1 | /* memmove/memcpy/mempcpy with unaligned load/store and rep movsb |
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 | /* 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 | #if VEC_SIZE == 16 || defined SHARED |
116 | ENTRY (MEMPCPY_SYMBOL (__mempcpy, unaligned)) |
117 | movq %rdi, %rax |
118 | addq %rdx, %rax |
119 | jmp L(start) |
120 | END (MEMPCPY_SYMBOL (__mempcpy, unaligned)) |
121 | #endif |
122 | |
123 | #if defined SHARED && IS_IN (libc) |
124 | ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned)) |
125 | cmpq %rdx, %rcx |
126 | jb HIDDEN_JUMPTARGET (__chk_fail) |
127 | END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned)) |
128 | #endif |
129 | |
130 | ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned)) |
131 | movq %rdi, %rax |
132 | L(start): |
133 | cmpq $VEC_SIZE, %rdx |
134 | jb L(less_vec) |
135 | cmpq $(VEC_SIZE * 2), %rdx |
136 | ja L(more_2x_vec) |
137 | #if !defined USE_MULTIARCH || !IS_IN (libc) |
138 | L(last_2x_vec): |
139 | #endif |
140 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
141 | VMOVU (%rsi), %VEC(0) |
142 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(1) |
143 | VMOVU %VEC(0), (%rdi) |
144 | VMOVU %VEC(1), -VEC_SIZE(%rdi,%rdx) |
145 | VZEROUPPER |
146 | #if !defined USE_MULTIARCH || !IS_IN (libc) |
147 | L(nop): |
148 | #endif |
149 | ret |
150 | #if defined USE_MULTIARCH && IS_IN (libc) |
151 | END (MEMMOVE_SYMBOL (__memmove, unaligned)) |
152 | |
153 | # if VEC_SIZE == 16 |
154 | # if defined SHARED |
155 | ENTRY (__mempcpy_chk_erms) |
156 | cmpq %rdx, %rcx |
157 | jb HIDDEN_JUMPTARGET (__chk_fail) |
158 | END (__mempcpy_chk_erms) |
159 | |
160 | /* Only used to measure performance of REP MOVSB. */ |
161 | ENTRY (__mempcpy_erms) |
162 | movq %rdi, %rax |
163 | addq %rdx, %rax |
164 | jmp L(start_movsb) |
165 | END (__mempcpy_erms) |
166 | # endif |
167 | |
168 | ENTRY (__memmove_chk_erms) |
169 | cmpq %rdx, %rcx |
170 | jb HIDDEN_JUMPTARGET (__chk_fail) |
171 | END (__memmove_chk_erms) |
172 | |
173 | ENTRY (__memmove_erms) |
174 | movq %rdi, %rax |
175 | L(start_movsb): |
176 | movq %rdx, %rcx |
177 | cmpq %rsi, %rdi |
178 | jb 1f |
179 | /* Source == destination is less common. */ |
180 | je 2f |
181 | leaq (%rsi,%rcx), %rdx |
182 | cmpq %rdx, %rdi |
183 | jb L(movsb_backward) |
184 | 1: |
185 | rep movsb |
186 | 2: |
187 | ret |
188 | L(movsb_backward): |
189 | leaq -1(%rdi,%rcx), %rdi |
190 | leaq -1(%rsi,%rcx), %rsi |
191 | std |
192 | rep movsb |
193 | cld |
194 | ret |
195 | END (__memmove_erms) |
196 | # if defined SHARED |
197 | strong_alias (__memmove_erms, __memcpy_erms) |
198 | strong_alias (__memmove_chk_erms, __memcpy_chk_erms) |
199 | # endif |
200 | # endif |
201 | |
202 | # ifdef SHARED |
203 | ENTRY (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned_erms)) |
204 | cmpq %rdx, %rcx |
205 | jb HIDDEN_JUMPTARGET (__chk_fail) |
206 | END (MEMMOVE_CHK_SYMBOL (__mempcpy_chk, unaligned_erms)) |
207 | |
208 | ENTRY (MEMMOVE_SYMBOL (__mempcpy, unaligned_erms)) |
209 | movq %rdi, %rax |
210 | addq %rdx, %rax |
211 | jmp L(start_erms) |
212 | END (MEMMOVE_SYMBOL (__mempcpy, unaligned_erms)) |
