1/* memmove/memcpy/mempcpy optimized with AVX512 for KNL hardware.
2 Copyright (C) 2016-2017 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 IS_IN (libc)
22
23# include "asm-syntax.h"
24
25 .section .text.avx512,"ax",@progbits
26# if defined SHARED && !defined USE_AS_MEMPCPY && !defined USE_AS_MEMMOVE
27ENTRY (__mempcpy_chk_avx512_no_vzeroupper)
28 cmp %RDX_LP, %RCX_LP
29 jb HIDDEN_JUMPTARGET (__chk_fail)
30END (__mempcpy_chk_avx512_no_vzeroupper)
31
32ENTRY (__mempcpy_avx512_no_vzeroupper)
33 mov %RDI_LP, %RAX_LP
34 add %RDX_LP, %RAX_LP
35 jmp L(start)
36END (__mempcpy_avx512_no_vzeroupper)
37# endif
38
39# ifdef SHARED
40ENTRY (__memmove_chk_avx512_no_vzeroupper)
41 cmp %RDX_LP, %RCX_LP
42 jb HIDDEN_JUMPTARGET (__chk_fail)
43END (__memmove_chk_avx512_no_vzeroupper)
44# endif
45
46ENTRY (__memmove_avx512_no_vzeroupper)
47 mov %RDI_LP, %RAX_LP
48# ifdef USE_AS_MEMPCPY
49 add %RDX_LP, %RAX_LP
50# endif
51L(start):
52# ifdef __ILP32__
53 /* Clear the upper 32 bits. */
54 mov %edx, %edx
55# endif
56 lea (%rsi, %rdx), %rcx
57 lea (%rdi, %rdx), %r9
58 cmp $512, %rdx
59 ja L(512bytesormore)
60
61L(check):
62 cmp $16, %rdx
63 jbe L(less_16bytes)
64 cmp $256, %rdx
65 jb L(less_256bytes)
66 vmovups (%rsi), %zmm0
67 vmovups 0x40(%rsi), %zmm1
68 vmovups 0x80(%rsi), %zmm2
69 vmovups 0xC0(%rsi), %zmm3
70 vmovups -0x100(%rcx), %zmm4
71 vmovups -0xC0(%rcx), %zmm5
72 vmovups -0x80(%rcx), %zmm6
73 vmovups -0x40(%rcx), %zmm7
74 vmovups %zmm0, (%rdi)
75 vmovups %zmm1, 0x40(%rdi)
76 vmovups %zmm2, 0x80(%rdi)
77 vmovups %zmm3, 0xC0(%rdi)
78 vmovups %zmm4, -0x100(%r9)
79 vmovups %zmm5, -0xC0(%r9)
80 vmovups %zmm6, -0x80(%r9)
81 vmovups %zmm7, -0x40(%r9)
82 ret
83
84L(less_256bytes):
85 cmp $128, %dl
86 jb L(less_128bytes)
87 vmovups (%rsi), %zmm0
88 vmovups 0x40(%rsi), %zmm1
89 vmovups -0x80(%rcx), %zmm2
90 vmovups -0x40(%rcx), %zmm3
91 vmovups %zmm0, (%rdi)
92 vmovups %zmm1, 0x40(%rdi)
93 vmovups %zmm2, -0x80(%r9)
94 vmovups %zmm3, -0x40(%r9)
95 ret
96
97L(less_128bytes):
98 cmp $64, %dl
99 jb L(less_64bytes)
100 vmovdqu (%rsi), %ymm0
101 vmovdqu 0x20(%rsi), %ymm1
102 vmovdqu -0x40(%rcx), %ymm2
103 vmovdqu -0x20(%rcx), %ymm3
104 vmovdqu %ymm0, (%rdi)
105 vmovdqu %ymm1, 0x20(%rdi)
106 vmovdqu %ymm2, -0x40(%r9)
107 vmovdqu %ymm3, -0x20(%r9)
108 ret
109
110L(less_64bytes):
111 cmp $32, %dl
112 jb L(less_32bytes)
113 vmovdqu (%rsi), %ymm0
114 vmovdqu -0x20(%rcx), %ymm1
115 vmovdqu %ymm0, (%rdi)
116 vmovdqu %ymm1, -0x20(%r9)
117 ret
118
119L(less_32bytes):
120 vmovdqu (%rsi), %xmm0
121 vmovdqu -0x10(%rcx), %xmm1
122 vmovdqu %xmm0, (%rdi)
123 vmovdqu %xmm1, -0x10(%r9)
124 ret
125
126L(less_16bytes):
