Signal handler messages moved to help_mp, so they can be translated
[mplayer/greg.git] / cpudetect.c
blobffc56d8f74d7a18930fcbf085986f87179d66948
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 //#define X86_FXSR_MAGIC
33 /* Thanks to the FreeBSD project for some of this cpuid code, and
34 * help understanding how to use it. Thanks to the Mesa
35 * team for SSE support detection and more cpu detect code.
38 /* I believe this code works. However, it has only been used on a PII and PIII */
40 static void check_os_katmai_support( void );
42 #if 1
43 // return TRUE if cpuid supported
44 static int has_cpuid()
46 int a, c;
48 // code from libavcodec:
49 __asm__ __volatile__ (
50 /* See if CPUID instruction is supported ... */
51 /* ... Get copies of EFLAGS into eax and ecx */
52 "pushf\n\t"
53 "popl %0\n\t"
54 "movl %0, %1\n\t"
56 /* ... Toggle the ID bit in one copy and store */
57 /* to the EFLAGS reg */
58 "xorl $0x200000, %0\n\t"
59 "push %0\n\t"
60 "popf\n\t"
62 /* ... Get the (hopefully modified) EFLAGS */
63 "pushf\n\t"
64 "popl %0\n\t"
65 : "=a" (a), "=c" (c)
67 : "cc"
70 return (a!=c);
72 #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 ("movl %%ebx, %%esi\n\t"
87 "cpuid\n\t"
88 "xchgl %%ebx, %%esi"
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;
114 unsigned cl_size;
116 do_cpuid(0x00000001, regs2);
118 tmpstr=GetCpuFriendlyName(regs, regs2);
119 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"CPU: %s ",tmpstr);
120 free(tmpstr);
122 caps->cpuType=(regs2[0] >> 8)&0xf;
123 if(caps->cpuType==0xf){
124 // use extended family (P4, IA64)
125 caps->cpuType=8+((regs2[0]>>20)&255);
127 caps->cpuStepping=regs2[0] & 0xf;
128 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"(Family: %d, Stepping: %d)\n",
129 caps->cpuType, caps->cpuStepping);
131 // general feature flags:
132 caps->hasMMX = (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
133 caps->hasSSE = (regs2[3] & (1 << 25 )) >> 25; // 0x2000000
134 caps->hasSSE2 = (regs2[3] & (1 << 26 )) >> 26; // 0x4000000
135 caps->hasMMX2 = caps->hasSSE; // SSE cpus supports mmxext too
136 cl_size = ((regs2[1] >> 8) & 0xFF)*8;
137 if(cl_size) caps->cl_size = cl_size;
139 do_cpuid(0x80000000, regs);
140 if (regs[0]>=0x80000001) {
141 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cpuid-level: %d\n",regs[0]&0x7FFFFFFF);
142 do_cpuid(0x80000001, regs2);
143 caps->hasMMX |= (regs2[3] & (1 << 23 )) >> 23; // 0x0800000
144 caps->hasMMX2 |= (regs2[3] & (1 << 22 )) >> 22; // 0x400000
145 caps->has3DNow = (regs2[3] & (1 << 31 )) >> 31; //0x80000000
146 caps->has3DNowExt = (regs2[3] & (1 << 30 )) >> 30;
148 if(regs[0]>=0x80000006)
150 do_cpuid(0x80000006, regs2);
151 mp_msg(MSGT_CPUDETECT,MSGL_V,"extended cache-info: %d\n",regs2[2]&0x7FFFFFFF);
152 caps->cl_size = regs2[2] & 0xFF;
154 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"Detected cache-line size is %u bytes\n",caps->cl_size);
155 #if 0
156 mp_msg(MSGT_CPUDETECT,MSGL_INFO,"cpudetect: MMX=%d MMX2=%d SSE=%d SSE2=%d 3DNow=%d 3DNowExt=%d\n",
157 gCpuCaps.hasMMX,
158 gCpuCaps.hasMMX2,
159 gCpuCaps.hasSSE,
160 gCpuCaps.hasSSE2,
161 gCpuCaps.has3DNow,
162 gCpuCaps.has3DNowExt );
163 #endif
165 /* FIXME: Does SSE2 need more OS support, too? */
166 #if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__)
167 if (caps->hasSSE)
168 check_os_katmai_support();
169 if (!caps->hasSSE)
170 caps->hasSSE2 = 0;
171 #else
172 caps->hasSSE=0;
173 caps->hasSSE2 = 0;
174 #endif
175 // caps->has3DNow=1;
176 // caps->hasMMX2 = 0;
