src/libFLAC/cpu.c

   1 /* libFLAC - Free Lossless Audio Codec library
   2  * Copyright (C) 2001,2002,2003,2004,2005,2006,2007  Josh Coalson
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  *
   8  * - Redistributions of source code must retain the above copyright
   9  * notice, this list of conditions and the following disclaimer.
  10  *
  11  * - Redistributions in binary form must reproduce the above copyright
  12  * notice, this list of conditions and the following disclaimer in the
  13  * documentation and/or other materials provided with the distribution.
  14  *
  15  * - Neither the name of the Xiph.org Foundation nor the names of its
  16  * contributors may be used to endorse or promote products derived from
  17  * this software without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  21  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  22  * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR
  23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
  26  * PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
  27  * LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
  28  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
  29  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  30  */
  31
  32 #if HAVE_CONFIG_H
  33 #  include <config.h>
  34 #endif
  35
  36 #include "private/cpu.h"
  37 #include <stdlib.h>
  38 #include <stdio.h>
  39
  40 #if defined FLAC__CPU_IA32
  41 # include <signal.h>
  42 #elif defined FLAC__CPU_PPC
  43 # if !defined FLAC__NO_ASM
  44 #  if defined FLAC__SYS_DARWIN
  45 #   include <sys/sysctl.h>
  46 #   include <mach/mach.h>
  47 #   include <mach/mach_host.h>
  48 #   include <mach/host_info.h>
  49 #   include <mach/machine.h>
  50 #   ifndef CPU_SUBTYPE_POWERPC_970
  51 #    define CPU_SUBTYPE_POWERPC_970 ((cpu_subtype_t) 100)
  52 #   endif
  53 #  else /* FLAC__SYS_DARWIN */
  54
  55 #   include <signal.h>
  56 #   include <setjmp.h>
  57
  58 static sigjmp_buf jmpbuf;
  59 static volatile sig_atomic_t canjump = 0;
  60
  61 static void sigill_handler (int sig)
  62 {
  63         if (!canjump) {
  64                 signal (sig, SIG_DFL);
  65                 raise (sig);
  66         }
  67         canjump = 0;
  68         siglongjmp (jmpbuf, 1);
  69 }
  70 #  endif /* FLAC__SYS_DARWIN */
  71 # endif /* FLAC__NO_ASM */
  72 #endif /* FLAC__CPU_PPC */
  73
  74 #if defined (__NetBSD__) || defined(__OpenBSD__)
  75 #include <sys/param.h>
  76 #include <sys/sysctl.h>
  77 #include <machine/cpu.h>
  78 #endif
  79
  80 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__)
  81 #include <sys/types.h>
  82 #include <sys/sysctl.h>
  83 #endif
  84
  85 #if defined(__APPLE__)
  86 /* how to get sysctlbyname()? */
  87 #endif
  88
  89 const unsigned FLAC__CPUINFO_IA32_CPUID_CMOV = 0x00008000;
  90 const unsigned FLAC__CPUINFO_IA32_CPUID_MMX = 0x00800000;
  91 const unsigned FLAC__CPUINFO_IA32_CPUID_FXSR = 0x01000000;
  92 const unsigned FLAC__CPUINFO_IA32_CPUID_SSE = 0x02000000;
  93 const unsigned FLAC__CPUINFO_IA32_CPUID_SSE2 = 0x04000000;
  94
  95 const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW = 0x80000000;
  96 const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW = 0x40000000;
  97 const unsigned FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX = 0x00400000;
  98
  99
 100 /*
 101  * Extra stuff needed for detection of OS support for SSE on IA-32
 102  */
 103 #if defined(FLAC__CPU_IA32) && !defined FLAC__NO_ASM && defined FLAC__HAS_NASM && !defined FLAC__NO_SSE_OS && !defined FLAC__SSE_OS
 104 # if defined(__linux__)
 105 /*
 106  * If the OS doesn't support SSE, we will get here with a SIGILL.  We
 107  * modify the return address to jump over the offending SSE instruction
 108  * and also the operation following it that indicates the instruction
 109  * executed successfully.  In this way we use no global variables and
 110  * stay thread-safe.
