1 | /* Initialize CPU feature data. |
2 | This file is part of the GNU C Library. |
3 | Copyright (C) 2008-2019 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 cpu |
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 | if (eax >= 0x80000007) |
57 | __cpuid (0x80000007, |
58 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000007].eax, |
59 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000007].ebx, |
60 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000007].ecx, |
61 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000007].edx); |
62 | if (eax >= 0x80000008) |
63 | __cpuid (0x80000008, |
64 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000008].eax, |
65 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000008].ebx, |
66 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000008].ecx, |
67 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000008].edx); |
68 | } |
69 | |
70 | static void |
71 | get_common_indices (struct cpu_features *cpu_features, |
72 | unsigned int *family, unsigned int *model, |
73 | unsigned int *extended_model, unsigned int *stepping) |
74 | { |
75 | if (family) |
76 | { |
77 | unsigned int eax; |
78 | __cpuid (1, eax, cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx, |
79 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx, |
80 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].edx); |
81 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].eax = eax; |
82 | *family = (eax >> 8) & 0x0f; |
83 | *model = (eax >> 4) & 0x0f; |
84 | *extended_model = (eax >> 12) & 0xf0; |
85 | *stepping = eax & 0x0f; |
86 | if (*family == 0x0f) |
87 | { |
88 | *family += (eax >> 20) & 0xff; |
89 | *model += *extended_model; |
90 | } |
91 | } |
92 | |
93 | if (cpu_features->basic.max_cpuid >= 7) |
94 | __cpuid_count (7, 0, |
95 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].eax, |
96 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ebx, |
97 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ecx, |
98 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].edx); |
99 | |
100 | if (cpu_features->basic.max_cpuid >= 0xd) |
101 | __cpuid_count (0xd, 1, |
102 | cpu_features->cpuid[COMMON_CPUID_INDEX_D_ECX_1].eax, |
103 | cpu_features->cpuid[COMMON_CPUID_INDEX_D_ECX_1].ebx, |
104 | cpu_features->cpuid[COMMON_CPUID_INDEX_D_ECX_1].ecx, |
105 | cpu_features->cpuid[COMMON_CPUID_INDEX_D_ECX_1].edx); |
106 | |
107 | /* Can we call xgetbv? */ |
108 | if (CPU_FEATURES_CPU_P (cpu_features, OSXSAVE)) |
109 | { |
110 | unsigned int xcrlow; |
111 | unsigned int xcrhigh; |
112 | asm ("xgetbv" : "=a" (xcrlow), "=d" (xcrhigh) : "c" (0)); |
113 | /* Is YMM and XMM state usable? */ |
114 | if ((xcrlow & (bit_YMM_state | bit_XMM_state)) == |
115 | (bit_YMM_state | bit_XMM_state)) |
116 | { |
117 | /* Determine if AVX is usable. */ |
118 | if (CPU_FEATURES_CPU_P (cpu_features, AVX)) |
119 | { |
120 | cpu_features->feature[index_arch_AVX_Usable] |
121 | |= bit_arch_AVX_Usable; |
122 | /* The following features depend on AVX being usable. */ |
123 | /* Determine if AVX2 is usable. */ |
124 | if (CPU_FEATURES_CPU_P (cpu_features, AVX2)) |
125 | { |
126 | cpu_features->feature[index_arch_AVX2_Usable] |
127 | |= bit_arch_AVX2_Usable; |
128 | |
129 | /* Unaligned load with 256-bit AVX registers are faster on |
130 | Intel/AMD processors with AVX2. */ |
