1 /* Subroutines for the gcc driver.
2 Copyright (C) 2006, 2007, 2008, 2010, 2011, 2012
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
26 const char *host_detect_local_cpu (int argc
, const char **argv
);
38 /* Returns command line parameters that describe size and
39 cache line size of the processor caches. */
42 describe_cache (struct cache_desc level1
, struct cache_desc level2
)
44 char size
[100], line
[100], size2
[100];
46 /* At the moment, gcc does not use the information
47 about the associativity of the cache. */
49 snprintf (size
, sizeof (size
),
50 "--param l1-cache-size=%u ", level1
.sizekb
);
51 snprintf (line
, sizeof (line
),
52 "--param l1-cache-line-size=%u ", level1
.line
);
54 snprintf (size2
, sizeof (size2
),
55 "--param l2-cache-size=%u ", level2
.sizekb
);
57 return concat (size
, line
, size2
, NULL
);
60 /* Detect L2 cache parameters using CPUID extended function 0x80000006. */
63 detect_l2_cache (struct cache_desc
*level2
)
65 unsigned eax
, ebx
, ecx
, edx
;
68 __cpuid (0x80000006, eax
, ebx
, ecx
, edx
);
70 level2
->sizekb
= (ecx
>> 16) & 0xffff;
71 level2
->line
= ecx
& 0xff;
73 assoc
= (ecx
>> 12) & 0xf;
78 else if (assoc
>= 0xa && assoc
<= 0xc)
79 assoc
= 32 + (assoc
- 0xa) * 16;
80 else if (assoc
>= 0xd && assoc
<= 0xe)
81 assoc
= 96 + (assoc
- 0xd) * 32;
83 level2
->assoc
= assoc
;
86 /* Returns the description of caches for an AMD processor. */
89 detect_caches_amd (unsigned max_ext_level
)
91 unsigned eax
, ebx
, ecx
, edx
;
93 struct cache_desc level1
, level2
= {0, 0, 0};
95 if (max_ext_level
< 0x80000005)
98 __cpuid (0x80000005, eax
, ebx
, ecx
, edx
);
100 level1
.sizekb
= (ecx
>> 24) & 0xff;
101 level1
.assoc
= (ecx
>> 16) & 0xff;
102 level1
.line
= ecx
& 0xff;
104 if (max_ext_level
>= 0x80000006)
105 detect_l2_cache (&level2
);
107 return describe_cache (level1
, level2
);
110 /* Decodes the size, the associativity and the cache line size of
111 L1/L2 caches of an Intel processor. Values are based on
112 "Intel Processor Identification and the CPUID Instruction"
113 [Application Note 485], revision -032, December 2007. */
116 decode_caches_intel (unsigned reg
, bool xeon_mp
,
117 struct cache_desc
*level1
, struct cache_desc
*level2
)
121 for (i
= 24; i
>= 0; i
-= 8)
122 switch ((reg
>> i
) & 0xff)
125 level1
->sizekb
= 8; level1
->assoc
= 2; level1
->line
= 32;
128 level1
->sizekb
= 16; level1
->assoc
= 4; level1
->line
= 32;
131 level1
->sizekb
= 32; level1
->assoc
= 8; level1
->line
= 64;
134 level2
->sizekb
= 128; level2
->assoc
= 4; level2
->line
= 64;
137 level2
->sizekb
= 192; level2
->assoc
= 6; level2
->line
= 64;
140 level2
->sizekb
= 128; level2
->assoc
= 2; level2
->line
= 64;
143 level2
->sizekb
= 256; level2
->assoc
= 4; level2
->line
= 64;
146 level2
->sizekb
= 384; level2
->assoc
= 6; level2
->line
= 64;
149 level2
->sizekb
= 512; level2
->assoc
= 4; level2
->line
= 64;
152 level2
->sizekb
= 128; level2
->assoc
= 4; level2
->line
= 32;
155 level2
->sizekb
= 256; level2
->assoc
= 4; level2
->line
= 32;
158 level2
->sizekb
= 512; level2
->assoc
= 4; level2
->line
= 32;
161 level2
->sizekb
= 1024; level2
->assoc
= 4; level2
->line
= 32;
164 level2
->sizekb
= 2048; level2
->assoc
= 4; level2
->line
= 32;
169 level2
->sizekb
= 4096; level2
->assoc
= 16; level2
->line
= 64;
172 level2
->sizekb
= 6144; level2
->assoc
= 24; level2