 111  *
 112  * 3 + 3 + 6:
 113  *   3 bytes for "xorps xmm0,xmm0"
 114  *   3 bytes for estimate of how long the follwing "inc var" instruction is
 115  *   6 bytes extra in case our estimate is wrong
 116  * 12 bytes puts us in the NOP "landing zone"
 117  */
 118 #  undef USE_OBSOLETE_SIGCONTEXT_FLAVOR /* #define this to use the older signal handler method */
 119 #  ifdef USE_OBSOLETE_SIGCONTEXT_FLAVOR
 120         static void sigill_handler_sse_os(int signal, struct sigcontext sc)
 121         {
 122                 (void)signal;
 123                 sc.eip += 3 + 3 + 6;
 124         }
 125 #  else
 126 #   include <sys/ucontext.h>
 127         static void sigill_handler_sse_os(int signal, siginfo_t *si, void *uc)
 128         {
 129                 (void)signal, (void)si;
 130                 ((ucontext_t*)uc)->uc_mcontext.gregs[14/*REG_EIP*/] += 3 + 3 + 6;
 131         }
 132 #  endif
 133 # elif defined(_MSC_VER)
 134 #  include <windows.h>
 135 #  undef USE_TRY_CATCH_FLAVOR /* #define this to use the try/catch method for catching illegal opcode exception */
 136 #  ifdef USE_TRY_CATCH_FLAVOR
 137 #  else
 138         LONG CALLBACK sigill_handler_sse_os(EXCEPTION_POINTERS *ep)
 139         {
 140                 if(ep->ExceptionRecord->ExceptionCode == EXCEPTION_ILLEGAL_INSTRUCTION) {
 141                         ep->ContextRecord->Eip += 3 + 3 + 6;
 142                         return EXCEPTION_CONTINUE_EXECUTION;
 143                 }
 144                 return EXCEPTION_CONTINUE_SEARCH;
 145         }
 146 #  endif
 147 # endif
 148 #endif
 149
 150
 151 void FLAC__cpu_info(FLAC__CPUInfo *info)
 152 {
 153 /*
 154  * IA32-specific
 155  */
 156 #ifdef FLAC__CPU_IA32
 157         info->type = FLAC__CPUINFO_TYPE_IA32;
 158 #if !defined FLAC__NO_ASM && defined FLAC__HAS_NASM
 159         info->use_asm = true;
 160         {
 161                 unsigned cpuid = FLAC__cpu_info_asm_ia32();
 162                 info->data.ia32.cmov = (cpuid & FLAC__CPUINFO_IA32_CPUID_CMOV)? true : false;
 163                 info->data.ia32.mmx  = (cpuid & FLAC__CPUINFO_IA32_CPUID_MMX )? true : false;
 164                 info->data.ia32.fxsr = (cpuid & FLAC__CPUINFO_IA32_CPUID_FXSR)? true : false;
 165                 info->data.ia32.sse  = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE )? true : false;
 166                 info->data.ia32.sse2 = (cpuid & FLAC__CPUINFO_IA32_CPUID_SSE2)? true : false;
 167
 168 #ifdef FLAC__USE_3DNOW
 169                 cpuid = FLAC__cpu_info_extended_amd_asm_ia32();
 170                 info->data.ia32._3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_3DNOW)? true : false;
 171                 info->data.ia32.ext3dnow = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXT3DNOW)? true : false;
 172                 info->data.ia32.extmmx = (cpuid & FLAC__CPUINFO_IA32_CPUID_EXTENDED_AMD_EXTMMX)? true : false;
 173 #else
 174                 info->data.ia32._3dnow = info->data.ia32.ext3dnow = info->data.ia32.extmmx = false;
 175 #endif
 176
 177                 /*
 178                  * now have to check for OS support of SSE/SSE2
 179                  */
 180                 if(info->data.ia32.fxsr || info->data.ia32.sse || info->data.ia32.sse2) {
 181 #if defined FLAC__NO_SSE_OS
 182                         /* assume user knows better than us; turn it off */
 183                         info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 184 #elif defined FLAC__SSE_OS
 185                         /* assume user knows better than us; leave as detected above */
 186 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
 187                         int sse = 0;
 188                         size_t len;
 189                         /* at least one of these must work: */
 190                         len = sizeof(sse); sse = sse || (sysctlbyname("hw.instruction_sse", &sse, &len, NULL, 0) == 0 && sse);
 191                         len = sizeof(sse); sse = sse || (sysctlbyname("hw.optional.sse"   , &sse, &len, NULL, 0) == 0 && sse); /* __APPLE__ ? */
 192                         if(!sse)
 193                                 info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 194 #elif defined(__NetBSD__) || defined (__OpenBSD__)
 195 # if __NetBSD_Version__ >= 105250000 || (defined __OpenBSD__)
 196                         int val = 0, mib[2] = { CTL_MACHDEP, CPU_SSE };
 197                         size_t len = sizeof(val);
