_t ZSTD_freeDCtx(ZSTD_DCtx* dctx)
{
    if (dctx==NULL) return 0;   /* support free on NULL */
    RETURN_ERROR_IF(dctx->staticSize, memory_allocation, "not compatible with static DCtx");
    {   ZSTD_customMem const cMem = dctx->customMem;
        ZSTD_clearDict(dctx);
        ZSTD_customFree(dctx->inBuff, cMem);
        dctx->inBuff = NULL;
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
        if (dctx->legacyContext)
            ZSTD_freeLegacyStreamContext(dctx->legacyContext, dctx->previousLegacyVersion);
#endif
        if (dctx->ddictSet) {
            ZSTD_freeDDictHashSet(dctx->ddictSet, cMem);
            dctx->ddictSet = NULL;
        }
        ZSTD_customFree(dctx, cMem);
        return 0;
    }
}

/* no longer useful */
void ZSTD_copyDCtx(ZSTD_DCtx* dstDCtx, const ZSTD_DCtx* srcDCtx)
{
    size_t const toCopy = (size_t)((char*)(&dstDCtx->inBuff) - (char*)dstDCtx);
    ZSTD_memcpy(dstDCtx, srcDCtx, toCopy);  /* no need to copy workspace */
}

/* Given a dctx with a digested frame params, re-selects the correct ZSTD_DDict based on
 * the requested dict ID from the frame. If there exists a reference to the correct ZSTD_DDict, then
 * accordingly sets the ddict to be used to decompress the frame.
 *
 * If no DDict is found, then no action is taken, and the ZSTD_DCtx::ddict remains as-is.
 *
 * ZSTD_d_refMultipleDDicts must be enabled for this function to be called.
 */
static void ZSTD_DCtx_selectFrameDDict(ZSTD_DCtx* dctx) {
    assert(dctx->refMultipleDDicts && dctx->ddictSet);
    DEBUGLOG(4, "Adjusting DDict based on requested dict ID from frame");
    if (dctx->ddict) {
        const ZSTD_DDict* frameDDict = ZSTD_DDictHashSet_getDDict(dctx->ddictSet, dctx->fParams.dictID);
        if (frameDDict) {
            DEBUGLOG(4, "DDict found!");
            ZSTD_clearDict(dctx);
            dctx->dictID = dctx->fParams.dictID;
            dctx->ddict = frameDDict;
            dctx->dictUses = ZSTD_use_indefinitely;
        }
    }
}


/*-*************************************************************
 *   Frame header decoding
 ***************************************************************/

/*! ZSTD_isFrame() :
 *  Tells if the content of `buffer` starts with a valid Frame Identifier.
 *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
 *  Note 2 : Legacy Frame Identifiers are considered valid only if Legacy Support is enabled.
 *  Note 3 : Skippable Frame Identifiers are considered valid. */
unsigned ZSTD_isFrame(const void* buffer, size_t size)
{
    if (size < ZSTD_FRAMEIDSIZE) return 0;
    {   U32 const magic = MEM_readLE32(buffer);
        if (magic == ZSTD_MAGICNUMBER) return 1;
        if ((magic & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) return 1;
    }
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
    if (ZSTD_isLegacy(buffer, size)) return 1;
#endif
    return 0;
}

/*! ZSTD_isSkippableFrame() :
 *  Tells if the content of `buffer` starts with a valid Frame Identifier for a skippable frame.
 *  Note : Frame Identifier is 4 bytes. If `size < 4`, @return will always be 0.
 */
unsigned ZSTD_isSkippableFrame(const void* buffer, size_t size)
{
    if (size < ZSTD_FRAMEIDSIZE) return 0;
    {   U32 const magic = MEM_readLE32(buffer);
        if ((magic & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) return 1;
    }
    return 0;
}

/** ZSTD_frameHeaderSize_internal() :
 *  srcSize must be large enough to reach header size fields.
 *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless.
 * @return : size of the Frame Header
 *           or an error code, which can be tested with ZSTD_isError() */
static size_t ZSTD_frameHeaderSize_internal(const void* src, size_t srcSize, ZSTD_format_e format)
{
    size_t const minInputSize = ZSTD_startingInputLength(format);
    RETURN_ERROR_IF(srcSize < minInputSize, srcSize_wrong, "");

    {   BYTE const fhd = ((const BYTE*)src)[minInputSize-1];
        U32 const dictID= fhd & 3;
        U32 const singleSegment = (fhd >> 5) & 1;
        U32 const fcsId = fhd >> 6;
        return minInputSize + !singleSegment
             + ZSTD_did_fieldSize[dictID] + ZSTD_fcs_fieldSize[fcsId]
             + (singleSegment && !fcsId);
    }
}

/** ZSTD_frameHeaderSize() :
 *  srcSize must be >= ZSTD_frameHeaderSize_prefix.
 * @return : size of the Frame Header,
 *           or an error code (if srcSize is too small) */
size_t ZSTD_frameHeaderSize(const void* src, size_t srcSize)
{
    return ZSTD_frameHeaderSize_internal(src, srcSize, ZSTD_f_zstd1);
}


/** ZSTD_getFrameHeader_advanced() :
 *  decode Frame Header, or require larger `srcSize`.
 *  note : only works for formats ZSTD_f_zstd1 and ZSTD_f_zstd1_magicless
 * @return : 0, `zfhPtr` is correctly filled,
 *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
**           or an error code, which can be tested using ZSTD_isError() */
size_t ZSTD_getFrameHeader_advanced(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize, ZSTD_format_e format)
{
    const BYTE* ip = (const BYTE*)src;
    size_t const minInputSize = ZSTD_startingInputLength(format);

    DEBUGLOG(5, "ZSTD_getFrameHeader_advanced: minInputSize = %zu, srcSize = %zu", minInputSize, srcSize);

    if (srcSize > 0) {
        /* note : technically could be considered an assert(), since it's an invalid entry */
        RETURN_ERROR_IF(src==NULL, GENERIC, "invalid parameter : src==NULL, but srcSize>0");
    }
    if (srcSize < minInputSize) {
        if (srcSize > 0 && format != ZSTD_f_zstd1_magicless) {
            /* when receiving less than @minInputSize bytes,
             * control these bytes at least correspond to a supported magic number
             * in order to error out early if they don't.
            **/
            size_t const toCopy = MIN(4, srcSize);
            unsigned char hbuf[4]; MEM_writeLE32(hbuf, ZSTD_MAGICNUMBER);
            assert(src != NULL);
            ZSTD_memcpy(hbuf, src, toCopy);
            if ( MEM_readLE32(hbuf) != ZSTD_MAGICNUMBER ) {
                /* not a zstd frame : let's check if it's a skippable frame */
                MEM_writeLE32(hbuf, ZSTD_MAGIC_SKIPPABLE_START);
                ZSTD_memcpy(hbuf, src, toCopy);
                if ((MEM_readLE32(hbuf) & ZSTD_MAGIC_SKIPPABLE_MASK) != ZSTD_MAGIC_SKIPPABLE_START) {
                    RETURN_ERROR(prefix_unknown,
                                "first bytes don't correspond to any supported magic number");
        }   }   }
        return minInputSize;
    }

    ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));   /* not strictly necessary, but static analyzers may not understand that zfhPtr will be read only if return value is zero, since they are 2 different signals */
    if ( (format != ZSTD_f_zstd1_magicless)
      && (MEM_readLE32(src) != ZSTD_MAGICNUMBER) ) {
        if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
            /* skippable frame */
            if (srcSize < ZSTD_SKIPPABLEHEADERSIZE)
                return ZSTD_SKIPPABLEHEADERSIZE; /* magic number + frame length */
            ZSTD_memset(zfhPtr, 0, sizeof(*zfhPtr));
            zfhPtr->frameType = ZSTD_skippableFrame;
            zfhPtr->dictID = MEM_readLE32(src) - ZSTD_MAGIC_SKIPPABLE_START;
            zfhPtr->headerSize = ZSTD_SKIPPABLEHEADERSIZE;
            zfhPtr->frameContentSize = MEM_readLE32((const char *)src + ZSTD_FRAMEIDSIZE);
            return 0;
        }
        RETURN_ERROR(prefix_unknown, "");
    }

    /* ensure there is enough `srcSize` to fully read/decode frame header */
    {   size_t const fhsize = ZSTD_frameHeaderSize_internal(src, srcSize, format);
        if (srcSize < fhsize) return fhsize;
        zfhPtr->headerSize = (U32)fhsize;
    }

    {   BYTE const fhdByte = ip[minInputSize-1];
        size_t pos = minInputSize;
        U32 const dictIDSizeCode = fhdByte&3;
        U32 const checksumFlag = (fhdByte>>2)&1;
        U32 const singleSegment = (fhdByte>>5)&1;
        U32 const fcsID = fhdByte>>6;
        U64 windowSize = 0;
        U32 dictID = 0;
        U64 frameContentSize = ZSTD_CONTENTSIZE_UNKNOWN;
        RETURN_ERROR_IF((fhdByte & 0x08) != 0, frameParameter_unsupported,
                        "reserved bits, must be zero");

        if (!singleSegment) {
            BYTE const wlByte = ip[pos++];
            U32 const windowLog = (wlByte >> 3) + ZSTD_WINDOWLOG_ABSOLUTEMIN;
            RETURN_ERROR_IF(windowLog > ZSTD_WINDOWLOG_MAX, frameParameter_windowTooLarge, "");
            windowSize = (1ULL << windowLog);
            windowSize += (windowSize >> 3) * (wlByte&7);
        }
        switch(dictIDSizeCode)
        {
            default:
                assert(0);  /* impossible */
                ZSTD_FALLTHROUGH;
            case 0 : break;
            case 1 : dictID = ip[pos]; pos++; break;
            case 2 : dictID = MEM_readLE16(ip+pos); pos+=2; break;
            case 3 : dictID = MEM_readLE32(ip+pos); pos+=4; break;
        }
        switch(fcsID)
        {
            default:
                assert(0);  /* impossible */
                ZSTD_FALLTHROUGH;
            case 0 : if (singleSegment) frameContentSize = ip[pos]; break;
            case 1 : frameContentSize = MEM_readLE16(ip+pos)+256; break;
            case 2 : frameContentSize = MEM_readLE32(ip+pos); break;
            case 3 : frameContentSize = MEM_readLE64(ip+pos); break;
        }
        if (singleSegment) windowSize = frameContentSize;

        zfhPtr->frameType = ZSTD_frame;
        zfhPtr->frameContentSize = frameContentSize;
        zfhPtr->windowSize = windowSize;
        zfhPtr->blockSizeMax = (unsigned) MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
        zfhPtr->dictID = dictID;
        zfhPtr->checksumFlag = checksumFlag;
    }
    return 0;
}

/** ZSTD_getFrameHeader() :
 *  decode Frame Header, or require larger `srcSize`.
 *  note : this function does not consume input, it only reads it.
 * @return : 0, `zfhPtr` is correctly filled,
 *          >0, `srcSize` is too small, value is wanted `srcSize` amount,
 *           or an error code, which can be tested using ZSTD_isError() */
size_t ZSTD_getFrameHeader(ZSTD_FrameHeader* zfhPtr, const void* src, size_t srcSize)
{
    return ZSTD_getFrameHeader_advanced(zfhPtr, src, srcSize, ZSTD_f_zstd1);
}

/** ZSTD_getFrameContentSize() :
 *  compatible with legacy mode
 * @return : decompressed size of the single frame pointed to be `src` if known, otherwise
 *         - ZSTD_CONTENTSIZE_UNKNOWN if the size cannot be determined
 *         - ZSTD_CONTENTSIZE_ERROR if an error occurred (e.g. invalid magic number, srcSize too small) */
unsigned long long ZSTD_getFrameContentSize(const void *src, size_t srcSize)
{
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
    if (ZSTD_isLegacy(src, srcSize)) {
        unsigned long long const ret = ZSTD_getDecompressedSize_legacy(src, srcSize);
        return ret == 0 ? ZSTD_CONTENTSIZE_UNKNOWN : ret;
    }
#endif
    {   ZSTD_FrameHeader zfh;
        if (ZSTD_getFrameHeader(&zfh, src, srcSize) != 0)
            return ZSTD_CONTENTSIZE_ERROR;
        if (zfh.frameType == ZSTD_skippableFrame) {
            return 0;
        } else {
            return zfh.frameContentSize;
    }   }
}

static size_t readSkippableFrameSize(void const* src, size_t srcSize)
{
    size_t const skippableHeaderSize = ZSTD_SKIPPABLEHEADERSIZE;
    U32 sizeU32;

    RETURN_ERROR_IF(srcSize < ZSTD_SKIPPABLEHEADERSIZE, srcSize_wrong, "");

    sizeU32 = MEM_readLE32((BYTE const*)src + ZSTD_FRAMEIDSIZE);
    RETURN_ERROR_IF((U32)(sizeU32 + ZSTD_SKIPPABLEHEADERSIZE) < sizeU32,
                    frameParameter_unsupported, "");
    {   size_t const skippableSize = skippableHeaderSize + sizeU32;
        RETURN_ERROR_IF(skippableSize > srcSize, srcSize_wrong, "");
        return skippableSize;
    }
}

/*! ZSTD_readSkippableFrame() :
 * Retrieves content of a skippable frame, and writes it to dst buffer.
 *
 * The parameter magicVariant will receive the magicVariant that was supplied when the frame was written,
 * i.e. magicNumber - ZSTD_MAGIC_SKIPPABLE_START.  This can be NULL if the caller is not interested
 * in the magicVariant.
 *
 * Returns an error if destination buffer is not large enough, or if this is not a valid skippable frame.
 *
 * @return : number of bytes written or a ZSTD error.
 */
size_t ZSTD_readSkippableFrame(void* dst, size_t dstCapacity,
                               unsigned* magicVariant,  /* optional, can be NULL */
                         const void* src, size_t srcSize)
{
    RETURN_ERROR_IF(srcSize < ZSTD_SKIPPABLEHEADERSIZE, srcSize_wrong, "");

    {   U32 const magicNumber = MEM_readLE32(src);
        size_t skippableFrameSize = readSkippableFrameSize(src, srcSize);
        size_t skippableContentSize = skippableFrameSize - ZSTD_SKIPPABLEHEADERSIZE;

        /* check input validity */
        RETURN_ERROR_IF(!ZSTD_isSkippableFrame(src, srcSize), frameParameter_unsupported, "");
        RETURN_ERROR_IF(skippableFrameSize < ZSTD_SKIPPABLEHEADERSIZE || skippableFrameSize > srcSize, srcSize_wrong, "");
        RETURN_ERROR_IF(skippableContentSize > dstCapacity, dstSize_tooSmall, "");

        /* deliver payload */
        if (skippableContentSize > 0  && dst != NULL)
            ZSTD_memcpy(dst, (const BYTE *)src + ZSTD_SKIPPABLEHEADERSIZE, skippableContentSize);
        if (magicVariant != NULL)
            *magicVariant = magicNumber - ZSTD_MAGIC_SKIPPABLE_START;
        return skippableContentSize;
    }
}

/** ZSTD_findDecompressedSize() :
 *  `srcSize` must be the exact length of some number of ZSTD compressed and/or
 *      skippable frames
 *  note: compatible with legacy mode
 * @return : decompressed size of the frames contained */
unsigned long long ZSTD_findDecompressedSize(const void* src, size_t srcSize)
{
    unsigned long long totalDstSize = 0;

    while (srcSize >= ZSTD_startingInputLength(ZSTD_f_zstd1)) {
        U32 const magicNumber = MEM_readLE32(src);

        if ((magicNumber & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
            size_t const skippableSize = readSkippableFrameSize(src, srcSize);
            if (ZSTD_isError(skippableSize)) return ZSTD_CONTENTSIZE_ERROR;
            assert(skippableSize <= srcSize);

            src = (const BYTE *)src + skippableSize;
            srcSize -= skippableSize;
            continue;
        }

        {   unsigned long long const fcs = ZSTD_getFrameContentSize(src, srcSize);
            if (fcs >= ZSTD_CONTENTSIZE_ERROR) return fcs;

            if (totalDstSize + fcs < totalDstSize)
                return ZSTD_CONTENTSIZE_ERROR; /* check for overflow */
            totalDstSize += fcs;
        }
        /* skip to next frame */
        {   size_t const frameSrcSize = ZSTD_findFrameCompressedSize(src, srcSize);
            if (ZSTD_isError(frameSrcSize)) return ZSTD_CONTENTSIZE_ERROR;
            assert(frameSrcSize <= srcSize);

            src = (const BYTE *)src + frameSrcSize;
            srcSize -= frameSrcSize;
        }
    }  /* while (srcSize >= ZSTD_frameHeaderSize_prefix) */

    if (srcSize) return ZSTD_CONTENTSIZE_ERROR;

    return totalDstSize;
}

/** ZSTD_getDecompressedSize() :
 *  compatible with legacy mode
 * @return : decompressed size if known, 0 otherwise
             note : 0 can mean any of the following :
                   - frame content is empty
                   - decompressed size field is not present in frame header
                   - frame header unknown / not supported
                   - frame header not complete (`srcSize` too small) */
unsigned long long ZSTD_getDecompressedSize(const void* src, size_t srcSize)
{
    unsigned long long const ret = ZSTD_getFrameContentSize(src, srcSize);
    ZSTD_STATIC_ASSERT(ZSTD_CONTENTSIZE_ERROR < ZSTD_CONTENTSIZE_UNKNOWN);
    return (ret >= ZSTD_CONTENTSIZE_ERROR) ? 0 : ret;
}


/** ZSTD_decodeFrameHeader() :
 * `headerSize` must be the size provided by ZSTD_frameHeaderSize().
 * If multiple DDict references are enabled, also will choose the correct DDict to use.
 * @return : 0 if success, or an error code, which can be tested using ZSTD_isError() */
static size_t ZSTD_decodeFrameHeader(ZSTD_DCtx* dctx, const void* src, size_t headerSize)
{
    size_t const result = ZSTD_getFrameHeader_advanced(&(dctx->fParams), src, headerSize, dctx->format);
    if (ZSTD_isError(result)) return result;    /* invalid header */
    RETURN_ERROR_IF(result>0, srcSize_wrong, "headerSize too small");

    /* Reference DDict requested by frame if dctx references multiple ddicts */
    if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts && dctx->ddictSet) {
        ZSTD_DCtx_selectFrameDDict(dctx);
    }

#ifndef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    /* Skip the dictID check in fuzzing mode, because it makes the search
     * harder.
     */
    RETURN_ERROR_IF(dctx->fParams.dictID && (dctx->dictID != dctx->fParams.dictID),
                    dictionary_wrong, "");
#endif
    dctx->validateChecksum = (dctx->fParams.checksumFlag && !dctx->forceIgnoreChecksum) ? 1 : 0;
    if (dctx->validateChecksum) XXH64_reset(&dctx->xxhState, 0);
    dctx->processedCSize += headerSize;
    return 0;
}

static ZSTD_frameSizeInfo ZSTD_errorFrameSizeInfo(size_t ret)
{
    ZSTD_frameSizeInfo frameSizeInfo;
    frameSizeInfo.compressedSize = ret;
    frameSizeInfo.decompressedBound = ZSTD_CONTENTSIZE_ERROR;
    return frameSizeInfo;
}

static ZSTD_frameSizeInfo ZSTD_findFrameSizeInfo(const void* src, size_t srcSize, ZSTD_format_e format)
{
    ZSTD_frameSizeInfo frameSizeInfo;
    ZSTD_memset(&frameSizeInfo, 0, sizeof(ZSTD_frameSizeInfo));

#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
    if (format == ZSTD_f_zstd1 && ZSTD_isLegacy(src, srcSize))
        return ZSTD_findFrameSizeInfoLegacy(src, srcSize);
#endif

    if (format == ZSTD_f_zstd1 && (srcSize >= ZSTD_SKIPPABLEHEADERSIZE)
        && (MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
        frameSizeInfo.compressedSize = readSkippableFrameSize(src, srcSize);
        assert(ZSTD_isError(frameSizeInfo.compressedSize) ||
               frameSizeInfo.compressedSize <= srcSize);
        return frameSizeInfo;
    } else {
        const BYTE* ip = (const BYTE*)src;
        const BYTE* const ipstart = ip;
        size_t remainingSize = srcSize;
        size_t nbBlocks = 0;
        ZSTD_FrameHeader zfh;

        /* Extract Frame Header */
        {   size_t const ret = ZSTD_getFrameHeader_advanced(&zfh, src, srcSize, format);
            if (ZSTD_isError(ret))
                return ZSTD_errorFrameSizeInfo(ret);
            if (ret > 0)
                return ZSTD_errorFrameSizeInfo(ERROR(srcSize_wrong));
        }

        ip += zfh.headerSize;
        remainingSize -= zfh.headerSize;

        /* Iterate over each block */
        while (1) {
            blockProperties_t blockProperties;
            size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSize, &blockProperties);
            if (ZSTD_isError(cBlockSize))
                return ZSTD_errorFrameSizeInfo(cBlockSize);

            if (ZSTD_blockHeaderSize + cBlockSize > remainingSize)
                return ZSTD_errorFrameSizeInfo(ERROR(srcSize_wrong));

            ip += ZSTD_blockHeaderSize + cBlockSize;
            remainingSize -= ZSTD_blockHeaderSize + cBlockSize;
            nbBlocks++;

            if (blockProperties.lastBlock) break;
        }

        /* Final frame content checksum */
        if (zfh.checksumFlag) {
            if (remainingSize < 4)
                return ZSTD_errorFrameSizeInfo(ERROR(srcSize_wrong));
            ip += 4;
        }

        frameSizeInfo.nbBlocks = nbBlocks;
        frameSizeInfo.compressedSize = (size_t)(ip - ipstart);
        frameSizeInfo.decompressedBound = (zfh.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN)
                                        ? zfh.frameContentSize
                                        : (unsigned long long)nbBlocks * zfh.blockSizeMax;
        return frameSizeInfo;
    }
}

static size_t ZSTD_findFrameCompressedSize_advanced(const void *src, size_t srcSize, ZSTD_format_e format) {
    ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, format);
    return frameSizeInfo.compressedSize;
}

/** ZSTD_findFrameCompressedSize() :
 * See docs in zstd.h
 * Note: compatible with legacy mode */
size_t ZSTD_findFrameCompressedSize(const void *src, size_t srcSize)
{
    return ZSTD_findFrameCompressedSize_advanced(src, srcSize, ZSTD_f_zstd1);
}

/** ZSTD_decompressBound() :
 *  compatible with legacy mode
 *  `src` must point to the start of a ZSTD frame or a skippable frame
 *  `srcSize` must be at least as large as the frame contained
 *  @return : the maximum decompressed size of the compressed source
 */
unsigned long long ZSTD_decompressBound(const void* src, size_t srcSize)
{
    unsigned long long bound = 0;
    /* Iterate over each frame */
    while (srcSize > 0) {
        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, ZSTD_f_zstd1);
        size_t const compressedSize = frameSizeInfo.compressedSize;
        unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
        if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
            return ZSTD_CONTENTSIZE_ERROR;
        assert(srcSize >= compressedSize);
        src = (const BYTE*)src + compressedSize;
        srcSize -= compressedSize;
        bound += decompressedBound;
    }
    return bound;
}

size_t ZSTD_decompressionMargin(void const* src, size_t srcSize)
{
    size_t margin = 0;
    unsigned maxBlockSize = 0;

    /* Iterate over each frame */
    while (srcSize > 0) {
        ZSTD_frameSizeInfo const frameSizeInfo = ZSTD_findFrameSizeInfo(src, srcSize, ZSTD_f_zstd1);
        size_t const compressedSize = frameSizeInfo.compressedSize;
        unsigned long long const decompressedBound = frameSizeInfo.decompressedBound;
        ZSTD_FrameHeader zfh;

        FORWARD_IF_ERROR(ZSTD_getFrameHeader(&zfh, src, srcSize), "");
        if (ZSTD_isError(compressedSize) || decompressedBound == ZSTD_CONTENTSIZE_ERROR)
            return ERROR(corruption_detected);

        if (zfh.frameType == ZSTD_frame) {
            /* Add the frame header to our margin */
            margin += zfh.headerSize;
            /* Add the checksum to our margin */
            margin += zfh.checksumFlag ? 4 : 0;
            /* Add 3 bytes per block */
            margin += 3 * frameSizeInfo.nbBlocks;

            /* Compute the max block size */
            maxBlockSize = MAX(maxBlockSize, zfh.blockSizeMax);
        } else {
            assert(zfh.frameType == ZSTD_skippableFrame);
            /* Add the entire skippable frame size to our margin. */
            margin += compressedSize;
        }

        assert(srcSize >= compressedSize);
        src = (const BYTE*)src + compressedSize;
        srcSize -= compressedSize;
    }

    /* Add the max block size back to the margin. */
    margin += maxBlockSize;

    return margin;
}

/*-*************************************************************
 *   Frame decoding
 ***************************************************************/

/** ZSTD_insertBlock() :
 *  insert `src` block into `dctx` history. Useful to track uncompressed blocks. */
size_t ZSTD_insertBlock(ZSTD_DCtx* dctx, const void* blockStart, size_t blockSize)
{
    DEBUGLOG(5, "ZSTD_insertBlock: %u bytes", (unsigned)blockSize);
    ZSTD_checkContinuity(dctx, blockStart, blockSize);
    dctx->previousDstEnd = (const char*)blockStart + blockSize;
    return blockSize;
}


static size_t ZSTD_copyRawBlock(void* dst, size_t dstCapacity,
                          const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_copyRawBlock");
    RETURN_ERROR_IF(srcSize > dstCapacity, dstSize_tooSmall, "");
    if (dst == NULL) {
        if (srcSize == 0) return 0;
        RETURN_ERROR(dstBuffer_null, "");
    }
    ZSTD_memmove(dst, src, srcSize);
    return srcSize;
}

static size_t ZSTD_setRleBlock(void* dst, size_t dstCapacity,
                               BYTE b,
                               size_t regenSize)
{
    RETURN_ERROR_IF(regenSize > dstCapacity, dstSize_tooSmall, "");
    if (dst == NULL) {
        if (regenSize == 0) return 0;
        RETURN_ERROR(dstBuffer_null, "");
    }
    ZSTD_memset(dst, b, regenSize);
    return regenSize;
}

static void ZSTD_DCtx_trace_end(ZSTD_DCtx const* dctx, U64 uncompressedSize, U64 compressedSize, int streaming)
{
#if ZSTD_TRACE
    if (dctx->traceCtx && ZSTD_trace_decompress_end != NULL) {
        ZSTD_Trace trace;
        ZSTD_memset(&trace, 0, sizeof(trace));
        trace.version = ZSTD_VERSION_NUMBER;
        trace.streaming = streaming;
        if (dctx->ddict) {
            trace.dictionaryID = ZSTD_getDictID_fromDDict(dctx->ddict);
            trace.dictionarySize = ZSTD_DDict_dictSize(dctx->ddict);
            trace.dictionaryIsCold = dctx->ddictIsCold;
        }
        trace.uncompressedSize = (size_t)uncompressedSize;
        trace.compressedSize = (size_t)compressedSize;
        trace.dctx = dctx;
        ZSTD_trace_decompress_end(dctx->traceCtx, &trace);
    }
#else
    (void)dctx;
    (void)uncompressedSize;
    (void)compressedSize;
    (void)streaming;
#endif
}


/*! ZSTD_decompressFrame() :
 * @dctx must be properly initialized
 *  will update *srcPtr and *srcSizePtr,
 *  to make *srcPtr progress by one frame. */
static size_t ZSTD_decompressFrame(ZSTD_DCtx* dctx,
                                   void* dst, size_t dstCapacity,
                             const void** srcPtr, size_t *srcSizePtr)
{
    const BYTE* const istart = (const BYTE*)(*srcPtr);
    const BYTE* ip = istart;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = dstCapacity != 0 ? ostart + dstCapacity : ostart;
    BYTE* op = ostart;
    size_t remainingSrcSize = *srcSizePtr;

    DEBUGLOG(4, "ZSTD_decompressFrame (srcSize:%i)", (int)*srcSizePtr);

    /* check */
    RETURN_ERROR_IF(
        remainingSrcSize < ZSTD_FRAMEHEADERSIZE_MIN(dctx->format)+ZSTD_blockHeaderSize,
        srcSize_wrong, "");

    /* Frame Header */
    {   size_t const frameHeaderSize = ZSTD_frameHeaderSize_internal(
                ip, ZSTD_FRAMEHEADERSIZE_PREFIX(dctx->format), dctx->format);
        if (ZSTD_isError(frameHeaderSize)) return frameHeaderSize;
        RETURN_ERROR_IF(remainingSrcSize < frameHeaderSize+ZSTD_blockHeaderSize,
                        srcSize_wrong, "");
        FORWARD_IF_ERROR( ZSTD_decodeFrameHeader(dctx, ip, frameHeaderSize) , "");
        ip += frameHeaderSize; remainingSrcSize -= frameHeaderSize;
    }

    /* Shrink the blockSizeMax if enabled */
    if (dctx->maxBlockSizeParam != 0)
        dctx->fParams.blockSizeMax = MIN(dctx->fParams.blockSizeMax, (unsigned)dctx->maxBlockSizeParam);

    /* Loop on each block */
    while (1) {
        BYTE* oBlockEnd = oend;
        size_t decodedSize;
        blockProperties_t blockProperties;
        memset(&blockProperties, 0, sizeof(blockProperties)); // shut up gcc warning
        size_t const cBlockSize = ZSTD_getcBlockSize(ip, remainingSrcSize, &blockProperties);
        if (ZSTD_isError(cBlockSize)) return cBlockSize;

        ip += ZSTD_blockHeaderSize;
        remainingSrcSize -= ZSTD_blockHeaderSize;
        RETURN_ERROR_IF(cBlockSize > remainingSrcSize, srcSize_wrong, "");

        if (ip >= op && ip < oBlockEnd) {
            /* We are decompressing in-place. Limit the output pointer so that we
             * don't overwrite the block that we are currently reading. This will
             * fail decompression if the input & output pointers aren't spaced
             * far enough apart.
             *
             * This is important to set, even when the pointers are far enough
             * apart, because ZSTD_decompressBlock_internal() can decide to store
             * literals in the output buffer, after the block it is decompressing.
             * Since we don't want anything to overwrite our input, we have to tell
             * ZSTD_decompressBlock_internal to never write past ip.
             *
             * See ZSTD_allocateLiteralsBuffer() for reference.
             */
            oBlockEnd = op + (ip - op);
        }

        switch(blockProperties.blockType)
        {
        case bt_compressed:
            assert(dctx->isFrameDecompression == 1);
            decodedSize = ZSTD_decompressBlock_internal(dctx, op, (size_t)(oBlockEnd-op), ip, cBlockSize, not_streaming);
            break;
        case bt_raw :
            /* Use oend instead of oBlockEnd because this function is safe to overlap. It uses memmove. */
            decodedSize = ZSTD_copyRawBlock(op, (size_t)(oend-op), ip, cBlockSize);
            break;
        case bt_rle :
            decodedSize = ZSTD_setRleBlock(op, (size_t)(oBlockEnd-op), *ip, blockProperties.origSize);
            break;
        case bt_reserved :
        default:
            RETURN_ERROR(corruption_detected, "invalid block type");
        }
        FORWARD_IF_ERROR(decodedSize, "Block decompression failure");
        DEBUGLOG(5, "Decompressed block of dSize = %u", (unsigned)decodedSize);
        if (dctx->validateChecksum) {
            XXH64_update(&dctx->xxhState, op, decodedSize);
        }
        if (decodedSize) /* support dst = NULL,0 */ {
            op += decodedSize;
        }
        assert(ip != NULL);
        ip += cBlockSize;
        remainingSrcSize -= cBlockSize;
        if (blockProperties.lastBlock) break;
    }

    if (dctx->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN) {
        RETURN_ERROR_IF((U64)(op-ostart) != dctx->fParams.frameContentSize,
                        corruption_detected, "");
    }
    if (dctx->fParams.checksumFlag) { /* Frame content checksum verification */
        RETURN_ERROR_IF(remainingSrcSize<4, checksum_wrong, "");
        if (!dctx->forceIgnoreChecksum) {
            U32 const checkCalc = (U32)XXH64_digest(&dctx->xxhState);
            U32 checkRead;
            checkRead = MEM_readLE32(ip);
            RETURN_ERROR_IF(checkRead != checkCalc, checksum_wrong, "");
        }
        ip += 4;
        remainingSrcSize -= 4;
    }
    ZSTD_DCtx_trace_end(dctx, (U64)(op-ostart), (U64)(ip-istart), /* streaming */ 0);
    /* Allow caller to get size read */
    DEBUGLOG(4, "ZSTD_decompressFrame: decompressed frame of size %i, consuming %i bytes of input", (int)(op-ostart), (int)(ip - (const BYTE*)*srcPtr));
    *srcPtr = ip;
    *srcSizePtr = remainingSrcSize;
    return (size_t)(op-ostart);
}

static
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_decompressMultiFrame(ZSTD_DCtx* dctx,
                                        void* dst, size_t dstCapacity,
                                  const void* src, size_t srcSize,
                                  const void* dict, size_t dictSize,
                                  const ZSTD_DDict* ddict)
{
    void* const dststart = dst;
    int moreThan1Frame = 0;

    DEBUGLOG(5, "ZSTD_decompressMultiFrame");
    assert(dict==NULL || ddict==NULL);  /* either dict or ddict set, not both */

    if (ddict) {
        dict = ZSTD_DDict_dictContent(ddict);
        dictSize = ZSTD_DDict_dictSize(ddict);
    }

    while (srcSize >= ZSTD_startingInputLength(dctx->format)) {

#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT >= 1)
        if (dctx->format == ZSTD_f_zstd1 && ZSTD_isLegacy(src, srcSize)) {
            size_t decodedSize;
            size_t const frameSize = ZSTD_findFrameCompressedSizeLegacy(src, srcSize);
            if (ZSTD_isError(frameSize)) return frameSize;
            RETURN_ERROR_IF(dctx->staticSize, memory_allocation,
                "legacy support is not compatible with static dctx");

            decodedSize = ZSTD_decompressLegacy(dst, dstCapacity, src, frameSize, dict, dictSize);
            if (ZSTD_isError(decodedSize)) return decodedSize;

            {
                unsigned long long const expectedSize = ZSTD_getFrameContentSize(src, srcSize);
                RETURN_ERROR_IF(expectedSize == ZSTD_CONTENTSIZE_ERROR, corruption_detected, "Corrupted frame header!");
                if (expectedSize != ZSTD_CONTENTSIZE_UNKNOWN) {
                    RETURN_ERROR_IF(expectedSize != decodedSize, corruption_detected,
                        "Frame header size does not match decoded size!");
                }
            }

            assert(decodedSize <= dstCapacity);
            dst = (BYTE*)dst + decodedSize;
            dstCapacity -= decodedSize;

            src = (const BYTE*)src + frameSize;
            srcSize -= frameSize;

            continue;
        }
#endif

        if (dctx->format == ZSTD_f_zstd1 && srcSize >= 4) {
            U32 const magicNumber = MEM_readLE32(src);
            DEBUGLOG(5, "reading magic number %08X", (unsigned)magicNumber);
            if ((magicNumber & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {
                /* skippable frame detected : skip it */
                size_t const skippableSize = readSkippableFrameSize(src, srcSize);
                FORWARD_IF_ERROR(skippableSize, "invalid skippable frame");
                assert(skippableSize <= srcSize);

                src = (const BYTE *)src + skippableSize;
                srcSize -= skippableSize;
                continue; /* check next frame */
        }   }

        if (ddict) {
            /* we were called from ZSTD_decompress_usingDDict */
            FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDDict(dctx, ddict), "");
        } else {
            /* this will initialize correctly with no dict if dict == NULL, so
             * use this in all cases but ddict */
            FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDict(dctx, dict, dictSize), "");
        }
        ZSTD_checkContinuity(dctx, dst, dstCapacity);

        {   const size_t res = ZSTD_decompressFrame(dctx, dst, dstCapacity,
                                                    &src, &srcSize);
            RETURN_ERROR_IF(
                (ZSTD_getErrorCode(res) == ZSTD_error_prefix_unknown)
             && (moreThan1Frame==1),
                srcSize_wrong,
                "At least one frame successfully completed, "
                "but following bytes are garbage: "
                "it's more likely to be a srcSize error, "
                "specifying more input bytes than size of frame(s). "
                "Note: one could be unlucky, it might be a corruption error instead, "
                "happening right at the place where we expect zstd magic bytes. "
                "But this is _much_ less likely than a srcSize field error.");
            if (ZSTD_isError(res)) return res;
            assert(res <= dstCapacity);
            if (res != 0)
                dst = (BYTE*)dst + res;
            dstCapacity -= res;
        }
        moreThan1Frame = 1;
    }  /* while (srcSize >= ZSTD_frameHeaderSize_prefix) */

    RETURN_ERROR_IF(srcSize, srcSize_wrong, "input not entirely consumed");

    return (size_t)((BYTE*)dst - (BYTE*)dststart);
}

size_t ZSTD_decompress_usingDict(ZSTD_DCtx* dctx,
                                 void* dst, size_t dstCapacity,
                           const void* src, size_t srcSize,
                           const void* dict, size_t dictSize)
{
    return ZSTD_decompressMultiFrame(dctx, dst, dstCapacity, src, srcSize, dict, dictSize, NULL);
}


static ZSTD_DDict const* ZSTD_getDDict(ZSTD_DCtx* dctx)
{
    switch (dctx->dictUses) {
    default:
        assert(0 /* Impossible */);
        ZSTD_FALLTHROUGH;
    case ZSTD_dont_use:
        ZSTD_clearDict(dctx);
        return NULL;
    case ZSTD_use_indefinitely:
        return dctx->ddict;
    case ZSTD_use_once:
        dctx->dictUses = ZSTD_dont_use;
        return dctx->ddict;
    }
}

size_t ZSTD_decompressDCtx(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    return ZSTD_decompress_usingDDict(dctx, dst, dstCapacity, src, srcSize, ZSTD_getDDict(dctx));
}


size_t ZSTD_decompress(void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
#if defined(ZSTD_HEAPMODE) && (ZSTD_HEAPMODE>=1)
    size_t regenSize;
    ZSTD_DCtx* const dctx =  ZSTD_createDCtx_internal(ZSTD_defaultCMem);
    RETURN_ERROR_IF(dctx==NULL, memory_allocation, "NULL pointer!");
    regenSize = ZSTD_decompressDCtx(dctx, dst, dstCapacity, src, srcSize);
    ZSTD_freeDCtx(dctx);
    return regenSize;
#else   /* stack mode */
    ZSTD_DCtx dctx;
    ZSTD_initDCtx_internal(&dctx);
    return ZSTD_decompressDCtx(&dctx, dst, dstCapacity, src, srcSize);
#endif
}


/*-**************************************
*   Advanced Streaming Decompression API
*   Bufferless and synchronous
****************************************/
size_t ZSTD_nextSrcSizeToDecompress(ZSTD_DCtx* dctx) { return dctx->expected; }

/**
 * Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed, we
 * allow taking a partial block as the input. Currently only raw uncompressed blocks can
 * be streamed.
 *
 * For blocks that can be streamed, this allows us to reduce the latency until we produce
 * output, and avoid copying the input.
 *
 * @param inputSize - The total amount of input that the caller currently has.
 */
static size_t ZSTD_nextSrcSizeToDecompressWithInputSize(ZSTD_DCtx* dctx, size_t inputSize) {
    if (!(dctx->stage == ZSTDds_decompressBlock || dctx->stage == ZSTDds_decompressLastBlock))
        return dctx->expected;
    if (dctx->bType != bt_raw)
        return dctx->expected;
    return BOUNDED(1, inputSize, dctx->expected);
}

ZSTD_nextInputType_e ZSTD_nextInputType(ZSTD_DCtx* dctx) {
    switch(dctx->stage)
    {
    default:   /* should not happen */
        assert(0);
        ZSTD_FALLTHROUGH;
    case ZSTDds_getFrameHeaderSize:
        ZSTD_FALLTHROUGH;
    case ZSTDds_decodeFrameHeader:
        return ZSTDnit_frameHeader;
    case ZSTDds_decodeBlockHeader:
        return ZSTDnit_blockHeader;
    case ZSTDds_decompressBlock:
        return ZSTDnit_block;
    case ZSTDds_decompressLastBlock:
        return ZSTDnit_lastBlock;
    case ZSTDds_checkChecksum:
        return ZSTDnit_checksum;
    case ZSTDds_decodeSkippableHeader:
        ZSTD_FALLTHROUGH;
    case ZSTDds_skipFrame:
        return ZSTDnit_skippableFrame;
    }
}

static int ZSTD_isSkipFrame(ZSTD_DCtx* dctx) { return dctx->stage == ZSTDds_skipFrame; }

/** ZSTD_decompressContinue() :
 *  srcSize : must be the exact nb of bytes expected (see ZSTD_nextSrcSizeToDecompress())
 *  @return : nb of bytes generated into `dst` (necessarily <= `dstCapacity)
 *            or an error code, which can be tested using ZSTD_isError() */
size_t ZSTD_decompressContinue(ZSTD_DCtx* dctx, void* dst, size_t dstCapacity, const void* src, size_t srcSize)
{
    DEBUGLOG(5, "ZSTD_decompressContinue (srcSize:%u)", (unsigned)srcSize);
    /* Sanity check */
    RETURN_ERROR_IF(srcSize != ZSTD_nextSrcSizeToDecompressWithInputSize(dctx, srcSize), srcSize_wrong, "not allowed");
    ZSTD_checkContinuity(dctx, dst, dstCapacity);

    dctx->processedCSize += srcSize;

    switch (dctx->stage)
    {
    case ZSTDds_getFrameHeaderSize :
        assert(src != NULL);
        if (dctx->format == ZSTD_f_zstd1) {  /* allows header */
            assert(srcSize >= ZSTD_FRAMEIDSIZE);  /* to read skippable magic number */
            if ((MEM_readLE32(src) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {        /* skippable frame */
                ZSTD_memcpy(dctx->headerBuffer, src, srcSize);
                dctx->expected = ZSTD_SKIPPABLEHEADERSIZE - srcSize;  /* remaining to load to get full skippable frame header */
                dctx->stage = ZSTDds_decodeSkippableHeader;
                return 0;
        }   }
        dctx->headerSize = ZSTD_frameHeaderSize_internal(src, srcSize, dctx->format);
        if (ZSTD_isError(dctx->headerSize)) return dctx->headerSize;
        ZSTD_memcpy(dctx->headerBuffer, src, srcSize);
        dctx->expected = dctx->headerSize - srcSize;
        dctx->stage = ZSTDds_decodeFrameHeader;
        return 0;

    case ZSTDds_decodeFrameHeader:
        assert(src != NULL);
        ZSTD_memcpy(dctx->headerBuffer + (dctx->headerSize - srcSize), src, srcSize);
        FORWARD_IF_ERROR(ZSTD_decodeFrameHeader(dctx, dctx->headerBuffer, dctx->headerSize), "");
        dctx->expected = ZSTD_blockHeaderSize;
        dctx->stage = ZSTDds_decodeBlockHeader;
        return 0;

    case ZSTDds_decodeBlockHeader:
        {   blockProperties_t bp;
            size_t const cBlockSize = ZSTD_getcBlockSize(src, ZSTD_blockHeaderSize, &bp);
            if (ZSTD_isError(cBlockSize)) return cBlockSize;
            RETURN_ERROR_IF(cBlockSize > dctx->fParams.blockSizeMax, corruption_detected, "Block Size Exceeds Maximum");
            dctx->expected = cBlockSize;
            dctx->bType = bp.blockType;
            dctx->rleSize = bp.origSize;
            if (cBlockSize) {
                dctx->stage = bp.lastBlock ? ZSTDds_decompressLastBlock : ZSTDds_decompressBlock;
                return 0;
            }
            /* empty block */
            if (bp.lastBlock) {
                if (dctx->fParams.checksumFlag) {
                    dctx->expected = 4;
                    dctx->stage = ZSTDds_checkChecksum;
                } else {
                    dctx->expected = 0; /* end of frame */
                    dctx->stage = ZSTDds_getFrameHeaderSize;
                }
            } else {
                dctx->expected = ZSTD_blockHeaderSize;  /* jump to next header */
                dctx->stage = ZSTDds_decodeBlockHeader;
            }
            return 0;
        }

    case ZSTDds_decompressLastBlock:
    case ZSTDds_decompressBlock:
        DEBUGLOG(5, "ZSTD_decompressContinue: case ZSTDds_decompressBlock");
        {   size_t rSize;
            switch(dctx->bType)
            {
            case bt_compressed:
                DEBUGLOG(5, "ZSTD_decompressContinue: case bt_compressed");
                assert(dctx->isFrameDecompression == 1);
                rSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, is_streaming);
                dctx->expected = 0;  /* Streaming not supported */
                break;
            case bt_raw :
                assert(srcSize <= dctx->expected);
                rSize = ZSTD_copyRawBlock(dst, dstCapacity, src, srcSize);
                FORWARD_IF_ERROR(rSize, "ZSTD_copyRawBlock failed");
                assert(rSize == srcSize);
                dctx->expected -= rSize;
                break;
            case bt_rle :
                rSize = ZSTD_setRleBlock(dst, dstCapacity, *(const BYTE*)src, dctx->rleSize);
                dctx->expected = 0;  /* Streaming not supported */
                break;
            case bt_reserved :   /* should never happen */
            default:
                RETURN_ERROR(corruption_detected, "invalid block type");
            }
            FORWARD_IF_ERROR(rSize, "");
            RETURN_ERROR_IF(rSize > dctx->fParams.blockSizeMax, corruption_detected, "Decompressed Block Size Exceeds Maximum");
            DEBUGLOG(5, "ZSTD_decompressContinue: decoded size from block : %u", (unsigned)rSize);
            dctx->decodedSize += rSize;
            if (dctx->validateChecksum) XXH64_update(&dctx->xxhState, dst, rSize);
            dctx->previousDstEnd = (char*)dst + rSize;

            /* Stay on the same stage until we are finished streaming the block. */
            if (dctx->expected > 0) {
                return rSize;
            }

            if (dctx->stage == ZSTDds_decompressLastBlock) {   /* end of frame */
                DEBUGLOG(4, "ZSTD_decompressContinue: decoded size from frame : %u", (unsigned)dctx->decodedSize);
                RETURN_ERROR_IF(
                    dctx->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                 && dctx->decodedSize != dctx->fParams.frameContentSize,
                    corruption_detected, "");
                if (dctx->fParams.checksumFlag) {  /* another round for frame checksum */
                    dctx->expected = 4;
                    dctx->stage = ZSTDds_checkChecksum;
                } else {
                    ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
                    dctx->expected = 0;   /* ends here */
                    dctx->stage = ZSTDds_getFrameHeaderSize;
                }
            } else {
                dctx->stage = ZSTDds_decodeBlockHeader;
                dctx->expected = ZSTD_blockHeaderSize;
            }
            return rSize;
        }

    case ZSTDds_checkChecksum:
        assert(srcSize == 4);  /* guaranteed by dctx->expected */
        {
            if (dctx->validateChecksum) {
                U32 const h32 = (U32)XXH64_digest(&dctx->xxhState);
                U32 const check32 = MEM_readLE32(src);
                DEBUGLOG(4, "ZSTD_decompressContinue: checksum : calculated %08X :: %08X read", (unsigned)h32, (unsigned)check32);
                RETURN_ERROR_IF(check32 != h32, checksum_wrong, "");
            }
            ZSTD_DCtx_trace_end(dctx, dctx->decodedSize, dctx->processedCSize, /* streaming */ 1);
            dctx->expected = 0;
            dctx->stage = ZSTDds_getFrameHeaderSize;
            return 0;
        }

    case ZSTDds_decodeSkippableHeader:
        assert(src != NULL);
        assert(srcSize <= ZSTD_SKIPPABLEHEADERSIZE);
        assert(dctx->format != ZSTD_f_zstd1_magicless);
        ZSTD_memcpy(dctx->headerBuffer + (ZSTD_SKIPPABLEHEADERSIZE - srcSize), src, srcSize);   /* complete skippable header */
        dctx->expected = MEM_readLE32(dctx->headerBuffer + ZSTD_FRAMEIDSIZE);   /* note : dctx->expected can grow seriously large, beyond local buffer size */
        dctx->stage = ZSTDds_skipFrame;
        return 0;

    case ZSTDds_skipFrame:
        dctx->expected = 0;
        dctx->stage = ZSTDds_getFrameHeaderSize;
        return 0;

    default:
        assert(0);   /* impossible */
        RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
    }
}


static size_t ZSTD_refDictContent(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
{
    dctx->dictEnd = dctx->previousDstEnd;
    dctx->virtualStart = (const char*)dict - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
    dctx->prefixStart = dict;
    dctx->previousDstEnd = (const char*)dict + dictSize;
#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
    dctx->dictContentBeginForFuzzing = dctx->prefixStart;
    dctx->dictContentEndForFuzzing = dctx->previousDstEnd;
#endif
    return 0;
}

