| 1 | /* -*-c-*- |
| 2 | * |
| 3 | * CPU-specific dispatch |
| 4 | * |
| 5 | * (c) 2015 Straylight/Edgeware |
| 6 | */ |
| 7 | |
| 8 | /*----- Licensing notice --------------------------------------------------* |
| 9 | * |
| 10 | * This file is part of Catacomb. |
| 11 | * |
| 12 | * Catacomb is free software; you can redistribute it and/or modify |
| 13 | * it under the terms of the GNU Library General Public License as |
| 14 | * published by the Free Software Foundation; either version 2 of the |
| 15 | * License, or (at your option) any later version. |
| 16 | * |
| 17 | * Catacomb is distributed in the hope that it will be useful, |
| 18 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 19 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 20 | * GNU Library General Public License for more details. |
| 21 | * |
| 22 | * You should have received a copy of the GNU Library General Public |
| 23 | * License along with Catacomb; if not, write to the Free |
| 24 | * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston, |
| 25 | * MA 02111-1307, USA. |
| 26 | */ |
| 27 | |
| 28 | /*----- Header files ------------------------------------------------------*/ |
| 29 | |
| 30 | #include "config.h" |
| 31 | |
| 32 | #include <ctype.h> |
| 33 | #include <stdarg.h> |
| 34 | #include <stdio.h> |
| 35 | #include <stdlib.h> |
| 36 | #include <string.h> |
| 37 | |
| 38 | #include <mLib/macros.h> |
| 39 | |
| 40 | #include "dispatch.h" |
| 41 | |
| 42 | /*----- Intel x86/AMD64 feature probing -----------------------------------*/ |
| 43 | |
| 44 | #if CPUFAM_X86 || CPUFAM_AMD64 |
| 45 | |
| 46 | # define EFLAGS_ID (1u << 21) |
| 47 | # define CPUID1D_SSE2 (1u << 26) |
| 48 | # define CPUID1D_FXSR (1u << 24) |
| 49 | # define CPUID1C_AESNI (1u << 25) |
| 50 | |
| 51 | struct cpuid { unsigned a, b, c, d; }; |
| 52 | |
| 53 | /* --- @cpuid@ --- * |
| 54 | * |
| 55 | * Arguments: @struct cpuid *cc@ = where to write the result |
| 56 | * @unsigned a, c@ = EAX and ECX registers to set |
| 57 | * |
| 58 | * Returns: --- |
| 59 | * |
| 60 | * Use: Minimal C wrapper around the x86 `CPUID' instruction. Checks |
| 61 | * that the instruction is actually available before invoking |
| 62 | * it; fills the output structure with zero if it's not going to |
| 63 | * work. |
| 64 | */ |
| 65 | |
| 66 | #ifdef __GNUC__ |
| 67 | # if CPUFAM_X86 |
| 68 | static __inline__ unsigned getflags(void) |
| 69 | { unsigned f; __asm__ ("pushf; popl %0" : "=g" (f)); return (f); } |
| 70 | static __inline__ unsigned setflags(unsigned f) |
| 71 | { |
| 72 | unsigned ff; |
| 73 | __asm__ ("pushf; pushl %1; popf; pushf; popl %0; popf" |
| 74 | : "=g" (ff) |
| 75 | : "g" (f)); |
| 76 | return (ff); |
| 77 | } |
| 78 | # else |
| 79 | static __inline__ unsigned long getflags(void) |
| 80 | { unsigned long f; __asm__ ("pushf; popq %0" : "=g" (f)); return (f); } |
| 81 | static __inline__ unsigned long long setflags(unsigned long f) |
| 82 | { |
| 83 | unsigned long ff; |
| 84 | __asm__ ("pushf; pushq %1; popf; pushf; popq %0; popf" |
| 85 | : "=g" (ff) |
| 86 | : "g" (f)); |
| 87 | return (ff); |
| 88 | } |
| 89 | # endif |
| 90 | #endif |
| 91 | |
| 92 | static void cpuid(struct cpuid *cc, unsigned a, unsigned c) |
| 93 | { |
| 94 | #ifdef __GNUC__ |
| 95 | unsigned f; |
| 96 | #endif |
| 97 | |
| 98 | cc->a = cc->b = cc->c = cc->d = 0; |
| 99 | |
| 100 | #ifdef __GNUC__ |
| 101 | /* Stupid dance to detect whether the CPUID instruction is available. */ |
| 102 | f = getflags(); |
| 103 | if (!(setflags(f | EFLAGS_ID) & EFLAGS_ID) || |
| 104 | setflags(f & ~EFLAGS_ID) & EFLAGS_ID) { |
