1 | /* memmove/memcpy/mempcpy with unaligned load/store and rep movsb |
2 | Copyright (C) 2016-2018 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 | /* memmove/memcpy/mempcpy is implemented as: |
20 | 1. Use overlapping load and store to avoid branch. |
21 | 2. Load all sources into registers and store them together to avoid |
22 | possible address overlap between source and destination. |
23 | 3. If size is 8 * VEC_SIZE or less, load all sources into registers |
24 | and store them together. |
25 | 4. If address of destination > address of source, backward copy |
26 | 4 * VEC_SIZE at a time with unaligned load and aligned store. |
27 | Load the first 4 * VEC and last VEC before the loop and store |
28 | them after the loop to support overlapping addresses. |
29 | 5. Otherwise, forward copy 4 * VEC_SIZE at a time with unaligned |
30 | load and aligned store. Load the last 4 * VEC and first VEC |
31 | before the loop and store them after the loop to support |
32 | overlapping addresses. |
33 | 6. If size >= __x86_shared_non_temporal_threshold and there is no |
34 | overlap between destination and source, use non-temporal store |
35 | instead of aligned store. */ |
36 | |
37 | #include <sysdep.h> |
38 | |
39 | #ifndef MEMCPY_SYMBOL |
40 | # define MEMCPY_SYMBOL(p,s) MEMMOVE_SYMBOL(p, s) |
41 | #endif |
42 | |
43 | #ifndef MEMPCPY_SYMBOL |
44 | # define MEMPCPY_SYMBOL(p,s) MEMMOVE_SYMBOL(p, s) |
45 | #endif |
46 | |
47 | #ifndef MEMMOVE_CHK_SYMBOL |
48 | # define MEMMOVE_CHK_SYMBOL(p,s) MEMMOVE_SYMBOL(p, s) |
49 | #endif |
50 | |
51 | #ifndef VZEROUPPER |
52 | # if VEC_SIZE > 16 |
53 | # define VZEROUPPER vzeroupper |
54 | # else |
55 | # define VZEROUPPER |
56 | # endif |
57 | #endif |
58 | |
59 | /* Threshold to use Enhanced REP MOVSB. Since there is overhead to set |
60 | up REP MOVSB operation, REP MOVSB isn't faster on short data. The |
61 | memcpy micro benchmark in glibc shows that 2KB is the approximate |
62 | value above which REP MOVSB becomes faster than SSE2 optimization |
63 | on processors with Enhanced REP MOVSB. Since larger register size |
64 | can move more data with a single load and store, the threshold is |
65 | higher with larger register size. */ |
66 | #ifndef REP_MOVSB_THRESHOLD |
67 | # define REP_MOVSB_THRESHOLD (2048 * (VEC_SIZE / 16)) |
68 | #endif |
69 | |
70 | #ifndef PREFETCH |
71 | # define PREFETCH(addr) prefetcht0 addr |
72 | #endif |
73 | |
74 | /* Assume 64-byte prefetch size. */ |
75 | #ifndef PREFETCH_SIZE |
76 | # define PREFETCH_SIZE 64 |
77 | #endif |
78 | |
79 | #define PREFETCHED_LOAD_SIZE (VEC_SIZE * 4) |
80 | |
81 | #if PREFETCH_SIZE == 64 |
82 | # if PREFETCHED_LOAD_SIZE == PREFETCH_SIZE |
83 | # define PREFETCH_ONE_SET(dir, base, offset) \ |
84 | PREFETCH ((offset)base) |
85 | # elif PREFETCHED_LOAD_SIZE == 2 * PREFETCH_SIZE |
86 | # define PREFETCH_ONE_SET(dir, base, offset) \ |
87 | PREFETCH ((offset)base); \ |
88 | PREFETCH ((offset + dir * PREFETCH_SIZE)base) |
89 | # elif PREFETCHED_LOAD_SIZE == 4 * PREFETCH_SIZE |
90 | # define PREFETCH_ONE_SET(dir, base, offset) \ |
91 | PREFETCH ((offset)base); \ |
92 | PREFETCH ((offset + dir * PREFETCH_SIZE)base); \ |
93 | PREFETCH ((offset + dir * PREFETCH_SIZE)base); \ |
94 | PREFETCH ((offset + dir * PREFETCH_SIZE * 2)base); \ |