->line
= 64;
175 level1
->sizekb
= 16; level1
->assoc
= 8; level1
->line
= 64;
178 level1
->sizekb
= 8; level1
->assoc
= 4; level1
->line
= 64;
181 level1
->sizekb
= 16; level1
->assoc
= 4; level1
->line
= 64;
184 level1
->sizekb
= 32; level1
->assoc
= 4; level1
->line
= 64;
187 level2
->sizekb
= 1024; level2
->assoc
= 4; level2
->line
= 64;
190 level2
->sizekb
= 128; level2
->assoc
= 8; level2
->line
= 64;
193 level2
->sizekb
= 256; level2
->assoc
= 8; level2
->line
= 64;
196 level2
->sizekb
= 512; level2
->assoc
= 8; level2
->line
= 64;
199 level2
->sizekb
= 1024; level2
->assoc
= 8; level2
->line
= 64;
202 level2
->sizekb
= 2048; level2
->assoc
= 8; level2
->line
= 64;
205 level2
->sizekb
= 512; level2
->assoc
= 2; level2
->line
= 64;
208 level2
->sizekb
= 256; level2
->assoc
= 8; level2
->line
= 32;
211 level2
->sizekb
= 512; level2
->assoc
= 8; level2
->line
= 32;
214 level2
->sizekb
= 1024; level2
->assoc
= 8; level2
->line
= 32;
217 level2
->sizekb
= 2048; level2
->assoc
= 8; level2
->line
= 32;
220 level2
->sizekb
= 512; level2
->assoc
= 4; level2
->line
= 64;
223 level2
->sizekb
= 1024; level2
->assoc
= 8; level2
->line
= 64;
230 /* Detect cache parameters using CPUID function 2. */
233 detect_caches_cpuid2 (bool xeon_mp
,
234 struct cache_desc
*level1
, struct cache_desc
*level2
)
239 __cpuid (2, regs
[0], regs
[1], regs
[2], regs
[3]);
241 nreps
= regs
[0] & 0x0f;
246 for (i
= 0; i
< 4; i
++)
247 if (regs
[i
] && !((regs
[i
] >> 31) & 1))
248 decode_caches_intel (regs
[i
], xeon_mp
, level1
, level2
);
251 __cpuid (2, regs
[0], regs
[1], regs
[2], regs
[3]);
255 /* Detect cache parameters using CPUID function 4. This
256 method doesn't require hardcoded tables. */
267 detect_caches_cpuid4 (struct cache_desc
*level1
, struct cache_desc
*level2
,
268 struct cache_desc
*level3
)
270 struct cache_desc
*cache
;
272 unsigned eax
, ebx
, ecx
, edx
;
275 for (count
= 0;; count
++)
277 __cpuid_count(4, count
, eax
, ebx
, ecx
, edx
);
285 switch ((eax
>> 5) & 0x07)
302 unsigned sets
= ecx
+ 1;
303 unsigned part
= ((ebx
>> 12) & 0x03ff) + 1;
305 cache
->assoc
= ((ebx
>> 22) & 0x03ff) + 1;
306 cache
->line
= (ebx
& 0x0fff) + 1;
308 cache
->sizekb
= (cache
->assoc
* part
309 * cache
->line
* sets
) / 1024;
318 /* Returns the description of caches for an Intel processor. */
321 detect_caches_intel (bool xeon_mp
, unsigned max_level
,
322 unsigned max_ext_level
, unsigned *l2sizekb
)
324 struct cache_desc level1
= {0, 0, 0}, level2
= {0, 0, 0}, level3
= {0, 0, 0};
327 detect_caches_cpuid4 (&level1
, &level2
, &level3
);
328 else if (max_level
>= 2)
329 detect_caches_cpuid2 (xeon_mp
, &level1
, &level2
);
333 if (level1
.sizekb
== 0)
336 /* Let the L3 replace the L2. This assumes inclusive caches
337 and single threaded program for now. */
341 /* Intel CPUs are equipped with AMD style L2 cache info. Try this
342 method if other methods fail to provide L2 cache parameters. */
343 if (level2
.sizekb
== 0 && max_ext_level
>= 0x80000006)
344 detect_l2_cache (&level2
);
346 *l2sizekb
= level2
.sizekb
;
348 return describe_cache (level1
, level2
);
351 /* This will be called by the spec parser in gcc.c when it sees
352 a %:local_cpu_detect(args) construct. Currently it will be called
353 with either "arch" or "tune" as argument depending on if -march=native
354 or -mtune=native is to be substituted.
