1 | /* memcpy optimized with AVX512 for KNL hardware. |
2 | Copyright (C) 2016 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 | <http://www.gnu.org/licenses/>. */ |
18 | |
19 | #include <sysdep.h> |
20 | |
21 | #if defined HAVE_AVX512_ASM_SUPPORT && IS_IN (libc) \ |
22 | && (defined SHARED \ |
23 | || defined USE_AS_MEMMOVE \ |
24 | || !defined USE_MULTIARCH) |
25 | |
26 | #include "asm-syntax.h" |
27 | #ifndef MEMCPY |
28 | # define MEMCPY __memcpy_avx512_no_vzeroupper |
29 | # define MEMCPY_CHK __memcpy_chk_avx512_no_vzeroupper |
30 | #endif |
31 | |
32 | .section .text,"ax" ,@progbits |
33 | #if !defined USE_AS_BCOPY |
34 | ENTRY (MEMCPY_CHK) |
35 | cmpq %rdx, %rcx |
36 | jb HIDDEN_JUMPTARGET (__chk_fail) |
37 | END (MEMCPY_CHK) |
38 | #endif |
39 | |
40 | ENTRY (MEMCPY) |
41 | mov %rdi, %rax |
42 | #ifdef USE_AS_MEMPCPY |
43 | add %rdx, %rax |
44 | #endif |
45 | lea (%rsi, %rdx), %rcx |
46 | lea (%rdi, %rdx), %r9 |
47 | cmp $512, %rdx |
48 | ja L(512bytesormore) |
49 | |
50 | L(check): |
51 | cmp $16, %rdx |
52 | jbe L(less_16bytes) |
53 | cmp $256, %rdx |
54 | jb L(less_256bytes) |
55 | vmovups (%rsi), %zmm0 |
56 | vmovups 0x40(%rsi), %zmm1 |
57 | vmovups 0x80(%rsi), %zmm2 |
58 | vmovups 0xC0(%rsi), %zmm3 |
59 | vmovups -0x100(%rcx), %zmm4 |
60 | vmovups -0xC0(%rcx), %zmm5 |
61 | vmovups -0x80(%rcx), %zmm6 |
62 | vmovups -0x40(%rcx), %zmm7 |
63 | vmovups %zmm0, (%rdi) |
64 | vmovups %zmm1, 0x40(%rdi) |
65 | vmovups %zmm2, 0x80(%rdi) |
66 | vmovups %zmm3, 0xC0(%rdi) |
67 | vmovups %zmm4, -0x100(%r9) |
68 | vmovups %zmm5, -0xC0(%r9) |
69 | vmovups %zmm6, -0x80(%r9) |
70 | vmovups %zmm7, -0x40(%r9) |
71 | ret |
72 | |
73 | L(less_256bytes): |
74 | cmp $128, %dl |
75 | jb L(less_128bytes) |
76 | vmovups (%rsi), %zmm0 |
77 | vmovups 0x40(%rsi), %zmm1 |
78 | vmovups -0x80(%rcx), %zmm2 |
79 | vmovups -0x40(%rcx), %zmm3 |
80 | vmovups %zmm0, (%rdi) |
81 | vmovups %zmm1, 0x40(%rdi) |
82 | vmovups %zmm2, -0x80(%r9) |
83 | vmovups %zmm3, -0x40(%r9) |
84 | ret |
85 | |
86 | L(less_128bytes): |
87 | cmp $64, %dl |
88 | jb L(less_64bytes) |
89 | vmovdqu (%rsi), %ymm0 |
90 | vmovdqu 0x20(%rsi), %ymm1 |
91 | vmovdqu -0x40(%rcx), %ymm2 |
92 | vmovdqu -0x20(%rcx), %ymm3 |
93 | vmovdqu %ymm0, (%rdi) |
94 | vmovdqu %ymm1, 0x20(%rdi) |
95 | vmovdqu %ymm2, -0x40(%r9) |
96 | vmovdqu %ymm3, -0x20(%r9) |