213 | |
214 | ENTRY (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms)) |
215 | cmpq %rdx, %rcx |
216 | jb HIDDEN_JUMPTARGET (__chk_fail) |
217 | END (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned_erms)) |
218 | # endif |
219 | |
220 | ENTRY (MEMMOVE_SYMBOL (__memmove, unaligned_erms)) |
221 | movq %rdi, %rax |
222 | L(start_erms): |
223 | cmpq $VEC_SIZE, %rdx |
224 | jb L(less_vec) |
225 | cmpq $(VEC_SIZE * 2), %rdx |
226 | ja L(movsb_more_2x_vec) |
227 | L(last_2x_vec): |
228 | /* From VEC and to 2 * VEC. No branch when size == VEC_SIZE. */ |
229 | VMOVU (%rsi), %VEC(0) |
230 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(1) |
231 | VMOVU %VEC(0), (%rdi) |
232 | VMOVU %VEC(1), -VEC_SIZE(%rdi,%rdx) |
233 | L(return): |
234 | VZEROUPPER |
235 | ret |
236 | |
237 | L(movsb): |
238 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
239 | jae L(more_8x_vec) |
240 | cmpq %rsi, %rdi |
241 | jb 1f |
242 | /* Source == destination is less common. */ |
243 | je L(nop) |
244 | leaq (%rsi,%rdx), %r9 |
245 | cmpq %r9, %rdi |
246 | /* Avoid slow backward REP MOVSB. */ |
247 | # if REP_MOVSB_THRESHOLD <= (VEC_SIZE * 8) |
248 | # error Unsupported REP_MOVSB_THRESHOLD and VEC_SIZE! |
249 | # endif |
250 | jb L(more_8x_vec_backward) |
251 | 1: |
252 | movq %rdx, %rcx |
253 | rep movsb |
254 | L(nop): |
255 | ret |
256 | #endif |
257 | |
258 | L(less_vec): |
259 | /* Less than 1 VEC. */ |
260 | #if VEC_SIZE != 16 && VEC_SIZE != 32 && VEC_SIZE != 64 |
261 | # error Unsupported VEC_SIZE! |
262 | #endif |
263 | #if VEC_SIZE > 32 |
264 | cmpb $32, %dl |
265 | jae L(between_32_63) |
266 | #endif |
267 | #if VEC_SIZE > 16 |
268 | cmpb $16, %dl |
269 | jae L(between_16_31) |
270 | #endif |
271 | cmpb $8, %dl |
272 | jae L(between_8_15) |
273 | cmpb $4, %dl |
274 | jae L(between_4_7) |
275 | cmpb $1, %dl |
276 | ja L(between_2_3) |
277 | jb 1f |
278 | movzbl (%rsi), %ecx |
279 | movb %cl, (%rdi) |
280 | 1: |
281 | ret |
282 | #if VEC_SIZE > 32 |
283 | L(between_32_63): |
284 | /* From 32 to 63. No branch when size == 32. */ |
285 | vmovdqu (%rsi), %ymm0 |
286 | vmovdqu -32(%rsi,%rdx), %ymm1 |
287 | vmovdqu %ymm0, (%rdi) |
288 | vmovdqu %ymm1, -32(%rdi,%rdx) |
289 | VZEROUPPER |
290 | ret |
291 | #endif |
292 | #if VEC_SIZE > 16 |
293 | /* From 16 to 31. No branch when size == 16. */ |
294 | L(between_16_31): |
295 | vmovdqu (%rsi), %xmm0 |
296 | vmovdqu -16(%rsi,%rdx), %xmm1 |
297 | vmovdqu %xmm0, (%rdi) |
298 | vmovdqu %xmm1, -16(%rdi,%rdx) |
299 | ret |
300 | #endif |
301 | L(between_8_15): |
302 | /* From 8 to 15. No branch when size == 8. */ |
303 | movq -8(%rsi,%rdx), %rcx |
304 | movq (%rsi), %rsi |
305 | movq %rcx, -8(%rdi,%rdx) |
306 | movq %rsi, (%rdi) |
307 | ret |
308 | L(between_4_7): |
309 | /* From 4 to 7. No branch when size == 4. */ |
310 | movl -4(%rsi,%rdx), %ecx |
311 | movl (%rsi), %esi |
312 | movl %ecx, -4(%rdi,%rdx) |
313 | movl %esi, (%rdi) |
314 | ret |
315 | L(between_2_3): |
316 | /* From 2 to 3. No branch when size == 2. */ |
317 | movzwl -2(%rsi,%rdx), %ecx |
318 | movzwl (%rsi), %esi |
319 | movw %cx, -2(%rdi,%rdx) |
320 | movw %si, (%rdi) |
321 | ret |
322 | |