| 105 | dispatch_debug("CPUID instruction not available"); |
| 106 | return; |
| 107 | } |
| 108 | setflags(f); |
| 109 | |
| 110 | /* Alas, EBX is magical in PIC code, so abuse ESI instead. This isn't |
| 111 | * pretty, but it works. |
| 112 | */ |
| 113 | # if CPUFAM_X86 |
| 114 | __asm__ ("pushl %%ebx; cpuid; movl %%ebx, %%esi; popl %%ebx" |
| 115 | : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d) |
| 116 | : "a" (a) , "c" (c)); |
| 117 | # elif CPUFAM_AMD64 |
| 118 | __asm__ ("pushq %%rbx; cpuid; movl %%ebx, %%esi; popq %%rbx" |
| 119 | : "=a" (cc->a), "=S" (cc->b), "=c" (cc->c), "=d" (cc->d) |
| 120 | : "a" (a) , "c" (c)); |
| 121 | # else |
| 122 | # error "I'm confused." |
| 123 | # endif |
| 124 | dispatch_debug("CPUID(%08x, %08x) -> %08x, %08x, %08x, %08x", |
| 125 | a, c, cc->a, cc->b, cc->c, cc->d); |
| 126 | #else |
| 127 | dispatch_debug("GNU inline assembler not available; can't CPUID"); |
| 128 | #endif |
| 129 | } |
| 130 | |
| 131 | static unsigned cpuid_maxleaf(void) |
| 132 | { struct cpuid c; cpuid(&c, 0, 0); return (c.a); } |
| 133 | |
| 134 | /* --- @cpuid_features_p@ --- * |
| 135 | * |
| 136 | * Arguments: @unsigned dbits@ = bits to check in EDX |
| 137 | * @unsigned cbits@ = bits to check in ECX |
| 138 | * |
| 139 | * Returns: Nonzero if all the requested bits are set in the CPUID result |
| 140 | * on leaf 1. |
| 141 | */ |
| 142 | |
| 143 | static int cpuid_features_p(unsigned dbits, unsigned cbits) |
| 144 | { |
| 145 | struct cpuid c; |
| 146 | if (cpuid_maxleaf() < 1) return (0); |
| 147 | cpuid(&c, 1, 0); |
| 148 | return ((c.d & dbits) == dbits && (c.c & cbits) == cbits); |
| 149 | } |
| 150 | |
| 151 | /* --- @xmm_registers_available_p@ --- * |
| 152 | * |
| 153 | * Arguments: --- |
| 154 | * |
| 155 | * Returns: Nonzero if the operating system has made the XMM registers |
| 156 | * available for use. |
| 157 | */ |
| 158 | |
| 159 | static int xmm_registers_available_p(void) |
| 160 | { |
| 161 | #ifdef __GNUC__ |
| 162 | unsigned f; |
| 163 | /* This hack is by Agner Fog. Use FXSAVE/FXRSTOR to figure out whether the |
| 164 | * XMM registers are actually alive. |
| 165 | */ |
| 166 | if (!cpuid_features_p(CPUID1D_FXSR, 0)) return (0); |
| 167 | # if CPUFAM_X86 |
| 168 | __asm__ ("movl %%esp, %%edx; subl $512, %%esp; andl $~15, %%esp\n" |
| 169 | "fxsave (%%esp)\n" |
| 170 | "movl 160(%%esp), %%eax; xorl $0xaaaa5555, 160(%%esp)\n" |
| 171 | "fxrstor (%%esp); fxsave (%%esp)\n" |
| 172 | "movl 160(%%esp), %%ecx; movl %%eax, 160(%%esp)\n" |
| 173 | "fxrstor (%%esp); movl %%edx, %%esp\n" |
| 174 | "xorl %%ecx, %%eax" |
| 175 | : "=a" (f) |
| 176 | : /* no inputs */ |
| 177 | : "%ecx", "%edx"); |
| 178 | # elif CPUFAM_AMD64 |
| 179 | __asm__ ("movq %%rsp, %%rdx; subq $512, %%rsp; andq $~15, %%rsp\n" |
| 180 | "fxsave (%%rsp)\n" |
| 181 | "movl 160(%%rsp), %%eax; xorl $0xaaaa5555, 160(%%rsp)\n" |
| 182 | "fxrstor (%%rsp); fxsave (%%rsp)\n" |
| 183 | "movl 160(%%rsp), %%ecx; movl %%eax, 160(%%rsp)\n" |
| 184 | "fxrstor (%%rsp); movq %%rdx, %%rsp\n" |
| 185 | "xorl %%ecx, %%eax" |
| 186 | : "=a" (f) |
| 187 | : /* no inputs */ |
| 188 | : "%ecx", "%rdx"); |
| 189 | # else |
| 190 | # error "I'm confused." |
| 191 | # endif |
| 192 | dispatch_debug("XMM registers %savailable", f ? "" : "not "); |
| 193 | return (f); |
| 194 | #else |
| 195 | dispatch_debug("GNU inline assembler not available; can't check for XMM"); |
| 196 | return (0); |
| 197 | #endif |
| 198 | } |
| 199 | |
| 200 | #endif |
| 201 | |
| 202 | /*----- External interface ------------------------------------------------*/ |
| 203 | |
| 204 | /* --- @dispatch_debug@ --- * |