97 | ret |
98 | |
99 | L(less_64bytes): |
100 | cmp $32, %dl |
101 | jb L(less_32bytes) |
102 | vmovdqu (%rsi), %ymm0 |
103 | vmovdqu -0x20(%rcx), %ymm1 |
104 | vmovdqu %ymm0, (%rdi) |
105 | vmovdqu %ymm1, -0x20(%r9) |
106 | ret |
107 | |
108 | L(less_32bytes): |
109 | vmovdqu (%rsi), %xmm0 |
110 | vmovdqu -0x10(%rcx), %xmm1 |
111 | vmovdqu %xmm0, (%rdi) |
112 | vmovdqu %xmm1, -0x10(%r9) |
113 | ret |
114 | |
115 | L(less_16bytes): |
116 | cmp $8, %dl |
117 | jb L(less_8bytes) |
118 | movq (%rsi), %rsi |
119 | movq -0x8(%rcx), %rcx |
120 | movq %rsi, (%rdi) |
121 | movq %rcx, -0x8(%r9) |
122 | ret |
123 | |
124 | L(less_8bytes): |
125 | cmp $4, %dl |
126 | jb L(less_4bytes) |
127 | mov (%rsi), %esi |
128 | mov -0x4(%rcx), %ecx |
129 | mov %esi, (%rdi) |
130 | mov %ecx, -0x4(%r9) |
131 | ret |
132 | |
133 | L(less_4bytes): |
134 | cmp $2, %dl |
135 | jb L(less_2bytes) |
136 | mov (%rsi), %si |
137 | mov -0x2(%rcx), %cx |
138 | mov %si, (%rdi) |
139 | mov %cx, -0x2(%r9) |
140 | ret |
141 | |
142 | L(less_2bytes): |
143 | cmp $1, %dl |
144 | jb L(less_1bytes) |
145 | mov (%rsi), %cl |
146 | mov %cl, (%rdi) |
147 | L(less_1bytes): |
148 | ret |
149 | |
150 | L(512bytesormore): |
151 | #ifdef SHARED_CACHE_SIZE_HALF |
152 | mov $SHARED_CACHE_SIZE_HALF, %r8 |
153 | #else |
154 | mov __x86_shared_cache_size_half(%rip), %r8 |
155 | #endif |
156 | cmp %r8, %rdx |
157 | jae L(preloop_large) |
158 | cmp $1024, %rdx |
159 | ja L(1024bytesormore) |
160 | prefetcht1 (%rsi) |
161 | prefetcht1 0x40(%rsi) |
162 | prefetcht1 0x80(%rsi) |
163 | prefetcht1 0xC0(%rsi) |
164 | prefetcht1 0x100(%rsi) |
165 | prefetcht1 0x140(%rsi) |
166 | prefetcht1 0x180(%rsi) |
167 | prefetcht1 0x1C0(%rsi) |
168 | prefetcht1 -0x200(%rcx) |
169 | prefetcht1 -0x1C0(%rcx) |
170 | prefetcht1 -0x180(%rcx) |
171 | prefetcht1 -0x140(%rcx) |
172 | prefetcht1 -0x100(%rcx) |
173 | prefetcht1 -0xC0(%rcx) |
174 | prefetcht1 -0x80(%rcx) |
175 | prefetcht1 -0x40(%rcx) |
176 | vmovups (%rsi), %zmm0 |
177 | vmovups 0x40(%rsi), %zmm1 |
178 | vmovups 0x80(%rsi), %zmm2 |
179 | vmovups 0xC0(%rsi), %zmm3 |
180 | vmovups 0x100(%rsi), %zmm4 |
181 | vmovups 0x140(%rsi), %zmm5 |
182 | vmovups 0x180(%rsi), %zmm6 |
183 | vmovups 0x1C0(%rsi), %zmm7 |
184 | vmovups -0x200(%rcx), %zmm8 |
185 | vmovups -0x1C0(%rcx), %zmm9 |
