1 | /* x86_64 cache info. |
2 | Copyright (C) 2003-2020 Free Software Foundation, Inc. |
3 | This file is part of the GNU C Library. |
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 | <https://www.gnu.org/licenses/>. */ |
18 | |
19 | #if IS_IN (libc) |
20 | |
21 | #include <assert.h> |
22 | #include <stdbool.h> |
23 | #include <stdlib.h> |
24 | #include <unistd.h> |
25 | #include <cpuid.h> |
26 | #include <init-arch.h> |
27 | |
28 | static const struct intel_02_cache_info |
29 | { |
30 | unsigned char idx; |
31 | unsigned char assoc; |
32 | unsigned char linesize; |
33 | unsigned char rel_name; |
34 | unsigned int size; |
35 | } intel_02_known [] = |
36 | { |
37 | #define M(sc) ((sc) - _SC_LEVEL1_ICACHE_SIZE) |
38 | { 0x06, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 8192 }, |
39 | { 0x08, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 16384 }, |
40 | { 0x09, 4, 32, M(_SC_LEVEL1_ICACHE_SIZE), 32768 }, |
41 | { 0x0a, 2, 32, M(_SC_LEVEL1_DCACHE_SIZE), 8192 }, |
42 | { 0x0c, 4, 32, M(_SC_LEVEL1_DCACHE_SIZE), 16384 }, |
43 | { 0x0d, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 }, |
44 | { 0x0e, 6, 64, M(_SC_LEVEL1_DCACHE_SIZE), 24576 }, |
45 | { 0x21, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
46 | { 0x22, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 524288 }, |
47 | { 0x23, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 }, |
48 | { 0x25, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 }, |
49 | { 0x29, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 }, |
50 | { 0x2c, 8, 64, M(_SC_LEVEL1_DCACHE_SIZE), 32768 }, |
51 | { 0x30, 8, 64, M(_SC_LEVEL1_ICACHE_SIZE), 32768 }, |
52 | { 0x39, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 }, |
53 | { 0x3a, 6, 64, M(_SC_LEVEL2_CACHE_SIZE), 196608 }, |
54 | { 0x3b, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 }, |
55 | { 0x3c, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
56 | { 0x3d, 6, 64, M(_SC_LEVEL2_CACHE_SIZE), 393216 }, |
57 | { 0x3e, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
58 | { 0x3f, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
59 | { 0x41, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 131072 }, |
60 | { 0x42, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
61 | { 0x43, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
62 | { 0x44, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 1048576 }, |
63 | { 0x45, 4, 32, M(_SC_LEVEL2_CACHE_SIZE), 2097152 }, |
64 | { 0x46, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 }, |
65 | { 0x47, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 }, |
66 | { 0x48, 12, 64, M(_SC_LEVEL2_CACHE_SIZE), 3145728 }, |
67 | { 0x49, 16, 64, M(_SC_LEVEL2_CACHE_SIZE), 4194304 }, |
68 | { 0x4a, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 6291456 }, |
69 | { 0x4b, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 }, |
70 | { 0x4c, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 }, |
71 | { 0x4d, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 16777216 }, |
72 | { 0x4e, 24, 64, M(_SC_LEVEL2_CACHE_SIZE), 6291456 }, |
73 | { 0x60, 8, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 }, |
74 | { 0x66, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 8192 }, |
75 | { 0x67, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 16384 }, |
76 | { 0x68, 4, 64, M(_SC_LEVEL1_DCACHE_SIZE), 32768 }, |
77 | { 0x78, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 }, |
78 | { 0x79, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 131072 }, |
79 | { 0x7a, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
80 | { 0x7b, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
