Restore Randen HWAES builds on CMake/MSVC/clang-cl. Build and preprocessor mismatches previously disabled hardware AES acceleration in Randen or broke non-AES builds: * CMake builds never populated HWAES compiler flags (silently falling back to software Randen across all CMake builds), and applied HWAES flags too broadly to detection, test, and benchmark targets instead of isolating them to `randen_hwaes.cc`. * Bazel and `platform.h` disabled HWAES on MSVC x64, `clang-cl` x86_64, MinGW, and Windows ARM64EC, while enabling `ABSL_HAVE_ACCELERATED_AES` under `-mavx -mno-aes` (breaking compilation) and passing unnecessary `-msse4.1`. * `randen_hwaes_test` silently passed even when a platform expected to build HWAES linked the unaccelerated stub, and `randen_benchmarks` crashed when HWAES was stubbed out. * 32-bit ARM builds passed `-mfpu=neon`, which never enables AES and overrides the crypto-capable default FPU of ARMv8 toolchains; big-endian ppc64 builds lacked `-mvsx`; and `platform.h` ignored `__ARM_FEATURE_AES`. Each left the unaccelerated stub in place. PiperOrigin-RevId: 1022546467
diff --git a/CMake/AbseilDll.cmake b/CMake/AbseilDll.cmake index 407b72a..1031168 100644 --- a/CMake/AbseilDll.cmake +++ b/CMake/AbseilDll.cmake
@@ -633,6 +633,18 @@ SHARED ${_dll_files} ) + if (NOT ABSL_INTERNAL_MAKE_DLL_TEST AND ABSL_RANDOM_RANDEN_COPTS) + set(_randen_hwaes_copts "") + foreach(_flag IN LISTS ABSL_RANDOM_RANDEN_COPTS) + string(REGEX REPLACE "^SHELL:" "" _flag "${_flag}") + string(APPEND _randen_hwaes_copts " ${_flag}") + endforeach() + string(STRIP "${_randen_hwaes_copts}" _randen_hwaes_copts) + set_source_files_properties( + "random/internal/randen_hwaes.cc" + PROPERTIES COMPILE_FLAGS "${_randen_hwaes_copts}" + ) + endif() target_link_libraries( ${_dll} PRIVATE
diff --git a/absl/copts/AbseilConfigureCopts.cmake b/absl/copts/AbseilConfigureCopts.cmake index f1d8b26..2d67569 100644 --- a/absl/copts/AbseilConfigureCopts.cmake +++ b/absl/copts/AbseilConfigureCopts.cmake
@@ -10,6 +10,17 @@ set(ABSL_BUILD_DLL FALSE) endif() +set(ABSL_RANDOM_HWAES_ARM64_FLAGS "-march=armv8-a+crypto") +set(ABSL_RANDOM_HWAES_X64_FLAGS "-maes") +# Big-endian ppc64 does not enable VSX by default, and -mcrypto alone does not +# define __VSX__. +set(ABSL_RANDOM_HWAES_PPC_FLAGS "-mcrypto" "-mvsx") + +set(_absl_randen_arch "${CMAKE_SYSTEM_PROCESSOR}") +if(MSVC AND CMAKE_CXX_COMPILER_ARCHITECTURE_ID) + set(_absl_randen_arch "${CMAKE_CXX_COMPILER_ARCHITECTURE_ID}") +endif() + if(APPLE AND CMAKE_CXX_COMPILER_ID MATCHES [[Clang]]) # Some CMake targets (not known at the moment of processing) could be set to # compile for multiple architectures as specified by the OSX_ARCHITECTURES @@ -52,23 +63,34 @@ if(ABSL_RANDOM_RANDEN_COPTS AND NOT ABSL_RANDOM_RANDEN_COPTS_WARNING) list(APPEND ABSL_RANDOM_RANDEN_COPTS "-Wno-unused-command-line-argument") endif() -elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64|AMD64") - if (MSVC) - set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_MSVC_X64_FLAGS}") +elseif(_absl_randen_arch MATCHES "^(x86_64|x64|amd64|AMD64)$") + if(NOT CMAKE_SIZEOF_VOID_P STREQUAL "8" OR + CMAKE_CXX_COMPILER_ID STREQUAL "MSVC") + # MSVC has no -maes equivalent; its AES intrinsics are always available. + set(ABSL_RANDOM_RANDEN_COPTS "") else() set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_X64_FLAGS}") endif() -elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "arm.