The large -help help_text should be const so it goes into .rodata
[mplayer/glamo.git] / cpudetect.c
blob5da169dfc0d16d235316d0722b5d21386c7577f3
1 #include "config.h"
2 #include "cpudetect.h"
3 #include "mp_msg.h"
5 CpuCaps gCpuCaps;
7 #if HAVE_MALLOC_H
8 #include <malloc.h>
9 #endif
10 #include <stdlib.h>
12 #if ARCH_X86
14 #include <stdio.h>
15 #include <string.h>
17 #if defined (__NetBSD__) || defined(__OpenBSD__)
18 #include <sys/param.h>
19 #include <sys/sysctl.h>
20 #include <machine/cpu.h>
21 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
22 #include <sys/types.h>
23 #include <sys/sysctl.h>
24 #elif defined(__linux__)
25 #include <signal.h>
26 #elif defined(__MINGW32__) || defined(__CYGWIN__)
27 #include <windows.h>
28 #elif defined(__OS2__)
29 #define INCL_DOS
30 #include <os2.h>
31 #elif defined(__AMIGAOS4__)
32 #include <proto/exec.h>
33 #endif
35 /* Thanks to the FreeBSD project for some of this cpuid code, and
36 * help understanding how to use it. Thanks to the Mesa
37 * team for SSE support detection and more cpu detect code.
40 /* I believe this code works. However, it has only been used on a PII and PIII */
42 static void check_os_katmai_support( void );
44 // return TRUE if cpuid supported
45 static int has_cpuid(void)
47 long a, c;
49 // code from libavcodec:
50 #if ARCH_X86_64
51 #define PUSHF "pushfq\n\t"
52 #define POPF "popfq\n\t"
53 #else
54 #define PUSHF "pushfl\n\t"
55 #define POPF "popfl\n\t"
56 #endif
57 __asm__ volatile (
58 /* See if CPUID instruction is supported ... */
59 /* ... Get copies of EFLAGS into eax and ecx */
60 PUSHF
61 "pop %0\n\t"
62 "mov %0, %1\n\t"
64 /* ... Toggle the ID bit in one copy and store */
65 /* to the EFLAGS reg */
66 "xor $0x200000, %0\n\t"
67 "push %0\n\t"
68 POPF
70 /* ... Get the (hopefully modified) EFLAGS */
71 PUSHF
72 "pop %0\n\t"
73 : "=a" (a), "=c" (c)
75 : "cc"
77 #undef PUSHF
78 #undef POPF
80 return a != c;
83 static void
84 do_cpuid(unsigned int ax, unsigned int *p)
86 #if 0
87 __asm__ volatile(
88 "cpuid;"
89 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
90 : "0" (ax)
92 #else
93 // code from libavcodec:
94 __asm__ volatile
95 ("mov %%"REG_b", %%"REG_S"\n\t"
96 "cpuid\n\t"
97 "xchg %%"REG_b", %%"REG_S
98 : "=a" (p[0]), "=S" (p[1]),
99 "=c" (p[2]), "=d" (p[3])
100 : "0" (ax));
101 #endif
105 void GetCpuCaps( CpuCaps *caps)
107 unsigned int regs[4];
108 unsigned int regs2[4];
110 memset(caps, 0, sizeof(*caps));
111 caps->isX86=1;
112 caps->cl_size=32; /* default */
113 if (!has_cpuid()) {
114 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"CPUID not supported!??? (maybe an old 486?)\n");
115 return;
117 do_cpuid(0x00000000, regs); // get _max_ cpuid level and vendor name
118 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU vendor name: %.4s%.4s%.4s max cpuid level: %d\n",
119 (char*) (regs+1),(char*) (regs+3),(char*) (regs+2), regs[0]);
120 if (regs[0]>=0x00000001)
122 char *tmpstr, *ptmpstr;
123 unsigned cl_size;
125 do_cpuid(0x00000001, regs2);
127 caps->cpuType=(regs2[0] >> 8)&0xf;
128 caps->cpuModel=(regs2[0] >> 4)&0xf;
130 // see AMD64 Architecture Programmer's Manual, Volume 3: General-purpose and
131 // System Instructions, Table 3-2: Effective family computation, page 120.
