xbyak/xbyak.h

#pragma once
#ifndef XBYAK_XBYAK_H_
#define XBYAK_XBYAK_H_
/*!
	@file xbyak.h
	@brief Xbyak ; JIT assembler for x86(IA32)/x64 by C++
	@author herumi
	@url https://github.com/herumi/xbyak
	@note modified new BSD license
	http://opensource.org/licenses/BSD-3-Clause
*/
#if (not +0) && !defined(XBYAK_NO_OP_NAMES) // trick to detect whether 'not' is operator or not
	#define XBYAK_NO_OP_NAMES
#endif

#include <stdio.h> // for debug print
#include <assert.h>
#include <list>
#include <string>
#include <algorithm>
#ifndef NDEBUG
#include <iostream>
#endif

// #define XBYAK_DISABLE_AVX512

#if !defined(XBYAK_USE_MMAP_ALLOCATOR) && !defined(XBYAK_DONT_USE_MMAP_ALLOCATOR)
	#define XBYAK_USE_MMAP_ALLOCATOR
#endif
#if !defined(__GNUC__) || defined(__MINGW32__)
	#undef XBYAK_USE_MMAP_ALLOCATOR
#endif

#ifdef __GNUC__
	#define XBYAK_GNUC_PREREQ(major, minor) ((__GNUC__) * 100 + (__GNUC_MINOR__) >= (major) * 100 + (minor))
#else
	#define XBYAK_GNUC_PREREQ(major, minor) 0
#endif

// This covers -std=(gnu|c)++(0x|11|1y), -stdlib=libc++, and modern Microsoft.
#if ((defined(_MSC_VER) && (_MSC_VER >= 1600)) || defined(_LIBCPP_VERSION) ||\
	 			 ((__cplusplus >= 201103) || defined(__GXX_EXPERIMENTAL_CXX0X__)))
	#include <unordered_set>
	#define XBYAK_STD_UNORDERED_SET std::unordered_set
	#include <unordered_map>
	#define XBYAK_STD_UNORDERED_MAP std::unordered_map
	#define XBYAK_STD_UNORDERED_MULTIMAP std::unordered_multimap

/*
	Clang/llvm-gcc and ICC-EDG in 'GCC-mode' always claim to be GCC 4.2, using
	libstdcxx 20070719 (from GCC 4.2.1, the last GPL 2 version).
*/
#elif XBYAK_GNUC_PREREQ(4, 5) || (XBYAK_GNUC_PREREQ(4, 2) && __GLIBCXX__ >= 20070719) || defined(__INTEL_COMPILER) || defined(__llvm__)
	#include <tr1/unordered_set>
	#define XBYAK_STD_UNORDERED_SET std::tr1::unordered_set
	#include <tr1/unordered_map>
	#define XBYAK_STD_UNORDERED_MAP std::tr1::unordered_map
	#define XBYAK_STD_UNORDERED_MULTIMAP std::tr1::unordered_multimap

#elif defined(_MSC_VER) && (_MSC_VER >= 1500) && (_MSC_VER < 1600)
	#include <unordered_set>
	#define XBYAK_STD_UNORDERED_SET std::tr1::unordered_set
	#include <unordered_map>
	#define XBYAK_STD_UNORDERED_MAP std::tr1::unordered_map
	#define XBYAK_STD_UNORDERED_MULTIMAP std::tr1::unordered_multimap

#else
	#include <set>
	#define XBYAK_STD_UNORDERED_SET std::set
	#include <map>
	#define XBYAK_STD_UNORDERED_MAP std::map
	#define XBYAK_STD_UNORDERED_MULTIMAP std::multimap
#endif
#ifdef _WIN32
	#ifndef WIN32_LEAN_AND_MEAN
		#define WIN32_LEAN_AND_MEAN
	#endif
	#include <windows.h>
	#include <malloc.h>
	#ifdef _MSC_VER
		#define XBYAK_TLS __declspec(thread)
	#else
		#define XBYAK_TLS __thread
	#endif
#elif defined(__GNUC__)
	#include <unistd.h>
	#include <sys/mman.h>
	#include <stdlib.h>
	#define XBYAK_TLS __thread
#endif
#if defined(__APPLE__) && !defined(XBYAK_DONT_USE_MAP_JIT)
	#define XBYAK_USE_MAP_JIT
	#include <sys/sysctl.h>
	#ifndef MAP_JIT
		#define MAP_JIT 0x800
	#endif
#endif
#if !defined(_MSC_VER) || (_MSC_VER >= 1600)
	#include <stdint.h>
#endif

// MFD_CLOEXEC defined only linux 3.17 or later.
// Android wraps the memfd_create syscall from API version 30.
#if !defined(MFD_CLOEXEC) || (defined(__ANDROID__) && __ANDROID_API__ < 30)
	#undef XBYAK_USE_MEMFD
#endif

#if defined(_WIN64) || defined(__MINGW64__) || (defined(__CYGWIN__) && defined(__x86_64__))
	#define XBYAK64_WIN
#elif defined(__x86_64__)
	#define XBYAK64_GCC
#endif
#if !defined(XBYAK64) && !defined(XBYAK32)
	#if defined(XBYAK64_GCC) || defined(XBYAK64_WIN)
		#define XBYAK64
	#else
		#define XBYAK32
	#endif
#endif

#if (__cplusplus >= 201103) || (defined(_MSC_VER) && _MSC_VER >= 1900)
	#undef XBYAK_TLS
	#define XBYAK_TLS thread_local
	#define XBYAK_VARIADIC_TEMPLATE
	#define XBYAK_NOEXCEPT noexcept
#else
	#define XBYAK_NOEXCEPT throw()
#endif

// require c++14 or later
// Visual Studio 2017 version 15.0 or later
// g++-6 or later
#if ((__cplusplus >= 201402L) && !(!defined(__clang__) && defined(__GNUC__) && (__GNUC__ <= 5))) || (defined(_MSC_VER) && _MSC_VER >= 1910)
	#define XBYAK_CONSTEXPR constexpr
#else
	#define XBYAK_CONSTEXPR
#endif

#ifdef _MSC_VER
	#pragma warning(push)
	#pragma warning(disable : 4514) /* remove inline function */
	#pragma warning(disable : 4786) /* identifier is too long */
	#pragma warning(disable : 4503) /* name is too long */
	#pragma warning(disable : 4127) /* constant expresison */
#endif

// disable -Warray-bounds because it may be a bug of gcc. https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104603
#if defined(__GNUC__) && !defined(__clang__)
	#define XBYAK_DISABLE_WARNING_ARRAY_BOUNDS
	#pragma GCC diagnostic push
	#pragma GCC diagnostic ignored "-Warray-bounds"
#endif

namespace Xbyak {

enum {
	DEFAULT_MAX_CODE_SIZE = 4096,
	VERSION = 0x7040 /* 0xABCD = A.BC(.D) */
};

#ifndef MIE_INTEGER_TYPE_DEFINED
#define MIE_INTEGER_TYPE_DEFINED
// for backward compatibility
typedef uint64_t uint64;
typedef int64_t sint64;
typedef uint32_t uint32;
typedef uint16_t uint16;
typedef uint8_t uint8;
#endif

#ifndef MIE_ALIGN
	#ifdef _MSC_VER
		#define MIE_ALIGN(x) __declspec(align(x))
	#else
		#define MIE_ALIGN(x) __attribute__((aligned(x)))
	#endif
#endif
#ifndef MIE_PACK // for shufps
	#define MIE_PACK(x, y, z, w) ((x) * 64 + (y) * 16 + (z) * 4 + (w))
#endif

enum {
	ERR_NONE = 0,
	ERR_BAD_ADDRESSING,
	ERR_CODE_IS_TOO_BIG,
	ERR_BAD_SCALE,
	ERR_ESP_CANT_BE_INDEX,
	ERR_BAD_COMBINATION,
	ERR_BAD_SIZE_OF_REGISTER,
	ERR_IMM_IS_TOO_BIG,
	ERR_BAD_ALIGN,
	ERR_LABEL_IS_REDEFINED,
	ERR_LABEL_IS_TOO_FAR,
	ERR_LABEL_IS_NOT_FOUND,
	ERR_CODE_ISNOT_COPYABLE,
	ERR_BAD_PARAMETER,
	ERR_CANT_PROTECT,
	ERR_CANT_USE_64BIT_DISP,
	ERR_OFFSET_IS_TOO_BIG,
	ERR_MEM_SIZE_IS_NOT_SPECIFIED,
	ERR_BAD_MEM_SIZE,
	ERR_BAD_ST_COMBINATION,
	ERR_OVER_LOCAL_LABEL, // not used
	ERR_UNDER_LOCAL_LABEL,
	ERR_CANT_ALLOC,
	ERR_ONLY_T_NEAR_IS_SUPPORTED_IN_AUTO_GROW,
	ERR_BAD_PROTECT_MODE,
	ERR_BAD_PNUM,
	ERR_BAD_TNUM,
	ERR_BAD_VSIB_ADDRESSING,
	ERR_CANT_CONVERT,
	ERR_LABEL_ISNOT_SET_BY_L,
	ERR_LABEL_IS_ALREADY_SET_BY_L,
	ERR_BAD_LABEL_STR,
	ERR_MUNMAP,
	ERR_OPMASK_IS_ALREADY_SET,
	ERR_ROUNDING_IS_ALREADY_SET,
	ERR_K0_IS_INVALID,
	ERR_EVEX_IS_INVALID,
	ERR_SAE_IS_INVALID,
	ERR_ER_IS_INVALID,
	ERR_INVALID_BROADCAST,
	ERR_INVALID_OPMASK_WITH_MEMORY,
	ERR_INVALID_ZERO,
	ERR_INVALID_RIP_IN_AUTO_GROW,
	ERR_INVALID_MIB_ADDRESS,
	ERR_X2APIC_IS_NOT_SUPPORTED,
	ERR_NOT_SUPPORTED,
	ERR_SAME_REGS_ARE_INVALID,
	ERR_INVALID_NF,
	ERR_INVALID_ZU,
	ERR_CANT_USE_REX2,
	ERR_INVALID_DFV,
	ERR_INVALID_REG_IDX,
	ERR_INTERNAL // Put it at last.
};

inline const char *ConvertErrorToString(int err)
{
	static const char *errTbl[] = {
		"none",
		"bad addressing",
		"code is too big",
		"bad scale",
		"esp can't be index",
		"bad combination",
		"bad size of register",
		"imm is too big",
		"bad align",
		"label is redefined",
		"label is too far",
		"label is not found",
		"code is not copyable",
		"bad parameter",
		"can't protect",
		"can't use 64bit disp(use (void*))",
		"offset is too big",
		"MEM size is not specified",
		"bad mem size",
		"bad st combination",
		"over local label",
		"under local label",
		"can't alloc",
		"T_SHORT is not supported in AutoGrow",
		"bad protect mode",
		"bad pNum",
		"bad tNum",
		"bad vsib addressing",
		"can't convert",
		"label is not set by L()",
		"label is already set by L()",
		"bad label string",
		"err munmap",
		"opmask is already set",
		"rounding is already set",
		"k0 is invalid",
		"evex is invalid",
		"sae(suppress all exceptions) is invalid",
		"er(embedded rounding) is invalid",
		"invalid broadcast",
		"invalid opmask with memory",
		"invalid zero",
		"invalid rip in AutoGrow",
		"invalid mib address",
		"x2APIC is not supported",
		"not supported",
		"same regs are invalid",
		"invalid NF",
		"invalid ZU",
		"can't use rex2",
		"invalid dfv",
		"invalid reg index",
		"internal error"
	};
	assert(ERR_INTERNAL + 1 == sizeof(errTbl) / sizeof(*errTbl));
	return err <= ERR_INTERNAL ? errTbl[err] : "unknown err";
}

#ifdef XBYAK_NO_EXCEPTION
namespace local {

inline int& GetErrorRef() {
	static XBYAK_TLS int err = 0;
	return err;
}

inline void SetError(int err) {
	if (local::GetErrorRef()) return; // keep the first err code
	local::GetErrorRef() = err;
}

} // local

inline void ClearError() {
	local::GetErrorRef() = 0;
}
inline int GetError() { return Xbyak::local::GetErrorRef(); }

#define XBYAK_THROW(err) { Xbyak::local::SetError(err); return; }
#define XBYAK_THROW_RET(err, r) { Xbyak::local::SetError(err); return r; }

#else
class Error : public std::exception {
	int err_;
public:
	explicit Error(int err) : err_(err)
	{
		if (err_ < 0 || err_ > ERR_INTERNAL) {
			err_ = ERR_INTERNAL;
		}
	}
	operator int() const { return err_; }
	const char *what() const XBYAK_NOEXCEPT
	{
		return ConvertErrorToString(err_);
	}
};

