1 | /* Initialize CPU feature data. |
2 | This file is part of the GNU C Library. |
3 | Copyright (C) 2008-2018 Free Software Foundation, Inc. |
4 | |
5 | The GNU C Library is free software; you can redistribute it and/or |
6 | modify it under the terms of the GNU Lesser General Public |
7 | License as published by the Free Software Foundation; either |
8 | version 2.1 of the License, or (at your option) any later version. |
9 | |
10 | The GNU C Library is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
13 | Lesser General Public License for more details. |
14 | |
15 | You should have received a copy of the GNU Lesser General Public |
16 | License along with the GNU C Library; if not, see |
17 | <http://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <cpuid.h> |
20 | #include <cpu-features.h> |
21 | #include <dl-hwcap.h> |
22 | #include <libc-pointer-arith.h> |
23 | |
24 | #if HAVE_TUNABLES |
25 | # define TUNABLE_NAMESPACE tune |
26 | # include <unistd.h> /* Get STDOUT_FILENO for _dl_printf. */ |
27 | # include <elf/dl-tunables.h> |
28 | |
29 | extern void TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *) |
30 | attribute_hidden; |
31 | |
32 | # if CET_ENABLED |
33 | extern void TUNABLE_CALLBACK (set_x86_ibt) (tunable_val_t *) |
34 | attribute_hidden; |
35 | extern void TUNABLE_CALLBACK (set_x86_shstk) (tunable_val_t *) |
36 | attribute_hidden; |
37 | # endif |
38 | #endif |
39 | |
40 | #if CET_ENABLED |
41 | # include <dl-cet.h> |
42 | # include <cet-tunables.h> |
43 | #endif |
44 | |
45 | static void |
46 | get_extended_indices (struct cpu_features *cpu_features) |
47 | { |
48 | unsigned int eax, ebx, ecx, edx; |
49 | __cpuid (0x80000000, eax, ebx, ecx, edx); |
50 | if (eax >= 0x80000001) |
51 | __cpuid (0x80000001, |
52 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].eax, |
53 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].ebx, |
54 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].ecx, |
55 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].edx); |
56 | |
57 | } |
58 | |
59 | static void |
60 | get_common_indeces (struct cpu_features *cpu_features, |
61 | unsigned int *family, unsigned int *model, |
62 | unsigned int *extended_model, unsigned int *stepping) |
63 | { |
64 | if (family) |
65 | { |
66 | unsigned int eax; |
67 | __cpuid (1, eax, cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx, |
68 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx, |
69 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].edx); |
70 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].eax = eax; |
71 | *family = (eax >> 8) & 0x0f; |
72 | *model = (eax >> 4) & 0x0f; |
73 | *extended_model = (eax >> 12) & 0xf0; |
74 | *stepping = eax & 0x0f; |
75 | if (*family == 0x0f) |
76 | { |
77 | *family += (eax >> 20) & 0xff; |
78 | *model += *extended_model; |
79 | } |
80 | } |
81 | |
82 | if (cpu_features->max_cpuid >= 7) |
83 | __cpuid_count (7, 0, |
84 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].eax, |
85 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ebx, |
86 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ecx, |
87 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].edx); |
88 | |
89 | /* Can we call xgetbv? */ |
90 | if (CPU_FEATURES_CPU_P (cpu_features, OSXSAVE)) |
91 | { |
92 | unsigned int xcrlow; |
93 | unsigned int xcrhigh; |
94 | asm ("xgetbv" : "=a" (xcrlow), "=d" (xcrhigh) : "c" (0)); |
95 | /* Is YMM and XMM state usable? */ |
96 | if ((xcrlow & (bit_YMM_state | bit_XMM_state)) == |
97 | (bit_YMM_state | bit_XMM_state)) |
98 | { |