132 if(caps->cpuType==0xf){
133 // use extended family (P4, IA64, K8)
134 caps->cpuType=0xf+((regs2[0]>>20)&255);
136 if(caps->cpuType==0xf || caps->cpuType==6)
137 caps->cpuModel |= ((regs2[0]>>16)&0xf) << 4;
139 caps->cpuStepping=regs2[0] & 0xf;
141 // general feature flags:
142 caps->hasTSC = (regs2[3] & (1 << 8 )) >> 8; // 0x0000010
143 caps->hasMMX = (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
144 caps->hasSSE = (regs2[3] & (1 << 25 )) >> 25; // 0x2000000
145 caps->hasSSE2 = (regs2[3] & (1 << 26 )) >> 26; // 0x4000000
146 caps->hasSSSE3 = (regs2[2] & (1 << 9 )) >> 9; // 0x0000200
147 caps->hasMMX2 = caps->hasSSE; // SSE cpus supports mmxext too
148 cl_size = ((regs2[1] >> 8) & 0xFF)*8;
149 if(cl_size) caps->cl_size = cl_size;
151 ptmpstr=tmpstr=GetCpuFriendlyName(regs, regs2);
152 while(*ptmpstr == ' ') // strip leading spaces
153 ptmpstr++;
154 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: %s ", ptmpstr);
155 free(tmpstr);
156 mp_msg(MSGT_CPUDETECT,MSGL_V,"(Family: %d, Model: %d, Stepping: %d)\n",
157 caps->cpuType, caps->cpuModel, caps->cpuStepping);
160 do_cpuid(0x80000000, regs);
161 if (regs[0]>=0x80000001) {
162 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cpuid-level: %d\n",regs[0]&0x7FFFFFFF);
163 do_cpuid(0x80000001, regs2);
164 caps->hasMMX |= (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
165 caps->hasMMX2 |= (regs2[3] & (1 << 22 )) >> 22; // 0x400000
166 caps->has3DNow = (regs2[3] & (1 << 31 )) >> 31; //0x80000000
167 caps->has3DNowExt = (regs2[3] & (1 << 30 )) >> 30;
168 caps->hasSSE4a = (regs2[2] & (1 << 6 )) >> 6; // 0x0000040
170 if(regs[0]>=0x80000006)
172 do_cpuid(0x80000006, regs2);
173 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cache-info: %d\n",regs2[2]&0x7FFFFFFF);
174 caps->cl_size = regs2[2] & 0xFF;
176 mp_msg(MSGT_CPUDETECT,MSGL_V,"Detected cache-line size is %u bytes\n",caps->cl_size);
177 #if 0
178 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"cpudetect: MMX=%d MMX2=%d SSE=%d SSE2=%d 3DNow=%d 3DNowExt=%d\n",
179 gCpuCaps.hasMMX,
180 gCpuCaps.hasMMX2,
181 gCpuCaps.hasSSE,
182 gCpuCaps.hasSSE2,
183 gCpuCaps.has3DNow,
184 gCpuCaps.has3DNowExt );
185 #endif
187 /* FIXME: Does SSE2 need more OS support, too? */
188 #if defined(__linux__) || defined(__FreeBSD__) || defined(__FreeBSD_kernel__) \
189 || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) \
190 || defined(__APPLE__) || defined(__CYGWIN__) || defined(__MINGW32__) \
191 || defined(__OS2__)
192 if (caps->hasSSE)
193 check_os_katmai_support();
194 if (!caps->hasSSE)
195 caps->hasSSE2 = 0;
196 #else
197 caps->hasSSE=0;
198 caps->hasSSE2 = 0;
199 #endif
200 // caps->has3DNow=1;
201 // caps->hasMMX2 = 0;
202 // caps->hasMMX = 0;
204 #ifndef RUNTIME_CPUDETECT
205 #if !HAVE_MMX
206 if(caps->hasMMX) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX supported but disabled\n");
207 caps->hasMMX=0;
208 #endif
209 #if !HAVE_MMX2
210 if(caps->hasMMX2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX2 supported but disabled\n");
211 caps->hasMMX2=0;
212 #endif
213 #if !HAVE_SSE
214 if(caps->hasSSE) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE supported but disabled\n");
