1 | /* memcpy with SSSE3 and REP string |
2 | Copyright (C) 2010-2017 Free Software Foundation, Inc. |
3 | Contributed by Intel Corporation. |
4 | This file is part of the GNU C Library. |
5 | |
6 | The GNU C Library is free software; you can redistribute it and/or |
7 | modify it under the terms of the GNU Lesser General Public |
8 | License as published by the Free Software Foundation; either |
9 | version 2.1 of the License, or (at your option) any later version. |
10 | |
11 | The GNU C Library is distributed in the hope that it will be useful, |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
14 | Lesser General Public License for more details. |
15 | |
16 | You should have received a copy of the GNU Lesser General Public |
17 | License along with the GNU C Library; if not, see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include <sysdep.h> |
21 | |
22 | #if IS_IN (libc) \ |
23 | && (defined SHARED \ |
24 | || defined USE_AS_MEMMOVE \ |
25 | || !defined USE_MULTIARCH) |
26 | |
27 | #include "asm-syntax.h" |
28 | |
29 | #ifndef MEMCPY |
30 | # define MEMCPY __memcpy_ssse3_back |
31 | # define MEMCPY_CHK __memcpy_chk_ssse3_back |
32 | # define MEMPCPY __mempcpy_ssse3_back |
33 | # define MEMPCPY_CHK __mempcpy_chk_ssse3_back |
34 | #endif |
35 | |
36 | #define JMPTBL(I, B) I - B |
37 | |
38 | /* Branch to an entry in a jump table. TABLE is a jump table with |
39 | relative offsets. INDEX is a register contains the index into the |
40 | jump table. SCALE is the scale of INDEX. */ |
41 | #define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ |
42 | lea TABLE(%rip), %r11; \ |
43 | movslq (%r11, INDEX, SCALE), INDEX; \ |
44 | lea (%r11, INDEX), INDEX; \ |
45 | jmp *INDEX; \ |
46 | ud2 |
47 | |
48 | .section .text.ssse3,"ax" ,@progbits |
49 | #if !defined USE_AS_MEMPCPY && !defined USE_AS_MEMMOVE |
50 | ENTRY (MEMPCPY_CHK) |
51 | cmpq %rdx, %rcx |
52 | jb HIDDEN_JUMPTARGET (__chk_fail) |
53 | END (MEMPCPY_CHK) |
54 | |
55 | ENTRY (MEMPCPY) |
56 | movq %rdi, %rax |
57 | addq %rdx, %rax |
58 | jmp L(start) |
59 | END (MEMPCPY) |
60 | #endif |
61 | |
62 | #if !defined USE_AS_BCOPY |
63 | ENTRY (MEMCPY_CHK) |
64 | cmpq %rdx, %rcx |
65 | jb HIDDEN_JUMPTARGET (__chk_fail) |
66 | END (MEMCPY_CHK) |
67 | #endif |
68 | |
69 | ENTRY (MEMCPY) |
70 | mov %rdi, %rax |
71 | #ifdef USE_AS_MEMPCPY |
72 | add %rdx, %rax |
73 | #endif |
74 | |
75 | #ifdef USE_AS_MEMMOVE |
76 | cmp %rsi, %rdi |
77 | jb L(copy_forward) |
78 | je L(bwd_write_0bytes) |