131 | cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load] |
132 | |= bit_arch_AVX_Fast_Unaligned_Load; |
133 | } |
134 | /* Determine if FMA is usable. */ |
135 | if (CPU_FEATURES_CPU_P (cpu_features, FMA)) |
136 | cpu_features->feature[index_arch_FMA_Usable] |
137 | |= bit_arch_FMA_Usable; |
138 | /* Determine if VAES is usable. */ |
139 | if (CPU_FEATURES_CPU_P (cpu_features, VAES)) |
140 | cpu_features->feature[index_arch_VAES_Usable] |
141 | |= bit_arch_VAES_Usable; |
142 | /* Determine if VPCLMULQDQ is usable. */ |
143 | if (CPU_FEATURES_CPU_P (cpu_features, VPCLMULQDQ)) |
144 | cpu_features->feature[index_arch_VPCLMULQDQ_Usable] |
145 | |= bit_arch_VPCLMULQDQ_Usable; |
146 | /* Determine if XOP is usable. */ |
147 | if (CPU_FEATURES_CPU_P (cpu_features, XOP)) |
148 | cpu_features->feature[index_arch_XOP_Usable] |
149 | |= bit_arch_XOP_Usable; |
150 | } |
151 | |
152 | /* Check if OPMASK state, upper 256-bit of ZMM0-ZMM15 and |
153 | ZMM16-ZMM31 state are enabled. */ |
154 | if ((xcrlow & (bit_Opmask_state | bit_ZMM0_15_state |
155 | | bit_ZMM16_31_state)) == |
156 | (bit_Opmask_state | bit_ZMM0_15_state | bit_ZMM16_31_state)) |
157 | { |
158 | /* Determine if AVX512F is usable. */ |
159 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512F)) |
160 | { |
161 | cpu_features->feature[index_arch_AVX512F_Usable] |
162 | |= bit_arch_AVX512F_Usable; |
163 | /* Determine if AVX512CD is usable. */ |
164 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512CD)) |
165 | cpu_features->feature[index_arch_AVX512CD_Usable] |
166 | |= bit_arch_AVX512CD_Usable; |
167 | /* Determine if AVX512ER is usable. */ |
168 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
169 | cpu_features->feature[index_arch_AVX512ER_Usable] |
170 | |= bit_arch_AVX512ER_Usable; |
171 | /* Determine if AVX512PF is usable. */ |
172 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512PF)) |
173 | cpu_features->feature[index_arch_AVX512PF_Usable] |
174 | |= bit_arch_AVX512PF_Usable; |
175 | /* Determine if AVX512VL is usable. */ |
176 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512VL)) |
177 | cpu_features->feature[index_arch_AVX512VL_Usable] |
178 | |= bit_arch_AVX512VL_Usable; |
179 | /* Determine if AVX512DQ is usable. */ |
180 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512DQ)) |
181 | cpu_features->feature[index_arch_AVX512DQ_Usable] |
182 | |= bit_arch_AVX512DQ_Usable; |
183 | /* Determine if AVX512BW is usable. */ |
184 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512BW)) |
185 | cpu_features->feature[index_arch_AVX512BW_Usable] |
186 | |= bit_arch_AVX512BW_Usable; |
187 | /* Determine if AVX512_4FMAPS is usable. */ |
188 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_4FMAPS)) |
189 | cpu_features->feature[index_arch_AVX512_4FMAPS_Usable] |
190 | |= bit_arch_AVX512_4FMAPS_Usable; |
191 | /* Determine if AVX512_4VNNIW is usable. */ |
192 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_4VNNIW)) |
193 | cpu_features->feature[index_arch_AVX512_4VNNIW_Usable] |
194 | |= bit_arch_AVX512_4VNNIW_Usable; |
195 | /* Determine if AVX512_BITALG is usable. */ |
196 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_BITALG)) |
197 | cpu_features->feature[index_arch_AVX512_BITALG_Usable] |
198 | |= bit_arch_AVX512_BITALG_Usable; |
199 | /* Determine if AVX512_IFMA is usable. */ |