215 caps->hasSSE=0;
216 #endif
217 #if !HAVE_SSE2
218 if(caps->hasSSE2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE2 supported but disabled\n");
219 caps->hasSSE2=0;
220 #endif
221 #if !HAVE_AMD3DNOW
222 if(caps->has3DNow) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNow supported but disabled\n");
223 caps->has3DNow=0;
224 #endif
225 #if !HAVE_AMD3DNOWEXT
226 if(caps->has3DNowExt) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNowExt supported but disabled\n");
227 caps->has3DNowExt=0;
228 #endif
229 #endif // RUNTIME_CPUDETECT
232 char *GetCpuFriendlyName(unsigned int regs[], unsigned int regs2[]){
233 char vendor[13];
234 char *retname;
235 int i;
237 if (NULL==(retname=malloc(256))) {
238 mp_msg(MSGT_CPUDETECT,MSGL_FATAL,"Error: GetCpuFriendlyName() not enough memory\n");
239 exit(1);
241 retname[0] = '\0';
243 sprintf(vendor,"%.4s%.4s%.4s",(char*)(regs+1),(char*)(regs+3),(char*)(regs+2));
245 do_cpuid(0x80000000,regs);
246 if (regs[0] >= 0x80000004)
248 // CPU has built-in namestring
249 for (i = 0x80000002; i <= 0x80000004; i++)
251 do_cpuid(i, regs);
252 strncat(retname, (char*)regs, 16);
255 return retname;
258 #if defined(__linux__) && defined(_POSIX_SOURCE) && !ARCH_X86_64
259 static void sigill_handler_sse( int signal, struct sigcontext sc )
261 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
263 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
264 * instructions are 3 bytes long. We must increment the instruction
265 * pointer manually to avoid repeated execution of the offending
266 * instruction.
268 * If the SIGILL is caused by a divide-by-zero when unmasked
269 * exceptions aren't supported, the SIMD FPU status and control
270 * word will be restored at the end of the test, so we don't need
271 * to worry about doing it here. Besides, we may not be able to...
273 sc.eip += 3;
275 gCpuCaps.hasSSE=0;
277 #endif /* __linux__ && _POSIX_SOURCE */
279 #if defined(__MINGW32__) || defined(__CYGWIN__)
280 LONG CALLBACK win32_sig_handler_sse(EXCEPTION_POINTERS* ep)
282 if(ep->ExceptionRecord->ExceptionCode==EXCEPTION_ILLEGAL_INSTRUCTION){
283 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
284 ep->ContextRecord->Eip +=3;
285 gCpuCaps.hasSSE=0;
286 return EXCEPTION_CONTINUE_EXECUTION;
288 return EXCEPTION_CONTINUE_SEARCH;
290 #endif /* defined(__MINGW32__) || defined(__CYGWIN__) */
292 #ifdef __OS2__
293 ULONG _System os2_sig_handler_sse( PEXCEPTIONREPORTRECORD p1,
294 PEXCEPTIONREGISTRATIONRECORD p2,
295 PCONTEXTRECORD p3,
296 PVOID p4 )
298 if(p1->ExceptionNum == XCPT_ILLEGAL_INSTRUCTION){
299 mp_msg(MSGT_CPUDETECT, MSGL_V, "SIGILL, ");
301 p3->ctx_RegEip += 3;
302 gCpuCaps.hasSSE = 0;
304 return XCPT_CONTINUE_EXECUTION;
306 return XCPT_CONTINUE_SEARCH;
308 #endif
310 /* If we're running on a processor that can do SSE, let's see if we
311 * are allowed to or not. This will catch 2.4.0 or later kernels that
312 * haven't been configured for a Pentium III but are running on one,
313 * and RedHat patched 2.2 kernels that have broken exception handling
314 * support for user space apps that do SSE.