99 | /* Determine if AVX is usable. */ |
100 | if (CPU_FEATURES_CPU_P (cpu_features, AVX)) |
101 | { |
102 | cpu_features->feature[index_arch_AVX_Usable] |
103 | |= bit_arch_AVX_Usable; |
104 | /* The following features depend on AVX being usable. */ |
105 | /* Determine if AVX2 is usable. */ |
106 | if (CPU_FEATURES_CPU_P (cpu_features, AVX2)) |
107 | { |
108 | cpu_features->feature[index_arch_AVX2_Usable] |
109 | |= bit_arch_AVX2_Usable; |
110 | |
111 | /* Unaligned load with 256-bit AVX registers are faster on |
112 | Intel/AMD processors with AVX2. */ |
113 | cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load] |
114 | |= bit_arch_AVX_Fast_Unaligned_Load; |
115 | } |
116 | /* Determine if FMA is usable. */ |
117 | if (CPU_FEATURES_CPU_P (cpu_features, FMA)) |
118 | cpu_features->feature[index_arch_FMA_Usable] |
119 | |= bit_arch_FMA_Usable; |
120 | } |
121 | |
122 | /* Check if OPMASK state, upper 256-bit of ZMM0-ZMM15 and |
123 | ZMM16-ZMM31 state are enabled. */ |
124 | if ((xcrlow & (bit_Opmask_state | bit_ZMM0_15_state |
125 | | bit_ZMM16_31_state)) == |
126 | (bit_Opmask_state | bit_ZMM0_15_state | bit_ZMM16_31_state)) |
127 | { |
128 | /* Determine if AVX512F is usable. */ |
129 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512F)) |
130 | { |
131 | cpu_features->feature[index_arch_AVX512F_Usable] |
132 | |= bit_arch_AVX512F_Usable; |
133 | /* Determine if AVX512DQ is usable. */ |
134 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512DQ)) |
135 | cpu_features->feature[index_arch_AVX512DQ_Usable] |
136 | |= bit_arch_AVX512DQ_Usable; |
137 | } |
138 | } |
139 | } |
140 | |
141 | /* For _dl_runtime_resolve, set xsave_state_size to xsave area |
142 | size + integer register save size and align it to 64 bytes. */ |
143 | if (cpu_features->max_cpuid >= 0xd) |
144 | { |
145 | unsigned int eax, ebx, ecx, edx; |
146 | |
147 | __cpuid_count (0xd, 0, eax, ebx, ecx, edx); |
148 | if (ebx != 0) |
149 | { |
150 | unsigned int xsave_state_full_size |
151 | = ALIGN_UP (ebx + STATE_SAVE_OFFSET, 64); |
152 | |
153 | cpu_features->xsave_state_size |
154 | = xsave_state_full_size; |
155 | cpu_features->xsave_state_full_size |
156 | = xsave_state_full_size; |
157 | |
158 | __cpuid_count (0xd, 1, eax, ebx, ecx, edx); |
159 | |
160 | /* Check if XSAVEC is available. */ |
161 | if ((eax & (1 << 1)) != 0) |
162 | { |
163 | unsigned int xstate_comp_offsets[32]; |
164 | unsigned int xstate_comp_sizes[32]; |
165 | unsigned int i; |
166 | |
167 | xstate_comp_offsets[0] = 0; |
168 | xstate_comp_offsets[1] = 160; |
169 | xstate_comp_offsets[2] = 576; |
170 | xstate_comp_sizes[0] = 160; |
171 | xstate_comp_sizes[1] = 256; |
172 | |
173 | for (i = 2; i < 32; i++) |
174 | { |
175 | if ((STATE_SAVE_MASK & (1 << i)) != 0) |
176 | { |
177 | __cpuid_count (0xd, i, eax, ebx, ecx, edx); |
178 | xstate_comp_sizes[i] = eax; |
179 | } |
180 | else |
181 | { |
182 | ecx = 0; |
183 | xstate_comp_sizes[i] = 0; |
184 | } |
185 | |
186 | if (i > 2) |
187 | { |
188 | xstate_comp_offsets[i] |
189 | = (xstate_comp_offsets[i - 1] |
190 | + xstate_comp_sizes[i -1]); |
191 | if ((ecx & (1 << 1)) != 0) |
192 | xstate_comp_offsets[i] |
193 | = ALIGN_UP (xstate_comp_offsets[i], 64); |
194 | } |
195 | } |
196 | |
197 | /* Use XSAVEC. */ |
198 | unsigned int size |
199 | = xstate_comp_offsets[31] + xstate_comp_sizes[31]; |
200 | if (size) |
201 | { |
202 | cpu_features->xsave_state_size |
203 | = ALIGN_UP (size + STATE_SAVE_OFFSET, 64); |
204 | cpu_features->feature[index_arch_XSAVEC_Usable] |
205 | |= bit_arch_XSAVEC_Usable; |
206 | } |
207 | } |
208 | } |
209 | } |