 198                         if(sysctl(mib, 2, &val, &len, NULL, 0) < 0 || !val)
 199                                 info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 200                         else { /* double-check SSE2 */
 201                                 mib[1] = CPU_SSE2;
 202                                 len = sizeof(val);
 203                                 if(sysctl(mib, 2, &val, &len, NULL, 0) < 0 || !val)
 204                                         info->data.ia32.sse2 = false;
 205                         }
 206 # else
 207                         info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 208 # endif
 209 #elif defined(__linux__)
 210                         int sse = 0;
 211                         struct sigaction sigill_save;
 212 #ifdef USE_OBSOLETE_SIGCONTEXT_FLAVOR
 213                         if(0 == sigaction(SIGILL, NULL, &sigill_save) && signal(SIGILL, (void (*)(int))sigill_handler_sse_os) != SIG_ERR)
 214 #else
 215                         struct sigaction sigill_sse;
 216                         sigill_sse.sa_sigaction = sigill_handler_sse_os;
 217                         __sigemptyset(&sigill_sse.sa_mask);
 218                         sigill_sse.sa_flags = SA_SIGINFO | SA_RESETHAND; /* SA_RESETHAND just in case our SIGILL return jump breaks, so we don't get stuck in a loop */
 219                         if(0 == sigaction(SIGILL, &sigill_sse, &sigill_save))
 220 #endif
 221                         {
 222                                 /* http://www.ibiblio.org/gferg/ldp/GCC-Inline-Assembly-HOWTO.html */
 223                                 /* see sigill_handler_sse_os() for an explanation of the following: */
 224                                 asm volatile (
 225                                         "xorl %0,%0\n\t"          /* for some reason, still need to do this to clear 'sse' var */
 226                                         "xorps %%xmm0,%%xmm0\n\t" /* will cause SIGILL if unsupported by OS */
 227                                         "incl %0\n\t"             /* SIGILL handler will jump over this */
 228                                         /* landing zone */
 229                                         "nop\n\t" /* SIGILL jump lands here if "inc" is 9 bytes */
 230                                         "nop\n\t"
 231                                         "nop\n\t"
 232                                         "nop\n\t"
 233                                         "nop\n\t"
 234                                         "nop\n\t"
 235                                         "nop\n\t" /* SIGILL jump lands here if "inc" is 3 bytes (expected) */
 236                                         "nop\n\t"
 237                                         "nop"     /* SIGILL jump lands here if "inc" is 1 byte */
 238                                         : "=r"(sse)
 239                                         : "r"(sse)
 240                                 );
 241
 242                                 sigaction(SIGILL, &sigill_save, NULL);
 243                         }
 244
 245                         if(!sse)
 246                                 info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 247 #elif defined(_MSC_VER)
 248 # ifdef USE_TRY_CATCH_FLAVOR
 249                         _try {
 250                                 __asm {
 251 #  if _MSC_VER <= 1200
 252                                         /* VC6 assembler doesn't know SSE, have to emit bytecode instead */
 253                                         _emit 0x0F
 254                                         _emit 0x57
 255                                         _emit 0xC0
 256 #  else
 257                                         xorps xmm0,xmm0
 258 #  endif
 259                                 }
 260                         }
 261                         _except(EXCEPTION_EXECUTE_HANDLER) {
 262                                 if (_exception_code() == STATUS_ILLEGAL_INSTRUCTION)
 263                                         info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 264                         }
 265 # else
 266                         int sse = 0;
 267                         LPTOP_LEVEL_EXCEPTION_FILTER save = SetUnhandledExceptionFilter(sigill_handler_sse_os);
 268                         /* see GCC version above for explanation */
 269                         //@@@@@@ http://msdn2.microsoft.com/en-us/library/4ks26t93.aspx