// dummy functions
inline void ClearError() { }
inline int GetError() { return 0; }

inline const char *ConvertErrorToString(const Error& err)
{
	return err.what();
}

#define XBYAK_THROW(err) { throw Error(err); }
#define XBYAK_THROW_RET(err, r) { throw Error(err); }

#endif

inline void *AlignedMalloc(size_t size, size_t alignment)
{
#ifdef __MINGW32__
	return __mingw_aligned_malloc(size, alignment);
#elif defined(_WIN32)
	return _aligned_malloc(size, alignment);
#else
	void *p;
	int ret = posix_memalign(&p, alignment, size);
	return (ret == 0) ? p : 0;
#endif
}

inline void AlignedFree(void *p)
{
#ifdef __MINGW32__
	__mingw_aligned_free(p);
#elif defined(_MSC_VER)
	_aligned_free(p);
#else
	free(p);
#endif
}

template<class To, class From>
inline const To CastTo(From p) XBYAK_NOEXCEPT
{
	return (const To)(size_t)(p);
}
namespace inner {

#ifdef _WIN32
struct SystemInfo {
	SYSTEM_INFO info;
	SystemInfo()
	{
		GetSystemInfo(&info);
	}
};
#endif
//static const size_t ALIGN_PAGE_SIZE = 4096;
inline size_t getPageSize()
{
#ifdef _WIN32
	static const SystemInfo si;
	return si.info.dwPageSize;
#else
#ifdef __GNUC__
	static const long pageSize = sysconf(_SC_PAGESIZE);
	if (pageSize > 0) {
		return (size_t)pageSize;
	}
#endif
	return 4096;
#endif
}

inline bool IsInDisp8(uint32_t x) { return 0xFFFFFF80 <= x || x <= 0x7F; }
inline bool IsInInt32(uint64_t x) { return ~uint64_t(0x7fffffffu) <= x || x <= 0x7FFFFFFFU; }

inline uint32_t VerifyInInt32(uint64_t x)
{
#if defined(XBYAK64) && !defined(__ILP32__)
	if (!IsInInt32(x)) XBYAK_THROW_RET(ERR_OFFSET_IS_TOO_BIG, 0)
#endif
	return static_cast<uint32_t>(x);
}

enum LabelMode {
	LasIs, // as is
	Labs, // absolute
	LaddTop // (addr + top) for mov(reg, label) with AutoGrow
};

} // inner

/*
	custom allocator
*/
struct Allocator {
	explicit Allocator(const std::string& = "") {} // same interface with MmapAllocator
	virtual uint8_t *alloc(size_t size) { return reinterpret_cast<uint8_t*>(AlignedMalloc(size, inner::getPageSize())); }
	virtual void free(uint8_t *p) { AlignedFree(p); }
	virtual ~Allocator() {}
	/* override to return false if you call protect() manually */
	virtual bool useProtect() const { return true; }
};

#ifdef XBYAK_USE_MMAP_ALLOCATOR
#ifdef XBYAK_USE_MAP_JIT
namespace util {

inline int getMacOsVersionPure()
{
	char buf[64];
	size_t size = sizeof(buf);
	int err = sysctlbyname("kern.osrelease", buf, &size, NULL, 0);
	if (err != 0) return 0;
	char *endp;
	int major = strtol(buf, &endp, 10);
	if (*endp != '.') return 0;
	return major;
}

inline int getMacOsVersion()
{
	static const int version = getMacOsVersionPure();
	return version;
}

} // util
#endif
class MmapAllocator : public Allocator {
	struct Allocation {
		size_t size;
#if defined(XBYAK_USE_MEMFD)
		// fd_ is only used with XBYAK_USE_MEMFD. We keep the file open
		// during the lifetime of each allocation in order to support
		// checkpoint/restore by unprivileged users.
		int fd;
#endif
	};
	const std::string name_; // only used with XBYAK_USE_MEMFD
	typedef XBYAK_STD_UNORDERED_MAP<uintptr_t, Allocation> AllocationList;
	AllocationList allocList_;
public:
	explicit MmapAllocator(const std::string& name = "xbyak") : name_(name) {}
	uint8_t *alloc(size_t size)
	{
		const size_t alignedSizeM1 = inner::getPageSize() - 1;
		size = (size + alignedSizeM1) & ~alignedSizeM1;
#if defined(MAP_ANONYMOUS)
		int mode = MAP_PRIVATE | MAP_ANONYMOUS;
#elif defined(MAP_ANON)
		int mode = MAP_PRIVATE | MAP_ANON;
#else
		#error "not supported"
#endif
#if defined(XBYAK_USE_MAP_JIT)
		const int mojaveVersion = 18;
		if (util::getMacOsVersion() >= mojaveVersion) mode |= MAP_JIT;
#endif
		int fd = -1;
#if defined(XBYAK_USE_MEMFD)
		fd = memfd_create(name_.c_str(), MFD_CLOEXEC);
		if (fd != -1) {
			mode = MAP_SHARED;
			if (ftruncate(fd, size) != 0) {
				close(fd);
				XBYAK_THROW_RET(ERR_CANT_ALLOC, 0)
			}
		}
#endif
		void *p = mmap(NULL, size, PROT_READ | PROT_WRITE, mode, fd, 0);
		if (p == MAP_FAILED) {
			if (fd != -1) close(fd);
			XBYAK_THROW_RET(ERR_CANT_ALLOC, 0)
		}
		assert(p);
		Allocation &alloc = allocList_[(uintptr_t)p];
		alloc.size = size;
#if defined(XBYAK_USE_MEMFD)
		alloc.fd = fd;
#endif
		return (uint8_t*)p;
	}
	void free(uint8_t *p)
	{
		if (p == 0) return;
		AllocationList::iterator i = allocList_.find((uintptr_t)p);
		if (i == allocList_.end()) XBYAK_THROW(ERR_BAD_PARAMETER)
		if (munmap((void*)i->first, i->second.size) < 0) XBYAK_THROW(ERR_MUNMAP)
#if defined(XBYAK_USE_MEMFD)
		if (i->second.fd != -1) close(i->second.fd);
#endif
		allocList_.erase(i);
	}
};
#else
typedef Allocator MmapAllocator;
#endif

class Address;
class Reg;

struct ApxFlagNF {};
struct ApxFlagZU {};

// dfv (default flags value) is or operation of these flags
static const int T_of = 8;
static const int T_sf = 4;
static const int T_zf = 2;
static const int T_cf = 1;

class Operand {
	static const uint8_t EXT8BIT = 0x20;
	unsigned int idx_:6; // 0..31 + EXT8BIT = 1 if spl/bpl/sil/dil
	unsigned int kind_:10;
	unsigned int bit_:14;
protected:
	unsigned int zero_:1;
	unsigned int mask_:3;
	unsigned int rounding_:3;
	unsigned int NF_:1;
	unsigned int ZU_:1; // ND=ZU
	void setIdx(int idx) { idx_ = idx; }
public:
	enum Kind {
		NONE = 0,
		MEM = 1 << 0,
		REG = 1 << 1,
		MMX = 1 << 2,
		FPU = 1 << 3,
		XMM = 1 << 4,
		YMM = 1 << 5,
		ZMM = 1 << 6,
		OPMASK = 1 << 7,
		BNDREG = 1 << 8,
		TMM = 1 << 9
	};
	enum Code {
#ifdef XBYAK64
		RAX = 0, RCX, RDX, RBX, RSP, RBP, RSI, RDI, R8, R9, R10, R11, R12, R13, R14, R15,
		R16, R17, R18, R19, R20, R21, R22, R23, R24, R25, R26, R27, R28, R29, R30, R31,
		R8D = 8, R9D, R10D, R11D, R12D, R13D, R14D, R15D,
		R16D, R17D, R18D, R19D, R20D, R21D, R22D, R23D, R24D, R25D, R26D, R27D, R28D, R29D, R30D, R31D,
		R8W = 8, R9W, R10W, R11W, R12W, R13W, R14W, R15W,
		R16W, R17W, R18W, R19W, R20W, R21W, R22W, R23W, R24W, R25W, R26W, R27W, R28W, R29W, R30W, R31W,
		R8B = 8, R9B, R10B, R11B, R12B, R13B, R14B, R15B,
		R16B, R17B, R18B, R19B, R20B, R21B, R22B, R23B, R24B, R25B, R26B, R27B, R28B, R29B, R30B, R31B,
		SPL = 4, BPL, SIL, DIL,
#endif
		EAX = 0, ECX, EDX, EBX, ESP, EBP, ESI, EDI,
		AX = 0, CX, DX, BX, SP, BP, SI, DI,
		AL = 0, CL, DL, BL, AH, CH, DH, BH
	};
	XBYAK_CONSTEXPR Operand() : idx_(0), kind_(0), bit_(0), zero_(0), mask_(0), rounding_(0), NF_(0), ZU_(0) { }
	XBYAK_CONSTEXPR Operand(int idx, Kind kind, int bit, bool ext8bit = 0)
		: idx_(static_cast<uint8_t>(idx | (ext8bit ? EXT8BIT : 0)))
		, kind_(kind)
		, bit_(bit)
		, zero_(0), mask_(0), rounding_(0), NF_(0), ZU_(0)
	{
		assert((bit_ & (bit_ - 1)) == 0); // bit must be power of two
	}
	XBYAK_CONSTEXPR Kind getKind() const { return static_cast<Kind>(kind_); }
	XBYAK_CONSTEXPR int getIdx() const { return idx_ & (EXT8BIT - 1); }
	XBYAK_CONSTEXPR bool hasIdxBit(int bit) const { return idx_ & (1<<bit); }
	XBYAK_CONSTEXPR bool isNone() const { return kind_ == 0; }
	XBYAK_CONSTEXPR bool isMMX() const { return is(MMX); }
	XBYAK_CONSTEXPR bool isXMM() const { return is(XMM); }
	XBYAK_CONSTEXPR bool isYMM() const { return is(YMM); }
	XBYAK_CONSTEXPR bool isZMM() const { return is(ZMM); }
	XBYAK_CONSTEXPR bool isSIMD() const { return is(XMM|YMM|ZMM); }
	XBYAK_CONSTEXPR bool isTMM() const { return is(TMM); }
	XBYAK_CONSTEXPR bool isXMEM() const { return is(XMM | MEM); }
	XBYAK_CONSTEXPR bool isYMEM() const { return is(YMM | MEM); }
	XBYAK_CONSTEXPR bool isZMEM() const { return is(ZMM | MEM); }
	XBYAK_CONSTEXPR bool isOPMASK() const { return is(OPMASK); }
	XBYAK_CONSTEXPR bool isBNDREG() const { return is(BNDREG); }
	XBYAK_CONSTEXPR bool isREG(int bit = 0) const { return is(REG, bit); }
	XBYAK_CONSTEXPR bool isMEM(int bit = 0) const { return is(MEM, bit); }
	XBYAK_CONSTEXPR bool isFPU() const { return is(FPU); }
	XBYAK_CONSTEXPR bool isExt8bit() const { return (idx_ & EXT8BIT) != 0; }
	XBYAK_CONSTEXPR bool isExtIdx() const { return (getIdx() & 8) != 0; }
	XBYAK_CONSTEXPR bool isExtIdx2() const { return (getIdx() & 16) != 0; }
	XBYAK_CONSTEXPR bool hasEvex() const { return isZMM() || isExtIdx2() || getOpmaskIdx() || getRounding(); }
	XBYAK_CONSTEXPR bool hasRex() const { return isExt8bit() || isREG(64) || isExtIdx(); }
	XBYAK_CONSTEXPR bool hasRex2() const;
	XBYAK_CONSTEXPR bool hasRex2NF() const { return hasRex2() || NF_; }
	XBYAK_CONSTEXPR bool hasRex2NFZU() const { return hasRex2() || NF_ || ZU_; }
	XBYAK_CONSTEXPR bool hasZero() const { return zero_; }
	XBYAK_CONSTEXPR int getOpmaskIdx() const { return mask_; }
	XBYAK_CONSTEXPR int getRounding() const { return rounding_; }
	void setKind(Kind kind)
	{
		if ((kind & (XMM|YMM|ZMM|TMM)) == 0) return;
		kind_ = kind;
		bit_ = kind == XMM ? 128 : kind == YMM ? 256 : kind == ZMM ? 512 : 8192;
	}
	// err if MMX/FPU/OPMASK/BNDREG
	void setBit(int bit);
	void setOpmaskIdx(int idx, bool /*ignore_idx0*/ = true)
	{
		if (mask_) XBYAK_THROW(ERR_OPMASK_IS_ALREADY_SET)
		mask_ = idx;
	}
	void setRounding(int idx)
	{
		if (rounding_) XBYAK_THROW(ERR_ROUNDING_IS_ALREADY_SET)
		rounding_ = idx;
	}
	void setZero() { zero_ = true; }
	void setNF() { NF_ = true; }
	int getNF() const { return NF_; }
	void setZU() { ZU_ = true; }
	int getZU() const { return ZU_; }
	// ah, ch, dh, bh?
	bool isHigh8bit() const
	{
		if (!isBit(8)) return false;
		if (isExt8bit()) return false;
		const int idx = getIdx();
		return AH <= idx && idx <= BH;
	}
	// any bit is accetable if bit == 0
	XBYAK_CONSTEXPR bool is(int kind, uint32_t bit = 0) const
	{
		return (kind == 0 || (kind_ & kind)) && (bit == 0 || (bit_ & bit)); // cf. you can set (8|16)
	}
	XBYAK_CONSTEXPR bool isBit(uint32_t bit) const { return (bit_ & bit) != 0; }
	XBYAK_CONSTEXPR uint32_t getBit() const { return bit_; }
	const char *toString() const
	{
		const int idx = getIdx();
		if (kind_ == REG) {
			if (isExt8bit()) {
				static const char *tbl[4] = { "spl", "bpl", "sil", "dil" };
				return tbl[idx - 4];
			}
			static const char *tbl[4][32] = {
				{ "al", "cl", "dl", "bl", "ah", "ch", "dh", "bh", "r8b", "r9b", "r10b",  "r11b", "r12b", "r13b", "r14b", "r15b",
				 "r16b", "r17b", "r18b", "r19b", "r20b", "r21b", "r22b", "r23b", "r24b", "r25b", "r26b", "r27b", "r28b", "r29b", "r30b", "r31b",
				},
				{ "ax", "cx", "dx", "bx", "sp", "bp", "si", "di", "r8w", "r9w", "r10w",  "r11w", "r12w", "r13w", "r14w", "r15w",
				 "r16w", "r17w", "r18w", "r19w", "r20w", "r21w", "r22w", "r23w", "r24w", "r25w", "r26w", "r27w", "r28w", "r29w", "r30w", "r31w",
				},
				{ "eax", "ecx", "edx", "ebx", "esp", "ebp", "esi", "edi", "r8d", "r9d", "r10d",  "r11d", "r12d", "r13d", "r14d", "r15d",
				 "r16d", "r17d", "r18d", "r19d", "r20d", "r21d", "r22d", "r23d", "r24d", "r25d", "r26d", "r27d", "r28d", "r29d", "r30d", "r31d",
				},
				{ "rax", "rcx", "rdx", "rbx", "rsp", "rbp", "rsi", "rdi", "r8", "r9", "r10",  "r11", "r12", "r13", "r14", "r15",
				 "r16", "r17", "r18", "r19", "r20", "r21", "r22", "r23", "r24", "r25", "r26", "r27", "r28", "r29", "r30", "r31",
				},
			};
			return tbl[bit_ == 8 ? 0 : bit_ == 16 ? 1 : bit_ == 32 ? 2 : 3][idx];
		} else if (isOPMASK()) {
			static const char *tbl[8] = { "k0", "k1", "k2", "k3", "k4", "k5", "k6", "k7" };
			return tbl[idx];
		} else if (isTMM()) {
			static const char *tbl[8] = {
				"tmm0", "tmm1", "tmm2", "tmm3", "tmm4", "tmm5", "tmm6", "tmm7"
			};
			return tbl[idx];
		} else if (isZMM()) {
			static const char *tbl[32] = {
				"zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5", "zmm6", "zmm7", "zmm8", "zmm9", "zmm10", "zmm11", "zmm12", "zmm13", "zmm14", "zmm15",
				"zmm16", "zmm17", "zmm18", "zmm19", "zmm20", "zmm21", "zmm22", "zmm23", "zmm24", "zmm25", "zmm26", "zmm27", "zmm28", "zmm29", "zmm30", "zmm31"
			};
			return tbl[idx];
		} else if (isYMM()) {
			static const char *tbl[32] = {
				"ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", "ymm7", "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13", "ymm14", "ymm15",
				"ymm16", "ymm17", "ymm18", "ymm19", "ymm20", "ymm21", "ymm22", "ymm23", "ymm24", "ymm25", "ymm26", "ymm27", "ymm28", "ymm29", "ymm30", "ymm31"
			};
			return tbl[idx];
		} else if (isXMM()) {
			static const char *tbl[32] = {
				"xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm14", "xmm15",
				"xmm16", "xmm17", "xmm18", "xmm19", "xmm20", "xmm21", "xmm22", "xmm23", "xmm24", "xmm25", "xmm26", "xmm27", "xmm28", "xmm29", "xmm30", "xmm31"
			};
			return tbl[idx];
		} else if (isMMX()) {
			static const char *tbl[8] = { "mm0", "mm1", "mm2", "mm3", "mm4", "mm5", "mm6", "mm7" };
			return tbl[idx];
		} else if (isFPU()) {
			static const char *tbl[8] = { "st0", "st1", "st2", "st3", "st4", "st5", "st6", "st7" };
			return tbl[idx];
		} else if (isBNDREG()) {
			static const char *tbl[4] = { "bnd0", "bnd1", "bnd2", "bnd3" };
			return tbl[idx];
		}
		XBYAK_THROW_RET(ERR_INTERNAL, 0);
	}
	bool isEqualIfNotInherited(const Operand& rhs) const { return idx_ == rhs.idx_ && kind_ == rhs.kind_ && bit_ == rhs.bit_ && zero_ == rhs.zero_ && mask_ == rhs.mask_ && rounding_ == rhs.rounding_; }
	bool operator==(const Operand& rhs) const;
	bool operator!=(const Operand& rhs) const { return !operator==(rhs); }
	const Address& getAddress() const;
	const Reg& getReg() const;
};

