Add P4 Northwood. (Roman Shiryaev <mih_val@mail.ru>; See <200308250434.38516.mih_val...
[mplayer/greg.git] / cpudetect.c
blobc81a96ba44eecc4f71ca349e7a4e9bdd4f6d3dd1
1 #include "config.h"
2 #include "cpudetect.h"
3 #include "mp_msg.h"
5 CpuCaps gCpuCaps;
7 #ifdef HAVE_MALLOC_H
8 #include <malloc.h>
9 #endif
10 #include <stdlib.h>
12 #ifdef ARCH_X86
14 #include <stdio.h>
15 #include <string.h>
17 #ifdef __NetBSD__
18 #include <sys/param.h>
19 #include <sys/sysctl.h>
20 #include <machine/cpu.h>
21 #endif
23 #ifdef __FreeBSD__
24 #include <sys/types.h>
25 #include <sys/sysctl.h>
26 #endif
28 #ifdef __linux__
29 #include <signal.h>
30 #endif
32 #ifdef WIN32
33 #include <windows.h>
34 #endif
36 //#define X86_FXSR_MAGIC
37 /* Thanks to the FreeBSD project for some of this cpuid code, and
38 * help understanding how to use it. Thanks to the Mesa
39 * team for SSE support detection and more cpu detect code.
42 /* I believe this code works. However, it has only been used on a PII and PIII */
44 static void check_os_katmai_support( void );
46 #if 1
47 // return TRUE if cpuid supported
48 static int has_cpuid()
50 int a, c;
52 // code from libavcodec:
53 __asm__ __volatile__ (
54 /* See if CPUID instruction is supported ... */
55 /* ... Get copies of EFLAGS into eax and ecx */
56 "pushf\n\t"
57 "popl %0\n\t"
58 "movl %0, %1\n\t"
60 /* ... Toggle the ID bit in one copy and store */
61 /* to the EFLAGS reg */
62 "xorl $0x200000, %0\n\t"
63 "push %0\n\t"
64 "popf\n\t"
66 /* ... Get the (hopefully modified) EFLAGS */
67 "pushf\n\t"
68 "popl %0\n\t"
69 : "=a" (a), "=c" (c)
71 : "cc"
74 return (a!=c);
76 #endif
78 static void
79 do_cpuid(unsigned int ax, unsigned int *p)
81 #if 0
82 __asm __volatile(
83 "cpuid;"
84 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
85 : "0" (ax)
87 #else
88 // code from libavcodec:
89 __asm __volatile
90 ("movl %%ebx, %%esi\n\t"
91 "cpuid\n\t"
92 "xchgl %%ebx, %%esi"
93 : "=a" (p[0]), "=S" (p[1]),
94 "=c" (p[2]), "=d" (p[3])
95 : "0" (ax));
96 #endif
100 void GetCpuCaps( CpuCaps *caps)
102 unsigned int regs[4];
103 unsigned int regs2[4];
105 memset(caps, 0, sizeof(*caps));
106 caps->isX86=1;
107 caps->cl_size=32; /* default */
108 if (!has_cpuid()) {
109 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"CPUID not supported!??? (maybe an old 486?)\n");
110 return;
112 do_cpuid(0x00000000, regs); // get _max_ cpuid level and vendor name
113 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU vendor name: %.4s%.4s%.4s max cpuid level: %d\n",
114 (char*) (regs+1),(char*) (regs+3),(char*) (regs+2), regs[0]);
115 if (regs[0]>=0x00000001)
117 char *tmpstr;
118 unsigned cl_size;
120 do_cpuid(0x00000001, regs2);
122 tmpstr=GetCpuFriendlyName(regs, regs2);
123 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"CPU: %s ",tmpstr);
124 free(tmpstr);
126 caps->cpuType=(regs2[0] >> 8)&0xf;
127 if(caps->cpuType==0xf){
128 // use extended family (P4, IA64)
129 caps->cpuType=8+((regs2[0]>>20)&255);
131 caps->cpuStepping=regs2[0] & 0xf;
132 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"(Family: %d, Stepping: %d)\n",
133 caps->cpuType, caps->cpuStepping);
135 // general feature flags:
136 caps->hasMMX = (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
137 caps->hasSSE = (regs2[3] & (1 << 25 )) >> 25; // 0x2000000
138 caps->hasSSE2 = (regs2[3] & (1 << 26 )) >> 26; // 0x4000000
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 do_cpuid(0x80000000, regs);
144 if (regs[0]>=0x80000001) {
145 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cpuid-level: %d\n",regs[0]&0x7FFFFFFF);
146 do_cpuid(0x80000001, regs2);
147 caps->hasMMX |= (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
148 caps->hasMMX2 |= (regs2[3] & (1 << 22 )) >> 22; // 0x400000
149 caps->has3DNow = (regs2[3] & (1 << 31 )) >> 31; //0x80000000
150 caps->has3DNowExt = (regs2[3] & (1 << 30 )) >> 30;
152 if(regs[0]>=0x80000006)
154 do_cpuid(0x80000006, regs2);
155 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cache-info: %d\n",regs2[2]&0x7FFFFFFF);
156 caps->cl_size = regs2[2] & 0xFF;
158 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"Detected cache-line size is %u bytes\n",caps->cl_size);