79 | cmp $144, %rdx |
80 | jae L(copy_backward) |
81 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
82 | L(copy_forward): |
83 | #endif |
84 | L(start): |
85 | cmp $144, %rdx |
86 | jae L(144bytesormore) |
87 | |
88 | L(fwd_write_less32bytes): |
89 | #ifndef USE_AS_MEMMOVE |
90 | cmp %dil, %sil |
91 | jbe L(bk_write) |
92 | #endif |
93 | add %rdx, %rsi |
94 | add %rdx, %rdi |
95 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
96 | #ifndef USE_AS_MEMMOVE |
97 | L(bk_write): |
98 | |
99 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
100 | #endif |
101 | |
102 | .p2align 4 |
103 | L(144bytesormore): |
104 | |
105 | #ifndef USE_AS_MEMMOVE |
106 | cmp %dil, %sil |
107 | jle L(copy_backward) |
108 | #endif |
109 | movdqu (%rsi), %xmm0 |
110 | mov %rdi, %r8 |
111 | and $-16, %rdi |
112 | add $16, %rdi |
113 | mov %rdi, %r9 |
114 | sub %r8, %r9 |
115 | sub %r9, %rdx |
116 | add %r9, %rsi |
117 | mov %rsi, %r9 |
118 | and $0xf, %r9 |
119 | jz L(shl_0) |
120 | #ifdef DATA_CACHE_SIZE |
121 | mov $DATA_CACHE_SIZE, %RCX_LP |
122 | #else |
123 | mov __x86_data_cache_size(%rip), %RCX_LP |
124 | #endif |
125 | cmp %rcx, %rdx |
126 | jae L(gobble_mem_fwd) |
127 | lea L(shl_table_fwd)(%rip), %r11 |
128 | sub $0x80, %rdx |
129 | movslq (%r11, %r9, 4), %r9 |
130 | add %r11, %r9 |
131 | jmp *%r9 |
132 | ud2 |
133 | |
134 | .p2align 4 |
135 | L(copy_backward): |
136 | #ifdef DATA_CACHE_SIZE |
137 | mov $DATA_CACHE_SIZE, %RCX_LP |
138 | #else |
139 | mov __x86_data_cache_size(%rip), %RCX_LP |
140 | #endif |
141 | shl $1, %rcx |
142 | cmp %rcx, %rdx |
143 | ja L(gobble_mem_bwd) |
144 | |
145 | add %rdx, %rdi |
146 | add %rdx, %rsi |
147 | movdqu -16(%rsi), %xmm0 |
148 | lea -16(%rdi), %r8 |
149 | mov %rdi, %r9 |
150 | and $0xf, %r9 |
151 | xor %r9, %rdi |
152 | sub %r9, %rsi |
153 | sub %r9, %rdx |
154 | mov %rsi, %r9 |
155 | and $0xf, %r9 |
156 | jz L(shl_0_bwd) |
157 | lea L(shl_table_bwd)(%rip), %r11 |
158 | sub $0x80, %rdx |
159 | movslq (%r11, %r9, 4), %r9 |
160 | add %r11, %r9 |
161 | jmp *%r9 |
162 | ud2 |
163 | |
164 | .p2align 4 |
165 | L(shl_0): |
166 | |
167 | mov %rdx, %r9 |
168 | shr $8, %r9 |
169 | add %rdx, %r9 |
170 | #ifdef DATA_CACHE_SIZE |
171 | cmp $DATA_CACHE_SIZE_HALF, %R9_LP |
172 | #else |
173 | cmp __x86_data_cache_size_half(%rip), %R9_LP |
174 | #endif |
175 | jae L(gobble_mem_fwd) |
176 | sub $0x80, %rdx |
177 | .p2align 4 |
178 | L(shl_0_loop): |
179 | movdqa (%rsi), %xmm1 |
180 | movdqa %xmm1, (%rdi) |
181 | movaps 0x10(%rsi), %xmm2 |