inline void Operand::setBit(int bit)
{
	if (bit != 8 && bit != 16 && bit != 32 && bit != 64 && bit != 128 && bit != 256 && bit != 512 && bit != 8192) goto ERR;
	if (isBit(bit)) return;
	if (is(MEM | OPMASK)) {
		bit_ = bit;
		return;
	}
	if (is(REG | XMM | YMM | ZMM | TMM)) {
		int idx = getIdx();
		// err if converting ah, bh, ch, dh
		if (isREG(8) && (4 <= idx && idx < 8) && !isExt8bit()) goto ERR;
		Kind kind = REG;
		switch (bit) {
		case 8:
#ifdef XBYAK32
			if (idx >= 4) goto ERR;
#else
			if (idx >= 32) goto ERR;
			if (4 <= idx && idx < 8) idx |= EXT8BIT;
#endif
			break;
		case 16:
		case 32:
		case 64:
#ifdef XBYAK32
			if (idx >= 16) goto ERR;
#else
			if (idx >= 32) goto ERR;
#endif
			break;
		case 128: kind = XMM; break;
		case 256: kind = YMM; break;
		case 512: kind = ZMM; break;
		case 8192: kind = TMM; break;
		}
		idx_ = idx;
		kind_ = kind;
		bit_ = bit;
		if (bit >= 128) return; // keep mask_ and rounding_
		mask_ = 0;
		rounding_ = 0;
		return;
	}
ERR:
	XBYAK_THROW(ERR_CANT_CONVERT)
}

class Label;

struct Reg8;
struct Reg16;
struct Reg32;
#ifdef XBYAK64
struct Reg64;
#endif
class Reg : public Operand {
public:
	XBYAK_CONSTEXPR Reg() { }
	XBYAK_CONSTEXPR Reg(int idx, Kind kind, int bit = 0, bool ext8bit = false) : Operand(idx, kind, bit, ext8bit) { }
	// convert to Reg8/Reg16/Reg32/Reg64/XMM/YMM/ZMM
	Reg changeBit(int bit) const { Reg r(*this); r.setBit(bit); return r; }
	Reg8 cvt8() const;
	Reg16 cvt16() const;
	Reg32 cvt32() const;
#ifdef XBYAK64
	Reg64 cvt64() const;
#endif
	Reg operator|(const ApxFlagNF&) const { Reg r(*this); r.setNF(); return r; }
	Reg operator|(const ApxFlagZU&) const { Reg r(*this); r.setZU(); return r; }
};

inline const Reg& Operand::getReg() const
{
	assert(!isMEM());
	return static_cast<const Reg&>(*this);
}

struct Reg8 : public Reg {
	explicit XBYAK_CONSTEXPR Reg8(int idx = 0, bool ext8bit = false) : Reg(idx, Operand::REG, 8, ext8bit) { }
};

struct Reg16 : public Reg {
	explicit XBYAK_CONSTEXPR Reg16(int idx = 0) : Reg(idx, Operand::REG, 16) { }
};

struct Mmx : public Reg {
	explicit XBYAK_CONSTEXPR Mmx(int idx = 0, Kind kind = Operand::MMX, int bit = 64) : Reg(idx, kind, bit) { }
};

struct EvexModifierRounding {
	enum {
		T_RN_SAE = 1,
		T_RD_SAE = 2,
		T_RU_SAE = 3,
		T_RZ_SAE = 4,
		T_SAE = 5
	};
	explicit XBYAK_CONSTEXPR EvexModifierRounding(int rounding) : rounding(rounding) {}
	int rounding;
};
struct EvexModifierZero{ XBYAK_CONSTEXPR EvexModifierZero() {}};

struct Xmm : public Mmx {
	explicit XBYAK_CONSTEXPR Xmm(int idx = 0, Kind kind = Operand::XMM, int bit = 128) : Mmx(idx, kind, bit) { }
	XBYAK_CONSTEXPR Xmm(Kind kind, int idx) : Mmx(idx, kind, kind == XMM ? 128 : kind == YMM ? 256 : 512) { }
	Xmm operator|(const EvexModifierRounding& emr) const { Xmm r(*this); r.setRounding(emr.rounding); return r; }
	Xmm copyAndSetIdx(int idx) const { Xmm ret(*this); ret.setIdx(idx); return ret; }
	Xmm copyAndSetKind(Operand::Kind kind) const { Xmm ret(*this); ret.setKind(kind); return ret; }
};

struct Ymm : public Xmm {
	explicit XBYAK_CONSTEXPR Ymm(int idx = 0, Kind kind = Operand::YMM, int bit = 256) : Xmm(idx, kind, bit) { }
	Ymm operator|(const EvexModifierRounding& emr) const { Ymm r(*this); r.setRounding(emr.rounding); return r; }
};

struct Zmm : public Ymm {
	explicit XBYAK_CONSTEXPR Zmm(int idx = 0) : Ymm(idx, Operand::ZMM, 512) { }
	Zmm operator|(const EvexModifierRounding& emr) const { Zmm r(*this); r.setRounding(emr.rounding); return r; }
};