356 It returns a string containing new command line parameters to be
357 put at the place of the above two options, depending on what CPU
358 this is executed. E.g. "-march=k8" on an AMD64 machine
361 ARGC and ARGV are set depending on the actual arguments given
364 const char *host_detect_local_cpu (int argc
, const char **argv
)
366 enum processor_type processor
= PROCESSOR_I386
;
367 const char *cpu
= "i386";
369 const char *cache
= "";
370 const char *options
= "";
372 unsigned int eax
, ebx
, ecx
, edx
;
374 unsigned int max_level
, ext_level
;
377 unsigned int model
, family
;
379 unsigned int has_sse3
, has_ssse3
, has_cmpxchg16b
;
380 unsigned int has_cmpxchg8b
, has_cmov
, has_mmx
, has_sse
, has_sse2
;
382 /* Extended features */
383 unsigned int has_lahf_lm
= 0, has_sse4a
= 0;
384 unsigned int has_longmode
= 0, has_3dnowp
= 0, has_3dnow
= 0;
385 unsigned int has_movbe
= 0, has_sse4_1
= 0, has_sse4_2
= 0;
386 unsigned int has_popcnt
= 0, has_aes
= 0, has_avx
= 0, has_avx2
= 0;
387 unsigned int has_pclmul
= 0, has_abm
= 0, has_lwp
= 0;
388 unsigned int has_fma
= 0, has_fma4
= 0, has_xop
= 0;
389 unsigned int has_bmi
= 0, has_bmi2
= 0, has_tbm
= 0, has_lzcnt
= 0;
390 unsigned int has_hle
= 0, has_rtm
= 0;
391 unsigned int has_rdrnd
= 0, has_f16c
= 0, has_fsgsbase
= 0;
392 unsigned int has_rdseed
= 0, has_prfchw
= 0, has_adx
= 0;
396 unsigned int l2sizekb
= 0;
401 arch
= !strcmp (argv
[0], "arch");
403 if (!arch
&& strcmp (argv
[0], "tune"))
406 max_level
= __get_cpuid_max (0, &vendor
);
410 __cpuid (1, eax
, ebx
, ecx
, edx
);
412 model
= (eax
>> 4) & 0x0f;
413 family
= (eax
>> 8) & 0x0f;
414 if (vendor
== signature_INTEL_ebx
)
416 unsigned int extended_model
, extended_family
;
418 extended_model
= (eax
>> 12) & 0xf0;
419 extended_family
= (eax
>> 20) & 0xff;
422 family
+= extended_family
;
423 model
+= extended_model
;
425 else if (family
== 0x06)
426 model
+= extended_model
;
429 has_sse3
= ecx
& bit_SSE3
;
430 has_ssse3
= ecx
& bit_SSSE3
;
431 has_sse4_1
= ecx
& bit_SSE4_1
;
432 has_sse4_2
= ecx
& bit_SSE4_2
;
433 has_avx
= ecx
& bit_AVX
;
434 has_cmpxchg16b
= ecx
& bit_CMPXCHG16B
;
435 has_movbe
= ecx
& bit_MOVBE
;
436 has_popcnt
= ecx
& bit_POPCNT
;
437 has_aes
= ecx
& bit_AES
;
438 has_pclmul
= ecx
& bit_PCLMUL
;
439 has_fma
= ecx
& bit_FMA
;
440 has_f16c
= ecx
& bit_F16C
;
441 has_rdrnd
= ecx
& bit_RDRND
;
443 has_cmpxchg8b
= edx
& bit_CMPXCHG8B
;
444 has_cmov
= edx
& bit_CMOV
;
445 has_mmx
= edx
& bit_MMX
;
446 has_sse
= edx
& bit_SSE
;
447 has_sse2
= edx
& bit_SSE2
;