323 | #if defined USE_MULTIARCH && IS_IN (libc) |
324 | L(movsb_more_2x_vec): |
325 | cmpq $REP_MOVSB_THRESHOLD, %rdx |
326 | ja L(movsb) |
327 | #endif |
328 | L(more_2x_vec): |
329 | /* More than 2 * VEC and there may be overlap between destination |
330 | and source. */ |
331 | cmpq $(VEC_SIZE * 8), %rdx |
332 | ja L(more_8x_vec) |
333 | cmpq $(VEC_SIZE * 4), %rdx |
334 | jb L(last_4x_vec) |
335 | /* Copy from 4 * VEC to 8 * VEC, inclusively. */ |
336 | VMOVU (%rsi), %VEC(0) |
337 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
338 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
339 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
340 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(4) |
341 | VMOVU -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(5) |
342 | VMOVU -(VEC_SIZE * 3)(%rsi,%rdx), %VEC(6) |
343 | VMOVU -(VEC_SIZE * 4)(%rsi,%rdx), %VEC(7) |
344 | VMOVU %VEC(0), (%rdi) |
345 | VMOVU %VEC(1), VEC_SIZE(%rdi) |
346 | VMOVU %VEC(2), (VEC_SIZE * 2)(%rdi) |
347 | VMOVU %VEC(3), (VEC_SIZE * 3)(%rdi) |
348 | VMOVU %VEC(4), -VEC_SIZE(%rdi,%rdx) |
349 | VMOVU %VEC(5), -(VEC_SIZE * 2)(%rdi,%rdx) |
350 | VMOVU %VEC(6), -(VEC_SIZE * 3)(%rdi,%rdx) |
351 | VMOVU %VEC(7), -(VEC_SIZE * 4)(%rdi,%rdx) |
352 | VZEROUPPER |
353 | ret |
354 | L(last_4x_vec): |
355 | /* Copy from 2 * VEC to 4 * VEC. */ |
356 | VMOVU (%rsi), %VEC(0) |
357 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
358 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(2) |
359 | VMOVU -(VEC_SIZE * 2)(%rsi,%rdx), %VEC(3) |
360 | VMOVU %VEC(0), (%rdi) |
361 | VMOVU %VEC(1), VEC_SIZE(%rdi) |
362 | VMOVU %VEC(2), -VEC_SIZE(%rdi,%rdx) |
363 | VMOVU %VEC(3), -(VEC_SIZE * 2)(%rdi,%rdx) |
364 | VZEROUPPER |
365 | ret |
366 | |
367 | L(more_8x_vec): |
368 | cmpq %rsi, %rdi |
369 | ja L(more_8x_vec_backward) |
370 | /* Source == destination is less common. */ |
371 | je L(nop) |
372 | /* Load the first VEC and last 4 * VEC to support overlapping |
373 | addresses. */ |
374 | VMOVU (%rsi), %VEC(4) |
375 | VMOVU -VEC_SIZE(%rsi, %rdx), %VEC(5) |
376 | VMOVU -(VEC_SIZE * 2)(%rsi, %rdx), %VEC(6) |
377 | VMOVU -(VEC_SIZE * 3)(%rsi, %rdx), %VEC(7) |
378 | VMOVU -(VEC_SIZE * 4)(%rsi, %rdx), %VEC(8) |
379 | /* Save start and stop of the destination buffer. */ |
380 | movq %rdi, %r11 |
381 | leaq -VEC_SIZE(%rdi, %rdx), %rcx |
382 | /* Align destination for aligned stores in the loop. Compute |
383 | how much destination is misaligned. */ |
384 | movq %rdi, %r8 |
385 | andq $(VEC_SIZE - 1), %r8 |
386 | /* Get the negative of offset for alignment. */ |
387 | subq $VEC_SIZE, %r8 |
388 | /* Adjust source. */ |
389 | subq %r8, %rsi |
390 | /* Adjust destination which should be aligned now. */ |
391 | subq %r8, %rdi |
392 | /* Adjust length. */ |
393 | addq %r8, %rdx |
394 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
395 | /* Check non-temporal store threshold. */ |
396 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
397 | ja L(large_forward) |
398 | #endif |
399 | L(loop_4x_vec_forward): |
400 | /* Copy 4 * VEC a time forward. */ |
401 | VMOVU (%rsi), %VEC(0) |
402 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