200 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_IFMA)) |
201 | cpu_features->feature[index_arch_AVX512_IFMA_Usable] |
202 | |= bit_arch_AVX512_IFMA_Usable; |
203 | /* Determine if AVX512_VBMI is usable. */ |
204 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_VBMI)) |
205 | cpu_features->feature[index_arch_AVX512_VBMI_Usable] |
206 | |= bit_arch_AVX512_VBMI_Usable; |
207 | /* Determine if AVX512_VBMI2 is usable. */ |
208 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_VBMI2)) |
209 | cpu_features->feature[index_arch_AVX512_VBMI2_Usable] |
210 | |= bit_arch_AVX512_VBMI2_Usable; |
211 | /* Determine if is AVX512_VNNI usable. */ |
212 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_VNNI)) |
213 | cpu_features->feature[index_arch_AVX512_VNNI_Usable] |
214 | |= bit_arch_AVX512_VNNI_Usable; |
215 | /* Determine if AVX512_VPOPCNTDQ is usable. */ |
216 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512_VPOPCNTDQ)) |
217 | cpu_features->feature[index_arch_AVX512_VPOPCNTDQ_Usable] |
218 | |= bit_arch_AVX512_VPOPCNTDQ_Usable; |
219 | } |
220 | } |
221 | } |
222 | |
223 | /* For _dl_runtime_resolve, set xsave_state_size to xsave area |
224 | size + integer register save size and align it to 64 bytes. */ |
225 | if (cpu_features->basic.max_cpuid >= 0xd) |
226 | { |
227 | unsigned int eax, ebx, ecx, edx; |
228 | |
229 | __cpuid_count (0xd, 0, eax, ebx, ecx, edx); |
230 | if (ebx != 0) |
231 | { |
232 | unsigned int xsave_state_full_size |
233 | = ALIGN_UP (ebx + STATE_SAVE_OFFSET, 64); |
234 | |
235 | cpu_features->xsave_state_size |
236 | = xsave_state_full_size; |
237 | cpu_features->xsave_state_full_size |
238 | = xsave_state_full_size; |
239 | |
240 | /* Check if XSAVEC is available. */ |
241 | if (CPU_FEATURES_CPU_P (cpu_features, XSAVEC)) |
242 | { |
243 | unsigned int xstate_comp_offsets[32]; |
244 | unsigned int xstate_comp_sizes[32]; |
245 | unsigned int i; |
246 | |
247 | xstate_comp_offsets[0] = 0; |
248 | xstate_comp_offsets[1] = 160; |
249 | xstate_comp_offsets[2] = 576; |
250 | xstate_comp_sizes[0] = 160; |
251 | xstate_comp_sizes[1] = 256; |
252 | |
253 | for (i = 2; i < 32; i++) |
254 | { |
255 | if ((STATE_SAVE_MASK & (1 << i)) != 0) |
256 | { |
257 | __cpuid_count (0xd, i, eax, ebx, ecx, edx); |
258 | xstate_comp_sizes[i] = eax; |
259 | } |
260 | else |
261 | { |
262 | ecx = 0; |
263 | xstate_comp_sizes[i] = 0; |
264 | } |
265 | |
266 | if (i > 2) |
267 | { |
268 | xstate_comp_offsets[i] |
269 | = (xstate_comp_offsets[i - 1] |
270 | + xstate_comp_sizes[i -1]); |
271 | if ((ecx & (1 << 1)) != 0) |
272 | xstate_comp_offsets[i] |
273 | = ALIGN_UP (xstate_comp_offsets[i], 64); |
274 | } |
275 | } |
276 | |
277 | /* Use XSAVEC. */ |
278 | unsigned int size |
279 | = xstate_comp_offsets[31] + xstate_comp_sizes[31]; |
280 | if (size) |
281 | { |
282 | cpu_features->xsave_state_size |
283 | = ALIGN_UP (size + STATE_SAVE_OFFSET, 64); |
284 | cpu_features->feature[index_arch_XSAVEC_Usable] |
285 | |= bit_arch_XSAVEC_Usable; |
286 | } |
287 | } |
288 | } |
289 | } |
290 | } |
291 | } |
292 | |
293 | _Static_assert (((index_arch_Fast_Unaligned_Load |
294 | == index_arch_Fast_Unaligned_Copy) |
295 | && (index_arch_Fast_Unaligned_Load |
296 | == index_arch_Prefer_PMINUB_for_stringop) |
297 | && (index_arch_Fast_Unaligned_Load |
298 | == index_arch_Slow_SSE4_2) |
299 | && (index_arch_Fast_Unaligned_Load |
