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__)
26 #elif defined(__MINGW32__) || defined(__CYGWIN__)
28 #elif defined(__OS2__)
31 #elif defined(__AMIGAOS4__)
32 #include <proto/exec.h>
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)
49 // code from libavcodec:
51 #define PUSHF "pushfq\n\t"
52 #define POPF "popfq\n\t"
54 #define PUSHF "pushfl\n\t"
55 #define POPF "popfl\n\t"
58 /* See if CPUID instruction is supported ... */
59 /* ... Get copies of EFLAGS into eax and ecx */
64 /* ... Toggle the ID bit in one copy and store */
65 /* to the EFLAGS reg */
66 "xor $0x200000, %0\n\t"
70 /* ... Get the (hopefully modified) EFLAGS */
84 do_cpuid(unsigned int ax
, unsigned int *p
)
89 : "=a" (p
[0]), "=b" (p
[1]), "=c" (p
[2]), "=d" (p
[3])
93 // code from libavcodec:
95 ("mov %%"REG_b
", %%"REG_S
"\n\t"
97 "xchg %%"REG_b
", %%"REG_S
98 : "=a" (p
[0]), "=S" (p
[1]),
99 "=c" (p
[2]), "=d" (p
[3])
105 void GetCpuCaps( CpuCaps
*caps
)
107 unsigned int regs
[4];
108 unsigned int regs2
[4];
110 memset(caps
, 0, sizeof(*caps
));
112 caps
->cl_size
=32; /* default */
114 mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"CPUID not supported!??? (maybe an old 486?)\n");
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
;
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
154 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: %s ", ptmpstr
);
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
);
178 mp_msg(MSGT_CPUDETECT
,MSGL_INFO
,"cpudetect: MMX=%d MMX2=%d SSE=%d SSE2=%d 3DNow=%d 3DNowExt=%d\n",
184 gCpuCaps
.has3DNowExt
);
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__) \
193 check_os_katmai_support();
201 // caps->hasMMX2 = 0;
204 #ifndef RUNTIME_CPUDETECT
206 if(caps
->hasMMX
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"MMX supported but disabled\n");
210 if(caps
->hasMMX2
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"MMX2 supported but disabled\n");
214 if(caps
->hasSSE
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"SSE supported but disabled\n");
218 if(caps
->hasSSE2
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"SSE2 supported but disabled\n");
222 if(caps
->has3DNow
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"3DNow supported but disabled\n");
225 #if !HAVE_AMD3DNOWEXT
226 if(caps
->has3DNowExt
) mp_msg(MSGT_CPUDETECT
,MSGL_WARN
,"3DNowExt supported but disabled\n");
229 #endif // RUNTIME_CPUDETECT
232 char *GetCpuFriendlyName(unsigned int regs
[], unsigned int regs2
[]){
237 if (NULL
==(retname
=malloc(256))) {
238 mp_msg(MSGT_CPUDETECT
,MSGL_FATAL
,"Error: GetCpuFriendlyName() not enough memory\n");
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
++)
252 strncat(retname
, (char*)regs
, 16);
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
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...
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;
286 return EXCEPTION_CONTINUE_EXECUTION
;
288 return EXCEPTION_CONTINUE_SEARCH
;
290 #endif /* defined(__MINGW32__) || defined(__CYGWIN__) */
293 ULONG _System
os2_sig_handler_sse( PEXCEPTIONREPORTRECORD p1
,
294 PEXCEPTIONREGISTRATIONRECORD p2
,
298 if(p1
->ExceptionNum
== XCPT_ILLEGAL_INSTRUCTION
){
299 mp_msg(MSGT_CPUDETECT
, MSGL_V
, "SIGILL, ");
304 return XCPT_CONTINUE_EXECUTION
;
306 return XCPT_CONTINUE_SEARCH
;
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"
323 static void check_os_katmai_support( void )
328 #elif defined(__FreeBSD__) || defined(__FreeBSD_kernel__) || defined(__DragonFly__) || defined(__APPLE__)
330 size_t len
=sizeof(has_sse
);
332 ret
= sysctlbyname(SSE_SYSCTL_NAME
, &has_sse
, &len
, NULL
, 0);
336 #elif defined(__NetBSD__) || defined (__OpenBSD__)
337 #if __NetBSD_Version__ >= 105250000 || (defined __OpenBSD__)
338 int has_sse
, has_sse2
, ret
, mib
[2];
341 mib
[0] = CTL_MACHDEP
;
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" );
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" );
358 mp_msg(MSGT_CPUDETECT
,MSGL_WARN
, "No OS support for SSE, disabling to be safe.\n" );
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
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!" );
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" );
417 #endif /* _POSIX_SOURCE */
419 /* Do nothing on other platforms for now.
421 mp_msg(MSGT_CPUDETECT
,MSGL_WARN
, "Cannot test OS support for SSE, leaving disabled.\n" );
423 #endif /* __linux__ */
428 #include <sys/sysctl.h>
429 #elif defined(__AMIGAOS4__)
435 static sigjmp_buf jmpbuf
;
436 static volatile sig_atomic_t canjump
= 0;
438 static void sigill_handler (int sig
)
441 signal (sig
, SIG_DFL
);
446 siglongjmp (jmpbuf
, 1);
448 #endif /* __APPLE__ */
450 void GetCpuCaps( CpuCaps
*caps
)
464 caps
->hasAltiVec
= 0;
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
};
474 size_t len
= sizeof(has_vu
);
477 err
= sysctl(sels
, 2, &has_vu
, &len
, NULL
, 0);
481 caps
->hasAltiVec
= 1;
483 #elif defined(__AMIGAOS4__)
486 GetCPUInfoTags(GCIT_VectorUnit
, &result
, TAG_DONE
);
487 if (result
== VECTORTYPE_ALTIVEC
)
488 caps
->hasAltiVec
= 1;
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
);
499 __asm__
volatile ("mtspr 256, %0\n\t"
500 "vand %%v0, %%v0, %%v0"
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 */
513 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: Intel Itanium\n");
516 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: Sun Sparc\n");
519 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: ARM\n");
522 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: PowerPC\n");
525 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: Digital Alpha\n");
528 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: SGI MIPS\n");
531 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: Hewlett-Packard PA-RISC\n");
534 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: IBM S/390\n");
537 mp_msg(MSGT_CPUDETECT
,MSGL_V
,"CPU: IBM S/390X\n");
540 mp_msg(MSGT_CPUDETECT
,MSGL_V
, "CPU: Digital VAX\n" );
543 mp_msg(MSGT_CPUDETECT
,MSGL_V
, "CPU: Tensilica Xtensa\n" );
545 #endif /* !ARCH_X86 */