210 | } |
211 | } |
212 | |
213 | static inline void |
214 | init_cpu_features (struct cpu_features *cpu_features) |
215 | { |
216 | unsigned int ebx, ecx, edx; |
217 | unsigned int family = 0; |
218 | unsigned int model = 0; |
219 | enum cpu_features_kind kind; |
220 | |
221 | #if !HAS_CPUID |
222 | if (__get_cpuid_max (0, 0) == 0) |
223 | { |
224 | kind = arch_kind_other; |
225 | goto no_cpuid; |
226 | } |
227 | #endif |
228 | |
229 | __cpuid (0, cpu_features->max_cpuid, ebx, ecx, edx); |
230 | |
231 | /* This spells out "GenuineIntel". */ |
232 | if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69) |
233 | { |
234 | unsigned int extended_model, stepping; |
235 | |
236 | kind = arch_kind_intel; |
237 | |
238 | get_common_indeces (cpu_features, &family, &model, &extended_model, |
239 | &stepping); |
240 | |
241 | get_extended_indices (cpu_features); |
242 | |
243 | if (family == 0x06) |
244 | { |
245 | model += extended_model; |
246 | switch (model) |
247 | { |
248 | case 0x1c: |
249 | case 0x26: |
250 | /* BSF is slow on Atom. */ |
251 | cpu_features->feature[index_arch_Slow_BSF] |
252 | |= bit_arch_Slow_BSF; |
253 | break; |
254 | |
255 | case 0x57: |
256 | /* Knights Landing. Enable Silvermont optimizations. */ |
257 | |
258 | case 0x5c: |
259 | case 0x5f: |
260 | /* Unaligned load versions are faster than SSSE3 |
261 | on Goldmont. */ |
262 | |
263 | case 0x4c: |
264 | /* Airmont is a die shrink of Silvermont. */ |
265 | |
266 | case 0x37: |
267 | case 0x4a: |
268 | case 0x4d: |
269 | case 0x5a: |
270 | case 0x5d: |
271 | /* Unaligned load versions are faster than SSSE3 |
272 | on Silvermont. */ |
273 | #if index_arch_Fast_Unaligned_Load != index_arch_Prefer_PMINUB_for_stringop |
274 | # error index_arch_Fast_Unaligned_Load != index_arch_Prefer_PMINUB_for_stringop |
275 | #endif |
276 | #if index_arch_Fast_Unaligned_Load != index_arch_Slow_SSE4_2 |
277 | # error index_arch_Fast_Unaligned_Load != index_arch_Slow_SSE4_2 |
278 | #endif |
279 | #if index_arch_Fast_Unaligned_Load != index_arch_Fast_Unaligned_Copy |
280 | # error index_arch_Fast_Unaligned_Load != index_arch_Fast_Unaligned_Copy |
281 | #endif |
282 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
283 | |= (bit_arch_Fast_Unaligned_Load |
284 | | bit_arch_Fast_Unaligned_Copy |
285 | | bit_arch_Prefer_PMINUB_for_stringop |
286 | | bit_arch_Slow_SSE4_2); |
287 | break; |
288 | |
289 | default: |
290 | /* Unknown family 0x06 processors. Assuming this is one |
291 | of Core i3/i5/i7 processors if AVX is available. */ |
292 | if (!CPU_FEATURES_CPU_P (cpu_features, AVX)) |
293 | break; |
294 | |
295 | case 0x1a: |
296 | case 0x1e: |
297 | case 0x1f: |
298 | case 0x25: |
299 | case 0x2c: |
300 | case 0x2e: |
301 | case 0x2f: |
302 | /* Rep string instructions, unaligned load, unaligned copy, |
303 | and pminub are fast on Intel Core i3, i5 and i7. */ |
304 | #if index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Load |
305 | # error index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Load |
306 | #endif |
307 | #if index_arch_Fast_Rep_String != index_arch_Prefer_PMINUB_for_stringop |
308 | # error index_arch_Fast_Rep_String != index_arch_Prefer_PMINUB_for_stringop |
309 | #endif |
310 | #if index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Copy |
311 | # error index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Copy |
312 | #endif |
313 | cpu_features->feature[index_arch_Fast_Rep_String] |
314 | |= (bit_arch_Fast_Rep_String |
315 | | bit_arch_Fast_Unaligned_Load |
316 | | bit_arch_Fast_Unaligned_Copy |
317 | | bit_arch_Prefer_PMINUB_for_stringop); |
