Add libavutil/internal.h #include, required for the DECLARE_ALIGNED macro.
[mplayer/glamo.git] / cpudetect.c
blob592eb65090b35592cc6142176c7e29d9bd54c3d5
1 #include "config.h"
2 #include "cpudetect.h"
3 #include "mp_msg.h"
5 CpuCaps gCpuCaps;
7 #include <stdlib.h>
9 #if ARCH_X86
11 #include <stdio.h>
12 #include <string.h>
14 #if defined (__NetBSD__) || defined(__OpenBSD__)
15 #include <sys/param.h>
16 #include <sys/sysctl.h>
17 #include <machine/cpu.h>
18 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
19 #include <sys/types.h>
20 #include <sys/sysctl.h>
21 #elif defined(__linux__)
22 #include <signal.h>
23 #elif defined(__MINGW32__) || defined(__CYGWIN__)
24 #include <windows.h>
25 #elif defined(__OS2__)
26 #define INCL_DOS
27 #include <os2.h>
28 #elif defined(__AMIGAOS4__)
29 #include <proto/exec.h>
30 #endif
32 /* Thanks to the FreeBSD project for some of this cpuid code, and
33 * help understanding how to use it. Thanks to the Mesa
34 * team for SSE support detection and more cpu detect code.
37 /* I believe this code works. However, it has only been used on a PII and PIII */
39 static void check_os_katmai_support( void );
41 // return TRUE if cpuid supported
42 static int has_cpuid(void)
44 // code from libavcodec:
45 #if ARCH_X86_64
46 return 1;
47 #else
48 long a, c;
49 __asm__ volatile (
50 /* See if CPUID instruction is supported ... */
51 /* ... Get copies of EFLAGS into eax and ecx */
52 "pushfl\n\t"
53 "pop %0\n\t"
54 "mov %0, %1\n\t"
56 /* ... Toggle the ID bit in one copy and store */
57 /* to the EFLAGS reg */
58 "xor $0x200000, %0\n\t"
59 "push %0\n\t"
60 "popfl\n\t"
62 /* ... Get the (hopefully modified) EFLAGS */
63 "pushfl\n\t"
64 "pop %0\n\t"
65 : "=a" (a), "=c" (c)
67 : "cc"
70 return a != c;
71 #endif
74 static void
75 do_cpuid(unsigned int ax, unsigned int *p)
77 #if 0
78 __asm__ volatile(
79 "cpuid;"
80 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
81 : "0" (ax)
83 #else
84 // code from libavcodec:
85 __asm__ volatile
86 ("mov %%"REG_b", %%"REG_S"\n\t"
87 "cpuid\n\t"
88 "xchg %%"REG_b", %%"REG_S
89 : "=a" (p[0]), "=S" (p[1]),
90 "=c" (p[2]), "=d" (p[3])
91 : "0" (ax));
92 #endif
96 void GetCpuCaps( CpuCaps *caps)
98 unsigned int regs[4];
99 unsigned int regs2[4];
101 memset(caps, 0, sizeof(*caps));
102 caps->isX86=1;
103 caps->cl_size=32; /* default */
104 if (!has_cpuid()) {
105 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"CPUID not supported!??? (maybe an old 486?)\n");
106 return;
108 do_cpuid(0x00000000, regs); // get _max_ cpuid level and vendor name
109 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU vendor name: %.4s%.4s%.4s max cpuid level: %d\n",
110 (char*) (regs+1),(char*) (regs+3),(char*) (regs+2), regs[0]);
111 if (regs[0]>=0x00000001)
113 char *tmpstr, *ptmpstr;
114 unsigned cl_size;
116 do_cpuid(0x00000001, regs2);
118 caps->cpuType=(regs2[0] >> 8)&0xf;
119 caps->cpuModel=(regs2[0] >> 4)&0xf;
121 // see AMD64 Architecture Programmer's Manual, Volume 3: General-purpose and
122 // System Instructions, Table 3-2: Effective family computation, page 120.