300 | == index_arch_Fast_Rep_String) |
301 | && (index_arch_Fast_Unaligned_Load |
302 | == index_arch_Fast_Copy_Backward)), |
303 | "Incorrect index_arch_Fast_Unaligned_Load" ); |
304 | |
305 | static inline void |
306 | init_cpu_features (struct cpu_features *cpu_features) |
307 | { |
308 | unsigned int ebx, ecx, edx; |
309 | unsigned int family = 0; |
310 | unsigned int model = 0; |
311 | unsigned int stepping = 0; |
312 | enum cpu_features_kind kind; |
313 | |
314 | #if !HAS_CPUID |
315 | if (__get_cpuid_max (0, 0) == 0) |
316 | { |
317 | kind = arch_kind_other; |
318 | goto no_cpuid; |
319 | } |
320 | #endif |
321 | |
322 | __cpuid (0, cpu_features->basic.max_cpuid, ebx, ecx, edx); |
323 | |
324 | /* This spells out "GenuineIntel". */ |
325 | if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69) |
326 | { |
327 | unsigned int extended_model; |
328 | |
329 | kind = arch_kind_intel; |
330 | |
331 | get_common_indices (cpu_features, &family, &model, &extended_model, |
332 | &stepping); |
333 | |
334 | get_extended_indices (cpu_features); |
335 | |
336 | if (family == 0x06) |
337 | { |
338 | model += extended_model; |
339 | switch (model) |
340 | { |
341 | case 0x1c: |
342 | case 0x26: |
343 | /* BSF is slow on Atom. */ |
344 | cpu_features->feature[index_arch_Slow_BSF] |
345 | |= bit_arch_Slow_BSF; |
346 | break; |
347 | |
348 | case 0x57: |
349 | /* Knights Landing. Enable Silvermont optimizations. */ |
350 | |
351 | case 0x5c: |
352 | case 0x5f: |
353 | /* Unaligned load versions are faster than SSSE3 |
354 | on Goldmont. */ |
355 | |
356 | case 0x4c: |
357 | /* Airmont is a die shrink of Silvermont. */ |
358 | |
359 | case 0x37: |
360 | case 0x4a: |
361 | case 0x4d: |
362 | case 0x5a: |
363 | case 0x5d: |
364 | /* Unaligned load versions are faster than SSSE3 |
365 | on Silvermont. */ |
366 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
367 | |= (bit_arch_Fast_Unaligned_Load |
368 | | bit_arch_Fast_Unaligned_Copy |
369 | | bit_arch_Prefer_PMINUB_for_stringop |
370 | | bit_arch_Slow_SSE4_2); |
371 | break; |
372 | |
373 | default: |
374 | /* Unknown family 0x06 processors. Assuming this is one |
375 | of Core i3/i5/i7 processors if AVX is available. */ |
376 | if (!CPU_FEATURES_CPU_P (cpu_features, AVX)) |
377 | break; |
378 | |
379 | case 0x1a: |
380 | case 0x1e: |
381 | case 0x1f: |
382 | case 0x25: |
383 | case 0x2c: |
384 | case 0x2e: |
385 | case 0x2f: |
386 | /* Rep string instructions, unaligned load, unaligned copy, |
387 | and pminub are fast on Intel Core i3, i5 and i7. */ |
388 | cpu_features->feature[index_arch_Fast_Rep_String] |
389 | |= (bit_arch_Fast_Rep_String |
390 | | bit_arch_Fast_Unaligned_Load |
391 | | bit_arch_Fast_Unaligned_Copy |
392 | | bit_arch_Prefer_PMINUB_for_stringop); |
393 | break; |
394 | } |
395 | |
396 | /* Disable TSX on some Haswell processors to avoid TSX on kernels that |
397 | weren't updated with the latest microcode package (which disables |
398 | broken feature by default). */ |
399 | switch (model) |
400 | { |
401 | case 0x3f: |
402 | /* Xeon E7 v3 with stepping >= 4 has working TSX. */ |
403 | if (stepping >= 4) |
404 | break; |
405 | case 0x3c: |
406 | case 0x45: |
407 | case 0x46: |
408 | /* Disable Intel TSX on Haswell processors (except Xeon E7 v3 |
409 | with stepping >= 4) to avoid TSX on kernels that weren't |
410 | updated with the latest microcode package (which disables |