/*! ZSTD_loadDEntropy() :
 *  dict : must point at beginning of a valid zstd dictionary.
 * @return : size of entropy tables read */
size_t
ZSTD_loadDEntropy(ZSTD_entropyDTables_t* entropy,
                  const void* const dict, size_t const dictSize)
{
    const BYTE* dictPtr = (const BYTE*)dict;
    const BYTE* const dictEnd = dictPtr + dictSize;

    RETURN_ERROR_IF(dictSize <= 8, dictionary_corrupted, "dict is too small");
    assert(MEM_readLE32(dict) == ZSTD_MAGIC_DICTIONARY);   /* dict must be valid */
    dictPtr += 8;   /* skip header = magic + dictID */

    ZSTD_STATIC_ASSERT(offsetof(ZSTD_entropyDTables_t, OFTable) == offsetof(ZSTD_entropyDTables_t, LLTable) + sizeof(entropy->LLTable));
    ZSTD_STATIC_ASSERT(offsetof(ZSTD_entropyDTables_t, MLTable) == offsetof(ZSTD_entropyDTables_t, OFTable) + sizeof(entropy->OFTable));
    ZSTD_STATIC_ASSERT(sizeof(entropy->LLTable) + sizeof(entropy->OFTable) + sizeof(entropy->MLTable) >= HUF_DECOMPRESS_WORKSPACE_SIZE);
    {   void* const workspace = &entropy->LLTable;   /* use fse tables as temporary workspace; implies fse tables are grouped together */
        size_t const workspaceSize = sizeof(entropy->LLTable) + sizeof(entropy->OFTable) + sizeof(entropy->MLTable);
#ifdef HUF_FORCE_DECOMPRESS_X1
        /* in minimal huffman, we always use X1 variants */
        size_t const hSize = HUF_readDTableX1_wksp(entropy->hufTable,
                                                dictPtr, dictEnd - dictPtr,
                                                workspace, workspaceSize, /* flags */ 0);
#else
        size_t const hSize = HUF_readDTableX2_wksp(entropy->hufTable,
                                                dictPtr, (size_t)(dictEnd - dictPtr),
                                                workspace, workspaceSize, /* flags */ 0);
#endif
        RETURN_ERROR_IF(HUF_isError(hSize), dictionary_corrupted, "");
        dictPtr += hSize;
    }

    {   short offcodeNCount[MaxOff+1];
        unsigned offcodeMaxValue = MaxOff, offcodeLog;
        size_t const offcodeHeaderSize = FSE_readNCount(offcodeNCount, &offcodeMaxValue, &offcodeLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(offcodeHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(offcodeMaxValue > MaxOff, dictionary_corrupted, "");
        RETURN_ERROR_IF(offcodeLog > OffFSELog, dictionary_corrupted, "");
        ZSTD_buildFSETable( entropy->OFTable,
                            offcodeNCount, offcodeMaxValue,
                            OF_base, OF_bits,
                            offcodeLog,
                            entropy->workspace, sizeof(entropy->workspace),
                            /* bmi2 */0);
        dictPtr += offcodeHeaderSize;
    }

    {   short matchlengthNCount[MaxML+1];
        unsigned matchlengthMaxValue = MaxML, matchlengthLog;
        size_t const matchlengthHeaderSize = FSE_readNCount(matchlengthNCount, &matchlengthMaxValue, &matchlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(matchlengthHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(matchlengthMaxValue > MaxML, dictionary_corrupted, "");
        RETURN_ERROR_IF(matchlengthLog > MLFSELog, dictionary_corrupted, "");
        ZSTD_buildFSETable( entropy->MLTable,
                            matchlengthNCount, matchlengthMaxValue,
                            ML_base, ML_bits,
                            matchlengthLog,
                            entropy->workspace, sizeof(entropy->workspace),
                            /* bmi2 */ 0);
        dictPtr += matchlengthHeaderSize;
    }

    {   short litlengthNCount[MaxLL+1];
        unsigned litlengthMaxValue = MaxLL, litlengthLog;
        size_t const litlengthHeaderSize = FSE_readNCount(litlengthNCount, &litlengthMaxValue, &litlengthLog, dictPtr, (size_t)(dictEnd-dictPtr));
        RETURN_ERROR_IF(FSE_isError(litlengthHeaderSize), dictionary_corrupted, "");
        RETURN_ERROR_IF(litlengthMaxValue > MaxLL, dictionary_corrupted, "");
        RETURN_ERROR_IF(litlengthLog > LLFSELog, dictionary_corrupted, "");
        ZSTD_buildFSETable( entropy->LLTable,
                            litlengthNCount, litlengthMaxValue,
                            LL_base, LL_bits,
                            litlengthLog,
                            entropy->workspace, sizeof(entropy->workspace),
                            /* bmi2 */ 0);
        dictPtr += litlengthHeaderSize;
    }

    RETURN_ERROR_IF(dictPtr+12 > dictEnd, dictionary_corrupted, "");
    {   int i;
        size_t const dictContentSize = (size_t)(dictEnd - (dictPtr+12));
        for (i=0; i<3; i++) {
            U32 const rep = MEM_readLE32(dictPtr); dictPtr += 4;
            RETURN_ERROR_IF(rep==0 || rep > dictContentSize,
                            dictionary_corrupted, "");
            entropy->rep[i] = rep;
    }   }

    return (size_t)(dictPtr - (const BYTE*)dict);
}

static size_t ZSTD_decompress_insertDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
{
    if (dictSize < 8) return ZSTD_refDictContent(dctx, dict, dictSize);
    {   U32 const magic = MEM_readLE32(dict);
        if (magic != ZSTD_MAGIC_DICTIONARY) {
            return ZSTD_refDictContent(dctx, dict, dictSize);   /* pure content mode */
    }   }
    dctx->dictID = MEM_readLE32((const char*)dict + ZSTD_FRAMEIDSIZE);

    /* load entropy tables */
    {   size_t const eSize = ZSTD_loadDEntropy(&dctx->entropy, dict, dictSize);
        RETURN_ERROR_IF(ZSTD_isError(eSize), dictionary_corrupted, "");
        dict = (const char*)dict + eSize;
        dictSize -= eSize;
    }
    dctx->litEntropy = dctx->fseEntropy = 1;

    /* reference dictionary content */
    return ZSTD_refDictContent(dctx, dict, dictSize);
}

size_t ZSTD_decompressBegin(ZSTD_DCtx* dctx)
{
    assert(dctx != NULL);
#if ZSTD_TRACE
    dctx->traceCtx = (ZSTD_trace_decompress_begin != NULL) ? ZSTD_trace_decompress_begin(dctx) : 0;
#endif
    dctx->expected = ZSTD_startingInputLength(dctx->format);  /* dctx->format must be properly set */
    dctx->stage = ZSTDds_getFrameHeaderSize;
    dctx->processedCSize = 0;
    dctx->decodedSize = 0;
    dctx->previousDstEnd = NULL;
    dctx->prefixStart = NULL;
    dctx->virtualStart = NULL;
    dctx->dictEnd = NULL;
    dctx->entropy.hufTable[0] = (HUF_DTable)((ZSTD_HUFFDTABLE_CAPACITY_LOG)*0x1000001);  /* cover both little and big endian */
    dctx->litEntropy = dctx->fseEntropy = 0;
    dctx->dictID = 0;
    dctx->bType = bt_reserved;
    dctx->isFrameDecompression = 1;
    ZSTD_STATIC_ASSERT(sizeof(dctx->entropy.rep) == sizeof(repStartValue));
    ZSTD_memcpy(dctx->entropy.rep, repStartValue, sizeof(repStartValue));  /* initial repcodes */
    dctx->LLTptr = dctx->entropy.LLTable;
    dctx->MLTptr = dctx->entropy.MLTable;
    dctx->OFTptr = dctx->entropy.OFTable;
    dctx->HUFptr = dctx->entropy.hufTable;
    return 0;
}

size_t ZSTD_decompressBegin_usingDict(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
{
    FORWARD_IF_ERROR( ZSTD_decompressBegin(dctx) , "");
    if (dict && dictSize)
        RETURN_ERROR_IF(
            ZSTD_isError(ZSTD_decompress_insertDictionary(dctx, dict, dictSize)),
            dictionary_corrupted, "");
    return 0;
}


/* ======   ZSTD_DDict   ====== */

size_t ZSTD_decompressBegin_usingDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
{
    DEBUGLOG(4, "ZSTD_decompressBegin_usingDDict");
    assert(dctx != NULL);
    if (ddict) {
        const char* const dictStart = (const char*)ZSTD_DDict_dictContent(ddict);
        size_t const dictSize = ZSTD_DDict_dictSize(ddict);
        const void* const dictEnd = dictStart + dictSize;
        dctx->ddictIsCold = (dctx->dictEnd != dictEnd);
        DEBUGLOG(4, "DDict is %s",
                    dctx->ddictIsCold ? "~cold~" : "hot!");
    }
    FORWARD_IF_ERROR( ZSTD_decompressBegin(dctx) , "");
    if (ddict) {   /* NULL ddict is equivalent to no dictionary */
        ZSTD_copyDDictParameters(dctx, ddict);
    }
    return 0;
}

/*! ZSTD_getDictID_fromDict() :
 *  Provides the dictID stored within dictionary.
 *  if @return == 0, the dictionary is not conformant with Zstandard specification.
 *  It can still be loaded, but as a content-only dictionary. */
unsigned ZSTD_getDictID_fromDict(const void* dict, size_t dictSize)
{
    if (dictSize < 8) return 0;
    if (MEM_readLE32(dict) != ZSTD_MAGIC_DICTIONARY) return 0;
    return MEM_readLE32((const char*)dict + ZSTD_FRAMEIDSIZE);
}

/*! ZSTD_getDictID_fromFrame() :
 *  Provides the dictID required to decompress frame stored within `src`.
 *  If @return == 0, the dictID could not be decoded.
 *  This could for one of the following reasons :
 *  - The frame does not require a dictionary (most common case).
 *  - The frame was built with dictID intentionally removed.
 *    Needed dictionary is a hidden piece of information.
 *    Note : this use case also happens when using a non-conformant dictionary.
 *  - `srcSize` is too small, and as a result, frame header could not be decoded.
 *    Note : possible if `srcSize < ZSTD_FRAMEHEADERSIZE_MAX`.
 *  - This is not a Zstandard frame.
 *  When identifying the exact failure cause, it's possible to use
 *  ZSTD_getFrameHeader(), which will provide a more precise error code. */
unsigned ZSTD_getDictID_fromFrame(const void* src, size_t srcSize)
{
    ZSTD_FrameHeader zfp = { 0, 0, 0, ZSTD_frame, 0, 0, 0, 0, 0 };
    size_t const hError = ZSTD_getFrameHeader(&zfp, src, srcSize);
    if (ZSTD_isError(hError)) return 0;
    return zfp.dictID;
}


/*! ZSTD_decompress_usingDDict() :
*   Decompression using a pre-digested Dictionary
*   Use dictionary without significant overhead. */
size_t ZSTD_decompress_usingDDict(ZSTD_DCtx* dctx,
                                  void* dst, size_t dstCapacity,
                            const void* src, size_t srcSize,
                            const ZSTD_DDict* ddict)
{
    /* pass content and size in case legacy frames are encountered */
    return ZSTD_decompressMultiFrame(dctx, dst, dstCapacity, src, srcSize,
                                     NULL, 0,
                                     ddict);
}


/*=====================================
*   Streaming decompression
*====================================*/

ZSTD_DStream* ZSTD_createDStream(void)
{
    DEBUGLOG(3, "ZSTD_createDStream");
    return ZSTD_createDCtx_internal(ZSTD_defaultCMem);
}

ZSTD_DStream* ZSTD_initStaticDStream(void *workspace, size_t workspaceSize)
{
    return ZSTD_initStaticDCtx(workspace, workspaceSize);
}

ZSTD_DStream* ZSTD_createDStream_advanced(ZSTD_customMem customMem)
{
    return ZSTD_createDCtx_internal(customMem);
}

size_t ZSTD_freeDStream(ZSTD_DStream* zds)
{
    return ZSTD_freeDCtx(zds);
}


/* ***  Initialization  *** */

size_t ZSTD_DStreamInSize(void)  { return ZSTD_BLOCKSIZE_MAX + ZSTD_blockHeaderSize; }
size_t ZSTD_DStreamOutSize(void) { return ZSTD_BLOCKSIZE_MAX; }

size_t ZSTD_DCtx_loadDictionary_advanced(ZSTD_DCtx* dctx,
                                   const void* dict, size_t dictSize,
                                         ZSTD_dictLoadMethod_e dictLoadMethod,
                                         ZSTD_dictContentType_e dictContentType)
{
    RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
    ZSTD_clearDict(dctx);
    if (dict && dictSize != 0) {
        dctx->ddictLocal = ZSTD_createDDict_advanced(dict, dictSize, dictLoadMethod, dictContentType, dctx->customMem);
        RETURN_ERROR_IF(dctx->ddictLocal == NULL, memory_allocation, "NULL pointer!");
        dctx->ddict = dctx->ddictLocal;
        dctx->dictUses = ZSTD_use_indefinitely;
    }
    return 0;
}

size_t ZSTD_DCtx_loadDictionary_byReference(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
{
    return ZSTD_DCtx_loadDictionary_advanced(dctx, dict, dictSize, ZSTD_dlm_byRef, ZSTD_dct_auto);
}

size_t ZSTD_DCtx_loadDictionary(ZSTD_DCtx* dctx, const void* dict, size_t dictSize)
{
    return ZSTD_DCtx_loadDictionary_advanced(dctx, dict, dictSize, ZSTD_dlm_byCopy, ZSTD_dct_auto);
}

size_t ZSTD_DCtx_refPrefix_advanced(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize, ZSTD_dictContentType_e dictContentType)
{
    FORWARD_IF_ERROR(ZSTD_DCtx_loadDictionary_advanced(dctx, prefix, prefixSize, ZSTD_dlm_byRef, dictContentType), "");
    dctx->dictUses = ZSTD_use_once;
    return 0;
}

size_t ZSTD_DCtx_refPrefix(ZSTD_DCtx* dctx, const void* prefix, size_t prefixSize)
{
    return ZSTD_DCtx_refPrefix_advanced(dctx, prefix, prefixSize, ZSTD_dct_rawContent);
}


/* ZSTD_initDStream_usingDict() :
 * return : expected size, aka ZSTD_startingInputLength().
 * this function cannot fail */
size_t ZSTD_initDStream_usingDict(ZSTD_DStream* zds, const void* dict, size_t dictSize)
{
    DEBUGLOG(4, "ZSTD_initDStream_usingDict");
    FORWARD_IF_ERROR( ZSTD_DCtx_reset(zds, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_DCtx_loadDictionary(zds, dict, dictSize) , "");
    return ZSTD_startingInputLength(zds->format);
}

/* note : this variant can't fail */
size_t ZSTD_initDStream(ZSTD_DStream* zds)
{
    DEBUGLOG(4, "ZSTD_initDStream");
    FORWARD_IF_ERROR(ZSTD_DCtx_reset(zds, ZSTD_reset_session_only), "");
    FORWARD_IF_ERROR(ZSTD_DCtx_refDDict(zds, NULL), "");
    return ZSTD_startingInputLength(zds->format);
}

/* ZSTD_initDStream_usingDDict() :
 * ddict will just be referenced, and must outlive decompression session
 * this function cannot fail */
size_t ZSTD_initDStream_usingDDict(ZSTD_DStream* dctx, const ZSTD_DDict* ddict)
{
    DEBUGLOG(4, "ZSTD_initDStream_usingDDict");
    FORWARD_IF_ERROR( ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only) , "");
    FORWARD_IF_ERROR( ZSTD_DCtx_refDDict(dctx, ddict) , "");
    return ZSTD_startingInputLength(dctx->format);
}

/* ZSTD_resetDStream() :
 * return : expected size, aka ZSTD_startingInputLength().
 * this function cannot fail */
size_t ZSTD_resetDStream(ZSTD_DStream* dctx)
{
    DEBUGLOG(4, "ZSTD_resetDStream");
    FORWARD_IF_ERROR(ZSTD_DCtx_reset(dctx, ZSTD_reset_session_only), "");
    return ZSTD_startingInputLength(dctx->format);
}


size_t ZSTD_DCtx_refDDict(ZSTD_DCtx* dctx, const ZSTD_DDict* ddict)
{
    RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
    ZSTD_clearDict(dctx);
    if (ddict) {
        dctx->ddict = ddict;
        dctx->dictUses = ZSTD_use_indefinitely;
        if (dctx->refMultipleDDicts == ZSTD_rmd_refMultipleDDicts) {
            if (dctx->ddictSet == NULL) {
                dctx->ddictSet = ZSTD_createDDictHashSet(dctx->customMem);
                if (!dctx->ddictSet) {
                    RETURN_ERROR(memory_allocation, "Failed to allocate memory for hash set!");
                }
            }
            assert(!dctx->staticSize);  /* Impossible: ddictSet cannot have been allocated if static dctx */
            FORWARD_IF_ERROR(ZSTD_DDictHashSet_addDDict(dctx->ddictSet, ddict, dctx->customMem), "");
        }
    }
    return 0;
}

/* ZSTD_DCtx_setMaxWindowSize() :
 * note : no direct equivalence in ZSTD_DCtx_setParameter,
 * since this version sets windowSize, and the other sets windowLog */
size_t ZSTD_DCtx_setMaxWindowSize(ZSTD_DCtx* dctx, size_t maxWindowSize)
{
    ZSTD_bounds const bounds = ZSTD_dParam_getBounds(ZSTD_d_windowLogMax);
    size_t const min = (size_t)1 << bounds.lowerBound;
    size_t const max = (size_t)1 << bounds.upperBound;
    RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
    RETURN_ERROR_IF(maxWindowSize < min, parameter_outOfBound, "");
    RETURN_ERROR_IF(maxWindowSize > max, parameter_outOfBound, "");
    dctx->maxWindowSize = maxWindowSize;
    return 0;
}

size_t ZSTD_DCtx_setFormat(ZSTD_DCtx* dctx, ZSTD_format_e format)
{
    return ZSTD_DCtx_setParameter(dctx, ZSTD_d_format, (int)format);
}

ZSTD_bounds ZSTD_dParam_getBounds(ZSTD_dParameter dParam)
{
    ZSTD_bounds bounds = { 0, 0, 0 };
    switch(dParam) {
        case ZSTD_d_windowLogMax:
            bounds.lowerBound = ZSTD_WINDOWLOG_ABSOLUTEMIN;
            bounds.upperBound = ZSTD_WINDOWLOG_MAX;
            return bounds;
        case ZSTD_d_format:
            bounds.lowerBound = (int)ZSTD_f_zstd1;
            bounds.upperBound = (int)ZSTD_f_zstd1_magicless;
            ZSTD_STATIC_ASSERT(ZSTD_f_zstd1 < ZSTD_f_zstd1_magicless);
            return bounds;
        case ZSTD_d_stableOutBuffer:
            bounds.lowerBound = (int)ZSTD_bm_buffered;
            bounds.upperBound = (int)ZSTD_bm_stable;
            return bounds;
        case ZSTD_d_forceIgnoreChecksum:
            bounds.lowerBound = (int)ZSTD_d_validateChecksum;
            bounds.upperBound = (int)ZSTD_d_ignoreChecksum;
            return bounds;
        case ZSTD_d_refMultipleDDicts:
            bounds.lowerBound = (int)ZSTD_rmd_refSingleDDict;
            bounds.upperBound = (int)ZSTD_rmd_refMultipleDDicts;
            return bounds;
        case ZSTD_d_disableHuffmanAssembly:
            bounds.lowerBound = 0;
            bounds.upperBound = 1;
            return bounds;
        case ZSTD_d_maxBlockSize:
            bounds.lowerBound = ZSTD_BLOCKSIZE_MAX_MIN;
            bounds.upperBound = ZSTD_BLOCKSIZE_MAX;
            return bounds;

        default:;
    }
    bounds.error = ERROR(parameter_unsupported);
    return bounds;
}

/* ZSTD_dParam_withinBounds:
 * @return 1 if value is within dParam bounds,
 * 0 otherwise */
static int ZSTD_dParam_withinBounds(ZSTD_dParameter dParam, int value)
{
    ZSTD_bounds const bounds = ZSTD_dParam_getBounds(dParam);
    if (ZSTD_isError(bounds.error)) return 0;
    if (value < bounds.lowerBound) return 0;
    if (value > bounds.upperBound) return 0;
    return 1;
}

#define CHECK_DBOUNDS(p,v) {                \
    RETURN_ERROR_IF(!ZSTD_dParam_withinBounds(p, v), parameter_outOfBound, ""); \
}

size_t ZSTD_DCtx_getParameter(ZSTD_DCtx* dctx, ZSTD_dParameter param, int* value)
{
    switch (param) {
        case ZSTD_d_windowLogMax:
            *value = (int)ZSTD_highbit32((U32)dctx->maxWindowSize);
            return 0;
        case ZSTD_d_format:
            *value = (int)dctx->format;
            return 0;
        case ZSTD_d_stableOutBuffer:
            *value = (int)dctx->outBufferMode;
            return 0;
        case ZSTD_d_forceIgnoreChecksum:
            *value = (int)dctx->forceIgnoreChecksum;
            return 0;
        case ZSTD_d_refMultipleDDicts:
            *value = (int)dctx->refMultipleDDicts;
            return 0;
        case ZSTD_d_disableHuffmanAssembly:
            *value = (int)dctx->disableHufAsm;
            return 0;
        case ZSTD_d_maxBlockSize:
            *value = dctx->maxBlockSizeParam;
            return 0;
        default:;
    }
    RETURN_ERROR(parameter_unsupported, "");
}

size_t ZSTD_DCtx_setParameter(ZSTD_DCtx* dctx, ZSTD_dParameter dParam, int value)
{
    RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
    switch(dParam) {
        case ZSTD_d_windowLogMax:
            if (value == 0) value = ZSTD_WINDOWLOG_LIMIT_DEFAULT;
            CHECK_DBOUNDS(ZSTD_d_windowLogMax, value);
            dctx->maxWindowSize = ((size_t)1) << value;
            return 0;
        case ZSTD_d_format:
            CHECK_DBOUNDS(ZSTD_d_format, value);
            dctx->format = (ZSTD_format_e)value;
            return 0;
        case ZSTD_d_stableOutBuffer:
            CHECK_DBOUNDS(ZSTD_d_stableOutBuffer, value);
            dctx->outBufferMode = (ZSTD_bufferMode_e)value;
            return 0;
        case ZSTD_d_forceIgnoreChecksum:
            CHECK_DBOUNDS(ZSTD_d_forceIgnoreChecksum, value);
            dctx->forceIgnoreChecksum = (ZSTD_forceIgnoreChecksum_e)value;
            return 0;
        case ZSTD_d_refMultipleDDicts:
            CHECK_DBOUNDS(ZSTD_d_refMultipleDDicts, value);
            if (dctx->staticSize != 0) {
                RETURN_ERROR(parameter_unsupported, "Static dctx does not support multiple DDicts!");
            }
            dctx->refMultipleDDicts = (ZSTD_refMultipleDDicts_e)value;
            return 0;
        case ZSTD_d_disableHuffmanAssembly:
            CHECK_DBOUNDS(ZSTD_d_disableHuffmanAssembly, value);
            dctx->disableHufAsm = value != 0;
            return 0;
        case ZSTD_d_maxBlockSize:
            if (value != 0) CHECK_DBOUNDS(ZSTD_d_maxBlockSize, value);
            dctx->maxBlockSizeParam = value;
            return 0;
        default:;
    }
    RETURN_ERROR(parameter_unsupported, "");
}

size_t ZSTD_DCtx_reset(ZSTD_DCtx* dctx, ZSTD_ResetDirective reset)
{
    if ( (reset == ZSTD_reset_session_only)
      || (reset == ZSTD_reset_session_and_parameters) ) {
        dctx->streamStage = zdss_init;
        dctx->noForwardProgress = 0;
        dctx->isFrameDecompression = 1;
    }
    if ( (reset == ZSTD_reset_parameters)
      || (reset == ZSTD_reset_session_and_parameters) ) {
        RETURN_ERROR_IF(dctx->streamStage != zdss_init, stage_wrong, "");
        ZSTD_clearDict(dctx);
        ZSTD_DCtx_resetParameters(dctx);
    }
    return 0;
}


size_t ZSTD_sizeof_DStream(const ZSTD_DStream* dctx)
{
    return ZSTD_sizeof_DCtx(dctx);
}

static size_t ZSTD_decodingBufferSize_internal(unsigned long long windowSize, unsigned long long frameContentSize, size_t blockSizeMax)
{
    size_t const blockSize = MIN((size_t)MIN(windowSize, ZSTD_BLOCKSIZE_MAX), blockSizeMax);
    /* We need blockSize + WILDCOPY_OVERLENGTH worth of buffer so that if a block
     * ends at windowSize + WILDCOPY_OVERLENGTH + 1 bytes, we can start writing
     * the block at the beginning of the output buffer, and maintain a full window.
     *
     * We need another blockSize worth of buffer so that we can store split
     * literals at the end of the block without overwriting the extDict window.
     */
    unsigned long long const neededRBSize = windowSize + (blockSize * 2) + (WILDCOPY_OVERLENGTH * 2);
    unsigned long long const neededSize = MIN(frameContentSize, neededRBSize);
    size_t const minRBSize = (size_t) neededSize;
    RETURN_ERROR_IF((unsigned long long)minRBSize != neededSize,
                    frameParameter_windowTooLarge, "");
    return minRBSize;
}

size_t ZSTD_decodingBufferSize_min(unsigned long long windowSize, unsigned long long frameContentSize)
{
    return ZSTD_decodingBufferSize_internal(windowSize, frameContentSize, ZSTD_BLOCKSIZE_MAX);
}

size_t ZSTD_estimateDStreamSize(size_t windowSize)
{
    size_t const blockSize = MIN(windowSize, ZSTD_BLOCKSIZE_MAX);
    size_t const inBuffSize = blockSize;  /* no block can be larger */
    size_t const outBuffSize = ZSTD_decodingBufferSize_min(windowSize, ZSTD_CONTENTSIZE_UNKNOWN);
    return ZSTD_estimateDCtxSize() + inBuffSize + outBuffSize;
}

size_t ZSTD_estimateDStreamSize_fromFrame(const void* src, size_t srcSize)
{
    U32 const windowSizeMax = 1U << ZSTD_WINDOWLOG_MAX;   /* note : should be user-selectable, but requires an additional parameter (or a dctx) */
    ZSTD_FrameHeader zfh;
    size_t const err = ZSTD_getFrameHeader(&zfh, src, srcSize);
    if (ZSTD_isError(err)) return err;
    RETURN_ERROR_IF(err>0, srcSize_wrong, "");
    RETURN_ERROR_IF(zfh.windowSize > windowSizeMax,
                    frameParameter_windowTooLarge, "");
    return ZSTD_estimateDStreamSize((size_t)zfh.windowSize);
}


/* *****   Decompression   ***** */

static int ZSTD_DCtx_isOverflow(ZSTD_DStream* zds, size_t const neededInBuffSize, size_t const neededOutBuffSize)
{
    return (zds->inBuffSize + zds->outBuffSize) >= (neededInBuffSize + neededOutBuffSize) * ZSTD_WORKSPACETOOLARGE_FACTOR;
}

static void ZSTD_DCtx_updateOversizedDuration(ZSTD_DStream* zds, size_t const neededInBuffSize, size_t const neededOutBuffSize)
{
    if (ZSTD_DCtx_isOverflow(zds, neededInBuffSize, neededOutBuffSize))
        zds->oversizedDuration++;
    else
        zds->oversizedDuration = 0;
}

static int ZSTD_DCtx_isOversizedTooLong(ZSTD_DStream* zds)
{
    return zds->oversizedDuration >= ZSTD_WORKSPACETOOLARGE_MAXDURATION;
}

/* Checks that the output buffer hasn't changed if ZSTD_obm_stable is used. */
static size_t ZSTD_checkOutBuffer(ZSTD_DStream const* zds, ZSTD_outBuffer const* output)
{
    ZSTD_outBuffer const expect = zds->expectedOutBuffer;
    /* No requirement when ZSTD_obm_stable is not enabled. */
    if (zds->outBufferMode != ZSTD_bm_stable)
        return 0;
    /* Any buffer is allowed in zdss_init, this must be the same for every other call until
     * the context is reset.
     */
    if (zds->streamStage == zdss_init)
        return 0;
    /* The buffer must match our expectation exactly. */
    if (expect.dst == output->dst && expect.pos == output->pos && expect.size == output->size)
        return 0;
    RETURN_ERROR(dstBuffer_wrong, "ZSTD_d_stableOutBuffer enabled but output differs!");
}

/* Calls ZSTD_decompressContinue() with the right parameters for ZSTD_decompressStream()
 * and updates the stage and the output buffer state. This call is extracted so it can be
 * used both when reading directly from the ZSTD_inBuffer, and in buffered input mode.
 * NOTE: You must break after calling this function since the streamStage is modified.
 */
static size_t ZSTD_decompressContinueStream(
            ZSTD_DStream* zds, char** op, char* oend,
            void const* src, size_t srcSize) {
    int const isSkipFrame = ZSTD_isSkipFrame(zds);
    if (zds->outBufferMode == ZSTD_bm_buffered) {
        size_t const dstSize = isSkipFrame ? 0 : zds->outBuffSize - zds->outStart;
        size_t const decodedSize = ZSTD_decompressContinue(zds,
                zds->outBuff + zds->outStart, dstSize, src, srcSize);
        FORWARD_IF_ERROR(decodedSize, "");
        if (!decodedSize && !isSkipFrame) {
            zds->streamStage = zdss_read;
        } else {
            zds->outEnd = zds->outStart + decodedSize;
            zds->streamStage = zdss_flush;
        }
    } else {
        /* Write directly into the output buffer */
        size_t const dstSize = isSkipFrame ? 0 : (size_t)(oend - *op);
        size_t const decodedSize = ZSTD_decompressContinue(zds, *op, dstSize, src, srcSize);
        FORWARD_IF_ERROR(decodedSize, "");
        *op += decodedSize;
        /* Flushing is not needed. */
        zds->streamStage = zdss_read;
        assert(*op <= oend);
        assert(zds->outBufferMode == ZSTD_bm_stable);
    }
    return 0;
}

size_t ZSTD_decompressStream(ZSTD_DStream* zds, ZSTD_outBuffer* output, ZSTD_inBuffer* input)
{
    const char* const src = (const char*)input->src;
    const char* const istart = input->pos != 0 ? src + input->pos : src;
    const char* const iend = input->size != 0 ? src + input->size : src;
    const char* ip = istart;
    char* const dst = (char*)output->dst;
    char* const ostart = output->pos != 0 ? dst + output->pos : dst;
    char* const oend = output->size != 0 ? dst + output->size : dst;
    char* op = ostart;
    U32 someMoreWork = 1;

    DEBUGLOG(5, "ZSTD_decompressStream");
    assert(zds != NULL);
    RETURN_ERROR_IF(
        input->pos > input->size,
        srcSize_wrong,
        "forbidden. in: pos: %u   vs size: %u",
        (U32)input->pos, (U32)input->size);
    RETURN_ERROR_IF(
        output->pos > output->size,
        dstSize_tooSmall,
        "forbidden. out: pos: %u   vs size: %u",
        (U32)output->pos, (U32)output->size);
    DEBUGLOG(5, "input size : %u", (U32)(input->size - input->pos));
    FORWARD_IF_ERROR(ZSTD_checkOutBuffer(zds, output), "");

    while (someMoreWork) {
        switch(zds->streamStage)
        {
        case zdss_init :
            DEBUGLOG(5, "stage zdss_init => transparent reset ");
            zds->streamStage = zdss_loadHeader;
            zds->lhSize = zds->inPos = zds->outStart = zds->outEnd = 0;
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
            zds->legacyVersion = 0;
#endif
            zds->hostageByte = 0;
            zds->expectedOutBuffer = *output;
            ZSTD_FALLTHROUGH;

        case zdss_loadHeader :
            DEBUGLOG(5, "stage zdss_loadHeader (srcSize : %u)", (U32)(iend - ip));
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
            if (zds->legacyVersion) {
                RETURN_ERROR_IF(zds->staticSize, memory_allocation,
                    "legacy support is incompatible with static dctx");
                {   size_t const hint = ZSTD_decompressLegacyStream(zds->legacyContext, zds->legacyVersion, output, input);
                    if (hint==0) zds->streamStage = zdss_init;
                    return hint;
            }   }
#endif
            {   size_t const hSize = ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format);
                if (zds->refMultipleDDicts && zds->ddictSet) {
                    ZSTD_DCtx_selectFrameDDict(zds);
                }
                if (ZSTD_isError(hSize)) {
#if defined(ZSTD_LEGACY_SUPPORT) && (ZSTD_LEGACY_SUPPORT>=1)
                    U32 const legacyVersion = ZSTD_isLegacy(istart, iend-istart);
                    if (legacyVersion) {
                        ZSTD_DDict const* const ddict = ZSTD_getDDict(zds);
                        const void* const dict = ddict ? ZSTD_DDict_dictContent(ddict) : NULL;
                        size_t const dictSize = ddict ? ZSTD_DDict_dictSize(ddict) : 0;
                        DEBUGLOG(5, "ZSTD_decompressStream: detected legacy version v0.%u", legacyVersion);
                        RETURN_ERROR_IF(zds->staticSize, memory_allocation,
                            "legacy support is incompatible with static dctx");
                        FORWARD_IF_ERROR(ZSTD_initLegacyStream(&zds->legacyContext,
                                    zds->previousLegacyVersion, legacyVersion,
                                    dict, dictSize), "");
                        zds->legacyVersion = zds->previousLegacyVersion = legacyVersion;
                        {   size_t const hint = ZSTD_decompressLegacyStream(zds->legacyContext, legacyVersion, output, input);
                            if (hint==0) zds->streamStage = zdss_init;   /* or stay in stage zdss_loadHeader */
                            return hint;
                    }   }
#endif
                    return hSize;   /* error */
                }
                if (hSize != 0) {   /* need more input */
                    size_t const toLoad = hSize - zds->lhSize;   /* if hSize!=0, hSize > zds->lhSize */
                    size_t const remainingInput = (size_t)(iend-ip);
                    assert(iend >= ip);
                    if (toLoad > remainingInput) {   /* not enough input to load full header */
                        if (remainingInput > 0) {
                            ZSTD_memcpy(zds->headerBuffer + zds->lhSize, ip, remainingInput);
                            zds->lhSize += remainingInput;
                        }
                        input->pos = input->size;
                        /* check first few bytes */
                        FORWARD_IF_ERROR(
                            ZSTD_getFrameHeader_advanced(&zds->fParams, zds->headerBuffer, zds->lhSize, zds->format),
                            "First few bytes detected incorrect" );
                        /* return hint input size */
                        return (MAX((size_t)ZSTD_FRAMEHEADERSIZE_MIN(zds->format), hSize) - zds->lhSize) + ZSTD_blockHeaderSize;   /* remaining header bytes + next block header */
                    }
                    assert(ip != NULL);
                    ZSTD_memcpy(zds->headerBuffer + zds->lhSize, ip, toLoad); zds->lhSize = hSize; ip += toLoad;
                    break;
            }   }

            /* check for single-pass mode opportunity */
            if (zds->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                && zds->fParams.frameType != ZSTD_skippableFrame
                && (U64)(size_t)(oend-op) >= zds->fParams.frameContentSize) {
                size_t const cSize = ZSTD_findFrameCompressedSize_advanced(istart, (size_t)(iend-istart), zds->format);
                if (cSize <= (size_t)(iend-istart)) {
                    /* shortcut : using single-pass mode */
                    size_t const decompressedSize = ZSTD_decompress_usingDDict(zds, op, (size_t)(oend-op), istart, cSize, ZSTD_getDDict(zds));
                    if (ZSTD_isError(decompressedSize)) return decompressedSize;
                    DEBUGLOG(4, "shortcut to single-pass ZSTD_decompress_usingDDict()");
                    assert(istart != NULL);
                    ip = istart + cSize;
                    op = op ? op + decompressedSize : op; /* can occur if frameContentSize = 0 (empty frame) */
                    zds->expected = 0;
                    zds->streamStage = zdss_init;
                    someMoreWork = 0;
                    break;
            }   }

            /* Check output buffer is large enough for ZSTD_odm_stable. */
            if (zds->outBufferMode == ZSTD_bm_stable
                && zds->fParams.frameType != ZSTD_skippableFrame
                && zds->fParams.frameContentSize != ZSTD_CONTENTSIZE_UNKNOWN
                && (U64)(size_t)(oend-op) < zds->fParams.frameContentSize) {
                RETURN_ERROR(dstSize_tooSmall, "ZSTD_obm_stable passed but ZSTD_outBuffer is too small");
            }

            /* Consume header (see ZSTDds_decodeFrameHeader) */
            DEBUGLOG(4, "Consume header");
            FORWARD_IF_ERROR(ZSTD_decompressBegin_usingDDict(zds, ZSTD_getDDict(zds)), "");

            if (zds->format == ZSTD_f_zstd1
                && (MEM_readLE32(zds->headerBuffer) & ZSTD_MAGIC_SKIPPABLE_MASK) == ZSTD_MAGIC_SKIPPABLE_START) {  /* skippable frame */
                zds->expected = MEM_readLE32(zds->headerBuffer + ZSTD_FRAMEIDSIZE);
                zds->stage = ZSTDds_skipFrame;
            } else {
                FORWARD_IF_ERROR(ZSTD_decodeFrameHeader(zds, zds->headerBuffer, zds->lhSize), "");
                zds->expected = ZSTD_blockHeaderSize;
                zds->stage = ZSTDds_decodeBlockHeader;
            }

            /* control buffer memory usage */
            DEBUGLOG(4, "Control max memory usage (%u KB <= max %u KB)",
                        (U32)(zds->fParams.windowSize >>10),
                        (U32)(zds->maxWindowSize >> 10) );
            zds->fParams.windowSize = MAX(zds->fParams.windowSize, 1U << ZSTD_WINDOWLOG_ABSOLUTEMIN);
            RETURN_ERROR_IF(zds->fParams.windowSize > zds->maxWindowSize,
                            frameParameter_windowTooLarge, "");
            if (zds->maxBlockSizeParam != 0)
                zds->fParams.blockSizeMax = MIN(zds->fParams.blockSizeMax, (unsigned)zds->maxBlockSizeParam);

            /* Adapt buffer sizes to frame header instructions */
            {   size_t const neededInBuffSize = MAX(zds->fParams.blockSizeMax, 4 /* frame checksum */);
                size_t const neededOutBuffSize = zds->outBufferMode == ZSTD_bm_buffered
                        ? ZSTD_decodingBufferSize_internal(zds->fParams.windowSize, zds->fParams.frameContentSize, zds->fParams.blockSizeMax)
                        : 0;

                ZSTD_DCtx_updateOversizedDuration(zds, neededInBuffSize, neededOutBuffSize);

                {   int const tooSmall = (zds->inBuffSize < neededInBuffSize) || (zds->outBuffSize < neededOutBuffSize);
                    int const tooLarge = ZSTD_DCtx_isOversizedTooLong(zds);

                    if (tooSmall || tooLarge) {
                        size_t const bufferSize = neededInBuffSize + neededOutBuffSize;
                        DEBUGLOG(4, "inBuff  : from %u to %u",
                                    (U32)zds->inBuffSize, (U32)neededInBuffSize);
                        DEBUGLOG(4, "outBuff : from %u to %u",
                                    (U32)zds->outBuffSize, (U32)neededOutBuffSize);
                        if (zds->staticSize) {  /* static DCtx */
                            DEBUGLOG(4, "staticSize : %u", (U32)zds->staticSize);
                            assert(zds->staticSize >= sizeof(ZSTD_DCtx));  /* controlled at init */
                            RETURN_ERROR_IF(
                                bufferSize > zds->staticSize - sizeof(ZSTD_DCtx),
                                memory_allocation, "");
                        } else {
                            ZSTD_customFree(zds->inBuff, zds->customMem);
                            zds->inBuffSize = 0;
                            zds->outBuffSize = 0;
                            zds->inBuff = (char*)ZSTD_customMalloc(bufferSize, zds->customMem);
                            RETURN_ERROR_IF(zds->inBuff == NULL, memory_allocation, "");
                        }
                        zds->inBuffSize = neededInBuffSize;
                        zds->outBuff = zds->inBuff + zds->inBuffSize;
                        zds->outBuffSize = neededOutBuffSize;
            }   }   }
            zds->streamStage = zdss_read;
            ZSTD_FALLTHROUGH;

        case zdss_read:
            DEBUGLOG(5, "stage zdss_read");
            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip));
                DEBUGLOG(5, "neededInSize = %u", (U32)neededInSize);
                if (neededInSize==0) {  /* end of frame */
                    zds->streamStage = zdss_init;
                    someMoreWork = 0;
                    break;
                }
                if ((size_t)(iend-ip) >= neededInSize) {  /* decode directly from src */
                    FORWARD_IF_ERROR(ZSTD_decompressContinueStream(zds, &op, oend, ip, neededInSize), "");
                    assert(ip != NULL);
                    ip += neededInSize;
                    /* Function modifies the stage so we must break */
                    break;
            }   }
            if (ip==iend) { someMoreWork = 0; break; }   /* no more input */
            zds->streamStage = zdss_load;
            ZSTD_FALLTHROUGH;

        case zdss_load:
            {   size_t const neededInSize = ZSTD_nextSrcSizeToDecompress(zds);
                size_t const toLoad = neededInSize - zds->inPos;
                int const isSkipFrame = ZSTD_isSkipFrame(zds);
                size_t loadedSize;
                /* At this point we shouldn't be decompressing a block that we can stream. */
                assert(neededInSize == ZSTD_nextSrcSizeToDecompressWithInputSize(zds, (size_t)(iend - ip)));
                if (isSkipFrame) {
                    loadedSize = MIN(toLoad, (size_t)(iend-ip));
                } else {
                    RETURN_ERROR_IF(toLoad > zds->inBuffSize - zds->inPos,
                                    corruption_detected,
                                    "should never happen");
                    loadedSize = ZSTD_limitCopy(zds->inBuff + zds->inPos, toLoad, ip, (size_t)(iend-ip));
                }
                if (loadedSize != 0) {
                    /* ip may be NULL */
                    ip += loadedSize;
                    zds->inPos += loadedSize;
                }
                if (loadedSize < toLoad) { someMoreWork = 0; break; }   /* not enough input, wait for more */

                /* decode loaded input */
                zds->inPos = 0;   /* input is consumed */
                FORWARD_IF_ERROR(ZSTD_decompressContinueStream(zds, &op, oend, zds->inBuff, neededInSize), "");
                /* Function modifies the stage so we must break */
                break;
            }
        case zdss_flush:
            {
                size_t const toFlushSize = zds->outEnd - zds->outStart;
                size_t const flushedSize = ZSTD_limitCopy(op, (size_t)(oend-op), zds->outBuff + zds->outStart, toFlushSize);

                op = op ? op + flushedSize : op;

                zds->outStart += flushedSize;
                if (flushedSize == toFlushSize) {  /* flush completed */
                    zds->streamStage = zdss_read;
                    if ( (zds->outBuffSize < zds->fParams.frameContentSize)
                        && (zds->outStart + zds->fParams.blockSizeMax > zds->outBuffSize) ) {
                        DEBUGLOG(5, "restart filling outBuff from beginning (left:%i, needed:%u)",
                                (int)(zds->outBuffSize - zds->outStart),
                                (U32)zds->fParams.blockSizeMax);
                        zds->outStart = zds->outEnd = 0;
                    }
                    break;
            }   }
            /* cannot complete flush */
            someMoreWork = 0;
            break;

        default:
            assert(0);    /* impossible */
            RETURN_ERROR(GENERIC, "impossible to reach");   /* some compilers require default to do something */
    }   }

    /* result */
    input->pos = (size_t)(ip - (const char*)(input->src));
    output->pos = (size_t)(op - (char*)(output->dst));

    /* Update the expected output buffer for ZSTD_obm_stable. */
    zds->expectedOutBuffer = *output;

    if ((ip==istart) && (op==ostart)) {  /* no forward progress */
        zds->noForwardProgress ++;
        if (zds->noForwardProgress >= ZSTD_NO_FORWARD_PROGRESS_MAX) {
            RETURN_ERROR_IF(op==oend, noForwardProgress_destFull, "");
            RETURN_ERROR_IF(ip==iend, noForwardProgress_inputEmpty, "");
            assert(0);
        }
    } else {
        zds->noForwardProgress = 0;
    }
    {   size_t nextSrcSizeHint = ZSTD_nextSrcSizeToDecompress(zds);
        if (!nextSrcSizeHint) {   /* frame fully decoded */
            if (zds->outEnd == zds->outStart) {  /* output fully flushed */
                if (zds->hostageByte) {
                    if (input->pos >= input->size) {
                        /* can't release hostage (not present) */
                        zds->streamStage = zdss_read;
                        return 1;
                    }
                    input->pos++;  /* release hostage */
                }   /* zds->hostageByte */
                return 0;
            }  /* zds->outEnd == zds->outStart */
            if (!zds->hostageByte) { /* output not fully flushed; keep last byte as hostage; will be released when all output is flushed */
                input->pos--;   /* note : pos > 0, otherwise, impossible to finish reading last block */
                zds->hostageByte=1;
            }
            return 1;
        }  /* nextSrcSizeHint==0 */
        nextSrcSizeHint += ZSTD_blockHeaderSize * (ZSTD_nextInputType(zds) == ZSTDnit_block);   /* preload header of next block */
        assert(zds->inPos <= nextSrcSizeHint);
        nextSrcSizeHint -= zds->inPos;   /* part already loaded*/
        return nextSrcSizeHint;
    }
}

size_t ZSTD_decompressStream_simpleArgs (
                            ZSTD_DCtx* dctx,
                            void* dst, size_t dstCapacity, size_t* dstPos,
                      const void* src, size_t srcSize, size_t* srcPos)
{
    ZSTD_outBuffer output;
    ZSTD_inBuffer  input;
    output.dst = dst;
    output.size = dstCapacity;
    output.pos = *dstPos;
    input.src = src;
    input.size = srcSize;
    input.pos = *srcPos;
    {   size_t const cErr = ZSTD_decompressStream(dctx, &output, &input);
        *dstPos = output.pos;
        *srcPos = input.pos;
        return cErr;
    }
}
/**** ended inlining decompress/zstd_decompress.c ****/
/**** start inlining decompress/zstd_decompress_block.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* zstd_decompress_block :
 * this module takes care of decompressing _compressed_ block */

/*-*******************************************************
*  Dependencies
*********************************************************/
/**** skipping file: ../common/zstd_deps.h ****/
/**** skipping file: ../common/compiler.h ****/
/**** skipping file: ../common/cpu.h ****/
/**** skipping file: ../common/mem.h ****/
#define FSE_STATIC_LINKING_ONLY
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: zstd_decompress_internal.h ****/
/**** skipping file: zstd_ddict.h ****/
/**** skipping file: zstd_decompress_block.h ****/
/**** skipping file: ../common/bits.h ****/