182 | movaps %xmm2, 0x10(%rdi) |
183 | movaps 0x20(%rsi), %xmm3 |
184 | movaps %xmm3, 0x20(%rdi) |
185 | movaps 0x30(%rsi), %xmm4 |
186 | movaps %xmm4, 0x30(%rdi) |
187 | movaps 0x40(%rsi), %xmm1 |
188 | movaps %xmm1, 0x40(%rdi) |
189 | movaps 0x50(%rsi), %xmm2 |
190 | movaps %xmm2, 0x50(%rdi) |
191 | movaps 0x60(%rsi), %xmm3 |
192 | movaps %xmm3, 0x60(%rdi) |
193 | movaps 0x70(%rsi), %xmm4 |
194 | movaps %xmm4, 0x70(%rdi) |
195 | sub $0x80, %rdx |
196 | lea 0x80(%rsi), %rsi |
197 | lea 0x80(%rdi), %rdi |
198 | jae L(shl_0_loop) |
199 | movdqu %xmm0, (%r8) |
200 | add $0x80, %rdx |
201 | add %rdx, %rsi |
202 | add %rdx, %rdi |
203 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
204 | |
205 | .p2align 4 |
206 | L(shl_0_bwd): |
207 | sub $0x80, %rdx |
208 | L(copy_backward_loop): |
209 | movaps -0x10(%rsi), %xmm1 |
210 | movaps %xmm1, -0x10(%rdi) |
211 | movaps -0x20(%rsi), %xmm2 |
212 | movaps %xmm2, -0x20(%rdi) |
213 | movaps -0x30(%rsi), %xmm3 |
214 | movaps %xmm3, -0x30(%rdi) |
215 | movaps -0x40(%rsi), %xmm4 |
216 | movaps %xmm4, -0x40(%rdi) |
217 | movaps -0x50(%rsi), %xmm5 |
218 | movaps %xmm5, -0x50(%rdi) |
219 | movaps -0x60(%rsi), %xmm5 |
220 | movaps %xmm5, -0x60(%rdi) |
221 | movaps -0x70(%rsi), %xmm5 |
222 | movaps %xmm5, -0x70(%rdi) |
223 | movaps -0x80(%rsi), %xmm5 |
224 | movaps %xmm5, -0x80(%rdi) |
225 | sub $0x80, %rdx |
226 | lea -0x80(%rdi), %rdi |
227 | lea -0x80(%rsi), %rsi |
228 | jae L(copy_backward_loop) |
229 | |
230 | movdqu %xmm0, (%r8) |
231 | add $0x80, %rdx |
232 | sub %rdx, %rdi |
233 | sub %rdx, %rsi |
234 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
235 | |
236 | .p2align 4 |
237 | L(shl_1): |
238 | sub $0x80, %rdx |
239 | movaps -0x01(%rsi), %xmm1 |
240 | movaps 0x0f(%rsi), %xmm2 |
241 | movaps 0x1f(%rsi), %xmm3 |
242 | movaps 0x2f(%rsi), %xmm4 |
243 | movaps 0x3f(%rsi), %xmm5 |
244 | movaps 0x4f(%rsi), %xmm6 |
245 | movaps 0x5f(%rsi), %xmm7 |
246 | movaps 0x6f(%rsi), %xmm8 |
247 | movaps 0x7f(%rsi), %xmm9 |
248 | lea 0x80(%rsi), %rsi |
249 | palignr $1, %xmm8, %xmm9 |
250 | movaps %xmm9, 0x70(%rdi) |
251 | palignr $1, %xmm7, %xmm8 |
252 | movaps %xmm8, 0x60(%rdi) |
253 | palignr $1, %xmm6, %xmm7 |
254 | movaps %xmm7, 0x50(%rdi) |
255 | palignr $1, %xmm5, %xmm6 |
256 | movaps %xmm6, 0x40(%rdi) |
257 | palignr $1, %xmm4, %xmm5 |
258 | movaps %xmm5, 0x30(%rdi) |
259 | palignr $1, %xmm3, %xmm4 |
260 | movaps %xmm4, 0x20(%rdi) |
261 | palignr $1, %xmm2, %xmm3 |
262 | movaps %xmm3, 0x10(%rdi) |
263 | palignr $1, %xmm1, %xmm2 |