*|aarch64") - if (CMAKE_SIZEOF_VOID_P STREQUAL "8") +elseif(_absl_randen_arch MATCHES "^(arm.*|ARM.*|aarch64|AARCH64)$") + if(WIN32) + # Windows ARM/ARM64/ARM64EC does not use runtime dispatch in platform.h. + set(ABSL_RANDOM_RANDEN_COPTS "") + elseif(CMAKE_SIZEOF_VOID_P STREQUAL "8") set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_ARM64_FLAGS}") elseif(CMAKE_SIZEOF_VOID_P STREQUAL "4") - set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_ARM32_FLAGS}") + # No flag is passed for 32-bit ARM. "-mfpu=neon" does not enable the AES + # extension and would override an ARMv8 toolchain's default FPU, which + # does; "-mfpu=crypto-neon-fp-armv8" is rejected for ARMv7 targets. + set(ABSL_RANDOM_RANDEN_COPTS "") else() message(WARNING "Value of CMAKE_SIZEOF_VOID_P (${CMAKE_SIZEOF_VOID_P}) is not supported.") endif() +elseif(_absl_randen_arch MATCHES "^(ppc64|powerpc64)(le)?$") + set(ABSL_RANDOM_RANDEN_COPTS "${ABSL_RANDOM_HWAES_PPC_FLAGS}") else() set(ABSL_RANDOM_RANDEN_COPTS "") endif() +unset(_absl_randen_arch) if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "QCC")
diff --git a/absl/random/CMakeLists.txt b/absl/random/CMakeLists.txt index cef4ab5..a91f021 100644 --- a/absl/random/CMakeLists.txt +++ b/absl/random/CMakeLists.txt
@@ -876,7 +876,6 @@ "internal/randen_hwaes.h" COPTS ${ABSL_DEFAULT_COPTS} - ${ABSL_RANDOM_RANDEN_COPTS} LINKOPTS ${ABSL_DEFAULT_LINKOPTS} DEPS
diff --git a/absl/random/internal/BUILD.bazel b/absl/random/internal/BUILD.bazel index 04f9a27..c39fddb 100644 --- a/absl/random/internal/BUILD.bazel +++ b/absl/random/internal/BUILD.bazel
@@ -43,22 +43,33 @@ licenses(["notice"]) -# Used to select on compilers that support GCC-compatible options -# (e.g. "-maes"). +# Matches compilers that reject GCC-style target feature flags +# (e.g. "-march=armv8-a+crypto", "-mfpu=neon", "-mcrypto"). selects.config_setting_group( - name = "gcc_compatible", + name = "non_gcc_compatible", match_any = [ - "@rules_cc//cc/compiler:clang", - "@rules_cc//cc/compiler:gcc", + "@rules_cc//cc/compiler:msvc-cl", + "@rules_cc//cc/compiler:clang-cl", + "@rules_cc//cc/compiler:emscripten", ], ) selects.config_setting_group( - name = "gcc_compatible-aarch32", - match_all = [ - ":gcc_compatible", - "@platforms//cpu:aarch32", - ], + name = "always", + match_any = ["//conditions:default"], +) + +# Matches every compiler that is not :non_gcc_compatible. This is an exclusion +# list rather than an allowlist of compilers so that toolchains reporting other +# compiler names (e.g. MinGW variants) still receive the hardware AES flags. +# select() has no negation, so when :non_gcc_compatible matches this resolves to +# compiler:gcc, which cannot also match because the compiler is single-valued. +alias( + name = "gcc_compatible", + actual = select({ + ":non_gcc_compatible": "@rules_cc//cc/compiler:gcc", + "//conditions:default": ":always", + }), ) selects.config_setting_group( @@ -73,7 +84,6 @@ name = "ppc_crypto", match_any = [ "@platforms//cpu:ppc", - "@platforms//cpu:ppc32", "@platforms//cpu:ppc64le", ], ) @@ -94,18 +104,31 @@ ], ) +# clang-cl rejects the other GCC-style flags above but accepts "-maes". +selects.config_setting_group( + name = "clang-cl-x86_64", + match_all = [ + "@rules_cc//cc/compiler:clang-cl", + "@platforms//cpu:x86_64", + ], +) + # Some libraries are compiled with options to generate AES-NI # instructions, and runtime dispatch is used to determine if the host # microarchitecture supports AES-NI or if a portable fallback library # should be called. +# +# No flag is passed for 32-bit ARM: "-mfpu=neon" does not enable the AES +# extension and would override an ARMv8 toolchain's default FPU, which does. ABSL_RANDOM_RANDEN_COPTS = select({ - ":gcc_compatible-aarch32": ["-mfpu=neon"], ":gcc_compatible-aarch64": ["-march=armv8-a+crypto"], - ":gcc_compatible-ppc_crypto": ["-mcrypto"], - ":gcc_compatible-x86_64": [ - "-maes", - "-msse4.1", + # Big-endian ppc64 does not enable VSX by default. + ":gcc_compatible-ppc_crypto": [ + "-mcrypto", + "-mvsx", ], + ":gcc_compatible-x86_64": ["-maes"], + ":clang-cl-x86_64": ["-maes"], "//conditions:default": [], }) @@ -743,7 +766,7 @@ name = "randen_hwaes_test", size = "small", srcs = ["randen_hwaes_test.cc"], - copts = ABSL_TEST_COPTS + ABSL_RANDOM_RANDEN_COPTS, + copts = ABSL_TEST_COPTS, linkopts = ABSL_DEFAULT_LINKOPTS, tags = ABSL_RANDOM_NONPORTABLE_TAGS, deps = [ @@ -838,7 +861,7 @@ size = "medium", timeout = "long", srcs = ["randen_benchmarks.cc"], - copts = ABSL_TEST_COPTS + ABSL_RANDOM_RANDEN_COPTS, + copts = ABSL_TEST_COPTS, flaky = 1, linkopts = ABSL_DEFAULT_LINKOPTS, tags = ABSL_RANDOM_NONPORTABLE_TAGS + ["benchmark"],
diff --git a/absl/random/internal/platform.h b/absl/random/internal/platform.h index a6fde4b..37a377a 100644 --- a/absl/random/internal/platform.h +++ b/absl/random/internal/platform.h
@@ -59,13 +59,17 @@ // including necessary headers to support hardware AES. // // ABSL_ARCH_{X86/PPC/ARM} macros determine the platform. -#if defined(__x86_64__) || defined(__x86_64) || defined(_M_AMD64) || \ +// +// AArch64 is checked first because MSVC defines both _M_ARM64EC and _M_X64 +// (and _M_AMD64) when targeting ARM64EC. +#if defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) || \ + defined(_M_ARM64EC) +#define ABSL_ARCH_AARCH64 +#elif defined(__x86_64__) || defined(__x86_64) || defined(_M_AMD64) || \ defined(_M_X64) #define ABSL_ARCH_X86_64 #elif defined(__i386) || defined(_M_IX86) #define ABSL_ARCH_X86_32 -#elif defined(__aarch64__) || defined(__arm64__) || defined(_M_ARM64) -#define ABSL_ARCH_AARCH64 #elif defined(__arm__) || defined(__ARMEL__) || defined(_M_ARM) #define ABSL_ARCH_ARM #elif defined(__powerpc64__) || defined(__PPC64__) || defined(__powerpc__) || \ @@ -99,7 +103,8 @@ #if defined(ABSL_ARCH_X86_64) -#if defined(__AES__) || defined(__AVX__) +// MSVC has no -maes equivalent; its AES intrinsics are always available. +#if defined(__AES__) || (defined(_MSC_VER) && !defined(__clang__)) #undef ABSL_HAVE_ACCELERATED_AES #define ABSL_HAVE_ACCELERATED_AES 1 #endif @@ -116,8 +121,10 @@ #elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64) // http://infocenter.arm.com/help/topic/com.arm.doc.ihi0053c/IHI0053C_acle_2_0.pdf -// Rely on NEON+CRYPTO extensions for ARM. -#if defined(__ARM_NEON) && defined(__ARM_FEATURE_CRYPTO) +// Rely on NEON+AES extensions for ARM. ACLE deprecated __ARM_FEATURE_CRYPTO +// (which also requires SHA1/SHA2) in favor of __ARM_FEATURE_AES. +#if defined(__ARM_NEON) && \ + (defined(__ARM_FEATURE_AES) || defined(__ARM_FEATURE_CRYPTO)) #undef ABSL_HAVE_ACCELERATED_AES #define ABSL_HAVE_ACCELERATED_AES 1 #endif