451 __cpuid_count (7, 0, eax
, ebx
, ecx
, edx
);
453 has_bmi
= ebx
& bit_BMI
;
454 has_hle
= ebx
& bit_HLE
;
455 has_rtm
= ebx
& bit_RTM
;
456 has_avx2
= ebx
& bit_AVX2
;
457 has_bmi2
= ebx
& bit_BMI2
;
458 has_fsgsbase
= ebx
& bit_FSGSBASE
;
459 has_rdseed
= ebx
& bit_RDSEED
;
460 has_adx
= ebx
& bit_ADX
;
463 /* Check cpuid level of extended features. */
464 __cpuid (0x80000000, ext_level
, ebx
, ecx
, edx
);
466 if (ext_level
> 0x80000000)
468 __cpuid (0x80000001, eax
, ebx
, ecx
, edx
);
470 has_lahf_lm
= ecx
& bit_LAHF_LM
;
471 has_sse4a
= ecx
& bit_SSE4a
;
472 has_abm
= ecx
& bit_ABM
;
473 has_lwp
= ecx
& bit_LWP
;
474 has_fma4
= ecx
& bit_FMA4
;
475 has_xop
= ecx
& bit_XOP
;
476 has_tbm
= ecx
& bit_TBM
;
477 has_lzcnt
= ecx
& bit_LZCNT
;
478 has_prfchw
= ecx
& bit_PRFCHW
;
480 has_longmode
= edx
& bit_LM
;
481 has_3dnowp
= edx
& bit_3DNOWP
;
482 has_3dnow
= edx
& bit_3DNOW
;
487 if (vendor
== signature_AMD_ebx
)
488 cache
= detect_caches_amd (ext_level
);
489 else if (vendor
== signature_INTEL_ebx
)
491 bool xeon_mp
= (family
== 15 && model
== 6);
492 cache
= detect_caches_intel (xeon_mp
, max_level
,
493 ext_level
, &l2sizekb
);
497 if (vendor
== signature_AMD_ebx
)
501 /* Detect geode processor by its processor signature. */
502 if (ext_level
> 0x80000001)
503 __cpuid (0x80000002, name
, ebx
, ecx
, edx
);
507 if (name
== signature_NSC_ebx
)
508 processor
= PROCESSOR_GEODE
;
510 processor
= PROCESSOR_BTVER2
;
512 processor
= PROCESSOR_BDVER2
;
514 processor
= PROCESSOR_BDVER1
;
515 else if (has_sse4a
&& has_ssse3
)
516 processor
= PROCESSOR_BTVER1
;
518 processor
= PROCESSOR_AMDFAM10
;
519 else if (has_sse2
|| has_longmode
)
520 processor
= PROCESSOR_K8
;
521 else if (has_3dnowp
&& family
== 6)
522 processor
= PROCESSOR_ATHLON
;
524 processor
= PROCESSOR_K6
;
526 processor
= PROCESSOR_PENTIUM
;
533 processor
= PROCESSOR_I486
;
536 processor
= PROCESSOR_PENTIUM
;
539 processor
= PROCESSOR_PENTIUMPRO
;
542 processor
= PROCESSOR_PENTIUM4
;
545 /* We have no idea. */
546 processor
= PROCESSOR_GENERIC32
;
558 case PROCESSOR_PENTIUM
:
564 case PROCESSOR_PENTIUMPRO
:
602 /* This is unknown family 0x6 CPU. */
604 /* Assume Sandy Bridge. */
607 /* Assume Core i7. */
619 /* It is Core Duo. */
622 /* It is Pentium M. */
625 /* It is Pentium III. */
628 /* It is Pentium II. */
631 /* Default to Pentium Pro. */
635 /* For -mtune, we default to -mtune=generic. */
640 case PROCESSOR_PENTIUM4
:
651 case PROCESSOR_GEODE
:
655 if (arch
&& has_3dnow
)
660 case PROCESSOR_ATHLON
:
667 if (arch