/*_*******************************************************
*  Macros
**********************************************************/

/* These two optional macros force the use one way or another of the two
 * ZSTD_decompressSequences implementations. You can't force in both directions
 * at the same time.
 */
#if defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
    defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
#error "Cannot force the use of the short and the long ZSTD_decompressSequences variants!"
#endif


/*_*******************************************************
*  Memory operations
**********************************************************/
static void ZSTD_copy4(void* dst, const void* src) { ZSTD_memcpy(dst, src, 4); }


/*-*************************************************************
 *   Block decoding
 ***************************************************************/

static size_t ZSTD_blockSizeMax(ZSTD_DCtx const* dctx)
{
    size_t const blockSizeMax = dctx->isFrameDecompression ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX;
    assert(blockSizeMax <= ZSTD_BLOCKSIZE_MAX);
    return blockSizeMax;
}

/*! ZSTD_getcBlockSize() :
 *  Provides the size of compressed block from block header `src` */
size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
                          blockProperties_t* bpPtr)
{
    RETURN_ERROR_IF(srcSize < ZSTD_blockHeaderSize, srcSize_wrong, "");

    {   U32 const cBlockHeader = MEM_readLE24(src);
        U32 const cSize = cBlockHeader >> 3;
        bpPtr->lastBlock = cBlockHeader & 1;
        bpPtr->blockType = (blockType_e)((cBlockHeader >> 1) & 3);
        bpPtr->origSize = cSize;   /* only useful for RLE */
        if (bpPtr->blockType == bt_rle) return 1;
        RETURN_ERROR_IF(bpPtr->blockType == bt_reserved, corruption_detected, "");
        return cSize;
    }
}

/* Allocate buffer for literals, either overlapping current dst, or split between dst and litExtraBuffer, or stored entirely within litExtraBuffer */
static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const size_t dstCapacity, const size_t litSize,
    const streaming_operation streaming, const size_t expectedWriteSize, const unsigned splitImmediately)
{
    size_t const blockSizeMax = ZSTD_blockSizeMax(dctx);
    assert(litSize <= blockSizeMax);
    assert(dctx->isFrameDecompression || streaming == not_streaming);
    assert(expectedWriteSize <= blockSizeMax);
    if (streaming == not_streaming && dstCapacity > blockSizeMax + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH) {
        /* If we aren't streaming, we can just put the literals after the output
         * of the current block. We don't need to worry about overwriting the
         * extDict of our window, because it doesn't exist.
         * So if we have space after the end of the block, just put it there.
         */
        dctx->litBuffer = (BYTE*)dst + blockSizeMax + WILDCOPY_OVERLENGTH;
        dctx->litBufferEnd = dctx->litBuffer + litSize;
        dctx->litBufferLocation = ZSTD_in_dst;
    } else if (litSize <= ZSTD_LITBUFFEREXTRASIZE) {
        /* Literals fit entirely within the extra buffer, put them there to avoid
         * having to split the literals.
         */
        dctx->litBuffer = dctx->litExtraBuffer;
        dctx->litBufferEnd = dctx->litBuffer + litSize;
        dctx->litBufferLocation = ZSTD_not_in_dst;
    } else {
        assert(blockSizeMax > ZSTD_LITBUFFEREXTRASIZE);
        /* Literals must be split between the output block and the extra lit
         * buffer. We fill the extra lit buffer with the tail of the literals,
         * and put the rest of the literals at the end of the block, with
         * WILDCOPY_OVERLENGTH of buffer room to allow for overreads.
         * This MUST not write more than our maxBlockSize beyond dst, because in
         * streaming mode, that could overwrite part of our extDict window.
         */
        if (splitImmediately) {
            /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
            dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
            dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
        } else {
            /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */
            dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
            dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
        }
        dctx->litBufferLocation = ZSTD_split;
        assert(dctx->litBufferEnd <= (BYTE*)dst + expectedWriteSize);
    }
}

/*! ZSTD_decodeLiteralsBlock() :
 * Where it is possible to do so without being stomped by the output during decompression, the literals block will be stored
 * in the dstBuffer.  If there is room to do so, it will be stored in full in the excess dst space after where the current
 * block will be output.  Otherwise it will be stored at the end of the current dst blockspace, with a small portion being
 * stored in dctx->litExtraBuffer to help keep it "ahead" of the current output write.
 *
 * @return : nb of bytes read from src (< srcSize )
 *  note : symbol not declared but exposed for fullbench */
static size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
                          const void* src, size_t srcSize,   /* note : srcSize < BLOCKSIZE */
                          void* dst, size_t dstCapacity, const streaming_operation streaming)
{
    DEBUGLOG(5, "ZSTD_decodeLiteralsBlock");
    RETURN_ERROR_IF(srcSize < MIN_CBLOCK_SIZE, corruption_detected, "");

    {   const BYTE* const istart = (const BYTE*) src;
        SymbolEncodingType_e const litEncType = (SymbolEncodingType_e)(istart[0] & 3);
        size_t const blockSizeMax = ZSTD_blockSizeMax(dctx);

        switch(litEncType)
        {
        case set_repeat:
            DEBUGLOG(5, "set_repeat flag : re-using stats from previous compressed literals block");
            RETURN_ERROR_IF(dctx->litEntropy==0, dictionary_corrupted, "");
            ZSTD_FALLTHROUGH;

        case set_compressed:
            RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need up to 5 for case 3");
            {   size_t lhSize, litSize, litCSize;
                U32 singleStream=0;
                U32 const lhlCode = (istart[0] >> 2) & 3;
                U32 const lhc = MEM_readLE32(istart);
                size_t hufSuccess;
                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
                int const flags = 0
                    | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0)
                    | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0);
                switch(lhlCode)
                {
                case 0: case 1: default:   /* note : default is impossible, since lhlCode into [0..3] */
                    /* 2 - 2 - 10 - 10 */
                    singleStream = !lhlCode;
                    lhSize = 3;
                    litSize  = (lhc >> 4) & 0x3FF;
                    litCSize = (lhc >> 14) & 0x3FF;
                    break;
                case 2:
                    /* 2 - 2 - 14 - 14 */
                    lhSize = 4;
                    litSize  = (lhc >> 4) & 0x3FFF;
                    litCSize = lhc >> 18;
                    break;
                case 3:
                    /* 2 - 2 - 18 - 18 */
                    lhSize = 5;
                    litSize  = (lhc >> 4) & 0x3FFFF;
                    litCSize = (lhc >> 22) + ((size_t)istart[4] << 10);
                    break;
                }
                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
                if (!singleStream)
                    RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong,
                        "Not enough literals (%zu) for the 4-streams mode (min %u)",
                        litSize, MIN_LITERALS_FOR_4_STREAMS);
                RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
                RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);

                /* prefetch huffman table if cold */
                if (dctx->ddictIsCold && (litSize > 768 /* heuristic */)) {
                    PREFETCH_AREA(dctx->HUFptr, sizeof(dctx->entropy.hufTable));
                }

                if (litEncType==set_repeat) {
                    if (singleStream) {
                        hufSuccess = HUF_decompress1X_usingDTable(
                            dctx->litBuffer, litSize, istart+lhSize, litCSize,
                            dctx->HUFptr, flags);
                    } else {
                        assert(litSize >= MIN_LITERALS_FOR_4_STREAMS);
                        hufSuccess = HUF_decompress4X_usingDTable(
                            dctx->litBuffer, litSize, istart+lhSize, litCSize,
                            dctx->HUFptr, flags);
                    }
                } else {
                    if (singleStream) {
#if defined(HUF_FORCE_DECOMPRESS_X2)
                        hufSuccess = HUF_decompress1X_DCtx_wksp(
                            dctx->entropy.hufTable, dctx->litBuffer, litSize,
                            istart+lhSize, litCSize, dctx->workspace,
                            sizeof(dctx->workspace), flags);
#else
                        hufSuccess = HUF_decompress1X1_DCtx_wksp(
                            dctx->entropy.hufTable, dctx->litBuffer, litSize,
                            istart+lhSize, litCSize, dctx->workspace,
                            sizeof(dctx->workspace), flags);
#endif
                    } else {
                        hufSuccess = HUF_decompress4X_hufOnly_wksp(
                            dctx->entropy.hufTable, dctx->litBuffer, litSize,
                            istart+lhSize, litCSize, dctx->workspace,
                            sizeof(dctx->workspace), flags);
                    }
                }
                if (dctx->litBufferLocation == ZSTD_split)
                {
                    assert(litSize > ZSTD_LITBUFFEREXTRASIZE);
                    ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
                    ZSTD_memmove(dctx->litBuffer + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH, dctx->litBuffer, litSize - ZSTD_LITBUFFEREXTRASIZE);
                    dctx->litBuffer += ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
                    dctx->litBufferEnd -= WILDCOPY_OVERLENGTH;
                    assert(dctx->litBufferEnd <= (BYTE*)dst + blockSizeMax);
                }

                RETURN_ERROR_IF(HUF_isError(hufSuccess), corruption_detected, "");

                dctx->litPtr = dctx->litBuffer;
                dctx->litSize = litSize;
                dctx->litEntropy = 1;
                if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
                return litCSize + lhSize;
            }

        case set_basic:
            {   size_t litSize, lhSize;
                U32 const lhlCode = ((istart[0]) >> 2) & 3;
                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
                switch(lhlCode)
                {
                case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
                    lhSize = 1;
                    litSize = istart[0] >> 3;
                    break;
                case 1:
                    lhSize = 2;
                    litSize = MEM_readLE16(istart) >> 4;
                    break;
                case 3:
                    lhSize = 3;
                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3");
                    litSize = MEM_readLE24(istart) >> 4;
                    break;
                }

                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
                RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
                if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) {  /* risk reading beyond src buffer with wildcopy */
                    RETURN_ERROR_IF(litSize+lhSize > srcSize, corruption_detected, "");
                    if (dctx->litBufferLocation == ZSTD_split)
                    {
                        ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize - ZSTD_LITBUFFEREXTRASIZE);
                        ZSTD_memcpy(dctx->litExtraBuffer, istart + lhSize + litSize - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
                    }
                    else
                    {
                        ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize);
                    }
                    dctx->litPtr = dctx->litBuffer;
                    dctx->litSize = litSize;
                    return lhSize+litSize;
                }
                /* direct reference into compressed stream */
                dctx->litPtr = istart+lhSize;
                dctx->litSize = litSize;
                dctx->litBufferEnd = dctx->litPtr + litSize;
                dctx->litBufferLocation = ZSTD_not_in_dst;
                return lhSize+litSize;
            }

        case set_rle:
            {   U32 const lhlCode = ((istart[0]) >> 2) & 3;
                size_t litSize, lhSize;
                size_t expectedWriteSize = MIN(blockSizeMax, dstCapacity);
                switch(lhlCode)
                {
                case 0: case 2: default:   /* note : default is impossible, since lhlCode into [0..3] */
                    lhSize = 1;
                    litSize = istart[0] >> 3;
                    break;
                case 1:
                    lhSize = 2;
                    RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3");
                    litSize = MEM_readLE16(istart) >> 4;
                    break;
                case 3:
                    lhSize = 3;
                    RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4");
                    litSize = MEM_readLE24(istart) >> 4;
                    break;
                }
                RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
                RETURN_ERROR_IF(litSize > blockSizeMax, corruption_detected, "");
                RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
                ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
                if (dctx->litBufferLocation == ZSTD_split)
                {
                    ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize - ZSTD_LITBUFFEREXTRASIZE);
                    ZSTD_memset(dctx->litExtraBuffer, istart[lhSize], ZSTD_LITBUFFEREXTRASIZE);
                }
                else
                {
                    ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize);
                }
                dctx->litPtr = dctx->litBuffer;
                dctx->litSize = litSize;
                return lhSize+1;
            }
        default:
            RETURN_ERROR(corruption_detected, "impossible");
        }
    }
}

/* Hidden declaration for fullbench */
size_t ZSTD_decodeLiteralsBlock_wrapper(ZSTD_DCtx* dctx,
                          const void* src, size_t srcSize,
                          void* dst, size_t dstCapacity);
size_t ZSTD_decodeLiteralsBlock_wrapper(ZSTD_DCtx* dctx,
                          const void* src, size_t srcSize,
                          void* dst, size_t dstCapacity)
{
    dctx->isFrameDecompression = 0;
    return ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, not_streaming);
}

/* Default FSE distribution tables.
 * These are pre-calculated FSE decoding tables using default distributions as defined in specification :
 * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#default-distributions
 * They were generated programmatically with following method :
 * - start from default distributions, present in /lib/common/zstd_internal.h
 * - generate tables normally, using ZSTD_buildFSETable()
 * - printout the content of tables
 * - prettify output, report below, test with fuzzer to ensure it's correct */

/* Default FSE distribution table for Literal Lengths */
static const ZSTD_seqSymbol LL_defaultDTable[(1<<LL_DEFAULTNORMLOG)+1] = {
     {  1,  1,  1, LL_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
     /* nextState, nbAddBits, nbBits, baseVal */
     {  0,  0,  4,    0},  { 16,  0,  4,    0},
     { 32,  0,  5,    1},  {  0,  0,  5,    3},
     {  0,  0,  5,    4},  {  0,  0,  5,    6},
     {  0,  0,  5,    7},  {  0,  0,  5,    9},
     {  0,  0,  5,   10},  {  0,  0,  5,   12},
     {  0,  0,  6,   14},  {  0,  1,  5,   16},
     {  0,  1,  5,   20},  {  0,  1,  5,   22},
     {  0,  2,  5,   28},  {  0,  3,  5,   32},
     {  0,  4,  5,   48},  { 32,  6,  5,   64},
     {  0,  7,  5,  128},  {  0,  8,  6,  256},
     {  0, 10,  6, 1024},  {  0, 12,  6, 4096},
     { 32,  0,  4,    0},  {  0,  0,  4,    1},
     {  0,  0,  5,    2},  { 32,  0,  5,    4},
     {  0,  0,  5,    5},  { 32,  0,  5,    7},
     {  0,  0,  5,    8},  { 32,  0,  5,   10},
     {  0,  0,  5,   11},  {  0,  0,  6,   13},
     { 32,  1,  5,   16},  {  0,  1,  5,   18},
     { 32,  1,  5,   22},  {  0,  2,  5,   24},
     { 32,  3,  5,   32},  {  0,  3,  5,   40},
     {  0,  6,  4,   64},  { 16,  6,  4,   64},
     { 32,  7,  5,  128},  {  0,  9,  6,  512},
     {  0, 11,  6, 2048},  { 48,  0,  4,    0},
     { 16,  0,  4,    1},  { 32,  0,  5,    2},
     { 32,  0,  5,    3},  { 32,  0,  5,    5},
     { 32,  0,  5,    6},  { 32,  0,  5,    8},
     { 32,  0,  5,    9},  { 32,  0,  5,   11},
     { 32,  0,  5,   12},  {  0,  0,  6,   15},
     { 32,  1,  5,   18},  { 32,  1,  5,   20},
     { 32,  2,  5,   24},  { 32,  2,  5,   28},
     { 32,  3,  5,   40},  { 32,  4,  5,   48},
     {  0, 16,  6,65536},  {  0, 15,  6,32768},
     {  0, 14,  6,16384},  {  0, 13,  6, 8192},
};   /* LL_defaultDTable */

/* Default FSE distribution table for Offset Codes */
static const ZSTD_seqSymbol OF_defaultDTable[(1<<OF_DEFAULTNORMLOG)+1] = {
    {  1,  1,  1, OF_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
    /* nextState, nbAddBits, nbBits, baseVal */
    {  0,  0,  5,    0},     {  0,  6,  4,   61},
    {  0,  9,  5,  509},     {  0, 15,  5,32765},
    {  0, 21,  5,2097149},   {  0,  3,  5,    5},
    {  0,  7,  4,  125},     {  0, 12,  5, 4093},
    {  0, 18,  5,262141},    {  0, 23,  5,8388605},
    {  0,  5,  5,   29},     {  0,  8,  4,  253},
    {  0, 14,  5,16381},     {  0, 20,  5,1048573},
    {  0,  2,  5,    1},     { 16,  7,  4,  125},
    {  0, 11,  5, 2045},     {  0, 17,  5,131069},
    {  0, 22,  5,4194301},   {  0,  4,  5,   13},
    { 16,  8,  4,  253},     {  0, 13,  5, 8189},
    {  0, 19,  5,524285},    {  0,  1,  5,    1},
    { 16,  6,  4,   61},     {  0, 10,  5, 1021},
    {  0, 16,  5,65533},     {  0, 28,  5,268435453},
    {  0, 27,  5,134217725}, {  0, 26,  5,67108861},
    {  0, 25,  5,33554429},  {  0, 24,  5,16777213},
};   /* OF_defaultDTable */


/* Default FSE distribution table for Match Lengths */
static const ZSTD_seqSymbol ML_defaultDTable[(1<<ML_DEFAULTNORMLOG)+1] = {
    {  1,  1,  1, ML_DEFAULTNORMLOG},  /* header : fastMode, tableLog */
    /* nextState, nbAddBits, nbBits, baseVal */
    {  0,  0,  6,    3},  {  0,  0,  4,    4},
    { 32,  0,  5,    5},  {  0,  0,  5,    6},
    {  0,  0,  5,    8},  {  0,  0,  5,    9},
    {  0,  0,  5,   11},  {  0,  0,  6,   13},
    {  0,  0,  6,   16},  {  0,  0,  6,   19},
    {  0,  0,  6,   22},  {  0,  0,  6,   25},
    {  0,  0,  6,   28},  {  0,  0,  6,   31},
    {  0,  0,  6,   34},  {  0,  1,  6,   37},
    {  0,  1,  6,   41},  {  0,  2,  6,   47},
    {  0,  3,  6,   59},  {  0,  4,  6,   83},
    {  0,  7,  6,  131},  {  0,  9,  6,  515},
    { 16,  0,  4,    4},  {  0,  0,  4,    5},
    { 32,  0,  5,    6},  {  0,  0,  5,    7},
    { 32,  0,  5,    9},  {  0,  0,  5,   10},
    {  0,  0,  6,   12},  {  0,  0,  6,   15},
    {  0,  0,  6,   18},  {  0,  0,  6,   21},
    {  0,  0,  6,   24},  {  0,  0,  6,   27},
    {  0,  0,  6,   30},  {  0,  0,  6,   33},
    {  0,  1,  6,   35},  {  0,  1,  6,   39},
    {  0,  2,  6,   43},  {  0,  3,  6,   51},
    {  0,  4,  6,   67},  {  0,  5,  6,   99},
    {  0,  8,  6,  259},  { 32,  0,  4,    4},
    { 48,  0,  4,    4},  { 16,  0,  4,    5},
    { 32,  0,  5,    7},  { 32,  0,  5,    8},
    { 32,  0,  5,   10},  { 32,  0,  5,   11},
    {  0,  0,  6,   14},  {  0,  0,  6,   17},
    {  0,  0,  6,   20},  {  0,  0,  6,   23},
    {  0,  0,  6,   26},  {  0,  0,  6,   29},
    {  0,  0,  6,   32},  {  0, 16,  6,65539},
    {  0, 15,  6,32771},  {  0, 14,  6,16387},
    {  0, 13,  6, 8195},  {  0, 12,  6, 4099},
    {  0, 11,  6, 2051},  {  0, 10,  6, 1027},
};   /* ML_defaultDTable */


static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U8 nbAddBits)
{
    void* ptr = dt;
    ZSTD_seqSymbol_header* const DTableH = (ZSTD_seqSymbol_header*)ptr;
    ZSTD_seqSymbol* const cell = dt + 1;

    DTableH->tableLog = 0;
    DTableH->fastMode = 0;

    cell->nbBits = 0;
    cell->nextState = 0;
    assert(nbAddBits < 255);
    cell->nbAdditionalBits = nbAddBits;
    cell->baseValue = baseValue;
}


/* ZSTD_buildFSETable() :
 * generate FSE decoding table for one symbol (ll, ml or off)
 * cannot fail if input is valid =>
 * all inputs are presumed validated at this stage */
FORCE_INLINE_TEMPLATE
void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
            const short* normalizedCounter, unsigned maxSymbolValue,
            const U32* baseValue, const U8* nbAdditionalBits,
            unsigned tableLog, void* wksp, size_t wkspSize)
{
    ZSTD_seqSymbol* const tableDecode = dt+1;
    U32 const maxSV1 = maxSymbolValue + 1;
    U32 const tableSize = 1 << tableLog;

    U16* symbolNext = (U16*)wksp;
    BYTE* spread = (BYTE*)(symbolNext + MaxSeq + 1);
    U32 highThreshold = tableSize - 1;


    /* Sanity Checks */
    assert(maxSymbolValue <= MaxSeq);
    assert(tableLog <= MaxFSELog);
    assert(wkspSize >= ZSTD_BUILD_FSE_TABLE_WKSP_SIZE);
    (void)wkspSize;
    /* Init, lay down lowprob symbols */
    {   ZSTD_seqSymbol_header DTableH;
        DTableH.tableLog = tableLog;
        DTableH.fastMode = 1;
        {   S16 const largeLimit= (S16)(1 << (tableLog-1));
            U32 s;
            for (s=0; s<maxSV1; s++) {
                if (normalizedCounter[s]==-1) {
                    tableDecode[highThreshold--].baseValue = s;
                    symbolNext[s] = 1;
                } else {
                    if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
                    assert(normalizedCounter[s]>=0);
                    symbolNext[s] = (U16)normalizedCounter[s];
        }   }   }
        ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
    }

    /* Spread symbols */
    assert(tableSize <= 512);
    /* Specialized symbol spreading for the case when there are
     * no low probability (-1 count) symbols. When compressing
     * small blocks we avoid low probability symbols to hit this
     * case, since header decoding speed matters more.
     */
    if (highThreshold == tableSize - 1) {
        size_t const tableMask = tableSize-1;
        size_t const step = FSE_TABLESTEP(tableSize);
        /* First lay down the symbols in order.
         * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
         * misses since small blocks generally have small table logs, so nearly
         * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
         * our buffer to handle the over-write.
         */
        {
            U64 const add = 0x0101010101010101ull;
            size_t pos = 0;
            U64 sv = 0;
            U32 s;
            for (s=0; s<maxSV1; ++s, sv += add) {
                int i;
                int const n = normalizedCounter[s];
                MEM_write64(spread + pos, sv);
                for (i = 8; i < n; i += 8) {
                    MEM_write64(spread + pos + i, sv);
                }
                assert(n>=0);
                pos += (size_t)n;
            }
        }
        /* Now we spread those positions across the table.
         * The benefit of doing it in two stages is that we avoid the
         * variable size inner loop, which caused lots of branch misses.
         * Now we can run through all the positions without any branch misses.
         * We unroll the loop twice, since that is what empirically worked best.
         */
        {
            size_t position = 0;
            size_t s;
            size_t const unroll = 2;
            assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
            for (s = 0; s < (size_t)tableSize; s += unroll) {
                size_t u;
                for (u = 0; u < unroll; ++u) {
                    size_t const uPosition = (position + (u * step)) & tableMask;
                    tableDecode[uPosition].baseValue = spread[s + u];
                }
                position = (position + (unroll * step)) & tableMask;
            }
            assert(position == 0);
        }
    } else {
        U32 const tableMask = tableSize-1;
        U32 const step = FSE_TABLESTEP(tableSize);
        U32 s, position = 0;
        for (s=0; s<maxSV1; s++) {
            int i;
            int const n = normalizedCounter[s];
            for (i=0; i<n; i++) {
                tableDecode[position].baseValue = s;
                position = (position + step) & tableMask;
                while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask;   /* lowprob area */
        }   }
        assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
    }

    /* Build Decoding table */
    {
        U32 u;
        for (u=0; u<tableSize; u++) {
            U32 const symbol = tableDecode[u].baseValue;
            U32 const nextState = symbolNext[symbol]++;
            tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
            tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
            assert(nbAdditionalBits[symbol] < 255);
            tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol];
            tableDecode[u].baseValue = baseValue[symbol];
        }
    }
}

/* Avoids the FORCE_INLINE of the _body() function. */
static void ZSTD_buildFSETable_body_default(ZSTD_seqSymbol* dt,
            const short* normalizedCounter, unsigned maxSymbolValue,
            const U32* baseValue, const U8* nbAdditionalBits,
            unsigned tableLog, void* wksp, size_t wkspSize)
{
    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
}

#if DYNAMIC_BMI2
BMI2_TARGET_ATTRIBUTE static void ZSTD_buildFSETable_body_bmi2(ZSTD_seqSymbol* dt,
            const short* normalizedCounter, unsigned maxSymbolValue,
            const U32* baseValue, const U8* nbAdditionalBits,
            unsigned tableLog, void* wksp, size_t wkspSize)
{
    ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
}
#endif

void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
            const short* normalizedCounter, unsigned maxSymbolValue,
            const U32* baseValue, const U8* nbAdditionalBits,
            unsigned tableLog, void* wksp, size_t wkspSize, int bmi2)
{
#if DYNAMIC_BMI2
    if (bmi2) {
        ZSTD_buildFSETable_body_bmi2(dt, normalizedCounter, maxSymbolValue,
                baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
        return;
    }
#endif
    (void)bmi2;
    ZSTD_buildFSETable_body_default(dt, normalizedCounter, maxSymbolValue,
            baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
}


/*! ZSTD_buildSeqTable() :
 * @return : nb bytes read from src,
 *           or an error code if it fails */
static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymbol** DTablePtr,
                                 SymbolEncodingType_e type, unsigned max, U32 maxLog,
                                 const void* src, size_t srcSize,
                                 const U32* baseValue, const U8* nbAdditionalBits,
                                 const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
                                 int ddictIsCold, int nbSeq, U32* wksp, size_t wkspSize,
                                 int bmi2)
{
    switch(type)
    {
    case set_rle :
        RETURN_ERROR_IF(!srcSize, srcSize_wrong, "");
        RETURN_ERROR_IF((*(const BYTE*)src) > max, corruption_detected, "");
        {   U32 const symbol = *(const BYTE*)src;
            U32 const baseline = baseValue[symbol];
            U8 const nbBits = nbAdditionalBits[symbol];
            ZSTD_buildSeqTable_rle(DTableSpace, baseline, nbBits);
        }
        *DTablePtr = DTableSpace;
        return 1;
    case set_basic :
        *DTablePtr = defaultTable;
        return 0;
    case set_repeat:
        RETURN_ERROR_IF(!flagRepeatTable, corruption_detected, "");
        /* prefetch FSE table if used */
        if (ddictIsCold && (nbSeq > 24 /* heuristic */)) {
            const void* const pStart = *DTablePtr;
            size_t const pSize = sizeof(ZSTD_seqSymbol) * (SEQSYMBOL_TABLE_SIZE(maxLog));
            PREFETCH_AREA(pStart, pSize);
        }
        return 0;
    case set_compressed :
        {   unsigned tableLog;
            S16 norm[MaxSeq+1];
            size_t const headerSize = FSE_readNCount(norm, &max, &tableLog, src, srcSize);
            RETURN_ERROR_IF(FSE_isError(headerSize), corruption_detected, "");
            RETURN_ERROR_IF(tableLog > maxLog, corruption_detected, "");
            ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog, wksp, wkspSize, bmi2);
            *DTablePtr = DTableSpace;
            return headerSize;
        }
    default :
        assert(0);
        RETURN_ERROR(GENERIC, "impossible");
    }
}

size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
                             const void* src, size_t srcSize)
{
    const BYTE* const istart = (const BYTE*)src;
    const BYTE* const iend = istart + srcSize;
    const BYTE* ip = istart;
    int nbSeq;
    DEBUGLOG(5, "ZSTD_decodeSeqHeaders");

    /* check */
    RETURN_ERROR_IF(srcSize < MIN_SEQUENCES_SIZE, srcSize_wrong, "");

    /* SeqHead */
    nbSeq = *ip++;
    if (nbSeq > 0x7F) {
        if (nbSeq == 0xFF) {
            RETURN_ERROR_IF(ip+2 > iend, srcSize_wrong, "");
            nbSeq = MEM_readLE16(ip) + LONGNBSEQ;
            ip+=2;
        } else {
            RETURN_ERROR_IF(ip >= iend, srcSize_wrong, "");
            nbSeq = ((nbSeq-0x80)<<8) + *ip++;
        }
    }
    *nbSeqPtr = nbSeq;

    if (nbSeq == 0) {
        /* No sequence : section ends immediately */
        RETURN_ERROR_IF(ip != iend, corruption_detected,
            "extraneous data present in the Sequences section");
        return (size_t)(ip - istart);
    }

    /* FSE table descriptors */
    RETURN_ERROR_IF(ip+1 > iend, srcSize_wrong, ""); /* minimum possible size: 1 byte for symbol encoding types */
    RETURN_ERROR_IF(*ip & 3, corruption_detected, ""); /* The last field, Reserved, must be all-zeroes. */
    {   SymbolEncodingType_e const LLtype = (SymbolEncodingType_e)(*ip >> 6);
        SymbolEncodingType_e const OFtype = (SymbolEncodingType_e)((*ip >> 4) & 3);
        SymbolEncodingType_e const MLtype = (SymbolEncodingType_e)((*ip >> 2) & 3);
        ip++;

        /* Build DTables */
        {   size_t const llhSize = ZSTD_buildSeqTable(dctx->entropy.LLTable, &dctx->LLTptr,
                                                      LLtype, MaxLL, LLFSELog,
                                                      ip, iend-ip,
                                                      LL_base, LL_bits,
                                                      LL_defaultDTable, dctx->fseEntropy,
                                                      dctx->ddictIsCold, nbSeq,
                                                      dctx->workspace, sizeof(dctx->workspace),
                                                      ZSTD_DCtx_get_bmi2(dctx));
            RETURN_ERROR_IF(ZSTD_isError(llhSize), corruption_detected, "ZSTD_buildSeqTable failed");
            ip += llhSize;
        }

        {   size_t const ofhSize = ZSTD_buildSeqTable(dctx->entropy.OFTable, &dctx->OFTptr,
                                                      OFtype, MaxOff, OffFSELog,
                                                      ip, iend-ip,
                                                      OF_base, OF_bits,
                                                      OF_defaultDTable, dctx->fseEntropy,
                                                      dctx->ddictIsCold, nbSeq,
                                                      dctx->workspace, sizeof(dctx->workspace),
                                                      ZSTD_DCtx_get_bmi2(dctx));
            RETURN_ERROR_IF(ZSTD_isError(ofhSize), corruption_detected, "ZSTD_buildSeqTable failed");
            ip += ofhSize;
        }

        {   size_t const mlhSize = ZSTD_buildSeqTable(dctx->entropy.MLTable, &dctx->MLTptr,
                                                      MLtype, MaxML, MLFSELog,
                                                      ip, iend-ip,
                                                      ML_base, ML_bits,
                                                      ML_defaultDTable, dctx->fseEntropy,
                                                      dctx->ddictIsCold, nbSeq,
                                                      dctx->workspace, sizeof(dctx->workspace),
                                                      ZSTD_DCtx_get_bmi2(dctx));
            RETURN_ERROR_IF(ZSTD_isError(mlhSize), corruption_detected, "ZSTD_buildSeqTable failed");
            ip += mlhSize;
        }
    }

    return ip-istart;
}


typedef struct {
    size_t litLength;
    size_t matchLength;
    size_t offset;
} seq_t;

typedef struct {
    size_t state;
    const ZSTD_seqSymbol* table;
} ZSTD_fseState;

typedef struct {
    BIT_DStream_t DStream;
    ZSTD_fseState stateLL;
    ZSTD_fseState stateOffb;
    ZSTD_fseState stateML;
    size_t prevOffset[ZSTD_REP_NUM];
} seqState_t;

/*! ZSTD_overlapCopy8() :
 *  Copies 8 bytes from ip to op and updates op and ip where ip <= op.
 *  If the offset is < 8 then the offset is spread to at least 8 bytes.
 *
 *  Precondition: *ip <= *op
 *  Postcondition: *op - *op >= 8
 */
HINT_INLINE void ZSTD_overlapCopy8(BYTE** op, BYTE const** ip, size_t offset) {
    assert(*ip <= *op);
    if (offset < 8) {
        /* close range match, overlap */
        static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 };   /* added */
        static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 };   /* subtracted */
        int const sub2 = dec64table[offset];
        (*op)[0] = (*ip)[0];
        (*op)[1] = (*ip)[1];
        (*op)[2] = (*ip)[2];
        (*op)[3] = (*ip)[3];
        *ip += dec32table[offset];
        ZSTD_copy4(*op+4, *ip);
        *ip -= sub2;
    } else {
        ZSTD_copy8(*op, *ip);
    }
    *ip += 8;
    *op += 8;
    assert(*op - *ip >= 8);
}

/*! ZSTD_safecopy() :
 *  Specialized version of memcpy() that is allowed to READ up to WILDCOPY_OVERLENGTH past the input buffer
 *  and write up to 16 bytes past oend_w (op >= oend_w is allowed).
 *  This function is only called in the uncommon case where the sequence is near the end of the block. It
 *  should be fast for a single long sequence, but can be slow for several short sequences.
 *
 *  @param ovtype controls the overlap detection
 *         - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
 *         - ZSTD_overlap_src_before_dst: The src and dst may overlap and may be any distance apart.
 *           The src buffer must be before the dst buffer.
 */
static void ZSTD_safecopy(BYTE* op, const BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
    ptrdiff_t const diff = op - ip;
    BYTE* const oend = op + length;

    assert((ovtype == ZSTD_no_overlap && (diff <= -8 || diff >= 8 || op >= oend_w)) ||
           (ovtype == ZSTD_overlap_src_before_dst && diff >= 0));

    if (length < 8) {
        /* Handle short lengths. */
        while (op < oend) *op++ = *ip++;
        return;
    }
    if (ovtype == ZSTD_overlap_src_before_dst) {
        /* Copy 8 bytes and ensure the offset >= 8 when there can be overlap. */
        assert(length >= 8);
        ZSTD_overlapCopy8(&op, &ip, diff);
        length -= 8;
        assert(op - ip >= 8);
        assert(op <= oend);
    }

    if (oend <= oend_w) {
        /* No risk of overwrite. */
        ZSTD_wildcopy(op, ip, length, ovtype);
        return;
    }
    if (op <= oend_w) {
        /* Wildcopy until we get close to the end. */
        assert(oend > oend_w);
        ZSTD_wildcopy(op, ip, oend_w - op, ovtype);
        ip += oend_w - op;
        op += oend_w - op;
    }
    /* Handle the leftovers. */
    while (op < oend) *op++ = *ip++;
}

/* ZSTD_safecopyDstBeforeSrc():
 * This version allows overlap with dst before src, or handles the non-overlap case with dst after src
 * Kept separate from more common ZSTD_safecopy case to avoid performance impact to the safecopy common case */
static void ZSTD_safecopyDstBeforeSrc(BYTE* op, const BYTE* ip, ptrdiff_t length) {
    ptrdiff_t const diff = op - ip;
    BYTE* const oend = op + length;

    if (length < 8 || diff > -8) {
        /* Handle short lengths, close overlaps, and dst not before src. */
        while (op < oend) *op++ = *ip++;
        return;
    }

    if (op <= oend - WILDCOPY_OVERLENGTH && diff < -WILDCOPY_VECLEN) {
        ZSTD_wildcopy(op, ip, oend - WILDCOPY_OVERLENGTH - op, ZSTD_no_overlap);
        ip += oend - WILDCOPY_OVERLENGTH - op;
        op += oend - WILDCOPY_OVERLENGTH - op;
    }

    /* Handle the leftovers. */
    while (op < oend) *op++ = *ip++;
}

/* ZSTD_execSequenceEnd():
 * This version handles cases that are near the end of the output buffer. It requires
 * more careful checks to make sure there is no overflow. By separating out these hard
 * and unlikely cases, we can speed up the common cases.
 *
 * NOTE: This function needs to be fast for a single long sequence, but doesn't need
 * to be optimized for many small sequences, since those fall into ZSTD_execSequence().
 */
FORCE_NOINLINE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_execSequenceEnd(BYTE* op,
    BYTE* const oend, seq_t sequence,
    const BYTE** litPtr, const BYTE* const litLimit,
    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
{
    BYTE* const oLitEnd = op + sequence.litLength;
    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
    const BYTE* match = oLitEnd - sequence.offset;
    BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;

    /* bounds checks : careful of address space overflow in 32-bit mode */
    RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
    RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
    assert(op < op + sequenceLength);
    assert(oLitEnd < op + sequenceLength);

    /* copy literals */
    ZSTD_safecopy(op, oend_w, *litPtr, sequence.litLength, ZSTD_no_overlap);
    op = oLitEnd;
    *litPtr = iLitEnd;

    /* copy Match */
    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
        /* offset beyond prefix */
        RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
        match = dictEnd - (prefixStart - match);
        if (match + sequence.matchLength <= dictEnd) {
            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
            return sequenceLength;
        }
        /* span extDict & currentPrefixSegment */
        {   size_t const length1 = dictEnd - match;
        ZSTD_memmove(oLitEnd, match, length1);
        op = oLitEnd + length1;
        sequence.matchLength -= length1;
        match = prefixStart;
        }
    }
    ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
    return sequenceLength;
}

/* ZSTD_execSequenceEndSplitLitBuffer():
 * This version is intended to be used during instances where the litBuffer is still split.  It is kept separate to avoid performance impact for the good case.
 */
FORCE_NOINLINE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_execSequenceEndSplitLitBuffer(BYTE* op,
    BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
    const BYTE** litPtr, const BYTE* const litLimit,
    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
{
    BYTE* const oLitEnd = op + sequence.litLength;
    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
    const BYTE* match = oLitEnd - sequence.offset;


    /* bounds checks : careful of address space overflow in 32-bit mode */
    RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
    RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
    assert(op < op + sequenceLength);
    assert(oLitEnd < op + sequenceLength);

    /* copy literals */
    RETURN_ERROR_IF(op > *litPtr && op < *litPtr + sequence.litLength, dstSize_tooSmall, "output should not catch up to and overwrite literal buffer");
    ZSTD_safecopyDstBeforeSrc(op, *litPtr, sequence.litLength);
    op = oLitEnd;
    *litPtr = iLitEnd;

    /* copy Match */
    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
        /* offset beyond prefix */
        RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
        match = dictEnd - (prefixStart - match);
        if (match + sequence.matchLength <= dictEnd) {
            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
            return sequenceLength;
        }
        /* span extDict & currentPrefixSegment */
        {   size_t const length1 = dictEnd - match;
        ZSTD_memmove(oLitEnd, match, length1);
        op = oLitEnd + length1;
        sequence.matchLength -= length1;
        match = prefixStart;
        }
    }
    ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
    return sequenceLength;
}

HINT_INLINE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_execSequence(BYTE* op,
    BYTE* const oend, seq_t sequence,
    const BYTE** litPtr, const BYTE* const litLimit,
    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
{
    BYTE* const oLitEnd = op + sequence.litLength;
    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
    BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;   /* risk : address space underflow on oend=NULL */
    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
    const BYTE* match = oLitEnd - sequence.offset;

    assert(op != NULL /* Precondition */);
    assert(oend_w < oend /* No underflow */);

#if defined(__aarch64__)
    /* prefetch sequence starting from match that will be used for copy later */
    PREFETCH_L1(match);
#endif
    /* Handle edge cases in a slow path:
     *   - Read beyond end of literals
     *   - Match end is within WILDCOPY_OVERLIMIT of oend
     *   - 32-bit mode and the match length overflows
     */
    if (UNLIKELY(
        iLitEnd > litLimit ||
        oMatchEnd > oend_w ||
        (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
        return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);

    /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
    assert(op <= oLitEnd /* No overflow */);
    assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
    assert(oMatchEnd <= oend /* No underflow */);
    assert(iLitEnd <= litLimit /* Literal length is in bounds */);
    assert(oLitEnd <= oend_w /* Can wildcopy literals */);
    assert(oMatchEnd <= oend_w /* Can wildcopy matches */);

    /* Copy Literals:
     * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
     * We likely don't need the full 32-byte wildcopy.
     */
    assert(WILDCOPY_OVERLENGTH >= 16);
    ZSTD_copy16(op, (*litPtr));
    if (UNLIKELY(sequence.litLength > 16)) {
        ZSTD_wildcopy(op + 16, (*litPtr) + 16, sequence.litLength - 16, ZSTD_no_overlap);
    }
    op = oLitEnd;
    *litPtr = iLitEnd;   /* update for next sequence */

    /* Copy Match */
    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
        /* offset beyond prefix -> go into extDict */
        RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
        match = dictEnd + (match - prefixStart);
        if (match + sequence.matchLength <= dictEnd) {
            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
            return sequenceLength;
        }
        /* span extDict & currentPrefixSegment */
        {   size_t const length1 = dictEnd - match;
        ZSTD_memmove(oLitEnd, match, length1);
        op = oLitEnd + length1;
        sequence.matchLength -= length1;
        match = prefixStart;
        }
    }
    /* Match within prefix of 1 or more bytes */
    assert(op <= oMatchEnd);
    assert(oMatchEnd <= oend_w);
    assert(match >= prefixStart);
    assert(sequence.matchLength >= 1);

    /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
     * without overlap checking.
     */
    if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
        /* We bet on a full wildcopy for matches, since we expect matches to be
         * longer than literals (in general). In silesia, ~10% of matches are longer
         * than 16 bytes.
         */
        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
        return sequenceLength;
    }
    assert(sequence.offset < WILDCOPY_VECLEN);

    /* Copy 8 bytes and spread the offset to be >= 8. */
    ZSTD_overlapCopy8(&op, &match, sequence.offset);

    /* If the match length is > 8 bytes, then continue with the wildcopy. */
    if (sequence.matchLength > 8) {
        assert(op < oMatchEnd);
        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength - 8, ZSTD_overlap_src_before_dst);
    }
    return sequenceLength;
}

HINT_INLINE
ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
size_t ZSTD_execSequenceSplitLitBuffer(BYTE* op,
    BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
    const BYTE** litPtr, const BYTE* const litLimit,
    const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
{
    BYTE* const oLitEnd = op + sequence.litLength;
    size_t const sequenceLength = sequence.litLength + sequence.matchLength;
    BYTE* const oMatchEnd = op + sequenceLength;   /* risk : address space overflow (32-bits) */
    const BYTE* const iLitEnd = *litPtr + sequence.litLength;
    const BYTE* match = oLitEnd - sequence.offset;

    assert(op != NULL /* Precondition */);
    assert(oend_w < oend /* No underflow */);
    /* Handle edge cases in a slow path:
     *   - Read beyond end of literals
     *   - Match end is within WILDCOPY_OVERLIMIT of oend
     *   - 32-bit mode and the match length overflows
     */
    if (UNLIKELY(
            iLitEnd > litLimit ||
            oMatchEnd > oend_w ||
            (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
        return ZSTD_execSequenceEndSplitLitBuffer(op, oend, oend_w, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);

    /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
    assert(op <= oLitEnd /* No overflow */);
    assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
    assert(oMatchEnd <= oend /* No underflow */);
    assert(iLitEnd <= litLimit /* Literal length is in bounds */);
    assert(oLitEnd <= oend_w /* Can wildcopy literals */);
    assert(oMatchEnd <= oend_w /* Can wildcopy matches */);

    /* Copy Literals:
     * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
     * We likely don't need the full 32-byte wildcopy.
     */
    assert(WILDCOPY_OVERLENGTH >= 16);
    ZSTD_copy16(op, (*litPtr));
    if (UNLIKELY(sequence.litLength > 16)) {
        ZSTD_wildcopy(op+16, (*litPtr)+16, sequence.litLength-16, ZSTD_no_overlap);
    }
    op = oLitEnd;
    *litPtr = iLitEnd;   /* update for next sequence */

    /* Copy Match */
    if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
        /* offset beyond prefix -> go into extDict */
        RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
        match = dictEnd + (match - prefixStart);
        if (match + sequence.matchLength <= dictEnd) {
            ZSTD_memmove(oLitEnd, match, sequence.matchLength);
            return sequenceLength;
        }
        /* span extDict & currentPrefixSegment */
        {   size_t const length1 = dictEnd - match;
            ZSTD_memmove(oLitEnd, match, length1);
            op = oLitEnd + length1;
            sequence.matchLength -= length1;
            match = prefixStart;
    }   }
    /* Match within prefix of 1 or more bytes */
    assert(op <= oMatchEnd);
    assert(oMatchEnd <= oend_w);
    assert(match >= prefixStart);
    assert(sequence.matchLength >= 1);

    /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
     * without overlap checking.
     */
    if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
        /* We bet on a full wildcopy for matches, since we expect matches to be
         * longer than literals (in general). In silesia, ~10% of matches are longer
         * than 16 bytes.
         */
        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
        return sequenceLength;
    }
    assert(sequence.offset < WILDCOPY_VECLEN);

    /* Copy 8 bytes and spread the offset to be >= 8. */
    ZSTD_overlapCopy8(&op, &match, sequence.offset);

    /* If the match length is > 8 bytes, then continue with the wildcopy. */
    if (sequence.matchLength > 8) {
        assert(op < oMatchEnd);
        ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8, ZSTD_overlap_src_before_dst);
    }
    return sequenceLength;
}


static void
ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqSymbol* dt)
{
    const void* ptr = dt;
    const ZSTD_seqSymbol_header* const DTableH = (const ZSTD_seqSymbol_header*)ptr;
    DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
    DEBUGLOG(6, "ZSTD_initFseState : val=%u using %u bits",
                (U32)DStatePtr->state, DTableH->tableLog);
    BIT_reloadDStream(bitD);
    DStatePtr->table = dt + 1;
}

FORCE_INLINE_TEMPLATE void
ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16 nextState, U32 nbBits)
{
    size_t const lowBits = BIT_readBits(bitD, nbBits);
    DStatePtr->state = nextState + lowBits;
}

/* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
 * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32
 * bits before reloading. This value is the maximum number of bytes we read
 * after reloading when we are decoding long offsets.
 */
#define LONG_OFFSETS_MAX_EXTRA_BITS_32                       \
    (ZSTD_WINDOWLOG_MAX_32 > STREAM_ACCUMULATOR_MIN_32       \
        ? ZSTD_WINDOWLOG_MAX_32 - STREAM_ACCUMULATOR_MIN_32  \
        : 0)

typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;

/**
 * ZSTD_decodeSequence():
 * @p longOffsets : tells the decoder to reload more bit while decoding large offsets
 *                  only used in 32-bit mode
 * @return : Sequence (litL + matchL + offset)
 */
FORCE_INLINE_TEMPLATE seq_t
ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets, const int isLastSeq)
{
    seq_t seq;
    /*
     * ZSTD_seqSymbol is a 64 bits wide structure.
     * It can be loaded in one operation
     * and its fields extracted by simply shifting or bit-extracting on aarch64.
     * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh
     * operations that cause performance drop. This can be avoided by using this
     * ZSTD_memcpy hack.
     */
#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__))
    ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS;
    ZSTD_seqSymbol* const llDInfo = &llDInfoS;
    ZSTD_seqSymbol* const mlDInfo = &mlDInfoS;
    ZSTD_seqSymbol* const ofDInfo = &ofDInfoS;
    ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol));
    ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol));
    ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol));
#else
    const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
    const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
    const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
#endif
    seq.matchLength = mlDInfo->baseValue;
    seq.litLength = llDInfo->baseValue;
    {   U32 const ofBase = ofDInfo->baseValue;
        BYTE const llBits = llDInfo->nbAdditionalBits;
        BYTE const mlBits = mlDInfo->nbAdditionalBits;
        BYTE const ofBits = ofDInfo->nbAdditionalBits;
        BYTE const totalBits = llBits+mlBits+ofBits;