127 cmp $8, %dl
128 jb L(less_8bytes)
129 movq (%rsi), %rsi
130 movq -0x8(%rcx), %rcx
131 movq %rsi, (%rdi)
132 movq %rcx, -0x8(%r9)
133 ret
134
135L(less_8bytes):
136 cmp $4, %dl
137 jb L(less_4bytes)
138 mov (%rsi), %esi
139 mov -0x4(%rcx), %ecx
140 mov %esi, (%rdi)
141 mov %ecx, -0x4(%r9)
142 ret
143
144L(less_4bytes):
145 cmp $2, %dl
146 jb L(less_2bytes)
147 mov (%rsi), %si
148 mov -0x2(%rcx), %cx
149 mov %si, (%rdi)
150 mov %cx, -0x2(%r9)
151 ret
152
153L(less_2bytes):
154 cmp $1, %dl
155 jb L(less_1bytes)
156 mov (%rsi), %cl
157 mov %cl, (%rdi)
158L(less_1bytes):
159 ret
160
161L(512bytesormore):
162# ifdef SHARED_CACHE_SIZE_HALF
163 mov $SHARED_CACHE_SIZE_HALF, %r8
164# else
165 mov __x86_shared_cache_size_half(%rip), %r8
166# endif
167 cmp %r8, %rdx
168 jae L(preloop_large)
169 cmp $1024, %rdx
170 ja L(1024bytesormore)
171 prefetcht1 (%rsi)
172 prefetcht1 0x40(%rsi)
173 prefetcht1 0x80(%rsi)
174 prefetcht1 0xC0(%rsi)
175 prefetcht1 0x100(%rsi)
176 prefetcht1 0x140(%rsi)
177 prefetcht1 0x180(%rsi)
178 prefetcht1 0x1C0(%rsi)
179 prefetcht1 -0x200(%rcx)
180 prefetcht1 -0x1C0(%rcx)
181 prefetcht1 -0x180(%rcx)
182 prefetcht1 -0x140(%rcx)
183 prefetcht1 -0x100(%rcx)
184 prefetcht1 -0xC0(%rcx)
185 prefetcht1 -0x80(%rcx)
186 prefetcht1 -0x40(%rcx)
187 vmovups (%rsi), %zmm0
188 vmovups 0x40(%rsi), %zmm1
189 vmovups 0x80(%rsi), %zmm2
190 vmovups 0xC0(%rsi), %zmm3
191 vmovups 0x100(%rsi), %zmm4
192 vmovups 0x140(%rsi), %zmm5
193 vmovups 0x180(%rsi), %zmm6
194 vmovups 0x1C0(%rsi), %zmm7
195 vmovups -0x200(%rcx), %zmm8
196 vmovups -0x1C0(%rcx), %zmm9
197 vmovups -0x180(%rcx), %zmm10
198 vmovups -0x140(%rcx), %zmm11
199 vmovups -0x100(%rcx), %zmm12
200 vmovups -0xC0(%rcx), %zmm13
201 vmovups -0x80(%rcx), %zmm14
202 vmovups -0x40(%rcx), %zmm15
203 vmovups %zmm0, (%rdi)
204 vmovups %zmm1, 0x40(%rdi)
205 vmovups %zmm2, 0x80(%rdi)
206 vmovups %zmm3, 0xC0(%rdi)
207 vmovups %zmm4, 0x100(%rdi)
208 vmovups %zmm5, 0x140(%rdi)
209 vmovups %zmm6, 0x180(%rdi)
210 vmovups %zmm7, 0x1C0(%rdi)
211 vmovups %zmm8, -0x200(%r9)
212 vmovups %zmm9, -0x1C0(%r9)
213 vmovups %zmm10, -0x180(%r9)
214 vmovups %zmm11, -0x140(%r9)
215 vmovups %zmm12, -0x100(%r9)
216 vmovups %zmm13, -0xC0(%r9)
217 vmovups %zmm14, -0x80(%r9)
218 vmovups %zmm15, -0x40(%r9)
219 ret
220
221L(1024bytesormore):
222 cmp %rsi, %rdi
223 ja L(1024bytesormore_bkw)
224 sub $512, %r9
225 vmovups -0x200(%rcx), %zmm8
226 vmovups -0x1C0(%rcx), %zmm9
227 vmovups -0x180(%rcx), %zmm10
228 vmovups -0x140(%rcx), %zmm11
229 vmovups -0x100(%rcx), %zmm12
230 vmovups -0xC0(%rcx), %zmm13
231 vmovups -0x80(%rcx), %zmm14
232 vmovups -0x40(%rcx), %zmm15
