1 | /* Initialize CPU feature data. |
2 | This file is part of the GNU C Library. |
3 | Copyright (C) 2008-2017 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 | |
23 | #if HAVE_TUNABLES |
24 | # define TUNABLE_NAMESPACE tune |
25 | # include <unistd.h> /* Get STDOUT_FILENO for _dl_printf. */ |
26 | # include <elf/dl-tunables.h> |
27 | |
28 | extern void TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *) |
29 | attribute_hidden; |
30 | #endif |
31 | |
32 | static void |
33 | get_common_indeces (struct cpu_features *cpu_features, |
34 | unsigned int *family, unsigned int *model, |
35 | unsigned int *extended_model, unsigned int *stepping) |
36 | { |
37 | if (family) |
38 | { |
39 | unsigned int eax; |
40 | __cpuid (1, eax, cpu_features->cpuid[COMMON_CPUID_INDEX_1].ebx, |
41 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx, |
42 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].edx); |
43 | cpu_features->cpuid[COMMON_CPUID_INDEX_1].eax = eax; |
44 | *family = (eax >> 8) & 0x0f; |
45 | *model = (eax >> 4) & 0x0f; |
46 | *extended_model = (eax >> 12) & 0xf0; |
47 | *stepping = eax & 0x0f; |
48 | if (*family == 0x0f) |
49 | { |
50 | *family += (eax >> 20) & 0xff; |
51 | *model += *extended_model; |
52 | } |
53 | } |
54 | |
55 | if (cpu_features->max_cpuid >= 7) |
56 | __cpuid_count (7, 0, |
57 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].eax, |
58 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ebx, |
59 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].ecx, |
60 | cpu_features->cpuid[COMMON_CPUID_INDEX_7].edx); |
61 | |
62 | /* Can we call xgetbv? */ |
63 | if (CPU_FEATURES_CPU_P (cpu_features, OSXSAVE)) |
64 | { |
65 | unsigned int xcrlow; |
66 | unsigned int xcrhigh; |
67 | asm ("xgetbv" : "=a" (xcrlow), "=d" (xcrhigh) : "c" (0)); |
68 | /* Is YMM and XMM state usable? */ |
69 | if ((xcrlow & (bit_YMM_state | bit_XMM_state)) == |
70 | (bit_YMM_state | bit_XMM_state)) |
71 | { |
72 | /* Determine if AVX is usable. */ |
73 | if (CPU_FEATURES_CPU_P (cpu_features, AVX)) |
74 | { |
75 | cpu_features->feature[index_arch_AVX_Usable] |
76 | |= bit_arch_AVX_Usable; |
77 | /* The following features depend on AVX being usable. */ |
78 | /* Determine if AVX2 is usable. */ |
79 | if (CPU_FEATURES_CPU_P (cpu_features, AVX2)) |
80 | cpu_features->feature[index_arch_AVX2_Usable] |
81 | |= bit_arch_AVX2_Usable; |
82 | /* Determine if FMA is usable. */ |
83 | if (CPU_FEATURES_CPU_P (cpu_features, FMA)) |
84 | cpu_features->feature[index_arch_FMA_Usable] |
85 | |= bit_arch_FMA_Usable; |
86 | } |
87 | |
88 | /* Check if OPMASK state, upper 256-bit of ZMM0-ZMM15 and |
89 | ZMM16-ZMM31 state are enabled. */ |
90 | if ((xcrlow & (bit_Opmask_state | bit_ZMM0_15_state |
91 | | bit_ZMM16_31_state)) == |
92 | (bit_Opmask_state | bit_ZMM0_15_state | bit_ZMM16_31_state)) |
93 | { |
94 | /* Determine if AVX512F is usable. */ |
95 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512F)) |
96 | { |
97 | cpu_features->feature[index_arch_AVX512F_Usable] |
98 | |= bit_arch_AVX512F_Usable; |
99 | /* Determine if AVX512DQ is usable. */ |
100 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512DQ)) |
101 | cpu_features->feature[index_arch_AVX512DQ_Usable] |
102 | |= bit_arch_AVX512DQ_Usable; |
103 | } |
104 | } |
105 | } |
106 | } |
107 | } |
108 | |