177 // caps->hasMMX = 0;
179 #ifndef HAVE_MMX
180 if(caps->hasMMX) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX supported but disabled\n");
181 caps->hasMMX=0;
182 #endif
183 #ifndef HAVE_MMX2
184 if(caps->hasMMX2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"MMX2 supported but disabled\n");
185 caps->hasMMX2=0;
186 #endif
187 #ifndef HAVE_SSE
188 if(caps->hasSSE) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE supported but disabled\n");
189 caps->hasSSE=0;
190 #endif
191 #ifndef HAVE_SSE2
192 if(caps->hasSSE2) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"SSE2 supported but disabled\n");
193 caps->hasSSE2=0;
194 #endif
195 #ifndef HAVE_3DNOW
196 if(caps->has3DNow) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNow supported but disabled\n");
197 caps->has3DNow=0;
198 #endif
199 #ifndef HAVE_3DNOWEX
200 if(caps->has3DNowExt) mp_msg(MSGT_CPUDETECT,MSGL_WARN,"3DNowExt supported but disabled\n");
201 caps->has3DNowExt=0;
202 #endif
206 #define CPUID_EXTFAMILY ((regs2[0] >> 20)&0xFF) /* 27..20 */
207 #define CPUID_EXTMODEL ((regs2[0] >> 16)&0x0F) /* 19..16 */
208 #define CPUID_TYPE ((regs2[0] >> 12)&0x04) /* 13..12 */
209 #define CPUID_FAMILY ((regs2[0] >> 8)&0x0F) /* 11..08 */
210 #define CPUID_MODEL ((regs2[0] >> 4)&0x0F) /* 07..04 */
211 #define CPUID_STEPPING ((regs2[0] >> 0)&0x0F) /* 03..00 */
213 char *GetCpuFriendlyName(unsigned int regs[], unsigned int regs2[]){
214 #include "cputable.h" /* get cpuname and cpuvendors */
215 char vendor[17];
216 char *retname;
217 int i;
219 if (NULL==(retname=(char*)malloc(256))) {
220 mp_msg(MSGT_CPUDETECT,MSGL_FATAL,"Error: GetCpuFriendlyName() not enough memory\n");
221 exit(1);
224 sprintf(vendor,"%.4s%.4s%.4s",(char*)(regs+1),(char*)(regs+3),(char*)(regs+2));
226 for(i=0; i<MAX_VENDORS; i++){
227 if(!strcmp(cpuvendors[i].string,vendor)){
228 if(cpuname[i][CPUID_FAMILY][CPUID_MODEL]){
229 snprintf(retname,255,"%s %s",cpuvendors[i].name,cpuname[i][CPUID_FAMILY][CPUID_MODEL]);
230 } else {
231 snprintf(retname,255,"unknown %s %d. Generation CPU",cpuvendors[i].name,CPUID_FAMILY);
232 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"unknown %s CPU:\n",cpuvendors[i].name);
233 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Vendor: %s\n",cpuvendors[i].string);
234 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Type: %d\n",CPUID_TYPE);
235 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Family: %d (ext: %d)\n",CPUID_FAMILY,CPUID_EXTFAMILY);
236 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Model: %d (ext: %d)\n",CPUID_MODEL,CPUID_EXTMODEL);
237 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Stepping: %d\n",CPUID_STEPPING);
238 mp_msg(MSGT_CPUDETECT,MSGL_WARN,"Please send the above info along with the exact CPU name"
239 "to the MPlayer-Developers, so we can add it to the list!\n");
244 //printf("Detected CPU: %s\n", retname);
245 return retname;
248 #undef CPUID_EXTFAMILY
249 #undef CPUID_EXTMODEL
250 #undef CPUID_TYPE
251 #undef CPUID_FAMILY
252 #undef CPUID_MODEL
253 #undef CPUID_STEPPING
256 #if defined(__linux__) && defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
257 static void sigill_handler_sse( int signal, struct sigcontext sc )
259 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGILL, " );
261 /* Both the "xorps %%xmm0,%%xmm0" and "divps %xmm0,%%xmm1"
262 * instructions are 3 bytes long. We must increment the instruction
263 * pointer manually to avoid repeated execution of the offending
264 * instruction.