81 | { 0x7c, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 }, |
82 | { 0x7d, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 2097152 }, |
83 | { 0x7f, 2, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
84 | { 0x80, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
85 | { 0x82, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 262144 }, |
86 | { 0x83, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
87 | { 0x84, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 1048576 }, |
88 | { 0x85, 8, 32, M(_SC_LEVEL2_CACHE_SIZE), 2097152 }, |
89 | { 0x86, 4, 64, M(_SC_LEVEL2_CACHE_SIZE), 524288 }, |
90 | { 0x87, 8, 64, M(_SC_LEVEL2_CACHE_SIZE), 1048576 }, |
91 | { 0xd0, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 524288 }, |
92 | { 0xd1, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 }, |
93 | { 0xd2, 4, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 }, |
94 | { 0xd6, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 1048576 }, |
95 | { 0xd7, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 }, |
96 | { 0xd8, 8, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 }, |
97 | { 0xdc, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 }, |
98 | { 0xdd, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 }, |
99 | { 0xde, 12, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 }, |
100 | { 0xe2, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 2097152 }, |
101 | { 0xe3, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 4194304 }, |
102 | { 0xe4, 16, 64, M(_SC_LEVEL3_CACHE_SIZE), 8388608 }, |
103 | { 0xea, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 12582912 }, |
104 | { 0xeb, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 18874368 }, |
105 | { 0xec, 24, 64, M(_SC_LEVEL3_CACHE_SIZE), 25165824 }, |
106 | }; |
107 | |
108 | #define nintel_02_known (sizeof (intel_02_known) / sizeof (intel_02_known [0])) |
109 | |
110 | static int |
111 | intel_02_known_compare (const void *p1, const void *p2) |
112 | { |
113 | const struct intel_02_cache_info *i1; |
114 | const struct intel_02_cache_info *i2; |
115 | |
116 | i1 = (const struct intel_02_cache_info *) p1; |
117 | i2 = (const struct intel_02_cache_info *) p2; |
118 | |
119 | if (i1->idx == i2->idx) |
120 | return 0; |
121 | |
122 | return i1->idx < i2->idx ? -1 : 1; |
123 | } |
124 | |
125 | |
126 | static long int |
127 | __attribute__ ((noinline)) |
128 | intel_check_word (int name, unsigned int value, bool *has_level_2, |
129 | bool *no_level_2_or_3, |
130 | const struct cpu_features *cpu_features) |
131 | { |
132 | if ((value & 0x80000000) != 0) |
133 | /* The register value is reserved. */ |
134 | return 0; |
135 | |
136 | /* Fold the name. The _SC_ constants are always in the order SIZE, |
137 | ASSOC, LINESIZE. */ |
138 | int folded_rel_name = (M(name) / 3) * 3; |
139 | |
140 | while (value != 0) |
141 | { |
142 | unsigned int byte = value & 0xff; |
143 | |
144 | if (byte == 0x40) |
145 | { |
146 | *no_level_2_or_3 = true; |
147 | |
148 | if (folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)) |
149 | /* No need to look further. */ |
150 | break; |
151 | } |
152 | else if (byte == 0xff) |
153 | { |
154 | /* CPUID leaf 0x4 contains all the information. We need to |
155 | iterate over it. */ |
156 | unsigned int eax; |
157 | unsigned int ebx; |
158 | unsigned int ecx; |
159 | unsigned int edx; |
160 | |
161 | unsigned int round = 0; |
162 | while (1) |
163 | { |
164 | __cpuid_count (4, round, eax, ebx, ecx, edx); |
165 | |
166 | enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f; |
167 | if (type == null) |
168 | /* That was the end. */ |
169 | break; |
170 | |