109 | static inline void |
110 | init_cpu_features (struct cpu_features *cpu_features) |
111 | { |
112 | unsigned int ebx, ecx, edx; |
113 | unsigned int family = 0; |
114 | unsigned int model = 0; |
115 | enum cpu_features_kind kind; |
116 | |
117 | #if !HAS_CPUID |
118 | if (__get_cpuid_max (0, 0) == 0) |
119 | { |
120 | kind = arch_kind_other; |
121 | goto no_cpuid; |
122 | } |
123 | #endif |
124 | |
125 | __cpuid (0, cpu_features->max_cpuid, ebx, ecx, edx); |
126 | |
127 | /* This spells out "GenuineIntel". */ |
128 | if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69) |
129 | { |
130 | unsigned int extended_model, stepping; |
131 | |
132 | kind = arch_kind_intel; |
133 | |
134 | get_common_indeces (cpu_features, &family, &model, &extended_model, |
135 | &stepping); |
136 | |
137 | if (family == 0x06) |
138 | { |
139 | model += extended_model; |
140 | switch (model) |
141 | { |
142 | case 0x1c: |
143 | case 0x26: |
144 | /* BSF is slow on Atom. */ |
145 | cpu_features->feature[index_arch_Slow_BSF] |
146 | |= bit_arch_Slow_BSF; |
147 | break; |
148 | |
149 | case 0x57: |
150 | /* Knights Landing. Enable Silvermont optimizations. */ |
151 | |
152 | case 0x5c: |
153 | case 0x5f: |
154 | /* Unaligned load versions are faster than SSSE3 |
155 | on Goldmont. */ |
156 | |
157 | case 0x4c: |
158 | /* Airmont is a die shrink of Silvermont. */ |
159 | |
160 | case 0x37: |
161 | case 0x4a: |
162 | case 0x4d: |
163 | case 0x5a: |
164 | case 0x5d: |
165 | /* Unaligned load versions are faster than SSSE3 |
166 | on Silvermont. */ |
167 | #if index_arch_Fast_Unaligned_Load != index_arch_Prefer_PMINUB_for_stringop |
168 | # error index_arch_Fast_Unaligned_Load != index_arch_Prefer_PMINUB_for_stringop |
169 | #endif |
170 | #if index_arch_Fast_Unaligned_Load != index_arch_Slow_SSE4_2 |
171 | # error index_arch_Fast_Unaligned_Load != index_arch_Slow_SSE4_2 |
172 | #endif |
173 | #if index_arch_Fast_Unaligned_Load != index_arch_Fast_Unaligned_Copy |
174 | # error index_arch_Fast_Unaligned_Load != index_arch_Fast_Unaligned_Copy |
175 | #endif |
176 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
177 | |= (bit_arch_Fast_Unaligned_Load |
178 | | bit_arch_Fast_Unaligned_Copy |
179 | | bit_arch_Prefer_PMINUB_for_stringop |
180 | | bit_arch_Slow_SSE4_2); |
181 | break; |
182 | |
183 | default: |
184 | /* Unknown family 0x06 processors. Assuming this is one |
185 | of Core i3/i5/i7 processors if AVX is available. */ |
186 | if (!CPU_FEATURES_CPU_P (cpu_features, AVX)) |
187 | break; |
188 | |
189 | case 0x1a: |
190 | case 0x1e: |
191 | case 0x1f: |
192 | case 0x25: |
193 | case 0x2c: |
194 | case 0x2e: |
195 | case 0x2f: |
196 | /* Rep string instructions, unaligned load, unaligned copy, |
197 | and pminub are fast on Intel Core i3, i5 and i7. */ |
198 | #if index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Load |
199 | # error index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Load |
200 | #endif |
201 | #if index_arch_Fast_Rep_String != index_arch_Prefer_PMINUB_for_stringop |
202 | # error index_arch_Fast_Rep_String != index_arch_Prefer_PMINUB_for_stringop |
203 | #endif |
204 | #if index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Copy |
205 | # error index_arch_Fast_Rep_String != index_arch_Fast_Unaligned_Copy |
206 | #endif |