318 | break; |
319 | |
320 | case 0x3f: |
321 | /* Xeon E7 v3 with stepping >= 4 has working TSX. */ |
322 | if (stepping >= 4) |
323 | break; |
324 | case 0x3c: |
325 | case 0x45: |
326 | case 0x46: |
327 | /* Disable Intel TSX on Haswell processors (except Xeon E7 v3 |
328 | with stepping >= 4) to avoid TSX on kernels that weren't |
329 | updated with the latest microcode package (which disables |
330 | broken feature by default). */ |
331 | cpu_features->cpuid[index_cpu_RTM].reg_RTM &= ~bit_cpu_RTM; |
332 | break; |
333 | } |
334 | } |
335 | |
336 | |
337 | /* Since AVX512ER is unique to Xeon Phi, set Prefer_No_VZEROUPPER |
338 | if AVX512ER is available. Don't use AVX512 to avoid lower CPU |
339 | frequency if AVX512ER isn't available. */ |
340 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
341 | cpu_features->feature[index_arch_Prefer_No_VZEROUPPER] |
342 | |= bit_arch_Prefer_No_VZEROUPPER; |
343 | else |
344 | cpu_features->feature[index_arch_Prefer_No_AVX512] |
345 | |= bit_arch_Prefer_No_AVX512; |
346 | } |
347 | /* This spells out "AuthenticAMD". */ |
348 | else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65) |
349 | { |
350 | unsigned int extended_model, stepping; |
351 | |
352 | kind = arch_kind_amd; |
353 | |
354 | get_common_indeces (cpu_features, &family, &model, &extended_model, |
355 | &stepping); |
356 | |
357 | get_extended_indices (cpu_features); |
358 | |
359 | ecx = cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx; |
360 | |
361 | if (HAS_ARCH_FEATURE (AVX_Usable)) |
362 | { |
363 | /* Since the FMA4 bit is in COMMON_CPUID_INDEX_80000001 and |
364 | FMA4 requires AVX, determine if FMA4 is usable here. */ |
365 | if (CPU_FEATURES_CPU_P (cpu_features, FMA4)) |
366 | cpu_features->feature[index_arch_FMA4_Usable] |
367 | |= bit_arch_FMA4_Usable; |
368 | } |
369 | |
370 | if (family == 0x15) |
371 | { |
372 | #if index_arch_Fast_Unaligned_Load != index_arch_Fast_Copy_Backward |
373 | # error index_arch_Fast_Unaligned_Load != index_arch_Fast_Copy_Backward |
374 | #endif |
375 | /* "Excavator" */ |
376 | if (model >= 0x60 && model <= 0x7f) |
377 | { |
378 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
379 | |= (bit_arch_Fast_Unaligned_Load |
380 | | bit_arch_Fast_Copy_Backward); |
381 | |
382 | /* Unaligned AVX loads are slower.*/ |
383 | cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load] |
384 | &= ~bit_arch_AVX_Fast_Unaligned_Load; |
385 | } |
386 | } |
387 | } |
388 | else |
389 | { |
390 | kind = arch_kind_other; |
391 | get_common_indeces (cpu_features, NULL, NULL, NULL, NULL); |
392 | } |
393 | |
394 | /* Support i586 if CX8 is available. */ |
395 | if (CPU_FEATURES_CPU_P (cpu_features, CX8)) |
396 | cpu_features->feature[index_arch_I586] |= bit_arch_I586; |
397 | |
398 | /* Support i686 if CMOV is available. */ |
399 | if (CPU_FEATURES_CPU_P (cpu_features, CMOV)) |
400 | cpu_features->feature[index_arch_I686] |= bit_arch_I686; |
401 | |
402 | #if !HAS_CPUID |
403 | no_cpuid: |
404 | #endif |
405 | |
406 | cpu_features->family = family; |
407 | cpu_features->model = model; |
408 | cpu_features->kind = kind; |
409 | |
410 | #if HAVE_TUNABLES |
411 | TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps)); |
412 | cpu_features->non_temporal_threshold |
413 | = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL); |
414 | cpu_features->data_cache_size |
415 | = TUNABLE_GET (x86_data_cache_size, long int, NULL); |
416 | cpu_features->shared_cache_size |
417 | = TUNABLE_GET (x86_shared_cache_size, long int, NULL); |
418 | #endif |
419 | |
