/* * Runtime CPU detection for x86 * (C) 2009,2010,2013,2017 Jack Lloyd * * Botan is released under the Simplified BSD License (see license.txt) */ #include #include #include #if defined(BOTAN_TARGET_CPU_IS_X86_FAMILY) #if defined(BOTAN_BUILD_COMPILER_IS_MSVC) #include #elif defined(BOTAN_BUILD_COMPILER_IS_INTEL) #include #elif defined(BOTAN_BUILD_COMPILER_IS_GCC) || defined(BOTAN_BUILD_COMPILER_IS_CLANG) #include #endif #endif namespace Botan { #if defined(BOTAN_TARGET_CPU_IS_X86_FAMILY) uint64_t CPUID::CPUID_Data::detect_cpu_features(size_t* cache_line_size) { #if defined(BOTAN_BUILD_COMPILER_IS_MSVC) #define X86_CPUID(type, out) do { __cpuid((int*)out, type); } while(0) #define X86_CPUID_SUBLEVEL(type, level, out) do { __cpuidex((int*)out, type, level); } while(0) #elif defined(BOTAN_BUILD_COMPILER_IS_INTEL) #define X86_CPUID(type, out) do { __cpuid(out, type); } while(0) #define X86_CPUID_SUBLEVEL(type, level, out) do { __cpuidex((int*)out, type, level); } while(0) #elif defined(BOTAN_TARGET_ARCH_IS_X86_64) && defined(BOTAN_USE_GCC_INLINE_ASM) #define X86_CPUID(type, out) \ asm("cpuid\n\t" : "=a" (out[0]), "=b" (out[1]), "=c" (out[2]), "=d" (out[3]) \ : "0" (type)) #define X86_CPUID_SUBLEVEL(type, level, out) \ asm("cpuid\n\t" : "=a" (out[0]), "=b" (out[1]), "=c" (out[2]), "=d" (out[3]) \ : "0" (type), "2" (level)) #elif defined(BOTAN_BUILD_COMPILER_IS_GCC) || defined(BOTAN_BUILD_COMPILER_IS_CLANG) #define X86_CPUID(type, out) do { __get_cpuid(type, out, out+1, out+2, out+3); } while(0) #define X86_CPUID_SUBLEVEL(type, level, out) \ do { __cpuid_count(type, level, out[0], out[1], out[2], out[3]); } while(0) #else #warning "No way of calling x86 cpuid instruction for this compiler" #define X86_CPUID(type, out) do { clear_mem(out, 4); } while(0) #define X86_CPUID_SUBLEVEL(type, level, out) do { clear_mem(out, 4); } while(0) #endif uint64_t features_detected = 0; uint32_t cpuid[4] = { 0 }; // CPUID 0: vendor identification, max sublevel X86_CPUID(0, cpuid); const uint32_t max_supported_sublevel = cpuid[0]; const uint32_t INTEL_CPUID[3] = { 0x756E6547, 0x6C65746E, 0x49656E69 }; const uint32_t AMD_CPUID[3] = { 0x68747541, 0x444D4163, 0x69746E65 }; const bool is_intel = same_mem(cpuid + 1, INTEL_CPUID, 3); const bool is_amd = same_mem(cpuid + 1, AMD_CPUID, 3); if(max_supported_sublevel >= 1) { // CPUID 1: feature bits X86_CPUID(1, cpuid); const uint64_t flags0 = (static_cast(cpuid[2]) << 32) | cpuid[3]; enum x86_CPUID_1_bits : uint64_t { RDTSC = (1ULL << 4), SSE2 = (1ULL << 26), CLMUL = (1ULL << 33), SSSE3 = (1ULL << 41), SSE41 = (1ULL << 51), SSE42 = (1ULL << 52), AESNI = (1ULL << 57), RDRAND = (1ULL << 62) }; if(flags0 & x86_CPUID_1_bits::RDTSC) features_detected |= CPUID::CPUID_RDTSC_BIT; if(flags0 & x86_CPUID_1_bits::SSE2) features_detected |= CPUID::CPUID_SSE2_BIT; if(flags0 & x86_CPUID_1_bits::CLMUL) features_detected |= CPUID::CPUID_CLMUL_BIT; if(flags0 & x86_CPUID_1_bits::SSSE3) features_detected |= CPUID::CPUID_SSSE3_BIT; if(flags0 & x86_CPUID_1_bits::SSE41) features_detected |= CPUID::CPUID_SSE41_BIT; if(flags0 & x86_CPUID_1_bits::SSE42) features_detected |= CPUID::CPUID_SSE42_BIT; if(flags0 & x86_CPUID_1_bits::AESNI) features_detected |= CPUID::CPUID_AESNI_BIT; if(flags0 & x86_CPUID_1_bits::RDRAND) features_detected |= CPUID::CPUID_RDRAND_BIT; } if(is_intel) { // Intel cache line size is in cpuid(1) output *cache_line_size = 8 * get_byte(2, cpuid[1]); } else if(is_amd) { // AMD puts it in vendor zone X86_CPUID(0x80000005, cpuid); *cache_line_size = get_byte(3, cpuid[2]); } if(max_supported_sublevel >= 7) { clear_mem(cpuid, 4); X86_CPUID_SUBLEVEL(7, 0, cpuid); enum x86_CPUID_7_bits : uint64_t { BMI1 = (1ULL << 3), AVX2 = (1ULL << 5), BMI2 = (1ULL << 8), AVX512_F = (1ULL << 16), AVX512_DQ = (1ULL << 17), RDSEED = (1ULL << 18), ADX = (1ULL << 19), AVX512_IFMA = (1ULL << 21), SHA = (1ULL << 29), AVX512_BW = (1ULL << 30), AVX512_VL = (1ULL << 31), AVX512_VBMI = (1ULL << 33), AVX512_VBMI2 = (1ULL << 38), AVX512_VAES = (1ULL << 41), AVX512_VCLMUL = (1ULL << 42), AVX512_VBITALG = (1ULL << 44), }; const uint64_t flags7 = (static_cast(cpuid[2]) << 32) | cpuid[1]; if(flags7 & x86_CPUID_7_bits::AVX2) features_detected |= CPUID::CPUID_AVX2_BIT; if(flags7 & x86_CPUID_7_bits::BMI1) { features_detected |= CPUID::CPUID_BMI1_BIT; /* We only set the BMI2 bit if BMI1 is also supported, so BMI2 code can safely use both extensions. No known processor implements BMI2 but not BMI1. */ if(flags7 & x86_CPUID_7_bits::BMI2) features_detected |= CPUID::CPUID_BMI2_BIT; } if(flags7 & x86_CPUID_7_bits::AVX512_F) { features_detected |= CPUID::CPUID_AVX512F_BIT; if(flags7 & x86_CPUID_7_bits::AVX512_DQ) features_detected |= CPUID::CPUID_AVX512DQ_BIT; if(flags7 & x86_CPUID_7_bits::AVX512_BW) features_detected |= CPUID::CPUID_AVX512BW_BIT; const uint64_t ICELAKE_FLAGS = x86_CPUID_7_bits::AVX512_F | x86_CPUID_7_bits::AVX512_DQ | x86_CPUID_7_bits::AVX512_IFMA | x86_CPUID_7_bits::AVX512_BW | x86_CPUID_7_bits::AVX512_VL | x86_CPUID_7_bits::AVX512_VBMI | x86_CPUID_7_bits::AVX512_VBMI2 | x86_CPUID_7_bits::AVX512_VBITALG; if((flags7 & ICELAKE_FLAGS) == ICELAKE_FLAGS) features_detected |= CPUID::CPUID_AVX512_ICL_BIT; if(flags7 & x86_CPUID_7_bits::AVX512_VAES) features_detected |= CPUID::CPUID_AVX512_AES_BIT; if(flags7 & x86_CPUID_7_bits::AVX512_VCLMUL) features_detected |= CPUID::CPUID_AVX512_CLMUL_BIT; } if(flags7 & x86_CPUID_7_bits::RDSEED) features_detected |= CPUID::CPUID_RDSEED_BIT; if(flags7 & x86_CPUID_7_bits::ADX) features_detected |= CPUID::CPUID_ADX_BIT; if(flags7 & x86_CPUID_7_bits::SHA) features_detected |= CPUID::CPUID_SHA_BIT; } #undef X86_CPUID #undef X86_CPUID_SUBLEVEL /* * If we don't have access to CPUID, we can still safely assume that * any x86-64 processor has SSE2 and RDTSC */ #if defined(BOTAN_TARGET_ARCH_IS_X86_64) if(features_detected == 0) { features_detected |= CPUID::CPUID_SSE2_BIT; features_detected |= CPUID::CPUID_RDTSC_BIT; } #endif return features_detected; } #endif }