        U16 const llNext = llDInfo->nextState;
        U16 const mlNext = mlDInfo->nextState;
        U16 const ofNext = ofDInfo->nextState;
        U32 const llnbBits = llDInfo->nbBits;
        U32 const mlnbBits = mlDInfo->nbBits;
        U32 const ofnbBits = ofDInfo->nbBits;

        assert(llBits <= MaxLLBits);
        assert(mlBits <= MaxMLBits);
        assert(ofBits <= MaxOff);
        /*
         * As gcc has better branch and block analyzers, sometimes it is only
         * valuable to mark likeliness for clang, it gives around 3-4% of
         * performance.
         */

        /* sequence */
        {   size_t offset;
            if (ofBits > 1) {
                ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
                ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32);
                ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits);
                if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
                    /* Always read extra bits, this keeps the logic simple,
                     * avoids branches, and avoids accidentally reading 0 bits.
                     */
                    U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32;
                    offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
                    BIT_reloadDStream(&seqState->DStream);
                    offset += BIT_readBitsFast(&seqState->DStream, extraBits);
                } else {
                    offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/);   /* <=  (ZSTD_WINDOWLOG_MAX-1) bits */
                    if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
                }
                seqState->prevOffset[2] = seqState->prevOffset[1];
                seqState->prevOffset[1] = seqState->prevOffset[0];
                seqState->prevOffset[0] = offset;
            } else {
                U32 const ll0 = (llDInfo->baseValue == 0);
                if (LIKELY((ofBits == 0))) {
                    offset = seqState->prevOffset[ll0];
                    seqState->prevOffset[1] = seqState->prevOffset[!ll0];
                    seqState->prevOffset[0] = offset;
                } else {
                    offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
                    {   size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
                        temp -= !temp; /* 0 is not valid: input corrupted => force offset to -1 => corruption detected at execSequence */
                        if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
                        seqState->prevOffset[1] = seqState->prevOffset[0];
                        seqState->prevOffset[0] = offset = temp;
            }   }   }
            seq.offset = offset;
        }

        if (mlBits > 0)
            seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);

        if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
            BIT_reloadDStream(&seqState->DStream);
        if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
            BIT_reloadDStream(&seqState->DStream);
        /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
        ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);

        if (llBits > 0)
            seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);

        if (MEM_32bits())
            BIT_reloadDStream(&seqState->DStream);

        DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
                    (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);

        if (!isLastSeq) {
            /* don't update FSE state for last Sequence */
            ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits);    /* <=  9 bits */
            ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits);    /* <=  9 bits */
            if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);    /* <= 18 bits */
            ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits);  /* <=  8 bits */
            BIT_reloadDStream(&seqState->DStream);
        }
    }

    return seq;
}

#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
#if DEBUGLEVEL >= 1
static int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
{
    size_t const windowSize = dctx->fParams.windowSize;
    /* No dictionary used. */
    if (dctx->dictContentEndForFuzzing == NULL) return 0;
    /* Dictionary is our prefix. */
    if (prefixStart == dctx->dictContentBeginForFuzzing) return 1;
    /* Dictionary is not our ext-dict. */
    if (dctx->dictEnd != dctx->dictContentEndForFuzzing) return 0;
    /* Dictionary is not within our window size. */
    if ((size_t)(oLitEnd - prefixStart) >= windowSize) return 0;
    /* Dictionary is active. */
    return 1;
}
#endif

static void ZSTD_assertValidSequence(
        ZSTD_DCtx const* dctx,
        BYTE const* op, BYTE const* oend,
        seq_t const seq,
        BYTE const* prefixStart, BYTE const* virtualStart)
{
#if DEBUGLEVEL >= 1
    if (dctx->isFrameDecompression) {
        size_t const windowSize = dctx->fParams.windowSize;
        size_t const sequenceSize = seq.litLength + seq.matchLength;
        BYTE const* const oLitEnd = op + seq.litLength;
        DEBUGLOG(6, "Checking sequence: litL=%u matchL=%u offset=%u",
                (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
        assert(op <= oend);
        assert((size_t)(oend - op) >= sequenceSize);
        assert(sequenceSize <= ZSTD_blockSizeMax(dctx));
        if (ZSTD_dictionaryIsActive(dctx, prefixStart, oLitEnd)) {
            size_t const dictSize = (size_t)((char const*)dctx->dictContentEndForFuzzing - (char const*)dctx->dictContentBeginForFuzzing);
            /* Offset must be within the dictionary. */
            assert(seq.offset <= (size_t)(oLitEnd - virtualStart));
            assert(seq.offset <= windowSize + dictSize);
        } else {
            /* Offset must be within our window. */
            assert(seq.offset <= windowSize);
        }
    }
#else
    (void)dctx, (void)op, (void)oend, (void)seq, (void)prefixStart, (void)virtualStart;
#endif
}
#endif

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG


FORCE_INLINE_TEMPLATE size_t
DONT_VECTORIZE
ZSTD_decompressSequences_bodySplitLitBuffer( ZSTD_DCtx* dctx,
                               void* dst, size_t maxDstSize,
                         const void* seqStart, size_t seqSize, int nbSeq,
                         const ZSTD_longOffset_e isLongOffset)
{
    const BYTE* ip = (const BYTE*)seqStart;
    const BYTE* const iend = ip + seqSize;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = ZSTD_maybeNullPtrAdd(ostart, maxDstSize);
    BYTE* op = ostart;
    const BYTE* litPtr = dctx->litPtr;
    const BYTE* litBufferEnd = dctx->litBufferEnd;
    const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
    const BYTE* const vBase = (const BYTE*) (dctx->virtualStart);
    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
    DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer (%i seqs)", nbSeq);

    /* Literals are split between internal buffer & output buffer */
    if (nbSeq) {
        seqState_t seqState;
        dctx->fseEntropy = 1;
        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
        RETURN_ERROR_IF(
            ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend-ip)),
            corruption_detected, "");
        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
        assert(dst != NULL);

        ZSTD_STATIC_ASSERT(
                BIT_DStream_unfinished < BIT_DStream_completed &&
                BIT_DStream_endOfBuffer < BIT_DStream_completed &&
                BIT_DStream_completed < BIT_DStream_overflow);

        /* decompress without overrunning litPtr begins */
        {   seq_t sequence = {0,0,0};  /* some static analyzer believe that @sequence is not initialized (it necessarily is, since for(;;) loop as at least one iteration) */
            /* Align the decompression loop to 32 + 16 bytes.
                *
                * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
                * speed swings based on the alignment of the decompression loop. This
                * performance swing is caused by parts of the decompression loop falling
                * out of the DSB. The entire decompression loop should fit in the DSB,
                * when it can't we get much worse performance. You can measure if you've
                * hit the good case or the bad case with this perf command for some
                * compressed file test.zst:
                *
                *   perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
                *             -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
                *
                * If you see most cycles served out of the MITE you've hit the bad case.
                * If you see most cycles served out of the DSB you've hit the good case.
                * If it is pretty even then you may be in an okay case.
                *
                * This issue has been reproduced on the following CPUs:
                *   - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
                *               Use Instruments->Counters to get DSB/MITE cycles.
                *               I never got performance swings, but I was able to
                *               go from the good case of mostly DSB to half of the
                *               cycles served from MITE.
                *   - Coffeelake: Intel i9-9900k
                *   - Coffeelake: Intel i7-9700k
                *
                * I haven't been able to reproduce the instability or DSB misses on any
                * of the following CPUS:
                *   - Haswell
                *   - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
                *   - Skylake
                *
                * Alignment is done for each of the three major decompression loops:
                *   - ZSTD_decompressSequences_bodySplitLitBuffer - presplit section of the literal buffer
                *   - ZSTD_decompressSequences_bodySplitLitBuffer - postsplit section of the literal buffer
                *   - ZSTD_decompressSequences_body
                * Alignment choices are made to minimize large swings on bad cases and influence on performance
                * from changes external to this code, rather than to overoptimize on the current commit.
                *
                * If you are seeing performance stability this script can help test.
                * It tests on 4 commits in zstd where I saw performance change.
                *
                *   https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
                */
#if defined(__GNUC__) && defined(__x86_64__)
            __asm__(".p2align 6");
#  if __GNUC__ >= 7
	    /* good for gcc-7, gcc-9, and gcc-11 */
            __asm__("nop");
            __asm__(".p2align 5");
            __asm__("nop");
            __asm__(".p2align 4");
#    if __GNUC__ == 8 || __GNUC__ == 10
	    /* good for gcc-8 and gcc-10 */
            __asm__("nop");
            __asm__(".p2align 3");
#    endif
#  endif
#endif

            /* Handle the initial state where litBuffer is currently split between dst and litExtraBuffer */
            for ( ; nbSeq; nbSeq--) {
                sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
                if (litPtr + sequence.litLength > dctx->litBufferEnd) break;
                {   size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                    assert(!ZSTD_isError(oneSeqSize));
                    ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
#endif
                    if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                        return oneSeqSize;
                    DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
                    op += oneSeqSize;
            }   }
            DEBUGLOG(6, "reached: (litPtr + sequence.litLength > dctx->litBufferEnd)");

            /* If there are more sequences, they will need to read literals from litExtraBuffer; copy over the remainder from dst and update litPtr and litEnd */
            if (nbSeq > 0) {
                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
                DEBUGLOG(6, "There are %i sequences left, and %zu/%zu literals left in buffer", nbSeq, leftoverLit, sequence.litLength);
                if (leftoverLit) {
                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
                    sequence.litLength -= leftoverLit;
                    op += leftoverLit;
                }
                litPtr = dctx->litExtraBuffer;
                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                dctx->litBufferLocation = ZSTD_not_in_dst;
                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                    assert(!ZSTD_isError(oneSeqSize));
                    ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
#endif
                    if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                        return oneSeqSize;
                    DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
                    op += oneSeqSize;
                }
                nbSeq--;
            }
        }

        if (nbSeq > 0) {
            /* there is remaining lit from extra buffer */

#if defined(__GNUC__) && defined(__x86_64__)
            __asm__(".p2align 6");
            __asm__("nop");
#  if __GNUC__ != 7
            /* worse for gcc-7 better for gcc-8, gcc-9, and gcc-10 and clang */
            __asm__(".p2align 4");
            __asm__("nop");
            __asm__(".p2align 3");
#  elif __GNUC__ >= 11
            __asm__(".p2align 3");
#  else
            __asm__(".p2align 5");
            __asm__("nop");
            __asm__(".p2align 3");
#  endif
#endif

            for ( ; nbSeq ; nbSeq--) {
                seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
                size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                assert(!ZSTD_isError(oneSeqSize));
                ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
#endif
                if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                    return oneSeqSize;
                DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
                op += oneSeqSize;
            }
        }

        /* check if reached exact end */
        DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer: after decode loop, remaining nbSeq : %i", nbSeq);
        RETURN_ERROR_IF(nbSeq, corruption_detected, "");
        DEBUGLOG(5, "bitStream : start=%p, ptr=%p, bitsConsumed=%u", seqState.DStream.start, seqState.DStream.ptr, seqState.DStream.bitsConsumed);
        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");
        /* save reps for next block */
        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
    }

    /* last literal segment */
    if (dctx->litBufferLocation == ZSTD_split) {
        /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
        size_t const lastLLSize = (size_t)(litBufferEnd - litPtr);
        DEBUGLOG(6, "copy last literals from segment : %u", (U32)lastLLSize);
        RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
        if (op != NULL) {
            ZSTD_memmove(op, litPtr, lastLLSize);
            op += lastLLSize;
        }
        litPtr = dctx->litExtraBuffer;
        litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
        dctx->litBufferLocation = ZSTD_not_in_dst;
    }
    /* copy last literals from internal buffer */
    {   size_t const lastLLSize = (size_t)(litBufferEnd - litPtr);
        DEBUGLOG(6, "copy last literals from internal buffer : %u", (U32)lastLLSize);
        RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
        if (op != NULL) {
            ZSTD_memcpy(op, litPtr, lastLLSize);
            op += lastLLSize;
    }   }

    DEBUGLOG(6, "decoded block of size %u bytes", (U32)(op - ostart));
    return (size_t)(op - ostart);
}

FORCE_INLINE_TEMPLATE size_t
DONT_VECTORIZE
ZSTD_decompressSequences_body(ZSTD_DCtx* dctx,
    void* dst, size_t maxDstSize,
    const void* seqStart, size_t seqSize, int nbSeq,
    const ZSTD_longOffset_e isLongOffset)
{
    const BYTE* ip = (const BYTE*)seqStart;
    const BYTE* const iend = ip + seqSize;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ZSTD_maybeNullPtrAdd(ostart, maxDstSize) : dctx->litBuffer;
    BYTE* op = ostart;
    const BYTE* litPtr = dctx->litPtr;
    const BYTE* const litEnd = litPtr + dctx->litSize;
    const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
    const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
    const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
    DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq);

    /* Regen sequences */
    if (nbSeq) {
        seqState_t seqState;
        dctx->fseEntropy = 1;
        { U32 i; for (i = 0; i < ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
        RETURN_ERROR_IF(
            ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend - ip)),
            corruption_detected, "");
        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
        assert(dst != NULL);

#if defined(__GNUC__) && defined(__x86_64__)
            __asm__(".p2align 6");
            __asm__("nop");
#  if __GNUC__ >= 7
            __asm__(".p2align 5");
            __asm__("nop");
            __asm__(".p2align 3");
#  else
            __asm__(".p2align 4");
            __asm__("nop");
            __asm__(".p2align 3");
#  endif
#endif

        for ( ; nbSeq ; nbSeq--) {
            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, nbSeq==1);
            size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
            assert(!ZSTD_isError(oneSeqSize));
            ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
#endif
            if (UNLIKELY(ZSTD_isError(oneSeqSize)))
                return oneSeqSize;
            DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
            op += oneSeqSize;
        }

        /* check if reached exact end */
        assert(nbSeq == 0);
        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");
        /* save reps for next block */
        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
    }

    /* last literal segment */
    {   size_t const lastLLSize = (size_t)(litEnd - litPtr);
        DEBUGLOG(6, "copy last literals : %u", (U32)lastLLSize);
        RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
        if (op != NULL) {
            ZSTD_memcpy(op, litPtr, lastLLSize);
            op += lastLLSize;
    }   }

    DEBUGLOG(6, "decoded block of size %u bytes", (U32)(op - ostart));
    return (size_t)(op - ostart);
}

static size_t
ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
                                 void* dst, size_t maxDstSize,
                           const void* seqStart, size_t seqSize, int nbSeq,
                           const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}

static size_t
ZSTD_decompressSequencesSplitLitBuffer_default(ZSTD_DCtx* dctx,
                                               void* dst, size_t maxDstSize,
                                         const void* seqStart, size_t seqSize, int nbSeq,
                                         const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT

FORCE_INLINE_TEMPLATE

size_t ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
                   const BYTE* const prefixStart, const BYTE* const dictEnd)
{
    prefetchPos += sequence.litLength;
    {   const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
        /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
         * No consequence though : memory address is only used for prefetching, not for dereferencing */
        const BYTE* const match = ZSTD_wrappedPtrSub(ZSTD_wrappedPtrAdd(matchBase, prefetchPos), sequence.offset);
        PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE);   /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
    }
    return prefetchPos + sequence.matchLength;
}

/* This decoding function employs prefetching
 * to reduce latency impact of cache misses.
 * It's generally employed when block contains a significant portion of long-distance matches
 * or when coupled with a "cold" dictionary */
FORCE_INLINE_TEMPLATE size_t
ZSTD_decompressSequencesLong_body(
                               ZSTD_DCtx* dctx,
                               void* dst, size_t maxDstSize,
                         const void* seqStart, size_t seqSize, int nbSeq,
                         const ZSTD_longOffset_e isLongOffset)
{
    const BYTE* ip = (const BYTE*)seqStart;
    const BYTE* const iend = ip + seqSize;
    BYTE* const ostart = (BYTE*)dst;
    BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ZSTD_maybeNullPtrAdd(ostart, maxDstSize);
    BYTE* op = ostart;
    const BYTE* litPtr = dctx->litPtr;
    const BYTE* litBufferEnd = dctx->litBufferEnd;
    const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
    const BYTE* const dictStart = (const BYTE*) (dctx->virtualStart);
    const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);

    /* Regen sequences */
    if (nbSeq) {
#define STORED_SEQS 8
#define STORED_SEQS_MASK (STORED_SEQS-1)
#define ADVANCED_SEQS STORED_SEQS
        seq_t sequences[STORED_SEQS];
        int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
        seqState_t seqState;
        int seqNb;
        size_t prefetchPos = (size_t)(op-prefixStart); /* track position relative to prefixStart */

        dctx->fseEntropy = 1;
        { int i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
        assert(dst != NULL);
        assert(iend >= ip);
        RETURN_ERROR_IF(
            ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend-ip)),
            corruption_detected, "");
        ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
        ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
        ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);

        /* prepare in advance */
        for (seqNb=0; seqNb<seqAdvance; seqNb++) {
            seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset, seqNb == nbSeq-1);
            prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
            sequences[seqNb] = sequence;
        }

        /* decompress without stomping litBuffer */
        for (; seqNb < nbSeq; seqNb++) {
            seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset, seqNb == nbSeq-1);

            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd) {
                /* lit buffer is reaching split point, empty out the first buffer and transition to litExtraBuffer */
                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
                if (leftoverLit)
                {
                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
                    sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength -= leftoverLit;
                    op += leftoverLit;
                }
                litPtr = dctx->litExtraBuffer;
                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                dctx->litBufferLocation = ZSTD_not_in_dst;
                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                    assert(!ZSTD_isError(oneSeqSize));
                    ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
#endif
                    if (ZSTD_isError(oneSeqSize)) return oneSeqSize;

                    prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
                    sequences[seqNb & STORED_SEQS_MASK] = sequence;
                    op += oneSeqSize;
            }   }
            else
            {
                /* lit buffer is either wholly contained in first or second split, or not split at all*/
                size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
                    ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength - WILDCOPY_OVERLENGTH, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
                    ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                assert(!ZSTD_isError(oneSeqSize));
                ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
#endif
                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;

                prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
                sequences[seqNb & STORED_SEQS_MASK] = sequence;
                op += oneSeqSize;
            }
        }
        RETURN_ERROR_IF(!BIT_endOfDStream(&seqState.DStream), corruption_detected, "");

        /* finish queue */
        seqNb -= seqAdvance;
        for ( ; seqNb<nbSeq ; seqNb++) {
            seq_t *sequence = &(sequences[seqNb&STORED_SEQS_MASK]);
            if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd) {
                const size_t leftoverLit = dctx->litBufferEnd - litPtr;
                if (leftoverLit) {
                    RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
                    ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
                    sequence->litLength -= leftoverLit;
                    op += leftoverLit;
                }
                litPtr = dctx->litExtraBuffer;
                litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
                dctx->litBufferLocation = ZSTD_not_in_dst;
                {   size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                    assert(!ZSTD_isError(oneSeqSize));
                    ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
#endif
                    if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
                    op += oneSeqSize;
                }
            }
            else
            {
                size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
                    ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence->litLength - WILDCOPY_OVERLENGTH, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
                    ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
                assert(!ZSTD_isError(oneSeqSize));
                ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
#endif
                if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
                op += oneSeqSize;
            }
        }

        /* save reps for next block */
        { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
    }

    /* last literal segment */
    if (dctx->litBufferLocation == ZSTD_split) { /* first deplete literal buffer in dst, then copy litExtraBuffer */
        size_t const lastLLSize = litBufferEnd - litPtr;
        RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
        if (op != NULL) {
            ZSTD_memmove(op, litPtr, lastLLSize);
            op += lastLLSize;
        }
        litPtr = dctx->litExtraBuffer;
        litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
    }
    {   size_t const lastLLSize = litBufferEnd - litPtr;
        RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
        if (op != NULL) {
            ZSTD_memmove(op, litPtr, lastLLSize);
            op += lastLLSize;
        }
    }

    return (size_t)(op - ostart);
}

static size_t
ZSTD_decompressSequencesLong_default(ZSTD_DCtx* dctx,
                                 void* dst, size_t maxDstSize,
                           const void* seqStart, size_t seqSize, int nbSeq,
                           const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */


#if DYNAMIC_BMI2

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
static BMI2_TARGET_ATTRIBUTE size_t
DONT_VECTORIZE
ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
                                 void* dst, size_t maxDstSize,
                           const void* seqStart, size_t seqSize, int nbSeq,
                           const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
static BMI2_TARGET_ATTRIBUTE size_t
DONT_VECTORIZE
ZSTD_decompressSequencesSplitLitBuffer_bmi2(ZSTD_DCtx* dctx,
                                 void* dst, size_t maxDstSize,
                           const void* seqStart, size_t seqSize, int nbSeq,
                           const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
static BMI2_TARGET_ATTRIBUTE size_t
ZSTD_decompressSequencesLong_bmi2(ZSTD_DCtx* dctx,
                                 void* dst, size_t maxDstSize,
                           const void* seqStart, size_t seqSize, int nbSeq,
                           const ZSTD_longOffset_e isLongOffset)
{
    return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */

#endif /* DYNAMIC_BMI2 */

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
static size_t
ZSTD_decompressSequences(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
                   const void* seqStart, size_t seqSize, int nbSeq,
                   const ZSTD_longOffset_e isLongOffset)
{
    DEBUGLOG(5, "ZSTD_decompressSequences");
#if DYNAMIC_BMI2
    if (ZSTD_DCtx_get_bmi2(dctx)) {
        return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
    }
#endif
    return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
static size_t
ZSTD_decompressSequencesSplitLitBuffer(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
                                 const void* seqStart, size_t seqSize, int nbSeq,
                                 const ZSTD_longOffset_e isLongOffset)
{
    DEBUGLOG(5, "ZSTD_decompressSequencesSplitLitBuffer");
#if DYNAMIC_BMI2
    if (ZSTD_DCtx_get_bmi2(dctx)) {
        return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
    }
#endif
    return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */


#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
/* ZSTD_decompressSequencesLong() :
 * decompression function triggered when a minimum share of offsets is considered "long",
 * aka out of cache.
 * note : "long" definition seems overloaded here, sometimes meaning "wider than bitstream register", and sometimes meaning "farther than memory cache distance".
 * This function will try to mitigate main memory latency through the use of prefetching */
static size_t
ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
                             void* dst, size_t maxDstSize,
                             const void* seqStart, size_t seqSize, int nbSeq,
                             const ZSTD_longOffset_e isLongOffset)
{
    DEBUGLOG(5, "ZSTD_decompressSequencesLong");
#if DYNAMIC_BMI2
    if (ZSTD_DCtx_get_bmi2(dctx)) {
        return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
    }
#endif
  return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset);
}
#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */


/**
 * @returns The total size of the history referenceable by zstd, including
 * both the prefix and the extDict. At @p op any offset larger than this
 * is invalid.
 */
static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart)
{
    return (size_t)(op - virtualStart);
}

typedef struct {
    unsigned longOffsetShare;
    unsigned maxNbAdditionalBits;
} ZSTD_OffsetInfo;

/* ZSTD_getOffsetInfo() :
 * condition : offTable must be valid
 * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
 *           compared to maximum possible of (1<<OffFSELog),
 *           as well as the maximum number additional bits required.
 */
static ZSTD_OffsetInfo
ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq)
{
    ZSTD_OffsetInfo info = {0, 0};
    /* If nbSeq == 0, then the offTable is uninitialized, but we have
     * no sequences, so both values should be 0.
     */
    if (nbSeq != 0) {
        const void* ptr = offTable;
        U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
        const ZSTD_seqSymbol* table = offTable + 1;
        U32 const max = 1 << tableLog;
        U32 u;
        DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);

        assert(max <= (1 << OffFSELog));  /* max not too large */
        for (u=0; u<max; u++) {
            info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits);
            if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1;
        }

        assert(tableLog <= OffFSELog);
        info.longOffsetShare <<= (OffFSELog - tableLog);  /* scale to OffFSELog */
    }

    return info;
}

/**
 * @returns The maximum offset we can decode in one read of our bitstream, without
 * reloading more bits in the middle of the offset bits read. Any offsets larger
 * than this must use the long offset decoder.
 */
static size_t ZSTD_maxShortOffset(void)
{
    if (MEM_64bits()) {
        /* We can decode any offset without reloading bits.
         * This might change if the max window size grows.
         */
        ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
        return (size_t)-1;
    } else {
        /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1.
         * This offBase would require STREAM_ACCUMULATOR_MIN extra bits.
         * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset.
         */
        size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1;
        size_t const maxOffset = maxOffbase - ZSTD_REP_NUM;
        assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN);
        return maxOffset;
    }
}

size_t
ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
                              void* dst, size_t dstCapacity,
                        const void* src, size_t srcSize, const streaming_operation streaming)
{   /* blockType == blockCompressed */
    const BYTE* ip = (const BYTE*)src;
    DEBUGLOG(5, "ZSTD_decompressBlock_internal (cSize : %u)", (unsigned)srcSize);

    /* Note : the wording of the specification
     * allows compressed block to be sized exactly ZSTD_blockSizeMax(dctx).
     * This generally does not happen, as it makes little sense,
     * since an uncompressed block would feature same size and have no decompression cost.
     * Also, note that decoder from reference libzstd before < v1.5.4
     * would consider this edge case as an error.
     * As a consequence, avoid generating compressed blocks of size ZSTD_blockSizeMax(dctx)
     * for broader compatibility with the deployed ecosystem of zstd decoders */
    RETURN_ERROR_IF(srcSize > ZSTD_blockSizeMax(dctx), srcSize_wrong, "");

    /* Decode literals section */
    {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
        DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize);
        if (ZSTD_isError(litCSize)) return litCSize;
        ip += litCSize;
        srcSize -= litCSize;
    }

    /* Build Decoding Tables */
    {
        /* Compute the maximum block size, which must also work when !frame and fParams are unset.
         * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t.
         */
        size_t const blockSizeMax = MIN(dstCapacity, ZSTD_blockSizeMax(dctx));
        size_t const totalHistorySize = ZSTD_totalHistorySize(ZSTD_maybeNullPtrAdd((BYTE*)dst, blockSizeMax), (BYTE const*)dctx->virtualStart);
        /* isLongOffset must be true if there are long offsets.
         * Offsets are long if they are larger than ZSTD_maxShortOffset().
         * We don't expect that to be the case in 64-bit mode.
         *
         * We check here to see if our history is large enough to allow long offsets.
         * If it isn't, then we can't possible have (valid) long offsets. If the offset
         * is invalid, then it is okay to read it incorrectly.
         *
         * If isLongOffsets is true, then we will later check our decoding table to see
         * if it is even possible to generate long offsets.
         */
        ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset()));
        /* These macros control at build-time which decompressor implementation
         * we use. If neither is defined, we do some inspection and dispatch at
         * runtime.
         */
#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
        int usePrefetchDecoder = dctx->ddictIsCold;
#else
        /* Set to 1 to avoid computing offset info if we don't need to.
         * Otherwise this value is ignored.
         */
        int usePrefetchDecoder = 1;
#endif
        int nbSeq;
        size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
        if (ZSTD_isError(seqHSize)) return seqHSize;
        ip += seqHSize;
        srcSize -= seqHSize;

        RETURN_ERROR_IF((dst == NULL || dstCapacity == 0) && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
        RETURN_ERROR_IF(MEM_64bits() && sizeof(size_t) == sizeof(void*) && (size_t)(-1) - (size_t)dst < (size_t)(1 << 20), dstSize_tooSmall,
                "invalid dst");

        /* If we could potentially have long offsets, or we might want to use the prefetch decoder,
         * compute information about the share of long offsets, and the maximum nbAdditionalBits.
         * NOTE: could probably use a larger nbSeq limit
         */
        if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) {
            ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq);
            if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) {
                /* If isLongOffset, but the maximum number of additional bits that we see in our table is small
                 * enough, then we know it is impossible to have too long an offset in this block, so we can
                 * use the regular offset decoder.
                 */
                isLongOffset = ZSTD_lo_isRegularOffset;
            }
            if (!usePrefetchDecoder) {
                U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
                usePrefetchDecoder = (info.longOffsetShare >= minShare);
            }
        }

        dctx->ddictIsCold = 0;

#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
    !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
        if (usePrefetchDecoder) {
#else
        (void)usePrefetchDecoder;
        {
#endif
#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
            return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
#endif
        }

#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
        /* else */
        if (dctx->litBufferLocation == ZSTD_split)
            return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
        else
            return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset);
#endif
    }
}


ZSTD_ALLOW_POINTER_OVERFLOW_ATTR
void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
{
    if (dst != dctx->previousDstEnd && dstSize > 0) {   /* not contiguous */
        dctx->dictEnd = dctx->previousDstEnd;
        dctx->virtualStart = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
        dctx->prefixStart = dst;
        dctx->previousDstEnd = dst;
    }
}


size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
                                       void* dst, size_t dstCapacity,
                                 const void* src, size_t srcSize)
{
    size_t dSize;
    dctx->isFrameDecompression = 0;
    ZSTD_checkContinuity(dctx, dst, dstCapacity);
    dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, not_streaming);
    FORWARD_IF_ERROR(dSize, "");
    dctx->previousDstEnd = (char*)dst + dSize;
    return dSize;
}


/* NOTE: Must just wrap ZSTD_decompressBlock_deprecated() */
size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
                            void* dst, size_t dstCapacity,
                      const void* src, size_t srcSize)
{
    return ZSTD_decompressBlock_deprecated(dctx, dst, dstCapacity, src, srcSize);
}
/**** ended inlining decompress/zstd_decompress_block.c ****/

/**** start inlining dictBuilder/cover.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/* *****************************************************************************
 * Constructs a dictionary using a heuristic based on the following paper:
 *
 * Liao, Petri, Moffat, Wirth
 * Effective Construction of Relative Lempel-Ziv Dictionaries
 * Published in WWW 2016.
 *
 * Adapted from code originally written by @ot (Giuseppe Ottaviano).
 ******************************************************************************/

/*-*************************************
*  Dependencies
***************************************/
/* qsort_r is an extension. */
#if defined(__linux) || defined(__linux__) || defined(linux) || defined(__gnu_linux__) || \
    defined(__CYGWIN__) || defined(__MSYS__)
#if !defined(_GNU_SOURCE) && !defined(__ANDROID__) /* NDK doesn't ship qsort_r(). */
#define _GNU_SOURCE
#endif
#endif

#include <stdio.h>  /* fprintf */
#include <stdlib.h> /* malloc, free, qsort_r */

#include <string.h> /* memset */
#include <time.h>   /* clock */

#ifndef ZDICT_STATIC_LINKING_ONLY
#  define ZDICT_STATIC_LINKING_ONLY
#endif

/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/pool.h ****/
/**** skipping file: ../common/threading.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../common/bits.h ****/
/**** start inlining ../zdict.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZSTD_ZDICT_H
#define ZSTD_ZDICT_H


/*======  Dependencies  ======*/
#include <stddef.h>  /* size_t */

#if defined (__cplusplus)
extern "C" {
#endif

/* =====   ZDICTLIB_API : control library symbols visibility   ===== */
#ifndef ZDICTLIB_VISIBLE
   /* Backwards compatibility with old macro name */
#  ifdef ZDICTLIB_VISIBILITY
#    define ZDICTLIB_VISIBLE ZDICTLIB_VISIBILITY
#  elif defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
#    define ZDICTLIB_VISIBLE __attribute__ ((visibility ("default")))
#  else
#    define ZDICTLIB_VISIBLE
#  endif
#endif

#ifndef ZDICTLIB_HIDDEN
#  if defined(__GNUC__) && (__GNUC__ >= 4) && !defined(__MINGW32__)
#    define ZDICTLIB_HIDDEN __attribute__ ((visibility ("hidden")))
#  else
#    define ZDICTLIB_HIDDEN
#  endif
#endif

#if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
#  define ZDICTLIB_API __declspec(dllexport) ZDICTLIB_VISIBLE
#elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
#  define ZDICTLIB_API __declspec(dllimport) ZDICTLIB_VISIBLE /* It isn't required but allows to generate better code, saving a function pointer load from the IAT and an indirect jump.*/
#else
#  define ZDICTLIB_API ZDICTLIB_VISIBLE
#endif

/*******************************************************************************
 * Zstd dictionary builder
 *
 * FAQ
 * ===
 * Why should I use a dictionary?
 * ------------------------------
 *
 * Zstd can use dictionaries to improve compression ratio of small data.
 * Traditionally small files don't compress well because there is very little
 * repetition in a single sample, since it is small. But, if you are compressing
 * many similar files, like a bunch of JSON records that share the same
 * structure, you can train a dictionary on ahead of time on some samples of
 * these files. Then, zstd can use the dictionary to find repetitions that are
 * present across samples. This can vastly improve compression ratio.
 *
 * When is a dictionary useful?
 * ----------------------------
 *
 * Dictionaries are useful when compressing many small files that are similar.
 * The larger a file is, the less benefit a dictionary will have. Generally,
 * we don't expect dictionary compression to be effective past 100KB. And the
 * smaller a file is, the more we would expect the dictionary to help.
 *
 * How do I use a dictionary?
 * --------------------------
 *
 * Simply pass the dictionary to the zstd compressor with
 * `ZSTD_CCtx_loadDictionary()`. The same dictionary must then be passed to
 * the decompressor, using `ZSTD_DCtx_loadDictionary()`. There are other
 * more advanced functions that allow selecting some options, see zstd.h for
 * complete documentation.
 *
 * What is a zstd dictionary?
 * --------------------------
 *
 * A zstd dictionary has two pieces: Its header, and its content. The header
 * contains a magic number, the dictionary ID, and entropy tables. These
 * entropy tables allow zstd to save on header costs in the compressed file,
 * which really matters for small data. The content is just bytes, which are
 * repeated content that is common across many samples.
 *
 * What is a raw content dictionary?
 * ---------------------------------
 *
 * A raw content dictionary is just bytes. It doesn't have a zstd dictionary
 * header, a dictionary ID, or entropy tables. Any buffer is a valid raw
 * content dictionary.
 *
 * How do I train a dictionary?
 * ----------------------------
 *
 * Gather samples from your use case. These samples should be similar to each
 * other. If you have several use cases, you could try to train one dictionary
 * per use case.
 *
 * Pass those samples to `ZDICT_trainFromBuffer()` and that will train your
 * dictionary. There are a few advanced versions of this function, but this
 * is a great starting point. If you want to further tune your dictionary
 * you could try `ZDICT_optimizeTrainFromBuffer_cover()`. If that is too slow
 * you can try `ZDICT_optimizeTrainFromBuffer_fastCover()`.
 *
 * If the dictionary training function fails, that is likely because you
 * either passed too few samples, or a dictionary would not be effective
 * for your data. Look at the messages that the dictionary trainer printed,
 * if it doesn't say too few samples, then a dictionary would not be effective.
 *
 * How large should my dictionary be?
 * ----------------------------------
 *
 * A reasonable dictionary size, the `dictBufferCapacity`, is about 100KB.
 * The zstd CLI defaults to a 110KB dictionary. You likely don't need a
 * dictionary larger than that. But, most use cases can get away with a
 * smaller dictionary. The advanced dictionary builders can automatically
 * shrink the dictionary for you, and select the smallest size that doesn't
 * hurt compression ratio too much. See the `shrinkDict` parameter.
 * A smaller dictionary can save memory, and potentially speed up
 * compression.
 *
 * How many samples should I provide to the dictionary builder?
 * ------------------------------------------------------------
 *
 * We generally recommend passing ~100x the size of the dictionary
 * in samples. A few thousand should suffice. Having too few samples
 * can hurt the dictionaries effectiveness. Having more samples will
 * only improve the dictionaries effectiveness. But having too many
 * samples can slow down the dictionary builder.
 *
 * How do I determine if a dictionary will be effective?
 * -----------------------------------------------------
 *
 * Simply train a dictionary and try it out. You can use zstd's built in
 * benchmarking tool to test the dictionary effectiveness.
 *
 *   # Benchmark levels 1-3 without a dictionary
 *   zstd -b1e3 -r /path/to/my/files
 *   # Benchmark levels 1-3 with a dictionary
 *   zstd -b1e3 -r /path/to/my/files -D /path/to/my/dictionary
 *
 * When should I retrain a dictionary?
 * -----------------------------------
 *
 * You should retrain a dictionary when its effectiveness drops. Dictionary
 * effectiveness drops as the data you are compressing changes. Generally, we do
 * expect dictionaries to "decay" over time, as your data changes, but the rate
 * at which they decay depends on your use case. Internally, we regularly
 * retrain dictionaries, and if the new dictionary performs significantly
 * better than the old dictionary, we will ship the new dictionary.
 *
 * I have a raw content dictionary, how do I turn it into a zstd dictionary?
 * -------------------------------------------------------------------------
 *
 * If you have a raw content dictionary, e.g. by manually constructing it, or
 * using a third-party dictionary builder, you can turn it into a zstd
 * dictionary by using `ZDICT_finalizeDictionary()`. You'll also have to
 * provide some samples of the data. It will add the zstd header to the
 * raw content, which contains a dictionary ID and entropy tables, which
 * will improve compression ratio, and allow zstd to write the dictionary ID
 * into the frame, if you so choose.
 *
 * Do I have to use zstd's dictionary builder?
 * -------------------------------------------
 *
 * No! You can construct dictionary content however you please, it is just
 * bytes. It will always be valid as a raw content dictionary. If you want
 * a zstd dictionary, which can improve compression ratio, use
 * `ZDICT_finalizeDictionary()`.
 *
 * What is the attack surface of a zstd dictionary?
 * ------------------------------------------------
 *
 * Zstd is heavily fuzz tested, including loading fuzzed dictionaries, so
 * zstd should never crash, or access out-of-bounds memory no matter what
 * the dictionary is. However, if an attacker can control the dictionary
 * during decompression, they can cause zstd to generate arbitrary bytes,
 * just like if they controlled the compressed data.
 *
 ******************************************************************************/


/*! ZDICT_trainFromBuffer():
 *  Train a dictionary from an array of samples.
 *  Redirect towards ZDICT_optimizeTrainFromBuffer_fastCover() single-threaded, with d=8, steps=4,
 *  f=20, and accel=1.
 *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
 *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
 *  The resulting dictionary will be saved into `dictBuffer`.
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *  Note:  Dictionary training will fail if there are not enough samples to construct a
 *         dictionary, or if most of the samples are too small (< 8 bytes being the lower limit).
 *         If dictionary training fails, you should use zstd without a dictionary, as the dictionary
 *         would've been ineffective anyways. If you believe your samples would benefit from a dictionary
 *         please open an issue with details, and we can look into it.
 *  Note: ZDICT_trainFromBuffer()'s memory usage is about 6 MB.
 *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
 *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
 *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
 *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
 */
ZDICTLIB_API size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
                                    const void* samplesBuffer,
                                    const size_t* samplesSizes, unsigned nbSamples);

typedef struct {
    int      compressionLevel;   /**< optimize for a specific zstd compression level; 0 means default */
    unsigned notificationLevel;  /**< Write log to stderr; 0 = none (default); 1 = errors; 2 = progression; 3 = details; 4 = debug; */
    unsigned dictID;             /**< force dictID value; 0 means auto mode (32-bits random value)
                                  *   NOTE: The zstd format reserves some dictionary IDs for future use.
                                  *         You may use them in private settings, but be warned that they
                                  *         may be used by zstd in a public dictionary registry in the future.
                                  *         These dictionary IDs are:
                                  *           - low range  : <= 32767
                                  *           - high range : >= (2^31)
                                  */
} ZDICT_params_t;

/*! ZDICT_finalizeDictionary():
 * Given a custom content as a basis for dictionary, and a set of samples,
 * finalize dictionary by adding headers and statistics according to the zstd
 * dictionary format.
 *
 * Samples must be stored concatenated in a flat buffer `samplesBuffer`,
 * supplied with an array of sizes `samplesSizes`, providing the size of each
 * sample in order. The samples are used to construct the statistics, so they
 * should be representative of what you will compress with this dictionary.
 *
 * The compression level can be set in `parameters`. You should pass the
 * compression level you expect to use in production. The statistics for each
 * compression level differ, so tuning the dictionary for the compression level
 * can help quite a bit.
 *
 * You can set an explicit dictionary ID in `parameters`, or allow us to pick
 * a random dictionary ID for you, but we can't guarantee no collisions.
 *
 * The dstDictBuffer and the dictContent may overlap, and the content will be
 * appended to the end of the header. If the header + the content doesn't fit in
 * maxDictSize the beginning of the content is truncated to make room, since it
 * is presumed that the most profitable content is at the end of the dictionary,
 * since that is the cheapest to reference.
 *
 * `maxDictSize` must be >= max(dictContentSize, ZDICT_DICTSIZE_MIN).
 *
 * @return: size of dictionary stored into `dstDictBuffer` (<= `maxDictSize`),
 *          or an error code, which can be tested by ZDICT_isError().
 * Note: ZDICT_finalizeDictionary() will push notifications into stderr if
 *       instructed to, using notificationLevel>0.
 * NOTE: This function currently may fail in several edge cases including:
 *         * Not enough samples
 *         * Samples are uncompressible
 *         * Samples are all exactly the same
 */
ZDICTLIB_API size_t ZDICT_finalizeDictionary(void* dstDictBuffer, size_t maxDictSize,
                                const void* dictContent, size_t dictContentSize,
                                const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
                                ZDICT_params_t parameters);


/*======   Helper functions   ======*/
ZDICTLIB_API unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize);  /**< extracts dictID; @return zero if error (not a valid dictionary) */
ZDICTLIB_API size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize);  /* returns dict header size; returns a ZSTD error code on failure */
ZDICTLIB_API unsigned ZDICT_isError(size_t errorCode);
ZDICTLIB_API const char* ZDICT_getErrorName(size_t errorCode);

#if defined (__cplusplus)
}
#endif

#endif   /* ZSTD_ZDICT_H */

#if defined(ZDICT_STATIC_LINKING_ONLY) && !defined(ZSTD_ZDICT_H_STATIC)
#define ZSTD_ZDICT_H_STATIC

#if defined (__cplusplus)
extern "C" {
#endif

/* This can be overridden externally to hide static symbols. */
#ifndef ZDICTLIB_STATIC_API
#  if defined(ZSTD_DLL_EXPORT) && (ZSTD_DLL_EXPORT==1)
#    define ZDICTLIB_STATIC_API __declspec(dllexport) ZDICTLIB_VISIBLE
#  elif defined(ZSTD_DLL_IMPORT) && (ZSTD_DLL_IMPORT==1)
#    define ZDICTLIB_STATIC_API __declspec(dllimport) ZDICTLIB_VISIBLE
#  else
#    define ZDICTLIB_STATIC_API ZDICTLIB_VISIBLE
#  endif
#endif

/* ====================================================================================
 * The definitions in this section are considered experimental.
 * They should never be used with a dynamic library, as they may change in the future.
 * They are provided for advanced usages.
 * Use them only in association with static linking.
 * ==================================================================================== */

#define ZDICT_DICTSIZE_MIN    256
/* Deprecated: Remove in v1.6.0 */
#define ZDICT_CONTENTSIZE_MIN 128

/*! ZDICT_cover_params_t:
 *  k and d are the only required parameters.
 *  For others, value 0 means default.
 */
typedef struct {
    unsigned k;                  /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
    unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
    unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */
    unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
    double splitPoint;           /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (1.0), 1.0 when all samples are used for both training and testing */
    unsigned shrinkDict;         /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking  */
    unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */
    ZDICT_params_t zParams;
} ZDICT_cover_params_t;

typedef struct {
    unsigned k;                  /* Segment size : constraint: 0 < k : Reasonable range [16, 2048+] */
    unsigned d;                  /* dmer size : constraint: 0 < d <= k : Reasonable range [6, 16] */
    unsigned f;                  /* log of size of frequency array : constraint: 0 < f <= 31 : 1 means default(20)*/
    unsigned steps;              /* Number of steps : Only used for optimization : 0 means default (40) : Higher means more parameters checked */
    unsigned nbThreads;          /* Number of threads : constraint: 0 < nbThreads : 1 means single-threaded : Only used for optimization : Ignored if ZSTD_MULTITHREAD is not defined */
    double splitPoint;           /* Percentage of samples used for training: Only used for optimization : the first nbSamples * splitPoint samples will be used to training, the last nbSamples * (1 - splitPoint) samples will be used for testing, 0 means default (0.75), 1.0 when all samples are used for both training and testing */
    unsigned accel;              /* Acceleration level: constraint: 0 < accel <= 10, higher means faster and less accurate, 0 means default(1) */
    unsigned shrinkDict;         /* Train dictionaries to shrink in size starting from the minimum size and selects the smallest dictionary that is shrinkDictMaxRegression% worse than the largest dictionary. 0 means no shrinking and 1 means shrinking  */
    unsigned shrinkDictMaxRegression; /* Sets shrinkDictMaxRegression so that a smaller dictionary can be at worse shrinkDictMaxRegression% worse than the max dict size dictionary. */

    ZDICT_params_t zParams;
} ZDICT_fastCover_params_t;

/*! ZDICT_trainFromBuffer_cover():
 *  Train a dictionary from an array of samples using the COVER algorithm.
 *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
 *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
 *  The resulting dictionary will be saved into `dictBuffer`.
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *          See ZDICT_trainFromBuffer() for details on failure modes.
 *  Note: ZDICT_trainFromBuffer_cover() requires about 9 bytes of memory for each input byte.
 *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
 *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
 *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
 *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
 */
ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_cover(
          void *dictBuffer, size_t dictBufferCapacity,
    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
          ZDICT_cover_params_t parameters);

/*! ZDICT_optimizeTrainFromBuffer_cover():
 * The same requirements as above hold for all the parameters except `parameters`.
 * This function tries many parameter combinations and picks the best parameters.
 * `*parameters` is filled with the best parameters found,
 * dictionary constructed with those parameters is stored in `dictBuffer`.
 *
 * All of the parameters d, k, steps are optional.
 * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
 * if steps is zero it defaults to its default value.
 * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
 *
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *          On success `*parameters` contains the parameters selected.
 *          See ZDICT_trainFromBuffer() for details on failure modes.
 * Note: ZDICT_optimizeTrainFromBuffer_cover() requires about 8 bytes of memory for each input byte and additionally another 5 bytes of memory for each byte of memory for each thread.
 */
ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_cover(
          void* dictBuffer, size_t dictBufferCapacity,
    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
          ZDICT_cover_params_t* parameters);

/*! ZDICT_trainFromBuffer_fastCover():
 *  Train a dictionary from an array of samples using a modified version of COVER algorithm.
 *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
 *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
 *  d and k are required.
 *  All other parameters are optional, will use default values if not provided
 *  The resulting dictionary will be saved into `dictBuffer`.
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *          See ZDICT_trainFromBuffer() for details on failure modes.
 *  Note: ZDICT_trainFromBuffer_fastCover() requires 6 * 2^f bytes of memory.
 *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
 *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
 *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
 *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
 */
ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_fastCover(void *dictBuffer,
                    size_t dictBufferCapacity, const void *samplesBuffer,
                    const size_t *samplesSizes, unsigned nbSamples,
                    ZDICT_fastCover_params_t parameters);

/*! ZDICT_optimizeTrainFromBuffer_fastCover():
 * The same requirements as above hold for all the parameters except `parameters`.
 * This function tries many parameter combinations (specifically, k and d combinations)
 * and picks the best parameters. `*parameters` is filled with the best parameters found,
 * dictionary constructed with those parameters is stored in `dictBuffer`.
 * All of the parameters d, k, steps, f, and accel are optional.
 * If d is non-zero then we don't check multiple values of d, otherwise we check d = {6, 8}.
 * if steps is zero it defaults to its default value.
 * If k is non-zero then we don't check multiple values of k, otherwise we check steps values in [50, 2000].
 * If f is zero, default value of 20 is used.
 * If accel is zero, default value of 1 is used.
 *
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *          On success `*parameters` contains the parameters selected.
 *          See ZDICT_trainFromBuffer() for details on failure modes.
 * Note: ZDICT_optimizeTrainFromBuffer_fastCover() requires about 6 * 2^f bytes of memory for each thread.
 */
ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_fastCover(void* dictBuffer,
                    size_t dictBufferCapacity, const void* samplesBuffer,
                    const size_t* samplesSizes, unsigned nbSamples,
                    ZDICT_fastCover_params_t* parameters);

typedef struct {
    unsigned selectivityLevel;   /* 0 means default; larger => select more => larger dictionary */
    ZDICT_params_t zParams;
} ZDICT_legacy_params_t;