420 | /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86. */ |
421 | #if !HAVE_TUNABLES && defined SHARED |
422 | /* The glibc.tune.hwcap_mask tunable is initialized already, so no need to do |
423 | this. */ |
424 | GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT; |
425 | #endif |
426 | |
427 | #ifdef __x86_64__ |
428 | GLRO(dl_hwcap) = HWCAP_X86_64; |
429 | if (cpu_features->kind == arch_kind_intel) |
430 | { |
431 | const char *platform = NULL; |
432 | |
433 | if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable) |
434 | && CPU_FEATURES_CPU_P (cpu_features, AVX512CD)) |
435 | { |
436 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
437 | { |
438 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512PF)) |
439 | platform = "xeon_phi" ; |
440 | } |
441 | else |
442 | { |
443 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512BW) |
444 | && CPU_FEATURES_CPU_P (cpu_features, AVX512DQ) |
445 | && CPU_FEATURES_CPU_P (cpu_features, AVX512VL)) |
446 | GLRO(dl_hwcap) |= HWCAP_X86_AVX512_1; |
447 | } |
448 | } |
449 | |
450 | if (platform == NULL |
451 | && CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable) |
452 | && CPU_FEATURES_ARCH_P (cpu_features, FMA_Usable) |
453 | && CPU_FEATURES_CPU_P (cpu_features, BMI1) |
454 | && CPU_FEATURES_CPU_P (cpu_features, BMI2) |
455 | && CPU_FEATURES_CPU_P (cpu_features, LZCNT) |
456 | && CPU_FEATURES_CPU_P (cpu_features, MOVBE) |
457 | && CPU_FEATURES_CPU_P (cpu_features, POPCNT)) |
458 | platform = "haswell" ; |
459 | |
460 | if (platform != NULL) |
461 | GLRO(dl_platform) = platform; |
462 | } |
463 | #else |
464 | GLRO(dl_hwcap) = 0; |
465 | if (CPU_FEATURES_CPU_P (cpu_features, SSE2)) |
466 | GLRO(dl_hwcap) |= HWCAP_X86_SSE2; |
467 | |
468 | if (CPU_FEATURES_ARCH_P (cpu_features, I686)) |
469 | GLRO(dl_platform) = "i686" ; |
470 | else if (CPU_FEATURES_ARCH_P (cpu_features, I586)) |
471 | GLRO(dl_platform) = "i586" ; |
472 | #endif |
473 | |
474 | #if CET_ENABLED |
475 | # if HAVE_TUNABLES |
476 | TUNABLE_GET (x86_ibt, tunable_val_t *, |
477 | TUNABLE_CALLBACK (set_x86_ibt)); |
478 | TUNABLE_GET (x86_shstk, tunable_val_t *, |
479 | TUNABLE_CALLBACK (set_x86_shstk)); |
480 | # endif |
481 | |
482 | /* Check CET status. */ |
483 | unsigned int cet_status = get_cet_status (); |
484 | |
485 | if (cet_status) |
486 | { |
487 | GL(dl_x86_feature_1)[0] = cet_status; |
488 | |
489 | # ifndef SHARED |
490 | /* Check if IBT and SHSTK are enabled by kernel. */ |
491 | if ((cet_status & GNU_PROPERTY_X86_FEATURE_1_IBT) |
492 | || (cet_status & GNU_PROPERTY_X86_FEATURE_1_SHSTK)) |
493 | { |
494 | /* Disable IBT and/or SHSTK if they are enabled by kernel, but |
495 | disabled by environment variable: |
496 | |
497 | GLIBC_TUNABLES=glibc.tune.hwcaps=-IBT,-SHSTK |
498 | */ |
499 | unsigned int cet_feature = 0; |
500 | if (!HAS_CPU_FEATURE (IBT)) |
501 | cet_feature |= GNU_PROPERTY_X86_FEATURE_1_IBT; |
502 | if (!HAS_CPU_FEATURE (SHSTK)) |
503 | cet_feature |= GNU_PROPERTY_X86_FEATURE_1_SHSTK; |
504 | |
505 | if (cet_feature) |
506 | { |
507 | int res = dl_cet_disable_cet (cet_feature); |
508 | |
509 | /* Clear the disabled bits in dl_x86_feature_1. */ |
510 | if (res == 0) |
511 | GL(dl_x86_feature_1)[0] &= ~cet_feature; |
512 | } |
513 | |
514 | /* Lock CET if IBT or SHSTK is enabled in executable. Don't |
515 | lock CET if SHSTK is enabled permissively. */ |
516 | if (((GL(dl_x86_feature_1)[1] >> CET_MAX) |
517 | & ((1 << CET_MAX) - 1)) |
518 | != CET_PERMISSIVE) |
519 | dl_cet_lock_cet (); |
520 | } |
521 | # endif |
522 | } |
523 | #endif |
524 | } |
525 | |