264 | movaps %xmm2, (%rdi) |
265 | lea 0x80(%rdi), %rdi |
266 | jae L(shl_1) |
267 | movdqu %xmm0, (%r8) |
268 | add $0x80, %rdx |
269 | add %rdx, %rdi |
270 | add %rdx, %rsi |
271 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
272 | |
273 | .p2align 4 |
274 | L(shl_1_bwd): |
275 | movaps -0x01(%rsi), %xmm1 |
276 | |
277 | movaps -0x11(%rsi), %xmm2 |
278 | palignr $1, %xmm2, %xmm1 |
279 | movaps %xmm1, -0x10(%rdi) |
280 | |
281 | movaps -0x21(%rsi), %xmm3 |
282 | palignr $1, %xmm3, %xmm2 |
283 | movaps %xmm2, -0x20(%rdi) |
284 | |
285 | movaps -0x31(%rsi), %xmm4 |
286 | palignr $1, %xmm4, %xmm3 |
287 | movaps %xmm3, -0x30(%rdi) |
288 | |
289 | movaps -0x41(%rsi), %xmm5 |
290 | palignr $1, %xmm5, %xmm4 |
291 | movaps %xmm4, -0x40(%rdi) |
292 | |
293 | movaps -0x51(%rsi), %xmm6 |
294 | palignr $1, %xmm6, %xmm5 |
295 | movaps %xmm5, -0x50(%rdi) |
296 | |
297 | movaps -0x61(%rsi), %xmm7 |
298 | palignr $1, %xmm7, %xmm6 |
299 | movaps %xmm6, -0x60(%rdi) |
300 | |
301 | movaps -0x71(%rsi), %xmm8 |
302 | palignr $1, %xmm8, %xmm7 |
303 | movaps %xmm7, -0x70(%rdi) |
304 | |
305 | movaps -0x81(%rsi), %xmm9 |
306 | palignr $1, %xmm9, %xmm8 |
307 | movaps %xmm8, -0x80(%rdi) |
308 | |
309 | sub $0x80, %rdx |
310 | lea -0x80(%rdi), %rdi |
311 | lea -0x80(%rsi), %rsi |
312 | jae L(shl_1_bwd) |
313 | movdqu %xmm0, (%r8) |
314 | add $0x80, %rdx |
315 | sub %rdx, %rdi |
316 | sub %rdx, %rsi |
317 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
318 | |
319 | .p2align 4 |
320 | L(shl_2): |
321 | sub $0x80, %rdx |
322 | movaps -0x02(%rsi), %xmm1 |
323 | movaps 0x0e(%rsi), %xmm2 |
324 | movaps 0x1e(%rsi), %xmm3 |
325 | movaps 0x2e(%rsi), %xmm4 |
326 | movaps 0x3e(%rsi), %xmm5 |
327 | movaps 0x4e(%rsi), %xmm6 |
328 | movaps 0x5e(%rsi), %xmm7 |
329 | movaps 0x6e(%rsi), %xmm8 |
330 | movaps 0x7e(%rsi), %xmm9 |
331 | lea 0x80(%rsi), %rsi |
332 | palignr $2, %xmm8, %xmm9 |
333 | movaps %xmm9, 0x70(%rdi) |
334 | palignr $2, %xmm7, %xmm8 |
335 | movaps %xmm8, 0x60(%rdi) |
336 | palignr $2, %xmm6, %xmm7 |
337 | movaps %xmm7, 0x50(%rdi) |
338 | palignr $2, %xmm5, %xmm6 |
339 | movaps %xmm6, 0x40(%rdi) |
340 | palignr $2, %xmm4, %xmm5 |
341 | movaps %xmm5, 0x30(%rdi) |
342 | palignr $2, %xmm3, %xmm4 |
343 | movaps %xmm4, 0x20(%rdi) |
344 | palignr $2, %xmm2, %xmm3 |
345 | movaps %xmm3, 0x10(%rdi) |
346 | palignr $2, %xmm1, %xmm2 |
347 | movaps %xmm2, (%rdi) |
348 | lea 0x80(%rdi), %rdi |
349 | jae L(shl_2) |
350 | movdqu %xmm0, (%r8) |
351 | add $0x80, %rdx |