/*! ZDICT_trainFromBuffer_legacy():
 *  Train a dictionary from an array of samples.
 *  Samples must be stored concatenated in a single flat buffer `samplesBuffer`,
 *  supplied with an array of sizes `samplesSizes`, providing the size of each sample, in order.
 *  The resulting dictionary will be saved into `dictBuffer`.
 * `parameters` is optional and can be provided with values set to 0 to mean "default".
 * @return: size of dictionary stored into `dictBuffer` (<= `dictBufferCapacity`)
 *          or an error code, which can be tested with ZDICT_isError().
 *          See ZDICT_trainFromBuffer() for details on failure modes.
 *  Tips: In general, a reasonable dictionary has a size of ~ 100 KB.
 *        It's possible to select smaller or larger size, just by specifying `dictBufferCapacity`.
 *        In general, it's recommended to provide a few thousands samples, though this can vary a lot.
 *        It's recommended that total size of all samples be about ~x100 times the target size of dictionary.
 *  Note: ZDICT_trainFromBuffer_legacy() will send notifications into stderr if instructed to, using notificationLevel>0.
 */
ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_legacy(
    void* dictBuffer, size_t dictBufferCapacity,
    const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
    ZDICT_legacy_params_t parameters);


/* Deprecation warnings */
/* It is generally possible to disable deprecation warnings from compiler,
   for example with -Wno-deprecated-declarations for gcc
   or _CRT_SECURE_NO_WARNINGS in Visual.
   Otherwise, it's also possible to manually define ZDICT_DISABLE_DEPRECATE_WARNINGS */
#ifdef ZDICT_DISABLE_DEPRECATE_WARNINGS
#  define ZDICT_DEPRECATED(message) /* disable deprecation warnings */
#else
#  define ZDICT_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
#  if defined (__cplusplus) && (__cplusplus >= 201402) /* C++14 or greater */
#    define ZDICT_DEPRECATED(message) [[deprecated(message)]]
#  elif defined(__clang__) || (ZDICT_GCC_VERSION >= 405)
#    define ZDICT_DEPRECATED(message) __attribute__((deprecated(message)))
#  elif (ZDICT_GCC_VERSION >= 301)
#    define ZDICT_DEPRECATED(message) __attribute__((deprecated))
#  elif defined(_MSC_VER)
#    define ZDICT_DEPRECATED(message) __declspec(deprecated(message))
#  else
#    pragma message("WARNING: You need to implement ZDICT_DEPRECATED for this compiler")
#    define ZDICT_DEPRECATED(message)
#  endif
#endif /* ZDICT_DISABLE_DEPRECATE_WARNINGS */

ZDICT_DEPRECATED("use ZDICT_finalizeDictionary() instead")
ZDICTLIB_STATIC_API
size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
                                  const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples);

#if defined (__cplusplus)
}
#endif

#endif   /* ZSTD_ZDICT_H_STATIC */
/**** ended inlining ../zdict.h ****/
/**** start inlining cover.h ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

#ifndef ZDICT_STATIC_LINKING_ONLY
#  define ZDICT_STATIC_LINKING_ONLY
#endif

/**** skipping file: ../common/threading.h ****/
/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../zdict.h ****/

/**
 * COVER_best_t is used for two purposes:
 * 1. Synchronizing threads.
 * 2. Saving the best parameters and dictionary.
 *
 * All of the methods except COVER_best_init() are thread safe if zstd is
 * compiled with multithreaded support.
 */
typedef struct COVER_best_s {
  ZSTD_pthread_mutex_t mutex;
  ZSTD_pthread_cond_t cond;
  size_t liveJobs;
  void *dict;
  size_t dictSize;
  ZDICT_cover_params_t parameters;
  size_t compressedSize;
} COVER_best_t;

/**
 * A segment is a range in the source as well as the score of the segment.
 */
typedef struct {
  U32 begin;
  U32 end;
  U32 score;
} COVER_segment_t;

/**
 *Number of epochs and size of each epoch.
 */
typedef struct {
  U32 num;
  U32 size;
} COVER_epoch_info_t;

/**
 * Struct used for the dictionary selection function.
 */
typedef struct COVER_dictSelection {
  BYTE* dictContent;
  size_t dictSize;
  size_t totalCompressedSize;
} COVER_dictSelection_t;

/**
 * Computes the number of epochs and the size of each epoch.
 * We will make sure that each epoch gets at least 10 * k bytes.
 *
 * The COVER algorithms divide the data up into epochs of equal size and
 * select one segment from each epoch.
 *
 * @param maxDictSize The maximum allowed dictionary size.
 * @param nbDmers     The number of dmers we are training on.
 * @param k           The parameter k (segment size).
 * @param passes      The target number of passes over the dmer corpus.
 *                    More passes means a better dictionary.
 */
COVER_epoch_info_t COVER_computeEpochs(U32 maxDictSize, U32 nbDmers,
                                       U32 k, U32 passes);

/**
 * Warns the user when their corpus is too small.
 */
void COVER_warnOnSmallCorpus(size_t maxDictSize, size_t nbDmers, int displayLevel);

/**
 *  Checks total compressed size of a dictionary
 */
size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters,
                                      const size_t *samplesSizes, const BYTE *samples,
                                      size_t *offsets,
                                      size_t nbTrainSamples, size_t nbSamples,
                                      BYTE *const dict, size_t dictBufferCapacity);

/**
 * Returns the sum of the sample sizes.
 */
size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) ;

/**
 * Initialize the `COVER_best_t`.
 */
void COVER_best_init(COVER_best_t *best);

/**
 * Wait until liveJobs == 0.
 */
void COVER_best_wait(COVER_best_t *best);

/**
 * Call COVER_best_wait() and then destroy the COVER_best_t.
 */
void COVER_best_destroy(COVER_best_t *best);

/**
 * Called when a thread is about to be launched.
 * Increments liveJobs.
 */
void COVER_best_start(COVER_best_t *best);

/**
 * Called when a thread finishes executing, both on error or success.
 * Decrements liveJobs and signals any waiting threads if liveJobs == 0.
 * If this dictionary is the best so far save it and its parameters.
 */
void COVER_best_finish(COVER_best_t *best, ZDICT_cover_params_t parameters,
                       COVER_dictSelection_t selection);
/**
 * Error function for COVER_selectDict function. Checks if the return
 * value is an error.
 */
unsigned COVER_dictSelectionIsError(COVER_dictSelection_t selection);

 /**
  * Error function for COVER_selectDict function. Returns a struct where
  * return.totalCompressedSize is a ZSTD error.
  */
COVER_dictSelection_t COVER_dictSelectionError(size_t error);

/**
 * Always call after selectDict is called to free up used memory from
 * newly created dictionary.
 */
void COVER_dictSelectionFree(COVER_dictSelection_t selection);

/**
 * Called to finalize the dictionary and select one based on whether or not
 * the shrink-dict flag was enabled. If enabled the dictionary used is the
 * smallest dictionary within a specified regression of the compressed size
 * from the largest dictionary.
 */
 COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBufferCapacity,
                       size_t dictContentSize, const BYTE* samplesBuffer, const size_t* samplesSizes, unsigned nbFinalizeSamples,
                       size_t nbCheckSamples, size_t nbSamples, ZDICT_cover_params_t params, size_t* offsets, size_t totalCompressedSize);
/**** ended inlining cover.h ****/

/*-*************************************
*  Constants
***************************************/
/**
* There are 32bit indexes used to ref samples, so limit samples size to 4GB
* on 64bit builds.
* For 32bit builds we choose 1 GB.
* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
* contiguous buffer, so 1GB is already a high limit.
*/
#define COVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
#define COVER_DEFAULT_SPLITPOINT 1.0

/*-*************************************
*  Console display
***************************************/
#ifndef LOCALDISPLAYLEVEL
static int g_displayLevel = 0;
#endif
#undef  DISPLAY
#define DISPLAY(...)                                                           \
  {                                                                            \
    fprintf(stderr, __VA_ARGS__);                                              \
    fflush(stderr);                                                            \
  }
#undef  LOCALDISPLAYLEVEL
#define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
  if (displayLevel >= l) {                                                     \
    DISPLAY(__VA_ARGS__);                                                      \
  } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
#undef  DISPLAYLEVEL
#define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)

#ifndef LOCALDISPLAYUPDATE
static const clock_t g_refreshRate = CLOCKS_PER_SEC * 15 / 100;
static clock_t g_time = 0;
#endif
#undef  LOCALDISPLAYUPDATE
#define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
  if (displayLevel >= l) {                                                     \
    if ((clock() - g_time > g_refreshRate) || (displayLevel >= 4)) {           \
      g_time = clock();                                                        \
      DISPLAY(__VA_ARGS__);                                                    \
    }                                                                          \
  }
#undef  DISPLAYUPDATE
#define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)

/*-*************************************
* Hash table
***************************************
* A small specialized hash map for storing activeDmers.
* The map does not resize, so if it becomes full it will loop forever.
* Thus, the map must be large enough to store every value.
* The map implements linear probing and keeps its load less than 0.5.
*/

#define MAP_EMPTY_VALUE ((U32)-1)
typedef struct COVER_map_pair_t_s {
  U32 key;
  U32 value;
} COVER_map_pair_t;

typedef struct COVER_map_s {
  COVER_map_pair_t *data;
  U32 sizeLog;
  U32 size;
  U32 sizeMask;
} COVER_map_t;

/**
 * Clear the map.
 */
static void COVER_map_clear(COVER_map_t *map) {
  memset(map->data, MAP_EMPTY_VALUE, map->size * sizeof(COVER_map_pair_t));
}

/**
 * Initializes a map of the given size.
 * Returns 1 on success and 0 on failure.
 * The map must be destroyed with COVER_map_destroy().
 * The map is only guaranteed to be large enough to hold size elements.
 */
static int COVER_map_init(COVER_map_t *map, U32 size) {
  map->sizeLog = ZSTD_highbit32(size) + 2;
  map->size = (U32)1 << map->sizeLog;
  map->sizeMask = map->size - 1;
  map->data = (COVER_map_pair_t *)malloc(map->size * sizeof(COVER_map_pair_t));
  if (!map->data) {
    map->sizeLog = 0;
    map->size = 0;
    return 0;
  }
  COVER_map_clear(map);
  return 1;
}

/**
 * Internal hash function
 */
static const U32 COVER_prime4bytes = 2654435761U;
static U32 COVER_map_hash(COVER_map_t *map, U32 key) {
  return (key * COVER_prime4bytes) >> (32 - map->sizeLog);
}

/**
 * Helper function that returns the index that a key should be placed into.
 */
static U32 COVER_map_index(COVER_map_t *map, U32 key) {
  const U32 hash = COVER_map_hash(map, key);
  U32 i;
  for (i = hash;; i = (i + 1) & map->sizeMask) {
    COVER_map_pair_t *pos = &map->data[i];
    if (pos->value == MAP_EMPTY_VALUE) {
      return i;
    }
    if (pos->key == key) {
      return i;
    }
  }
}

/**
 * Returns the pointer to the value for key.
 * If key is not in the map, it is inserted and the value is set to 0.
 * The map must not be full.
 */
static U32 *COVER_map_at(COVER_map_t *map, U32 key) {
  COVER_map_pair_t *pos = &map->data[COVER_map_index(map, key)];
  if (pos->value == MAP_EMPTY_VALUE) {
    pos->key = key;
    pos->value = 0;
  }
  return &pos->value;
}

/**
 * Deletes key from the map if present.
 */
static void COVER_map_remove(COVER_map_t *map, U32 key) {
  U32 i = COVER_map_index(map, key);
  COVER_map_pair_t *del = &map->data[i];
  U32 shift = 1;
  if (del->value == MAP_EMPTY_VALUE) {
    return;
  }
  for (i = (i + 1) & map->sizeMask;; i = (i + 1) & map->sizeMask) {
    COVER_map_pair_t *const pos = &map->data[i];
    /* If the position is empty we are done */
    if (pos->value == MAP_EMPTY_VALUE) {
      del->value = MAP_EMPTY_VALUE;
      return;
    }
    /* If pos can be moved to del do so */
    if (((i - COVER_map_hash(map, pos->key)) & map->sizeMask) >= shift) {
      del->key = pos->key;
      del->value = pos->value;
      del = pos;
      shift = 1;
    } else {
      ++shift;
    }
  }
}

/**
 * Destroys a map that is inited with COVER_map_init().
 */
static void COVER_map_destroy(COVER_map_t *map) {
  if (map->data) {
    free(map->data);
  }
  map->data = NULL;
  map->size = 0;
}

/*-*************************************
* Context
***************************************/

typedef struct {
  const BYTE *samples;
  size_t *offsets;
  const size_t *samplesSizes;
  size_t nbSamples;
  size_t nbTrainSamples;
  size_t nbTestSamples;
  U32 *suffix;
  size_t suffixSize;
  U32 *freqs;
  U32 *dmerAt;
  unsigned d;
} COVER_ctx_t;

#if !defined(_GNU_SOURCE) && !defined(__APPLE__) && !defined(_MSC_VER)
/* C90 only offers qsort() that needs a global context. */
static COVER_ctx_t *g_coverCtx = NULL;
#endif

/*-*************************************
*  Helper functions
***************************************/

/**
 * Returns the sum of the sample sizes.
 */
size_t COVER_sum(const size_t *samplesSizes, unsigned nbSamples) {
  size_t sum = 0;
  unsigned i;
  for (i = 0; i < nbSamples; ++i) {
    sum += samplesSizes[i];
  }
  return sum;
}

/**
 * Returns -1 if the dmer at lp is less than the dmer at rp.
 * Return 0 if the dmers at lp and rp are equal.
 * Returns 1 if the dmer at lp is greater than the dmer at rp.
 */
static int COVER_cmp(COVER_ctx_t *ctx, const void *lp, const void *rp) {
  U32 const lhs = *(U32 const *)lp;
  U32 const rhs = *(U32 const *)rp;
  return memcmp(ctx->samples + lhs, ctx->samples + rhs, ctx->d);
}
/**
 * Faster version for d <= 8.
 */
static int COVER_cmp8(COVER_ctx_t *ctx, const void *lp, const void *rp) {
  U64 const mask = (ctx->d == 8) ? (U64)-1 : (((U64)1 << (8 * ctx->d)) - 1);
  U64 const lhs = MEM_readLE64(ctx->samples + *(U32 const *)lp) & mask;
  U64 const rhs = MEM_readLE64(ctx->samples + *(U32 const *)rp) & mask;
  if (lhs < rhs) {
    return -1;
  }
  return (lhs > rhs);
}

/**
 * Same as COVER_cmp() except ties are broken by pointer value
 */
#if (defined(_WIN32) && defined(_MSC_VER)) || defined(__APPLE__)
static int WIN_CDECL COVER_strict_cmp(void* g_coverCtx, const void* lp, const void* rp) {
#elif defined(_GNU_SOURCE)
static int COVER_strict_cmp(const void *lp, const void *rp, void *g_coverCtx) {
#else /* C90 fallback.*/
static int COVER_strict_cmp(const void *lp, const void *rp) {
#endif
  int result = COVER_cmp((COVER_ctx_t*)g_coverCtx, lp, rp);
  if (result == 0) {
    result = lp < rp ? -1 : 1;
  }
  return result;
}
/**
 * Faster version for d <= 8.
 */
#if (defined(_WIN32) && defined(_MSC_VER)) || defined(__APPLE__)
static int WIN_CDECL COVER_strict_cmp8(void* g_coverCtx, const void* lp, const void* rp) {
#elif defined(_GNU_SOURCE)
static int COVER_strict_cmp8(const void *lp, const void *rp, void *g_coverCtx) {
#else /* C90 fallback.*/
static int COVER_strict_cmp8(const void *lp, const void *rp) {
#endif
  int result = COVER_cmp8((COVER_ctx_t*)g_coverCtx, lp, rp);
  if (result == 0) {
    result = lp < rp ? -1 : 1;
  }
  return result;
}

/**
 * Abstract away divergence of qsort_r() parameters.
 * Hopefully when C11 become the norm, we will be able
 * to clean it up.
 */
static void stableSort(COVER_ctx_t *ctx) {
#if defined(__APPLE__)
    qsort_r(ctx->suffix, ctx->suffixSize, sizeof(U32),
            ctx,
            (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
#elif defined(_GNU_SOURCE)
    qsort_r(ctx->suffix, ctx->suffixSize, sizeof(U32),
            (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp),
            ctx);
#elif defined(_WIN32) && defined(_MSC_VER)
    qsort_s(ctx->suffix, ctx->suffixSize, sizeof(U32),
            (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp),
            ctx);
#elif defined(__OpenBSD__)
    g_coverCtx = ctx;
    mergesort(ctx->suffix, ctx->suffixSize, sizeof(U32),
          (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
#else /* C90 fallback.*/
    g_coverCtx = ctx;
    /* TODO(cavalcanti): implement a reentrant qsort() when is not available. */
    qsort(ctx->suffix, ctx->suffixSize, sizeof(U32),
          (ctx->d <= 8 ? &COVER_strict_cmp8 : &COVER_strict_cmp));
#endif
}

/**
 * Returns the first pointer in [first, last) whose element does not compare
 * less than value.  If no such element exists it returns last.
 */
static const size_t *COVER_lower_bound(const size_t* first, const size_t* last,
                                       size_t value) {
  size_t count = (size_t)(last - first);
  assert(last >= first);
  while (count != 0) {
    size_t step = count / 2;
    const size_t *ptr = first;
    ptr += step;
    if (*ptr < value) {
      first = ++ptr;
      count -= step + 1;
    } else {
      count = step;
    }
  }
  return first;
}

/**
 * Generic groupBy function.
 * Groups an array sorted by cmp into groups with equivalent values.
 * Calls grp for each group.
 */
static void
COVER_groupBy(const void *data, size_t count, size_t size, COVER_ctx_t *ctx,
              int (*cmp)(COVER_ctx_t *, const void *, const void *),
              void (*grp)(COVER_ctx_t *, const void *, const void *)) {
  const BYTE *ptr = (const BYTE *)data;
  size_t num = 0;
  while (num < count) {
    const BYTE *grpEnd = ptr + size;
    ++num;
    while (num < count && cmp(ctx, ptr, grpEnd) == 0) {
      grpEnd += size;
      ++num;
    }
    grp(ctx, ptr, grpEnd);
    ptr = grpEnd;
  }
}

/*-*************************************
*  Cover functions
***************************************/

/**
 * Called on each group of positions with the same dmer.
 * Counts the frequency of each dmer and saves it in the suffix array.
 * Fills `ctx->dmerAt`.
 */
static void COVER_group(COVER_ctx_t *ctx, const void *group,
                        const void *groupEnd) {
  /* The group consists of all the positions with the same first d bytes. */
  const U32 *grpPtr = (const U32 *)group;
  const U32 *grpEnd = (const U32 *)groupEnd;
  /* The dmerId is how we will reference this dmer.
   * This allows us to map the whole dmer space to a much smaller space, the
   * size of the suffix array.
   */
  const U32 dmerId = (U32)(grpPtr - ctx->suffix);
  /* Count the number of samples this dmer shows up in */
  U32 freq = 0;
  /* Details */
  const size_t *curOffsetPtr = ctx->offsets;
  const size_t *offsetsEnd = ctx->offsets + ctx->nbSamples;
  /* Once *grpPtr >= curSampleEnd this occurrence of the dmer is in a
   * different sample than the last.
   */
  size_t curSampleEnd = ctx->offsets[0];
  for (; grpPtr != grpEnd; ++grpPtr) {
    /* Save the dmerId for this position so we can get back to it. */
    ctx->dmerAt[*grpPtr] = dmerId;
    /* Dictionaries only help for the first reference to the dmer.
     * After that zstd can reference the match from the previous reference.
     * So only count each dmer once for each sample it is in.
     */
    if (*grpPtr < curSampleEnd) {
      continue;
    }
    freq += 1;
    /* Binary search to find the end of the sample *grpPtr is in.
     * In the common case that grpPtr + 1 == grpEnd we can skip the binary
     * search because the loop is over.
     */
    if (grpPtr + 1 != grpEnd) {
      const size_t *sampleEndPtr =
          COVER_lower_bound(curOffsetPtr, offsetsEnd, *grpPtr);
      curSampleEnd = *sampleEndPtr;
      curOffsetPtr = sampleEndPtr + 1;
    }
  }
  /* At this point we are never going to look at this segment of the suffix
   * array again.  We take advantage of this fact to save memory.
   * We store the frequency of the dmer in the first position of the group,
   * which is dmerId.
   */
  ctx->suffix[dmerId] = freq;
}


/**
 * Selects the best segment in an epoch.
 * Segments of are scored according to the function:
 *
 * Let F(d) be the frequency of dmer d.
 * Let S_i be the dmer at position i of segment S which has length k.
 *
 *     Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1})
 *
 * Once the dmer d is in the dictionary we set F(d) = 0.
 */
static COVER_segment_t COVER_selectSegment(const COVER_ctx_t *ctx, U32 *freqs,
                                           COVER_map_t *activeDmers, U32 begin,
                                           U32 end,
                                           ZDICT_cover_params_t parameters) {
  /* Constants */
  const U32 k = parameters.k;
  const U32 d = parameters.d;
  const U32 dmersInK = k - d + 1;
  /* Try each segment (activeSegment) and save the best (bestSegment) */
  COVER_segment_t bestSegment = {0, 0, 0};
  COVER_segment_t activeSegment;
  /* Reset the activeDmers in the segment */
  COVER_map_clear(activeDmers);
  /* The activeSegment starts at the beginning of the epoch. */
  activeSegment.begin = begin;
  activeSegment.end = begin;
  activeSegment.score = 0;
  /* Slide the activeSegment through the whole epoch.
   * Save the best segment in bestSegment.
   */
  while (activeSegment.end < end) {
    /* The dmerId for the dmer at the next position */
    U32 newDmer = ctx->dmerAt[activeSegment.end];
    /* The entry in activeDmers for this dmerId */
    U32 *newDmerOcc = COVER_map_at(activeDmers, newDmer);
    /* If the dmer isn't already present in the segment add its score. */
    if (*newDmerOcc == 0) {
      /* The paper suggest using the L-0.5 norm, but experiments show that it
       * doesn't help.
       */
      activeSegment.score += freqs[newDmer];
    }
    /* Add the dmer to the segment */
    activeSegment.end += 1;
    *newDmerOcc += 1;

    /* If the window is now too large, drop the first position */
    if (activeSegment.end - activeSegment.begin == dmersInK + 1) {
      U32 delDmer = ctx->dmerAt[activeSegment.begin];
      U32 *delDmerOcc = COVER_map_at(activeDmers, delDmer);
      activeSegment.begin += 1;
      *delDmerOcc -= 1;
      /* If this is the last occurrence of the dmer, subtract its score */
      if (*delDmerOcc == 0) {
        COVER_map_remove(activeDmers, delDmer);
        activeSegment.score -= freqs[delDmer];
      }
    }

    /* If this segment is the best so far save it */
    if (activeSegment.score > bestSegment.score) {
      bestSegment = activeSegment;
    }
  }
  {
    /* Trim off the zero frequency head and tail from the segment. */
    U32 newBegin = bestSegment.end;
    U32 newEnd = bestSegment.begin;
    U32 pos;
    for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
      U32 freq = freqs[ctx->dmerAt[pos]];
      if (freq != 0) {
        newBegin = MIN(newBegin, pos);
        newEnd = pos + 1;
      }
    }
    bestSegment.begin = newBegin;
    bestSegment.end = newEnd;
  }
  {
    /* Zero out the frequency of each dmer covered by the chosen segment. */
    U32 pos;
    for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
      freqs[ctx->dmerAt[pos]] = 0;
    }
  }
  return bestSegment;
}

/**
 * Check the validity of the parameters.
 * Returns non-zero if the parameters are valid and 0 otherwise.
 */
static int COVER_checkParameters(ZDICT_cover_params_t parameters,
                                 size_t maxDictSize) {
  /* k and d are required parameters */
  if (parameters.d == 0 || parameters.k == 0) {
    return 0;
  }
  /* k <= maxDictSize */
  if (parameters.k > maxDictSize) {
    return 0;
  }
  /* d <= k */
  if (parameters.d > parameters.k) {
    return 0;
  }
  /* 0 < splitPoint <= 1 */
  if (parameters.splitPoint <= 0 || parameters.splitPoint > 1){
    return 0;
  }
  return 1;
}

/**
 * Clean up a context initialized with `COVER_ctx_init()`.
 */
static void COVER_ctx_destroy(COVER_ctx_t *ctx) {
  if (!ctx) {
    return;
  }
  if (ctx->suffix) {
    free(ctx->suffix);
    ctx->suffix = NULL;
  }
  if (ctx->freqs) {
    free(ctx->freqs);
    ctx->freqs = NULL;
  }
  if (ctx->dmerAt) {
    free(ctx->dmerAt);
    ctx->dmerAt = NULL;
  }
  if (ctx->offsets) {
    free(ctx->offsets);
    ctx->offsets = NULL;
  }
}

/**
 * Prepare a context for dictionary building.
 * The context is only dependent on the parameter `d` and can be used multiple
 * times.
 * Returns 0 on success or error code on error.
 * The context must be destroyed with `COVER_ctx_destroy()`.
 */
static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
                          const size_t *samplesSizes, unsigned nbSamples,
                          unsigned d, double splitPoint)
{
  const BYTE *const samples = (const BYTE *)samplesBuffer;
  const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
  /* Split samples into testing and training sets */
  const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples;
  const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
  const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
  const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;
  /* Checks */
  if (totalSamplesSize < MAX(d, sizeof(U64)) ||
      totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) {
    DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
                 (unsigned)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20));
    return ERROR(srcSize_wrong);
  }
  /* Check if there are at least 5 training samples */
  if (nbTrainSamples < 5) {
    DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid.", nbTrainSamples);
    return ERROR(srcSize_wrong);
  }
  /* Check if there's testing sample */
  if (nbTestSamples < 1) {
    DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.", nbTestSamples);
    return ERROR(srcSize_wrong);
  }
  /* Zero the context */
  memset(ctx, 0, sizeof(*ctx));
  DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
               (unsigned)trainingSamplesSize);
  DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
               (unsigned)testSamplesSize);
  ctx->samples = samples;
  ctx->samplesSizes = samplesSizes;
  ctx->nbSamples = nbSamples;
  ctx->nbTrainSamples = nbTrainSamples;
  ctx->nbTestSamples = nbTestSamples;
  /* Partial suffix array */
  ctx->suffixSize = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
  ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
  /* Maps index to the dmerID */
  ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
  /* The offsets of each file */
  ctx->offsets = (size_t *)malloc((nbSamples + 1) * sizeof(size_t));
  if (!ctx->suffix || !ctx->dmerAt || !ctx->offsets) {
    DISPLAYLEVEL(1, "Failed to allocate scratch buffers\n");
    COVER_ctx_destroy(ctx);
    return ERROR(memory_allocation);
  }
  ctx->freqs = NULL;
  ctx->d = d;

  /* Fill offsets from the samplesSizes */
  {
    U32 i;
    ctx->offsets[0] = 0;
    for (i = 1; i <= nbSamples; ++i) {
      ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1];
    }
  }
  DISPLAYLEVEL(2, "Constructing partial suffix array\n");
  {
    /* suffix is a partial suffix array.
     * It only sorts suffixes by their first parameters.d bytes.
     * The sort is stable, so each dmer group is sorted by position in input.
     */
    U32 i;
    for (i = 0; i < ctx->suffixSize; ++i) {
      ctx->suffix[i] = i;
    }
    stableSort(ctx);
  }
  DISPLAYLEVEL(2, "Computing frequencies\n");
  /* For each dmer group (group of positions with the same first d bytes):
   * 1. For each position we set dmerAt[position] = dmerID.  The dmerID is
   *    (groupBeginPtr - suffix).  This allows us to go from position to
   *    dmerID so we can look up values in freq.
   * 2. We calculate how many samples the dmer occurs in and save it in
   *    freqs[dmerId].
   */
  COVER_groupBy(ctx->suffix, ctx->suffixSize, sizeof(U32), ctx,
                (ctx->d <= 8 ? &COVER_cmp8 : &COVER_cmp), &COVER_group);
  ctx->freqs = ctx->suffix;
  ctx->suffix = NULL;
  return 0;
}

void COVER_warnOnSmallCorpus(size_t maxDictSize, size_t nbDmers, int displayLevel)
{
  const double ratio = (double)nbDmers / (double)maxDictSize;
  if (ratio >= 10) {
      return;
  }
  LOCALDISPLAYLEVEL(displayLevel, 1,
                    "WARNING: The maximum dictionary size %u is too large "
                    "compared to the source size %u! "
                    "size(source)/size(dictionary) = %f, but it should be >= "
                    "10! This may lead to a subpar dictionary! We recommend "
                    "training on sources at least 10x, and preferably 100x "
                    "the size of the dictionary! \n", (U32)maxDictSize,
                    (U32)nbDmers, ratio);
}

COVER_epoch_info_t COVER_computeEpochs(U32 maxDictSize,
                                       U32 nbDmers, U32 k, U32 passes)
{
  const U32 minEpochSize = k * 10;
  COVER_epoch_info_t epochs;
  epochs.num = MAX(1, maxDictSize / k / passes);
  epochs.size = nbDmers / epochs.num;
  if (epochs.size >= minEpochSize) {
      assert(epochs.size * epochs.num <= nbDmers);
      return epochs;
  }
  epochs.size = MIN(minEpochSize, nbDmers);
  epochs.num = nbDmers / epochs.size;
  assert(epochs.size * epochs.num <= nbDmers);
  return epochs;
}

/**
 * Given the prepared context build the dictionary.
 */
static size_t COVER_buildDictionary(const COVER_ctx_t *ctx, U32 *freqs,
                                    COVER_map_t *activeDmers, void *dictBuffer,
                                    size_t dictBufferCapacity,
                                    ZDICT_cover_params_t parameters) {
  BYTE *const dict = (BYTE *)dictBuffer;
  size_t tail = dictBufferCapacity;
  /* Divide the data into epochs. We will select one segment from each epoch. */
  const COVER_epoch_info_t epochs = COVER_computeEpochs(
      (U32)dictBufferCapacity, (U32)ctx->suffixSize, parameters.k, 4);
  const size_t maxZeroScoreRun = MAX(10, MIN(100, epochs.num >> 3));
  size_t zeroScoreRun = 0;
  size_t epoch;
  DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n",
                (U32)epochs.num, (U32)epochs.size);
  /* Loop through the epochs until there are no more segments or the dictionary
   * is full.
   */
  for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs.num) {
    const U32 epochBegin = (U32)(epoch * epochs.size);
    const U32 epochEnd = epochBegin + epochs.size;
    size_t segmentSize;
    /* Select a segment */
    COVER_segment_t segment = COVER_selectSegment(
        ctx, freqs, activeDmers, epochBegin, epochEnd, parameters);
    /* If the segment covers no dmers, then we are out of content.
     * There may be new content in other epochs, for continue for some time.
     */
    if (segment.score == 0) {
      if (++zeroScoreRun >= maxZeroScoreRun) {
          break;
      }
      continue;
    }
    zeroScoreRun = 0;
    /* Trim the segment if necessary and if it is too small then we are done */
    segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail);
    if (segmentSize < parameters.d) {
      break;
    }
    /* We fill the dictionary from the back to allow the best segments to be
     * referenced with the smallest offsets.
     */
    tail -= segmentSize;
    memcpy(dict + tail, ctx->samples + segment.begin, segmentSize);
    DISPLAYUPDATE(
        2, "\r%u%%       ",
        (unsigned)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
  }
  DISPLAYLEVEL(2, "\r%79s\r", "");
  return tail;
}

ZDICTLIB_STATIC_API size_t ZDICT_trainFromBuffer_cover(
    void *dictBuffer, size_t dictBufferCapacity,
    const void *samplesBuffer, const size_t *samplesSizes, unsigned nbSamples,
    ZDICT_cover_params_t parameters)
{
  BYTE* const dict = (BYTE*)dictBuffer;
  COVER_ctx_t ctx;
  COVER_map_t activeDmers;
  parameters.splitPoint = 1.0;
  /* Initialize global data */
  g_displayLevel = (int)parameters.zParams.notificationLevel;
  /* Checks */
  if (!COVER_checkParameters(parameters, dictBufferCapacity)) {
    DISPLAYLEVEL(1, "Cover parameters incorrect\n");
    return ERROR(parameter_outOfBound);
  }
  if (nbSamples == 0) {
    DISPLAYLEVEL(1, "Cover must have at least one input file\n");
    return ERROR(srcSize_wrong);
  }
  if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
    DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
                 ZDICT_DICTSIZE_MIN);
    return ERROR(dstSize_tooSmall);
  }
  /* Initialize context and activeDmers */
  {
    size_t const initVal = COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
                      parameters.d, parameters.splitPoint);
    if (ZSTD_isError(initVal)) {
      return initVal;
    }
  }
  COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.suffixSize, g_displayLevel);
  if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
    DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
    COVER_ctx_destroy(&ctx);
    return ERROR(memory_allocation);
  }

  DISPLAYLEVEL(2, "Building dictionary\n");
  {
    const size_t tail =
        COVER_buildDictionary(&ctx, ctx.freqs, &activeDmers, dictBuffer,
                              dictBufferCapacity, parameters);
    const size_t dictionarySize = ZDICT_finalizeDictionary(
        dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
        samplesBuffer, samplesSizes, nbSamples, parameters.zParams);
    if (!ZSTD_isError(dictionarySize)) {
      DISPLAYLEVEL(2, "Constructed dictionary of size %u\n",
                   (unsigned)dictionarySize);
    }
    COVER_ctx_destroy(&ctx);
    COVER_map_destroy(&activeDmers);
    return dictionarySize;
  }
}


size_t COVER_checkTotalCompressedSize(const ZDICT_cover_params_t parameters,
                                    const size_t *samplesSizes, const BYTE *samples,
                                    size_t *offsets,
                                    size_t nbTrainSamples, size_t nbSamples,
                                    BYTE *const dict, size_t dictBufferCapacity) {
  size_t totalCompressedSize = ERROR(GENERIC);
  /* Pointers */
  ZSTD_CCtx *cctx;
  ZSTD_CDict *cdict;
  void *dst;
  /* Local variables */
  size_t dstCapacity;
  size_t i;
  /* Allocate dst with enough space to compress the maximum sized sample */
  {
    size_t maxSampleSize = 0;
    i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
    for (; i < nbSamples; ++i) {
      maxSampleSize = MAX(samplesSizes[i], maxSampleSize);
    }
    dstCapacity = ZSTD_compressBound(maxSampleSize);
    dst = malloc(dstCapacity);
  }
  /* Create the cctx and cdict */
  cctx = ZSTD_createCCtx();
  cdict = ZSTD_createCDict(dict, dictBufferCapacity,
                           parameters.zParams.compressionLevel);
  if (!dst || !cctx || !cdict) {
    goto _compressCleanup;
  }
  /* Compress each sample and sum their sizes (or error) */
  totalCompressedSize = dictBufferCapacity;
  i = parameters.splitPoint < 1.0 ? nbTrainSamples : 0;
  for (; i < nbSamples; ++i) {
    const size_t size = ZSTD_compress_usingCDict(
        cctx, dst, dstCapacity, samples + offsets[i],
        samplesSizes[i], cdict);
    if (ZSTD_isError(size)) {
      totalCompressedSize = size;
      goto _compressCleanup;
    }
    totalCompressedSize += size;
  }
_compressCleanup:
  ZSTD_freeCCtx(cctx);
  ZSTD_freeCDict(cdict);
  if (dst) {
    free(dst);
  }
  return totalCompressedSize;
}


/**
 * Initialize the `COVER_best_t`.
 */
void COVER_best_init(COVER_best_t *best) {
  if (best==NULL) return; /* compatible with init on NULL */
  (void)ZSTD_pthread_mutex_init(&best->mutex, NULL);
  (void)ZSTD_pthread_cond_init(&best->cond, NULL);
  best->liveJobs = 0;
  best->dict = NULL;
  best->dictSize = 0;
  best->compressedSize = (size_t)-1;
  memset(&best->parameters, 0, sizeof(best->parameters));
}

/**
 * Wait until liveJobs == 0.
 */
void COVER_best_wait(COVER_best_t *best) {
  if (!best) {
    return;
  }
  ZSTD_pthread_mutex_lock(&best->mutex);
  while (best->liveJobs != 0) {
    ZSTD_pthread_cond_wait(&best->cond, &best->mutex);
  }
  ZSTD_pthread_mutex_unlock(&best->mutex);
}

/**
 * Call COVER_best_wait() and then destroy the COVER_best_t.
 */
void COVER_best_destroy(COVER_best_t *best) {
  if (!best) {
    return;
  }
  COVER_best_wait(best);
  if (best->dict) {
    free(best->dict);
  }
  ZSTD_pthread_mutex_destroy(&best->mutex);
  ZSTD_pthread_cond_destroy(&best->cond);
}

/**
 * Called when a thread is about to be launched.
 * Increments liveJobs.
 */
void COVER_best_start(COVER_best_t *best) {
  if (!best) {
    return;
  }
  ZSTD_pthread_mutex_lock(&best->mutex);
  ++best->liveJobs;
  ZSTD_pthread_mutex_unlock(&best->mutex);
}

/**
 * Called when a thread finishes executing, both on error or success.
 * Decrements liveJobs and signals any waiting threads if liveJobs == 0.
 * If this dictionary is the best so far save it and its parameters.
 */
void COVER_best_finish(COVER_best_t* best,
                      ZDICT_cover_params_t parameters,
                      COVER_dictSelection_t selection)
{
  void* dict = selection.dictContent;
  size_t compressedSize = selection.totalCompressedSize;
  size_t dictSize = selection.dictSize;
  if (!best) {
    return;
  }
  {
    size_t liveJobs;
    ZSTD_pthread_mutex_lock(&best->mutex);
    --best->liveJobs;
    liveJobs = best->liveJobs;
    /* If the new dictionary is better */
    if (compressedSize < best->compressedSize) {
      /* Allocate space if necessary */
      if (!best->dict || best->dictSize < dictSize) {
        if (best->dict) {
          free(best->dict);
        }
        best->dict = malloc(dictSize);
        if (!best->dict) {
          best->compressedSize = ERROR(GENERIC);
          best->dictSize = 0;
          ZSTD_pthread_cond_signal(&best->cond);
          ZSTD_pthread_mutex_unlock(&best->mutex);
          return;
        }
      }
      /* Save the dictionary, parameters, and size */
      if (dict) {
        memcpy(best->dict, dict, dictSize);
        best->dictSize = dictSize;
        best->parameters = parameters;
        best->compressedSize = compressedSize;
      }
    }
    if (liveJobs == 0) {
      ZSTD_pthread_cond_broadcast(&best->cond);
    }
    ZSTD_pthread_mutex_unlock(&best->mutex);
  }
}

static COVER_dictSelection_t setDictSelection(BYTE* buf, size_t s, size_t csz)
{
    COVER_dictSelection_t ds;
    ds.dictContent = buf;
    ds.dictSize = s;
    ds.totalCompressedSize = csz;
    return ds;
}

COVER_dictSelection_t COVER_dictSelectionError(size_t error) {
    return setDictSelection(NULL, 0, error);
}

unsigned COVER_dictSelectionIsError(COVER_dictSelection_t selection) {
  return (ZSTD_isError(selection.totalCompressedSize) || !selection.dictContent);
}

void COVER_dictSelectionFree(COVER_dictSelection_t selection){
  free(selection.dictContent);
}

COVER_dictSelection_t COVER_selectDict(BYTE* customDictContent, size_t dictBufferCapacity,
        size_t dictContentSize, const BYTE* samplesBuffer, const size_t* samplesSizes, unsigned nbFinalizeSamples,
        size_t nbCheckSamples, size_t nbSamples, ZDICT_cover_params_t params, size_t* offsets, size_t totalCompressedSize) {

  size_t largestDict = 0;
  size_t largestCompressed = 0;
  BYTE* customDictContentEnd = customDictContent + dictContentSize;

  BYTE* largestDictbuffer = (BYTE*)malloc(dictBufferCapacity);
  BYTE* candidateDictBuffer = (BYTE*)malloc(dictBufferCapacity);
  double regressionTolerance = ((double)params.shrinkDictMaxRegression / 100.0) + 1.00;

  if (!largestDictbuffer || !candidateDictBuffer) {
    free(largestDictbuffer);
    free(candidateDictBuffer);
    return COVER_dictSelectionError(dictContentSize);
  }

  /* Initial dictionary size and compressed size */
  memcpy(largestDictbuffer, customDictContent, dictContentSize);
  dictContentSize = ZDICT_finalizeDictionary(
    largestDictbuffer, dictBufferCapacity, customDictContent, dictContentSize,
    samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams);

  if (ZDICT_isError(dictContentSize)) {
    free(largestDictbuffer);
    free(candidateDictBuffer);
    return COVER_dictSelectionError(dictContentSize);
  }

  totalCompressedSize = COVER_checkTotalCompressedSize(params, samplesSizes,
                                                       samplesBuffer, offsets,
                                                       nbCheckSamples, nbSamples,
                                                       largestDictbuffer, dictContentSize);

  if (ZSTD_isError(totalCompressedSize)) {
    free(largestDictbuffer);
    free(candidateDictBuffer);
    return COVER_dictSelectionError(totalCompressedSize);
  }

  if (params.shrinkDict == 0) {
    free(candidateDictBuffer);
    return setDictSelection(largestDictbuffer, dictContentSize, totalCompressedSize);
  }

  largestDict = dictContentSize;
  largestCompressed = totalCompressedSize;
  dictContentSize = ZDICT_DICTSIZE_MIN;

  /* Largest dict is initially at least ZDICT_DICTSIZE_MIN */
  while (dictContentSize < largestDict) {
    memcpy(candidateDictBuffer, largestDictbuffer, largestDict);
    dictContentSize = ZDICT_finalizeDictionary(
      candidateDictBuffer, dictBufferCapacity, customDictContentEnd - dictContentSize, dictContentSize,
      samplesBuffer, samplesSizes, nbFinalizeSamples, params.zParams);

    if (ZDICT_isError(dictContentSize)) {
      free(largestDictbuffer);
      free(candidateDictBuffer);
      return COVER_dictSelectionError(dictContentSize);

    }

    totalCompressedSize = COVER_checkTotalCompressedSize(params, samplesSizes,
                                                         samplesBuffer, offsets,
                                                         nbCheckSamples, nbSamples,
                                                         candidateDictBuffer, dictContentSize);

    if (ZSTD_isError(totalCompressedSize)) {
      free(largestDictbuffer);
      free(candidateDictBuffer);
      return COVER_dictSelectionError(totalCompressedSize);
    }

    if ((double)totalCompressedSize <= (double)largestCompressed * regressionTolerance) {
      free(largestDictbuffer);
      return setDictSelection( candidateDictBuffer, dictContentSize, totalCompressedSize );
    }
    dictContentSize *= 2;
  }
  dictContentSize = largestDict;
  totalCompressedSize = largestCompressed;
  free(candidateDictBuffer);
  return setDictSelection( largestDictbuffer, dictContentSize, totalCompressedSize );
}

/**
 * Parameters for COVER_tryParameters().
 */
typedef struct COVER_tryParameters_data_s {
  const COVER_ctx_t *ctx;
  COVER_best_t *best;
  size_t dictBufferCapacity;
  ZDICT_cover_params_t parameters;
} COVER_tryParameters_data_t;

/**
 * Tries a set of parameters and updates the COVER_best_t with the results.
 * This function is thread safe if zstd is compiled with multithreaded support.
 * It takes its parameters as an *OWNING* opaque pointer to support threading.
 */
static void COVER_tryParameters(void *opaque)
{
  /* Save parameters as local variables */
  COVER_tryParameters_data_t *const data = (COVER_tryParameters_data_t*)opaque;
  const COVER_ctx_t *const ctx = data->ctx;
  const ZDICT_cover_params_t parameters = data->parameters;
  size_t dictBufferCapacity = data->dictBufferCapacity;
  size_t totalCompressedSize = ERROR(GENERIC);
  /* Allocate space for hash table, dict, and freqs */
  COVER_map_t activeDmers;
  BYTE* const dict = (BYTE*)malloc(dictBufferCapacity);
  COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
  U32* const freqs = (U32*)malloc(ctx->suffixSize * sizeof(U32));
  if (!COVER_map_init(&activeDmers, parameters.k - parameters.d + 1)) {
    DISPLAYLEVEL(1, "Failed to allocate dmer map: out of memory\n");
    goto _cleanup;
  }
  if (!dict || !freqs) {
    DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
    goto _cleanup;
  }
  /* Copy the frequencies because we need to modify them */
  memcpy(freqs, ctx->freqs, ctx->suffixSize * sizeof(U32));
  /* Build the dictionary */
  {
    const size_t tail = COVER_buildDictionary(ctx, freqs, &activeDmers, dict,
                                              dictBufferCapacity, parameters);
    selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail,
        ctx->samples, ctx->samplesSizes, (unsigned)ctx->nbTrainSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets,
        totalCompressedSize);

    if (COVER_dictSelectionIsError(selection)) {
      DISPLAYLEVEL(1, "Failed to select dictionary\n");
      goto _cleanup;
    }
  }
_cleanup:
  free(dict);
  COVER_best_finish(data->best, parameters, selection);
  free(data);
  COVER_map_destroy(&activeDmers);
  COVER_dictSelectionFree(selection);
  free(freqs);
}

ZDICTLIB_STATIC_API size_t ZDICT_optimizeTrainFromBuffer_cover(
    void* dictBuffer, size_t dictBufferCapacity, const void* samplesBuffer,
    const size_t* samplesSizes, unsigned nbSamples,
    ZDICT_cover_params_t* parameters)
{
  /* constants */
  const unsigned nbThreads = parameters->nbThreads;
  const double splitPoint =
      parameters->splitPoint <= 0.0 ? COVER_DEFAULT_SPLITPOINT : parameters->splitPoint;
  const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
  const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
  const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
  const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k;
  const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps;
  const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1);
  const unsigned kIterations =
      (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize);
  const unsigned shrinkDict = 0;
  /* Local variables */
  const int displayLevel = parameters->zParams.notificationLevel;
  unsigned iteration = 1;
  unsigned d;
  unsigned k;
  COVER_best_t best;
  POOL_ctx *pool = NULL;
  int warned = 0;