diff --git a/absl/random/internal/randen_benchmarks.cc b/absl/random/internal/randen_benchmarks.cc index 51e6a50..1a5be6d 100644 --- a/absl/random/internal/randen_benchmarks.cc +++ b/absl/random/internal/randen_benchmarks.cc
@@ -28,6 +28,7 @@ namespace { using absl::random_internal::CPUSupportsRandenHwAes; +using absl::random_internal::HasRandenHwAesImplementation; using absl::random_internal::Randen; using absl::random_internal::RandenHwAes; using absl::random_internal::RandenSlow; @@ -151,12 +152,16 @@ const FuncInput unpredictable = (argc != 999); static const FuncInput inputs[] = {unpredictable * 100, unpredictable * 1000}; - if (CPUSupportsRandenHwAes()) { + // RandenHwAes is a stub that exits when the implementation is not built. + const bool measure_hwaes = + HasRandenHwAesImplementation() && CPUSupportsRandenHwAes(); + + if (measure_hwaes) { Measure<AbsorbFn<RandenHwAes>>("Absorb (HwAes)", inputs); } Measure<AbsorbFn<RandenSlow>>("Absorb (Slow)", inputs); - if (CPUSupportsRandenHwAes()) { + if (measure_hwaes) { Measure<GenerateFn<RandenHwAes>>("Generate (HwAes)", inputs); } Measure<GenerateFn<RandenSlow>>("Generate (Slow)", inputs);
diff --git a/absl/random/internal/randen_hwaes_test.cc b/absl/random/internal/randen_hwaes_test.cc index 956e9d6..6890214 100644 --- a/absl/random/internal/randen_hwaes_test.cc +++ b/absl/random/internal/randen_hwaes_test.cc
@@ -28,7 +28,22 @@ using absl::random_internal::RandenHwAes; using absl::random_internal::RandenTraits; +// Whether the build is expected to compile the accelerated implementation +// into randen_hwaes.cc. On these configurations a missing implementation means +// that the hardware AES compiler flags were lost, so the test must fail rather +// than be skipped. ILP32 ABIs (such as x32) are excluded because the CMake +// build does not pass the flags for them. +#if ABSL_RANDOM_INTERNAL_AES_DISPATCH && !defined(__ILP32__) && \ + (defined(ABSL_ARCH_X86_64) || \ + (defined(ABSL_ARCH_AARCH64) && !defined(__ANDROID__))) && \ + (defined(__clang__) || defined(__GNUC__) || defined(_MSC_VER)) +constexpr bool kExpectHwAesImplementation = true; +#else +constexpr bool kExpectHwAesImplementation = false; +#endif + TEST(RandenHwAesTest, Default) { + ASSERT_TRUE(absl::random_internal::HasRandenHwAesImplementation()); EXPECT_TRUE(absl::random_internal::CPUSupportsRandenHwAes()); constexpr uint8_t kGolden[] = { @@ -92,7 +107,9 @@ int y = absl::random_internal::CPUSupportsRandenHwAes(); LOG(INFO) << "CPUSupportsRandenHwAes = " << y; - if (!x || !y) { + // When the implementation is expected, run the tests even if it is missing + // so that RandenHwAesTest.Default reports the failure. + if (!y || (!x && !kExpectHwAesImplementation)) { LOG(INFO) << "Skipping Randen HWAES tests."; return 0; }