#ifdef XBYAK64
struct Tmm : public Reg {
	explicit XBYAK_CONSTEXPR Tmm(int idx = 0, Kind kind = Operand::TMM, int bit = 8192) : Reg(idx, kind, bit) { }
};
#endif

struct Opmask : public Reg {
	explicit XBYAK_CONSTEXPR Opmask(int idx = 0) : Reg(idx, Operand::OPMASK, 64) {}
};

struct BoundsReg : public Reg {
	explicit XBYAK_CONSTEXPR BoundsReg(int idx = 0) : Reg(idx, Operand::BNDREG, 128) {}
};

template<class T>T operator|(const T& x, const Opmask& k) { T r(x); r.setOpmaskIdx(k.getIdx()); return r; }
template<class T>T operator|(const T& x, const EvexModifierZero&) { T r(x); r.setZero(); return r; }
template<class T>T operator|(const T& x, const EvexModifierRounding& emr) { T r(x); r.setRounding(emr.rounding); return r; }

struct Fpu : public Reg {
	explicit XBYAK_CONSTEXPR Fpu(int idx = 0) : Reg(idx, Operand::FPU, 32) { }
};

struct Reg32e : public Reg {
	explicit XBYAK_CONSTEXPR Reg32e(int idx, int bit) : Reg(idx, Operand::REG, bit) {}
	Reg32e operator|(const ApxFlagNF&) const { Reg32e r(*this); r.setNF(); return r; }
	Reg32e operator|(const ApxFlagZU&) const { Reg32e r(*this); r.setZU(); return r; }
};
struct Reg32 : public Reg32e {
	explicit XBYAK_CONSTEXPR Reg32(int idx = 0) : Reg32e(idx, 32) {}
};
#ifdef XBYAK64
struct Reg64 : public Reg32e {
	explicit XBYAK_CONSTEXPR Reg64(int idx = 0) : Reg32e(idx, 64) {}
};
struct RegRip {
	int64_t disp_;
	const Label* label_;
	bool isAddr_;
	explicit XBYAK_CONSTEXPR RegRip(int64_t disp = 0, const Label* label = 0, bool isAddr = false) : disp_(disp), label_(label), isAddr_(isAddr) {}
	friend const RegRip operator+(const RegRip& r, int disp) {
		return RegRip(r.disp_ + disp, r.label_, r.isAddr_);
	}
	friend const RegRip operator-(const RegRip& r, int disp) {
		return RegRip(r.disp_ - disp, r.label_, r.isAddr_);
	}
	friend const RegRip operator+(const RegRip& r, int64_t disp) {
		return RegRip(r.disp_ + disp, r.label_, r.isAddr_);
	}
	friend const RegRip operator-(const RegRip& r, int64_t disp) {
		return RegRip(r.disp_ - disp, r.label_, r.isAddr_);
	}
	friend const RegRip operator+(const RegRip& r, const Label& label) {
		if (r.label_ || r.isAddr_) XBYAK_THROW_RET(ERR_BAD_ADDRESSING, RegRip());
		return RegRip(r.disp_, &label);
	}
	friend const RegRip operator+(const RegRip& r, const void *addr) {
		if (r.label_ || r.isAddr_) XBYAK_THROW_RET(ERR_BAD_ADDRESSING, RegRip());
		return RegRip(r.disp_ + (int64_t)addr, 0, true);
	}
};
#endif

inline Reg8 Reg::cvt8() const
{
	Reg r = changeBit(8); return Reg8(r.getIdx(), r.isExt8bit());
}

inline Reg16 Reg::cvt16() const
{
	return Reg16(changeBit(16).getIdx());
}

inline Reg32 Reg::cvt32() const
{
	return Reg32(changeBit(32).getIdx());
}

#ifdef XBYAK64
inline Reg64 Reg::cvt64() const
{
	return Reg64(changeBit(64).getIdx());
}
#endif

#ifndef XBYAK_DISABLE_SEGMENT
// not derived from Reg
class Segment {
	int idx_;
public:
	enum {
		es, cs, ss, ds, fs, gs
	};
	explicit XBYAK_CONSTEXPR Segment(int idx) : idx_(idx) { assert(0 <= idx_ && idx_ < 6); }
	int getIdx() const { return idx_; }
	const char *toString() const
	{
		static const char tbl[][3] = {
			"es", "cs", "ss", "ds", "fs", "gs"
		};
		return tbl[idx_];
	}
};
#endif

class RegExp {
public:
#ifdef XBYAK64
	enum { i32e = 32 | 64 };
#else
	enum { i32e = 32 };
#endif
	XBYAK_CONSTEXPR RegExp(size_t disp = 0) : scale_(0), disp_(disp) { }
	XBYAK_CONSTEXPR RegExp(const Reg& r, int scale = 1)
		: scale_(scale)
		, disp_(0)
	{
		if (!r.isREG(i32e) && !r.is(Reg::XMM|Reg::YMM|Reg::ZMM|Reg::TMM)) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		if (scale == 0) return;
		if (scale != 1 && scale != 2 && scale != 4 && scale != 8) XBYAK_THROW(ERR_BAD_SCALE)
		if (r.getBit() >= 128 || scale != 1) { // xmm/ymm is always index
			index_ = r;
		} else {
			base_ = r;
		}
	}
	bool isVsib(int bit = 128 | 256 | 512) const { return index_.isBit(bit); }
	RegExp optimize() const
	{
		RegExp exp = *this;
		// [reg * 2] => [reg + reg]
		if (index_.isBit(i32e) && !base_.getBit() && scale_ == 2) {
			exp.base_ = index_;
			exp.scale_ = 1;
		}
		return exp;
	}
	bool operator==(const RegExp& rhs) const
	{
		return base_ == rhs.base_ && index_ == rhs.index_ && disp_ == rhs.disp_ && scale_ == rhs.scale_;
	}
	const Reg& getBase() const { return base_; }
	const Reg& getIndex() const { return index_; }
	int getScale() const { return scale_; }
	size_t getDisp() const { return disp_; }
	XBYAK_CONSTEXPR void verify() const
	{
		if (base_.getBit() >= 128) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		if (index_.getBit() && index_.getBit() <= 64) {
			if (index_.getIdx() == Operand::ESP) XBYAK_THROW(ERR_ESP_CANT_BE_INDEX)
			if (base_.getBit() && base_.getBit() != index_.getBit()) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		}
	}
	friend RegExp operator+(const RegExp& a, const RegExp& b);
	friend RegExp operator-(const RegExp& e, size_t disp);
private:
	/*
		[base_ + index_ * scale_ + disp_]
		base : Reg32e, index : Reg32e(w/o esp), Xmm, Ymm
	*/
	Reg base_;
	Reg index_;
	int scale_;
	size_t disp_;
};

inline RegExp operator+(const RegExp& a, const RegExp& b)
{
	if (a.index_.getBit() && b.index_.getBit()) XBYAK_THROW_RET(ERR_BAD_ADDRESSING, RegExp())
	RegExp ret = a;
	if (!ret.index_.getBit()) { ret.index_ = b.index_; ret.scale_ = b.scale_; }
	if (b.base_.getBit()) {
		if (ret.base_.getBit()) {
			if (ret.index_.getBit()) XBYAK_THROW_RET(ERR_BAD_ADDRESSING, RegExp())
			// base + base => base + index * 1
			ret.index_ = b.base_;
			// [reg + esp] => [esp + reg]
			if (ret.index_.getIdx() == Operand::ESP) std::swap(ret.base_, ret.index_);
			ret.scale_ = 1;
		} else {
			ret.base_ = b.base_;
		}
	}
	ret.disp_ += b.disp_;
	return ret;
}
inline RegExp operator*(const Reg& r, int scale)
{
	return RegExp(r, scale);
}
inline RegExp operator*(int scale, const Reg& r)
{
	return r * scale;
}
inline RegExp operator-(const RegExp& e, size_t disp)
{
	RegExp ret = e;
	ret.disp_ -= disp;
	return ret;
}

// 2nd parameter for constructor of CodeArray(maxSize, userPtr, alloc)
void *const AutoGrow = (void*)1; //-V566
void *const DontSetProtectRWE = (void*)2; //-V566

class CodeArray {
	enum Type {
		USER_BUF = 1, // use userPtr(non alignment, non protect)
		ALLOC_BUF, // use new(alignment, protect)
		AUTO_GROW // automatically move and grow memory if necessary
	};
	CodeArray(const CodeArray& rhs);
	void operator=(const CodeArray&);
	bool isAllocType() const { return type_ == ALLOC_BUF || type_ == AUTO_GROW; }
	struct AddrInfo {
		size_t codeOffset; // position to write
		size_t jmpAddr; // value to write
		int jmpSize; // size of jmpAddr
		inner::LabelMode mode;
		AddrInfo(size_t _codeOffset, size_t _jmpAddr, int _jmpSize, inner::LabelMode _mode)
			: codeOffset(_codeOffset), jmpAddr(_jmpAddr), jmpSize(_jmpSize), mode(_mode) {}
		uint64_t getVal(const uint8_t *top) const
		{
			uint64_t disp = (mode == inner::LaddTop) ? jmpAddr + size_t(top) : (mode == inner::LasIs) ? jmpAddr : jmpAddr - size_t(top);
			if (jmpSize == 4) disp = inner::VerifyInInt32(disp);
			return disp;
		}
	};
	typedef std::list<AddrInfo> AddrInfoList;
	AddrInfoList addrInfoList_;
	const Type type_;
#ifdef XBYAK_USE_MMAP_ALLOCATOR
	MmapAllocator defaultAllocator_;
#else
	Allocator defaultAllocator_;
#endif
	Allocator *alloc_;
protected:
	size_t maxSize_;
	uint8_t *top_;
	size_t size_;
	bool isCalledCalcJmpAddress_;

	bool useProtect() const { return alloc_->useProtect(); }
	/*
		allocate new memory and copy old data to the new area
	*/
	void growMemory()
	{
		const size_t newSize = (std::max<size_t>)(DEFAULT_MAX_CODE_SIZE, maxSize_ * 2);
		uint8_t *newTop = alloc_->alloc(newSize);
		if (newTop == 0) XBYAK_THROW(ERR_CANT_ALLOC)
		for (size_t i = 0; i < size_; i++) newTop[i] = top_[i];
		alloc_->free(top_);
		top_ = newTop;
		maxSize_ = newSize;
	}
	/*
		calc jmp address for AutoGrow mode
	*/
	void calcJmpAddress()
	{
		if (isCalledCalcJmpAddress_) return;
		for (AddrInfoList::const_iterator i = addrInfoList_.begin(), ie = addrInfoList_.end(); i != ie; ++i) {
			uint64_t disp = i->getVal(top_);
			rewrite(i->codeOffset, disp, i->jmpSize);
		}
		isCalledCalcJmpAddress_ = true;
	}
public:
	enum ProtectMode {
		PROTECT_RW = 0, // read/write
		PROTECT_RWE = 1, // read/write/exec
		PROTECT_RE = 2 // read/exec
	};
	explicit CodeArray(size_t maxSize, void *userPtr = 0, Allocator *allocator = 0)
		: type_(userPtr == AutoGrow ? AUTO_GROW : (userPtr == 0 || userPtr == DontSetProtectRWE) ? ALLOC_BUF : USER_BUF)
		, alloc_(allocator ? allocator : (Allocator*)&defaultAllocator_)
		, maxSize_(maxSize)
		, top_(type_ == USER_BUF ? reinterpret_cast<uint8_t*>(userPtr) : alloc_->alloc((std::max<size_t>)(maxSize, 1)))
		, size_(0)
		, isCalledCalcJmpAddress_(false)
	{
		if (maxSize_ > 0 && top_ == 0) XBYAK_THROW(ERR_CANT_ALLOC)
		if ((type_ == ALLOC_BUF && userPtr != DontSetProtectRWE && useProtect()) && !setProtectMode(PROTECT_RWE, false)) {
			alloc_->free(top_);
			XBYAK_THROW(ERR_CANT_PROTECT)
		}
	}
	virtual ~CodeArray()
	{
		if (isAllocType()) {
			if (useProtect()) setProtectModeRW(false);
			alloc_->free(top_);
		}
	}
	bool setProtectMode(ProtectMode mode, bool throwException = true)
	{
		bool isOK = protect(top_, maxSize_, mode);
		if (isOK) return true;
		if (throwException) XBYAK_THROW_RET(ERR_CANT_PROTECT, false)
		return false;
	}
	bool setProtectModeRE(bool throwException = true) { return setProtectMode(PROTECT_RE, throwException); }
	bool setProtectModeRW(bool throwException = true) { return setProtectMode(PROTECT_RW, throwException); }
	void resetSize()
	{
		size_ = 0;
		addrInfoList_.clear();
		isCalledCalcJmpAddress_ = false;
	}
	void db(int code)
	{
		if (size_ >= maxSize_) {
			if (type_ == AUTO_GROW) {
				growMemory();
			} else {
				XBYAK_THROW(ERR_CODE_IS_TOO_BIG)
			}
		}
		top_[size_++] = static_cast<uint8_t>(code);
	}
	void db(const uint8_t *code, size_t codeSize)
	{
		for (size_t i = 0; i < codeSize; i++) db(code[i]);
	}
	void db(uint64_t code, size_t codeSize)
	{
		if (codeSize > 8) XBYAK_THROW(ERR_BAD_PARAMETER)
		for (size_t i = 0; i < codeSize; i++) db(static_cast<uint8_t>(code >> (i * 8)));
	}
	void dw(uint32_t code) { db(code, 2); }
	void dd(uint32_t code) { db(code, 4); }
	void dq(uint64_t code) { db(code, 8); }
	const uint8_t *getCode() const { return top_; }
	template<class F>
	const F getCode() const { return reinterpret_cast<F>(top_); }
	const uint8_t *getCurr() const { return &top_[size_]; }
	template<class F>
	const F getCurr() const { return reinterpret_cast<F>(&top_[size_]); }
	size_t getSize() const { return size_; }
	void setSize(size_t size)
	{
		if (size > maxSize_) XBYAK_THROW(ERR_OFFSET_IS_TOO_BIG)
		size_ = size;
	}
	void dump() const
	{
		const uint8_t *p = getCode();
		size_t bufSize = getSize();
		size_t remain = bufSize;
		for (int i = 0; i < 4; i++) {
			size_t disp = 16;
			if (remain < 16) {
				disp = remain;
			}
			for (size_t j = 0; j < 16; j++) {
				if (j < disp) {
					printf("%02X", p[i * 16 + j]);
				}
			}
			putchar('\n');
			remain -= disp;
			if (remain == 0) {
				break;
			}
		}
	}
	/*
		@param offset [in] offset from top
		@param disp [in] offset from the next of jmp
		@param size [in] write size(1, 2, 4, 8)
	*/
	void rewrite(size_t offset, uint64_t disp, size_t size)
	{
		assert(offset < maxSize_);
		if (size != 1 && size != 2 && size != 4 && size != 8) XBYAK_THROW(ERR_BAD_PARAMETER)
		uint8_t *const data = top_ + offset;
		for (size_t i = 0; i < size; i++) {
			data[i] = static_cast<uint8_t>(disp >> (i * 8));
		}
	}
	void save(size_t offset, size_t val, int size, inner::LabelMode mode)
	{
		addrInfoList_.push_back(AddrInfo(offset, val, size, mode));
	}
	bool isAutoGrow() const { return type_ == AUTO_GROW; }
	bool isCalledCalcJmpAddress() const { return isCalledCalcJmpAddress_; }
	/**
		change exec permission of memory
		@param addr [in] buffer address
		@param size [in] buffer size
		@param protectMode [in] mode(RW/RWE/RE)
		@return true(success), false(failure)
	*/
	static inline bool protect(const void *addr, size_t size, int protectMode)
	{
#if defined(_WIN32)
		const DWORD c_rw = PAGE_READWRITE;
		const DWORD c_rwe = PAGE_EXECUTE_READWRITE;
		const DWORD c_re = PAGE_EXECUTE_READ;
		DWORD mode;
#else
		const int c_rw = PROT_READ | PROT_WRITE;
		const int c_rwe = PROT_READ | PROT_WRITE | PROT_EXEC;
		const int c_re = PROT_READ | PROT_EXEC;
		int mode;
#endif
		switch (protectMode) {
		case PROTECT_RW: mode = c_rw; break;
		case PROTECT_RWE: mode = c_rwe; break;
		case PROTECT_RE: mode = c_re; break;
		default:
			return false;
		}
#if defined(_WIN32)
		DWORD oldProtect;
		return VirtualProtect(const_cast<void*>(addr), size, mode, &oldProtect) != 0;
#elif defined(__GNUC__)
		size_t pageSize = sysconf(_SC_PAGESIZE);
		size_t iaddr = reinterpret_cast<size_t>(addr);
		size_t roundAddr = iaddr & ~(pageSize - static_cast<size_t>(1));
		return mprotect(reinterpret_cast<void*>(roundAddr), size + (iaddr - roundAddr), mode) == 0;
#else
		return true;
#endif
	}
	/**
		get aligned memory pointer
		@param addr [in] address
		@param alignedSize [in] power of two
		@return aligned addr by alingedSize
	*/
	static inline uint8_t *getAlignedAddress(uint8_t *addr, size_t alignedSize = 16)
	{
		return reinterpret_cast<uint8_t*>((reinterpret_cast<size_t>(addr) + alignedSize - 1) & ~(alignedSize - static_cast<size_t>(1)));
	}
};