95 | PREFETCH ((offset + dir * PREFETCH_SIZE * 3)base) |
96 | # else |
97 | # error Unsupported PREFETCHED_LOAD_SIZE! |
98 | # endif |
99 | #else |
100 | # error Unsupported PREFETCH_SIZE! |
101 | #endif |
102 | |
103 | #ifndef SECTION |
104 | # error SECTION is not defined! |
105 | #endif |
106 | |
107 | .section SECTION(.text),"ax" ,@progbits |
108 | #if defined SHARED && IS_IN (libc) |
109 | ENTRY (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned)) |
110 | cmpq %rdx, %rcx |
111 | jb HIDDEN_JUMPTARGET (__chk_fail) |
112 | END (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned)) |
113 | #endif |
114 | |
115 | ENTRY (MEMPCPY_SYMBOL (__mempcpy, unaligned)) |
116 | movq %rdi, %rax |
117 | addq %rdx, %rax |
118 | jmp L(start) |
119 | END (MEMPCPY_SYMBOL (__mempcpy, unaligned)) |
120 | |
121 | #if defined SHARED && IS_IN (libc) |
122 | ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned)) |
123 | cmpq %rdx, %rcx |
124 | jb HIDDEN_JUMPTARGET (__chk_fail) |
125 | END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned)) |
126 | #endif |
127 | |
128 | ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned)) |
129 | movq %rdi, %rax |
130 | L(start): |
131 | cmpq $VEC_SIZE, %rdx |
132 | jb L(less_vec) |
133 | cmpq $(VEC_SIZE * 2), %rdx |
134 | ja L(more_2x_vec) |
135 | #if !defined USE_MULTIARCH || !IS_IN (libc) |
136 | L(last_2x_vec): |
137 | #endif |
138 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
139 | VMOVU (%rsi), %VEC(0) |
140 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(1) |
141 | VMOVU %VEC(0), (%rdi) |
142 | VMOVU %VEC(1), -VEC_SIZE(%rdi,%rdx) |
143 | VZEROUPPER |
144 | #if !defined USE_MULTIARCH || !IS_IN (libc) |
145 | L(nop): |
146 | #endif |
147 | ret |
148 | #if defined USE_MULTIARCH && IS_IN (libc) |
149 | END (MEMMOVE_SYMBOL (__memmove, unaligned)) |
150 | |
151 | # if VEC_SIZE == 16 |
152 | ENTRY (__mempcpy_chk_erms) |
153 | cmpq %rdx, %rcx |
154 | jb HIDDEN_JUMPTARGET (__chk_fail) |
155 | END (__mempcpy_chk_erms) |
156 | |
157 | /* Only used to measure performance of REP MOVSB. */ |
158 | ENTRY (__mempcpy_erms) |
159 | movq %rdi, %rax |
160 | addq %rdx, %rax |
161 | jmp L(start_movsb) |
162 | END (__mempcpy_erms) |
163 | |
164 | ENTRY (__memmove_chk_erms) |
165 | cmpq %rdx, %rcx |
166 | jb HIDDEN_JUMPTARGET (__chk_fail) |
167 | END (__memmove_chk_erms) |
168 | |
169 | ENTRY (__memmove_erms) |
170 | movq %rdi, %rax |
171 | L(start_movsb): |
172 | movq %rdx, %rcx |
173 | cmpq %rsi, %rdi |
174 | jb 1f |
175 | /* Source == destination is less common. */ |
176 | je 2f |
177 | leaq (%rsi,%rcx), %rdx |
178 | cmpq %rdx, %rdi |
179 | jb L(movsb_backward) |
180 | 1: |
181 | rep movsb |
182 | 2: |
183 | ret |
184 | L(movsb_backward): |
185 | leaq -1(%rdi,%rcx), %rdi |
186 | leaq -1(%rsi,%rcx), %rsi |
187 | std |
188 | rep movsb |
189 | cld |
190 | ret |
191 | END (__memmove_erms) |
192 | strong_alias (__memmove_erms, __memcpy_erms) |
193 | strong_alias (__memmove_chk_erms, __memcpy_chk_erms) |
194 | # endif |
195 | |
196 | # ifdef SHARED |
197 | ENTRY (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned_erms)) |
198 | cmpq %rdx, %rcx |
199 | jb HIDDEN_JUMPTARGET (__chk_fail) |
200 | END (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned_erms)) |