123 if(caps->cpuType==0xf){
124 // use extended family (P4, IA64, K8)
125 caps->cpuType=0xf+((regs2[0]>>20)&255);
127 if(caps->cpuType==0xf || caps->cpuType==6)
128 caps->cpuModel |= ((regs2[0]>>16)&0xf) << 4;
130 caps->cpuStepping=regs2[0] & 0xf;
132 // general feature flags:
133 caps->hasTSC = (regs2[3] & (1 << 8 )) >> 8; // 0x0000010
134 caps->hasMMX = (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
135 caps->hasSSE = (regs2[3] & (1 << 25 )) >> 25; // 0x2000000
136 caps->hasSSE2 = (regs2[3] & (1 << 26 )) >> 26; // 0x4000000
137 caps->hasSSE3 = (regs2[2] & 1); // 0x0000001
138 caps->hasSSSE3 = (regs2[2] & (1 << 9 )) >> 9; // 0x0000200
139 caps->hasMMX2 = caps->hasSSE; // SSE cpus supports mmxext too
140 cl_size = ((regs2[1] >> 8) & 0xFF)*8;
141 if(cl_size) caps->cl_size = cl_size;
143 ptmpstr=tmpstr=GetCpuFriendlyName(regs, regs2);
144 while(*ptmpstr == ' ') // strip leading spaces
145 ptmpstr++;
146 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: %s ", ptmpstr);
147 free(tmpstr);
148 mp_msg(MSGT_CPUDETECT,MSGL_V,"(Family: %d, Model: %d, Stepping: %d)\n",
149 caps->cpuType, caps->cpuModel, caps->cpuStepping);
152 do_cpuid(0x80000000, regs);
153 if (regs[0]>=0x80000001) {
154 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cpuid-level: %d\n",regs[0]&0x7FFFFFFF);
155 do_cpuid(0x80000001, regs2);
156 caps->hasMMX |= (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
157 caps->hasMMX2 |= (regs2[3] & (1 << 22 )) >> 22; // 0x400000
158 caps->has3DNow = (regs2[3] & (1 << 31 )) >> 31; //0x80000000
159 caps->has3DNowExt = (regs2[3] & (1 << 30 )) >> 30;
160 caps->hasSSE4a = (regs2[2] & (1 << 6 )) >> 6; // 0x0000040
162 if(regs[0]>=0x80000006)
164 do_cpuid(0x80000006, regs2);
165 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cache-info: %d\n",regs2[2]&0x7FFFFFFF);
166 caps->cl_size = regs2[2] & 0xFF;
168 mp_msg(MSGT_CPUDETECT,MSGL_V,"Detected cache-line size is %u bytes\n",caps->cl_size);
169 #if 0
170 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"cpudetect: MMX=%d MMX2=%d SSE=%d SSE2=%d 3DNow=%d 3DNowExt=%d\n",
171 gCpuCaps.hasMMX,
172 gCpuCaps.hasMMX2,
173 gCpuCaps.hasSSE,
174 gCpuCaps.hasSSE2,
175 gCpuCaps.has3DNow,
176 gCpuCaps.has3DNowExt );
177 #endif
179 /* FIXME: Does SSE2 need more OS support, too? */
180 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__) \
181 || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) \
182 || defined(__APPLE__) || defined(__CYGWIN__) || defined(__MINGW32__) \
183 || defined(__OS2__)
184 if (caps->hasSSE)
185 check_os_katmai_support();
186 if (!caps->hasSSE)
187 caps->hasSSE2 = 0;
188 #else
189 caps->hasSSE=0;
190 caps->hasSSE2 = 0;
191 #endif
192 // caps->has3DNow=1;
193 // caps->hasMMX2 = 0;
194 // caps->hasMMX = 0;
196 #if !CONFIG_RUNTIME_CPUDETECT
197 #if !HAVE_MMX
198 if(caps->hasMMX) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX supported but disabled\n");
199 caps->hasMMX=0;
200 #endif
201 #if !HAVE_MMX2
202 if(caps->hasMMX2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX2 supported but disabled\n");
203 caps->hasMMX2=0;
204 #endif
205 #if !HAVE_SSE
206 if(caps->hasSSE) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE supported but disabled\n");