411 | broken feature by default). */ |
412 | cpu_features->cpuid[index_cpu_RTM].reg_RTM &= ~bit_cpu_RTM; |
413 | break; |
414 | } |
415 | } |
416 | |
417 | |
418 | /* Since AVX512ER is unique to Xeon Phi, set Prefer_No_VZEROUPPER |
419 | if AVX512ER is available. Don't use AVX512 to avoid lower CPU |
420 | frequency if AVX512ER isn't available. */ |
421 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
422 | cpu_features->feature[index_arch_Prefer_No_VZEROUPPER] |
423 | |= bit_arch_Prefer_No_VZEROUPPER; |
424 | else |
425 | cpu_features->feature[index_arch_Prefer_No_AVX512] |
426 | |= bit_arch_Prefer_No_AVX512; |
427 | } |
428 | /* This spells out "AuthenticAMD" or "HygonGenuine". */ |
429 | else if ((ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65) |
430 | || (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)) |
431 | { |
432 | unsigned int extended_model; |
433 | |
434 | kind = arch_kind_amd; |
435 | |
436 | get_common_indices (cpu_features, &family, &model, &extended_model, |
437 | &stepping); |
438 | |
439 | get_extended_indices (cpu_features); |
440 | |
441 | ecx = cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx; |
442 | |
443 | if (HAS_ARCH_FEATURE (AVX_Usable)) |
444 | { |
445 | /* Since the FMA4 bit is in COMMON_CPUID_INDEX_80000001 and |
446 | FMA4 requires AVX, determine if FMA4 is usable here. */ |
447 | if (CPU_FEATURES_CPU_P (cpu_features, FMA4)) |
448 | cpu_features->feature[index_arch_FMA4_Usable] |
449 | |= bit_arch_FMA4_Usable; |
450 | } |
451 | |
452 | if (family == 0x15) |
453 | { |
454 | /* "Excavator" */ |
455 | if (model >= 0x60 && model <= 0x7f) |
456 | { |
457 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
458 | |= (bit_arch_Fast_Unaligned_Load |
459 | | bit_arch_Fast_Copy_Backward); |
460 | |
461 | /* Unaligned AVX loads are slower.*/ |
462 | cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load] |
463 | &= ~bit_arch_AVX_Fast_Unaligned_Load; |
464 | } |
465 | } |
466 | } |
467 | else |
468 | { |
469 | kind = arch_kind_other; |
470 | get_common_indices (cpu_features, NULL, NULL, NULL, NULL); |
471 | } |
472 | |
473 | /* Support i586 if CX8 is available. */ |
474 | if (CPU_FEATURES_CPU_P (cpu_features, CX8)) |
475 | cpu_features->feature[index_arch_I586] |= bit_arch_I586; |
476 | |
477 | /* Support i686 if CMOV is available. */ |
478 | if (CPU_FEATURES_CPU_P (cpu_features, CMOV)) |
479 | cpu_features->feature[index_arch_I686] |= bit_arch_I686; |
480 | |
481 | #if !HAS_CPUID |
482 | no_cpuid: |
483 | #endif |
484 | |
485 | cpu_features->basic.kind = kind; |
486 | cpu_features->basic.family = family; |
487 | cpu_features->basic.model = model; |
488 | cpu_features->basic.stepping = stepping; |
489 | |
490 | #if HAVE_TUNABLES |
491 | TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps)); |
492 | cpu_features->non_temporal_threshold |
493 | = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL); |
494 | cpu_features->data_cache_size |
495 | = TUNABLE_GET (x86_data_cache_size, long int, NULL); |
496 | cpu_features->shared_cache_size |
497 | = TUNABLE_GET (x86_shared_cache_size, long int, NULL); |
498 | #endif |
499 | |
500 | /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86. */ |
501 | #if !HAVE_TUNABLES && defined SHARED |
502 | /* The glibc.cpu.hwcap_mask tunable is initialized already, so no need to do |
503 | this. */ |