159 #if 0
160 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"cpudetect: MMX=%d MMX2=%d SSE=%d SSE2=%d 3DNow=%d 3DNowExt=%d\n",
161 gCpuCaps.hasMMX,
162 gCpuCaps.hasMMX2,
163 gCpuCaps.hasSSE,
164 gCpuCaps.hasSSE2,
165 gCpuCaps.has3DNow,
166 gCpuCaps.has3DNowExt );
167 #endif
169 /* FIXME: Does SSE2 need more OS support, too? */
170 #if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(WIN32)
171 if (caps->hasSSE)
172 check_os_katmai_support();
173 if (!caps->hasSSE)
174 caps->hasSSE2 = 0;
175 #else
176 caps->hasSSE=0;
177 caps->hasSSE2 = 0;
178 #endif
179 // caps->has3DNow=1;
180 // caps->hasMMX2 = 0;
181 // caps->hasMMX = 0;
183 #ifndef HAVE_MMX
184 if(caps->hasMMX) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX supported but disabled\n");
185 caps->hasMMX=0;
186 #endif
187 #ifndef HAVE_MMX2
188 if(caps->hasMMX2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX2 supported but disabled\n");
189 caps->hasMMX2=0;
190 #endif
191 #ifndef HAVE_SSE
192 if(caps->hasSSE) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE supported but disabled\n");
193 caps->hasSSE=0;
194 #endif
195 #ifndef HAVE_SSE2
196 if(caps->hasSSE2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE2 supported but disabled\n");
197 caps->hasSSE2=0;
198 #endif
199 #ifndef HAVE_3DNOW
200 if(caps->has3DNow) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNow supported but disabled\n");
201 caps->has3DNow=0;
202 #endif
203 #ifndef HAVE_3DNOWEX
204 if(caps->has3DNowExt) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNowExt supported but disabled\n");
205 caps->has3DNowExt=0;
206 #endif
210 #define CPUID_EXTFAMILY ((regs2[0] >> 20)&0xFF) /* 27..20 */
211 #define CPUID_EXTMODEL ((regs2[0] >> 16)&0x0F) /* 19..16 */
212 #define CPUID_TYPE ((regs2[0] >> 12)&0x04) /* 13..12 */
213 #define CPUID_FAMILY ((regs2[0] >> 8)&0x0F) /* 11..08 */
214 #define CPUID_MODEL ((regs2[0] >> 4)&0x0F) /* 07..04 */
215 #define CPUID_STEPPING ((regs2[0] >> 0)&0x0F) /* 03..00 */
217 char *GetCpuFriendlyName(unsigned int regs[], unsigned int regs2[]){
218 #include "cputable.h" /* get cpuname and cpuvendors */
219 char vendor[17];
220 char *retname;
221 int i;
223 if (NULL==(retname=(char*)malloc(256))) {
224 mp_msg(MSGT_CPUDETECT,MSGL_FATAL,"Error: GetCpuFriendlyName() not enough memory\n");
225 exit(1);
228 sprintf(vendor,"%.4s%.4s%.4s",(char*)(regs+1),(char*)(regs+3),(char*)(regs+2));
230 for(i=0; i<MAX_VENDORS; i++){
231 if(!strcmp(cpuvendors[i].string,vendor)){
232 if(cpuname[i][CPUID_FAMILY][CPUID_MODEL]){
233 snprintf(retname,255,"%s %s",cpuvendors[i].name,cpuname[i][CPUID_FAMILY][CPUID_MODEL]);
234 } else {
235 snprintf(retname,255,"unknown %s %d. Generation CPU",cpuvendors[i].name,CPUID_FAMILY);
236 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"unknown %s CPU:\n",cpuvendors[i].name);
237 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Vendor: %s\n",cpuvendors[i].string);
238 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Type: %d\n",CPUID_TYPE);
239 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Family: %d (ext: %d)\n",CPUID_FAMILY,CPUID_EXTFAMILY);
240 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Model: %d (ext: %d)\n",CPUID_MODEL,CPUID_EXTMODEL);
241 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Stepping: %d\n",CPUID_STEPPING);
242 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Please send the above info along with the exact CPU name"
243 "to the MPlayer-Developers, so we can add it to the list!\n");
248 //printf("Detected CPU: %s\n", retname);
249 return retname;
252 #undef CPUID_EXTFAMILY
253 #undef CPUID_EXTMODEL
254 #undef CPUID_TYPE
255 #undef CPUID_FAMILY
256 #undef CPUID_MODEL
257 #undef CPUID_STEPPING
260 #if defined(__linux__) && defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
261 static void sigill_handler_sse( int signal, struct sigcontext sc )
263 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
265 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
266 * instructions are 3 bytes long. We must increment the instruction
267 * pointer manually to avoid repeated execution of the offending
268 * instruction.