class Address : public Operand {
public:
	enum Mode {
		M_ModRM,
		M_64bitDisp,
		M_rip,
		M_ripAddr
	};
	XBYAK_CONSTEXPR Address(uint32_t sizeBit, bool broadcast, const RegExp& e)
		: Operand(0, MEM, sizeBit), e_(e), label_(0), mode_(M_ModRM), broadcast_(broadcast), optimize_(true)
	{
		e_.verify();
	}
#ifdef XBYAK64
	explicit XBYAK_CONSTEXPR Address(size_t disp)
		: Operand(0, MEM, 64), e_(disp), label_(0), mode_(M_64bitDisp), broadcast_(false), optimize_(true) { }
	XBYAK_CONSTEXPR Address(uint32_t sizeBit, bool broadcast, const RegRip& addr)
		: Operand(0, MEM, sizeBit), e_(addr.disp_), label_(addr.label_), mode_(addr.isAddr_ ? M_ripAddr : M_rip), broadcast_(broadcast), optimize_(true) { }
#endif
	RegExp getRegExp() const
	{
		return optimize_ ? e_.optimize() : e_;
	}
	Address cloneNoOptimize() const { Address addr = *this; addr.optimize_ = false; return addr; }
	Mode getMode() const { return mode_; }
	bool is32bit() const { return e_.getBase().getBit() == 32 || e_.getIndex().getBit() == 32; }
	bool isOnlyDisp() const { return !e_.getBase().getBit() && !e_.getIndex().getBit(); } // for mov eax
	size_t getDisp() const { return e_.getDisp(); }
	bool is64bitDisp() const { return mode_ == M_64bitDisp; } // for moffset
	bool isBroadcast() const { return broadcast_; }
	bool hasRex2() const { return e_.getBase().hasRex2() || e_.getIndex().hasRex2(); }
	const Label* getLabel() const { return label_; }
	bool operator==(const Address& rhs) const
	{
		return getBit() == rhs.getBit() && e_ == rhs.e_ && label_ == rhs.label_ && mode_ == rhs.mode_ && broadcast_ == rhs.broadcast_;
	}
	bool operator!=(const Address& rhs) const { return !operator==(rhs); }
	bool isVsib() const { return e_.isVsib(); }
private:
	RegExp e_;
	const Label* label_;
	Mode mode_;
	bool broadcast_;
	bool optimize_;
};

inline const Address& Operand::getAddress() const
{
	assert(isMEM());
	return static_cast<const Address&>(*this);
}

inline bool Operand::operator==(const Operand& rhs) const
{
	if (isMEM() && rhs.isMEM()) return this->getAddress() == rhs.getAddress();
	return isEqualIfNotInherited(rhs);
}

inline XBYAK_CONSTEXPR bool Operand::hasRex2() const
{
	return (isREG() && isExtIdx2()) || (isMEM() && static_cast<const Address&>(*this).hasRex2());
}

class AddressFrame {
	void operator=(const AddressFrame&);
	AddressFrame(const AddressFrame&);
public:
	const uint32_t bit_;
	const bool broadcast_;
	explicit XBYAK_CONSTEXPR AddressFrame(uint32_t bit, bool broadcast = false) : bit_(bit), broadcast_(broadcast) { }
	Address operator[](const RegExp& e) const
	{
		return Address(bit_, broadcast_, e);
	}
	Address operator[](const void *disp) const
	{
		return Address(bit_, broadcast_, RegExp(reinterpret_cast<size_t>(disp)));
	}
#ifdef XBYAK64
	Address operator[](uint64_t disp) const { return Address(disp); }
	Address operator[](const RegRip& addr) const { return Address(bit_, broadcast_, addr); }
#endif
};

struct JmpLabel {
	size_t endOfJmp; /* offset from top to the end address of jmp */
	int jmpSize;
	inner::LabelMode mode;
	size_t disp; // disp for [rip + disp]
	explicit JmpLabel(size_t endOfJmp = 0, int jmpSize = 0, inner::LabelMode mode = inner::LasIs, size_t disp = 0)
		: endOfJmp(endOfJmp), jmpSize(jmpSize), mode(mode), disp(disp)
	{
	}
};

class LabelManager;

class Label {
	mutable LabelManager *mgr;
	mutable int id;
	friend class LabelManager;
public:
	Label() : mgr(0), id(0) {}
	Label(const Label& rhs);
	Label& operator=(const Label& rhs);
	~Label();
	void clear() { mgr = 0; id = 0; }
	int getId() const { return id; }
	const uint8_t *getAddress() const;

	// backward compatibility
	static inline std::string toStr(int num)
	{
		char buf[16];
#if defined(_MSC_VER) && (_MSC_VER < 1900)
		_snprintf_s
#else
		snprintf
#endif
		(buf, sizeof(buf), ".%08x", num);
		return buf;
	}
};

class LabelManager {
	// for string label
	struct SlabelVal {
		size_t offset;
		SlabelVal(size_t offset) : offset(offset) {}
	};
	typedef XBYAK_STD_UNORDERED_MAP<std::string, SlabelVal> SlabelDefList;
	typedef XBYAK_STD_UNORDERED_MULTIMAP<std::string, const JmpLabel> SlabelUndefList;
	struct SlabelState {
		SlabelDefList defList;
		SlabelUndefList undefList;
	};
	typedef std::list<SlabelState> StateList;
	// for Label class
	struct ClabelVal {
		ClabelVal(size_t offset = 0) : offset(offset), refCount(1) {}
		size_t offset;
		int refCount;
	};
	typedef XBYAK_STD_UNORDERED_MAP<int, ClabelVal> ClabelDefList;
	typedef XBYAK_STD_UNORDERED_MULTIMAP<int, const JmpLabel> ClabelUndefList;
	typedef XBYAK_STD_UNORDERED_SET<Label*> LabelPtrList;

	CodeArray *base_;
	// global : stateList_.front(), local : stateList_.back()
	StateList stateList_;
	mutable int labelId_;
	ClabelDefList clabelDefList_;
	ClabelUndefList clabelUndefList_;
	LabelPtrList labelPtrList_;

	int getId(const Label& label) const
	{
		if (label.id == 0) label.id = labelId_++;
		return label.id;
	}
	template<class DefList, class UndefList, class T>
	void define_inner(DefList& defList, UndefList& undefList, const T& labelId, size_t addrOffset)
	{
		// add label
		typename DefList::value_type item(labelId, addrOffset);
		std::pair<typename DefList::iterator, bool> ret = defList.insert(item);
		if (!ret.second) XBYAK_THROW(ERR_LABEL_IS_REDEFINED)
		// search undefined label
		for (;;) {
			typename UndefList::iterator itr = undefList.find(labelId);
			if (itr == undefList.end()) break;
			const JmpLabel *jmp = &itr->second;
			const size_t offset = jmp->endOfJmp - jmp->jmpSize;
			size_t disp;
			if (jmp->mode == inner::LaddTop) {
				disp = addrOffset;
			} else if (jmp->mode == inner::Labs) {
				disp = size_t(base_->getCurr());
			} else {
				disp = addrOffset - jmp->endOfJmp + jmp->disp;
#ifdef XBYAK64
				if (jmp->jmpSize <= 4 && !inner::IsInInt32(disp)) XBYAK_THROW(ERR_OFFSET_IS_TOO_BIG)
#endif
				if (jmp->jmpSize == 1 && !inner::IsInDisp8((uint32_t)disp)) XBYAK_THROW(ERR_LABEL_IS_TOO_FAR)
			}
			if (base_->isAutoGrow()) {
				base_->save(offset, disp, jmp->jmpSize, jmp->mode);
			} else {
				base_->rewrite(offset, disp, jmp->jmpSize);
			}
			undefList.erase(itr);
		}
	}
	template<class DefList, class T>
	bool getOffset_inner(const DefList& defList, size_t *offset, const T& label) const
	{
		typename DefList::const_iterator i = defList.find(label);
		if (i == defList.end()) return false;
		*offset = i->second.offset;
		return true;
	}
	friend class Label;
	void incRefCount(int id, Label *label)
	{
		clabelDefList_[id].refCount++;
		labelPtrList_.insert(label);
	}
	void decRefCount(int id, Label *label)
	{
		labelPtrList_.erase(label);
		ClabelDefList::iterator i = clabelDefList_.find(id);
		if (i == clabelDefList_.end()) return;
		if (i->second.refCount == 1) {
			clabelDefList_.erase(id);
		} else {
			--i->second.refCount;
		}
	}
	template<class T>
	bool hasUndefinedLabel_inner(const T& list) const
	{
#ifndef NDEBUG
		for (typename T::const_iterator i = list.begin(); i != list.end(); ++i) {
			std::cerr << "undefined label:" << i->first << std::endl;
		}
#endif
		return !list.empty();
	}
	// detach all labels linked to LabelManager
	void resetLabelPtrList()
	{
		for (LabelPtrList::iterator i = labelPtrList_.begin(), ie = labelPtrList_.end(); i != ie; ++i) {
			(*i)->clear();
		}
		labelPtrList_.clear();
	}
public:
	LabelManager()
	{
		reset();
	}
	~LabelManager()
	{
		resetLabelPtrList();
	}
	void reset()
	{
		base_ = 0;
		labelId_ = 1;
		stateList_.clear();
		stateList_.push_back(SlabelState());
		stateList_.push_back(SlabelState());
		clabelDefList_.clear();
		clabelUndefList_.clear();
		resetLabelPtrList();
	}
	void enterLocal()
	{
		stateList_.push_back(SlabelState());
	}
	void leaveLocal()
	{
		if (stateList_.size() <= 2) XBYAK_THROW(ERR_UNDER_LOCAL_LABEL)
		if (hasUndefinedLabel_inner(stateList_.back().undefList)) XBYAK_THROW(ERR_LABEL_IS_NOT_FOUND)
		stateList_.pop_back();
	}
	void set(CodeArray *base) { base_ = base; }
	void defineSlabel(std::string label)
	{
		if (label == "@b" || label == "@f") XBYAK_THROW(ERR_BAD_LABEL_STR)
		if (label == "@@") {
			SlabelDefList& defList = stateList_.front().defList;
			SlabelDefList::iterator i = defList.find("@f");
			if (i != defList.end()) {
				defList.erase(i);
				label = "@b";
			} else {
				i = defList.find("@b");
				if (i != defList.end()) {
					defList.erase(i);
				}
				label = "@f";
			}
		}
		SlabelState& st = *label.c_str() == '.' ? stateList_.back() : stateList_.front();
		define_inner(st.defList, st.undefList, label, base_->getSize());
	}
	void defineClabel(Label& label)
	{
		define_inner(clabelDefList_, clabelUndefList_, getId(label), base_->getSize());
		label.mgr = this;
		labelPtrList_.insert(&label);
	}
	void assign(Label& dst, const Label& src)
	{
		ClabelDefList::const_iterator i = clabelDefList_.find(src.id);
		if (i == clabelDefList_.end()) XBYAK_THROW(ERR_LABEL_ISNOT_SET_BY_L)
		define_inner(clabelDefList_, clabelUndefList_, dst.id, i->second.offset);
		dst.mgr = this;
		labelPtrList_.insert(&dst);
	}
	bool getOffset(size_t *offset, std::string& label) const
	{
		const SlabelDefList& defList = stateList_.front().defList;
		if (label == "@b") {
			if (defList.find("@f") != defList.end()) {
				label = "@f";
			} else if (defList.find("@b") == defList.end()) {
				XBYAK_THROW_RET(ERR_LABEL_IS_NOT_FOUND, false)
			}
		} else if (label == "@f") {
			if (defList.find("@f") != defList.end()) {
				label = "@b";
			}
		}
		const SlabelState& st = *label.c_str() == '.' ? stateList_.back() : stateList_.front();
		return getOffset_inner(st.defList, offset, label);
	}
	bool getOffset(size_t *offset, const Label& label) const
	{
		return getOffset_inner(clabelDefList_, offset, getId(label));
	}
	void addUndefinedLabel(const std::string& label, const JmpLabel& jmp)
	{
		SlabelState& st = *label.c_str() == '.' ? stateList_.back() : stateList_.front();
		st.undefList.insert(SlabelUndefList::value_type(label, jmp));
	}
	void addUndefinedLabel(const Label& label, const JmpLabel& jmp)
	{
		clabelUndefList_.insert(ClabelUndefList::value_type(label.id, jmp));
	}
	bool hasUndefSlabel() const
	{
		for (StateList::const_iterator i = stateList_.begin(), ie = stateList_.end(); i != ie; ++i) {
			if (hasUndefinedLabel_inner(i->undefList)) return true;
		}
		return false;
	}
	bool hasUndefClabel() const { return hasUndefinedLabel_inner(clabelUndefList_); }
	const uint8_t *getCode() const { return base_->getCode(); }
	bool isReady() const { return !base_->isAutoGrow() || base_->isCalledCalcJmpAddress(); }
};