201 | # endif |
202 | |
203 | ENTRY (MEMMOVE_SYMBOL (__mempcpy, unaligned_erms)) |
204 | movq %rdi, %rax |
205 | addq %rdx, %rax |
206 | jmp L(start_erms) |
207 | END (MEMMOVE_SYMBOL (__mempcpy, unaligned_erms)) |
208 | |
209 | # ifdef SHARED |
210 | ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms)) |
211 | cmpq %rdx, %rcx |
212 | jb HIDDEN_JUMPTARGET (__chk_fail) |
213 | END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms)) |
214 | # endif |
215 | |
216 | ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms)) |
217 | movq %rdi, %rax |
218 | L(start_erms): |
219 | cmpq $VEC_SIZE, %rdx |
220 | jb L(less_vec) |
221 | cmpq $(VEC_SIZE * 2), %rdx |
222 | ja L(movsb_more_2x_vec) |
223 | L(last_2x_vec): |
224 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
225 | VMOVU (%rsi), %VEC(0) |
226 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(1) |
227 | VMOVU %VEC(0), (%rdi) |
228 | VMOVU %VEC(1), -VEC_SIZE(%rdi,%rdx) |
229 | L(return): |
230 | VZEROUPPER |
231 | ret |
232 | |
233 | L(movsb): |
234 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
235 | jae L(more_8x_vec) |
236 | cmpq %rsi, %rdi |
237 | jb 1f |
238 | /* Source == destination is less common. */ |
239 | je L(nop) |
240 | leaq (%rsi,%rdx), %r9 |
241 | cmpq %r9, %rdi |
242 | /* Avoid slow backward REP MOVSB. */ |
243 | # if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8) |
244 | # error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE! |
245 | # endif |
246 | jb L(more_8x_vec_backward) |
247 | 1: |
248 | movq %rdx, %rcx |
249 | rep movsb |
250 | L(nop): |
251 | ret |
252 | #endif |
253 | |
254 | L(less_vec): |
255 | /* Less than 1 VEC. */ |
256 | #if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64 |
257 | # error Unsupported VEC_SIZE! |
258 | #endif |
259 | #if VEC_SIZE > 32 |
260 | cmpb $32, %dl |
261 | jae L(between_32_63) |
262 | #endif |
263 | #if VEC_SIZE > 16 |
264 | cmpb $16, %dl |
265 | jae L(between_16_31) |
266 | #endif |
267 | cmpb $8, %dl |
268 | jae L(between_8_15) |
269 | cmpb $4, %dl |
270 | jae L(between_4_7) |
271 | cmpb $1, %dl |
272 | ja L(between_2_3) |
273 | jb 1f |
274 | movzbl (%rsi), %ecx |
275 | movb %cl, (%rdi) |
276 | 1: |
277 | ret |
278 | #if VEC_SIZE > 32 |
279 | L(between_32_63): |
280 | /* From 32 to 63. No branch when size == 32. */ |
281 | vmovdqu (%rsi), %ymm0 |
282 | vmovdqu -32(%rsi,%rdx), %ymm1 |
283 | vmovdqu %ymm0, (%rdi) |
284 | vmovdqu %ymm1, -32(%rdi,%rdx) |
285 | VZEROUPPER |
286 | ret |
287 | #endif |
288 | #if VEC_SIZE > 16 |
289 | /* From 16 to 31. No branch when size == 16. */ |
290 | L(between_16_31): |
291 | vmovdqu (%rsi), %xmm0 |
292 | vmovdqu -16(%rsi,%rdx), %xmm1 |
293 | vmovdqu %xmm0, (%rdi) |
294 | vmovdqu %xmm1, -16(%rdi,%rdx) |
295 | ret |
296 | #endif |
297 | L(between_8_15): |
298 | /* From 8 to 15. No branch when size == 8. */ |
299 | movq -8(%rsi,%rdx), %rcx |
300 | movq (%rsi), %rsi |
301 | movq %rcx, -8(%rdi,%rdx) |
302 | movq %rsi, (%rdi) |
303 | ret |
304 | L(between_4_7): |
305 | /* From 4 to 7. No branch when size == 4. */ |
306 | movl -4(%rsi,%rdx), %ecx |
307 | movl (%rsi), %esi |
308 | movl %ecx, -4(%rdi,%rdx) |
309 | movl %esi, (%rdi) |
310 | ret |
311 | L(between_2_3): |