186 | vmovups -0x180(%rcx), %zmm10 |
187 | vmovups -0x140(%rcx), %zmm11 |
188 | vmovups -0x100(%rcx), %zmm12 |
189 | vmovups -0xC0(%rcx), %zmm13 |
190 | vmovups -0x80(%rcx), %zmm14 |
191 | vmovups -0x40(%rcx), %zmm15 |
192 | vmovups %zmm0, (%rdi) |
193 | vmovups %zmm1, 0x40(%rdi) |
194 | vmovups %zmm2, 0x80(%rdi) |
195 | vmovups %zmm3, 0xC0(%rdi) |
196 | vmovups %zmm4, 0x100(%rdi) |
197 | vmovups %zmm5, 0x140(%rdi) |
198 | vmovups %zmm6, 0x180(%rdi) |
199 | vmovups %zmm7, 0x1C0(%rdi) |
200 | vmovups %zmm8, -0x200(%r9) |
201 | vmovups %zmm9, -0x1C0(%r9) |
202 | vmovups %zmm10, -0x180(%r9) |
203 | vmovups %zmm11, -0x140(%r9) |
204 | vmovups %zmm12, -0x100(%r9) |
205 | vmovups %zmm13, -0xC0(%r9) |
206 | vmovups %zmm14, -0x80(%r9) |
207 | vmovups %zmm15, -0x40(%r9) |
208 | ret |
209 | |
210 | L(1024bytesormore): |
211 | cmp %rsi, %rdi |
212 | ja L(1024bytesormore_bkw) |
213 | sub $512, %r9 |
214 | vmovups -0x200(%rcx), %zmm8 |
215 | vmovups -0x1C0(%rcx), %zmm9 |
216 | vmovups -0x180(%rcx), %zmm10 |
217 | vmovups -0x140(%rcx), %zmm11 |
218 | vmovups -0x100(%rcx), %zmm12 |
219 | vmovups -0xC0(%rcx), %zmm13 |
220 | vmovups -0x80(%rcx), %zmm14 |
221 | vmovups -0x40(%rcx), %zmm15 |
222 | prefetcht1 (%rsi) |
223 | prefetcht1 0x40(%rsi) |
224 | prefetcht1 0x80(%rsi) |
225 | prefetcht1 0xC0(%rsi) |
226 | prefetcht1 0x100(%rsi) |
227 | prefetcht1 0x140(%rsi) |
228 | prefetcht1 0x180(%rsi) |
229 | prefetcht1 0x1C0(%rsi) |
230 | |
231 | /* Loop with unaligned memory access. */ |
232 | L(gobble_512bytes_loop): |
233 | vmovups (%rsi), %zmm0 |
234 | vmovups 0x40(%rsi), %zmm1 |
235 | vmovups 0x80(%rsi), %zmm2 |
236 | vmovups 0xC0(%rsi), %zmm3 |
237 | vmovups 0x100(%rsi), %zmm4 |
238 | vmovups 0x140(%rsi), %zmm5 |
239 | vmovups 0x180(%rsi), %zmm6 |
240 | vmovups 0x1C0(%rsi), %zmm7 |
241 | add $512, %rsi |
242 | prefetcht1 (%rsi) |
243 | prefetcht1 0x40(%rsi) |
244 | prefetcht1 0x80(%rsi) |
245 | prefetcht1 0xC0(%rsi) |
246 | prefetcht1 0x100(%rsi) |
247 | prefetcht1 0x140(%rsi) |
248 | prefetcht1 0x180(%rsi) |
249 | prefetcht1 0x1C0(%rsi) |
250 | vmovups %zmm0, (%rdi) |
251 | vmovups %zmm1, 0x40(%rdi) |
252 | vmovups %zmm2, 0x80(%rdi) |
253 | vmovups %zmm3, 0xC0(%rdi) |
254 | vmovups %zmm4, 0x100(%rdi) |
255 | vmovups %zmm5, 0x140(%rdi) |
256 | vmovups %zmm6, 0x180(%rdi) |
257 | vmovups %zmm7, 0x1C0(%rdi) |