270 * If the SIGILL is caused by a divide-by-zero when unmasked
271 * exceptions aren't supported, the SIMD FPU status and control
272 * word will be restored at the end of the test, so we don't need
273 * to worry about doing it here. Besides, we may not be able to...
275 sc.eip += 3;
277 gCpuCaps.hasSSE=0;
280 static void sigfpe_handler_sse( int signal, struct sigcontext sc )
282 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGFPE, " );
284 if ( sc.fpstate->magic != 0xffff ) {
285 /* Our signal context has the extended FPU state, so reset the
286 * divide-by-zero exception mask and clear the divide-by-zero
287 * exception bit.
289 sc.fpstate->mxcsr |= 0x00000200;
290 sc.fpstate->mxcsr &= 0xfffffffb;
291 } else {
292 /* If we ever get here, we're completely hosed.
294 mp_msg(MSGT_CPUDETECT,MSGL_V, "\n\n" );
295 mp_msg(MSGT_CPUDETECT,MSGL_V, "SSE enabling test failed badly!" );
298 #endif /* __linux__ && _POSIX_SOURCE && X86_FXSR_MAGIC */
300 #ifdef WIN32
301 LONG CALLBACK win32_sig_handler_sse(EXCEPTION_POINTERS* ep)
303 if(ep->ExceptionRecord->ExceptionCode==EXCEPTION_ILLEGAL_INSTRUCTION){
304 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
305 ep->ContextRecord->Eip +=3;
306 gCpuCaps.hasSSE=0;
307 return EXCEPTION_CONTINUE_EXECUTION;
309 return EXCEPTION_CONTINUE_SEARCH;
311 #endif /* WIN32 */
313 /* If we're running on a processor that can do SSE, let's see if we
314 * are allowed to or not. This will catch 2.4.0 or later kernels that
315 * haven't been configured for a Pentium III but are running on one,
316 * and RedHat patched 2.2 kernels that have broken exception handling
317 * support for user space apps that do SSE.
319 static void check_os_katmai_support( void )
321 #if defined(__FreeBSD__)
322 int has_sse=0, ret;
323 size_t len=sizeof(has_sse);
325 ret = sysctlbyname("hw.instruction_sse", &has_sse, &len, NULL, 0);
326 if (ret || !has_sse)
327 gCpuCaps.hasSSE=0;
329 #elif defined(__NetBSD__)
330 #if __NetBSD_Version__ >= 105250000
331 int has_sse, has_sse2, ret, mib[2];
332 size_t varlen;
334 mib[0] = CTL_MACHDEP;
335 mib[1] = CPU_SSE;
336 varlen = sizeof(has_sse);
338 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
339 ret = sysctl(mib, 2, &has_sse, &varlen, NULL, 0);
340 if (ret < 0 || !has_sse) {
341 gCpuCaps.hasSSE=0;
342 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
343 } else {
344 gCpuCaps.hasSSE=1;
345 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes!\n" );
348 mib[1] = CPU_SSE2;
349 varlen = sizeof(has_sse2);
350 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE2... " );
351 ret = sysctl(mib, 2, &has_sse2, &varlen, NULL, 0);
352 if (ret < 0 || !has_sse2) {
353 gCpuCaps.hasSSE2=0;
354 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
355 } else {
356 gCpuCaps.hasSSE2=1;
357 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes!\n" );
359 #else
360 gCpuCaps.hasSSE = 0;
361 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "No OS support for SSE, disabling to be safe.\n" );
362 #endif
363 #elif defined(WIN32)
364 LPTOP_LEVEL_EXCEPTION_FILTER exc_fil;
365 if ( gCpuCaps.hasSSE ) {
366 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
367 exc_fil = SetUnhandledExceptionFilter(win32_sig_handler_sse);
368 __asm __volatile ("xorps %xmm0, %xmm0");
369 SetUnhandledExceptionFilter(exc_fil);
370 if ( gCpuCaps.hasSSE ) mp_msg(MSGT_CPUDETECT,MSGL_V, "yes.\n" );
371 else mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
373 #elif defined(__linux__)
374 #if defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
375 struct sigaction saved_sigill;
376 struct sigaction saved_sigfpe;
378 /* Save the original signal handlers.