312 | /* From 2 to 3. No branch when size == 2. */ |
313 | movzwl -2(%rsi,%rdx), %ecx |
314 | movzwl (%rsi), %esi |
315 | movw %cx, -2(%rdi,%rdx) |
316 | movw %si, (%rdi) |
317 | ret |
318 | |
319 | #if defined USE_MULTIARCH && IS_IN (libc) |
320 | L(movsb_more_2x_vec): |
321 | cmpq $REP_MOVSB_THRESHOLD, %rdx |
322 | ja L(movsb) |
323 | #endif |
324 | L(more_2x_vec): |
325 | /* More than 2 * VEC and there may be overlap between destination |
326 | and source. */ |
327 | cmpq $(VEC_SIZE * 8), %rdx |
328 | ja L(more_8x_vec) |
329 | cmpq $(VEC_SIZE * 4), %rdx |
330 | jb L(last_4x_vec) |
331 | /* Copy from 4 * VEC to 8 * VEC, inclusively. */ |
332 | VMOVU (%rsi), %VEC(0) |
333 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
334 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
335 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
336 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(4) |
337 | VMOVU -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(5) |
338 | VMOVU -(VEC_SIZE * 3)(%rsi,%rdx), %VEC(6) |
339 | VMOVU -(VEC_SIZE * 4)(%rsi,%rdx), %VEC(7) |
340 | VMOVU %VEC(0), (%rdi) |
341 | VMOVU %VEC(1), VEC_SIZE(%rdi) |
342 | VMOVU %VEC(2), (VEC_SIZE * 2)(%rdi) |
343 | VMOVU %VEC(3), (VEC_SIZE * 3)(%rdi) |
344 | VMOVU %VEC(4), -VEC_SIZE(%rdi,%rdx) |
345 | VMOVU %VEC(5), -(VEC_SIZE * 2)(%rdi,%rdx) |
346 | VMOVU %VEC(6), -(VEC_SIZE * 3)(%rdi,%rdx) |
347 | VMOVU %VEC(7), -(VEC_SIZE * 4)(%rdi,%rdx) |
348 | VZEROUPPER |
349 | ret |
350 | L(last_4x_vec): |
351 | /* Copy from 2 * VEC to 4 * VEC. */ |
352 | VMOVU (%rsi), %VEC(0) |
353 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
354 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(2) |
355 | VMOVU -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(3) |
356 | VMOVU %VEC(0), (%rdi) |
357 | VMOVU %VEC(1), VEC_SIZE(%rdi) |
358 | VMOVU %VEC(2), -VEC_SIZE(%rdi,%rdx) |
359 | VMOVU %VEC(3), -(VEC_SIZE * 2)(%rdi,%rdx) |
360 | VZEROUPPER |
361 | ret |
362 | |
363 | L(more_8x_vec): |
364 | cmpq %rsi, %rdi |
365 | ja L(more_8x_vec_backward) |
366 | /* Source == destination is less common. */ |
367 | je L(nop) |
368 | /* Load the first VEC and last 4 * VEC to support overlapping |
369 | addresses. */ |
370 | VMOVU (%rsi), %VEC(4) |
371 | VMOVU -VEC_SIZE(%rsi, %rdx), %VEC(5) |
372 | VMOVU -(VEC_SIZE * 2)(%rsi, %rdx), %VEC(6) |
373 | VMOVU -(VEC_SIZE * 3)(%rsi, %rdx), %VEC(7) |
374 | VMOVU -(VEC_SIZE * 4)(%rsi, %rdx), %VEC(8) |
375 | /* Save start and stop of the destination buffer. */ |
376 | movq %rdi, %r11 |
377 | leaq -VEC_SIZE(%rdi, %rdx), %rcx |
378 | /* Align destination for aligned stores in the loop. Compute |
379 | how much destination is misaligned. */ |
380 | movq %rdi, %r8 |
381 | andq $(VEC_SIZE - 1), %r8 |
382 | /* Get the negative of offset for alignment. */ |
383 | subq $VEC_SIZE, %r8 |
384 | /* Adjust source. */ |
385 | subq %r8, %rsi |
386 | /* Adjust destination which should be aligned now. */ |
387 | subq %r8, %rdi |
388 | /* Adjust length. */ |
389 | addq %r8, %rdx |
390 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
391 | /* Check non-temporal store threshold. */ |
392 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
393 | ja L(large_forward) |
394 | #endif |
395 | L(loop_4x_vec_forward): |
396 | /* Copy 4 * VEC a time forward. */ |