233 prefetcht1 (%rsi)
234 prefetcht1 0x40(%rsi)
235 prefetcht1 0x80(%rsi)
236 prefetcht1 0xC0(%rsi)
237 prefetcht1 0x100(%rsi)
238 prefetcht1 0x140(%rsi)
239 prefetcht1 0x180(%rsi)
240 prefetcht1 0x1C0(%rsi)
241
242/* Loop with unaligned memory access. */
243L(gobble_512bytes_loop):
244 vmovups (%rsi), %zmm0
245 vmovups 0x40(%rsi), %zmm1
246 vmovups 0x80(%rsi), %zmm2
247 vmovups 0xC0(%rsi), %zmm3
248 vmovups 0x100(%rsi), %zmm4
249 vmovups 0x140(%rsi), %zmm5
250 vmovups 0x180(%rsi), %zmm6
251 vmovups 0x1C0(%rsi), %zmm7
252 add $512, %rsi
253 prefetcht1 (%rsi)
254 prefetcht1 0x40(%rsi)
255 prefetcht1 0x80(%rsi)
256 prefetcht1 0xC0(%rsi)
257 prefetcht1 0x100(%rsi)
258 prefetcht1 0x140(%rsi)
259 prefetcht1 0x180(%rsi)
260 prefetcht1 0x1C0(%rsi)
261 vmovups %zmm0, (%rdi)
262 vmovups %zmm1, 0x40(%rdi)
263 vmovups %zmm2, 0x80(%rdi)
264 vmovups %zmm3, 0xC0(%rdi)
265 vmovups %zmm4, 0x100(%rdi)
266 vmovups %zmm5, 0x140(%rdi)
267 vmovups %zmm6, 0x180(%rdi)
268 vmovups %zmm7, 0x1C0(%rdi)
269 add $512, %rdi
270 cmp %r9, %rdi
271 jb L(gobble_512bytes_loop)
272 vmovups %zmm8, (%r9)
273 vmovups %zmm9, 0x40(%r9)
274 vmovups %zmm10, 0x80(%r9)
275 vmovups %zmm11, 0xC0(%r9)
276 vmovups %zmm12, 0x100(%r9)
277 vmovups %zmm13, 0x140(%r9)
278 vmovups %zmm14, 0x180(%r9)
279 vmovups %zmm15, 0x1C0(%r9)
280 ret
281
282L(1024bytesormore_bkw):
283 add $512, %rdi
284 vmovups 0x1C0(%rsi), %zmm8
285 vmovups 0x180(%rsi), %zmm9
286 vmovups 0x140(%rsi), %zmm10
287 vmovups 0x100(%rsi), %zmm11
288 vmovups 0xC0(%rsi), %zmm12
289 vmovups 0x80(%rsi), %zmm13
290 vmovups 0x40(%rsi), %zmm14
291 vmovups (%rsi), %zmm15
292 prefetcht1 -0x40(%rcx)
293 prefetcht1 -0x80(%rcx)
294 prefetcht1 -0xC0(%rcx)
295 prefetcht1 -0x100(%rcx)
296 prefetcht1 -0x140(%rcx)
297 prefetcht1 -0x180(%rcx)
298 prefetcht1 -0x1C0(%rcx)
299 prefetcht1 -0x200(%rcx)
300
301/* Backward loop with unaligned memory access. */
302L(gobble_512bytes_loop_bkw):
303 vmovups -0x40(%rcx), %zmm0
304 vmovups -0x80(%rcx), %zmm1
305 vmovups -0xC0(%rcx), %zmm2
306 vmovups -0x100(%rcx), %zmm3
307 vmovups -0x140(%rcx), %zmm4
308 vmovups -0x180(%rcx), %zmm5
309 vmovups -0x1C0(%rcx), %zmm6
310 vmovups -0x200(%rcx), %zmm7
311 sub $512, %rcx
312 prefetcht1 -0x40(%rcx)
313 prefetcht1 -0x80(%rcx)
314 prefetcht1 -0xC0(%rcx)
315 prefetcht1 -0x100(%rcx)
316 prefetcht1 -0x140(%rcx)
317 prefetcht1 -0x180(%rcx)
318 prefetcht1 -0x1C0(%rcx)
319 prefetcht1 -0x200(%rcx)
320 vmovups %zmm0, -0x40(%r9)
321 vmovups %zmm1, -0x80(%r9)
322 vmovups %zmm2, -0xC0(%r9)
323 vmovups %zmm3, -0x100(%r9)
324 vmovups %zmm4, -0x140(%r9)
325 vmovups %zmm5, -0x180(%r9)
326 vmovups %zmm6, -0x1C0(%r9)