380 sigaction( SIGILL, NULL, &saved_sigill );
381 sigaction( SIGFPE, NULL, &saved_sigfpe );
383 signal( SIGILL, (void (*)(int))sigill_handler_sse );
384 signal( SIGFPE, (void (*)(int))sigfpe_handler_sse );
386 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
387 * supports the extended FPU save and restore required for SSE. If
388 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
389 * doesn't support Streaming SIMD Exceptions, even if the processor
390 * does.
392 if ( gCpuCaps.hasSSE ) {
393 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
395 // __asm __volatile ("xorps %%xmm0, %%xmm0");
396 __asm __volatile ("xorps %xmm0, %xmm0");
398 if ( gCpuCaps.hasSSE ) {
399 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes.\n" );
400 } else {
401 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
405 /* Emulate test for OSXMMEXCPT in CR4. The OS will set this bit if
406 * it supports unmasked SIMD FPU exceptions. If we unmask the
407 * exceptions, do a SIMD divide-by-zero and get a SIGILL, the OS
408 * doesn't support unmasked SIMD FPU exceptions. If we get a SIGFPE
409 * as expected, we're okay but we need to clean up after it.
411 * Are we being too stringent in our requirement that the OS support
412 * unmasked exceptions? Certain RedHat 2.2 kernels enable SSE by
413 * setting CR4.OSFXSR but don't support unmasked exceptions. Win98
414 * doesn't even support them. We at least know the user-space SSE
415 * support is good in kernels that do support unmasked exceptions,
416 * and therefore to be safe I'm going to leave this test in here.
418 if ( gCpuCaps.hasSSE ) {
419 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE unmasked exceptions... " );
421 // test_os_katmai_exception_support();
423 if ( gCpuCaps.hasSSE ) {
424 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes.\n" );
425 } else {
426 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
430 /* Restore the original signal handlers.
432 sigaction( SIGILL, &saved_sigill, NULL );
433 sigaction( SIGFPE, &saved_sigfpe, NULL );
435 /* If we've gotten to here and the XMM CPUID bit is still set, we're
436 * safe to go ahead and hook out the SSE code throughout Mesa.
438 if ( gCpuCaps.hasSSE ) {
439 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE passed.\n" );
440 } else {
441 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE failed!\n" );
443 #else
444 /* We can't use POSIX signal handling to test the availability of
445 * SSE, so we disable it by default.
447 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, disabling to be safe.\n" );
448 gCpuCaps.hasSSE=0;
449 #endif /* _POSIX_SOURCE && X86_FXSR_MAGIC */
450 #else
451 /* Do nothing on other platforms for now.
453 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, leaving disabled.\n" );
454 gCpuCaps.hasSSE=0;
455 #endif /* __linux__ */
457 #else /* ARCH_X86 */
459 #ifdef SYS_DARWIN
460 #include <sys/sysctl.h>
461 #else
462 #include <signal.h>
463 #include <setjmp.h>
465 static sigjmp_buf jmpbuf;
466 static volatile sig_atomic_t canjump = 0;
468 static void sigill_handler (int sig)
470 if (!canjump) {
471 signal (sig, SIG_DFL);
472 raise (sig);
475 canjump = 0;
476 siglongjmp (jmpbuf, 1);
478 #endif
480 void GetCpuCaps( CpuCaps *caps)
482 caps->cpuType=0;
483 caps->cpuStepping=0;
484 caps->hasMMX=0;
485 caps->hasMMX2=0;
486 caps->has3DNow=0;
487 caps->has3DNowExt=0;
488 caps->hasSSE=0;
489 caps->hasSSE2=0;
490 caps->isX86=0;
491 caps->hasAltiVec = 0;
492 #ifdef HAVE_ALTIVEC
493 #ifdef SYS_DARWIN
495 rip-off from ffmpeg altivec detection code.
496 this code also appears on Apple's AltiVec pages.
499 int sels[2] = {CTL_HW, HW_VECTORUNIT};
500 int has_vu = 0;
501 size_t len = sizeof(has_vu);
502 int err;
504 err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
506 if (err == 0)
507 if (has_vu != 0)
508 caps->hasAltiVec = 1;
510 #else /* SYS_DARWIN */
511 /* no Darwin, do it the brute-force way */
512 /* this is borrowed from the libmpeg2 library */
514 signal (SIGILL, sigill_handler);
515 if (sigsetjmp (jmpbuf, 1)) {
516 signal (SIGILL, SIG_DFL);
517 } else {
518 canjump = 1;
520 asm volatile ("mtspr 256, %0\n\t"
521 "vand %%v0, %%v0, %%v0"
523 : "r" (-1));
525 signal (SIGILL, SIG_DFL);
526 caps->hasAltiVec = 1;
529 #endif /* SYS_DARWIN */
530 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"AltiVec %sfound\n", (caps->hasAltiVec ? "" : "not "));
531 #endif /* HAVE_ALTIVEC */
533 #endif /* !ARCH_X86 */