171 | unsigned int level = (eax >> 5) & 0x7; |
172 | |
173 | if ((level == 1 && type == data |
174 | && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE)) |
175 | || (level == 1 && type == inst |
176 | && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE)) |
177 | || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE)) |
178 | || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)) |
179 | || (level == 4 && folded_rel_name == M(_SC_LEVEL4_CACHE_SIZE))) |
180 | { |
181 | unsigned int offset = M(name) - folded_rel_name; |
182 | |
183 | if (offset == 0) |
184 | /* Cache size. */ |
185 | return (((ebx >> 22) + 1) |
186 | * (((ebx >> 12) & 0x3ff) + 1) |
187 | * ((ebx & 0xfff) + 1) |
188 | * (ecx + 1)); |
189 | if (offset == 1) |
190 | return (ebx >> 22) + 1; |
191 | |
192 | assert (offset == 2); |
193 | return (ebx & 0xfff) + 1; |
194 | } |
195 | |
196 | ++round; |
197 | } |
198 | /* There is no other cache information anywhere else. */ |
199 | break; |
200 | } |
201 | else |
202 | { |
203 | if (byte == 0x49 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE)) |
204 | { |
205 | /* Intel reused this value. For family 15, model 6 it |
206 | specifies the 3rd level cache. Otherwise the 2nd |
207 | level cache. */ |
208 | unsigned int family = cpu_features->basic.family; |
209 | unsigned int model = cpu_features->basic.model; |
210 | |
211 | if (family == 15 && model == 6) |
212 | { |
213 | /* The level 3 cache is encoded for this model like |
214 | the level 2 cache is for other models. Pretend |
215 | the caller asked for the level 2 cache. */ |
216 | name = (_SC_LEVEL2_CACHE_SIZE |
217 | + (name - _SC_LEVEL3_CACHE_SIZE)); |
218 | folded_rel_name = M(_SC_LEVEL2_CACHE_SIZE); |
219 | } |
220 | } |
221 | |
222 | struct intel_02_cache_info *found; |
223 | struct intel_02_cache_info search; |
224 | |
225 | search.idx = byte; |
226 | found = bsearch (&search, intel_02_known, nintel_02_known, |
227 | sizeof (intel_02_known[0]), intel_02_known_compare); |
228 | if (found != NULL) |
229 | { |
230 | if (found->rel_name == folded_rel_name) |
231 | { |
232 | unsigned int offset = M(name) - folded_rel_name; |
233 | |
234 | if (offset == 0) |
235 | /* Cache size. */ |
236 | return found->size; |
237 | if (offset == 1) |
238 | return found->assoc; |
239 | |
240 | assert (offset == 2); |
241 | return found->linesize; |
242 | } |
243 | |
244 | if (found->rel_name == M(_SC_LEVEL2_CACHE_SIZE)) |
245 | *has_level_2 = true; |
246 | } |
247 | } |
248 | |
249 | /* Next byte for the next round. */ |
250 | value >>= 8; |
251 | } |
252 | |
253 | /* Nothing found. */ |
254 | return 0; |
255 | } |
256 | |
257 | |
258 | static long int __attribute__ ((noinline)) |
259 | handle_intel (int name, const struct cpu_features *cpu_features) |
260 | { |
261 | unsigned int maxidx = cpu_features->basic.max_cpuid; |
262 | |
263 | /* Return -1 for older CPUs. */ |
264 | if (maxidx < 2) |
265 | return -1; |
266 | |
267 | /* OK, we can use the CPUID instruction to get all info about the |
268 | caches. */ |
269 | unsigned int cnt = 0; |
270 | unsigned int max = 1; |
271 | long int result = 0; |
272 | bool no_level_2_or_3 = false; |
273 | bool has_level_2 = false; |
274 | |
275 | while (cnt++ < max) |
276 | { |
277 | unsigned int eax; |
278 | unsigned int ebx; |
279 | unsigned int ecx; |
280 | unsigned int edx; |
281 | __cpuid (2, eax, ebx, ecx, edx); |
282 | |
283 | /* The low byte of EAX in the first round contain the number of |
284 | rounds we have to make. At least one, the one we are already |
285 | doing. */ |
286 | if (cnt == 1) |
287 | { |
288 | max = eax & 0xff; |
289 | eax &= 0xffffff00; |
290 | } |
291 | |
292 | /* Process the individual registers' value. */ |