327 vmovups %zmm7, -0x200(%r9)
328 sub $512, %r9
329 cmp %rdi, %r9
330 ja L(gobble_512bytes_loop_bkw)
331 vmovups %zmm8, -0x40(%rdi)
332 vmovups %zmm9, -0x80(%rdi)
333 vmovups %zmm10, -0xC0(%rdi)
334 vmovups %zmm11, -0x100(%rdi)
335 vmovups %zmm12, -0x140(%rdi)
336 vmovups %zmm13, -0x180(%rdi)
337 vmovups %zmm14, -0x1C0(%rdi)
338 vmovups %zmm15, -0x200(%rdi)
339 ret
340
341L(preloop_large):
342 cmp %rsi, %rdi
343 ja L(preloop_large_bkw)
344 vmovups (%rsi), %zmm4
345 vmovups 0x40(%rsi), %zmm5
346
347/* Align destination for access with non-temporal stores in the loop. */
348 mov %rdi, %r8
349 and $-0x80, %rdi
350 add $0x80, %rdi
351 sub %rdi, %r8
352 sub %r8, %rsi
353 add %r8, %rdx
354L(gobble_256bytes_nt_loop):
355 prefetcht1 0x200(%rsi)
356 prefetcht1 0x240(%rsi)
357 prefetcht1 0x280(%rsi)
358 prefetcht1 0x2C0(%rsi)
359 prefetcht1 0x300(%rsi)
360 prefetcht1 0x340(%rsi)
361 prefetcht1 0x380(%rsi)
362 prefetcht1 0x3C0(%rsi)
363 vmovdqu64 (%rsi), %zmm0
364 vmovdqu64 0x40(%rsi), %zmm1
365 vmovdqu64 0x80(%rsi), %zmm2
366 vmovdqu64 0xC0(%rsi), %zmm3
367 vmovntdq %zmm0, (%rdi)
368 vmovntdq %zmm1, 0x40(%rdi)
369 vmovntdq %zmm2, 0x80(%rdi)
370 vmovntdq %zmm3, 0xC0(%rdi)
371 sub $256, %rdx
372 add $256, %rsi
373 add $256, %rdi
374 cmp $256, %rdx
375 ja L(gobble_256bytes_nt_loop)
376 sfence
377 vmovups %zmm4, (%rax)
378 vmovups %zmm5, 0x40(%rax)
379 jmp L(check)
380
381L(preloop_large_bkw):
382 vmovups -0x80(%rcx), %zmm4
383 vmovups -0x40(%rcx), %zmm5
384
385/* Align end of destination for access with non-temporal stores. */
386 mov %r9, %r8
387 and $-0x80, %r9
388 sub %r9, %r8
389 sub %r8, %rcx
390 sub %r8, %rdx
391 add %r9, %r8
392L(gobble_256bytes_nt_loop_bkw):
393 prefetcht1 -0x400(%rcx)
394 prefetcht1 -0x3C0(%rcx)
395 prefetcht1 -0x380(%rcx)
396 prefetcht1 -0x340(%rcx)
397 prefetcht1 -0x300(%rcx)
398 prefetcht1 -0x2C0(%rcx)
399 prefetcht1 -0x280(%rcx)
400 prefetcht1 -0x240(%rcx)
401 vmovdqu64 -0x100(%rcx), %zmm0
402 vmovdqu64 -0xC0(%rcx), %zmm1
403 vmovdqu64 -0x80(%rcx), %zmm2
404 vmovdqu64 -0x40(%rcx), %zmm3
405 vmovntdq %zmm0, -0x100(%r9)
406 vmovntdq %zmm1, -0xC0(%r9)
407 vmovntdq %zmm2, -0x80(%r9)
408 vmovntdq %zmm3, -0x40(%r9)
409 sub $256, %rdx
410 sub $256, %rcx
411 sub $256, %r9
412 cmp $256, %rdx
413 ja L(gobble_256bytes_nt_loop_bkw)
414 sfence
415 vmovups %zmm4, -0x80(%r8)
416 vmovups %zmm5, -0x40(%r8)
417 jmp L(check)
418END (__memmove_avx512_no_vzeroupper)
419
420# ifdef SHARED
421strong_alias (__memmove_avx512_no_vzeroupper, __memcpy_avx512_no_vzeroupper)
422strong_alias (__memmove_chk_avx512_no_vzeroupper, __memcpy_chk_avx512_no_vzeroupper)
423# endif
424#endif
425