207 | cpu_features->feature[index_arch_Fast_Rep_String] |
208 | |= (bit_arch_Fast_Rep_String |
209 | | bit_arch_Fast_Unaligned_Load |
210 | | bit_arch_Fast_Unaligned_Copy |
211 | | bit_arch_Prefer_PMINUB_for_stringop); |
212 | break; |
213 | |
214 | case 0x3f: |
215 | /* Xeon E7 v3 with stepping >= 4 has working TSX. */ |
216 | if (stepping >= 4) |
217 | break; |
218 | case 0x3c: |
219 | case 0x45: |
220 | case 0x46: |
221 | /* Disable Intel TSX on Haswell processors (except Xeon E7 v3 |
222 | with stepping >= 4) to avoid TSX on kernels that weren't |
223 | updated with the latest microcode package (which disables |
224 | broken feature by default). */ |
225 | cpu_features->cpuid[index_cpu_RTM].reg_RTM &= ~bit_cpu_RTM; |
226 | break; |
227 | } |
228 | } |
229 | |
230 | /* Unaligned load with 256-bit AVX registers are faster on |
231 | Intel processors with AVX2. */ |
232 | if (CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable)) |
233 | cpu_features->feature[index_arch_AVX_Fast_Unaligned_Load] |
234 | |= bit_arch_AVX_Fast_Unaligned_Load; |
235 | |
236 | /* Since AVX512ER is unique to Xeon Phi, set Prefer_No_VZEROUPPER |
237 | if AVX512ER is available. Don't use AVX512 to avoid lower CPU |
238 | frequency if AVX512ER isn't available. */ |
239 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
240 | cpu_features->feature[index_arch_Prefer_No_VZEROUPPER] |
241 | |= bit_arch_Prefer_No_VZEROUPPER; |
242 | else |
243 | cpu_features->feature[index_arch_Prefer_No_AVX512] |
244 | |= bit_arch_Prefer_No_AVX512; |
245 | |
246 | /* To avoid SSE transition penalty, use _dl_runtime_resolve_slow. |
247 | If XGETBV suports ECX == 1, use _dl_runtime_resolve_opt. */ |
248 | cpu_features->feature[index_arch_Use_dl_runtime_resolve_slow] |
249 | |= bit_arch_Use_dl_runtime_resolve_slow; |
250 | if (cpu_features->max_cpuid >= 0xd) |
251 | { |
252 | unsigned int eax; |
253 | |
254 | __cpuid_count (0xd, 1, eax, ebx, ecx, edx); |
255 | if ((eax & (1 << 2)) != 0) |
256 | cpu_features->feature[index_arch_Use_dl_runtime_resolve_opt] |
257 | |= bit_arch_Use_dl_runtime_resolve_opt; |
258 | } |
259 | } |
260 | /* This spells out "AuthenticAMD". */ |
261 | else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65) |
262 | { |
263 | unsigned int extended_model, stepping; |
264 | |
265 | kind = arch_kind_amd; |
266 | |
267 | get_common_indeces (cpu_features, &family, &model, &extended_model, |
268 | &stepping); |
269 | |
270 | ecx = cpu_features->cpuid[COMMON_CPUID_INDEX_1].ecx; |
271 | |
272 | unsigned int eax; |
273 | __cpuid (0x80000000, eax, ebx, ecx, edx); |
274 | if (eax >= 0x80000001) |
275 | __cpuid (0x80000001, |
276 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].eax, |
277 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].ebx, |
278 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].ecx, |
279 | cpu_features->cpuid[COMMON_CPUID_INDEX_80000001].edx); |
280 | |
281 | if (HAS_ARCH_FEATURE (AVX_Usable)) |
282 | { |
283 | /* Since the FMA4 bit is in COMMON_CPUID_INDEX_80000001 and |
284 | FMA4 requires AVX, determine if FMA4 is usable here. */ |
285 | if (CPU_FEATURES_CPU_P (cpu_features, FMA4)) |
286 | cpu_features->feature[index_arch_FMA4_Usable] |
287 | |= bit_arch_FMA4_Usable; |
288 | } |
289 | |
290 | if (family == 0x15) |
291 | { |