258 | add $512, %rdi |
259 | cmp %r9, %rdi |
260 | jb L(gobble_512bytes_loop) |
261 | vmovups %zmm8, (%r9) |
262 | vmovups %zmm9, 0x40(%r9) |
263 | vmovups %zmm10, 0x80(%r9) |
264 | vmovups %zmm11, 0xC0(%r9) |
265 | vmovups %zmm12, 0x100(%r9) |
266 | vmovups %zmm13, 0x140(%r9) |
267 | vmovups %zmm14, 0x180(%r9) |
268 | vmovups %zmm15, 0x1C0(%r9) |
269 | ret |
270 | |
271 | L(1024bytesormore_bkw): |
272 | add $512, %rdi |
273 | vmovups 0x1C0(%rsi), %zmm8 |
274 | vmovups 0x180(%rsi), %zmm9 |
275 | vmovups 0x140(%rsi), %zmm10 |
276 | vmovups 0x100(%rsi), %zmm11 |
277 | vmovups 0xC0(%rsi), %zmm12 |
278 | vmovups 0x80(%rsi), %zmm13 |
279 | vmovups 0x40(%rsi), %zmm14 |
280 | vmovups (%rsi), %zmm15 |
281 | prefetcht1 -0x40(%rcx) |
282 | prefetcht1 -0x80(%rcx) |
283 | prefetcht1 -0xC0(%rcx) |
284 | prefetcht1 -0x100(%rcx) |
285 | prefetcht1 -0x140(%rcx) |
286 | prefetcht1 -0x180(%rcx) |
287 | prefetcht1 -0x1C0(%rcx) |
288 | prefetcht1 -0x200(%rcx) |
289 | |
290 | /* Backward loop with unaligned memory access. */ |
291 | L(gobble_512bytes_loop_bkw): |
292 | vmovups -0x40(%rcx), %zmm0 |
293 | vmovups -0x80(%rcx), %zmm1 |
294 | vmovups -0xC0(%rcx), %zmm2 |
295 | vmovups -0x100(%rcx), %zmm3 |
296 | vmovups -0x140(%rcx), %zmm4 |
297 | vmovups -0x180(%rcx), %zmm5 |
298 | vmovups -0x1C0(%rcx), %zmm6 |
299 | vmovups -0x200(%rcx), %zmm7 |
300 | sub $512, %rcx |
301 | prefetcht1 -0x40(%rcx) |
302 | prefetcht1 -0x80(%rcx) |
303 | prefetcht1 -0xC0(%rcx) |
304 | prefetcht1 -0x100(%rcx) |
305 | prefetcht1 -0x140(%rcx) |
306 | prefetcht1 -0x180(%rcx) |
307 | prefetcht1 -0x1C0(%rcx) |
308 | prefetcht1 -0x200(%rcx) |
309 | vmovups %zmm0, -0x40(%r9) |
310 | vmovups %zmm1, -0x80(%r9) |
311 | vmovups %zmm2, -0xC0(%r9) |
312 | vmovups %zmm3, -0x100(%r9) |
313 | vmovups %zmm4, -0x140(%r9) |
314 | vmovups %zmm5, -0x180(%r9) |
315 | vmovups %zmm6, -0x1C0(%r9) |
316 | vmovups %zmm7, -0x200(%r9) |
317 | sub $512, %r9 |
318 | cmp %rdi, %r9 |
319 | ja L(gobble_512bytes_loop_bkw) |
320 | vmovups %zmm8, -0x40(%rdi) |
321 | vmovups %zmm9, -0x80(%rdi) |
322 | vmovups %zmm10, -0xC0(%rdi) |
323 | vmovups %zmm11, -0x100(%rdi) |
324 | vmovups %zmm12, -0x140(%rdi) |
325 | vmovups %zmm13, -0x180(%rdi) |
326 | vmovups %zmm14, -0x1C0(%rdi) |
327 | vmovups %zmm15, -0x200(%rdi) |
328 | ret |
329 | |
330 | L(preloop_large): |