266 * If the SIGILL is caused by a divide-by-zero when unmasked
267 * exceptions aren't supported, the SIMD FPU status and control
268 * word will be restored at the end of the test, so we don't need
269 * to worry about doing it here. Besides, we may not be able to...
271 sc.eip += 3;
273 gCpuCaps.hasSSE=0;
276 static void sigfpe_handler_sse( int signal, struct sigcontext sc )
278 mp_msg(MSGT_CPUDETECT,MSGL_V, "SIGFPE, " );
280 if ( sc.fpstate->magic != 0xffff ) {
281 /* Our signal context has the extended FPU state, so reset the
282 * divide-by-zero exception mask and clear the divide-by-zero
283 * exception bit.
285 sc.fpstate->mxcsr |= 0x00000200;
286 sc.fpstate->mxcsr &= 0xfffffffb;
287 } else {
288 /* If we ever get here, we're completely hosed.
290 mp_msg(MSGT_CPUDETECT,MSGL_V, "\n\n" );
291 mp_msg(MSGT_CPUDETECT,MSGL_V, "SSE enabling test failed badly!" );
294 #endif /* __linux__ && _POSIX_SOURCE && X86_FXSR_MAGIC */
296 /* If we're running on a processor that can do SSE, let's see if we
297 * are allowed to or not. This will catch 2.4.0 or later kernels that
298 * haven't been configured for a Pentium III but are running on one,
299 * and RedHat patched 2.2 kernels that have broken exception handling
300 * support for user space apps that do SSE.
302 static void check_os_katmai_support( void )
304 #if defined(__FreeBSD__)
305 int has_sse=0, ret;
306 size_t len=sizeof(has_sse);
308 ret = sysctlbyname("hw.instruction_sse", &has_sse, &len, NULL, 0);
309 if (ret || !has_sse)
310 gCpuCaps.hasSSE=0;
312 #elif defined(__NetBSD__)
313 #if __NetBSD_Version__ >= 105250000
314 int has_sse, has_sse2, ret, mib[2];
315 size_t varlen;
317 mib[0] = CTL_MACHDEP;
318 mib[1] = CPU_SSE;
319 varlen = sizeof(has_sse);
321 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
322 ret = sysctl(mib, 2, &has_sse, &varlen, NULL, 0);
323 if (ret < 0 || !has_sse) {
324 gCpuCaps.hasSSE=0;
325 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
326 } else {
327 gCpuCaps.hasSSE=1;
328 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes!\n" );
331 mib[1] = CPU_SSE2;
332 varlen = sizeof(has_sse2);
333 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE2... " );
334 ret = sysctl(mib, 2, &has_sse2, &varlen, NULL, 0);
335 if (ret < 0 || !has_sse2) {
336 gCpuCaps.hasSSE2=0;
337 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
338 } else {
339 gCpuCaps.hasSSE2=1;
340 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes!\n" );
342 #else
343 gCpuCaps.hasSSE = 0;
344 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "No OS support for SSE, disabling to be safe.\n" );
345 #endif
346 #elif defined(__linux__)
347 #if defined(_POSIX_SOURCE) && defined(X86_FXSR_MAGIC)
348 struct sigaction saved_sigill;
349 struct sigaction saved_sigfpe;
351 /* Save the original signal handlers.