 270                         //@@@@@@ http://www.daniweb.com/techtalkforums/thread8072.html
 271                         //@@@@@@ http://www.codeproject.com/cpp/gccasm.asp
 272                         //@@@@@@ http://www.hick.org/~mmiller/msvc_inline_asm.html
 273                         __asm {
 274 #  if _MSC_VER <= 1200
 275                                 /* VC6 assembler doesn't know SSE, have to emit bytecode instead */
 276                                 _emit 0x0F
 277                                 _emit 0x57
 278                                 _emit 0xC0
 279 #  else
 280                                 xorps xmm0,xmm0
 281 #  endif
 282                                 inc sse
 283                                 nop
 284                                 nop
 285                                 nop
 286                                 nop
 287                                 nop
 288                                 nop
 289                                 nop
 290                                 nop
 291                                 nop
 292                         }
 293                         SetUnhandledExceptionFilter(save);
 294                         if(!sse)
 295                                 info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 296 fprintf(stderr,"\n@@@@@@ SSE OS=%d %d\n",sse?1:0,sse);
 297 # endif
 298 #else
 299                         /* no way to test, disable to be safe */
 300                         info->data.ia32.fxsr = info->data.ia32.sse = info->data.ia32.sse2 = false;
 301 #endif
 302                 }
 303         }
 304 #else
 305         info->use_asm = false;
 306 #endif
 307
 308 /*
 309  * PPC-specific
 310  */
 311 #elif defined FLAC__CPU_PPC
 312         info->type = FLAC__CPUINFO_TYPE_PPC;
 313 # if !defined FLAC__NO_ASM
 314         info->use_asm = true;
 315 #  ifdef FLAC__USE_ALTIVEC
 316 #   if defined FLAC__SYS_DARWIN
 317         {
 318                 int val = 0, mib[2] = { CTL_HW, HW_VECTORUNIT };
 319                 size_t len = sizeof(val);
 320                 info->data.ppc.altivec = !(sysctl(mib, 2, &val, &len, NULL, 0) || !val);
 321         }
 322         {
 323                 host_basic_info_data_t hostInfo;
 324                 mach_msg_type_number_t infoCount;
 325
 326                 infoCount = HOST_BASIC_INFO_COUNT;
 327                 host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&hostInfo, &infoCount);
 328
 329                 info->data.ppc.ppc64 = (hostInfo.cpu_type == CPU_TYPE_POWERPC) && (hostInfo.cpu_subtype == CPU_SUBTYPE_POWERPC_970);
 330         }
 331 #   else /* FLAC__USE_ALTIVEC && !FLAC__SYS_DARWIN */
 332         {
 333                 /* no Darwin, do it the brute-force way */
 334                 /* @@@@@@ this is not thread-safe; replace with SSE OS method above or remove */
 335                 info->data.ppc.altivec = 0;
 336                 info->data.ppc.ppc64 = 0;
 337
 338                 signal (SIGILL, sigill_handler);
 339                 canjump = 0;
 340                 if (!sigsetjmp (jmpbuf, 1)) {
 341                         canjump = 1;
 342
 343                         asm volatile (
 344                                 "mtspr 256, %0\n\t"
 345                                 "vand %%v0, %%v0, %%v0"
 346                                 :
 347                                 : "r" (-1)
 348                         );
 349
 350                         info->data.ppc.altivec = 1;
 351                 }
 352                 canjump = 0;
 353                 if (!sigsetjmp (jmpbuf, 1)) {
 354                         int x = 0;
 355                         canjump = 1;
 356
 357                         /* PPC64 hardware implements the cntlzd instruction */
 358                         asm volatile ("cntlzd %0, %1" : "=r" (x) : "r" (x) );
 359
 360                         info->data.ppc.ppc64 = 1;
 361                 }
 362                 signal (SIGILL, SIG_DFL); /*@@@@@@ should save and restore old signal */
 363         }
 364 #   endif
 365 #  else /* !FLAC__USE_ALTIVEC */
 366         info->data.ppc.altivec = 0;
 367         info->data.ppc.ppc64 = 0;
 368 #  endif
 369 # else
 370         info->use_asm = false;
 371 # endif
 372
 373 /*
 374  * unknown CPI
 375  */
 376 #else
 377         info->type = FLAC__CPUINFO_TYPE_UNKNOWN;
 378         info->use_asm = false;
 379 #endif
 380 }