403 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
404 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
405 | addq $(VEC_SIZE * 4), %rsi |
406 | subq $(VEC_SIZE * 4), %rdx |
407 | VMOVA %VEC(0), (%rdi) |
408 | VMOVA %VEC(1), VEC_SIZE(%rdi) |
409 | VMOVA %VEC(2), (VEC_SIZE * 2)(%rdi) |
410 | VMOVA %VEC(3), (VEC_SIZE * 3)(%rdi) |
411 | addq $(VEC_SIZE * 4), %rdi |
412 | cmpq $(VEC_SIZE * 4), %rdx |
413 | ja L(loop_4x_vec_forward) |
414 | /* Store the last 4 * VEC. */ |
415 | VMOVU %VEC(5), (%rcx) |
416 | VMOVU %VEC(6), -VEC_SIZE(%rcx) |
417 | VMOVU %VEC(7), -(VEC_SIZE * 2)(%rcx) |
418 | VMOVU %VEC(8), -(VEC_SIZE * 3)(%rcx) |
419 | /* Store the first VEC. */ |
420 | VMOVU %VEC(4), (%r11) |
421 | VZEROUPPER |
422 | ret |
423 | |
424 | L(more_8x_vec_backward): |
425 | /* Load the first 4 * VEC and last VEC to support overlapping |
426 | addresses. */ |
427 | VMOVU (%rsi), %VEC(4) |
428 | VMOVU VEC_SIZE(%rsi), %VEC(5) |
429 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(6) |
430 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(7) |
431 | VMOVU -VEC_SIZE(%rsi,%rdx), %VEC(8) |
432 | /* Save stop of the destination buffer. */ |
433 | leaq -VEC_SIZE(%rdi, %rdx), %r11 |
434 | /* Align destination end for aligned stores in the loop. Compute |
435 | how much destination end is misaligned. */ |
436 | leaq -VEC_SIZE(%rsi, %rdx), %rcx |
437 | movq %r11, %r9 |
438 | movq %r11, %r8 |
439 | andq $(VEC_SIZE - 1), %r8 |
440 | /* Adjust source. */ |
441 | subq %r8, %rcx |
442 | /* Adjust the end of destination which should be aligned now. */ |
443 | subq %r8, %r9 |
444 | /* Adjust length. */ |
445 | subq %r8, %rdx |
446 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
447 | /* Check non-temporal store threshold. */ |
448 | cmpq __x86_shared_non_temporal_threshold(%rip), %rdx |
449 | ja L(large_backward) |
450 | #endif |
451 | L(loop_4x_vec_backward): |
452 | /* Copy 4 * VEC a time backward. */ |
453 | VMOVU (%rcx), %VEC(0) |
454 | VMOVU -VEC_SIZE(%rcx), %VEC(1) |
455 | VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2) |
456 | VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3) |
457 | subq $(VEC_SIZE * 4), %rcx |
458 | subq $(VEC_SIZE * 4), %rdx |
459 | VMOVA %VEC(0), (%r9) |
460 | VMOVA %VEC(1), -VEC_SIZE(%r9) |
461 | VMOVA %VEC(2), -(VEC_SIZE * 2)(%r9) |
462 | VMOVA %VEC(3), -(VEC_SIZE * 3)(%r9) |
463 | subq $(VEC_SIZE * 4), %r9 |
464 | cmpq $(VEC_SIZE * 4), %rdx |
465 | ja L(loop_4x_vec_backward) |
466 | /* Store the first 4 * VEC. */ |
467 | VMOVU %VEC(4), (%rdi) |
468 | VMOVU %VEC(5), VEC_SIZE(%rdi) |
469 | VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi) |
470 | VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi) |
471 | /* Store the last VEC. */ |
472 | VMOVU %VEC(8), (%r11) |
473 | VZEROUPPER |
474 | ret |
475 | |
476 | #if (defined USE_MULTIARCH || VEC_SIZE == 16) && IS_IN (libc) |
477 | L(large_forward): |
478 | /* Don't use non-temporal store if there is overlap between |
479 | destination and source since destination may be in cache |
480 | when source is loaded. */ |
481 | leaq (%rdi, %rdx), %r10 |
482 | cmpq %r10, %rsi |