293 | result = intel_check_word (name, eax, &has_level_2, |
294 | &no_level_2_or_3, cpu_features); |
295 | if (result != 0) |
296 | return result; |
297 | |
298 | result = intel_check_word (name, ebx, &has_level_2, |
299 | &no_level_2_or_3, cpu_features); |
300 | if (result != 0) |
301 | return result; |
302 | |
303 | result = intel_check_word (name, ecx, &has_level_2, |
304 | &no_level_2_or_3, cpu_features); |
305 | if (result != 0) |
306 | return result; |
307 | |
308 | result = intel_check_word (name, edx, &has_level_2, |
309 | &no_level_2_or_3, cpu_features); |
310 | if (result != 0) |
311 | return result; |
312 | } |
313 | |
314 | if (name >= _SC_LEVEL2_CACHE_SIZE && name <= _SC_LEVEL3_CACHE_LINESIZE |
315 | && no_level_2_or_3) |
316 | return -1; |
317 | |
318 | return 0; |
319 | } |
320 | |
321 | |
322 | static long int __attribute__ ((noinline)) |
323 | handle_amd (int name) |
324 | { |
325 | unsigned int eax; |
326 | unsigned int ebx; |
327 | unsigned int ecx; |
328 | unsigned int edx; |
329 | __cpuid (0x80000000, eax, ebx, ecx, edx); |
330 | |
331 | /* No level 4 cache (yet). */ |
332 | if (name > _SC_LEVEL3_CACHE_LINESIZE) |
333 | return 0; |
334 | |
335 | unsigned int fn = 0x80000005 + (name >= _SC_LEVEL2_CACHE_SIZE); |
336 | if (eax < fn) |
337 | return 0; |
338 | |
339 | __cpuid (fn, eax, ebx, ecx, edx); |
340 | |
341 | if (name < _SC_LEVEL1_DCACHE_SIZE) |
342 | { |
343 | name += _SC_LEVEL1_DCACHE_SIZE - _SC_LEVEL1_ICACHE_SIZE; |
344 | ecx = edx; |
345 | } |
346 | |
347 | switch (name) |
348 | { |
349 | case _SC_LEVEL1_DCACHE_SIZE: |
350 | return (ecx >> 14) & 0x3fc00; |
351 | |
352 | case _SC_LEVEL1_DCACHE_ASSOC: |
353 | ecx >>= 16; |
354 | if ((ecx & 0xff) == 0xff) |
355 | /* Fully associative. */ |
356 | return (ecx << 2) & 0x3fc00; |
357 | return ecx & 0xff; |
358 | |
359 | case _SC_LEVEL1_DCACHE_LINESIZE: |
360 | return ecx & 0xff; |
361 | |
362 | case _SC_LEVEL2_CACHE_SIZE: |
363 | return (ecx & 0xf000) == 0 ? 0 : (ecx >> 6) & 0x3fffc00; |
364 | |
365 | case _SC_LEVEL2_CACHE_ASSOC: |
366 | switch ((ecx >> 12) & 0xf) |
367 | { |
368 | case 0: |
369 | case 1: |
370 | case 2: |
371 | case 4: |
372 | return (ecx >> 12) & 0xf; |
373 | case 6: |
374 | return 8; |
375 | case 8: |
376 | return 16; |
377 | case 10: |
378 | return 32; |
379 | case 11: |
380 | return 48; |
381 | case 12: |
382 | return 64; |
383 | case 13: |
384 | return 96; |
385 | case 14: |
386 | return 128; |
387 | case 15: |
388 | return ((ecx >> 6) & 0x3fffc00) / (ecx & 0xff); |
389 | default: |
390 | return 0; |
391 | } |
392 | /* NOTREACHED */ |
393 | |
394 | case _SC_LEVEL2_CACHE_LINESIZE: |
395 | return (ecx & 0xf000) == 0 ? 0 : ecx & 0xff; |
396 | |
397 | case _SC_LEVEL3_CACHE_SIZE: |
398 | return (edx & 0xf000) == 0 ? 0 : (edx & 0x3ffc0000) << 1; |
399 | |
400 | case _SC_LEVEL3_CACHE_ASSOC: |
401 | switch ((edx >> 12) & 0xf) |
402 | { |
403 | case 0: |
404 | case 1: |
405 | case 2: |
406 | case 4: |
407 | return (edx >> 12) & 0xf; |
408 | case 6: |
409 | return 8; |
410 | case 8: |
411 | return 16; |
412 | case 10: |
413 | return 32; |
414 | case 11: |
415 | return 48; |
416 | case 12: |
417 | return 64; |
418 | case 13: |
419 | return 96; |
420 | case 14: |
421 | return 128; |
422 | case 15: |
423 | return ((edx & 0x3ffc0000) << 1) / (edx & 0xff); |
424 | default: |
425 | return 0; |
426 | } |
427 | /* NOTREACHED */ |
428 | |
429 | case _SC_LEVEL3_CACHE_LINESIZE: |
430 | return (edx & 0xf000) == 0 ? 0 : edx & 0xff; |
431 | |
432 | default: |
433 | assert (! "cannot happen" ); |
434 | } |
435 | return -1; |
436 | } |
437 | |
438 | |
439 | static long int __attribute__ ((noinline)) |