331 | cmp %rsi, %rdi |
332 | ja L(preloop_large_bkw) |
333 | vmovups (%rsi), %zmm4 |
334 | vmovups 0x40(%rsi), %zmm5 |
335 | |
336 | /* Align destination for access with non-temporal stores in the loop. */ |
337 | mov %rdi, %r8 |
338 | and $-0x80, %rdi |
339 | add $0x80, %rdi |
340 | sub %rdi, %r8 |
341 | sub %r8, %rsi |
342 | add %r8, %rdx |
343 | L(gobble_256bytes_nt_loop): |
344 | prefetcht1 0x200(%rsi) |
345 | prefetcht1 0x240(%rsi) |
346 | prefetcht1 0x280(%rsi) |
347 | prefetcht1 0x2C0(%rsi) |
348 | prefetcht1 0x300(%rsi) |
349 | prefetcht1 0x340(%rsi) |
350 | prefetcht1 0x380(%rsi) |
351 | prefetcht1 0x3C0(%rsi) |
352 | vmovdqu64 (%rsi), %zmm0 |
353 | vmovdqu64 0x40(%rsi), %zmm1 |
354 | vmovdqu64 0x80(%rsi), %zmm2 |
355 | vmovdqu64 0xC0(%rsi), %zmm3 |
356 | vmovntdq %zmm0, (%rdi) |
357 | vmovntdq %zmm1, 0x40(%rdi) |
358 | vmovntdq %zmm2, 0x80(%rdi) |
359 | vmovntdq %zmm3, 0xC0(%rdi) |
360 | sub $256, %rdx |
361 | add $256, %rsi |
362 | add $256, %rdi |
363 | cmp $256, %rdx |
364 | ja L(gobble_256bytes_nt_loop) |
365 | sfence |
366 | vmovups %zmm4, (%rax) |
367 | vmovups %zmm5, 0x40(%rax) |
368 | jmp L(check) |
369 | |
370 | L(preloop_large_bkw): |
371 | vmovups -0x80(%rcx), %zmm4 |
372 | vmovups -0x40(%rcx), %zmm5 |
373 | |
374 | /* Align end of destination for access with non-temporal stores. */ |
375 | mov %r9, %r8 |
376 | and $-0x80, %r9 |
377 | sub %r9, %r8 |
378 | sub %r8, %rcx |
379 | sub %r8, %rdx |
380 | add %r9, %r8 |
381 | L(gobble_256bytes_nt_loop_bkw): |
382 | prefetcht1 -0x400(%rcx) |
383 | prefetcht1 -0x3C0(%rcx) |
384 | prefetcht1 -0x380(%rcx) |
385 | prefetcht1 -0x340(%rcx) |
386 | prefetcht1 -0x300(%rcx) |
387 | prefetcht1 -0x2C0(%rcx) |
388 | prefetcht1 -0x280(%rcx) |
389 | prefetcht1 -0x240(%rcx) |
390 | vmovdqu64 -0x100(%rcx), %zmm0 |
391 | vmovdqu64 -0xC0(%rcx), %zmm1 |
392 | vmovdqu64 -0x80(%rcx), %zmm2 |
393 | vmovdqu64 -0x40(%rcx), %zmm3 |
394 | vmovntdq %zmm0, -0x100(%r9) |
395 | vmovntdq %zmm1, -0xC0(%r9) |
396 | vmovntdq %zmm2, -0x80(%r9) |
397 | vmovntdq %zmm3, -0x40(%r9) |
398 | sub $256, %rdx |
399 | sub $256, %rcx |
400 | sub $256, %r9 |
401 | cmp $256, %rdx |
402 | ja L(gobble_256bytes_nt_loop_bkw) |
403 | sfence |
404 | vmovups %zmm4, -0x80(%r8) |
405 | vmovups %zmm5, -0x40(%r8) |
406 | jmp L(check) |
407 | END (MEMCPY) |
408 | #endif |
409 | |