207 caps->hasSSE=0;
208 #endif
209 #if !HAVE_SSE2
210 if(caps->hasSSE2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE2 supported but disabled\n");
211 caps->hasSSE2=0;
212 #endif
213 #if !HAVE_AMD3DNOW
214 if(caps->has3DNow) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNow supported but disabled\n");
215 caps->has3DNow=0;
216 #endif
217 #if !HAVE_AMD3DNOWEXT
218 if(caps->has3DNowExt) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNowExt supported but disabled\n");
219 caps->has3DNowExt=0;
220 #endif
221 #endif // CONFIG_RUNTIME_CPUDETECT
224 char *GetCpuFriendlyName(unsigned int regs[], unsigned int regs2[]){
225 char vendor[13];
226 char *retname;
227 int i;
229 if (NULL==(retname=malloc(256))) {
230 mp_msg(MSGT_CPUDETECT,MSGL_FATAL,"Error: GetCpuFriendlyName() not enough memory\n");
231 exit(1);
233 retname[0] = '\0';
235 sprintf(vendor,"%.4s%.4s%.4s",(char*)(regs+1),(char*)(regs+3),(char*)(regs+2));
237 do_cpuid(0x80000000,regs);
238 if (regs[0] >= 0x80000004)
240 // CPU has built-in namestring
241 for (i = 0x80000002; i <= 0x80000004; i++)
243 do_cpuid(i, regs);
244 strncat(retname, (char*)regs, 16);
247 return retname;
250 #if defined(__linux__) && defined(_POSIX_SOURCE) && !ARCH_X86_64
251 static void sigill_handler_sse( int signal, struct sigcontext sc )
253 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
255 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
256 * instructions are 3 bytes long. We must increment the instruction
257 * pointer manually to avoid repeated execution of the offending
258 * instruction.
260 * If the SIGILL is caused by a divide-by-zero when unmasked
261 * exceptions aren't supported, the SIMD FPU status and control
262 * word will be restored at the end of the test, so we don't need
263 * to worry about doing it here. Besides, we may not be able to...
265 sc.eip += 3;
267 gCpuCaps.hasSSE=0;
269 #endif /* __linux__ && _POSIX_SOURCE */
271 #if (defined(__MINGW32__) || defined(__CYGWIN__)) && !ARCH_X86_64
272 LONG CALLBACK win32_sig_handler_sse(EXCEPTION_POINTERS* ep)
274 if(ep->ExceptionRecord->ExceptionCode==EXCEPTION_ILLEGAL_INSTRUCTION){
275 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
276 ep->ContextRecord->Eip +=3;
277 gCpuCaps.hasSSE=0;
278 return EXCEPTION_CONTINUE_EXECUTION;
280 return EXCEPTION_CONTINUE_SEARCH;
282 #endif /* defined(__MINGW32__) || defined(__CYGWIN__) */
284 #ifdef __OS2__
285 ULONG _System os2_sig_handler_sse( PEXCEPTIONREPORTRECORD p1,
286 PEXCEPTIONREGISTRATIONRECORD p2,
287 PCONTEXTRECORD p3,
288 PVOID p4 )
290 if(p1->ExceptionNum == XCPT_ILLEGAL_INSTRUCTION){
291 mp_msg(MSGT_CPUDETECT, MSGL_V, "SIGILL, ");
293 p3->ctx_RegEip += 3;
294 gCpuCaps.hasSSE = 0;
296 return XCPT_CONTINUE_EXECUTION;
298 return XCPT_CONTINUE_SEARCH;
300 #endif
302 /* If we're running on a processor that can do SSE, let's see if we
303 * are allowed to or not. This will catch 2.4.0 or later kernels that
304 * haven't been configured for a Pentium III but are running on one,
305 * and RedHat patched 2.2 kernels that have broken exception handling
306 * support for user space apps that do SSE.