440 | handle_zhaoxin (int name) |
441 | { |
442 | unsigned int eax; |
443 | unsigned int ebx; |
444 | unsigned int ecx; |
445 | unsigned int edx; |
446 | |
447 | int folded_rel_name = (M(name) / 3) * 3; |
448 | |
449 | unsigned int round = 0; |
450 | while (1) |
451 | { |
452 | __cpuid_count (4, round, eax, ebx, ecx, edx); |
453 | |
454 | enum { null = 0, data = 1, inst = 2, uni = 3 } type = eax & 0x1f; |
455 | if (type == null) |
456 | break; |
457 | |
458 | unsigned int level = (eax >> 5) & 0x7; |
459 | |
460 | if ((level == 1 && type == data |
461 | && folded_rel_name == M(_SC_LEVEL1_DCACHE_SIZE)) |
462 | || (level == 1 && type == inst |
463 | && folded_rel_name == M(_SC_LEVEL1_ICACHE_SIZE)) |
464 | || (level == 2 && folded_rel_name == M(_SC_LEVEL2_CACHE_SIZE)) |
465 | || (level == 3 && folded_rel_name == M(_SC_LEVEL3_CACHE_SIZE))) |
466 | { |
467 | unsigned int offset = M(name) - folded_rel_name; |
468 | |
469 | if (offset == 0) |
470 | /* Cache size. */ |
471 | return (((ebx >> 22) + 1) |
472 | * (((ebx >> 12) & 0x3ff) + 1) |
473 | * ((ebx & 0xfff) + 1) |
474 | * (ecx + 1)); |
475 | if (offset == 1) |
476 | return (ebx >> 22) + 1; |
477 | |
478 | assert (offset == 2); |
479 | return (ebx & 0xfff) + 1; |
480 | } |
481 | |
482 | ++round; |
483 | } |
484 | |
485 | /* Nothing found. */ |
486 | return 0; |
487 | } |
488 | |
489 | |
490 | /* Get the value of the system variable NAME. */ |
491 | long int |
492 | attribute_hidden |
493 | __cache_sysconf (int name) |
494 | { |
495 | const struct cpu_features *cpu_features = __get_cpu_features (); |
496 | |
497 | if (cpu_features->basic.kind == arch_kind_intel) |
498 | return handle_intel (name, cpu_features); |
499 | |
500 | if (cpu_features->basic.kind == arch_kind_amd) |
501 | return handle_amd (name); |
502 | |
503 | if (cpu_features->basic.kind == arch_kind_zhaoxin) |
504 | return handle_zhaoxin (name); |
505 | |
506 | // XXX Fill in more vendors. |
507 | |
508 | /* CPU not known, we have no information. */ |
509 | return 0; |
510 | } |
511 | |
512 | |
513 | /* Data cache size for use in memory and string routines, typically |
514 | L1 size, rounded to multiple of 256 bytes. */ |
515 | long int __x86_data_cache_size_half attribute_hidden = 32 * 1024 / 2; |
516 | long int __x86_data_cache_size attribute_hidden = 32 * 1024; |
517 | /* Similar to __x86_data_cache_size_half, but not rounded. */ |
518 | long int __x86_raw_data_cache_size_half attribute_hidden = 32 * 1024 / 2; |
519 | /* Similar to __x86_data_cache_size, but not rounded. */ |
520 | long int __x86_raw_data_cache_size attribute_hidden = 32 * 1024; |
521 | /* Shared cache size for use in memory and string routines, typically |
522 | L2 or L3 size, rounded to multiple of 256 bytes. */ |
523 | long int __x86_shared_cache_size_half attribute_hidden = 1024 * 1024 / 2; |
524 | long int __x86_shared_cache_size attribute_hidden = 1024 * 1024; |
525 | /* Similar to __x86_shared_cache_size_half, but not rounded. */ |
526 | long int __x86_raw_shared_cache_size_half attribute_hidden = 1024 * 1024 / 2; |
527 | /* Similar to __x86_shared_cache_size, but not rounded. */ |
528 | long int __x86_raw_shared_cache_size attribute_hidden = 1024 * 1024; |
529 | |
530 | /* Threshold to use non temporal store. */ |
531 | long int __x86_shared_non_temporal_threshold attribute_hidden; |
532 | |
533 | /* Threshold to use Enhanced REP MOVSB. */ |
534 | long int __x86_rep_movsb_threshold attribute_hidden = 2048; |
535 | |
536 | /* Threshold to use Enhanced REP STOSB. */ |
537 | long int __x86_rep_stosb_threshold attribute_hidden = 2048; |
538 | |
539 | |
540 | static void |
541 | get_common_cache_info (long int *shared_ptr, unsigned int *threads_ptr, |