353 sigaction( SIGILL, NULL, &saved_sigill );
354 sigaction( SIGFPE, NULL, &saved_sigfpe );
356 signal( SIGILL, (void (*)(int))sigill_handler_sse );
357 signal( SIGFPE, (void (*)(int))sigfpe_handler_sse );
359 /* Emulate test for OSFXSR in CR4. The OS will set this bit if it
360 * supports the extended FPU save and restore required for SSE. If
361 * we execute an SSE instruction on a PIII and get a SIGILL, the OS
362 * doesn't support Streaming SIMD Exceptions, even if the processor
363 * does.
365 if ( gCpuCaps.hasSSE ) {
366 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE... " );
368 // __asm __volatile ("xorps %%xmm0, %%xmm0");
369 __asm __volatile ("xorps %xmm0, %xmm0");
371 if ( gCpuCaps.hasSSE ) {
372 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes.\n" );
373 } else {
374 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
378 /* Emulate test for OSXMMEXCPT in CR4. The OS will set this bit if
379 * it supports unmasked SIMD FPU exceptions. If we unmask the
380 * exceptions, do a SIMD divide-by-zero and get a SIGILL, the OS
381 * doesn't support unmasked SIMD FPU exceptions. If we get a SIGFPE
382 * as expected, we're okay but we need to clean up after it.
384 * Are we being too stringent in our requirement that the OS support
385 * unmasked exceptions? Certain RedHat 2.2 kernels enable SSE by
386 * setting CR4.OSFXSR but don't support unmasked exceptions. Win98
387 * doesn't even support them. We at least know the user-space SSE
388 * support is good in kernels that do support unmasked exceptions,
389 * and therefore to be safe I'm going to leave this test in here.
391 if ( gCpuCaps.hasSSE ) {
392 mp_msg(MSGT_CPUDETECT,MSGL_V, "Testing OS support for SSE unmasked exceptions... " );
394 // test_os_katmai_exception_support();
396 if ( gCpuCaps.hasSSE ) {
397 mp_msg(MSGT_CPUDETECT,MSGL_V, "yes.\n" );
398 } else {
399 mp_msg(MSGT_CPUDETECT,MSGL_V, "no!\n" );
403 /* Restore the original signal handlers.
405 sigaction( SIGILL, &saved_sigill, NULL );
406 sigaction( SIGFPE, &saved_sigfpe, NULL );
408 /* If we've gotten to here and the XMM CPUID bit is still set, we're
409 * safe to go ahead and hook out the SSE code throughout Mesa.
411 if ( gCpuCaps.hasSSE ) {
412 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE passed.\n" );
413 } else {
414 mp_msg(MSGT_CPUDETECT,MSGL_V, "Tests of OS support for SSE failed!\n" );
416 #else
417 /* We can't use POSIX signal handling to test the availability of
418 * SSE, so we disable it by default.
420 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, disabling to be safe.\n" );
421 gCpuCaps.hasSSE=0;
422 #endif /* _POSIX_SOURCE && X86_FXSR_MAGIC */
423 #else
424 /* Do nothing on other platforms for now.
426 mp_msg(MSGT_CPUDETECT,MSGL_WARN, "Cannot test OS support for SSE, leaving disabled.\n" );
427 gCpuCaps.hasSSE=0;
428 #endif /* __linux__ */
430 #else /* ARCH_X86 */
432 void GetCpuCaps( CpuCaps *caps)
434 caps->cpuType=0;
435 caps->cpuStepping=0;
436 caps->hasMMX=0;
437 caps->hasMMX2=0;
438 caps->has3DNow=0;
439 caps->has3DNowExt=0;
440 caps->hasSSE=0;
441 caps->hasSSE2=0;
442 caps->isX86=0;
444 #endif /* !ARCH_X86 */