&& has_sse3
)
672 case PROCESSOR_AMDFAM10
:
675 case PROCESSOR_BDVER1
:
678 case PROCESSOR_BDVER2
:
681 case PROCESSOR_BTVER1
:
684 case PROCESSOR_BTVER2
:
689 /* Use something reasonable. */
707 else if (has_cmpxchg8b
)
716 const char *cx16
= has_cmpxchg16b
? " -mcx16" : " -mno-cx16";
717 const char *sahf
= has_lahf_lm
? " -msahf" : " -mno-sahf";
718 const char *movbe
= has_movbe
? " -mmovbe" : " -mno-movbe";
719 const char *ase
= has_aes
? " -maes" : " -mno-aes";
720 const char *pclmul
= has_pclmul
? " -mpclmul" : " -mno-pclmul";
721 const char *popcnt
= has_popcnt
? " -mpopcnt" : " -mno-popcnt";
722 const char *abm
= has_abm
? " -mabm" : " -mno-abm";
723 const char *lwp
= has_lwp
? " -mlwp" : " -mno-lwp";
724 const char *fma
= has_fma
? " -mfma" : " -mno-fma";
725 const char *fma4
= has_fma4
? " -mfma4" : " -mno-fma4";
726 const char *xop
= has_xop
? " -mxop" : " -mno-xop";
727 const char *bmi
= has_bmi
? " -mbmi" : " -mno-bmi";
728 const char *bmi2
= has_bmi2
? " -mbmi2" : " -mno-bmi2";
729 const char *tbm
= has_tbm
? " -mtbm" : " -mno-tbm";
730 const char *avx
= has_avx
? " -mavx" : " -mno-avx";
731 const char *avx2
= has_avx2
? " -mavx2" : " -mno-avx2";
732 const char *sse4_2
= has_sse4_2
? " -msse4.2" : " -mno-sse4.2";
733 const char *sse4_1
= has_sse4_1
? " -msse4.1" : " -mno-sse4.1";
734 const char *lzcnt
= has_lzcnt
? " -mlzcnt" : " -mno-lzcnt";
735 const char *hle
= has_hle
? " -mhle" : " -mno-hle";
736 const char *rtm
= has_rtm
? " -mrtm" : " -mno-rtm";
737 const char *rdrnd
= has_rdrnd
? " -mrdrnd" : " -mno-rdrnd";
738 const char *f16c
= has_f16c
? " -mf16c" : " -mno-f16c";
739 const char *fsgsbase
= has_fsgsbase
? " -mfsgsbase" : " -mno-fsgsbase";
740 const char *rdseed
= has_rdseed
? " -mrdseed" : " -mno-rdseed";
741 const char *prfchw
= has_prfchw
? " -mprfchw" : " -mno-prfchw";
742 const char *adx
= has_adx
? " -madx" : " -mno-adx";
744 options
= concat (options
, cx16
, sahf
, movbe
, ase
, pclmul
,
745 popcnt
, abm
, lwp
, fma
, fma4
, xop
, bmi
, bmi2
,
746 tbm
, avx
, avx2
, sse4_2
, sse4_1
, lzcnt
, rtm
,
747 hle
, rdrnd
, f16c
, fsgsbase
, rdseed
, prfchw
, adx
, NULL
);
751 return concat (cache
, "-m", argv
[0], "=", cpu
, options
, NULL
);
755 /* If we aren't compiling with GCC then the driver will just ignore
756 -march and -mtune "native" target and will leave to the newly
757 built compiler to generate code for its default target. */
759 const char *host_detect_local_cpu (int argc ATTRIBUTE_UNUSED
,
760 const char **argv ATTRIBUTE_UNUSED
)
764 #endif /* __GNUC__ */