292 | #if index_arch_Fast_Unaligned_Load != index_arch_Fast_Copy_Backward |
293 | # error index_arch_Fast_Unaligned_Load != index_arch_Fast_Copy_Backward |
294 | #endif |
295 | /* "Excavator" */ |
296 | if (model >= 0x60 && model <= 0x7f) |
297 | cpu_features->feature[index_arch_Fast_Unaligned_Load] |
298 | |= (bit_arch_Fast_Unaligned_Load |
299 | | bit_arch_Fast_Copy_Backward); |
300 | } |
301 | } |
302 | else |
303 | { |
304 | kind = arch_kind_other; |
305 | get_common_indeces (cpu_features, NULL, NULL, NULL, NULL); |
306 | } |
307 | |
308 | /* Support i586 if CX8 is available. */ |
309 | if (CPU_FEATURES_CPU_P (cpu_features, CX8)) |
310 | cpu_features->feature[index_arch_I586] |= bit_arch_I586; |
311 | |
312 | /* Support i686 if CMOV is available. */ |
313 | if (CPU_FEATURES_CPU_P (cpu_features, CMOV)) |
314 | cpu_features->feature[index_arch_I686] |= bit_arch_I686; |
315 | |
316 | #if !HAS_CPUID |
317 | no_cpuid: |
318 | #endif |
319 | |
320 | cpu_features->family = family; |
321 | cpu_features->model = model; |
322 | cpu_features->kind = kind; |
323 | |
324 | #if HAVE_TUNABLES |
325 | TUNABLE_GET (hwcaps, tunable_val_t *, TUNABLE_CALLBACK (set_hwcaps)); |
326 | cpu_features->non_temporal_threshold |
327 | = TUNABLE_GET (x86_non_temporal_threshold, long int, NULL); |
328 | cpu_features->data_cache_size |
329 | = TUNABLE_GET (x86_data_cache_size, long int, NULL); |
330 | cpu_features->shared_cache_size |
331 | = TUNABLE_GET (x86_shared_cache_size, long int, NULL); |
332 | #endif |
333 | |
334 | /* Reuse dl_platform, dl_hwcap and dl_hwcap_mask for x86. */ |
335 | GLRO(dl_platform) = NULL; |
336 | GLRO(dl_hwcap) = 0; |
337 | #if !HAVE_TUNABLES && defined SHARED |
338 | /* The glibc.tune.hwcap_mask tunable is initialized already, so no need to do |
339 | this. */ |
340 | GLRO(dl_hwcap_mask) = HWCAP_IMPORTANT; |
341 | #endif |
342 | |
343 | #ifdef __x86_64__ |
344 | if (cpu_features->kind == arch_kind_intel) |
345 | { |
346 | if (CPU_FEATURES_ARCH_P (cpu_features, AVX512F_Usable) |
347 | && CPU_FEATURES_CPU_P (cpu_features, AVX512CD)) |
348 | { |
349 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512ER)) |
350 | { |
351 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512PF)) |
352 | GLRO(dl_platform) = "xeon_phi" ; |
353 | } |
354 | else |
355 | { |
356 | if (CPU_FEATURES_CPU_P (cpu_features, AVX512BW) |
357 | && CPU_FEATURES_CPU_P (cpu_features, AVX512DQ) |
358 | && CPU_FEATURES_CPU_P (cpu_features, AVX512VL)) |
359 | GLRO(dl_hwcap) |= HWCAP_X86_AVX512_1; |
360 | } |
361 | } |
362 | |
363 | if (GLRO(dl_platform) == NULL |
364 | && CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable) |
365 | && CPU_FEATURES_ARCH_P (cpu_features, FMA_Usable) |
366 | && CPU_FEATURES_CPU_P (cpu_features, BMI1) |
367 | && CPU_FEATURES_CPU_P (cpu_features, BMI2) |
368 | && CPU_FEATURES_CPU_P (cpu_features, LZCNT) |
369 | && CPU_FEATURES_CPU_P (cpu_features, MOVBE) |
370 | && CPU_FEATURES_CPU_P (cpu_features, POPCNT)) |
371 | GLRO(dl_platform) = "haswell" ; |
372 | } |
373 | #else |
374 | if (CPU_FEATURES_CPU_P (cpu_features, SSE2)) |
375 | GLRO(dl_hwcap) |= HWCAP_X86_SSE2; |
376 | |
377 | if (CPU_FEATURES_ARCH_P (cpu_features, I686)) |
378 | GLRO(dl_platform) = "i686" ; |
379 | else if (CPU_FEATURES_ARCH_P (cpu_features, I586)) |
380 | GLRO(dl_platform) = "i586" ; |
381 | #endif |
382 | } |
383 | |