317 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__)
318 #define SSE_SYSCTL_NAME "hw.instruction_sse"
319 #elif defined(__APPLE__)
320 #define SSE_SYSCTL_NAME "hw.optional.sse"
321 #endif
323 static void check_os_katmai_support( void )
325 #if ARCH_X86_64
326 gCpuCaps.hasSSE=1;
327 gCpuCaps.hasSSE2=1;
328 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
329 int has_sse=0, ret;
330 size_t len=sizeof(has_sse);
332 ret = sysctlbyname(SSE_SYSCTL_NAME, &has_sse, &len, NULL, 0);
333 if (ret || !has_sse)
334 gCpuCaps.hasSSE=0;
336 #elif defined(__NetBSD__) || defined (__OpenBSD__)
337 #if __NetBSD_Version__ >= 105250000 || (defined __OpenBSD__)
338 int has_sse, has_sse2, ret, mib[2];
339 size_t varlen;
341 mib[0] = CTL_MACHDEP;
342 mib[1] = CPU_SSE;
343 varlen = sizeof(has_sse);
345 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
346 ret = sysctl(mib, 2, &has_sse, &varlen, NULL, 0);
347 gCpuCaps.hasSSE = ret >= 0 && has_sse;
348 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
350 mib[1] = CPU_SSE2;
351 varlen = sizeof(has_sse2);
352 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE2... " );
353 ret = sysctl(mib, 2, &has_sse2, &varlen, NULL, 0);
354 gCpuCaps.hasSSE2 = ret >= 0 && has_sse2;
355 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE2 ? "yes.\n" : "no!\n" );
356 #else
357 gCpuCaps.hasSSE = 0;
358 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "No OS support for SSE, disabling to be safe.\n" );
359 #endif
360 #elif defined(__MINGW32__) || defined(__CYGWIN__)
361 LPTOP_LEVEL_EXCEPTION_FILTER exc_fil;
362 if ( gCpuCaps.hasSSE ) {
363 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
364 exc_fil = SetUnhandledExceptionFilter(win32_sig_handler_sse);
365 __asm__ volatile ("xorps %xmm0, %xmm0");
366 SetUnhandledExceptionFilter(exc_fil);
367 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
369 #elif defined(__OS2__)
370 EXCEPTIONREGISTRATIONRECORD RegRec = { 0, &os2_sig_handler_sse };
371 if ( gCpuCaps.hasSSE ) {
372 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
373 DosSetExceptionHandler( &RegRec );
374 __asm__ volatile ("xorps %xmm0, %xmm0");
375 DosUnsetExceptionHandler( &RegRec );
376 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
378 #elif defined(__linux__)
379 #if defined(_POSIX_SOURCE)
380 struct sigaction saved_sigill;
382 /* Save the original signal handlers.
384 sigaction( SIGILL, NULL, &saved_sigill );
386 signal( SIGILL, (void (*)(int))sigill_handler_sse );
388 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
389 * supports the extended FPU save and restore required for SSE. If
390 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
391 * doesn't support Streaming SIMD Exceptions, even if the processor
392 * does.
394 if ( gCpuCaps.hasSSE ) {
395 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
397 // __asm__ volatile ("xorps %%xmm0, %%xmm0");
398 __asm__ volatile ("xorps %xmm0, %xmm0");
400 mp_msg(MSGT_CPUDETECT,MSGL_V, gCpuCaps.hasSSE ? "yes.\n" : "no!\n" );
403 /* Restore the original signal handlers.
405 sigaction( SIGILL, &saved_sigill, NULL );
407 /* If we've gotten to here and the XMM CPUID bit is still set, we're
408 * safe to go ahead and hook out the SSE code throughout Mesa.
410 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE %s\n", gCpuCaps.hasSSE ? "passed." : "failed!" );
411 #else
412 /* We can't use POSIX signal handling to test the availability of
413 * SSE, so we disable it by default.
415 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, disabling to be safe.\n" );
416 gCpuCaps.hasSSE=0;
417 #endif /* _POSIX_SOURCE */
418 #else
419 /* Do nothing on other platforms for now.