397 | VMOVU (%rsi), %VEC(0) |
398 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
399 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
400 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
401 | addq $(VEC_SIZE * 4), %rsi |
402 | subq $(VEC_SIZE * 4), %rdx |
403 | VMOVA %VEC(0), (%rdi) |
404 | VMOVA %VEC(1), VEC_SIZE(%rdi) |
405 | VMOVA %VEC(2), (VEC_SIZE * 2)(%rdi) |
406 | VMOVA %VEC(3), (VEC_SIZE * 3)(%rdi) |
407 | addq $(VEC_SIZE * 4), %rdi |
408 | cmpq $(VEC_SIZE * 4), %rdx |
409 | ja L(loop_4x_vec_forward) |
410 | /* Store the last 4 * VEC. */ |
411 | VMOVU %VEC(5), (%rcx) |
412 | VMOVU %VEC(6), -VEC_SIZE(%rcx) |
413 | VMOVU %VEC(7), -(VEC_SIZE * 2)(%rcx) |
414 | VMOVU %VEC(8), -(VEC_SIZE * 3)(%rcx) |
415 | /* Store the first VEC. */ |
416 | VMOVU %VEC(4), (%r11) |
417 | VZEROUPPER |
418 | ret |
419 | |
420 | L(more_8x_vec_backward): |
421 | /* Load the first 4 * VEC and last VEC to support overlapping |
422 | addresses. */ |
423 | VMOVU (%rsi), %VEC(4) |
424 | VMOVU VEC_SIZE(%rsi), %VEC(5) |
425 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(6) |
426 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(7) |
427 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(8) |
428 | /* Save stop of the destination buffer. */ |
429 | leaq -VEC_SIZE(%rdi, %rdx), %r11 |
430 | /* Align destination end for aligned stores in the loop. Compute |
431 | how much destination end is misaligned. */ |
432 | leaq -VEC_SIZE(%rsi, %rdx), %rcx |
433 | movq %r11, %r9 |
434 | movq %r11, %r8 |
435 | andq $(VEC_SIZE - 1), %r8 |
436 | /* Adjust source. */ |
437 | subq %r8, %rcx |
438 | /* Adjust the end of destination which should be aligned now. */ |
439 | subq %r8, %r9 |
440 | /* Adjust length. */ |
441 | subq %r8, %rdx |
442 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
443 | /* Check non-temporal store threshold. */ |
444 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
445 | ja L(large_backward) |
446 | #endif |
447 | L(loop_4x_vec_backward): |
448 | /* Copy 4 * VEC a time backward. */ |
449 | VMOVU (%rcx), %VEC(0) |
450 | VMOVU -VEC_SIZE(%rcx), %VEC(1) |
451 | VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2) |
452 | VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3) |
453 | subq $(VEC_SIZE * 4), %rcx |
454 | subq $(VEC_SIZE * 4), %rdx |
455 | VMOVA %VEC(0), (%r9) |
456 | VMOVA %VEC(1), -VEC_SIZE(%r9) |
457 | VMOVA %VEC(2), -(VEC_SIZE * 2)(%r9) |
458 | VMOVA %VEC(3), -(VEC_SIZE * 3)(%r9) |
459 | subq $(VEC_SIZE * 4), %r9 |
460 | cmpq $(VEC_SIZE * 4), %rdx |
461 | ja L(loop_4x_vec_backward) |
462 | /* Store the first 4 * VEC. */ |
463 | VMOVU %VEC(4), (%rdi) |
464 | VMOVU %VEC(5), VEC_SIZE(%rdi) |
465 | VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi) |
466 | VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi) |
467 | /* Store the last VEC. */ |
468 | VMOVU %VEC(8), (%r11) |
469 | VZEROUPPER |
470 | ret |
471 | |
472 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
473 | L(large_forward): |
474 | /* Don't use non-temporal store if there is overlap between |
475 | destination and source since destination may be in cache |
476 | when source is loaded. */ |
477 | leaq (%rdi, %rdx), %r10 |
478 | cmpq %r10, %rsi |
479 | jb L(loop_4x_vec_forward) |
