1 // SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0-only 2 /* 3 * Copyright (c) 2018-2020, Facebook, Inc. 4 * All rights reserved. 5 * 6 * This source code is licensed under both the BSD-style license (found in the 7 * LICENSE file in the root directory of this source tree) and the GPLv2 (found 8 * in the COPYING file in the root directory of this source tree). 9 * You may select, at your option, one of the above-listed licenses. 10 */ 11 12 #ifndef ZSTD_COMMON_CPU_H 13 #define ZSTD_COMMON_CPU_H 14 15 /** 16 * Implementation taken from folly/CpuId.h 17 * https://github.com/facebook/folly/blob/master/folly/CpuId.h 18 */ 19 20 #include <string.h> 21 22 #include "mem.h" 23 24 #ifdef _MSC_VER 25 #include <intrin.h> 26 #endif 27 28 typedef struct { 29 U32 f1c; 30 U32 f1d; 31 U32 f7b; 32 U32 f7c; 33 } ZSTD_cpuid_t; 34 35 MEM_STATIC ZSTD_cpuid_t ZSTD_cpuid(void) { 36 U32 f1c = 0; 37 U32 f1d = 0; 38 U32 f7b = 0; 39 U32 f7c = 0; 40 #if defined(_MSC_VER) && (defined(_M_X64) || defined(_M_IX86)) 41 int reg[4]; 42 __cpuid((int*)reg, 0); 43 { 44 int const n = reg[0]; 45 if (n >= 1) { 46 __cpuid((int*)reg, 1); 47 f1c = (U32)reg[2]; 48 f1d = (U32)reg[3]; 49 } 50 if (n >= 7) { 51 __cpuidex((int*)reg, 7, 0); 52 f7b = (U32)reg[1]; 53 f7c = (U32)reg[2]; 54 } 55 } 56 #elif defined(__i386__) && defined(__PIC__) && !defined(__clang__) && defined(__GNUC__) 57 /* The following block like the normal cpuid branch below, but gcc 58 * reserves ebx for use of its pic register so we must specially 59 * handle the save and restore to avoid clobbering the register 60 */ 61 U32 n; 62 __asm__( 63 "pushl %%ebx\n\t" 64 "cpuid\n\t" 65 "popl %%ebx\n\t" 66 : "=a"(n) 67 : "a"(0) 68 : "ecx", "edx"); 69 if (n >= 1) { 70 U32 f1a; 71 __asm__( 72 "pushl %%ebx\n\t" 73 "cpuid\n\t" 74 "popl %%ebx\n\t" 75 : "=a"(f1a), "=c"(f1c), "=d"(f1d) 76 : "a"(1)); 77 } 78 if (n >= 7) { 79 __asm__( 80 "pushl %%ebx\n\t" 81 "cpuid\n\t" 82 "movl %%ebx, %%eax\n\t" 83 "popl %%ebx" 84 : "=a"(f7b), "=c"(f7c) 85 : "a"(7), "c"(0) 86 : "edx"); 87 } 88 #elif defined(__x86_64__) || defined(_M_X64) || defined(__i386__) 89 U32 n; 90 __asm__("cpuid" : "=a"(n) : "a"(0) : "ebx", "ecx", "edx"); 91 if (n >= 1) { 92 U32 f1a; 93 __asm__("cpuid" : "=a"(f1a), "=c"(f1c), "=d"(f1d) : "a"(1) : "ebx"); 94 } 95 if (n >= 7) { 96 U32 f7a; 97 __asm__("cpuid" 98 : "=a"(f7a), "=b"(f7b), "=c"(f7c) 99 : "a"(7), "c"(0) 100 : "edx"); 101 } 102 #endif 103 { 104 ZSTD_cpuid_t cpuid; 105 cpuid.f1c = f1c; 106 cpuid.f1d = f1d; 107 cpuid.f7b = f7b; 108 cpuid.f7c = f7c; 109 return cpuid; 110 } 111 } 112 113 #define X(name, r, bit) \ 114 MEM_STATIC int ZSTD_cpuid_##name(ZSTD_cpuid_t const cpuid) { \ 115 return ((cpuid.r) & (1U << bit)) != 0; \ 116 } 117 118 /* cpuid(1): Processor Info and Feature Bits. */ 119 #define C(name, bit) X(name, f1c, bit) 120 C(sse3, 0) 121 C(pclmuldq, 1) 122 C(dtes64, 2) 123 C(monitor, 3) 124 C(dscpl, 4) 125 C(vmx, 5) 126 C(smx, 6) 127 C(eist, 7) 128 C(tm2, 8) 129 C(ssse3, 9) 130 C(cnxtid, 10) 131 C(fma, 12) 132 C(cx16, 13) 133 C(xtpr, 14) 134 C(pdcm, 15) 135 C(pcid, 17) 136 C(dca, 18) 137 C(sse41, 19) 138 C(sse42, 20) 139 C(x2apic, 21) 140 C(movbe, 22) 141 C(popcnt, 23) 142 C(tscdeadline, 24) 143 C(aes, 25) 144 C(xsave, 26) 145 C(osxsave, 27) 146 C(avx, 28) 147 C(f16c, 29) 148 C(rdrand, 30) 149 #undef C 150 #define D(name, bit) X(name, f1d, bit) 151 D(fpu, 0) 152 D(vme, 1) 153 D(de, 2) 154 D(pse, 3) 155 D(tsc, 4) 156 D(msr, 5) 157 D(pae, 6) 158 D(mce, 7) 159 D(cx8, 8) 160 D(apic, 9) 161 D(sep, 11) 162 D(mtrr, 12) 163 D(pge, 13) 164 D(mca, 14) 165 D(cmov, 15) 166 D(pat, 16) 167 D(pse36, 17) 168 D(psn, 18) 169 D(clfsh, 19) 170 D(ds, 21) 171 D(acpi, 22) 172 D(mmx, 23) 173 D(fxsr, 24) 174 D(sse, 25) 175 D(sse2, 26) 176 D(ss, 27) 177 D(htt, 28) 178 D(tm, 29) 179 D(pbe, 31) 180 #undef D 181 182 /* cpuid(7): Extended Features. */ 183 #define B(name, bit) X(name, f7b, bit) 184 B(bmi1, 3) 185 B(hle, 4) 186 B(avx2, 5) 187 B(smep, 7) 188 B(bmi2, 8) 189 B(erms, 9) 190 B(invpcid, 10) 191 B(rtm, 11) 192 B(mpx, 14) 193 B(avx512f, 16) 194 B(avx512dq, 17) 195 B(rdseed, 18) 196 B(adx, 19) 197 B(smap, 20) 198 B(avx512ifma, 21) 199 B(pcommit, 22) 200 B(clflushopt, 23) 201 B(clwb, 24) 202 B(avx512pf, 26) 203 B(avx512er, 27) 204 B(avx512cd, 28) 205 B(sha, 29) 206 B(avx512bw, 30) 207 B(avx512vl, 31) 208 #undef B 209 #define C(name, bit) X(name, f7c, bit) 210 C(prefetchwt1, 0) 211 C(avx512vbmi, 1) 212 #undef C 213 214 #undef X 215 216 #endif /* ZSTD_COMMON_CPU_H */ 217