309 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__)
310 #define SSE_SYSCTL_NAME "hw.instruction_sse"
311 #elif defined(__APPLE__)
312 #define SSE_SYSCTL_NAME "hw.optional.sse"
313 #endif
315 static void check_os_katmai_support( void )
317 #if ARCH_X86_64
318 gCpuCaps.hasSSE=1;
319 gCpuCaps.hasSSE2=1;
320 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
321 int has_sse=0, ret;
322 size_t len=sizeof(has_sse);
324 ret = sysctlbyname(SSE_SYSCTL_NAME, &has_sse, &len, NULL, 0);
325 if (ret || !has_sse)
326 gCpuCaps.hasSSE=0;
328 #elif defined(__NetBSD__) || defined (__OpenBSD__)
329 #if __NetBSD_Version__ >= 105250000 || (defined __OpenBSD__)
330 int has_sse, has_sse2, ret, mib[2];
331 size_t varlen;
333 mib[0] = CTL_MACHDEP;
334 mib[1] = CPU_SSE;
335 varlen = sizeof(has_sse);
337 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
338 ret = sysctl(mib, 2, &has_sse, &varlen, NULL, 0);
339 gCpuCaps.hasSSE = ret >= 0 && has_sse;
340 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
342 mib[1] = CPU_SSE2;
343 varlen = sizeof(has_sse2);
344 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE2... " );
345 ret = sysctl(mib, 2, &has_sse2, &varlen, NULL, 0);
346 gCpuCaps.hasSSE2 = ret >= 0 && has_sse2;
347 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE2 ? "yes.\n" : "no!\n" );
348 #else
349 gCpuCaps.hasSSE = 0;
350 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "No OS support for SSE, disabling to be safe.\n" );
351 #endif
352 #elif defined(__MINGW32__) || defined(__CYGWIN__)
353 LPTOP_LEVEL_EXCEPTION_FILTER exc_fil;
354 if ( gCpuCaps.hasSSE ) {
355 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
356 exc_fil = SetUnhandledExceptionFilter(win32_sig_handler_sse);
357 __asm__ volatile ("xorps %xmm0, %xmm0");
358 SetUnhandledExceptionFilter(exc_fil);
359 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
361 #elif defined(__OS2__)
362 EXCEPTIONREGISTRATIONRECORD RegRec = { 0, &os2_sig_handler_sse };
363 if ( gCpuCaps.hasSSE ) {
364 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
365 DosSetExceptionHandler( &RegRec );
366 __asm__ volatile ("xorps %xmm0, %xmm0");
367 DosUnsetExceptionHandler( &RegRec );
368 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
370 #elif defined(__linux__)
371 #if defined(_POSIX_SOURCE)
372 struct sigaction saved_sigill;
374 /* Save the original signal handlers.
376 sigaction( SIGILL, NULL, &saved_sigill );
378 signal( SIGILL, (void (*)(int))sigill_handler_sse );
380 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
381 * supports the extended FPU save and restore required for SSE. If
382 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
383 * doesn't support Streaming SIMD Exceptions, even if the processor
384 * does.
386 if ( gCpuCaps.hasSSE ) {
387 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
389 // __asm__ volatile ("xorps %%xmm0, %%xmm0");
390 __asm__ volatile ("xorps %xmm0, %xmm0");
392 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
395 /* Restore the original signal handlers.
397 sigaction( SIGILL, &saved_sigill, NULL );
399 /* If we've gotten to here and the XMM CPUID bit is still set, we're
400 * safe to go ahead and hook out the SSE code throughout Mesa.
402 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE %s\n", gCpuCaps.hasSSE ? "passed." : "failed!" );
403 #else
404 /* We can't use POSIX signal handling to test the availability of
405 * SSE, so we disable it by default.