483 | jb L(loop_4x_vec_forward) |
484 | L(loop_large_forward): |
485 | /* Copy 4 * VEC a time forward with non-temporal stores. */ |
486 | PREFETCH_ONE_SET (1, (%rsi), PREFETCHED_LOAD_SIZE * 2) |
487 | PREFETCH_ONE_SET (1, (%rsi), PREFETCHED_LOAD_SIZE * 3) |
488 | VMOVU (%rsi), %VEC(0) |
489 | VMOVU VEC_SIZE(%rsi), %VEC(1) |
490 | VMOVU (VEC_SIZE * 2)(%rsi), %VEC(2) |
491 | VMOVU (VEC_SIZE * 3)(%rsi), %VEC(3) |
492 | addq $PREFETCHED_LOAD_SIZE, %rsi |
493 | subq $PREFETCHED_LOAD_SIZE, %rdx |
494 | VMOVNT %VEC(0), (%rdi) |
495 | VMOVNT %VEC(1), VEC_SIZE(%rdi) |
496 | VMOVNT %VEC(2), (VEC_SIZE * 2)(%rdi) |
497 | VMOVNT %VEC(3), (VEC_SIZE * 3)(%rdi) |
498 | addq $PREFETCHED_LOAD_SIZE, %rdi |
499 | cmpq $PREFETCHED_LOAD_SIZE, %rdx |
500 | ja L(loop_large_forward) |
501 | sfence |
502 | /* Store the last 4 * VEC. */ |
503 | VMOVU %VEC(5), (%rcx) |
504 | VMOVU %VEC(6), -VEC_SIZE(%rcx) |
505 | VMOVU %VEC(7), -(VEC_SIZE * 2)(%rcx) |
506 | VMOVU %VEC(8), -(VEC_SIZE * 3)(%rcx) |
507 | /* Store the first VEC. */ |
508 | VMOVU %VEC(4), (%r11) |
509 | VZEROUPPER |
510 | ret |
511 | |
512 | L(large_backward): |
513 | /* Don't use non-temporal store if there is overlap between |
514 | destination and source since destination may be in cache |
515 | when source is loaded. */ |
516 | leaq (%rcx, %rdx), %r10 |
517 | cmpq %r10, %r9 |
518 | jb L(loop_4x_vec_backward) |
519 | L(loop_large_backward): |
520 | /* Copy 4 * VEC a time backward with non-temporal stores. */ |
521 | PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 2) |
522 | PREFETCH_ONE_SET (-1, (%rcx), -PREFETCHED_LOAD_SIZE * 3) |
523 | VMOVU (%rcx), %VEC(0) |
524 | VMOVU -VEC_SIZE(%rcx), %VEC(1) |
525 | VMOVU -(VEC_SIZE * 2)(%rcx), %VEC(2) |
526 | VMOVU -(VEC_SIZE * 3)(%rcx), %VEC(3) |
527 | subq $PREFETCHED_LOAD_SIZE, %rcx |
528 | subq $PREFETCHED_LOAD_SIZE, %rdx |
529 | VMOVNT %VEC(0), (%r9) |
530 | VMOVNT %VEC(1), -VEC_SIZE(%r9) |
531 | VMOVNT %VEC(2), -(VEC_SIZE * 2)(%r9) |
532 | VMOVNT %VEC(3), -(VEC_SIZE * 3)(%r9) |
533 | subq $PREFETCHED_LOAD_SIZE, %r9 |
534 | cmpq $PREFETCHED_LOAD_SIZE, %rdx |
535 | ja L(loop_large_backward) |
536 | sfence |
537 | /* Store the first 4 * VEC. */ |
538 | VMOVU %VEC(4), (%rdi) |
539 | VMOVU %VEC(5), VEC_SIZE(%rdi) |
540 | VMOVU %VEC(6), (VEC_SIZE * 2)(%rdi) |
541 | VMOVU %VEC(7), (VEC_SIZE * 3)(%rdi) |
542 | /* Store the last VEC. */ |
543 | VMOVU %VEC(8), (%r11) |
544 | VZEROUPPER |
545 | ret |
546 | #endif |
547 | END (MEMMOVE_SYMBOL (__memmove, unaligned_erms)) |
548 | |
549 | #ifdef SHARED |
550 | # if IS_IN (libc) |
551 | # ifdef USE_MULTIARCH |
552 | strong_alias (MEMMOVE_SYMBOL (__memmove, unaligned_erms), |
553 | MEMMOVE_SYMBOL (__memcpy, unaligned_erms)) |
554 | strong_alias (MEMMOVE_SYMBOL (__memmove_chk, unaligned_erms), |
555 | MEMMOVE_SYMBOL (__memcpy_chk, unaligned_erms)) |
556 | # endif |
557 | strong_alias (MEMMOVE_CHK_SYMBOL (__memmove_chk, unaligned), |
558 | MEMMOVE_CHK_SYMBOL (__memcpy_chk, unaligned)) |
559 | # endif |
560 | #endif |
561 | #if VEC_SIZE == 16 || defined SHARED |
562 | strong_alias (MEMMOVE_SYMBOL (__memmove, unaligned), |
563 | MEMCPY_SYMBOL (__memcpy, unaligned)) |
564 | #endif |
565 | |