inline Label::Label(const Label& rhs)
{
	id = rhs.id;
	mgr = rhs.mgr;
	if (mgr) mgr->incRefCount(id, this);
}
inline Label& Label::operator=(const Label& rhs)
{
	if (id) XBYAK_THROW_RET(ERR_LABEL_IS_ALREADY_SET_BY_L, *this)
	id = rhs.id;
	mgr = rhs.mgr;
	if (mgr) mgr->incRefCount(id, this);
	return *this;
}
inline Label::~Label()
{
	if (id && mgr) mgr->decRefCount(id, this);
}
inline const uint8_t* Label::getAddress() const
{
	if (mgr == 0 || !mgr->isReady()) return 0;
	size_t offset;
	if (!mgr->getOffset(&offset, *this)) return 0;
	return mgr->getCode() + offset;
}

typedef enum {
	DefaultEncoding,
	VexEncoding,
	EvexEncoding
} PreferredEncoding;

class CodeGenerator : public CodeArray {
public:
	enum LabelType {
		T_SHORT,
		T_NEAR,
		T_FAR, // far jump
		T_AUTO // T_SHORT if possible
	};
private:
	CodeGenerator operator=(const CodeGenerator&); // don't call
#ifdef XBYAK64
	enum { i32e = 32 | 64, BIT = 64 };
	static const uint64_t dummyAddr = uint64_t(0x1122334455667788ull);
	typedef Reg64 NativeReg;
#else
	enum { i32e = 32, BIT = 32 };
	static const size_t dummyAddr = 0x12345678;
	typedef Reg32 NativeReg;
#endif
	// (XMM, XMM|MEM)
	static inline bool isXMM_XMMorMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isXMM() && (op2.isXMM() || op2.isMEM());
	}
	// (MMX, MMX|MEM) or (XMM, XMM|MEM)
	static inline bool isXMMorMMX_MEM(const Operand& op1, const Operand& op2)
	{
		return (op1.isMMX() && (op2.isMMX() || op2.isMEM())) || isXMM_XMMorMEM(op1, op2);
	}
	// (XMM, MMX|MEM)
	static inline bool isXMM_MMXorMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isXMM() && (op2.isMMX() || op2.isMEM());
	}
	// (MMX, XMM|MEM)
	static inline bool isMMX_XMMorMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isMMX() && (op2.isXMM() || op2.isMEM());
	}
	// (XMM, REG32|MEM)
	static inline bool isXMM_REG32orMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isXMM() && (op2.isREG(i32e) || op2.isMEM());
	}
	// (REG32, XMM|MEM)
	static inline bool isREG32_XMMorMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isREG(i32e) && (op2.isXMM() || op2.isMEM());
	}
	// (REG32, REG32|MEM)
	static inline bool isREG32_REG32orMEM(const Operand& op1, const Operand& op2)
	{
		return op1.isREG(i32e) && ((op2.isREG(i32e) && op1.getBit() == op2.getBit()) || op2.isMEM());
	}
	static inline bool isValidSSE(const Operand& op1)
	{
		// SSE instructions do not support XMM16 - XMM31
		return !(op1.isXMM() && op1.getIdx() >= 16);
	}
	static inline uint8_t rexRXB(int bit, int bit3, const Reg& r, const Reg& b, const Reg& x = Reg())
	{
		int v = bit3 ? 8 : 0;
		if (r.hasIdxBit(bit)) v |= 4;
		if (x.hasIdxBit(bit)) v |= 2;
		if (b.hasIdxBit(bit)) v |= 1;
		return uint8_t(v);
	}
	void rex2(int bit3, int rex4bit, const Reg& r, const Reg& b, const Reg& x = Reg())
	{
		db(0xD5);
		db((rexRXB(4, bit3, r, b, x) << 4) | rex4bit);
	}
	// return true if rex2 is selected
	bool rex(const Operand& op1, const Operand& op2 = Operand(), uint64_t type = 0)
	{
		if (op1.getNF() | op2.getNF()) XBYAK_THROW_RET(ERR_INVALID_NF, false)
		if (op1.getZU() | op2.getZU()) XBYAK_THROW_RET(ERR_INVALID_ZU, false)
		uint8_t rex = 0;
		const Operand *p1 = &op1, *p2 = &op2;
		if (p1->isMEM()) std::swap(p1, p2);
		if (p1->isMEM()) XBYAK_THROW_RET(ERR_BAD_COMBINATION, false)
		// except movsx(16bit, 32/64bit)
		bool p66 = (op1.isBit(16) && !op2.isBit(i32e)) || (op2.isBit(16) && !op1.isBit(i32e));
		if ((type & T_66) || p66) db(0x66);
		if (type & T_F2) {
			db(0xF2);
		}
		if (type & T_F3) {
			db(0xF3);
		}
		bool is0F = type & T_0F;
		if (p2->isMEM()) {
			const Reg& r = *static_cast<const Reg*>(p1);
			const Address& addr = p2->getAddress();
			const RegExp e = addr.getRegExp();
			const Reg& base = e.getBase();
			const Reg& idx = e.getIndex();
			if (BIT == 64 && addr.is32bit()) db(0x67);
			rex = rexRXB(3, r.isREG(64), r, base, idx);
			if (r.hasRex2() || addr.hasRex2()) {
				if (type & (T_0F38|T_0F3A)) XBYAK_THROW_RET(ERR_CANT_USE_REX2, false)
				rex2(is0F, rex, r, base, idx);
				return true;
			}
			if (rex || r.isExt8bit()) rex |= 0x40;
		} else {
			const Reg& r1 = static_cast<const Reg&>(op1);
			const Reg& r2 = static_cast<const Reg&>(op2);
			// ModRM(reg, base);
			rex = rexRXB(3, r1.isREG(64) || r2.isREG(64), r2, r1);
			if (r1.hasRex2() || r2.hasRex2()) {
				if (type & (T_0F38|T_0F3A)) XBYAK_THROW_RET(ERR_CANT_USE_REX2, 0)
				rex2(is0F, rex, r2, r1);
				return true;
			}
			if (rex || r1.isExt8bit() || r2.isExt8bit()) rex |= 0x40;
		}
		if (rex) db(rex);
		return false;
	}
	// @@@begin of avx_type_def.h
	static const uint64_t T_NONE = 0ull;
		// low 3 bit
	static const uint64_t T_N1 = 1ull;
	static const uint64_t T_N2 = 2ull;
	static const uint64_t T_N4 = 3ull;
	static const uint64_t T_N8 = 4ull;
	static const uint64_t T_N16 = 5ull;
	static const uint64_t T_N32 = 6ull;
	static const uint64_t T_NX_MASK = 7ull;
	static const uint64_t T_DUP = T_NX_MASK;//1 << 4, // N = (8, 32, 64)
	static const uint64_t T_N_VL = 1ull << 3; // N * (1, 2, 4) for VL
	static const uint64_t T_APX = 1ull << 4;
	static const uint64_t T_66 = 1ull << 5; // pp = 1
	static const uint64_t T_F3 = 1ull << 6; // pp = 2
	static const uint64_t T_ER_R = 1ull << 7; // reg{er}
	static const uint64_t T_0F = 1ull << 8;
	static const uint64_t T_0F38 = 1ull << 9;
	static const uint64_t T_0F3A = 1ull << 10;
	static const uint64_t T_L0 = 1ull << 11;
	static const uint64_t T_L1 = 1ull << 12;
	static const uint64_t T_W0 = 1ull << 13;
	static const uint64_t T_W1 = 1ull << 14;
	static const uint64_t T_EW0 = 1ull << 15;
	static const uint64_t T_EW1 = 1ull << 16;
	static const uint64_t T_YMM = 1ull << 17; // support YMM, ZMM
	static const uint64_t T_EVEX = 1ull << 18;
	static const uint64_t T_ER_X = 1ull << 19; // xmm{er}
	static const uint64_t T_ER_Y = 1ull << 20; // ymm{er}
	static const uint64_t T_ER_Z = 1ull << 21; // zmm{er}
	static const uint64_t T_SAE_X = 1ull << 22; // xmm{sae}
	static const uint64_t T_SAE_Y = 1ull << 23; // ymm{sae}
	static const uint64_t T_SAE_Z = 1ull << 24; // zmm{sae}
	static const uint64_t T_MUST_EVEX = 1ull << 25; // contains T_EVEX
	static const uint64_t T_B32 = 1ull << 26; // m32bcst
	static const uint64_t T_B64 = 1ull << 27; // m64bcst
	static const uint64_t T_B16 = T_B32 | T_B64; // m16bcst (Be careful)
	static const uint64_t T_M_K = 1ull << 28; // mem{k}
	static const uint64_t T_VSIB = 1ull << 29;
	static const uint64_t T_MEM_EVEX = 1ull << 30; // use evex if mem
	static const uint64_t T_FP16 = 1ull << 31; // avx512-fp16
	static const uint64_t T_MAP5 = T_FP16 | T_0F;
	static const uint64_t T_MAP6 = T_FP16 | T_0F38;
	static const uint64_t T_NF = 1ull << 32; // T_nf
	static const uint64_t T_CODE1_IF1 = 1ull << 33; // code|=1 if !r.isBit(8)