  /* Checks */
  if (splitPoint <= 0 || splitPoint > 1) {
    LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
    return ERROR(parameter_outOfBound);
  }
  if (kMinK < kMaxD || kMaxK < kMinK) {
    LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect parameters\n");
    return ERROR(parameter_outOfBound);
  }
  if (nbSamples == 0) {
    DISPLAYLEVEL(1, "Cover must have at least one input file\n");
    return ERROR(srcSize_wrong);
  }
  if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
    DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
                 ZDICT_DICTSIZE_MIN);
    return ERROR(dstSize_tooSmall);
  }
  if (nbThreads > 1) {
    pool = POOL_create(nbThreads, 1);
    if (!pool) {
      return ERROR(memory_allocation);
    }
  }
  /* Initialization */
  COVER_best_init(&best);
  /* Turn down global display level to clean up display at level 2 and below */
  g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1;
  /* Loop through d first because each new value needs a new context */
  LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n",
                    kIterations);
  for (d = kMinD; d <= kMaxD; d += 2) {
    /* Initialize the context for this value of d */
    COVER_ctx_t ctx;
    LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
    {
      const size_t initVal = COVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint);
      if (ZSTD_isError(initVal)) {
        LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
        COVER_best_destroy(&best);
        POOL_free(pool);
        return initVal;
      }
    }
    if (!warned) {
      COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.suffixSize, displayLevel);
      warned = 1;
    }
    /* Loop through k reusing the same context */
    for (k = kMinK; k <= kMaxK; k += kStepSize) {
      /* Prepare the arguments */
      COVER_tryParameters_data_t *data = (COVER_tryParameters_data_t *)malloc(
          sizeof(COVER_tryParameters_data_t));
      LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k);
      if (!data) {
        LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n");
        COVER_best_destroy(&best);
        COVER_ctx_destroy(&ctx);
        POOL_free(pool);
        return ERROR(memory_allocation);
      }
      data->ctx = &ctx;
      data->best = &best;
      data->dictBufferCapacity = dictBufferCapacity;
      data->parameters = *parameters;
      data->parameters.k = k;
      data->parameters.d = d;
      data->parameters.splitPoint = splitPoint;
      data->parameters.steps = kSteps;
      data->parameters.shrinkDict = shrinkDict;
      data->parameters.zParams.notificationLevel = g_displayLevel;
      /* Check the parameters */
      if (!COVER_checkParameters(data->parameters, dictBufferCapacity)) {
        DISPLAYLEVEL(1, "Cover parameters incorrect\n");
        free(data);
        continue;
      }
      /* Call the function and pass ownership of data to it */
      COVER_best_start(&best);
      if (pool) {
        POOL_add(pool, &COVER_tryParameters, data);
      } else {
        COVER_tryParameters(data);
      }
      /* Print status */
      LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%%       ",
                         (unsigned)((iteration * 100) / kIterations));
      ++iteration;
    }
    COVER_best_wait(&best);
    COVER_ctx_destroy(&ctx);
  }
  LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", "");
  /* Fill the output buffer and parameters with output of the best parameters */
  {
    const size_t dictSize = best.dictSize;
    if (ZSTD_isError(best.compressedSize)) {
      const size_t compressedSize = best.compressedSize;
      COVER_best_destroy(&best);
      POOL_free(pool);
      return compressedSize;
    }
    *parameters = best.parameters;
    memcpy(dictBuffer, best.dict, dictSize);
    COVER_best_destroy(&best);
    POOL_free(pool);
    return dictSize;
  }
}
/**** ended inlining dictBuilder/cover.c ****/
/**** start inlining dictBuilder/divsufsort.c ****/
/*
 * divsufsort.c for libdivsufsort-lite
 * Copyright (c) 2003-2008 Yuta Mori All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person
 * obtaining a copy of this software and associated documentation
 * files (the "Software"), to deal in the Software without
 * restriction, including without limitation the rights to use,
 * copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following
 * conditions:
 *
 * The above copyright notice and this permission notice shall be
 * included in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
 * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 * OTHER DEALINGS IN THE SOFTWARE.
 */

/*- Compiler specifics -*/
#ifdef __clang__
#pragma clang diagnostic ignored "-Wshorten-64-to-32"
#endif

#if defined(_MSC_VER)
#  pragma warning(disable : 4244)
#  pragma warning(disable : 4127)    /* C4127 : Condition expression is constant */
#endif


/*- Dependencies -*/
#include <assert.h>
#include <stdio.h>
#include <stdlib.h>

/**** start inlining divsufsort.h ****/
/*
 * divsufsort.h for libdivsufsort-lite
 * Copyright (c) 2003-2008 Yuta Mori All Rights Reserved.
 *
 * Permission is hereby granted, free of charge, to any person
 * obtaining a copy of this software and associated documentation
 * files (the "Software"), to deal in the Software without
 * restriction, including without limitation the rights to use,
 * copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the
 * Software is furnished to do so, subject to the following
 * conditions:
 *
 * The above copyright notice and this permission notice shall be
 * included in all copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
 * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
 * OTHER DEALINGS IN THE SOFTWARE.
 */

#ifndef _DIVSUFSORT_H
#define _DIVSUFSORT_H 1

/*- Prototypes -*/

/**
 * Constructs the suffix array of a given string.
 * @param T [0..n-1] The input string.
 * @param SA [0..n-1] The output array of suffixes.
 * @param n The length of the given string.
 * @param openMP enables OpenMP optimization.
 * @return 0 if no error occurred, -1 or -2 otherwise.
 */
int
divsufsort(const unsigned char *T, int *SA, int n, int openMP);

/**
 * Constructs the burrows-wheeler transformed string of a given string.
 * @param T [0..n-1] The input string.
 * @param U [0..n-1] The output string. (can be T)
 * @param A [0..n-1] The temporary array. (can be NULL)
 * @param n The length of the given string.
 * @param num_indexes The length of secondary indexes array. (can be NULL)
 * @param indexes The secondary indexes array. (can be NULL)
 * @param openMP enables OpenMP optimization.
 * @return The primary index if no error occurred, -1 or -2 otherwise.
 */
int
divbwt(const unsigned char *T, unsigned char *U, int *A, int n, unsigned char * num_indexes, int * indexes, int openMP);

#endif /* _DIVSUFSORT_H */
/**** ended inlining divsufsort.h ****/

/*- Constants -*/
#if defined(INLINE)
# undef INLINE
#endif
#if !defined(INLINE)
# define INLINE __inline
#endif
#if defined(ALPHABET_SIZE) && (ALPHABET_SIZE < 1)
# undef ALPHABET_SIZE
#endif
#if !defined(ALPHABET_SIZE)
# define ALPHABET_SIZE (256)
#endif
#define BUCKET_A_SIZE (ALPHABET_SIZE)
#define BUCKET_B_SIZE (ALPHABET_SIZE * ALPHABET_SIZE)
#if defined(SS_INSERTIONSORT_THRESHOLD)
# if SS_INSERTIONSORT_THRESHOLD < 1
#  undef SS_INSERTIONSORT_THRESHOLD
#  define SS_INSERTIONSORT_THRESHOLD (1)
# endif
#else
# define SS_INSERTIONSORT_THRESHOLD (8)
#endif
#if defined(SS_BLOCKSIZE)
# if SS_BLOCKSIZE < 0
#  undef SS_BLOCKSIZE
#  define SS_BLOCKSIZE (0)
# elif 32768 <= SS_BLOCKSIZE
#  undef SS_BLOCKSIZE
#  define SS_BLOCKSIZE (32767)
# endif
#else
# define SS_BLOCKSIZE (1024)
#endif
/* minstacksize = log(SS_BLOCKSIZE) / log(3) * 2 */
#if SS_BLOCKSIZE == 0
# define SS_MISORT_STACKSIZE (96)
#elif SS_BLOCKSIZE <= 4096
# define SS_MISORT_STACKSIZE (16)
#else
# define SS_MISORT_STACKSIZE (24)
#endif
#define SS_SMERGE_STACKSIZE (32)
#define TR_INSERTIONSORT_THRESHOLD (8)
#define TR_STACKSIZE (64)


/*- Macros -*/
#ifndef SWAP
# define SWAP(_a, _b) do { t = (_a); (_a) = (_b); (_b) = t; } while(0)
#endif /* SWAP */
#ifndef MIN
# define MIN(_a, _b) (((_a) < (_b)) ? (_a) : (_b))
#endif /* MIN */
#ifndef MAX
# define MAX(_a, _b) (((_a) > (_b)) ? (_a) : (_b))
#endif /* MAX */
#define STACK_PUSH(_a, _b, _c, _d)\
  do {\
    assert(ssize < STACK_SIZE);\
    stack[ssize].a = (_a), stack[ssize].b = (_b),\
    stack[ssize].c = (_c), stack[ssize++].d = (_d);\
  } while(0)
#define STACK_PUSH5(_a, _b, _c, _d, _e)\
  do {\
    assert(ssize < STACK_SIZE);\
    stack[ssize].a = (_a), stack[ssize].b = (_b),\
    stack[ssize].c = (_c), stack[ssize].d = (_d), stack[ssize++].e = (_e);\
  } while(0)
#define STACK_POP(_a, _b, _c, _d)\
  do {\
    assert(0 <= ssize);\
    if(ssize == 0) { return; }\
    (_a) = stack[--ssize].a, (_b) = stack[ssize].b,\
    (_c) = stack[ssize].c, (_d) = stack[ssize].d;\
  } while(0)
#define STACK_POP5(_a, _b, _c, _d, _e)\
  do {\
    assert(0 <= ssize);\
    if(ssize == 0) { return; }\
    (_a) = stack[--ssize].a, (_b) = stack[ssize].b,\
    (_c) = stack[ssize].c, (_d) = stack[ssize].d, (_e) = stack[ssize].e;\
  } while(0)
#define BUCKET_A(_c0) bucket_A[(_c0)]
#if ALPHABET_SIZE == 256
#define BUCKET_B(_c0, _c1) (bucket_B[((_c1) << 8) | (_c0)])
#define BUCKET_BSTAR(_c0, _c1) (bucket_B[((_c0) << 8) | (_c1)])
#else
#define BUCKET_B(_c0, _c1) (bucket_B[(_c1) * ALPHABET_SIZE + (_c0)])
#define BUCKET_BSTAR(_c0, _c1) (bucket_B[(_c0) * ALPHABET_SIZE + (_c1)])
#endif


/*- Private Functions -*/

static const int lg_table[256]= {
 -1,0,1,1,2,2,2,2,3,3,3,3,3,3,3,3,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,4,
  5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,
  6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,
  6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,6,
  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,
  7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7
};

#if (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE)

static INLINE
int
ss_ilg(int n) {
#if SS_BLOCKSIZE == 0
  return (n & 0xffff0000) ?
          ((n & 0xff000000) ?
            24 + lg_table[(n >> 24) & 0xff] :
            16 + lg_table[(n >> 16) & 0xff]) :
          ((n & 0x0000ff00) ?
             8 + lg_table[(n >>  8) & 0xff] :
             0 + lg_table[(n >>  0) & 0xff]);
#elif SS_BLOCKSIZE < 256
  return lg_table[n];
#else
  return (n & 0xff00) ?
          8 + lg_table[(n >> 8) & 0xff] :
          0 + lg_table[(n >> 0) & 0xff];
#endif
}

#endif /* (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE) */

#if SS_BLOCKSIZE != 0

static const int sqq_table[256] = {
  0,  16,  22,  27,  32,  35,  39,  42,  45,  48,  50,  53,  55,  57,  59,  61,
 64,  65,  67,  69,  71,  73,  75,  76,  78,  80,  81,  83,  84,  86,  87,  89,
 90,  91,  93,  94,  96,  97,  98,  99, 101, 102, 103, 104, 106, 107, 108, 109,
110, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126,
128, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142,
143, 144, 144, 145, 146, 147, 148, 149, 150, 150, 151, 152, 153, 154, 155, 155,
156, 157, 158, 159, 160, 160, 161, 162, 163, 163, 164, 165, 166, 167, 167, 168,
169, 170, 170, 171, 172, 173, 173, 174, 175, 176, 176, 177, 178, 178, 179, 180,
181, 181, 182, 183, 183, 184, 185, 185, 186, 187, 187, 188, 189, 189, 190, 191,
192, 192, 193, 193, 194, 195, 195, 196, 197, 197, 198, 199, 199, 200, 201, 201,
202, 203, 203, 204, 204, 205, 206, 206, 207, 208, 208, 209, 209, 210, 211, 211,
212, 212, 213, 214, 214, 215, 215, 216, 217, 217, 218, 218, 219, 219, 220, 221,
221, 222, 222, 223, 224, 224, 225, 225, 226, 226, 227, 227, 228, 229, 229, 230,
230, 231, 231, 232, 232, 233, 234, 234, 235, 235, 236, 236, 237, 237, 238, 238,
239, 240, 240, 241, 241, 242, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247,
247, 248, 248, 249, 249, 250, 250, 251, 251, 252, 252, 253, 253, 254, 254, 255
};

static INLINE
int
ss_isqrt(int x) {
  int y, e;

  if(x >= (SS_BLOCKSIZE * SS_BLOCKSIZE)) { return SS_BLOCKSIZE; }
  e = (x & 0xffff0000) ?
        ((x & 0xff000000) ?
          24 + lg_table[(x >> 24) & 0xff] :
          16 + lg_table[(x >> 16) & 0xff]) :
        ((x & 0x0000ff00) ?
           8 + lg_table[(x >>  8) & 0xff] :
           0 + lg_table[(x >>  0) & 0xff]);

  if(e >= 16) {
    y = sqq_table[x >> ((e - 6) - (e & 1))] << ((e >> 1) - 7);
    if(e >= 24) { y = (y + 1 + x / y) >> 1; }
    y = (y + 1 + x / y) >> 1;
  } else if(e >= 8) {
    y = (sqq_table[x >> ((e - 6) - (e & 1))] >> (7 - (e >> 1))) + 1;
  } else {
    return sqq_table[x] >> 4;
  }

  return (x < (y * y)) ? y - 1 : y;
}

#endif /* SS_BLOCKSIZE != 0 */


/*---------------------------------------------------------------------------*/

/* Compares two suffixes. */
static INLINE
int
ss_compare(const unsigned char *T,
           const int *p1, const int *p2,
           int depth) {
  const unsigned char *U1, *U2, *U1n, *U2n;

  for(U1 = T + depth + *p1,
      U2 = T + depth + *p2,
      U1n = T + *(p1 + 1) + 2,
      U2n = T + *(p2 + 1) + 2;
      (U1 < U1n) && (U2 < U2n) && (*U1 == *U2);
      ++U1, ++U2) {
  }

  return U1 < U1n ?
        (U2 < U2n ? *U1 - *U2 : 1) :
        (U2 < U2n ? -1 : 0);
}


/*---------------------------------------------------------------------------*/

#if (SS_BLOCKSIZE != 1) && (SS_INSERTIONSORT_THRESHOLD != 1)

/* Insertionsort for small size groups */
static
void
ss_insertionsort(const unsigned char *T, const int *PA,
                 int *first, int *last, int depth) {
  int *i, *j;
  int t;
  int r;

  for(i = last - 2; first <= i; --i) {
    for(t = *i, j = i + 1; 0 < (r = ss_compare(T, PA + t, PA + *j, depth));) {
      do { *(j - 1) = *j; } while((++j < last) && (*j < 0));
      if(last <= j) { break; }
    }
    if(r == 0) { *j = ~*j; }
    *(j - 1) = t;
  }
}

#endif /* (SS_BLOCKSIZE != 1) && (SS_INSERTIONSORT_THRESHOLD != 1) */


/*---------------------------------------------------------------------------*/

#if (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE)

static INLINE
void
ss_fixdown(const unsigned char *Td, const int *PA,
           int *SA, int i, int size) {
  int j, k;
  int v;
  int c, d, e;

  for(v = SA[i], c = Td[PA[v]]; (j = 2 * i + 1) < size; SA[i] = SA[k], i = k) {
    d = Td[PA[SA[k = j++]]];
    if(d < (e = Td[PA[SA[j]]])) { k = j; d = e; }
    if(d <= c) { break; }
  }
  SA[i] = v;
}

/* Simple top-down heapsort. */
static
void
ss_heapsort(const unsigned char *Td, const int *PA, int *SA, int size) {
  int i, m;
  int t;

  m = size;
  if((size % 2) == 0) {
    m--;
    if(Td[PA[SA[m / 2]]] < Td[PA[SA[m]]]) { SWAP(SA[m], SA[m / 2]); }
  }

  for(i = m / 2 - 1; 0 <= i; --i) { ss_fixdown(Td, PA, SA, i, m); }
  if((size % 2) == 0) { SWAP(SA[0], SA[m]); ss_fixdown(Td, PA, SA, 0, m); }
  for(i = m - 1; 0 < i; --i) {
    t = SA[0], SA[0] = SA[i];
    ss_fixdown(Td, PA, SA, 0, i);
    SA[i] = t;
  }
}


/*---------------------------------------------------------------------------*/

/* Returns the median of three elements. */
static INLINE
int *
ss_median3(const unsigned char *Td, const int *PA,
           int *v1, int *v2, int *v3) {
  int *t;
  if(Td[PA[*v1]] > Td[PA[*v2]]) { SWAP(v1, v2); }
  if(Td[PA[*v2]] > Td[PA[*v3]]) {
    if(Td[PA[*v1]] > Td[PA[*v3]]) { return v1; }
    else { return v3; }
  }
  return v2;
}

/* Returns the median of five elements. */
static INLINE
int *
ss_median5(const unsigned char *Td, const int *PA,
           int *v1, int *v2, int *v3, int *v4, int *v5) {
  int *t;
  if(Td[PA[*v2]] > Td[PA[*v3]]) { SWAP(v2, v3); }
  if(Td[PA[*v4]] > Td[PA[*v5]]) { SWAP(v4, v5); }
  if(Td[PA[*v2]] > Td[PA[*v4]]) { SWAP(v2, v4); SWAP(v3, v5); }
  if(Td[PA[*v1]] > Td[PA[*v3]]) { SWAP(v1, v3); }
  if(Td[PA[*v1]] > Td[PA[*v4]]) { SWAP(v1, v4); SWAP(v3, v5); }
  if(Td[PA[*v3]] > Td[PA[*v4]]) { return v4; }
  return v3;
}

/* Returns the pivot element. */
static INLINE
int *
ss_pivot(const unsigned char *Td, const int *PA, int *first, int *last) {
  int *middle;
  int t;

  t = last - first;
  middle = first + t / 2;

  if(t <= 512) {
    if(t <= 32) {
      return ss_median3(Td, PA, first, middle, last - 1);
    } else {
      t >>= 2;
      return ss_median5(Td, PA, first, first + t, middle, last - 1 - t, last - 1);
    }
  }
  t >>= 3;
  first  = ss_median3(Td, PA, first, first + t, first + (t << 1));
  middle = ss_median3(Td, PA, middle - t, middle, middle + t);
  last   = ss_median3(Td, PA, last - 1 - (t << 1), last - 1 - t, last - 1);
  return ss_median3(Td, PA, first, middle, last);
}


/*---------------------------------------------------------------------------*/

/* Binary partition for substrings. */
static INLINE
int *
ss_partition(const int *PA,
                    int *first, int *last, int depth) {
  int *a, *b;
  int t;
  for(a = first - 1, b = last;;) {
    for(; (++a < b) && ((PA[*a] + depth) >= (PA[*a + 1] + 1));) { *a = ~*a; }
    for(; (a < --b) && ((PA[*b] + depth) <  (PA[*b + 1] + 1));) { }
    if(b <= a) { break; }
    t = ~*b;
    *b = *a;
    *a = t;
  }
  if(first < a) { *first = ~*first; }
  return a;
}

/* Multikey introsort for medium size groups. */
static
void
ss_mintrosort(const unsigned char *T, const int *PA,
              int *first, int *last,
              int depth) {
#define STACK_SIZE SS_MISORT_STACKSIZE
  struct { int *a, *b, c; int d; } stack[STACK_SIZE];
  const unsigned char *Td;
  int *a, *b, *c, *d, *e, *f;
  int s, t;
  int ssize;
  int limit;
  int v, x = 0;

  for(ssize = 0, limit = ss_ilg(last - first);;) {

    if((last - first) <= SS_INSERTIONSORT_THRESHOLD) {
#if 1 < SS_INSERTIONSORT_THRESHOLD
      if(1 < (last - first)) { ss_insertionsort(T, PA, first, last, depth); }
#endif
      STACK_POP(first, last, depth, limit);
      continue;
    }

    Td = T + depth;
    if(limit-- == 0) { ss_heapsort(Td, PA, first, last - first); }
    if(limit < 0) {
      for(a = first + 1, v = Td[PA[*first]]; a < last; ++a) {
        if((x = Td[PA[*a]]) != v) {
          if(1 < (a - first)) { break; }
          v = x;
          first = a;
        }
      }
      if(Td[PA[*first] - 1] < v) {
        first = ss_partition(PA, first, a, depth);
      }
      if((a - first) <= (last - a)) {
        if(1 < (a - first)) {
          STACK_PUSH(a, last, depth, -1);
          last = a, depth += 1, limit = ss_ilg(a - first);
        } else {
          first = a, limit = -1;
        }
      } else {
        if(1 < (last - a)) {
          STACK_PUSH(first, a, depth + 1, ss_ilg(a - first));
          first = a, limit = -1;
        } else {
          last = a, depth += 1, limit = ss_ilg(a - first);
        }
      }
      continue;
    }

    /* choose pivot */
    a = ss_pivot(Td, PA, first, last);
    v = Td[PA[*a]];
    SWAP(*first, *a);

    /* partition */
    for(b = first; (++b < last) && ((x = Td[PA[*b]]) == v);) { }
    if(((a = b) < last) && (x < v)) {
      for(; (++b < last) && ((x = Td[PA[*b]]) <= v);) {
        if(x == v) { SWAP(*b, *a); ++a; }
      }
    }
    for(c = last; (b < --c) && ((x = Td[PA[*c]]) == v);) { }
    if((b < (d = c)) && (x > v)) {
      for(; (b < --c) && ((x = Td[PA[*c]]) >= v);) {
        if(x == v) { SWAP(*c, *d); --d; }
      }
    }
    for(; b < c;) {
      SWAP(*b, *c);
      for(; (++b < c) && ((x = Td[PA[*b]]) <= v);) {
        if(x == v) { SWAP(*b, *a); ++a; }
      }
      for(; (b < --c) && ((x = Td[PA[*c]]) >= v);) {
        if(x == v) { SWAP(*c, *d); --d; }
      }
    }

    if(a <= d) {
      c = b - 1;

      if((s = a - first) > (t = b - a)) { s = t; }
      for(e = first, f = b - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
      if((s = d - c) > (t = last - d - 1)) { s = t; }
      for(e = b, f = last - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }

      a = first + (b - a), c = last - (d - c);
      b = (v <= Td[PA[*a] - 1]) ? a : ss_partition(PA, a, c, depth);

      if((a - first) <= (last - c)) {
        if((last - c) <= (c - b)) {
          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
          STACK_PUSH(c, last, depth, limit);
          last = a;
        } else if((a - first) <= (c - b)) {
          STACK_PUSH(c, last, depth, limit);
          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
          last = a;
        } else {
          STACK_PUSH(c, last, depth, limit);
          STACK_PUSH(first, a, depth, limit);
          first = b, last = c, depth += 1, limit = ss_ilg(c - b);
        }
      } else {
        if((a - first) <= (c - b)) {
          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
          STACK_PUSH(first, a, depth, limit);
          first = c;
        } else if((last - c) <= (c - b)) {
          STACK_PUSH(first, a, depth, limit);
          STACK_PUSH(b, c, depth + 1, ss_ilg(c - b));
          first = c;
        } else {
          STACK_PUSH(first, a, depth, limit);
          STACK_PUSH(c, last, depth, limit);
          first = b, last = c, depth += 1, limit = ss_ilg(c - b);
        }
      }
    } else {
      limit += 1;
      if(Td[PA[*first] - 1] < v) {
        first = ss_partition(PA, first, last, depth);
        limit = ss_ilg(last - first);
      }
      depth += 1;
    }
  }
#undef STACK_SIZE
}

#endif /* (SS_BLOCKSIZE == 0) || (SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE) */


/*---------------------------------------------------------------------------*/

#if SS_BLOCKSIZE != 0

static INLINE
void
ss_blockswap(int *a, int *b, int n) {
  int t;
  for(; 0 < n; --n, ++a, ++b) {
    t = *a, *a = *b, *b = t;
  }
}

static INLINE
void
ss_rotate(int *first, int *middle, int *last) {
  int *a, *b, t;
  int l, r;
  l = middle - first, r = last - middle;
  for(; (0 < l) && (0 < r);) {
    if(l == r) { ss_blockswap(first, middle, l); break; }
    if(l < r) {
      a = last - 1, b = middle - 1;
      t = *a;
      do {
        *a-- = *b, *b-- = *a;
        if(b < first) {
          *a = t;
          last = a;
          if((r -= l + 1) <= l) { break; }
          a -= 1, b = middle - 1;
          t = *a;
        }
      } while(1);
    } else {
      a = first, b = middle;
      t = *a;
      do {
        *a++ = *b, *b++ = *a;
        if(last <= b) {
          *a = t;
          first = a + 1;
          if((l -= r + 1) <= r) { break; }
          a += 1, b = middle;
          t = *a;
        }
      } while(1);
    }
  }
}


/*---------------------------------------------------------------------------*/

static
void
ss_inplacemerge(const unsigned char *T, const int *PA,
                int *first, int *middle, int *last,
                int depth) {
  const int *p;
  int *a, *b;
  int len, half;
  int q, r;
  int x;

  for(;;) {
    if(*(last - 1) < 0) { x = 1; p = PA + ~*(last - 1); }
    else                { x = 0; p = PA +  *(last - 1); }
    for(a = first, len = middle - first, half = len >> 1, r = -1;
        0 < len;
        len = half, half >>= 1) {
      b = a + half;
      q = ss_compare(T, PA + ((0 <= *b) ? *b : ~*b), p, depth);
      if(q < 0) {
        a = b + 1;
        half -= (len & 1) ^ 1;
      } else {
        r = q;
      }
    }
    if(a < middle) {
      if(r == 0) { *a = ~*a; }
      ss_rotate(a, middle, last);
      last -= middle - a;
      middle = a;
      if(first == middle) { break; }
    }
    --last;
    if(x != 0) { while(*--last < 0) { } }
    if(middle == last) { break; }
  }
}


/*---------------------------------------------------------------------------*/

/* Merge-forward with internal buffer. */
static
void
ss_mergeforward(const unsigned char *T, const int *PA,
                int *first, int *middle, int *last,
                int *buf, int depth) {
  int *a, *b, *c, *bufend;
  int t;
  int r;

  bufend = buf + (middle - first) - 1;
  ss_blockswap(buf, first, middle - first);

  for(t = *(a = first), b = buf, c = middle;;) {
    r = ss_compare(T, PA + *b, PA + *c, depth);
    if(r < 0) {
      do {
        *a++ = *b;
        if(bufend <= b) { *bufend = t; return; }
        *b++ = *a;
      } while(*b < 0);
    } else if(r > 0) {
      do {
        *a++ = *c, *c++ = *a;
        if(last <= c) {
          while(b < bufend) { *a++ = *b, *b++ = *a; }
          *a = *b, *b = t;
          return;
        }
      } while(*c < 0);
    } else {
      *c = ~*c;
      do {
        *a++ = *b;
        if(bufend <= b) { *bufend = t; return; }
        *b++ = *a;
      } while(*b < 0);

      do {
        *a++ = *c, *c++ = *a;
        if(last <= c) {
          while(b < bufend) { *a++ = *b, *b++ = *a; }
          *a = *b, *b = t;
          return;
        }
      } while(*c < 0);
    }
  }
}

/* Merge-backward with internal buffer. */
static
void
ss_mergebackward(const unsigned char *T, const int *PA,
                 int *first, int *middle, int *last,
                 int *buf, int depth) {
  const int *p1, *p2;
  int *a, *b, *c, *bufend;
  int t;
  int r;
  int x;

  bufend = buf + (last - middle) - 1;
  ss_blockswap(buf, middle, last - middle);

  x = 0;
  if(*bufend < 0)       { p1 = PA + ~*bufend; x |= 1; }
  else                  { p1 = PA +  *bufend; }
  if(*(middle - 1) < 0) { p2 = PA + ~*(middle - 1); x |= 2; }
  else                  { p2 = PA +  *(middle - 1); }
  for(t = *(a = last - 1), b = bufend, c = middle - 1;;) {
    r = ss_compare(T, p1, p2, depth);
    if(0 < r) {
      if(x & 1) { do { *a-- = *b, *b-- = *a; } while(*b < 0); x ^= 1; }
      *a-- = *b;
      if(b <= buf) { *buf = t; break; }
      *b-- = *a;
      if(*b < 0) { p1 = PA + ~*b; x |= 1; }
      else       { p1 = PA +  *b; }
    } else if(r < 0) {
      if(x & 2) { do { *a-- = *c, *c-- = *a; } while(*c < 0); x ^= 2; }
      *a-- = *c, *c-- = *a;
      if(c < first) {
        while(buf < b) { *a-- = *b, *b-- = *a; }
        *a = *b, *b = t;
        break;
      }
      if(*c < 0) { p2 = PA + ~*c; x |= 2; }
      else       { p2 = PA +  *c; }
    } else {
      if(x & 1) { do { *a-- = *b, *b-- = *a; } while(*b < 0); x ^= 1; }
      *a-- = ~*b;
      if(b <= buf) { *buf = t; break; }
      *b-- = *a;
      if(x & 2) { do { *a-- = *c, *c-- = *a; } while(*c < 0); x ^= 2; }
      *a-- = *c, *c-- = *a;
      if(c < first) {
        while(buf < b) { *a-- = *b, *b-- = *a; }
        *a = *b, *b = t;
        break;
      }
      if(*b < 0) { p1 = PA + ~*b; x |= 1; }
      else       { p1 = PA +  *b; }
      if(*c < 0) { p2 = PA + ~*c; x |= 2; }
      else       { p2 = PA +  *c; }
    }
  }
}

/* D&C based merge. */
static
void
ss_swapmerge(const unsigned char *T, const int *PA,
             int *first, int *middle, int *last,
             int *buf, int bufsize, int depth) {
#define STACK_SIZE SS_SMERGE_STACKSIZE
#define GETIDX(a) ((0 <= (a)) ? (a) : (~(a)))
#define MERGE_CHECK(a, b, c)\
  do {\
    if(((c) & 1) ||\
       (((c) & 2) && (ss_compare(T, PA + GETIDX(*((a) - 1)), PA + *(a), depth) == 0))) {\
      *(a) = ~*(a);\
    }\
    if(((c) & 4) && ((ss_compare(T, PA + GETIDX(*((b) - 1)), PA + *(b), depth) == 0))) {\
      *(b) = ~*(b);\
    }\
  } while(0)
  struct { int *a, *b, *c; int d; } stack[STACK_SIZE];
  int *l, *r, *lm, *rm;
  int m, len, half;
  int ssize;
  int check, next;

  for(check = 0, ssize = 0;;) {
    if((last - middle) <= bufsize) {
      if((first < middle) && (middle < last)) {
        ss_mergebackward(T, PA, first, middle, last, buf, depth);
      }
      MERGE_CHECK(first, last, check);
      STACK_POP(first, middle, last, check);
      continue;
    }

    if((middle - first) <= bufsize) {
      if(first < middle) {
        ss_mergeforward(T, PA, first, middle, last, buf, depth);
      }
      MERGE_CHECK(first, last, check);
      STACK_POP(first, middle, last, check);
      continue;
    }

    for(m = 0, len = MIN(middle - first, last - middle), half = len >> 1;
        0 < len;
        len = half, half >>= 1) {
      if(ss_compare(T, PA + GETIDX(*(middle + m + half)),
                       PA + GETIDX(*(middle - m - half - 1)), depth) < 0) {
        m += half + 1;
        half -= (len & 1) ^ 1;
      }
    }

    if(0 < m) {
      lm = middle - m, rm = middle + m;
      ss_blockswap(lm, middle, m);
      l = r = middle, next = 0;
      if(rm < last) {
        if(*rm < 0) {
          *rm = ~*rm;
          if(first < lm) { for(; *--l < 0;) { } next |= 4; }
          next |= 1;
        } else if(first < lm) {
          for(; *r < 0; ++r) { }
          next |= 2;
        }
      }

      if((l - first) <= (last - r)) {
        STACK_PUSH(r, rm, last, (next & 3) | (check & 4));
        middle = lm, last = l, check = (check & 3) | (next & 4);
      } else {
        if((next & 2) && (r == middle)) { next ^= 6; }
        STACK_PUSH(first, lm, l, (check & 3) | (next & 4));
        first = r, middle = rm, check = (next & 3) | (check & 4);
      }
    } else {
      if(ss_compare(T, PA + GETIDX(*(middle - 1)), PA + *middle, depth) == 0) {
        *middle = ~*middle;
      }
      MERGE_CHECK(first, last, check);
      STACK_POP(first, middle, last, check);
    }
  }
#undef STACK_SIZE
}

#endif /* SS_BLOCKSIZE != 0 */


/*---------------------------------------------------------------------------*/

/* Substring sort */
static
void
sssort(const unsigned char *T, const int *PA,
       int *first, int *last,
       int *buf, int bufsize,
       int depth, int n, int lastsuffix) {
  int *a;
#if SS_BLOCKSIZE != 0
  int *b, *middle, *curbuf;
  int j, k, curbufsize, limit;
#endif
  int i;

  if(lastsuffix != 0) { ++first; }

#if SS_BLOCKSIZE == 0
  ss_mintrosort(T, PA, first, last, depth);
#else
  if((bufsize < SS_BLOCKSIZE) &&
      (bufsize < (last - first)) &&
      (bufsize < (limit = ss_isqrt(last - first)))) {
    if(SS_BLOCKSIZE < limit) { limit = SS_BLOCKSIZE; }
    buf = middle = last - limit, bufsize = limit;
  } else {
    middle = last, limit = 0;
  }
  for(a = first, i = 0; SS_BLOCKSIZE < (middle - a); a += SS_BLOCKSIZE, ++i) {
#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
    ss_mintrosort(T, PA, a, a + SS_BLOCKSIZE, depth);
#elif 1 < SS_BLOCKSIZE
    ss_insertionsort(T, PA, a, a + SS_BLOCKSIZE, depth);
#endif
    curbufsize = last - (a + SS_BLOCKSIZE);
    curbuf = a + SS_BLOCKSIZE;
    if(curbufsize <= bufsize) { curbufsize = bufsize, curbuf = buf; }
    for(b = a, k = SS_BLOCKSIZE, j = i; j & 1; b -= k, k <<= 1, j >>= 1) {
      ss_swapmerge(T, PA, b - k, b, b + k, curbuf, curbufsize, depth);
    }
  }
#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
  ss_mintrosort(T, PA, a, middle, depth);
#elif 1 < SS_BLOCKSIZE
  ss_insertionsort(T, PA, a, middle, depth);
#endif
  for(k = SS_BLOCKSIZE; i != 0; k <<= 1, i >>= 1) {
    if(i & 1) {
      ss_swapmerge(T, PA, a - k, a, middle, buf, bufsize, depth);
      a -= k;
    }
  }
  if(limit != 0) {
#if SS_INSERTIONSORT_THRESHOLD < SS_BLOCKSIZE
    ss_mintrosort(T, PA, middle, last, depth);
#elif 1 < SS_BLOCKSIZE
    ss_insertionsort(T, PA, middle, last, depth);
#endif
    ss_inplacemerge(T, PA, first, middle, last, depth);
  }
#endif

  if(lastsuffix != 0) {
    /* Insert last type B* suffix. */
    int PAi[2]; PAi[0] = PA[*(first - 1)], PAi[1] = n - 2;
    for(a = first, i = *(first - 1);
        (a < last) && ((*a < 0) || (0 < ss_compare(T, &(PAi[0]), PA + *a, depth)));
        ++a) {
      *(a - 1) = *a;
    }
    *(a - 1) = i;
  }
}


/*---------------------------------------------------------------------------*/

static INLINE
int
tr_ilg(int n) {
  return (n & 0xffff0000) ?
          ((n & 0xff000000) ?
            24 + lg_table[(n >> 24) & 0xff] :
            16 + lg_table[(n >> 16) & 0xff]) :
          ((n & 0x0000ff00) ?
             8 + lg_table[(n >>  8) & 0xff] :
             0 + lg_table[(n >>  0) & 0xff]);
}


/*---------------------------------------------------------------------------*/

/* Simple insertionsort for small size groups. */
static
void
tr_insertionsort(const int *ISAd, int *first, int *last) {
  int *a, *b;
  int t, r;

  for(a = first + 1; a < last; ++a) {
    for(t = *a, b = a - 1; 0 > (r = ISAd[t] - ISAd[*b]);) {
      do { *(b + 1) = *b; } while((first <= --b) && (*b < 0));
      if(b < first) { break; }
    }
    if(r == 0) { *b = ~*b; }
    *(b + 1) = t;
  }
}


/*---------------------------------------------------------------------------*/

static INLINE
void
tr_fixdown(const int *ISAd, int *SA, int i, int size) {
  int j, k;
  int v;
  int c, d, e;

  for(v = SA[i], c = ISAd[v]; (j = 2 * i + 1) < size; SA[i] = SA[k], i = k) {
    d = ISAd[SA[k = j++]];
    if(d < (e = ISAd[SA[j]])) { k = j; d = e; }
    if(d <= c) { break; }
  }
  SA[i] = v;
}

/* Simple top-down heapsort. */
static
void
tr_heapsort(const int *ISAd, int *SA, int size) {
  int i, m;
  int t;

  m = size;
  if((size % 2) == 0) {
    m--;
    if(ISAd[SA[m / 2]] < ISAd[SA[m]]) { SWAP(SA[m], SA[m / 2]); }
  }

  for(i = m / 2 - 1; 0 <= i; --i) { tr_fixdown(ISAd, SA, i, m); }
  if((size % 2) == 0) { SWAP(SA[0], SA[m]); tr_fixdown(ISAd, SA, 0, m); }
  for(i = m - 1; 0 < i; --i) {
    t = SA[0], SA[0] = SA[i];
    tr_fixdown(ISAd, SA, 0, i);
    SA[i] = t;
  }
}


/*---------------------------------------------------------------------------*/

/* Returns the median of three elements. */
static INLINE
int *
tr_median3(const int *ISAd, int *v1, int *v2, int *v3) {
  int *t;
  if(ISAd[*v1] > ISAd[*v2]) { SWAP(v1, v2); }
  if(ISAd[*v2] > ISAd[*v3]) {
    if(ISAd[*v1] > ISAd[*v3]) { return v1; }
    else { return v3; }
  }
  return v2;
}

/* Returns the median of five elements. */
static INLINE
int *
tr_median5(const int *ISAd,
           int *v1, int *v2, int *v3, int *v4, int *v5) {
  int *t;
  if(ISAd[*v2] > ISAd[*v3]) { SWAP(v2, v3); }
  if(ISAd[*v4] > ISAd[*v5]) { SWAP(v4, v5); }
  if(ISAd[*v2] > ISAd[*v4]) { SWAP(v2, v4); SWAP(v3, v5); }
  if(ISAd[*v1] > ISAd[*v3]) { SWAP(v1, v3); }
  if(ISAd[*v1] > ISAd[*v4]) { SWAP(v1, v4); SWAP(v3, v5); }
  if(ISAd[*v3] > ISAd[*v4]) { return v4; }
  return v3;
}

/* Returns the pivot element. */
static INLINE
int *
tr_pivot(const int *ISAd, int *first, int *last) {
  int *middle;
  int t;

  t = last - first;
  middle = first + t / 2;

  if(t <= 512) {
    if(t <= 32) {
      return tr_median3(ISAd, first, middle, last - 1);
    } else {
      t >>= 2;
      return tr_median5(ISAd, first, first + t, middle, last - 1 - t, last - 1);
    }
  }
  t >>= 3;
  first  = tr_median3(ISAd, first, first + t, first + (t << 1));
  middle = tr_median3(ISAd, middle - t, middle, middle + t);
  last   = tr_median3(ISAd, last - 1 - (t << 1), last - 1 - t, last - 1);
  return tr_median3(ISAd, first, middle, last);
}


/*---------------------------------------------------------------------------*/

typedef struct _trbudget_t trbudget_t;
struct _trbudget_t {
  int chance;
  int remain;
  int incval;
  int count;
};

static INLINE
void
trbudget_init(trbudget_t *budget, int chance, int incval) {
  budget->chance = chance;
  budget->remain = budget->incval = incval;
}

static INLINE
int
trbudget_check(trbudget_t *budget, int size) {
  if(size <= budget->remain) { budget->remain -= size; return 1; }
  if(budget->chance == 0) { budget->count += size; return 0; }
  budget->remain += budget->incval - size;
  budget->chance -= 1;
  return 1;
}


/*---------------------------------------------------------------------------*/

static INLINE
void
tr_partition(const int *ISAd,
             int *first, int *middle, int *last,
             int **pa, int **pb, int v) {
  int *a, *b, *c, *d, *e, *f;
  int t, s;
  int x = 0;

  for(b = middle - 1; (++b < last) && ((x = ISAd[*b]) == v);) { }
  if(((a = b) < last) && (x < v)) {
    for(; (++b < last) && ((x = ISAd[*b]) <= v);) {
      if(x == v) { SWAP(*b, *a); ++a; }
    }
  }
  for(c = last; (b < --c) && ((x = ISAd[*c]) == v);) { }
  if((b < (d = c)) && (x > v)) {
    for(; (b < --c) && ((x = ISAd[*c]) >= v);) {
      if(x == v) { SWAP(*c, *d); --d; }
    }
  }
  for(; b < c;) {
    SWAP(*b, *c);
    for(; (++b < c) && ((x = ISAd[*b]) <= v);) {
      if(x == v) { SWAP(*b, *a); ++a; }
    }
    for(; (b < --c) && ((x = ISAd[*c]) >= v);) {
      if(x == v) { SWAP(*c, *d); --d; }
    }
  }

  if(a <= d) {
    c = b - 1;
    if((s = a - first) > (t = b - a)) { s = t; }
    for(e = first, f = b - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
    if((s = d - c) > (t = last - d - 1)) { s = t; }
    for(e = b, f = last - s; 0 < s; --s, ++e, ++f) { SWAP(*e, *f); }
    first += (b - a), last -= (d - c);
  }
  *pa = first, *pb = last;
}

static
void
tr_copy(int *ISA, const int *SA,
        int *first, int *a, int *b, int *last,
        int depth) {
  /* sort suffixes of middle partition
     by using sorted order of suffixes of left and right partition. */
  int *c, *d, *e;
  int s, v;

  v = b - SA - 1;
  for(c = first, d = a - 1; c <= d; ++c) {
    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
      *++d = s;
      ISA[s] = d - SA;
    }
  }
  for(c = last - 1, e = d + 1, d = b; e < d; --c) {
    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
      *--d = s;
      ISA[s] = d - SA;
    }
  }
}

static
void
tr_partialcopy(int *ISA, const int *SA,
               int *first, int *a, int *b, int *last,
               int depth) {
  int *c, *d, *e;
  int s, v;
  int rank, lastrank, newrank = -1;

  v = b - SA - 1;
  lastrank = -1;
  for(c = first, d = a - 1; c <= d; ++c) {
    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
      *++d = s;
      rank = ISA[s + depth];
      if(lastrank != rank) { lastrank = rank; newrank = d - SA; }
      ISA[s] = newrank;
    }
  }

  lastrank = -1;
  for(e = d; first <= e; --e) {
    rank = ISA[*e];
    if(lastrank != rank) { lastrank = rank; newrank = e - SA; }
    if(newrank != rank) { ISA[*e] = newrank; }
  }

  lastrank = -1;
  for(c = last - 1, e = d + 1, d = b; e < d; --c) {
    if((0 <= (s = *c - depth)) && (ISA[s] == v)) {
      *--d = s;
      rank = ISA[s + depth];
      if(lastrank != rank) { lastrank = rank; newrank = d - SA; }
      ISA[s] = newrank;
    }
  }
}

static
void
tr_introsort(int *ISA, const int *ISAd,
             int *SA, int *first, int *last,
             trbudget_t *budget) {
#define STACK_SIZE TR_STACKSIZE
  struct { const int *a; int *b, *c; int d, e; }stack[STACK_SIZE];
  int *a, *b, *c;
  int t;
  int v, x = 0;
  int incr = ISAd - ISA;
  int limit, next;
  int ssize, trlink = -1;

  for(ssize = 0, limit = tr_ilg(last - first);;) {

    if(limit < 0) {
      if(limit == -1) {
        /* tandem repeat partition */
        tr_partition(ISAd - incr, first, first, last, &a, &b, last - SA - 1);

        /* update ranks */
        if(a < last) {
          for(c = first, v = a - SA - 1; c < a; ++c) { ISA[*c] = v; }
        }
        if(b < last) {
          for(c = a, v = b - SA - 1; c < b; ++c) { ISA[*c] = v; }
        }

        /* push */
        if(1 < (b - a)) {
          STACK_PUSH5(NULL, a, b, 0, 0);
          STACK_PUSH5(ISAd - incr, first, last, -2, trlink);
          trlink = ssize - 2;
        }
        if((a - first) <= (last - b)) {
          if(1 < (a - first)) {
            STACK_PUSH5(ISAd, b, last, tr_ilg(last - b), trlink);
            last = a, limit = tr_ilg(a - first);
          } else if(1 < (last - b)) {
            first = b, limit = tr_ilg(last - b);
          } else {
            STACK_POP5(ISAd, first, last, limit, trlink);
          }
        } else {
          if(1 < (last - b)) {
            STACK_PUSH5(ISAd, first, a, tr_ilg(a - first), trlink);
            first = b, limit = tr_ilg(last - b);
          } else if(1 < (a - first)) {
            last = a, limit = tr_ilg(a - first);
          } else {
            STACK_POP5(ISAd, first, last, limit, trlink);
          }
        }
      } else if(limit == -2) {
        /* tandem repeat copy */
        a = stack[--ssize].b, b = stack[ssize].c;
        if(stack[ssize].d == 0) {
          tr_copy(ISA, SA, first, a, b, last, ISAd - ISA);
        } else {
          if(0 <= trlink) { stack[trlink].d = -1; }
          tr_partialcopy(ISA, SA, first, a, b, last, ISAd - ISA);
        }
        STACK_POP5(ISAd, first, last, limit, trlink);
      } else {
        /* sorted partition */
        if(0 <= *first) {
          a = first;
          do { ISA[*a] = a - SA; } while((++a < last) && (0 <= *a));
          first = a;
        }
        if(first < last) {
          a = first; do { *a = ~*a; } while(*++a < 0);
          next = (ISA[*a] != ISAd[*a]) ? tr_ilg(a - first + 1) : -1;
          if(++a < last) { for(b = first, v = a - SA - 1; b < a; ++b) { ISA[*b] = v; } }

          /* push */
          if(trbudget_check(budget, a - first)) {
            if((a - first) <= (last - a)) {
              STACK_PUSH5(ISAd, a, last, -3, trlink);
              ISAd += incr, last = a, limit = next;
            } else {
              if(1 < (last - a)) {
                STACK_PUSH5(ISAd + incr, first, a, next, trlink);
                first = a, limit = -3;
              } else {
                ISAd += incr, last = a, limit = next;
              }
            }
          } else {
            if(0 <= trlink) { stack[trlink].d = -1; }
            if(1 < (last - a)) {
              first = a, limit = -3;
            } else {
              STACK_POP5(ISAd, first, last, limit, trlink);
            }
          }
        } else {
          STACK_POP5(ISAd, first, last, limit, trlink);
        }
      }
      continue;
    }

    if((last - first) <= TR_INSERTIONSORT_THRESHOLD) {
      tr_insertionsort(ISAd, first, last);
      limit = -3;
      continue;
    }

    if(limit-- == 0) {
      tr_heapsort(ISAd, first, last - first);
      for(a = last - 1; first < a; a = b) {
        for(x = ISAd[*a], b = a - 1; (first <= b) && (ISAd[*b] == x); --b) { *b = ~*b; }
      }
      limit = -3;
      continue;
    }

    /* choose pivot */
    a = tr_pivot(ISAd, first, last);
    SWAP(*first, *a);
    v = ISAd[*first];

    /* partition */
    tr_partition(ISAd, first, first + 1, last, &a, &b, v);
    if((last - first) != (b - a)) {
      next = (ISA[*a] != v) ? tr_ilg(b - a) : -1;

      /* update ranks */
      for(c = first, v = a - SA - 1; c < a; ++c) { ISA[*c] = v; }
      if(b < last) { for(c = a, v = b - SA - 1; c < b; ++c) { ISA[*c] = v; } }