542 | long int core) |
543 | { |
544 | unsigned int eax; |
545 | unsigned int ebx; |
546 | unsigned int ecx; |
547 | unsigned int edx; |
548 | |
549 | /* Number of logical processors sharing L2 cache. */ |
550 | int threads_l2; |
551 | |
552 | /* Number of logical processors sharing L3 cache. */ |
553 | int threads_l3; |
554 | |
555 | const struct cpu_features *cpu_features = __get_cpu_features (); |
556 | int max_cpuid = cpu_features->basic.max_cpuid; |
557 | unsigned int family = cpu_features->basic.family; |
558 | unsigned int model = cpu_features->basic.model; |
559 | long int shared = *shared_ptr; |
560 | unsigned int threads = *threads_ptr; |
561 | bool inclusive_cache = true; |
562 | bool support_count_mask = true; |
563 | |
564 | /* Try L3 first. */ |
565 | unsigned int level = 3; |
566 | |
567 | if (cpu_features->basic.kind == arch_kind_zhaoxin && family == 6) |
568 | support_count_mask = false; |
569 | |
570 | if (shared <= 0) |
571 | { |
572 | /* Try L2 otherwise. */ |
573 | level = 2; |
574 | shared = core; |
575 | threads_l2 = 0; |
576 | threads_l3 = -1; |
577 | } |
578 | else |
579 | { |
580 | threads_l2 = 0; |
581 | threads_l3 = 0; |
582 | } |
583 | |
584 | /* A value of 0 for the HTT bit indicates there is only a single |
585 | logical processor. */ |
586 | if (CPU_FEATURE_USABLE (HTT)) |
587 | { |
588 | /* Figure out the number of logical threads that share the |
589 | highest cache level. */ |
590 | if (max_cpuid >= 4) |
591 | { |
592 | int i = 0; |
593 | |
594 | /* Query until cache level 2 and 3 are enumerated. */ |
595 | int check = 0x1 | (threads_l3 == 0) << 1; |
596 | do |
597 | { |
598 | __cpuid_count (4, i++, eax, ebx, ecx, edx); |
599 | |
600 | /* There seems to be a bug in at least some Pentium Ds |
601 | which sometimes fail to iterate all cache parameters. |
602 | Do not loop indefinitely here, stop in this case and |
603 | assume there is no such information. */ |
604 | if (cpu_features->basic.kind == arch_kind_intel |
605 | && (eax & 0x1f) == 0 ) |
606 | goto intel_bug_no_cache_info; |
607 | |
608 | switch ((eax >> 5) & 0x7) |
609 | { |
610 | default: |
611 | break; |
612 | case 2: |
613 | if ((check & 0x1)) |
614 | { |
615 | /* Get maximum number of logical processors |
616 | sharing L2 cache. */ |
617 | threads_l2 = (eax >> 14) & 0x3ff; |
618 | check &= ~0x1; |
619 | } |
620 | break; |
621 | case 3: |
622 | if ((check & (0x1 << 1))) |
623 | { |
624 | /* Get maximum number of logical processors |
625 | sharing L3 cache. */ |
626 | threads_l3 = (eax >> 14) & 0x3ff; |
627 | |
628 | /* Check if L2 and L3 caches are inclusive. */ |
629 | inclusive_cache = (edx & 0x2) != 0; |
630 | check &= ~(0x1 << 1); |
631 | } |
632 | break; |
633 | } |
634 | } |
635 | while (check); |
636 | |
637 | /* If max_cpuid >= 11, THREADS_L2/THREADS_L3 are the maximum |
638 | numbers of addressable IDs for logical processors sharing |
639 | the cache, instead of the maximum number of threads |
640 | sharing the cache. */ |
641 | if (max_cpuid >= 11 && support_count_mask) |
642 | { |
643 | /* Find the number of logical processors shipped in |
644 | one core and apply count mask. */ |
645 | i = 0; |
646 | |
647 | /* Count SMT only if there is L3 cache. Always count |
648 | core if there is no L3 cache. */ |
649 | int count = ((threads_l2 > 0 && level == 3) |
650 | | ((threads_l3 > 0 |
651 | || (threads_l2 > 0 && level == 2)) << 1)); |
652 | |
653 | while (count) |
654 | { |
655 | __cpuid_count (11, i++, eax, ebx, ecx, edx); |
656 | |
657 | int shipped = ebx & 0xff; |
658 | int type = ecx & 0xff00; |
659 | if (shipped == 0 || type == 0) |
660 | break; |