	static const uint64_t T_ND1 = 1ull << 35; // ND=1
	static const uint64_t T_ZU = 1ull << 36; // ND=ZU
	static const uint64_t T_F2 = 1ull << 37; // pp = 3
	// T_66 = 1, T_F3 = 2, T_F2 = 3
	static inline uint32_t getPP(uint64_t type) { return (type & T_66) ? 1 : (type & T_F3) ? 2 : (type & T_F2) ? 3 : 0; }
	// @@@end of avx_type_def.h
	static inline uint32_t getMap(uint64_t type) { return (type & T_0F) ? 1 : (type & T_0F38) ? 2 : (type & T_0F3A) ? 3 : 0; }
	void vex(const Reg& reg, const Reg& base, const Operand *v, uint64_t type, int code, bool x = false)
	{
		int w = (type & T_W1) ? 1 : 0;
		bool is256 = (type & T_L1) ? true : (type & T_L0) ? false : reg.isYMM();
		bool r = reg.isExtIdx();
		bool b = base.isExtIdx();
		int idx = v ? v->getIdx() : 0;
		if ((idx | reg.getIdx() | base.getIdx()) >= 16) XBYAK_THROW(ERR_BAD_COMBINATION)
		uint32_t pp = getPP(type);
		uint32_t vvvv = (((~idx) & 15) << 3) | (is256 ? 4 : 0) | pp;
		if (!b && !x && !w && (type & T_0F)) {
			db(0xC5); db((r ? 0 : 0x80) | vvvv);
		} else {
			uint32_t mmmm = getMap(type);
			db(0xC4); db((r ? 0 : 0x80) | (x ? 0 : 0x40) | (b ? 0 : 0x20) | mmmm); db((w << 7) | vvvv);
		}
		db(code);
	}
	void verifySAE(const Reg& r, uint64_t type) const
	{
		if (((type & T_SAE_X) && r.isXMM()) || ((type & T_SAE_Y) && r.isYMM()) || ((type & T_SAE_Z) && r.isZMM())) return;
		XBYAK_THROW(ERR_SAE_IS_INVALID)
	}
	void verifyER(const Reg& r, uint64_t type) const
	{
		if ((type & T_ER_R) && r.isREG(32|64)) return;
		if (((type & T_ER_X) && r.isXMM()) || ((type & T_ER_Y) && r.isYMM()) || ((type & T_ER_Z) && r.isZMM())) return;
		XBYAK_THROW(ERR_ER_IS_INVALID)
	}
	// (a, b, c) contains non zero two or three values then err
	int verifyDuplicate(int a, int b, int c, int err)
	{
		int v = a | b | c;
		if ((a > 0 && a != v) + (b > 0 && b != v) + (c > 0 && c != v) > 0) XBYAK_THROW_RET(err, 0)
		return v;
	}
	int evex(const Reg& reg, const Reg& base, const Operand *v, uint64_t type, int code, const Reg *x = 0, bool b = false, int aaa = 0, uint32_t VL = 0, bool Hi16Vidx = false)
	{
		if (!(type & (T_EVEX | T_MUST_EVEX))) XBYAK_THROW_RET(ERR_EVEX_IS_INVALID, 0)
		int w = (type & T_EW1) ? 1 : 0;
		uint32_t mmm = getMap(type);
		if (type & T_FP16) mmm |= 4;
		uint32_t pp = getPP(type);
		int idx = v ? v->getIdx() : 0;
		uint32_t vvvv = ~idx;

		bool R = reg.isExtIdx();
		bool X3 = (x && x->isExtIdx()) || (base.isSIMD() && base.isExtIdx2());
		bool B4 = base.isREG() && base.isExtIdx2();
		bool X4 = x && (x->isREG() && x->isExtIdx2());
		bool B = base.isExtIdx();
		bool Rp = reg.isExtIdx2();
		int LL;
		int rounding = verifyDuplicate(reg.getRounding(), base.getRounding(), v ? v->getRounding() : 0, ERR_ROUNDING_IS_ALREADY_SET);
		int disp8N = 1;
		if (rounding) {
			if (rounding == EvexModifierRounding::T_SAE) {
				verifySAE(base, type); LL = 0;
			} else {
				verifyER(base, type); LL = rounding - 1;
			}
			b = true;
		} else {
			if (v) VL = (std::max)(VL, v->getBit());
			VL = (std::max)((std::max)(reg.getBit(), base.getBit()), VL);
			LL = (VL == 512) ? 2 : (VL == 256) ? 1 : 0;
			if (b) {
				disp8N = ((type & T_B16) == T_B16) ? 2 : (type & T_B32) ? 4 : 8;
			} else if ((type & T_NX_MASK) == T_DUP) {
				disp8N = VL == 128 ? 8 : VL == 256 ? 32 : 64;
			} else {
				if ((type & (T_NX_MASK | T_N_VL)) == 0) {
					type |= T_N16 | T_N_VL; // default
				}
				int low = type & T_NX_MASK;
				if (low > 0) {
					disp8N = 1 << (low - 1);
					if (type & T_N_VL) disp8N *= (VL == 512 ? 4 : VL == 256 ? 2 : 1);
				}
			}
		}
		bool V4 = ((v ? v->isExtIdx2() : 0) || Hi16Vidx);
		bool z = reg.hasZero() || base.hasZero() || (v ? v->hasZero() : false);
		if (aaa == 0) aaa = verifyDuplicate(base.getOpmaskIdx(), reg.getOpmaskIdx(), (v ? v->getOpmaskIdx() : 0), ERR_OPMASK_IS_ALREADY_SET);
		if (aaa == 0) z = 0; // clear T_z if mask is not set
		db(0x62);
		db((R ? 0 : 0x80) | (X3 ? 0 : 0x40) | (B ? 0 : 0x20) | (Rp ? 0 : 0x10) | (B4 ? 8 : 0) | mmm);
		db((w == 1 ? 0x80 : 0) | ((vvvv & 15) << 3) | (X4 ? 0 : 4) | (pp & 3));
		db((z ? 0x80 : 0) | ((LL & 3) << 5) | (b ? 0x10 : 0) | (V4 ? 0 : 8) | (aaa & 7));
		db(code);
		return disp8N;
	}
	// evex of Legacy
	void evexLeg(const Reg& r, const Reg& b, const Reg& x, const Reg& v, uint64_t type, int sc = NONE)
	{
		int M = getMap(type); if (M == 0) M = 4; // legacy
		int R3 = !r.isExtIdx();
		int X3 = !x.isExtIdx();
		int B3 = b.isExtIdx() ? 0 : 0x20;
		int R4 = r.isExtIdx2() ? 0 : 0x10;
		int B4 = b.isExtIdx2() ? 0x08 : 0;
		int w = (type & T_W0) ? 0 : (r.isBit(64) || v.isBit(64) || (type & T_W1));
		int V = (~v.getIdx() & 15) << 3;
		int X4 = x.isExtIdx2() ? 0 : 0x04;
		int pp = (type & (T_F2|T_F3|T_66)) ? getPP(type) : (r.isBit(16) || v.isBit(16));
		int V4 = !v.isExtIdx2();
		int ND = (type & T_ZU) ? (r.getZU() || b.getZU()) : (type & T_ND1) ? 1 : (type & T_APX) ? 0 : v.isREG();
		int NF = r.getNF() | b.getNF() | x.getNF() | v.getNF();
		int L = 0;
		if ((type & T_NF) == 0 && NF) XBYAK_THROW(ERR_INVALID_NF)
		if ((type & T_ZU) == 0 && r.getZU()) XBYAK_THROW(ERR_INVALID_ZU)
		db(0x62);
		db((R3<<7) | (X3<<6) | B3 | R4 | B4 | M);
		db((w<<7) | V | X4 | pp);
		if (sc != NONE) {
			db((L<<5) | (ND<<4) | sc);
		} else {
			db((L<<5) | (ND<<4) | (V4<<3) | (NF<<2));
		}
	}
	void setModRM(int mod, int r1, int r2)
	{
		db(static_cast<uint8_t>((mod << 6) | ((r1 & 7) << 3) | (r2 & 7)));
	}
	void setSIB(const RegExp& e, int reg, int disp8N = 0)
	{
		uint64_t disp64 = e.getDisp();
#if defined(XBYAK64) && !defined(__ILP32__)
#ifdef XBYAK_OLD_DISP_CHECK
		// treat 0xffffffff as 0xffffffffffffffff
		uint64_t high = disp64 >> 32;
		if (high != 0 && high != 0xFFFFFFFF) XBYAK_THROW(ERR_OFFSET_IS_TOO_BIG)
#else
		// displacement should be a signed 32-bit value, so also check sign bit
		uint64_t high = disp64 >> 31;
		if (high != 0 && high != 0x1FFFFFFFF) XBYAK_THROW(ERR_OFFSET_IS_TOO_BIG)
#endif
#endif
		uint32_t disp = static_cast<uint32_t>(disp64);
		const Reg& base = e.getBase();
		const Reg& index = e.getIndex();
		const int baseIdx = base.getIdx();
		const int baseBit = base.getBit();
		const int indexBit = index.getBit();
		enum {
			mod00 = 0, mod01 = 1, mod10 = 2
		};
		int mod = mod10; // disp32
		if (!baseBit || ((baseIdx & 7) != Operand::EBP && disp == 0)) {
			mod = mod00;
		} else {
			if (disp8N == 0) {
				if (inner::IsInDisp8(disp)) {
					mod = mod01;
				}
			} else {
				// disp must be casted to signed
				uint32_t t = static_cast<uint32_t>(static_cast<int>(disp) / disp8N);
				if ((disp % disp8N) == 0 && inner::IsInDisp8(t)) {
					disp = t;
					mod = mod01;
				}
			}
		}
		const int newBaseIdx = baseBit ? (baseIdx & 7) : Operand::EBP;
		/* ModR/M = [2:3:3] = [Mod:reg/code:R/M] */
		bool hasSIB = indexBit || (baseIdx & 7) == Operand::ESP;
#ifdef XBYAK64
		if (!baseBit && !indexBit) hasSIB = true;
#endif
		if (hasSIB) {
			setModRM(mod, reg, Operand::ESP);
			/* SIB = [2:3:3] = [SS:index:base(=rm)] */
			const int idx = indexBit ? (index.getIdx() & 7) : Operand::ESP;
			const int scale = e.getScale();
			const int SS = (scale == 8) ? 3 : (scale == 4) ? 2 : (scale == 2) ? 1 : 0;
			setModRM(SS, idx, newBaseIdx);
		} else {
			setModRM(mod, reg, newBaseIdx);
		}
		if (mod == mod01) {
			db(disp);
		} else if (mod == mod10 || (mod == mod00 && !baseBit)) {
			dd(disp);
		}
	}
	LabelManager labelMgr_;
	bool isInDisp16(uint32_t x) const { return 0xFFFF8000 <= x || x <= 0x7FFF; }
	void writeCode(uint64_t type, const Reg& r, int code, bool rex2 = false)
	{
		if (!(type&T_APX || rex2)) {
			if (type & T_0F) {
				db(0x0F);
			} else if (type & T_0F38) {
				db(0x0F); db(0x38);
			} else if (type & T_0F3A) {
				db(0x0F); db(0x3A);
			}
		}
		db(code | ((type == 0 || (type & T_CODE1_IF1)) && !r.isBit(8)));
	}
	void opRR(const Reg& reg1, const Reg& reg2, uint64_t type, int code)
	{
		bool rex2 = rex(reg2, reg1, type);
		writeCode(type, reg1, code, rex2);
		setModRM(3, reg1.getIdx(), reg2.getIdx());
	}
	void opMR(const Address& addr, const Reg& r, uint64_t type, int code, int immSize = 0)
	{
		if (addr.is64bitDisp()) XBYAK_THROW(ERR_CANT_USE_64BIT_DISP)
		bool rex2 = rex(addr, r, type);
		writeCode(type, r, code, rex2);
		opAddr(addr, r.getIdx(), immSize);
	}
	void opLoadSeg(const Address& addr, const Reg& reg, uint64_t type, int code)
	{
		if (reg.isBit(8)) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		if (addr.is64bitDisp()) XBYAK_THROW(ERR_CANT_USE_64BIT_DISP)
		// can't use opMR
		rex(addr, reg, type);
		if (type & T_0F) db(0x0F);
		db(code);
		opAddr(addr, reg.getIdx());
	}
	// for only MPX(bnd*)
	void opMIB(const Address& addr, const Reg& reg, uint64_t type, int code)
	{
		if (addr.getMode() != Address::M_ModRM) XBYAK_THROW(ERR_INVALID_MIB_ADDRESS)
		opMR(addr.cloneNoOptimize(), reg, type, code);
	}
	void makeJmp(uint32_t disp, LabelType type, uint8_t shortCode, uint8_t longCode, uint8_t longPref)
	{
		const int shortJmpSize = 2;
		const int longHeaderSize = longPref ? 2 : 1;
		const int longJmpSize = longHeaderSize + 4;
		if (type != T_NEAR && inner::IsInDisp8(disp - shortJmpSize)) {
			db(shortCode); db(disp - shortJmpSize);
		} else {
			if (type == T_SHORT) XBYAK_THROW(ERR_LABEL_IS_TOO_FAR)
			if (longPref) db(longPref);
			db(longCode); dd(disp - longJmpSize);
		}
	}
	bool isNEAR(LabelType type) const { return type == T_NEAR || (type == T_AUTO && isDefaultJmpNEAR_); }
	template<class T>
	void opJmp(T& label, LabelType type, uint8_t shortCode, uint8_t longCode, uint8_t longPref)
	{
		if (type == T_FAR) XBYAK_THROW(ERR_NOT_SUPPORTED)
		if (isAutoGrow() && size_ + 16 >= maxSize_) growMemory(); /* avoid splitting code of jmp */
		size_t offset = 0;
		if (labelMgr_.getOffset(&offset, label)) { /* label exists */
			makeJmp(inner::VerifyInInt32(offset - size_), type, shortCode, longCode, longPref);
		} else {
			int jmpSize = 0;
			if (isNEAR(type)) {
				jmpSize = 4;
				if (longPref) db(longPref);
				db(longCode); dd(0);
			} else {
				jmpSize = 1;
				db(shortCode); db(0);
			}
			JmpLabel jmp(size_, jmpSize, inner::LasIs);
			labelMgr_.addUndefinedLabel(label, jmp);
		}
	}
	void opJmpAbs(const void *addr, LabelType type, uint8_t shortCode, uint8_t longCode, uint8_t longPref = 0)
	{
		if (type == T_FAR) XBYAK_THROW(ERR_NOT_SUPPORTED)
		if (isAutoGrow()) {
			if (!isNEAR(type)) XBYAK_THROW(ERR_ONLY_T_NEAR_IS_SUPPORTED_IN_AUTO_GROW)
			if (size_ + 16 >= maxSize_) growMemory();
			if (longPref) db(longPref);
			db(longCode);
			dd(0);
			save(size_ - 4, size_t(addr) - size_, 4, inner::Labs);
		} else {
			makeJmp(inner::VerifyInInt32(reinterpret_cast<const uint8_t*>(addr) - getCurr()), type, shortCode, longCode, longPref);
		}