      /* push */
      if((1 < (b - a)) && (trbudget_check(budget, b - a))) {
        if((a - first) <= (last - b)) {
          if((last - b) <= (b - a)) {
            if(1 < (a - first)) {
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              STACK_PUSH5(ISAd, b, last, limit, trlink);
              last = a;
            } else if(1 < (last - b)) {
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              first = b;
            } else {
              ISAd += incr, first = a, last = b, limit = next;
            }
          } else if((a - first) <= (b - a)) {
            if(1 < (a - first)) {
              STACK_PUSH5(ISAd, b, last, limit, trlink);
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              last = a;
            } else {
              STACK_PUSH5(ISAd, b, last, limit, trlink);
              ISAd += incr, first = a, last = b, limit = next;
            }
          } else {
            STACK_PUSH5(ISAd, b, last, limit, trlink);
            STACK_PUSH5(ISAd, first, a, limit, trlink);
            ISAd += incr, first = a, last = b, limit = next;
          }
        } else {
          if((a - first) <= (b - a)) {
            if(1 < (last - b)) {
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              STACK_PUSH5(ISAd, first, a, limit, trlink);
              first = b;
            } else if(1 < (a - first)) {
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              last = a;
            } else {
              ISAd += incr, first = a, last = b, limit = next;
            }
          } else if((last - b) <= (b - a)) {
            if(1 < (last - b)) {
              STACK_PUSH5(ISAd, first, a, limit, trlink);
              STACK_PUSH5(ISAd + incr, a, b, next, trlink);
              first = b;
            } else {
              STACK_PUSH5(ISAd, first, a, limit, trlink);
              ISAd += incr, first = a, last = b, limit = next;
            }
          } else {
            STACK_PUSH5(ISAd, first, a, limit, trlink);
            STACK_PUSH5(ISAd, b, last, limit, trlink);
            ISAd += incr, first = a, last = b, limit = next;
          }
        }
      } else {
        if((1 < (b - a)) && (0 <= trlink)) { stack[trlink].d = -1; }
        if((a - first) <= (last - b)) {
          if(1 < (a - first)) {
            STACK_PUSH5(ISAd, b, last, limit, trlink);
            last = a;
          } else if(1 < (last - b)) {
            first = b;
          } else {
            STACK_POP5(ISAd, first, last, limit, trlink);
          }
        } else {
          if(1 < (last - b)) {
            STACK_PUSH5(ISAd, first, a, limit, trlink);
            first = b;
          } else if(1 < (a - first)) {
            last = a;
          } else {
            STACK_POP5(ISAd, first, last, limit, trlink);
          }
        }
      }
    } else {
      if(trbudget_check(budget, last - first)) {
        limit = tr_ilg(last - first), ISAd += incr;
      } else {
        if(0 <= trlink) { stack[trlink].d = -1; }
        STACK_POP5(ISAd, first, last, limit, trlink);
      }
    }
  }
#undef STACK_SIZE
}


/*---------------------------------------------------------------------------*/

/* Tandem repeat sort */
static
void
trsort(int *ISA, int *SA, int n, int depth) {
  int *ISAd;
  int *first, *last;
  trbudget_t budget;
  int t, skip, unsorted;

  trbudget_init(&budget, tr_ilg(n) * 2 / 3, n);
/*  trbudget_init(&budget, tr_ilg(n) * 3 / 4, n); */
  for(ISAd = ISA + depth; -n < *SA; ISAd += ISAd - ISA) {
    first = SA;
    skip = 0;
    unsorted = 0;
    do {
      if((t = *first) < 0) { first -= t; skip += t; }
      else {
        if(skip != 0) { *(first + skip) = skip; skip = 0; }
        last = SA + ISA[t] + 1;
        if(1 < (last - first)) {
          budget.count = 0;
          tr_introsort(ISA, ISAd, SA, first, last, &budget);
          if(budget.count != 0) { unsorted += budget.count; }
          else { skip = first - last; }
        } else if((last - first) == 1) {
          skip = -1;
        }
        first = last;
      }
    } while(first < (SA + n));
    if(skip != 0) { *(first + skip) = skip; }
    if(unsorted == 0) { break; }
  }
}


/*---------------------------------------------------------------------------*/

/* Sorts suffixes of type B*. */
static
int
sort_typeBstar(const unsigned char *T, int *SA,
               int *bucket_A, int *bucket_B,
               int n, int openMP) {
  int *PAb, *ISAb, *buf;
#ifdef LIBBSC_OPENMP
  int *curbuf;
  int l;
#endif
  int i, j, k, t, m, bufsize;
  int c0, c1;
#ifdef LIBBSC_OPENMP
  int d0, d1;
#endif
  (void)openMP;

  /* Initialize bucket arrays. */
  for(i = 0; i < BUCKET_A_SIZE; ++i) { bucket_A[i] = 0; }
  for(i = 0; i < BUCKET_B_SIZE; ++i) { bucket_B[i] = 0; }

  /* Count the number of occurrences of the first one or two characters of each
     type A, B and B* suffix. Moreover, store the beginning position of all
     type B* suffixes into the array SA. */
  for(i = n - 1, m = n, c0 = T[n - 1]; 0 <= i;) {
    /* type A suffix. */
    do { ++BUCKET_A(c1 = c0); } while((0 <= --i) && ((c0 = T[i]) >= c1));
    if(0 <= i) {
      /* type B* suffix. */
      ++BUCKET_BSTAR(c0, c1);
      SA[--m] = i;
      /* type B suffix. */
      for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) <= c1); --i, c1 = c0) {
        ++BUCKET_B(c0, c1);
      }
    }
  }
  m = n - m;
/*
note:
  A type B* suffix is lexicographically smaller than a type B suffix that
  begins with the same first two characters.
*/

  /* Calculate the index of start/end point of each bucket. */
  for(c0 = 0, i = 0, j = 0; c0 < ALPHABET_SIZE; ++c0) {
    t = i + BUCKET_A(c0);
    BUCKET_A(c0) = i + j; /* start point */
    i = t + BUCKET_B(c0, c0);
    for(c1 = c0 + 1; c1 < ALPHABET_SIZE; ++c1) {
      j += BUCKET_BSTAR(c0, c1);
      BUCKET_BSTAR(c0, c1) = j; /* end point */
      i += BUCKET_B(c0, c1);
    }
  }

  if(0 < m) {
    /* Sort the type B* suffixes by their first two characters. */
    PAb = SA + n - m; ISAb = SA + m;
    for(i = m - 2; 0 <= i; --i) {
      t = PAb[i], c0 = T[t], c1 = T[t + 1];
      SA[--BUCKET_BSTAR(c0, c1)] = i;
    }
    t = PAb[m - 1], c0 = T[t], c1 = T[t + 1];
    SA[--BUCKET_BSTAR(c0, c1)] = m - 1;

    /* Sort the type B* substrings using sssort. */
#ifdef LIBBSC_OPENMP
    if (openMP)
    {
        buf = SA + m;
        c0 = ALPHABET_SIZE - 2, c1 = ALPHABET_SIZE - 1, j = m;
#pragma omp parallel default(shared) private(bufsize, curbuf, k, l, d0, d1)
        {
          bufsize = (n - (2 * m)) / omp_get_num_threads();
          curbuf = buf + omp_get_thread_num() * bufsize;
          k = 0;
          for(;;) {
            #pragma omp critical(sssort_lock)
            {
              if(0 < (l = j)) {
                d0 = c0, d1 = c1;
                do {
                  k = BUCKET_BSTAR(d0, d1);
                  if(--d1 <= d0) {
                    d1 = ALPHABET_SIZE - 1;
                    if(--d0 < 0) { break; }
                  }
                } while(((l - k) <= 1) && (0 < (l = k)));
                c0 = d0, c1 = d1, j = k;
              }
            }
            if(l == 0) { break; }
            sssort(T, PAb, SA + k, SA + l,
                   curbuf, bufsize, 2, n, *(SA + k) == (m - 1));
          }
        }
    }
    else
    {
        buf = SA + m, bufsize = n - (2 * m);
        for(c0 = ALPHABET_SIZE - 2, j = m; 0 < j; --c0) {
          for(c1 = ALPHABET_SIZE - 1; c0 < c1; j = i, --c1) {
            i = BUCKET_BSTAR(c0, c1);
            if(1 < (j - i)) {
              sssort(T, PAb, SA + i, SA + j,
                     buf, bufsize, 2, n, *(SA + i) == (m - 1));
            }
          }
        }
    }
#else
    buf = SA + m, bufsize = n - (2 * m);
    for(c0 = ALPHABET_SIZE - 2, j = m; 0 < j; --c0) {
      for(c1 = ALPHABET_SIZE - 1; c0 < c1; j = i, --c1) {
        i = BUCKET_BSTAR(c0, c1);
        if(1 < (j - i)) {
          sssort(T, PAb, SA + i, SA + j,
                 buf, bufsize, 2, n, *(SA + i) == (m - 1));
        }
      }
    }
#endif

    /* Compute ranks of type B* substrings. */
    for(i = m - 1; 0 <= i; --i) {
      if(0 <= SA[i]) {
        j = i;
        do { ISAb[SA[i]] = i; } while((0 <= --i) && (0 <= SA[i]));
        SA[i + 1] = i - j;
        if(i <= 0) { break; }
      }
      j = i;
      do { ISAb[SA[i] = ~SA[i]] = j; } while(SA[--i] < 0);
      ISAb[SA[i]] = j;
    }

    /* Construct the inverse suffix array of type B* suffixes using trsort. */
    trsort(ISAb, SA, m, 1);

    /* Set the sorted order of type B* suffixes. */
    for(i = n - 1, j = m, c0 = T[n - 1]; 0 <= i;) {
      for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) >= c1); --i, c1 = c0) { }
      if(0 <= i) {
        t = i;
        for(--i, c1 = c0; (0 <= i) && ((c0 = T[i]) <= c1); --i, c1 = c0) { }
        SA[ISAb[--j]] = ((t == 0) || (1 < (t - i))) ? t : ~t;
      }
    }

    /* Calculate the index of start/end point of each bucket. */
    BUCKET_B(ALPHABET_SIZE - 1, ALPHABET_SIZE - 1) = n; /* end point */
    for(c0 = ALPHABET_SIZE - 2, k = m - 1; 0 <= c0; --c0) {
      i = BUCKET_A(c0 + 1) - 1;
      for(c1 = ALPHABET_SIZE - 1; c0 < c1; --c1) {
        t = i - BUCKET_B(c0, c1);
        BUCKET_B(c0, c1) = i; /* end point */

        /* Move all type B* suffixes to the correct position. */
        for(i = t, j = BUCKET_BSTAR(c0, c1);
            j <= k;
            --i, --k) { SA[i] = SA[k]; }
      }
      BUCKET_BSTAR(c0, c0 + 1) = i - BUCKET_B(c0, c0) + 1; /* start point */
      BUCKET_B(c0, c0) = i; /* end point */
    }
  }

  return m;
}

/* Constructs the suffix array by using the sorted order of type B* suffixes. */
static
void
construct_SA(const unsigned char *T, int *SA,
             int *bucket_A, int *bucket_B,
             int n, int m) {
  int *i, *j, *k;
  int s;
  int c0, c1, c2;

  if(0 < m) {
    /* Construct the sorted order of type B suffixes by using
       the sorted order of type B* suffixes. */
    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
      /* Scan the suffix array from right to left. */
      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
          i <= j;
          --j) {
        if(0 < (s = *j)) {
          assert(T[s] == c1);
          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
          assert(T[s - 1] <= T[s]);
          *j = ~s;
          c0 = T[--s];
          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
          if(c0 != c2) {
            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
            k = SA + BUCKET_B(c2 = c0, c1);
          }
          assert(k < j); assert(k != NULL);
          *k-- = s;
        } else {
          assert(((s == 0) && (T[s] == c1)) || (s < 0));
          *j = ~s;
        }
      }
    }
  }

  /* Construct the suffix array by using
     the sorted order of type B suffixes. */
  k = SA + BUCKET_A(c2 = T[n - 1]);
  *k++ = (T[n - 2] < c2) ? ~(n - 1) : (n - 1);
  /* Scan the suffix array from left to right. */
  for(i = SA, j = SA + n; i < j; ++i) {
    if(0 < (s = *i)) {
      assert(T[s - 1] >= T[s]);
      c0 = T[--s];
      if((s == 0) || (T[s - 1] < c0)) { s = ~s; }
      if(c0 != c2) {
        BUCKET_A(c2) = k - SA;
        k = SA + BUCKET_A(c2 = c0);
      }
      assert(i < k);
      *k++ = s;
    } else {
      assert(s < 0);
      *i = ~s;
    }
  }
}

/* Constructs the burrows-wheeler transformed string directly
   by using the sorted order of type B* suffixes. */
static
int
construct_BWT(const unsigned char *T, int *SA,
              int *bucket_A, int *bucket_B,
              int n, int m) {
  int *i, *j, *k, *orig;
  int s;
  int c0, c1, c2;

  if(0 < m) {
    /* Construct the sorted order of type B suffixes by using
       the sorted order of type B* suffixes. */
    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
      /* Scan the suffix array from right to left. */
      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
          i <= j;
          --j) {
        if(0 < (s = *j)) {
          assert(T[s] == c1);
          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
          assert(T[s - 1] <= T[s]);
          c0 = T[--s];
          *j = ~((int)c0);
          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
          if(c0 != c2) {
            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
            k = SA + BUCKET_B(c2 = c0, c1);
          }
          assert(k < j); assert(k != NULL);
          *k-- = s;
        } else if(s != 0) {
          *j = ~s;
#ifndef NDEBUG
        } else {
          assert(T[s] == c1);
#endif
        }
      }
    }
  }

  /* Construct the BWTed string by using
     the sorted order of type B suffixes. */
  k = SA + BUCKET_A(c2 = T[n - 1]);
  *k++ = (T[n - 2] < c2) ? ~((int)T[n - 2]) : (n - 1);
  /* Scan the suffix array from left to right. */
  for(i = SA, j = SA + n, orig = SA; i < j; ++i) {
    if(0 < (s = *i)) {
      assert(T[s - 1] >= T[s]);
      c0 = T[--s];
      *i = c0;
      if((0 < s) && (T[s - 1] < c0)) { s = ~((int)T[s - 1]); }
      if(c0 != c2) {
        BUCKET_A(c2) = k - SA;
        k = SA + BUCKET_A(c2 = c0);
      }
      assert(i < k);
      *k++ = s;
    } else if(s != 0) {
      *i = ~s;
    } else {
      orig = i;
    }
  }

  return orig - SA;
}

/* Constructs the burrows-wheeler transformed string directly
   by using the sorted order of type B* suffixes. */
static
int
construct_BWT_indexes(const unsigned char *T, int *SA,
                      int *bucket_A, int *bucket_B,
                      int n, int m,
                      unsigned char * num_indexes, int * indexes) {
  int *i, *j, *k, *orig;
  int s;
  int c0, c1, c2;

  int mod = n / 8;
  {
      mod |= mod >> 1;  mod |= mod >> 2;
      mod |= mod >> 4;  mod |= mod >> 8;
      mod |= mod >> 16; mod >>= 1;

      *num_indexes = (unsigned char)((n - 1) / (mod + 1));
  }

  if(0 < m) {
    /* Construct the sorted order of type B suffixes by using
       the sorted order of type B* suffixes. */
    for(c1 = ALPHABET_SIZE - 2; 0 <= c1; --c1) {
      /* Scan the suffix array from right to left. */
      for(i = SA + BUCKET_BSTAR(c1, c1 + 1),
          j = SA + BUCKET_A(c1 + 1) - 1, k = NULL, c2 = -1;
          i <= j;
          --j) {
        if(0 < (s = *j)) {
          assert(T[s] == c1);
          assert(((s + 1) < n) && (T[s] <= T[s + 1]));
          assert(T[s - 1] <= T[s]);

          if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = j - SA;

          c0 = T[--s];
          *j = ~((int)c0);
          if((0 < s) && (T[s - 1] > c0)) { s = ~s; }
          if(c0 != c2) {
            if(0 <= c2) { BUCKET_B(c2, c1) = k - SA; }
            k = SA + BUCKET_B(c2 = c0, c1);
          }
          assert(k < j); assert(k != NULL);
          *k-- = s;
        } else if(s != 0) {
          *j = ~s;
#ifndef NDEBUG
        } else {
          assert(T[s] == c1);
#endif
        }
      }
    }
  }

  /* Construct the BWTed string by using
     the sorted order of type B suffixes. */
  k = SA + BUCKET_A(c2 = T[n - 1]);
  if (T[n - 2] < c2) {
    if (((n - 1) & mod) == 0) indexes[(n - 1) / (mod + 1) - 1] = k - SA;
    *k++ = ~((int)T[n - 2]);
  }
  else {
    *k++ = n - 1;
  }

  /* Scan the suffix array from left to right. */
  for(i = SA, j = SA + n, orig = SA; i < j; ++i) {
    if(0 < (s = *i)) {
      assert(T[s - 1] >= T[s]);

      if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = i - SA;

      c0 = T[--s];
      *i = c0;
      if(c0 != c2) {
        BUCKET_A(c2) = k - SA;
        k = SA + BUCKET_A(c2 = c0);
      }
      assert(i < k);
      if((0 < s) && (T[s - 1] < c0)) {
          if ((s & mod) == 0) indexes[s / (mod + 1) - 1] = k - SA;
          *k++ = ~((int)T[s - 1]);
      } else
        *k++ = s;
    } else if(s != 0) {
      *i = ~s;
    } else {
      orig = i;
    }
  }

  return orig - SA;
}


/*---------------------------------------------------------------------------*/

/*- Function -*/

int
divsufsort(const unsigned char *T, int *SA, int n, int openMP) {
  int *bucket_A, *bucket_B;
  int m;
  int err = 0;

  /* Check arguments. */
  if((T == NULL) || (SA == NULL) || (n < 0)) { return -1; }
  else if(n == 0) { return 0; }
  else if(n == 1) { SA[0] = 0; return 0; }
  else if(n == 2) { m = (T[0] < T[1]); SA[m ^ 1] = 0, SA[m] = 1; return 0; }

  bucket_A = (int *)malloc(BUCKET_A_SIZE * sizeof(int));
  bucket_B = (int *)malloc(BUCKET_B_SIZE * sizeof(int));

  /* Suffixsort. */
  if((bucket_A != NULL) && (bucket_B != NULL)) {
    m = sort_typeBstar(T, SA, bucket_A, bucket_B, n, openMP);
    construct_SA(T, SA, bucket_A, bucket_B, n, m);
  } else {
    err = -2;
  }

  free(bucket_B);
  free(bucket_A);

  return err;
}

int
divbwt(const unsigned char *T, unsigned char *U, int *A, int n, unsigned char * num_indexes, int * indexes, int openMP) {
  int *B;
  int *bucket_A, *bucket_B;
  int m, pidx, i;

  /* Check arguments. */
  if((T == NULL) || (U == NULL) || (n < 0)) { return -1; }
  else if(n <= 1) { if(n == 1) { U[0] = T[0]; } return n; }

  if((B = A) == NULL) { B = (int *)malloc((size_t)(n + 1) * sizeof(int)); }
  bucket_A = (int *)malloc(BUCKET_A_SIZE * sizeof(int));
  bucket_B = (int *)malloc(BUCKET_B_SIZE * sizeof(int));

  /* Burrows-Wheeler Transform. */
  if((B != NULL) && (bucket_A != NULL) && (bucket_B != NULL)) {
    m = sort_typeBstar(T, B, bucket_A, bucket_B, n, openMP);

    if (num_indexes == NULL || indexes == NULL) {
        pidx = construct_BWT(T, B, bucket_A, bucket_B, n, m);
    } else {
        pidx = construct_BWT_indexes(T, B, bucket_A, bucket_B, n, m, num_indexes, indexes);
    }

    /* Copy to output string. */
    U[0] = T[n - 1];
    for(i = 0; i < pidx; ++i) { U[i + 1] = (unsigned char)B[i]; }
    for(i += 1; i < n; ++i) { U[i] = (unsigned char)B[i]; }
    pidx += 1;
  } else {
    pidx = -2;
  }

  free(bucket_B);
  free(bucket_A);
  if(A == NULL) { free(B); }

  return pidx;
}
/**** ended inlining dictBuilder/divsufsort.c ****/
/**** start inlining dictBuilder/fastcover.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */

/*-*************************************
*  Dependencies
***************************************/
#include <stdio.h>  /* fprintf */
#include <stdlib.h> /* malloc, free, qsort */
#include <string.h> /* memset */
#include <time.h>   /* clock */

#ifndef ZDICT_STATIC_LINKING_ONLY
#  define ZDICT_STATIC_LINKING_ONLY
#endif

/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/pool.h ****/
/**** skipping file: ../common/threading.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../compress/zstd_compress_internal.h ****/
/**** skipping file: ../zdict.h ****/
/**** skipping file: cover.h ****/


/*-*************************************
*  Constants
***************************************/
/**
* There are 32bit indexes used to ref samples, so limit samples size to 4GB
* on 64bit builds.
* For 32bit builds we choose 1 GB.
* Most 32bit platforms have 2GB user-mode addressable space and we allocate a large
* contiguous buffer, so 1GB is already a high limit.
*/
#define FASTCOVER_MAX_SAMPLES_SIZE (sizeof(size_t) == 8 ? ((unsigned)-1) : ((unsigned)1 GB))
#define FASTCOVER_MAX_F 31
#define FASTCOVER_MAX_ACCEL 10
#define FASTCOVER_DEFAULT_SPLITPOINT 0.75
#define DEFAULT_F 20
#define DEFAULT_ACCEL 1


/*-*************************************
*  Console display
***************************************/
#ifndef LOCALDISPLAYLEVEL
static int g_displayLevel = 0;
#endif
#undef  DISPLAY
#define DISPLAY(...)                                                           \
  {                                                                            \
    fprintf(stderr, __VA_ARGS__);                                              \
    fflush(stderr);                                                            \
  }
#undef  LOCALDISPLAYLEVEL
#define LOCALDISPLAYLEVEL(displayLevel, l, ...)                                \
  if (displayLevel >= l) {                                                     \
    DISPLAY(__VA_ARGS__);                                                      \
  } /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */
#undef  DISPLAYLEVEL
#define DISPLAYLEVEL(l, ...) LOCALDISPLAYLEVEL(g_displayLevel, l, __VA_ARGS__)

#ifndef LOCALDISPLAYUPDATE
static const clock_t g_refreshRate = CLOCKS_PER_SEC * 15 / 100;
static clock_t g_time = 0;
#endif
#undef  LOCALDISPLAYUPDATE
#define LOCALDISPLAYUPDATE(displayLevel, l, ...)                               \
  if (displayLevel >= l) {                                                     \
    if ((clock() - g_time > g_refreshRate) || (displayLevel >= 4)) {             \
      g_time = clock();                                                        \
      DISPLAY(__VA_ARGS__);                                                    \
    }                                                                          \
  }
#undef  DISPLAYUPDATE
#define DISPLAYUPDATE(l, ...) LOCALDISPLAYUPDATE(g_displayLevel, l, __VA_ARGS__)


/*-*************************************
* Hash Functions
***************************************/
/**
 * Hash the d-byte value pointed to by p and mod 2^f into the frequency vector
 */
static size_t FASTCOVER_hashPtrToIndex(const void* p, U32 f, unsigned d) {
  if (d == 6) {
    return ZSTD_hash6Ptr(p, f);
  }
  return ZSTD_hash8Ptr(p, f);
}


/*-*************************************
* Acceleration
***************************************/
typedef struct {
  unsigned finalize;    /* Percentage of training samples used for ZDICT_finalizeDictionary */
  unsigned skip;        /* Number of dmer skipped between each dmer counted in computeFrequency */
} FASTCOVER_accel_t;


static const FASTCOVER_accel_t FASTCOVER_defaultAccelParameters[FASTCOVER_MAX_ACCEL+1] = {
  { 100, 0 },   /* accel = 0, should not happen because accel = 0 defaults to accel = 1 */
  { 100, 0 },   /* accel = 1 */
  { 50, 1 },   /* accel = 2 */
  { 34, 2 },   /* accel = 3 */
  { 25, 3 },   /* accel = 4 */
  { 20, 4 },   /* accel = 5 */
  { 17, 5 },   /* accel = 6 */
  { 14, 6 },   /* accel = 7 */
  { 13, 7 },   /* accel = 8 */
  { 11, 8 },   /* accel = 9 */
  { 10, 9 },   /* accel = 10 */
};


/*-*************************************
* Context
***************************************/
typedef struct {
  const BYTE *samples;
  size_t *offsets;
  const size_t *samplesSizes;
  size_t nbSamples;
  size_t nbTrainSamples;
  size_t nbTestSamples;
  size_t nbDmers;
  U32 *freqs;
  unsigned d;
  unsigned f;
  FASTCOVER_accel_t accelParams;
} FASTCOVER_ctx_t;


/*-*************************************
*  Helper functions
***************************************/
/**
 * Selects the best segment in an epoch.
 * Segments of are scored according to the function:
 *
 * Let F(d) be the frequency of all dmers with hash value d.
 * Let S_i be hash value of the dmer at position i of segment S which has length k.
 *
 *     Score(S) = F(S_1) + F(S_2) + ... + F(S_{k-d+1})
 *
 * Once the dmer with hash value d is in the dictionary we set F(d) = 0.
 */
static COVER_segment_t FASTCOVER_selectSegment(const FASTCOVER_ctx_t *ctx,
                                              U32 *freqs, U32 begin, U32 end,
                                              ZDICT_cover_params_t parameters,
                                              U16* segmentFreqs) {
  /* Constants */
  const U32 k = parameters.k;
  const U32 d = parameters.d;
  const U32 f = ctx->f;
  const U32 dmersInK = k - d + 1;

  /* Try each segment (activeSegment) and save the best (bestSegment) */
  COVER_segment_t bestSegment = {0, 0, 0};
  COVER_segment_t activeSegment;

  /* Reset the activeDmers in the segment */
  /* The activeSegment starts at the beginning of the epoch. */
  activeSegment.begin = begin;
  activeSegment.end = begin;
  activeSegment.score = 0;

  /* Slide the activeSegment through the whole epoch.
   * Save the best segment in bestSegment.
   */
  while (activeSegment.end < end) {
    /* Get hash value of current dmer */
    const size_t idx = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.end, f, d);

    /* Add frequency of this index to score if this is the first occurrence of index in active segment */
    if (segmentFreqs[idx] == 0) {
      activeSegment.score += freqs[idx];
    }
    /* Increment end of segment and segmentFreqs*/
    activeSegment.end += 1;
    segmentFreqs[idx] += 1;
    /* If the window is now too large, drop the first position */
    if (activeSegment.end - activeSegment.begin == dmersInK + 1) {
      /* Get hash value of the dmer to be eliminated from active segment */
      const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d);
      segmentFreqs[delIndex] -= 1;
      /* Subtract frequency of this index from score if this is the last occurrence of this index in active segment */
      if (segmentFreqs[delIndex] == 0) {
        activeSegment.score -= freqs[delIndex];
      }
      /* Increment start of segment */
      activeSegment.begin += 1;
    }

    /* If this segment is the best so far save it */
    if (activeSegment.score > bestSegment.score) {
      bestSegment = activeSegment;
    }
  }

  /* Zero out rest of segmentFreqs array */
  while (activeSegment.begin < end) {
    const size_t delIndex = FASTCOVER_hashPtrToIndex(ctx->samples + activeSegment.begin, f, d);
    segmentFreqs[delIndex] -= 1;
    activeSegment.begin += 1;
  }

  {
    /*  Zero the frequency of hash value of each dmer covered by the chosen segment. */
    U32 pos;
    for (pos = bestSegment.begin; pos != bestSegment.end; ++pos) {
      const size_t i = FASTCOVER_hashPtrToIndex(ctx->samples + pos, f, d);
      freqs[i] = 0;
    }
  }

  return bestSegment;
}


static int FASTCOVER_checkParameters(ZDICT_cover_params_t parameters,
                                     size_t maxDictSize, unsigned f,
                                     unsigned accel) {
  /* k, d, and f are required parameters */
  if (parameters.d == 0 || parameters.k == 0) {
    return 0;
  }
  /* d has to be 6 or 8 */
  if (parameters.d != 6 && parameters.d != 8) {
    return 0;
  }
  /* k <= maxDictSize */
  if (parameters.k > maxDictSize) {
    return 0;
  }
  /* d <= k */
  if (parameters.d > parameters.k) {
    return 0;
  }
  /* 0 < f <= FASTCOVER_MAX_F*/
  if (f > FASTCOVER_MAX_F || f == 0) {
    return 0;
  }
  /* 0 < splitPoint <= 1 */
  if (parameters.splitPoint <= 0 || parameters.splitPoint > 1) {
    return 0;
  }
  /* 0 < accel <= 10 */
  if (accel > 10 || accel == 0) {
    return 0;
  }
  return 1;
}


/**
 * Clean up a context initialized with `FASTCOVER_ctx_init()`.
 */
static void
FASTCOVER_ctx_destroy(FASTCOVER_ctx_t* ctx)
{
    if (!ctx) return;

    free(ctx->freqs);
    ctx->freqs = NULL;

    free(ctx->offsets);
    ctx->offsets = NULL;
}


/**
 * Calculate for frequency of hash value of each dmer in ctx->samples
 */
static void
FASTCOVER_computeFrequency(U32* freqs, const FASTCOVER_ctx_t* ctx)
{
    const unsigned f = ctx->f;
    const unsigned d = ctx->d;
    const unsigned skip = ctx->accelParams.skip;
    const unsigned readLength = MAX(d, 8);
    size_t i;
    assert(ctx->nbTrainSamples >= 5);
    assert(ctx->nbTrainSamples <= ctx->nbSamples);
    for (i = 0; i < ctx->nbTrainSamples; i++) {
        size_t start = ctx->offsets[i];  /* start of current dmer */
        size_t const currSampleEnd = ctx->offsets[i+1];
        while (start + readLength <= currSampleEnd) {
            const size_t dmerIndex = FASTCOVER_hashPtrToIndex(ctx->samples + start, f, d);
            freqs[dmerIndex]++;
            start = start + skip + 1;
        }
    }
}


/**
 * Prepare a context for dictionary building.
 * The context is only dependent on the parameter `d` and can be used multiple
 * times.
 * Returns 0 on success or error code on error.
 * The context must be destroyed with `FASTCOVER_ctx_destroy()`.
 */
static size_t
FASTCOVER_ctx_init(FASTCOVER_ctx_t* ctx,
                   const void* samplesBuffer,
                   const size_t* samplesSizes, unsigned nbSamples,
                   unsigned d, double splitPoint, unsigned f,
                   FASTCOVER_accel_t accelParams)
{
    const BYTE* const samples = (const BYTE*)samplesBuffer;
    const size_t totalSamplesSize = COVER_sum(samplesSizes, nbSamples);
    /* Split samples into testing and training sets */
    const unsigned nbTrainSamples = splitPoint < 1.0 ? (unsigned)((double)nbSamples * splitPoint) : nbSamples;
    const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
    const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
    const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;

    /* Checks */
    if (totalSamplesSize < MAX(d, sizeof(U64)) ||
        totalSamplesSize >= (size_t)FASTCOVER_MAX_SAMPLES_SIZE) {
        DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
                    (unsigned)(totalSamplesSize >> 20), (FASTCOVER_MAX_SAMPLES_SIZE >> 20));
        return ERROR(srcSize_wrong);
    }

    /* Check if there are at least 5 training samples */
    if (nbTrainSamples < 5) {
        DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid\n", nbTrainSamples);
        return ERROR(srcSize_wrong);
    }

    /* Check if there's testing sample */
    if (nbTestSamples < 1) {
        DISPLAYLEVEL(1, "Total number of testing samples is %u and is invalid.\n", nbTestSamples);
        return ERROR(srcSize_wrong);
    }

    /* Zero the context */
    memset(ctx, 0, sizeof(*ctx));
    DISPLAYLEVEL(2, "Training on %u samples of total size %u\n", nbTrainSamples,
                    (unsigned)trainingSamplesSize);
    DISPLAYLEVEL(2, "Testing on %u samples of total size %u\n", nbTestSamples,
                    (unsigned)testSamplesSize);

    ctx->samples = samples;
    ctx->samplesSizes = samplesSizes;
    ctx->nbSamples = nbSamples;
    ctx->nbTrainSamples = nbTrainSamples;
    ctx->nbTestSamples = nbTestSamples;
    ctx->nbDmers = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
    ctx->d = d;
    ctx->f = f;
    ctx->accelParams = accelParams;

    /* The offsets of each file */
    ctx->offsets = (size_t*)calloc((nbSamples + 1), sizeof(size_t));
    if (ctx->offsets == NULL) {
        DISPLAYLEVEL(1, "Failed to allocate scratch buffers \n");
        FASTCOVER_ctx_destroy(ctx);
        return ERROR(memory_allocation);
    }

    /* Fill offsets from the samplesSizes */
    {   U32 i;
        ctx->offsets[0] = 0;
        assert(nbSamples >= 5);
        for (i = 1; i <= nbSamples; ++i) {
            ctx->offsets[i] = ctx->offsets[i - 1] + samplesSizes[i - 1];
        }
    }

    /* Initialize frequency array of size 2^f */
    ctx->freqs = (U32*)calloc(((U64)1 << f), sizeof(U32));
    if (ctx->freqs == NULL) {
        DISPLAYLEVEL(1, "Failed to allocate frequency table \n");
        FASTCOVER_ctx_destroy(ctx);
        return ERROR(memory_allocation);
    }

    DISPLAYLEVEL(2, "Computing frequencies\n");
    FASTCOVER_computeFrequency(ctx->freqs, ctx);

    return 0;
}


/**
 * Given the prepared context build the dictionary.
 */
static size_t
FASTCOVER_buildDictionary(const FASTCOVER_ctx_t* ctx,
                          U32* freqs,
                          void* dictBuffer, size_t dictBufferCapacity,
                          ZDICT_cover_params_t parameters,
                          U16* segmentFreqs)
{
  BYTE *const dict = (BYTE *)dictBuffer;
  size_t tail = dictBufferCapacity;
  /* Divide the data into epochs. We will select one segment from each epoch. */
  const COVER_epoch_info_t epochs = COVER_computeEpochs(
      (U32)dictBufferCapacity, (U32)ctx->nbDmers, parameters.k, 1);
  const size_t maxZeroScoreRun = 10;
  size_t zeroScoreRun = 0;
  size_t epoch;
  DISPLAYLEVEL(2, "Breaking content into %u epochs of size %u\n",
                (U32)epochs.num, (U32)epochs.size);
  /* Loop through the epochs until there are no more segments or the dictionary
   * is full.
   */
  for (epoch = 0; tail > 0; epoch = (epoch + 1) % epochs.num) {
    const U32 epochBegin = (U32)(epoch * epochs.size);
    const U32 epochEnd = epochBegin + epochs.size;
    size_t segmentSize;
    /* Select a segment */
    COVER_segment_t segment = FASTCOVER_selectSegment(
        ctx, freqs, epochBegin, epochEnd, parameters, segmentFreqs);

    /* If the segment covers no dmers, then we are out of content.
     * There may be new content in other epochs, for continue for some time.
     */
    if (segment.score == 0) {
      if (++zeroScoreRun >= maxZeroScoreRun) {
          break;
      }
      continue;
    }
    zeroScoreRun = 0;

    /* Trim the segment if necessary and if it is too small then we are done */
    segmentSize = MIN(segment.end - segment.begin + parameters.d - 1, tail);
    if (segmentSize < parameters.d) {
      break;
    }

    /* We fill the dictionary from the back to allow the best segments to be
     * referenced with the smallest offsets.
     */
    tail -= segmentSize;
    memcpy(dict + tail, ctx->samples + segment.begin, segmentSize);
    DISPLAYUPDATE(
        2, "\r%u%%       ",
        (unsigned)(((dictBufferCapacity - tail) * 100) / dictBufferCapacity));
  }
  DISPLAYLEVEL(2, "\r%79s\r", "");
  return tail;
}

/**
 * Parameters for FASTCOVER_tryParameters().
 */
typedef struct FASTCOVER_tryParameters_data_s {
    const FASTCOVER_ctx_t* ctx;
    COVER_best_t* best;
    size_t dictBufferCapacity;
    ZDICT_cover_params_t parameters;
} FASTCOVER_tryParameters_data_t;


/**
 * Tries a set of parameters and updates the COVER_best_t with the results.
 * This function is thread safe if zstd is compiled with multithreaded support.
 * It takes its parameters as an *OWNING* opaque pointer to support threading.
 */
static void FASTCOVER_tryParameters(void* opaque)
{
  /* Save parameters as local variables */
  FASTCOVER_tryParameters_data_t *const data = (FASTCOVER_tryParameters_data_t*)opaque;
  const FASTCOVER_ctx_t *const ctx = data->ctx;
  const ZDICT_cover_params_t parameters = data->parameters;
  size_t dictBufferCapacity = data->dictBufferCapacity;
  size_t totalCompressedSize = ERROR(GENERIC);
  /* Initialize array to keep track of frequency of dmer within activeSegment */
  U16* segmentFreqs = (U16*)calloc(((U64)1 << ctx->f), sizeof(U16));
  /* Allocate space for hash table, dict, and freqs */
  BYTE *const dict = (BYTE*)malloc(dictBufferCapacity);
  COVER_dictSelection_t selection = COVER_dictSelectionError(ERROR(GENERIC));
  U32* freqs = (U32*) malloc(((U64)1 << ctx->f) * sizeof(U32));
  if (!segmentFreqs || !dict || !freqs) {
    DISPLAYLEVEL(1, "Failed to allocate buffers: out of memory\n");
    goto _cleanup;
  }
  /* Copy the frequencies because we need to modify them */
  memcpy(freqs, ctx->freqs, ((U64)1 << ctx->f) * sizeof(U32));
  /* Build the dictionary */
  { const size_t tail = FASTCOVER_buildDictionary(ctx, freqs, dict, dictBufferCapacity,
                                                    parameters, segmentFreqs);

    const unsigned nbFinalizeSamples = (unsigned)(ctx->nbTrainSamples * ctx->accelParams.finalize / 100);
    selection = COVER_selectDict(dict + tail, dictBufferCapacity, dictBufferCapacity - tail,
         ctx->samples, ctx->samplesSizes, nbFinalizeSamples, ctx->nbTrainSamples, ctx->nbSamples, parameters, ctx->offsets,
         totalCompressedSize);

    if (COVER_dictSelectionIsError(selection)) {
      DISPLAYLEVEL(1, "Failed to select dictionary\n");
      goto _cleanup;
    }
  }
_cleanup:
  free(dict);
  COVER_best_finish(data->best, parameters, selection);
  free(data);
  free(segmentFreqs);
  COVER_dictSelectionFree(selection);
  free(freqs);
}


static void
FASTCOVER_convertToCoverParams(ZDICT_fastCover_params_t fastCoverParams,
                               ZDICT_cover_params_t* coverParams)
{
    coverParams->k = fastCoverParams.k;
    coverParams->d = fastCoverParams.d;
    coverParams->steps = fastCoverParams.steps;
    coverParams->nbThreads = fastCoverParams.nbThreads;
    coverParams->splitPoint = fastCoverParams.splitPoint;
    coverParams->zParams = fastCoverParams.zParams;
    coverParams->shrinkDict = fastCoverParams.shrinkDict;
}


static void
FASTCOVER_convertToFastCoverParams(ZDICT_cover_params_t coverParams,
                                   ZDICT_fastCover_params_t* fastCoverParams,
                                   unsigned f, unsigned accel)
{
    fastCoverParams->k = coverParams.k;
    fastCoverParams->d = coverParams.d;
    fastCoverParams->steps = coverParams.steps;
    fastCoverParams->nbThreads = coverParams.nbThreads;
    fastCoverParams->splitPoint = coverParams.splitPoint;
    fastCoverParams->f = f;
    fastCoverParams->accel = accel;
    fastCoverParams->zParams = coverParams.zParams;
    fastCoverParams->shrinkDict = coverParams.shrinkDict;
}


ZDICTLIB_STATIC_API size_t
ZDICT_trainFromBuffer_fastCover(void* dictBuffer, size_t dictBufferCapacity,
                                const void* samplesBuffer,
                                const size_t* samplesSizes, unsigned nbSamples,
                                ZDICT_fastCover_params_t parameters)
{
    BYTE* const dict = (BYTE*)dictBuffer;
    FASTCOVER_ctx_t ctx;
    ZDICT_cover_params_t coverParams;
    FASTCOVER_accel_t accelParams;
    /* Initialize global data */
    g_displayLevel = (int)parameters.zParams.notificationLevel;
    /* Assign splitPoint and f if not provided */
    parameters.splitPoint = 1.0;
    parameters.f = parameters.f == 0 ? DEFAULT_F : parameters.f;
    parameters.accel = parameters.accel == 0 ? DEFAULT_ACCEL : parameters.accel;
    /* Convert to cover parameter */
    memset(&coverParams, 0 , sizeof(coverParams));
    FASTCOVER_convertToCoverParams(parameters, &coverParams);
    /* Checks */
    if (!FASTCOVER_checkParameters(coverParams, dictBufferCapacity, parameters.f,
                                   parameters.accel)) {
      DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n");
      return ERROR(parameter_outOfBound);
    }
    if (nbSamples == 0) {
      DISPLAYLEVEL(1, "FASTCOVER must have at least one input file\n");
      return ERROR(srcSize_wrong);
    }
    if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
      DISPLAYLEVEL(1, "dictBufferCapacity must be at least %u\n",
                   ZDICT_DICTSIZE_MIN);
      return ERROR(dstSize_tooSmall);
    }
    /* Assign corresponding FASTCOVER_accel_t to accelParams*/
    accelParams = FASTCOVER_defaultAccelParameters[parameters.accel];
    /* Initialize context */
    {
      size_t const initVal = FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples,
                            coverParams.d, parameters.splitPoint, parameters.f,
                            accelParams);
      if (ZSTD_isError(initVal)) {
        DISPLAYLEVEL(1, "Failed to initialize context\n");
        return initVal;
      }
    }
    COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.nbDmers, g_displayLevel);
    /* Build the dictionary */
    DISPLAYLEVEL(2, "Building dictionary\n");
    {
      /* Initialize array to keep track of frequency of dmer within activeSegment */
      U16* segmentFreqs = (U16 *)calloc(((U64)1 << parameters.f), sizeof(U16));
      const size_t tail = FASTCOVER_buildDictionary(&ctx, ctx.freqs, dictBuffer,
                                                dictBufferCapacity, coverParams, segmentFreqs);
      const unsigned nbFinalizeSamples = (unsigned)(ctx.nbTrainSamples * ctx.accelParams.finalize / 100);
      const size_t dictionarySize = ZDICT_finalizeDictionary(
          dict, dictBufferCapacity, dict + tail, dictBufferCapacity - tail,
          samplesBuffer, samplesSizes, nbFinalizeSamples, coverParams.zParams);
      if (!ZSTD_isError(dictionarySize)) {
          DISPLAYLEVEL(2, "Constructed dictionary of size %u\n",
                      (unsigned)dictionarySize);
      }
      FASTCOVER_ctx_destroy(&ctx);
      free(segmentFreqs);
      return dictionarySize;
    }
}


ZDICTLIB_STATIC_API size_t
ZDICT_optimizeTrainFromBuffer_fastCover(
                    void* dictBuffer, size_t dictBufferCapacity,
                    const void* samplesBuffer,
                    const size_t* samplesSizes, unsigned nbSamples,
                    ZDICT_fastCover_params_t* parameters)
{
    ZDICT_cover_params_t coverParams;
    FASTCOVER_accel_t accelParams;
    /* constants */
    const unsigned nbThreads = parameters->nbThreads;
    const double splitPoint =
        parameters->splitPoint <= 0.0 ? FASTCOVER_DEFAULT_SPLITPOINT : parameters->splitPoint;
    const unsigned kMinD = parameters->d == 0 ? 6 : parameters->d;
    const unsigned kMaxD = parameters->d == 0 ? 8 : parameters->d;
    const unsigned kMinK = parameters->k == 0 ? 50 : parameters->k;
    const unsigned kMaxK = parameters->k == 0 ? 2000 : parameters->k;
    const unsigned kSteps = parameters->steps == 0 ? 40 : parameters->steps;
    const unsigned kStepSize = MAX((kMaxK - kMinK) / kSteps, 1);
    const unsigned kIterations =
        (1 + (kMaxD - kMinD) / 2) * (1 + (kMaxK - kMinK) / kStepSize);
    const unsigned f = parameters->f == 0 ? DEFAULT_F : parameters->f;
    const unsigned accel = parameters->accel == 0 ? DEFAULT_ACCEL : parameters->accel;
    const unsigned shrinkDict = 0;
    /* Local variables */
    const int displayLevel = (int)parameters->zParams.notificationLevel;
    unsigned iteration = 1;
    unsigned d;
    unsigned k;
    COVER_best_t best;
    POOL_ctx *pool = NULL;
    int warned = 0;
    /* Checks */
    if (splitPoint <= 0 || splitPoint > 1) {
      LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect splitPoint\n");
      return ERROR(parameter_outOfBound);
    }
    if (accel == 0 || accel > FASTCOVER_MAX_ACCEL) {
      LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect accel\n");
      return ERROR(parameter_outOfBound);
    }
    if (kMinK < kMaxD || kMaxK < kMinK) {
      LOCALDISPLAYLEVEL(displayLevel, 1, "Incorrect k\n");
      return ERROR(parameter_outOfBound);
    }
    if (nbSamples == 0) {
      LOCALDISPLAYLEVEL(displayLevel, 1, "FASTCOVER must have at least one input file\n");
      return ERROR(srcSize_wrong);
    }
    if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) {
      LOCALDISPLAYLEVEL(displayLevel, 1, "dictBufferCapacity must be at least %u\n",
                   ZDICT_DICTSIZE_MIN);
      return ERROR(dstSize_tooSmall);
    }
    if (nbThreads > 1) {
      pool = POOL_create(nbThreads, 1);
      if (!pool) {
        return ERROR(memory_allocation);
      }
    }
    /* Initialization */
    COVER_best_init(&best);
    memset(&coverParams, 0 , sizeof(coverParams));
    FASTCOVER_convertToCoverParams(*parameters, &coverParams);
    accelParams = FASTCOVER_defaultAccelParameters[accel];
    /* Turn down global display level to clean up display at level 2 and below */
    g_displayLevel = displayLevel == 0 ? 0 : displayLevel - 1;
    /* Loop through d first because each new value needs a new context */
    LOCALDISPLAYLEVEL(displayLevel, 2, "Trying %u different sets of parameters\n",
                      kIterations);
    for (d = kMinD; d <= kMaxD; d += 2) {
      /* Initialize the context for this value of d */
      FASTCOVER_ctx_t ctx;
      LOCALDISPLAYLEVEL(displayLevel, 3, "d=%u\n", d);
      {
        size_t const initVal = FASTCOVER_ctx_init(&ctx, samplesBuffer, samplesSizes, nbSamples, d, splitPoint, f, accelParams);
        if (ZSTD_isError(initVal)) {
          LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to initialize context\n");
          COVER_best_destroy(&best);
          POOL_free(pool);
          return initVal;
        }
      }
      if (!warned) {
        COVER_warnOnSmallCorpus(dictBufferCapacity, ctx.nbDmers, displayLevel);
        warned = 1;
      }
      /* Loop through k reusing the same context */
      for (k = kMinK; k <= kMaxK; k += kStepSize) {
        /* Prepare the arguments */
        FASTCOVER_tryParameters_data_t *data = (FASTCOVER_tryParameters_data_t *)malloc(
            sizeof(FASTCOVER_tryParameters_data_t));
        LOCALDISPLAYLEVEL(displayLevel, 3, "k=%u\n", k);
        if (!data) {
          LOCALDISPLAYLEVEL(displayLevel, 1, "Failed to allocate parameters\n");
          COVER_best_destroy(&best);
          FASTCOVER_ctx_destroy(&ctx);
          POOL_free(pool);
          return ERROR(memory_allocation);
        }
        data->ctx = &ctx;
        data->best = &best;
        data->dictBufferCapacity = dictBufferCapacity;
        data->parameters = coverParams;
        data->parameters.k = k;
        data->parameters.d = d;
        data->parameters.splitPoint = splitPoint;
        data->parameters.steps = kSteps;
        data->parameters.shrinkDict = shrinkDict;
        data->parameters.zParams.notificationLevel = (unsigned)g_displayLevel;
        /* Check the parameters */
        if (!FASTCOVER_checkParameters(data->parameters, dictBufferCapacity,
                                       data->ctx->f, accel)) {
          DISPLAYLEVEL(1, "FASTCOVER parameters incorrect\n");
          free(data);
          continue;
        }
        /* Call the function and pass ownership of data to it */
        COVER_best_start(&best);
        if (pool) {
          POOL_add(pool, &FASTCOVER_tryParameters, data);
        } else {
          FASTCOVER_tryParameters(data);
        }
        /* Print status */
        LOCALDISPLAYUPDATE(displayLevel, 2, "\r%u%%       ",
                           (unsigned)((iteration * 100) / kIterations));
        ++iteration;
      }
      COVER_best_wait(&best);
      FASTCOVER_ctx_destroy(&ctx);
    }
    LOCALDISPLAYLEVEL(displayLevel, 2, "\r%79s\r", "");
    /* Fill the output buffer and parameters with output of the best parameters */
    {
      const size_t dictSize = best.dictSize;
      if (ZSTD_isError(best.compressedSize)) {
        const size_t compressedSize = best.compressedSize;
        COVER_best_destroy(&best);
        POOL_free(pool);
        return compressedSize;
      }
      FASTCOVER_convertToFastCoverParams(best.parameters, parameters, f, accel);
      memcpy(dictBuffer, best.dict, dictSize);
      COVER_best_destroy(&best);
      POOL_free(pool);
      return dictSize;
    }