480 | L(loop_large_forward): |
481 | /* Copy 4 * VEC a time forward with non-temporal stores. */ |
482 | PREFETCH_ONE_SET (1, (%rsi), PREFETCHED_LOAD_SIZE * 2) |
483 | PREFETCH_ONE_SET (1, (%rsi), PREFETCHED_LOAD_SIZE * 3) |
484 | VMOVU (%rsi), %VEC(0) |
485 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
486 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
487 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
488 | addq $PREFETCHED_LOAD_SIZE, %rsi |
489 | subq $PREFETCHED_LOAD_SIZE, %rdx |
490 | VMOVNT %VEC(0), (%rdi) |
491 | VMOVNT %VEC(1), VEC_SIZE(%rdi) |
492 | VMOVNT %VEC(2), (VEC_SIZE * 2)(%rdi) |
493 | VMOVNT %VEC(3), (VEC_SIZE * 3)(%rdi) |
494 | addq $PREFETCHED_LOAD_SIZE, %rdi |
495 | cmpq $PREFETCHED_LOAD_SIZE, %rdx |
496 | ja L(loop_large_forward) |
497 | sfence |
498 | /* Store the last 4 * VEC. */ |
499 | VMOVU %VEC(5), (%rcx) |
500 | VMOVU %VEC(6), -VEC_SIZE(%rcx) |
501 | VMOVU %VEC(7), -(VEC_SIZE * 2)(%rcx) |
502 | VMOVU %VEC(8), -(VEC_SIZE * 3)(%rcx) |
503 | /* Store the first VEC. */ |
504 | VMOVU %VEC(4), (%r11) |
505 | VZEROUPPER |
506 | ret |
507 | |
508 | L(large_backward): |
509 | /* Don't use non-temporal store if there is overlap between |
510 | destination and source since destination may be in cache |
511 | when source is loaded. */ |
512 | leaq (%rcx, %rdx), %r10 |
513 | cmpq %r10, %r9 |
514 | jb L(loop_4x_vec_backward) |
515 | L(loop_large_backward): |
516 | /* Copy 4 * VEC a time backward with non-temporal stores. */ |
517 | PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 2) |
518 | PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 3) |
519 | VMOVU (%rcx), %VEC(0) |
520 | VMOVU -VEC_SIZE(%rcx), %VEC(1) |
521 | VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2) |
522 | VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3) |
523 | subq $PREFETCHED_LOAD_SIZE, %rcx |
524 | subq $PREFETCHED_LOAD_SIZE, %rdx |
525 | VMOVNT %VEC(0), (%r9) |
526 | VMOVNT %VEC(1), -VEC_SIZE(%r9) |
527 | VMOVNT %VEC(2), -(VEC_SIZE * 2)(%r9) |
528 | VMOVNT %VEC(3), -(VEC_SIZE * 3)(%r9) |
529 | subq $PREFETCHED_LOAD_SIZE, %r9 |
530 | cmpq $PREFETCHED_LOAD_SIZE, %rdx |
531 | ja L(loop_large_backward) |
532 | sfence |
533 | /* Store the first 4 * VEC. */ |
534 | VMOVU %VEC(4), (%rdi) |
535 | VMOVU %VEC(5), VEC_SIZE(%rdi) |
536 | VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi) |
537 | VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi) |
538 | /* Store the last VEC. */ |
539 | VMOVU %VEC(8), (%r11) |
540 | VZEROUPPER |
541 | ret |
542 | #endif |
543 | END (MEMMOVE_SYMBOL (__memmove, unaligned_erms)) |
544 | |
545 | #if IS_IN (libc) |
546 | # ifdef USE_MULTIARCH |
547 | strong_alias (MEMMOVE_SYMBOL (__memmove, unaligned_erms), |
548 | MEMMOVE_SYMBOL (__memcpy, unaligned_erms)) |
549 | # ifdef SHARED |
550 | strong_alias (MEMMOVE_SYMBOL (__memmove_chk, unaligned_erms), |
551 | MEMMOVE_SYMBOL (__memcpy_chk, unaligned_erms)) |
552 | # endif |
553 | # endif |
554 | # ifdef SHARED |
555 | strong_alias (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned), |
556 | MEMMOVE_CHK_SYMBOL (__memcpy_chk, unaligned)) |
557 | # endif |
558 | #endif |
559 | strong_alias (MEMMOVE_SYMBOL (__memmove, unaligned), |
560 | MEMCPY_SYMBOL (__memcpy, unaligned)) |
561 | |