	}
	void opJmpOp(const Operand& op, LabelType type, int ext)
	{
		const int bit = 16|i32e;
		if (type == T_FAR) {
			if (!op.isMEM(bit)) XBYAK_THROW(ERR_NOT_SUPPORTED)
			opRext(op, bit, ext + 1, 0, 0xFF, false);
		} else {
			opRext(op, bit, ext, 0, 0xFF, true);
		}
	}
	// reg is reg field of ModRM
	// immSize is the size for immediate value
	// disp8N = 0(normal), disp8N = 1(force disp32), disp8N = {2, 4, 8} ; compressed displacement
	void opAddr(const Address &addr, int reg, int immSize = 0, int disp8N = 0, bool permitVisb = false)
	{
		if (!permitVisb && addr.isVsib()) XBYAK_THROW(ERR_BAD_VSIB_ADDRESSING)
		if (addr.getMode() == Address::M_ModRM) {
			setSIB(addr.getRegExp(), reg, disp8N);
		} else if (addr.getMode() == Address::M_rip || addr.getMode() == Address::M_ripAddr) {
			setModRM(0, reg, 5);
			if (addr.getLabel()) { // [rip + Label]
				putL_inner(*addr.getLabel(), true, addr.getDisp() - immSize);
			} else {
				size_t disp = addr.getDisp();
				if (addr.getMode() == Address::M_ripAddr) {
					if (isAutoGrow()) XBYAK_THROW(ERR_INVALID_RIP_IN_AUTO_GROW)
					disp -= (size_t)getCurr() + 4 + immSize;
				}
				dd(inner::VerifyInInt32(disp));
			}
		}
	}
	void opSSE(const Reg& r, const Operand& op, uint64_t type, int code, bool isValid(const Operand&, const Operand&), int imm8 = NONE)
	{
		if (isValid && !isValid(r, op)) XBYAK_THROW(ERR_BAD_COMBINATION)
		if (!isValidSSE(r) || !isValidSSE(op)) XBYAK_THROW(ERR_NOT_SUPPORTED)
		opRO(r, op, type, code, true, (imm8 != NONE) ? 1 : 0);
		if (imm8 != NONE) db(imm8);
	}
	void opMMX_IMM(const Mmx& mmx, int imm8, int code, int ext)
	{
		if (!isValidSSE(mmx)) XBYAK_THROW(ERR_NOT_SUPPORTED)
		uint64_t type = T_0F;
		if (mmx.isXMM()) type |= T_66;
		opRR(Reg32(ext), mmx, type, code);
		db(imm8);
	}
	void opMMX(const Mmx& mmx, const Operand& op, int code, uint64_t type = T_0F, uint64_t pref = T_66, int imm8 = NONE)
	{
		if (mmx.isXMM()) type |= pref;
		opSSE(mmx, op, type, code, isXMMorMMX_MEM, imm8);
	}
	void opMovXMM(const Operand& op1, const Operand& op2, uint64_t type, int code)
	{
		if (!isValidSSE(op1) || !isValidSSE(op2)) XBYAK_THROW(ERR_NOT_SUPPORTED)
		if (op1.isXMM() && op2.isMEM()) {
			opMR(op2.getAddress(), op1.getReg(), type, code);
		} else if (op1.isMEM() && op2.isXMM()) {
			opMR(op1.getAddress(), op2.getReg(), type, code | 1);
		} else {
			XBYAK_THROW(ERR_BAD_COMBINATION)
		}
	}
	// pextr{w,b,d}, extractps
	void opExt(const Operand& op, const Mmx& mmx, int code, int imm, bool hasMMX2 = false)
	{
		if (!isValidSSE(op) || !isValidSSE(mmx)) XBYAK_THROW(ERR_NOT_SUPPORTED)
		if (hasMMX2 && op.isREG(i32e)) { /* pextrw is special */
			if (mmx.isXMM()) db(0x66);
			opRR(op.getReg(), mmx, T_0F, 0xC5); db(imm);
		} else {
			opSSE(mmx, op, T_66 | T_0F3A, code, isXMM_REG32orMEM, imm);
		}
	}
	// (r, r, m) or (r, m, r)
	bool opROO(const Reg& d, const Operand& op1, const Operand& op2, uint64_t type, int code, int immSize = 0, int sc = NONE)
	{
		if (!(type & T_MUST_EVEX) && !d.isREG() && !(d.hasRex2NFZU() || op1.hasRex2NFZU() || op2.hasRex2NFZU())) return false;
		const Operand *p1 = &op1, *p2 = &op2;
		if (p1->isMEM()) { std::swap(p1, p2); } else { if (p2->isMEM()) code |= 2; }
		if (p1->isMEM()) XBYAK_THROW_RET(ERR_BAD_COMBINATION, false)
		if (p2->isMEM()) {
			const Reg& r = *static_cast<const Reg*>(p1);
			const Address& addr = p2->getAddress();
			const RegExp e = addr.getRegExp();
			evexLeg(r, e.getBase(), e.getIndex(), d, type, sc);
			writeCode(type, d, code);
			opAddr(addr, r.getIdx(), immSize);
		} else {
			evexLeg(static_cast<const Reg&>(op2), static_cast<const Reg&>(op1), Reg(), d, type, sc);
			writeCode(type, d, code);
			setModRM(3, op2.getIdx(), op1.getIdx());
		}
		return true;
	}
	void opRext(const Operand& op, int bit, int ext, uint64_t type, int code, bool disableRex = false, int immSize = 0, const Reg *d = 0)
	{
		int opBit = op.getBit();
		if (disableRex && opBit == 64) opBit = 32;
		const Reg r(ext, Operand::REG, opBit);
		if ((type & T_APX) && op.hasRex2NFZU() && opROO(d ? *d : Reg(0, Operand::REG, opBit), op, r, type, code)) return;
		if (op.isMEM()) {
			opMR(op.getAddress(), r, type, code, immSize);
		} else if (op.isREG(bit)) {
			opRR(r, op.getReg().changeBit(opBit), type, code);
		} else {
			XBYAK_THROW(ERR_BAD_COMBINATION)
		}
	}
	void opShift(const Operand& op, int imm, int ext, const Reg *d = 0)
	{
		if (d == 0) verifyMemHasSize(op);
		if (d && op.getBit() != 0 && d->getBit() != op.getBit()) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		uint64_t type = T_APX|T_CODE1_IF1; if (ext & 8) type |= T_NF; if (d) type |= T_ND1;
		opRext(op, 0, ext&7, type, (0xC0 | ((imm == 1 ? 1 : 0) << 4)), false, (imm != 1) ? 1 : 0, d);
		if (imm != 1) db(imm);
	}
	void opShift(const Operand& op, const Reg8& _cl, int ext, const Reg *d = 0)
	{
		if (_cl.getIdx() != Operand::CL) XBYAK_THROW(ERR_BAD_COMBINATION)
		if (d && op.getBit() != 0 && d->getBit() != op.getBit()) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		uint64_t type = T_APX|T_CODE1_IF1; if (ext & 8) type |= T_NF; if (d) type |= T_ND1;
		opRext(op, 0, ext&7, type, 0xD2, false, 0, d);
	}
	// condR assumes that op.isREG() is true
	void opRO(const Reg& r, const Operand& op, uint64_t type, int code, bool condR = true, int immSize = 0)
	{
		if (op.isMEM()) {
			opMR(op.getAddress(), r, type, code, immSize);
		} else if (condR) {
			opRR(r, op.getReg(), type, code);
		} else {
			XBYAK_THROW(ERR_BAD_COMBINATION)
		}
	}
	void opShxd(const Reg& d, const Operand& op, const Reg& reg, uint8_t imm, int code, int code2, const Reg8 *_cl = 0)
	{
		if (_cl && _cl->getIdx() != Operand::CL) XBYAK_THROW(ERR_BAD_COMBINATION)
		if (!reg.isREG(16|i32e)) XBYAK_THROW(ERR_BAD_SIZE_OF_REGISTER)
		int immSize = _cl ? 0 : 1;
		if (_cl) code |= 1;
		uint64_t type = T_APX | T_NF;
		if (d.isREG()) type |= T_ND1;
		if (!opROO(d, op, reg, type, _cl ? code : code2, immSize)) {
			opRO(reg, op, T_0F, code, true, immSize);
		}
		if (!_cl) db(imm);
	}
	// (REG, REG|MEM), (MEM, REG)
	void opRO_MR(const Operand& op1, const Operand& op2, int code)
	{
		if (op2.isMEM()) {
			if (!op1.isREG()) XBYAK_THROW(ERR_BAD_COMBINATION)
			opMR(op2.getAddress(), op1.getReg(), 0, code | 2);
		} else {
			opRO(static_cast<const Reg&>(op2), op1, 0, code, op1.getKind() == op2.getKind());
		}
	}
	uint32_t getImmBit(const Operand& op, uint32_t imm)
	{
		verifyMemHasSize(op);
		uint32_t immBit = inner::IsInDisp8(imm) ? 8 : isInDisp16(imm) ? 16 : 32;
		if (op.isBit(8)) immBit = 8;
		if (op.getBit() < immBit) XBYAK_THROW_RET(ERR_IMM_IS_TOO_BIG, 0)
		if (op.isBit(32|64) && immBit == 16) immBit = 32; /* don't use MEM16 if 32/64bit mode */
		return immBit;
	}
	// (REG|MEM, IMM)
	void opOI(const Operand& op, uint32_t imm, int code, int ext)
	{
		uint32_t immBit = getImmBit(op, imm);
		if (op.isREG() && op.getIdx() == 0 && (op.getBit() == immBit || (op.isBit(64) && immBit == 32))) { // rax, eax, ax, al
			rex(op);
			db(code | 4 | (immBit == 8 ? 0 : 1));
		} else {
			int tmp = immBit < (std::min)(op.getBit(), 32U) ? 2 : 0;
			opRext(op, 0, ext, 0, 0x80 | tmp, false, immBit / 8);
		}
		db(imm, immBit / 8);
	}
	// (r, r/m, imm)
	void opROI(const Reg& d, const Operand& op, uint32_t imm, uint64_t type, int ext)
	{
		uint32_t immBit = getImmBit(d, imm);
		int code = immBit < (std::min)(d.getBit()