}
/**** ended inlining dictBuilder/fastcover.c ****/
/**** start inlining dictBuilder/zdict.c ****/
/*
 * Copyright (c) Meta Platforms, Inc. and affiliates.
 * All rights reserved.
 *
 * This source code is licensed under both the BSD-style license (found in the
 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
 * in the COPYING file in the root directory of this source tree).
 * You may select, at your option, one of the above-listed licenses.
 */


/*-**************************************
*  Tuning parameters
****************************************/
#define MINRATIO 4   /* minimum nb of apparition to be selected in dictionary */
#define ZDICT_MAX_SAMPLES_SIZE (2000U << 20)
#define ZDICT_MIN_SAMPLES_SIZE (ZDICT_CONTENTSIZE_MIN * MINRATIO)


/*-**************************************
*  Compiler Options
****************************************/
/* Unix Large Files support (>4GB) */
#define _FILE_OFFSET_BITS 64
#if (defined(__sun__) && (!defined(__LP64__)))   /* Sun Solaris 32-bits requires specific definitions */
#  ifndef _LARGEFILE_SOURCE
#  define _LARGEFILE_SOURCE
#  endif
#elif ! defined(__LP64__)                        /* No point defining Large file for 64 bit */
#  ifndef _LARGEFILE64_SOURCE
#  define _LARGEFILE64_SOURCE
#  endif
#endif


/*-*************************************
*  Dependencies
***************************************/
#include <stdlib.h>        /* malloc, free */
#include <string.h>        /* memset */
#include <stdio.h>         /* fprintf, fopen, ftello64 */
#include <time.h>          /* clock */

#ifndef ZDICT_STATIC_LINKING_ONLY
#  define ZDICT_STATIC_LINKING_ONLY
#endif

/**** skipping file: ../common/mem.h ****/
/**** skipping file: ../common/fse.h ****/
/**** skipping file: ../common/huf.h ****/
/**** skipping file: ../common/zstd_internal.h ****/
/**** skipping file: ../common/xxhash.h ****/
/**** skipping file: ../compress/zstd_compress_internal.h ****/
/**** skipping file: ../zdict.h ****/
/**** skipping file: divsufsort.h ****/
/**** skipping file: ../common/bits.h ****/


/*-*************************************
*  Constants
***************************************/
#define KB *(1 <<10)
#define MB *(1 <<20)
#define GB *(1U<<30)

#define DICTLISTSIZE_DEFAULT 10000

#define NOISELENGTH 32

static const U32 g_selectivity_default = 9;


/*-*************************************
*  Console display
***************************************/
#undef  DISPLAY
#define DISPLAY(...)         do { fprintf(stderr, __VA_ARGS__); fflush( stderr ); } while (0)
#undef  DISPLAYLEVEL
#define DISPLAYLEVEL(l, ...) do { if (notificationLevel>=l) { DISPLAY(__VA_ARGS__); } } while (0)    /* 0 : no display;   1: errors;   2: default;  3: details;  4: debug */

static clock_t ZDICT_clockSpan(clock_t nPrevious) { return clock() - nPrevious; }

static void ZDICT_printHex(const void* ptr, size_t length)
{
    const BYTE* const b = (const BYTE*)ptr;
    size_t u;
    for (u=0; u<length; u++) {
        BYTE c = b[u];
        if (c<32 || c>126) c = '.';   /* non-printable char */
        DISPLAY("%c", c);
    }
}


/*-********************************************************
*  Helper functions
**********************************************************/
unsigned ZDICT_isError(size_t errorCode) { return ERR_isError(errorCode); }

const char* ZDICT_getErrorName(size_t errorCode) { return ERR_getErrorName(errorCode); }

unsigned ZDICT_getDictID(const void* dictBuffer, size_t dictSize)
{
    if (dictSize < 8) return 0;
    if (MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return 0;
    return MEM_readLE32((const char*)dictBuffer + 4);
}

size_t ZDICT_getDictHeaderSize(const void* dictBuffer, size_t dictSize)
{
    size_t headerSize;
    if (dictSize <= 8 || MEM_readLE32(dictBuffer) != ZSTD_MAGIC_DICTIONARY) return ERROR(dictionary_corrupted);

    {   ZSTD_compressedBlockState_t* bs = (ZSTD_compressedBlockState_t*)malloc(sizeof(ZSTD_compressedBlockState_t));
        U32* wksp = (U32*)malloc(HUF_WORKSPACE_SIZE);
        if (!bs || !wksp) {
            headerSize = ERROR(memory_allocation);
        } else {
            ZSTD_reset_compressedBlockState(bs);
            headerSize = ZSTD_loadCEntropy(bs, wksp, dictBuffer, dictSize);
        }

        free(bs);
        free(wksp);
    }

    return headerSize;
}

/*-********************************************************
*  Dictionary training functions
**********************************************************/
/*! ZDICT_count() :
    Count the nb of common bytes between 2 pointers.
    Note : this function presumes end of buffer followed by noisy guard band.
*/
static size_t ZDICT_count(const void* pIn, const void* pMatch)
{
    const char* const pStart = (const char*)pIn;
    for (;;) {
        size_t const diff = MEM_readST(pMatch) ^ MEM_readST(pIn);
        if (!diff) {
            pIn = (const char*)pIn+sizeof(size_t);
            pMatch = (const char*)pMatch+sizeof(size_t);
            continue;
        }
        pIn = (const char*)pIn+ZSTD_NbCommonBytes(diff);
        return (size_t)((const char*)pIn - pStart);
    }
}


typedef struct {
    U32 pos;
    U32 length;
    U32 savings;
} dictItem;

static void ZDICT_initDictItem(dictItem* d)
{
    d->pos = 1;
    d->length = 0;
    d->savings = (U32)(-1);
}


#define LLIMIT 64          /* heuristic determined experimentally */
#define MINMATCHLENGTH 7   /* heuristic determined experimentally */
static dictItem ZDICT_analyzePos(
                       BYTE* doneMarks,
                       const int* suffix, U32 start,
                       const void* buffer, U32 minRatio, U32 notificationLevel)
{
    U32 lengthList[LLIMIT] = {0};
    U32 cumulLength[LLIMIT] = {0};
    U32 savings[LLIMIT] = {0};
    const BYTE* b = (const BYTE*)buffer;
    size_t maxLength = LLIMIT;
    size_t pos = (size_t)suffix[start];
    U32 end = start;
    dictItem solution;

    /* init */
    memset(&solution, 0, sizeof(solution));
    doneMarks[pos] = 1;

    /* trivial repetition cases */
    if ( (MEM_read16(b+pos+0) == MEM_read16(b+pos+2))
       ||(MEM_read16(b+pos+1) == MEM_read16(b+pos+3))
       ||(MEM_read16(b+pos+2) == MEM_read16(b+pos+4)) ) {
        /* skip and mark segment */
        U16 const pattern16 = MEM_read16(b+pos+4);
        U32 u, patternEnd = 6;
        while (MEM_read16(b+pos+patternEnd) == pattern16) patternEnd+=2 ;
        if (b[pos+patternEnd] == b[pos+patternEnd-1]) patternEnd++;
        for (u=1; u<patternEnd; u++)
            doneMarks[pos+u] = 1;
        return solution;
    }

    /* look forward */
    {   size_t length;
        do {
            end++;
            length = ZDICT_count(b + pos, b + suffix[end]);
        } while (length >= MINMATCHLENGTH);
    }

    /* look backward */
    {   size_t length;
        do {
            length = ZDICT_count(b + pos, b + *(suffix+start-1));
            if (length >=MINMATCHLENGTH) start--;
        } while(length >= MINMATCHLENGTH);
    }

    /* exit if not found a minimum nb of repetitions */
    if (end-start < minRatio) {
        U32 idx;
        for(idx=start; idx<end; idx++)
            doneMarks[suffix[idx]] = 1;
        return solution;
    }

    {   int i;
        U32 mml;
        U32 refinedStart = start;
        U32 refinedEnd = end;

        DISPLAYLEVEL(4, "\n");
        DISPLAYLEVEL(4, "found %3u matches of length >= %i at pos %7u  ", (unsigned)(end-start), MINMATCHLENGTH, (unsigned)pos);
        DISPLAYLEVEL(4, "\n");

        for (mml = MINMATCHLENGTH ; ; mml++) {
            BYTE currentChar = 0;
            U32 currentCount = 0;
            U32 currentID = refinedStart;
            U32 id;
            U32 selectedCount = 0;
            U32 selectedID = currentID;
            for (id =refinedStart; id < refinedEnd; id++) {
                if (b[suffix[id] + mml] != currentChar) {
                    if (currentCount > selectedCount) {
                        selectedCount = currentCount;
                        selectedID = currentID;
                    }
                    currentID = id;
                    currentChar = b[ suffix[id] + mml];
                    currentCount = 0;
                }
                currentCount ++;
            }
            if (currentCount > selectedCount) {  /* for last */
                selectedCount = currentCount;
                selectedID = currentID;
            }

            if (selectedCount < minRatio)
                break;
            refinedStart = selectedID;
            refinedEnd = refinedStart + selectedCount;
        }

        /* evaluate gain based on new dict */
        start = refinedStart;
        pos = suffix[refinedStart];
        end = start;
        memset(lengthList, 0, sizeof(lengthList));

        /* look forward */
        {   size_t length;
            do {
                end++;
                length = ZDICT_count(b + pos, b + suffix[end]);
                if (length >= LLIMIT) length = LLIMIT-1;
                lengthList[length]++;
            } while (length >=MINMATCHLENGTH);
        }

        /* look backward */
        {   size_t length = MINMATCHLENGTH;
            while ((length >= MINMATCHLENGTH) & (start > 0)) {
                length = ZDICT_count(b + pos, b + suffix[start - 1]);
                if (length >= LLIMIT) length = LLIMIT - 1;
                lengthList[length]++;
                if (length >= MINMATCHLENGTH) start--;
            }
        }

        /* largest useful length */
        memset(cumulLength, 0, sizeof(cumulLength));
        cumulLength[maxLength-1] = lengthList[maxLength-1];
        for (i=(int)(maxLength-2); i>=0; i--)
            cumulLength[i] = cumulLength[i+1] + lengthList[i];

        for (i=LLIMIT-1; i>=MINMATCHLENGTH; i--) if (cumulLength[i]>=minRatio) break;
        maxLength = i;

        /* reduce maxLength in case of final into repetitive data */
        {   U32 l = (U32)maxLength;
            BYTE const c = b[pos + maxLength-1];
            while (b[pos+l-2]==c) l--;
            maxLength = l;
        }
        if (maxLength < MINMATCHLENGTH) return solution;   /* skip : no long-enough solution */

        /* calculate savings */
        savings[5] = 0;
        for (i=MINMATCHLENGTH; i<=(int)maxLength; i++)
            savings[i] = savings[i-1] + (lengthList[i] * (i-3));

        DISPLAYLEVEL(4, "Selected dict at position %u, of length %u : saves %u (ratio: %.2f)  \n",
                     (unsigned)pos, (unsigned)maxLength, (unsigned)savings[maxLength], (double)savings[maxLength] / (double)maxLength);

        solution.pos = (U32)pos;
        solution.length = (U32)maxLength;
        solution.savings = savings[maxLength];

        /* mark positions done */
        {   U32 id;
            for (id=start; id<end; id++) {
                U32 p, pEnd, length;
                U32 const testedPos = (U32)suffix[id];
                if (testedPos == pos)
                    length = solution.length;
                else {
                    length = (U32)ZDICT_count(b+pos, b+testedPos);
                    if (length > solution.length) length = solution.length;
                }
                pEnd = (U32)(testedPos + length);
                for (p=testedPos; p<pEnd; p++)
                    doneMarks[p] = 1;
    }   }   }

    return solution;
}


static int isIncluded(const void* in, const void* container, size_t length)
{
    const char* const ip = (const char*) in;
    const char* const into = (const char*) container;
    size_t u;

    for (u=0; u<length; u++) {  /* works because end of buffer is a noisy guard band */
        if (ip[u] != into[u]) break;
    }

    return u==length;
}

/*! ZDICT_tryMerge() :
    check if dictItem can be merged, do it if possible
    @return : id of destination elt, 0 if not merged
*/
static U32 ZDICT_tryMerge(dictItem* table, dictItem elt, U32 eltNbToSkip, const void* buffer)
{
    const U32 tableSize = table->pos;
    const U32 eltEnd = elt.pos + elt.length;
    const char* const buf = (const char*) buffer;

    /* tail overlap */
    U32 u; for (u=1; u<tableSize; u++) {
        if (u==eltNbToSkip) continue;
        if ((table[u].pos > elt.pos) && (table[u].pos <= eltEnd)) {  /* overlap, existing > new */
            /* append */
            U32 const addedLength = table[u].pos - elt.pos;
            table[u].length += addedLength;
            table[u].pos = elt.pos;
            table[u].savings += elt.savings * addedLength / elt.length;   /* rough approx */
            table[u].savings += elt.length / 8;    /* rough approx bonus */
            elt = table[u];
            /* sort : improve rank */
            while ((u>1) && (table[u-1].savings < elt.savings))
                table[u] = table[u-1], u--;
            table[u] = elt;
            return u;
    }   }

    /* front overlap */
    for (u=1; u<tableSize; u++) {
        if (u==eltNbToSkip) continue;

        if ((table[u].pos + table[u].length >= elt.pos) && (table[u].pos < elt.pos)) {  /* overlap, existing < new */
            /* append */
            int const addedLength = (int)eltEnd - (int)(table[u].pos + table[u].length);
            table[u].savings += elt.length / 8;    /* rough approx bonus */
            if (addedLength > 0) {   /* otherwise, elt fully included into existing */
                table[u].length += addedLength;
                table[u].savings += elt.savings * addedLength / elt.length;   /* rough approx */
            }
            /* sort : improve rank */
            elt = table[u];
            while ((u>1) && (table[u-1].savings < elt.savings))
                table[u] = table[u-1], u--;
            table[u] = elt;
            return u;
        }

        if (MEM_read64(buf + table[u].pos) == MEM_read64(buf + elt.pos + 1)) {
            if (isIncluded(buf + table[u].pos, buf + elt.pos + 1, table[u].length)) {
                size_t const addedLength = MAX( (int)elt.length - (int)table[u].length , 1 );
                table[u].pos = elt.pos;
                table[u].savings += (U32)(elt.savings * addedLength / elt.length);
                table[u].length = MIN(elt.length, table[u].length + 1);
                return u;
            }
        }
    }

    return 0;
}


static void ZDICT_removeDictItem(dictItem* table, U32 id)
{
    /* convention : table[0].pos stores nb of elts */
    U32 const max = table[0].pos;
    U32 u;
    if (!id) return;   /* protection, should never happen */
    for (u=id; u<max-1; u++)
        table[u] = table[u+1];
    table->pos--;
}


static void ZDICT_insertDictItem(dictItem* table, U32 maxSize, dictItem elt, const void* buffer)
{
    /* merge if possible */
    U32 mergeId = ZDICT_tryMerge(table, elt, 0, buffer);
    if (mergeId) {
        U32 newMerge = 1;
        while (newMerge) {
            newMerge = ZDICT_tryMerge(table, table[mergeId], mergeId, buffer);
            if (newMerge) ZDICT_removeDictItem(table, mergeId);
            mergeId = newMerge;
        }
        return;
    }

    /* insert */
    {   U32 current;
        U32 nextElt = table->pos;
        if (nextElt >= maxSize) nextElt = maxSize-1;
        current = nextElt-1;
        while (table[current].savings < elt.savings) {
            table[current+1] = table[current];
            current--;
        }
        table[current+1] = elt;
        table->pos = nextElt+1;
    }
}


static U32 ZDICT_dictSize(const dictItem* dictList)
{
    U32 u, dictSize = 0;
    for (u=1; u<dictList[0].pos; u++)
        dictSize += dictList[u].length;
    return dictSize;
}


static size_t ZDICT_trainBuffer_legacy(dictItem* dictList, U32 dictListSize,
                            const void* const buffer, size_t bufferSize,   /* buffer must end with noisy guard band */
                            const size_t* fileSizes, unsigned nbFiles,
                            unsigned minRatio, U32 notificationLevel)
{
    int* const suffix0 = (int*)malloc((bufferSize+2)*sizeof(*suffix0));
    int* const suffix = suffix0+1;
    U32* reverseSuffix = (U32*)malloc((bufferSize)*sizeof(*reverseSuffix));
    BYTE* doneMarks = (BYTE*)malloc((bufferSize+16)*sizeof(*doneMarks));   /* +16 for overflow security */
    U32* filePos = (U32*)malloc(nbFiles * sizeof(*filePos));
    size_t result = 0;
    clock_t displayClock = 0;
    clock_t const refreshRate = CLOCKS_PER_SEC * 3 / 10;

#   undef  DISPLAYUPDATE
#   define DISPLAYUPDATE(l, ...)                                   \
        do {                                                       \
            if (notificationLevel>=l) {                            \
                if (ZDICT_clockSpan(displayClock) > refreshRate) { \
                    displayClock = clock();                        \
                    DISPLAY(__VA_ARGS__);                          \
                }                                                  \
                if (notificationLevel>=4) fflush(stderr);          \
            }                                                      \
        } while (0)

    /* init */
    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
    if (!suffix0 || !reverseSuffix || !doneMarks || !filePos) {
        result = ERROR(memory_allocation);
        goto _cleanup;
    }
    if (minRatio < MINRATIO) minRatio = MINRATIO;
    memset(doneMarks, 0, bufferSize+16);

    /* limit sample set size (divsufsort limitation)*/
    if (bufferSize > ZDICT_MAX_SAMPLES_SIZE) DISPLAYLEVEL(3, "sample set too large : reduced to %u MB ...\n", (unsigned)(ZDICT_MAX_SAMPLES_SIZE>>20));
    while (bufferSize > ZDICT_MAX_SAMPLES_SIZE) bufferSize -= fileSizes[--nbFiles];

    /* sort */
    DISPLAYLEVEL(2, "sorting %u files of total size %u MB ...\n", nbFiles, (unsigned)(bufferSize>>20));
    {   int const divSuftSortResult = divsufsort((const unsigned char*)buffer, suffix, (int)bufferSize, 0);
        if (divSuftSortResult != 0) { result = ERROR(GENERIC); goto _cleanup; }
    }
    suffix[bufferSize] = (int)bufferSize;   /* leads into noise */
    suffix0[0] = (int)bufferSize;           /* leads into noise */
    /* build reverse suffix sort */
    {   size_t pos;
        for (pos=0; pos < bufferSize; pos++)
            reverseSuffix[suffix[pos]] = (U32)pos;
        /* note filePos tracks borders between samples.
           It's not used at this stage, but planned to become useful in a later update */
        filePos[0] = 0;
        for (pos=1; pos<nbFiles; pos++)
            filePos[pos] = (U32)(filePos[pos-1] + fileSizes[pos-1]);
    }

    DISPLAYLEVEL(2, "finding patterns ... \n");
    DISPLAYLEVEL(3, "minimum ratio : %u \n", minRatio);

    {   U32 cursor; for (cursor=0; cursor < bufferSize; ) {
            dictItem solution;
            if (doneMarks[cursor]) { cursor++; continue; }
            solution = ZDICT_analyzePos(doneMarks, suffix, reverseSuffix[cursor], buffer, minRatio, notificationLevel);
            if (solution.length==0) { cursor++; continue; }
            ZDICT_insertDictItem(dictList, dictListSize, solution, buffer);
            cursor += solution.length;
            DISPLAYUPDATE(2, "\r%4.2f %% \r", (double)cursor / (double)bufferSize * 100.0);
    }   }

_cleanup:
    free(suffix0);
    free(reverseSuffix);
    free(doneMarks);
    free(filePos);
    return result;
}


static void ZDICT_fillNoise(void* buffer, size_t length)
{
    unsigned const prime1 = 2654435761U;
    unsigned const prime2 = 2246822519U;
    unsigned acc = prime1;
    size_t p=0;
    for (p=0; p<length; p++) {
        acc *= prime2;
        ((unsigned char*)buffer)[p] = (unsigned char)(acc >> 21);
    }
}


typedef struct
{
    ZSTD_CDict* dict;    /* dictionary */
    ZSTD_CCtx* zc;     /* working context */
    void* workPlace;   /* must be ZSTD_BLOCKSIZE_MAX allocated */
} EStats_ress_t;

#define MAXREPOFFSET 1024

static void ZDICT_countEStats(EStats_ress_t esr, const ZSTD_parameters* params,
                              unsigned* countLit, unsigned* offsetcodeCount, unsigned* matchlengthCount, unsigned* litlengthCount, U32* repOffsets,
                              const void* src, size_t srcSize,
                              U32 notificationLevel)
{
    size_t const blockSizeMax = MIN (ZSTD_BLOCKSIZE_MAX, 1 << params->cParams.windowLog);
    size_t cSize;

    if (srcSize > blockSizeMax) srcSize = blockSizeMax;   /* protection vs large samples */
    {   size_t const errorCode = ZSTD_compressBegin_usingCDict_deprecated(esr.zc, esr.dict);
        if (ZSTD_isError(errorCode)) { DISPLAYLEVEL(1, "warning : ZSTD_compressBegin_usingCDict failed \n"); return; }

    }
    cSize = ZSTD_compressBlock_deprecated(esr.zc, esr.workPlace, ZSTD_BLOCKSIZE_MAX, src, srcSize);
    if (ZSTD_isError(cSize)) { DISPLAYLEVEL(3, "warning : could not compress sample size %u \n", (unsigned)srcSize); return; }

    if (cSize) {  /* if == 0; block is not compressible */
        const SeqStore_t* const seqStorePtr = ZSTD_getSeqStore(esr.zc);

        /* literals stats */
        {   const BYTE* bytePtr;
            for(bytePtr = seqStorePtr->litStart; bytePtr < seqStorePtr->lit; bytePtr++)
                countLit[*bytePtr]++;
        }

        /* seqStats */
        {   U32 const nbSeq = (U32)(seqStorePtr->sequences - seqStorePtr->sequencesStart);
            ZSTD_seqToCodes(seqStorePtr);

            {   const BYTE* codePtr = seqStorePtr->ofCode;
                U32 u;
                for (u=0; u<nbSeq; u++) offsetcodeCount[codePtr[u]]++;
            }

            {   const BYTE* codePtr = seqStorePtr->mlCode;
                U32 u;
                for (u=0; u<nbSeq; u++) matchlengthCount[codePtr[u]]++;
            }

            {   const BYTE* codePtr = seqStorePtr->llCode;
                U32 u;
                for (u=0; u<nbSeq; u++) litlengthCount[codePtr[u]]++;
            }

            if (nbSeq >= 2) { /* rep offsets */
                const SeqDef* const seq = seqStorePtr->sequencesStart;
                U32 offset1 = seq[0].offBase - ZSTD_REP_NUM;
                U32 offset2 = seq[1].offBase - ZSTD_REP_NUM;
                if (offset1 >= MAXREPOFFSET) offset1 = 0;
                if (offset2 >= MAXREPOFFSET) offset2 = 0;
                repOffsets[offset1] += 3;
                repOffsets[offset2] += 1;
    }   }   }
}

static size_t ZDICT_totalSampleSize(const size_t* fileSizes, unsigned nbFiles)
{
    size_t total=0;
    unsigned u;
    for (u=0; u<nbFiles; u++) total += fileSizes[u];
    return total;
}

typedef struct { U32 offset; U32 count; } offsetCount_t;

static void ZDICT_insertSortCount(offsetCount_t table[ZSTD_REP_NUM+1], U32 val, U32 count)
{
    U32 u;
    table[ZSTD_REP_NUM].offset = val;
    table[ZSTD_REP_NUM].count = count;
    for (u=ZSTD_REP_NUM; u>0; u--) {
        offsetCount_t tmp;
        if (table[u-1].count >= table[u].count) break;
        tmp = table[u-1];
        table[u-1] = table[u];
        table[u] = tmp;
    }
}

/* ZDICT_flatLit() :
 * rewrite `countLit` to contain a mostly flat but still compressible distribution of literals.
 * necessary to avoid generating a non-compressible distribution that HUF_writeCTable() cannot encode.
 */
static void ZDICT_flatLit(unsigned* countLit)
{
    int u;
    for (u=1; u<256; u++) countLit[u] = 2;
    countLit[0]   = 4;
    countLit[253] = 1;
    countLit[254] = 1;
}

#define OFFCODE_MAX 30  /* only applicable to first block */
static size_t ZDICT_analyzeEntropy(void*  dstBuffer, size_t maxDstSize,
                                   int compressionLevel,
                             const void*  srcBuffer, const size_t* fileSizes, unsigned nbFiles,
                             const void* dictBuffer, size_t  dictBufferSize,
                                   unsigned notificationLevel)
{
    unsigned countLit[256];
    HUF_CREATE_STATIC_CTABLE(hufTable, 255);
    unsigned offcodeCount[OFFCODE_MAX+1];
    short offcodeNCount[OFFCODE_MAX+1];
    U32 offcodeMax = ZSTD_highbit32((U32)(dictBufferSize + 128 KB));
    unsigned matchLengthCount[MaxML+1];
    short matchLengthNCount[MaxML+1];
    unsigned litLengthCount[MaxLL+1];
    short litLengthNCount[MaxLL+1];
    U32 repOffset[MAXREPOFFSET];
    offsetCount_t bestRepOffset[ZSTD_REP_NUM+1];
    EStats_ress_t esr = { NULL, NULL, NULL };
    ZSTD_parameters params;
    U32 u, huffLog = 11, Offlog = OffFSELog, mlLog = MLFSELog, llLog = LLFSELog, total;
    size_t pos = 0, errorCode;
    size_t eSize = 0;
    size_t const totalSrcSize = ZDICT_totalSampleSize(fileSizes, nbFiles);
    size_t const averageSampleSize = totalSrcSize / (nbFiles + !nbFiles);
    BYTE* dstPtr = (BYTE*)dstBuffer;
    U32 wksp[HUF_CTABLE_WORKSPACE_SIZE_U32];

    /* init */
    DEBUGLOG(4, "ZDICT_analyzeEntropy");
    if (offcodeMax>OFFCODE_MAX) { eSize = ERROR(dictionaryCreation_failed); goto _cleanup; }   /* too large dictionary */
    for (u=0; u<256; u++) countLit[u] = 1;   /* any character must be described */
    for (u=0; u<=offcodeMax; u++) offcodeCount[u] = 1;
    for (u=0; u<=MaxML; u++) matchLengthCount[u] = 1;
    for (u=0; u<=MaxLL; u++) litLengthCount[u] = 1;
    memset(repOffset, 0, sizeof(repOffset));
    repOffset[1] = repOffset[4] = repOffset[8] = 1;
    memset(bestRepOffset, 0, sizeof(bestRepOffset));
    if (compressionLevel==0) compressionLevel = ZSTD_CLEVEL_DEFAULT;
    params = ZSTD_getParams(compressionLevel, averageSampleSize, dictBufferSize);

    esr.dict = ZSTD_createCDict_advanced(dictBuffer, dictBufferSize, ZSTD_dlm_byRef, ZSTD_dct_rawContent, params.cParams, ZSTD_defaultCMem);
    esr.zc = ZSTD_createCCtx();
    esr.workPlace = malloc(ZSTD_BLOCKSIZE_MAX);
    if (!esr.dict || !esr.zc || !esr.workPlace) {
        eSize = ERROR(memory_allocation);
        DISPLAYLEVEL(1, "Not enough memory \n");
        goto _cleanup;
    }

    /* collect stats on all samples */
    for (u=0; u<nbFiles; u++) {
        ZDICT_countEStats(esr, &params,
                          countLit, offcodeCount, matchLengthCount, litLengthCount, repOffset,
                         (const char*)srcBuffer + pos, fileSizes[u],
                          notificationLevel);
        pos += fileSizes[u];
    }

    if (notificationLevel >= 4) {
        /* writeStats */
        DISPLAYLEVEL(4, "Offset Code Frequencies : \n");
        for (u=0; u<=offcodeMax; u++) {
            DISPLAYLEVEL(4, "%2u :%7u \n", u, offcodeCount[u]);
    }   }

    /* analyze, build stats, starting with literals */
    {   size_t maxNbBits = HUF_buildCTable_wksp(hufTable, countLit, 255, huffLog, wksp, sizeof(wksp));
        if (HUF_isError(maxNbBits)) {
            eSize = maxNbBits;
            DISPLAYLEVEL(1, " HUF_buildCTable error \n");
            goto _cleanup;
        }
        if (maxNbBits==8) {  /* not compressible : will fail on HUF_writeCTable() */
            DISPLAYLEVEL(2, "warning : pathological dataset : literals are not compressible : samples are noisy or too regular \n");
            ZDICT_flatLit(countLit);  /* replace distribution by a fake "mostly flat but still compressible" distribution, that HUF_writeCTable() can encode */
            maxNbBits = HUF_buildCTable_wksp(hufTable, countLit, 255, huffLog, wksp, sizeof(wksp));
            assert(maxNbBits==9);
        }
        huffLog = (U32)maxNbBits;
    }

    /* looking for most common first offsets */
    {   U32 offset;
        for (offset=1; offset<MAXREPOFFSET; offset++)
            ZDICT_insertSortCount(bestRepOffset, offset, repOffset[offset]);
    }
    /* note : the result of this phase should be used to better appreciate the impact on statistics */

    total=0; for (u=0; u<=offcodeMax; u++) total+=offcodeCount[u];
    errorCode = FSE_normalizeCount(offcodeNCount, Offlog, offcodeCount, total, offcodeMax, /* useLowProbCount */ 1);
    if (FSE_isError(errorCode)) {
        eSize = errorCode;
        DISPLAYLEVEL(1, "FSE_normalizeCount error with offcodeCount \n");
        goto _cleanup;
    }
    Offlog = (U32)errorCode;

    total=0; for (u=0; u<=MaxML; u++) total+=matchLengthCount[u];
    errorCode = FSE_normalizeCount(matchLengthNCount, mlLog, matchLengthCount, total, MaxML, /* useLowProbCount */ 1);
    if (FSE_isError(errorCode)) {
        eSize = errorCode;
        DISPLAYLEVEL(1, "FSE_normalizeCount error with matchLengthCount \n");
        goto _cleanup;
    }
    mlLog = (U32)errorCode;

    total=0; for (u=0; u<=MaxLL; u++) total+=litLengthCount[u];
    errorCode = FSE_normalizeCount(litLengthNCount, llLog, litLengthCount, total, MaxLL, /* useLowProbCount */ 1);
    if (FSE_isError(errorCode)) {
        eSize = errorCode;
        DISPLAYLEVEL(1, "FSE_normalizeCount error with litLengthCount \n");
        goto _cleanup;
    }
    llLog = (U32)errorCode;

    /* write result to buffer */
    {   size_t const hhSize = HUF_writeCTable_wksp(dstPtr, maxDstSize, hufTable, 255, huffLog, wksp, sizeof(wksp));
        if (HUF_isError(hhSize)) {
            eSize = hhSize;
            DISPLAYLEVEL(1, "HUF_writeCTable error \n");
            goto _cleanup;
        }
        dstPtr += hhSize;
        maxDstSize -= hhSize;
        eSize += hhSize;
    }

    {   size_t const ohSize = FSE_writeNCount(dstPtr, maxDstSize, offcodeNCount, OFFCODE_MAX, Offlog);
        if (FSE_isError(ohSize)) {
            eSize = ohSize;
            DISPLAYLEVEL(1, "FSE_writeNCount error with offcodeNCount \n");
            goto _cleanup;
        }
        dstPtr += ohSize;
        maxDstSize -= ohSize;
        eSize += ohSize;
    }

    {   size_t const mhSize = FSE_writeNCount(dstPtr, maxDstSize, matchLengthNCount, MaxML, mlLog);
        if (FSE_isError(mhSize)) {
            eSize = mhSize;
            DISPLAYLEVEL(1, "FSE_writeNCount error with matchLengthNCount \n");
            goto _cleanup;
        }
        dstPtr += mhSize;
        maxDstSize -= mhSize;
        eSize += mhSize;
    }

    {   size_t const lhSize = FSE_writeNCount(dstPtr, maxDstSize, litLengthNCount, MaxLL, llLog);
        if (FSE_isError(lhSize)) {
            eSize = lhSize;
            DISPLAYLEVEL(1, "FSE_writeNCount error with litlengthNCount \n");
            goto _cleanup;
        }
        dstPtr += lhSize;
        maxDstSize -= lhSize;
        eSize += lhSize;
    }

    if (maxDstSize<12) {
        eSize = ERROR(dstSize_tooSmall);
        DISPLAYLEVEL(1, "not enough space to write RepOffsets \n");
        goto _cleanup;
    }
# if 0
    MEM_writeLE32(dstPtr+0, bestRepOffset[0].offset);
    MEM_writeLE32(dstPtr+4, bestRepOffset[1].offset);
    MEM_writeLE32(dstPtr+8, bestRepOffset[2].offset);
#else
    /* at this stage, we don't use the result of "most common first offset",
     * as the impact of statistics is not properly evaluated */
    MEM_writeLE32(dstPtr+0, repStartValue[0]);
    MEM_writeLE32(dstPtr+4, repStartValue[1]);
    MEM_writeLE32(dstPtr+8, repStartValue[2]);
#endif
    eSize += 12;

_cleanup:
    ZSTD_freeCDict(esr.dict);
    ZSTD_freeCCtx(esr.zc);
    free(esr.workPlace);

    return eSize;
}


/**
 * @returns the maximum repcode value
 */
static U32 ZDICT_maxRep(U32 const reps[ZSTD_REP_NUM])
{
    U32 maxRep = reps[0];
    int r;
    for (r = 1; r < ZSTD_REP_NUM; ++r)
        maxRep = MAX(maxRep, reps[r]);
    return maxRep;
}

size_t ZDICT_finalizeDictionary(void* dictBuffer, size_t dictBufferCapacity,
                          const void* customDictContent, size_t dictContentSize,
                          const void* samplesBuffer, const size_t* samplesSizes,
                          unsigned nbSamples, ZDICT_params_t params)
{
    size_t hSize;
#define HBUFFSIZE 256   /* should prove large enough for all entropy headers */
    BYTE header[HBUFFSIZE];
    int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
    U32 const notificationLevel = params.notificationLevel;
    /* The final dictionary content must be at least as large as the largest repcode */
    size_t const minContentSize = (size_t)ZDICT_maxRep(repStartValue);
    size_t paddingSize;

    /* check conditions */
    DEBUGLOG(4, "ZDICT_finalizeDictionary");
    if (dictBufferCapacity < dictContentSize) return ERROR(dstSize_tooSmall);
    if (dictBufferCapacity < ZDICT_DICTSIZE_MIN) return ERROR(dstSize_tooSmall);

    /* dictionary header */
    MEM_writeLE32(header, ZSTD_MAGIC_DICTIONARY);
    {   U64 const randomID = XXH64(customDictContent, dictContentSize, 0);
        U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
        U32 const dictID = params.dictID ? params.dictID : compliantID;
        MEM_writeLE32(header+4, dictID);
    }
    hSize = 8;

    /* entropy tables */
    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
    DISPLAYLEVEL(2, "statistics ... \n");
    {   size_t const eSize = ZDICT_analyzeEntropy(header+hSize, HBUFFSIZE-hSize,
                                  compressionLevel,
                                  samplesBuffer, samplesSizes, nbSamples,
                                  customDictContent, dictContentSize,
                                  notificationLevel);
        if (ZDICT_isError(eSize)) return eSize;
        hSize += eSize;
    }

    /* Shrink the content size if it doesn't fit in the buffer */
    if (hSize + dictContentSize > dictBufferCapacity) {
        dictContentSize = dictBufferCapacity - hSize;
    }

    /* Pad the dictionary content with zeros if it is too small */
    if (dictContentSize < minContentSize) {
        RETURN_ERROR_IF(hSize + minContentSize > dictBufferCapacity, dstSize_tooSmall,
                        "dictBufferCapacity too small to fit max repcode");
        paddingSize = minContentSize - dictContentSize;
    } else {
        paddingSize = 0;
    }

    {
        size_t const dictSize = hSize + paddingSize + dictContentSize;

        /* The dictionary consists of the header, optional padding, and the content.
         * The padding comes before the content because the "best" position in the
         * dictionary is the last byte.
         */
        BYTE* const outDictHeader = (BYTE*)dictBuffer;
        BYTE* const outDictPadding = outDictHeader + hSize;
        BYTE* const outDictContent = outDictPadding + paddingSize;

        assert(dictSize <= dictBufferCapacity);
        assert(outDictContent + dictContentSize == (BYTE*)dictBuffer + dictSize);

        /* First copy the customDictContent into its final location.
         * `customDictContent` and `dictBuffer` may overlap, so we must
         * do this before any other writes into the output buffer.
         * Then copy the header & padding into the output buffer.
         */
        memmove(outDictContent, customDictContent, dictContentSize);
        memcpy(outDictHeader, header, hSize);
        memset(outDictPadding, 0, paddingSize);

        return dictSize;
    }
}


static size_t ZDICT_addEntropyTablesFromBuffer_advanced(
        void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
        const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
        ZDICT_params_t params)
{
    int const compressionLevel = (params.compressionLevel == 0) ? ZSTD_CLEVEL_DEFAULT : params.compressionLevel;
    U32 const notificationLevel = params.notificationLevel;
    size_t hSize = 8;

    /* calculate entropy tables */
    DISPLAYLEVEL(2, "\r%70s\r", "");   /* clean display line */
    DISPLAYLEVEL(2, "statistics ... \n");
    {   size_t const eSize = ZDICT_analyzeEntropy((char*)dictBuffer+hSize, dictBufferCapacity-hSize,
                                  compressionLevel,
                                  samplesBuffer, samplesSizes, nbSamples,
                                  (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize,
                                  notificationLevel);
        if (ZDICT_isError(eSize)) return eSize;
        hSize += eSize;
    }

    /* add dictionary header (after entropy tables) */
    MEM_writeLE32(dictBuffer, ZSTD_MAGIC_DICTIONARY);
    {   U64 const randomID = XXH64((char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize, 0);
        U32 const compliantID = (randomID % ((1U<<31)-32768)) + 32768;
        U32 const dictID = params.dictID ? params.dictID : compliantID;
        MEM_writeLE32((char*)dictBuffer+4, dictID);
    }

    if (hSize + dictContentSize < dictBufferCapacity)
        memmove((char*)dictBuffer + hSize, (char*)dictBuffer + dictBufferCapacity - dictContentSize, dictContentSize);
    return MIN(dictBufferCapacity, hSize+dictContentSize);
}

/*! ZDICT_trainFromBuffer_unsafe_legacy() :
*   Warning : `samplesBuffer` must be followed by noisy guard band !!!
*   @return : size of dictionary, or an error code which can be tested with ZDICT_isError()
*/
static size_t ZDICT_trainFromBuffer_unsafe_legacy(
                            void* dictBuffer, size_t maxDictSize,
                            const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
                            ZDICT_legacy_params_t params)
{
    U32 const dictListSize = MAX(MAX(DICTLISTSIZE_DEFAULT, nbSamples), (U32)(maxDictSize/16));
    dictItem* const dictList = (dictItem*)malloc(dictListSize * sizeof(*dictList));
    unsigned const selectivity = params.selectivityLevel == 0 ? g_selectivity_default : params.selectivityLevel;
    unsigned const minRep = (selectivity > 30) ? MINRATIO : nbSamples >> selectivity;
    size_t const targetDictSize = maxDictSize;
    size_t const samplesBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
    size_t dictSize = 0;
    U32 const notificationLevel = params.zParams.notificationLevel;

    /* checks */
    if (!dictList) return ERROR(memory_allocation);
    if (maxDictSize < ZDICT_DICTSIZE_MIN) { free(dictList); return ERROR(dstSize_tooSmall); }   /* requested dictionary size is too small */
    if (samplesBuffSize < ZDICT_MIN_SAMPLES_SIZE) { free(dictList); return ERROR(dictionaryCreation_failed); }   /* not enough source to create dictionary */

    /* init */
    ZDICT_initDictItem(dictList);

    /* build dictionary */
    ZDICT_trainBuffer_legacy(dictList, dictListSize,
                       samplesBuffer, samplesBuffSize,
                       samplesSizes, nbSamples,
                       minRep, notificationLevel);

    /* display best matches */
    if (params.zParams.notificationLevel>= 3) {
        unsigned const nb = MIN(25, dictList[0].pos);
        unsigned const dictContentSize = ZDICT_dictSize(dictList);
        unsigned u;
        DISPLAYLEVEL(3, "\n %u segments found, of total size %u \n", (unsigned)dictList[0].pos-1, dictContentSize);
        DISPLAYLEVEL(3, "list %u best segments \n", nb-1);
        for (u=1; u<nb; u++) {
            unsigned const pos = dictList[u].pos;
            unsigned const length = dictList[u].length;
            U32 const printedLength = MIN(40, length);
            if ((pos > samplesBuffSize) || ((pos + length) > samplesBuffSize)) {
                free(dictList);
                return ERROR(GENERIC);   /* should never happen */
            }
            DISPLAYLEVEL(3, "%3u:%3u bytes at pos %8u, savings %7u bytes |",
                         u, length, pos, (unsigned)dictList[u].savings);
            ZDICT_printHex((const char*)samplesBuffer+pos, printedLength);
            DISPLAYLEVEL(3, "| \n");
    }   }


    /* create dictionary */
    {   unsigned dictContentSize = ZDICT_dictSize(dictList);
        if (dictContentSize < ZDICT_CONTENTSIZE_MIN) { free(dictList); return ERROR(dictionaryCreation_failed); }   /* dictionary content too small */
        if (dictContentSize < targetDictSize/4) {
            DISPLAYLEVEL(2, "!  warning : selected content significantly smaller than requested (%u < %u) \n", dictContentSize, (unsigned)maxDictSize);
            if (samplesBuffSize < 10 * targetDictSize)
                DISPLAYLEVEL(2, "!  consider increasing the number of samples (total size : %u MB)\n", (unsigned)(samplesBuffSize>>20));
            if (minRep > MINRATIO) {
                DISPLAYLEVEL(2, "!  consider increasing selectivity to produce larger dictionary (-s%u) \n", selectivity+1);
                DISPLAYLEVEL(2, "!  note : larger dictionaries are not necessarily better, test its efficiency on samples \n");
            }
        }

        if ((dictContentSize > targetDictSize*3) && (nbSamples > 2*MINRATIO) && (selectivity>1)) {
            unsigned proposedSelectivity = selectivity-1;
            while ((nbSamples >> proposedSelectivity) <= MINRATIO) { proposedSelectivity--; }
            DISPLAYLEVEL(2, "!  note : calculated dictionary significantly larger than requested (%u > %u) \n", dictContentSize, (unsigned)maxDictSize);
            DISPLAYLEVEL(2, "!  consider increasing dictionary size, or produce denser dictionary (-s%u) \n", proposedSelectivity);
            DISPLAYLEVEL(2, "!  always test dictionary efficiency on real samples \n");
        }

        /* limit dictionary size */
        {   U32 const max = dictList->pos;   /* convention : nb of useful elts within dictList */
            U32 currentSize = 0;
            U32 n; for (n=1; n<max; n++) {
                currentSize += dictList[n].length;
                if (currentSize > targetDictSize) { currentSize -= dictList[n].length; break; }
            }
            dictList->pos = n;
            dictContentSize = currentSize;
        }

        /* build dict content */
        {   U32 u;
            BYTE* ptr = (BYTE*)dictBuffer + maxDictSize;
            for (u=1; u<dictList->pos; u++) {
                U32 l = dictList[u].length;
                ptr -= l;
                if (ptr<(BYTE*)dictBuffer) { free(dictList); return ERROR(GENERIC); }   /* should not happen */
                memcpy(ptr, (const char*)samplesBuffer+dictList[u].pos, l);
        }   }

        dictSize = ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, maxDictSize,
                                                             samplesBuffer, samplesSizes, nbSamples,
                                                             params.zParams);
    }

    /* clean up */
    free(dictList);
    return dictSize;
}


/* ZDICT_trainFromBuffer_legacy() :
 * issue : samplesBuffer need to be followed by a noisy guard band.
 * work around : duplicate the buffer, and add the noise */
size_t ZDICT_trainFromBuffer_legacy(void* dictBuffer, size_t dictBufferCapacity,
                              const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples,
                              ZDICT_legacy_params_t params)
{
    size_t result;
    void* newBuff;
    size_t const sBuffSize = ZDICT_totalSampleSize(samplesSizes, nbSamples);
    if (sBuffSize < ZDICT_MIN_SAMPLES_SIZE) return 0;   /* not enough content => no dictionary */

    newBuff = malloc(sBuffSize + NOISELENGTH);
    if (!newBuff) return ERROR(memory_allocation);

    memcpy(newBuff, samplesBuffer, sBuffSize);
    ZDICT_fillNoise((char*)newBuff + sBuffSize, NOISELENGTH);   /* guard band, for end of buffer condition */

    result =
        ZDICT_trainFromBuffer_unsafe_legacy(dictBuffer, dictBufferCapacity, newBuff,
                                            samplesSizes, nbSamples, params);
    free(newBuff);
    return result;
}


size_t ZDICT_trainFromBuffer(void* dictBuffer, size_t dictBufferCapacity,
                             const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
{
    ZDICT_fastCover_params_t params;
    DEBUGLOG(3, "ZDICT_trainFromBuffer");
    memset(&params, 0, sizeof(params));
    params.d = 8;
    params.steps = 4;
    /* Use default level since no compression level information is available */
    params.zParams.compressionLevel = ZSTD_CLEVEL_DEFAULT;
#if defined(DEBUGLEVEL) && (DEBUGLEVEL>=1)
    params.zParams.notificationLevel = DEBUGLEVEL;
#endif
    return ZDICT_optimizeTrainFromBuffer_fastCover(dictBuffer, dictBufferCapacity,
                                               samplesBuffer, samplesSizes, nbSamples,
                                               &params);
}

size_t ZDICT_addEntropyTablesFromBuffer(void* dictBuffer, size_t dictContentSize, size_t dictBufferCapacity,
                                  const void* samplesBuffer, const size_t* samplesSizes, unsigned nbSamples)
{
    ZDICT_params_t params;
    memset(&params, 0, sizeof(params));
    return ZDICT_addEntropyTablesFromBuffer_advanced(dictBuffer, dictContentSize, dictBufferCapacity,
                                                     samplesBuffer, samplesSizes, nbSamples,
                                                     params);
}
/**** ended inlining dictBuilder/zdict.c ****/