352 | add %rdx, %rdi |
353 | add %rdx, %rsi |
354 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
355 | |
356 | .p2align 4 |
357 | L(shl_2_bwd): |
358 | movaps -0x02(%rsi), %xmm1 |
359 | |
360 | movaps -0x12(%rsi), %xmm2 |
361 | palignr $2, %xmm2, %xmm1 |
362 | movaps %xmm1, -0x10(%rdi) |
363 | |
364 | movaps -0x22(%rsi), %xmm3 |
365 | palignr $2, %xmm3, %xmm2 |
366 | movaps %xmm2, -0x20(%rdi) |
367 | |
368 | movaps -0x32(%rsi), %xmm4 |
369 | palignr $2, %xmm4, %xmm3 |
370 | movaps %xmm3, -0x30(%rdi) |
371 | |
372 | movaps -0x42(%rsi), %xmm5 |
373 | palignr $2, %xmm5, %xmm4 |
374 | movaps %xmm4, -0x40(%rdi) |
375 | |
376 | movaps -0x52(%rsi), %xmm6 |
377 | palignr $2, %xmm6, %xmm5 |
378 | movaps %xmm5, -0x50(%rdi) |
379 | |
380 | movaps -0x62(%rsi), %xmm7 |
381 | palignr $2, %xmm7, %xmm6 |
382 | movaps %xmm6, -0x60(%rdi) |
383 | |
384 | movaps -0x72(%rsi), %xmm8 |
385 | palignr $2, %xmm8, %xmm7 |
386 | movaps %xmm7, -0x70(%rdi) |
387 | |
388 | movaps -0x82(%rsi), %xmm9 |
389 | palignr $2, %xmm9, %xmm8 |
390 | movaps %xmm8, -0x80(%rdi) |
391 | |
392 | sub $0x80, %rdx |
393 | lea -0x80(%rdi), %rdi |
394 | lea -0x80(%rsi), %rsi |
395 | jae L(shl_2_bwd) |
396 | movdqu %xmm0, (%r8) |
397 | add $0x80, %rdx |
398 | sub %rdx, %rdi |
399 | sub %rdx, %rsi |
400 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
401 | |
402 | .p2align 4 |
403 | L(shl_3): |
404 | sub $0x80, %rdx |
405 | movaps -0x03(%rsi), %xmm1 |
406 | movaps 0x0d(%rsi), %xmm2 |
407 | movaps 0x1d(%rsi), %xmm3 |
408 | movaps 0x2d(%rsi), %xmm4 |
409 | movaps 0x3d(%rsi), %xmm5 |
410 | movaps 0x4d(%rsi), %xmm6 |
411 | movaps 0x5d(%rsi), %xmm7 |
412 | movaps 0x6d(%rsi), %xmm8 |
413 | movaps 0x7d(%rsi), %xmm9 |
414 | lea 0x80(%rsi), %rsi |
415 | palignr $3, %xmm8, %xmm9 |
416 | movaps %xmm9, 0x70(%rdi) |
417 | palignr $3, %xmm7, %xmm8 |
418 | movaps %xmm8, 0x60(%rdi) |
419 | palignr $3, %xmm6, %xmm7 |
420 | movaps %xmm7, 0x50(%rdi) |
421 | palignr $3, %xmm5, %xmm6 |
422 | movaps %xmm6, 0x40(%rdi) |
423 | palignr $3, %xmm4, %xmm5 |
424 | movaps %xmm5, 0x30(%rdi) |
425 | palignr $3, %xmm3, %xmm4 |
426 | movaps %xmm4, 0x20(%rdi) |
427 | palignr $3, %xmm2, %xmm3 |
428 | movaps %xmm3, 0x10(%rdi) |
429 | palignr $3, %xmm1, %xmm2 |
430 | movaps %xmm2, (%rdi) |
431 | lea 0x80(%rdi), %rdi |
432 | jae L(shl_3) |
433 | movdqu %xmm0, (%r8) |
434 | add $0x80, %rdx |
435 | add %rdx, %rdi |
436 | add %rdx, %rsi |
437 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
438 | |
439 | .p2align 4 |
440 | L(shl_3_bwd): |