661 | else if (type == 0x100) |
662 | { |
663 | /* Count SMT. */ |
664 | if ((count & 0x1)) |
665 | { |
666 | int count_mask; |
667 | |
668 | /* Compute count mask. */ |
669 | asm ("bsr %1, %0" |
670 | : "=r" (count_mask) : "g" (threads_l2)); |
671 | count_mask = ~(-1 << (count_mask + 1)); |
672 | threads_l2 = (shipped - 1) & count_mask; |
673 | count &= ~0x1; |
674 | } |
675 | } |
676 | else if (type == 0x200) |
677 | { |
678 | /* Count core. */ |
679 | if ((count & (0x1 << 1))) |
680 | { |
681 | int count_mask; |
682 | int threads_core |
683 | = (level == 2 ? threads_l2 : threads_l3); |
684 | |
685 | /* Compute count mask. */ |
686 | asm ("bsr %1, %0" |
687 | : "=r" (count_mask) : "g" (threads_core)); |
688 | count_mask = ~(-1 << (count_mask + 1)); |
689 | threads_core = (shipped - 1) & count_mask; |
690 | if (level == 2) |
691 | threads_l2 = threads_core; |
692 | else |
693 | threads_l3 = threads_core; |
694 | count &= ~(0x1 << 1); |
695 | } |
696 | } |
697 | } |
698 | } |
699 | if (threads_l2 > 0) |
700 | threads_l2 += 1; |
701 | if (threads_l3 > 0) |
702 | threads_l3 += 1; |
703 | if (level == 2) |
704 | { |
705 | if (threads_l2) |
706 | { |
707 | threads = threads_l2; |
708 | if (cpu_features->basic.kind == arch_kind_intel |
709 | && threads > 2 |
710 | && family == 6) |
711 | switch (model) |
712 | { |
713 | case 0x37: |
714 | case 0x4a: |
715 | case 0x4d: |
716 | case 0x5a: |
717 | case 0x5d: |
718 | /* Silvermont has L2 cache shared by 2 cores. */ |
719 | threads = 2; |
720 | break; |
721 | default: |
722 | break; |
723 | } |
724 | } |
725 | } |
726 | else if (threads_l3) |
727 | threads = threads_l3; |
728 | } |
729 | else |
730 | { |
731 | intel_bug_no_cache_info: |
732 | /* Assume that all logical threads share the highest cache |
733 | level. */ |
734 | threads |
735 | = ((cpu_features->features[COMMON_CPUID_INDEX_1].cpuid.ebx |
736 | >> 16) & 0xff); |
737 | } |
738 | |
739 | /* Cap usage of highest cache level to the number of supported |
740 | threads. */ |
741 | if (shared > 0 && threads > 0) |
742 | shared /= threads; |
743 | } |
744 | |
745 | /* Account for non-inclusive L2 and L3 caches. */ |
746 | if (!inclusive_cache) |
747 | { |
748 | if (threads_l2 > 0) |
749 | core /= threads_l2; |
750 | shared += core; |
751 | } |
752 | |
753 | *shared_ptr = shared; |
754 | *threads_ptr = threads; |
755 | } |
756 | |
757 | |
758 | static void |
759 | __attribute__((constructor)) |
760 | init_cacheinfo (void) |
761 | { |
762 | /* Find out what brand of processor. */ |
763 | unsigned int ebx; |
764 | unsigned int ecx; |
765 | unsigned int edx; |
766 | int max_cpuid_ex; |
767 | long int data = -1; |
768 | long int shared = -1; |
769 | long int core; |
770 | unsigned int threads = 0; |
771 | const struct cpu_features *cpu_features = __get_cpu_features (); |
772 | |
773 | if (cpu_features->basic.kind == arch_kind_intel) |
774 | { |
775 | data = handle_intel (_SC_LEVEL1_DCACHE_SIZE, cpu_features); |
776 | core = handle_intel (_SC_LEVEL2_CACHE_SIZE, cpu_features); |
777 | shared = handle_intel (_SC_LEVEL3_CACHE_SIZE, cpu_features); |
778 | |
779 | get_common_cache_info (&shared, &threads, core); |
780 | } |
781 | else if (cpu_features->basic.kind == arch_kind_zhaoxin) |
782 | { |
783 | data = handle_zhaoxin (_SC_LEVEL1_DCACHE_SIZE); |
784 | core = handle_zhaoxin (_SC_LEVEL2_CACHE_SIZE); |
785 | shared = handle_zhaoxin (_SC_LEVEL3_CACHE_SIZE); |
786 | |
787 | get_common_cache_info (&shared, &threads, core); |
788 | } |
789 | else if (cpu_features->basic.kind == arch_kind_amd) |
790 | { |
791 | data = handle_amd (_SC_LEVEL1_DCACHE_SIZE); |
792 | long int core = handle_amd (_SC_LEVEL2_CACHE_SIZE); |