407 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, disabling to be safe.\n" );
408 gCpuCaps.hasSSE=0;
409 #endif /* _POSIX_SOURCE */
410 #else
411 /* Do nothing on other platforms for now.
413 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, leaving disabled.\n" );
414 gCpuCaps.hasSSE=0;
415 #endif /* __linux__ */
417 #else /* ARCH_X86 */
419 #ifdef __APPLE__
420 #include <sys/sysctl.h>
421 #elif defined(__AMIGAOS4__)
422 /* nothing */
423 #else
424 #include <signal.h>
425 #include <setjmp.h>
427 static sigjmp_buf jmpbuf;
428 static volatile sig_atomic_t canjump = 0;
430 static void sigill_handler (int sig)
432 if (!canjump) {
433 signal (sig, SIG_DFL);
434 raise (sig);
437 canjump = 0;
438 siglongjmp (jmpbuf, 1);
440 #endif /* __APPLE__ */
442 void GetCpuCaps( CpuCaps *caps)
444 caps->cpuType=0;
445 caps->cpuModel=0;
446 caps->cpuStepping=0;
447 caps->hasMMX=0;
448 caps->hasMMX2=0;
449 caps->has3DNow=0;
450 caps->has3DNowExt=0;
451 caps->hasSSE=0;
452 caps->hasSSE2=0;
453 caps->hasSSE3=0;
454 caps->hasSSSE3=0;
455 caps->hasSSE4a=0;
456 caps->isX86=0;
457 caps->hasAltiVec = 0;
458 #if HAVE_ALTIVEC
459 #ifdef __APPLE__
461 rip-off from ffmpeg altivec detection code.
462 this code also appears on Apple's AltiVec pages.
465 int sels[2] = {CTL_HW, HW_VECTORUNIT};
466 int has_vu = 0;
467 size_t len = sizeof(has_vu);
468 int err;
470 err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
472 if (err == 0)
473 if (has_vu != 0)
474 caps->hasAltiVec = 1;
476 #elif defined(__AMIGAOS4__)
477 ULONG result = 0;
479 GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
480 if (result == VECTORTYPE_ALTIVEC)
481 caps->hasAltiVec = 1;
482 #else
483 /* no Darwin, do it the brute-force way */
484 /* this is borrowed from the libmpeg2 library */
486 signal (SIGILL, sigill_handler);
487 if (sigsetjmp (jmpbuf, 1)) {
488 signal (SIGILL, SIG_DFL);
489 } else {
490 canjump = 1;
492 __asm__ volatile ("mtspr 256, %0\n\t"
493 "vand %%v0, %%v0, %%v0"
495 : "r" (-1));
497 signal (SIGILL, SIG_DFL);
498 caps->hasAltiVec = 1;
501 #endif /* __APPLE__ */
502 mp_msg(MSGT_CPUDETECT,MSGL_V,"AltiVec %sfound\n", (caps->hasAltiVec ? "" : "not "));
503 #endif /* HAVE_ALTIVEC */
505 if (ARCH_IA64)
506 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Intel Itanium\n");
508 if (ARCH_SPARC)
509 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Sun Sparc\n");
511 if (ARCH_ARM)
512 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: ARM\n");
514 if (ARCH_PPC)
515 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: PowerPC\n");
517 if (ARCH_ALPHA)
518 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Digital Alpha\n");
520 if (ARCH_SGI_MIPS)
521 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: SGI MIPS\n");
523 if (ARCH_PA_RISC)
524 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Hewlett-Packard PA-RISC\n");
526 if (ARCH_S390)
527 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: IBM S/390\n");
529 if (ARCH_S390X)
530 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: IBM S/390X\n");
532 if (ARCH_VAX)
533 mp_msg(MSGT_CPUDETECT,MSGL_V, "CPU: Digital VAX\n" );
535 if (ARCH_XTENSA)
536 mp_msg(MSGT_CPUDETECT,MSGL_V, "CPU: Tensilica Xtensa\n" );
538 #endif /* !ARCH_X86 */