441 | movaps -0x03(%rsi), %xmm1 |
442 | |
443 | movaps -0x13(%rsi), %xmm2 |
444 | palignr $3, %xmm2, %xmm1 |
445 | movaps %xmm1, -0x10(%rdi) |
446 | |
447 | movaps -0x23(%rsi), %xmm3 |
448 | palignr $3, %xmm3, %xmm2 |
449 | movaps %xmm2, -0x20(%rdi) |
450 | |
451 | movaps -0x33(%rsi), %xmm4 |
452 | palignr $3, %xmm4, %xmm3 |
453 | movaps %xmm3, -0x30(%rdi) |
454 | |
455 | movaps -0x43(%rsi), %xmm5 |
456 | palignr $3, %xmm5, %xmm4 |
457 | movaps %xmm4, -0x40(%rdi) |
458 | |
459 | movaps -0x53(%rsi), %xmm6 |
460 | palignr $3, %xmm6, %xmm5 |
461 | movaps %xmm5, -0x50(%rdi) |
462 | |
463 | movaps -0x63(%rsi), %xmm7 |
464 | palignr $3, %xmm7, %xmm6 |
465 | movaps %xmm6, -0x60(%rdi) |
466 | |
467 | movaps -0x73(%rsi), %xmm8 |
468 | palignr $3, %xmm8, %xmm7 |
469 | movaps %xmm7, -0x70(%rdi) |
470 | |
471 | movaps -0x83(%rsi), %xmm9 |
472 | palignr $3, %xmm9, %xmm8 |
473 | movaps %xmm8, -0x80(%rdi) |
474 | |
475 | sub $0x80, %rdx |
476 | lea -0x80(%rdi), %rdi |
477 | lea -0x80(%rsi), %rsi |
478 | jae L(shl_3_bwd) |
479 | movdqu %xmm0, (%r8) |
480 | add $0x80, %rdx |
481 | sub %rdx, %rdi |
482 | sub %rdx, %rsi |
483 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
484 | |
485 | .p2align 4 |
486 | L(shl_4): |
487 | sub $0x80, %rdx |
488 | movaps -0x04(%rsi), %xmm1 |
489 | movaps 0x0c(%rsi), %xmm2 |
490 | movaps 0x1c(%rsi), %xmm3 |
491 | movaps 0x2c(%rsi), %xmm4 |
492 | movaps 0x3c(%rsi), %xmm5 |
493 | movaps 0x4c(%rsi), %xmm6 |
494 | movaps 0x5c(%rsi), %xmm7 |
495 | movaps 0x6c(%rsi), %xmm8 |
496 | movaps 0x7c(%rsi), %xmm9 |
497 | lea 0x80(%rsi), %rsi |
498 | palignr $4, %xmm8, %xmm9 |
499 | movaps %xmm9, 0x70(%rdi) |
500 | palignr $4, %xmm7, %xmm8 |
501 | movaps %xmm8, 0x60(%rdi) |
502 | palignr $4, %xmm6, %xmm7 |
503 | movaps %xmm7, 0x50(%rdi) |
504 | palignr $4, %xmm5, %xmm6 |
505 | movaps %xmm6, 0x40(%rdi) |
506 | palignr $4, %xmm4, %xmm5 |
507 | movaps %xmm5, 0x30(%rdi) |
508 | palignr $4, %xmm3, %xmm4 |
509 | movaps %xmm4, 0x20(%rdi) |
510 | palignr $4, %xmm2, %xmm3 |
511 | movaps %xmm3, 0x10(%rdi) |
512 | palignr $4, %xmm1, %xmm2 |
513 | movaps %xmm2, (%rdi) |
514 | lea 0x80(%rdi), %rdi |
515 | jae L(shl_4) |
516 | movdqu %xmm0, (%r8) |
517 | add $0x80, %rdx |
518 | add %rdx, %rdi |
519 | add %rdx, %rsi |
520 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
521 | |
522 | .p2align 4 |
523 | L(shl_4_bwd): |
524 | movaps -0x04(%rsi), %xmm1 |
525 | |
526 | movaps -0x14(%rsi), %xmm2 |
527 | palignr $4, %xmm2, %xmm1 |
528 | movaps %xmm1, -0x10(%rdi) |