421 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, leaving disabled.\n" );
422 gCpuCaps.hasSSE=0;
423 #endif /* __linux__ */
425 #else /* ARCH_X86 */
427 #ifdef __APPLE__
428 #include <sys/sysctl.h>
429 #elif defined(__AMIGAOS4__)
430 /* nothing */
431 #else
432 #include <signal.h>
433 #include <setjmp.h>
435 static sigjmp_buf jmpbuf;
436 static volatile sig_atomic_t canjump = 0;
438 static void sigill_handler (int sig)
440 if (!canjump) {
441 signal (sig, SIG_DFL);
442 raise (sig);
445 canjump = 0;
446 siglongjmp (jmpbuf, 1);
448 #endif /* __APPLE__ */
450 void GetCpuCaps( CpuCaps *caps)
452 caps->cpuType=0;
453 caps->cpuModel=0;
454 caps->cpuStepping=0;
455 caps->hasMMX=0;
456 caps->hasMMX2=0;
457 caps->has3DNow=0;
458 caps->has3DNowExt=0;
459 caps->hasSSE=0;
460 caps->hasSSE2=0;
461 caps->hasSSSE3=0;
462 caps->hasSSE4a=0;
463 caps->isX86=0;
464 caps->hasAltiVec = 0;
465 #if HAVE_ALTIVEC
466 #ifdef __APPLE__
468 rip-off from ffmpeg altivec detection code.
469 this code also appears on Apple's AltiVec pages.
472 int sels[2] = {CTL_HW, HW_VECTORUNIT};
473 int has_vu = 0;
474 size_t len = sizeof(has_vu);
475 int err;
477 err = sysctl(sels, 2, &has_vu, &len, NULL, 0);
479 if (err == 0)
480 if (has_vu != 0)
481 caps->hasAltiVec = 1;
483 #elif defined(__AMIGAOS4__)
484 ULONG result = 0;
486 GetCPUInfoTags(GCIT_VectorUnit, &result, TAG_DONE);
487 if (result == VECTORTYPE_ALTIVEC)
488 caps->hasAltiVec = 1;
489 #else
490 /* no Darwin, do it the brute-force way */
491 /* this is borrowed from the libmpeg2 library */
493 signal (SIGILL, sigill_handler);
494 if (sigsetjmp (jmpbuf, 1)) {
495 signal (SIGILL, SIG_DFL);
496 } else {
497 canjump = 1;
499 __asm__ volatile ("mtspr 256, %0\n\t"
500 "vand %%v0, %%v0, %%v0"
502 : "r" (-1));
504 signal (SIGILL, SIG_DFL);
505 caps->hasAltiVec = 1;
508 #endif /* __APPLE__ */
509 mp_msg(MSGT_CPUDETECT,MSGL_V,"AltiVec %sfound\n", (caps->hasAltiVec ? "" : "not "));
510 #endif /* HAVE_ALTIVEC */
512 if (ARCH_IA64)
513 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Intel Itanium\n");
515 if (ARCH_SPARC)
516 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Sun Sparc\n");
518 if (ARCH_ARM)
519 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: ARM\n");
521 if (ARCH_PPC)
522 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: PowerPC\n");
524 if (ARCH_ALPHA)
525 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Digital Alpha\n");
527 if (ARCH_SGI_MIPS)
528 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: SGI MIPS\n");
530 if (ARCH_PA_RISC)
531 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: Hewlett-Packard PA-RISC\n");
533 if (ARCH_S390)
534 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: IBM S/390\n");
536 if (ARCH_S390X)
537 mp_msg(MSGT_CPUDETECT,MSGL_V,"CPU: IBM S/390X\n");
539 if (ARCH_VAX)
540 mp_msg(MSGT_CPUDETECT,MSGL_V, "CPU: Digital VAX\n" );
542 if (ARCH_XTENSA)
543 mp_msg(MSGT_CPUDETECT,MSGL_V, "CPU: Tensilica Xtensa\n" );
545 #endif /* !ARCH_X86 */