793 | shared = handle_amd (_SC_LEVEL3_CACHE_SIZE); |
794 | |
795 | /* Get maximum extended function. */ |
796 | __cpuid (0x80000000, max_cpuid_ex, ebx, ecx, edx); |
797 | |
798 | if (shared <= 0) |
799 | /* No shared L3 cache. All we have is the L2 cache. */ |
800 | shared = core; |
801 | else |
802 | { |
803 | /* Figure out the number of logical threads that share L3. */ |
804 | if (max_cpuid_ex >= 0x80000008) |
805 | { |
806 | /* Get width of APIC ID. */ |
807 | __cpuid (0x80000008, max_cpuid_ex, ebx, ecx, edx); |
808 | threads = 1 << ((ecx >> 12) & 0x0f); |
809 | } |
810 | |
811 | if (threads == 0) |
812 | { |
813 | /* If APIC ID width is not available, use logical |
814 | processor count. */ |
815 | __cpuid (0x00000001, max_cpuid_ex, ebx, ecx, edx); |
816 | |
817 | if ((edx & (1 << 28)) != 0) |
818 | threads = (ebx >> 16) & 0xff; |
819 | } |
820 | |
821 | /* Cap usage of highest cache level to the number of |
822 | supported threads. */ |
823 | if (threads > 0) |
824 | shared /= threads; |
825 | |
826 | /* Account for exclusive L2 and L3 caches. */ |
827 | shared += core; |
828 | } |
829 | } |
830 | |
831 | if (cpu_features->data_cache_size != 0) |
832 | data = cpu_features->data_cache_size; |
833 | |
834 | if (data > 0) |
835 | { |
836 | __x86_raw_data_cache_size_half = data / 2; |
837 | __x86_raw_data_cache_size = data; |
838 | /* Round data cache size to multiple of 256 bytes. */ |
839 | data = data & ~255L; |
840 | __x86_data_cache_size_half = data / 2; |
841 | __x86_data_cache_size = data; |
842 | } |
843 | |
844 | if (cpu_features->shared_cache_size != 0) |
845 | shared = cpu_features->shared_cache_size; |
846 | |
847 | if (shared > 0) |
848 | { |
849 | __x86_raw_shared_cache_size_half = shared / 2; |
850 | __x86_raw_shared_cache_size = shared; |
851 | /* Round shared cache size to multiple of 256 bytes. */ |
852 | shared = shared & ~255L; |
853 | __x86_shared_cache_size_half = shared / 2; |
854 | __x86_shared_cache_size = shared; |
855 | } |
856 | |
857 | /* The large memcpy micro benchmark in glibc shows that 6 times of |
858 | shared cache size is the approximate value above which non-temporal |
859 | store becomes faster on a 8-core processor. This is the 3/4 of the |
860 | total shared cache size. */ |
861 | __x86_shared_non_temporal_threshold |
862 | = (cpu_features->non_temporal_threshold != 0 |
863 | ? cpu_features->non_temporal_threshold |
864 | : __x86_shared_cache_size * threads * 3 / 4); |
865 | |
866 | /* NB: The REP MOVSB threshold must be greater than VEC_SIZE * 8. */ |
867 | unsigned int minimum_rep_movsb_threshold; |
868 | /* NB: The default REP MOVSB threshold is 2048 * (VEC_SIZE / 16). */ |
869 | unsigned int rep_movsb_threshold; |
870 | if (CPU_FEATURE_USABLE_P (cpu_features, AVX512F) |
871 | && !CPU_FEATURE_PREFERRED_P (cpu_features, Prefer_No_AVX512)) |
872 | { |
873 | rep_movsb_threshold = 2048 * (64 / 16); |
874 | minimum_rep_movsb_threshold = 64 * 8; |
875 | } |
876 | else if (CPU_FEATURE_PREFERRED_P (cpu_features, |
877 | AVX_Fast_Unaligned_Load)) |
878 | { |
879 | rep_movsb_threshold = 2048 * (32 / 16); |
880 | minimum_rep_movsb_threshold = 32 * 8; |
881 | } |
882 | else |
883 | { |
884 | rep_movsb_threshold = 2048 * (16 / 16); |
885 | minimum_rep_movsb_threshold = 16 * 8; |
886 | } |
887 | if (cpu_features->rep_movsb_threshold > minimum_rep_movsb_threshold) |
888 | __x86_rep_movsb_threshold = cpu_features->rep_movsb_threshold; |
889 | else |
890 | __x86_rep_movsb_threshold = rep_movsb_threshold; |
891 | |
892 | # if HAVE_TUNABLES |
893 | __x86_rep_stosb_threshold = cpu_features->rep_stosb_threshold; |
894 | # endif |
895 | } |
896 | |
897 | #endif |
898 | |