| 205 | * |
| 206 | * Arguments: @const char *fmt@ = a format string |
| 207 | * @...@ = additional arguments |
| 208 | * |
| 209 | * Returns: --- |
| 210 | * |
| 211 | * Use: Writes a formatted message to standard output if dispatch |
| 212 | * debugging is enabled. |
| 213 | */ |
| 214 | |
| 215 | void dispatch_debug(const char *fmt, ...) |
| 216 | { |
| 217 | va_list ap; |
| 218 | const char *e = getenv("CATACOMB_CPUDISPATCH_DEBUG"); |
| 219 | |
| 220 | if (e && *e != 'n' && *e != '0') { |
| 221 | va_start(ap, fmt); |
| 222 | fputs("Catacomb CPUDISPATCH: ", stderr); |
| 223 | vfprintf(stderr, fmt, ap); |
| 224 | fputc('\n', stderr); |
| 225 | va_end(ap); |
| 226 | } |
| 227 | } |
| 228 | |
| 229 | /* --- @check_env@ --- * |
| 230 | * |
| 231 | * Arguments: @const char *ftok@ = feature token |
| 232 | * |
| 233 | * Returns: Zero if the feature is forced off; positive if it's forced |
| 234 | * on; negative if the user hasn't decided. |
| 235 | * |
| 236 | * Use: Checks the environment variable `CATACOMB_CPUFEAT' for the |
| 237 | * feature token @ftok@. The variable, if it exists, should be |
| 238 | * a space-separated sequence of `+tok' and `-tok' items. These |
| 239 | * tokens may end in `*', which matches any suffix. |
| 240 | */ |
| 241 | |
| 242 | static int IGNORABLE check_env(const char *ftok) |
| 243 | { |
| 244 | const char *p, *q, *pp; |
| 245 | int d; |
| 246 | |
| 247 | p = getenv("CATACOMB_CPUFEAT"); |
| 248 | if (!p) return (-1); |
| 249 | |
| 250 | for (;;) { |
| 251 | while (isspace((unsigned char)*p)) p++; |
| 252 | if (!*p) return (-1); |
| 253 | switch (*p) { |
| 254 | case '+': d = +1; p++; break; |
| 255 | case '-': d = 0; p++; break; |
| 256 | default: d = -1; break; |
| 257 | } |
| 258 | for (q = p; *q && !isspace((unsigned char)*q); q++); |
| 259 | if (d >= 0) { |
| 260 | for (pp = ftok; p < q && *pp && *p == *pp; p++, pp++); |
| 261 | if ((p == q && !*pp) || (*p == '*' && p + 1 == q)) return (d); |
| 262 | } |
| 263 | p = q; |
| 264 | } |
| 265 | return (-1); |
| 266 | } |
| 267 | |
| 268 | /* --- @cpu_feature_p@ --- * |
| 269 | * |
| 270 | * Arguments: @unsigned feat@ = a @CPUFEAT_...@ code |
| 271 | * |
| 272 | * Returns: Nonzero if the feature is available. |
| 273 | */ |
| 274 | |
| 275 | #include <stdio.h> |
| 276 | |
| 277 | static int IGNORABLE |
| 278 | feat_debug(const char *ftok, const char *check, int verdict) |
| 279 | { |
| 280 | if (verdict >= 0) { |
| 281 | dispatch_debug("feature `%s': %s -> %s", ftok, check, |
| 282 | verdict ? "available" : "absent"); |
| 283 | } |
| 284 | return (verdict); |
| 285 | } |
| 286 | |
| 287 | int cpu_feature_p(int feat) |
| 288 | { |
| 289 | int IGNORABLE f; |
| 290 | IGNORE(f); |
| 291 | #define CASE_CPUFEAT(feat, ftok, cond) case CPUFEAT_##feat: \ |
| 292 | if ((f = feat_debug(ftok, "environment override", \ |
| 293 | check_env(ftok))) >= 0) \ |
| 294 | return (f); \ |
| 295 | else \ |
| 296 | return (feat_debug(ftok, "runtime probe", cond)); |
| 297 | |
| 298 | switch (feat) { |
| 299 | #if CPUFAM_X86 || CPUFAM_AMD64 |
| 300 | CASE_CPUFEAT(X86_SSE2, "x86:sse2", |
| 301 | xmm_registers_available_p() && |
| 302 | cpuid_features_p(CPUID1D_SSE2, 0)); |
| 303 | CASE_CPUFEAT(X86_AESNI, "x86:aesni", |
| 304 | xmm_registers_available_p() && |
| 305 | cpuid_features_p(CPUID1D_SSE2, CPUID1C_AESNI)); |
| 306 | #endif |
| 307 | default: |
| 308 | dispatch_debug("denying unknown feature %d", feat); |
| 309 | return (0); |
| 310 | } |
| 311 | #undef CASE_CPUFEAT |
| 312 | } |
| 313 | |
| 314 | /*----- That's all, folks -------------------------------------------------*/ |