504 | GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT; |
505 | #endif |
506 | |
507 | #ifdef __x86_64__ |
508 | GLRO(dl_hwcap) = HWCAP_X86_64; |
509 | if (cpu_features->basic.kind == arch_kind_intel) |
510 | { |
511 | const char *platform = NULL; |
512 | |
513 | if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable) |
514 | && CPU_FEATURES_CPU_P (cpu_features, AVX512CD)) |
515 | { |
516 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
517 | { |
518 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512PF)) |
519 | platform = "xeon_phi" ; |
520 | } |
521 | else |
522 | { |
523 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512BW) |
524 | && CPU_FEATURES_CPU_P (cpu_features, AVX512DQ) |
525 | && CPU_FEATURES_CPU_P (cpu_features, AVX512VL)) |
526 | GLRO(dl_hwcap) |= HWCAP_X86_AVX512_1; |
527 | } |
528 | } |
529 | |
530 | if (platform == NULL |
531 | && CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable) |
532 | && CPU_FEATURES_ARCH_P (cpu_features, FMA_Usable) |
533 | && CPU_FEATURES_CPU_P (cpu_features, BMI1) |
534 | && CPU_FEATURES_CPU_P (cpu_features, BMI2) |
535 | && CPU_FEATURES_CPU_P (cpu_features, LZCNT) |
536 | && CPU_FEATURES_CPU_P (cpu_features, MOVBE) |
537 | && CPU_FEATURES_CPU_P (cpu_features, POPCNT)) |
538 | platform = "haswell" ; |
539 | |
540 | if (platform != NULL) |
541 | GLRO(dl_platform) = platform; |
542 | } |
543 | #else |
544 | GLRO(dl_hwcap) = 0; |
545 | if (CPU_FEATURES_CPU_P (cpu_features, SSE2)) |
546 | GLRO(dl_hwcap) |= HWCAP_X86_SSE2; |
547 | |
548 | if (CPU_FEATURES_ARCH_P (cpu_features, I686)) |
549 | GLRO(dl_platform) = "i686" ; |
550 | else if (CPU_FEATURES_ARCH_P (cpu_features, I586)) |
551 | GLRO(dl_platform) = "i586" ; |
552 | #endif |
553 | |
554 | #if CET_ENABLED |
555 | # if HAVE_TUNABLES |
556 | TUNABLE_GET (x86_ibt, tunable_val_t *, |
557 | TUNABLE_CALLBACK (set_x86_ibt)); |
558 | TUNABLE_GET (x86_shstk, tunable_val_t *, |
559 | TUNABLE_CALLBACK (set_x86_shstk)); |
560 | # endif |
561 | |
562 | /* Check CET status. */ |
563 | unsigned int cet_status = get_cet_status (); |
564 | |
565 | if (cet_status) |
566 | { |
567 | GL(dl_x86_feature_1)[0] = cet_status; |
568 | |
569 | # ifndef SHARED |
570 | /* Check if IBT and SHSTK are enabled by kernel. */ |
571 | if ((cet_status & GNU_PROPERTY_X86_FEATURE_1_IBT) |
572 | || (cet_status & GNU_PROPERTY_X86_FEATURE_1_SHSTK)) |
573 | { |
574 | /* Disable IBT and/or SHSTK if they are enabled by kernel, but |
575 | disabled by environment variable: |
576 | |
577 | GLIBC_TUNABLES=glibc.cpu.hwcaps=-IBT,-SHSTK |
578 | */ |
579 | unsigned int cet_feature = 0; |
580 | if (!HAS_CPU_FEATURE (IBT)) |
581 | cet_feature |= GNU_PROPERTY_X86_FEATURE_1_IBT; |
582 | if (!HAS_CPU_FEATURE (SHSTK)) |
583 | cet_feature |= GNU_PROPERTY_X86_FEATURE_1_SHSTK; |
584 | |
585 | if (cet_feature) |
586 | { |
587 | int res = dl_cet_disable_cet (cet_feature); |
588 | |
589 | /* Clear the disabled bits in dl_x86_feature_1. */ |
590 | if (res == 0) |
591 | GL(dl_x86_feature_1)[0] &= ~cet_feature; |
592 | } |
593 | |
594 | /* Lock CET if IBT or SHSTK is enabled in executable. Don't |
595 | lock CET if SHSTK is enabled permissively. */ |
596 | if (((GL(dl_x86_feature_1)[1] >> CET_MAX) |
597 | & ((1 << CET_MAX) - 1)) |
598 | != CET_PERMISSIVE) |
599 | dl_cet_lock_cet (); |
600 | } |
601 | # endif |
602 | } |
603 | #endif |
604 | } |
605 | |