529 | |
530 | movaps -0x24(%rsi), %xmm3 |
531 | palignr $4, %xmm3, %xmm2 |
532 | movaps %xmm2, -0x20(%rdi) |
533 | |
534 | movaps -0x34(%rsi), %xmm4 |
535 | palignr $4, %xmm4, %xmm3 |
536 | movaps %xmm3, -0x30(%rdi) |
537 | |
538 | movaps -0x44(%rsi), %xmm5 |
539 | palignr $4, %xmm5, %xmm4 |
540 | movaps %xmm4, -0x40(%rdi) |
541 | |
542 | movaps -0x54(%rsi), %xmm6 |
543 | palignr $4, %xmm6, %xmm5 |
544 | movaps %xmm5, -0x50(%rdi) |
545 | |
546 | movaps -0x64(%rsi), %xmm7 |
547 | palignr $4, %xmm7, %xmm6 |
548 | movaps %xmm6, -0x60(%rdi) |
549 | |
550 | movaps -0x74(%rsi), %xmm8 |
551 | palignr $4, %xmm8, %xmm7 |
552 | movaps %xmm7, -0x70(%rdi) |
553 | |
554 | movaps -0x84(%rsi), %xmm9 |
555 | palignr $4, %xmm9, %xmm8 |
556 | movaps %xmm8, -0x80(%rdi) |
557 | |
558 | sub $0x80, %rdx |
559 | lea -0x80(%rdi), %rdi |
560 | lea -0x80(%rsi), %rsi |
561 | jae L(shl_4_bwd) |
562 | movdqu %xmm0, (%r8) |
563 | add $0x80, %rdx |
564 | sub %rdx, %rdi |
565 | sub %rdx, %rsi |
566 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
567 | |
568 | .p2align 4 |
569 | L(shl_5): |
570 | sub $0x80, %rdx |
571 | movaps -0x05(%rsi), %xmm1 |
572 | movaps 0x0b(%rsi), %xmm2 |
573 | movaps 0x1b(%rsi), %xmm3 |
574 | movaps 0x2b(%rsi), %xmm4 |
575 | movaps 0x3b(%rsi), %xmm5 |
576 | movaps 0x4b(%rsi), %xmm6 |
577 | movaps 0x5b(%rsi), %xmm7 |
578 | movaps 0x6b(%rsi), %xmm8 |
579 | movaps 0x7b(%rsi), %xmm9 |
580 | lea 0x80(%rsi), %rsi |
581 | palignr $5, %xmm8, %xmm9 |
582 | movaps %xmm9, 0x70(%rdi) |
583 | palignr $5, %xmm7, %xmm8 |
584 | movaps %xmm8, 0x60(%rdi) |
585 | palignr $5, %xmm6, %xmm7 |
586 | movaps %xmm7, 0x50(%rdi) |
587 | palignr $5, %xmm5, %xmm6 |
588 | movaps %xmm6, 0x40(%rdi) |
589 | palignr $5, %xmm4, %xmm5 |
590 | movaps %xmm5, 0x30(%rdi) |
591 | palignr $5, %xmm3, %xmm4 |
592 | movaps %xmm4, 0x20(%rdi) |
593 | palignr $5, %xmm2, %xmm3 |
594 | movaps %xmm3, 0x10(%rdi) |
595 | palignr $5, %xmm1, %xmm2 |
596 | movaps %xmm2, (%rdi) |
597 | lea 0x80(%rdi), %rdi |
598 | jae L(shl_5) |
599 | movdqu %xmm0, (%r8) |
600 | add $0x80, %rdx |
601 | add %rdx, %rdi |
602 | add %rdx, %rsi |
603 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
604 | |
605 | .p2align 4 |
606 | L(shl_5_bwd): |
607 | movaps -0x05(%rsi), %xmm1 |
608 | |
609 | movaps -0x15(%rsi), %xmm2 |
610 | palignr $5, %xmm2, %xmm1 |
611 | movaps %xmm1, -0x10(%rdi) |
612 | |
613 | movaps -0x25(%rsi), %xmm3 |
614 | palignr $5, %xmm3, %xmm2 |
615 | movaps %xmm2, -0x20(%rdi) |
616 | |
617 | movaps -0x35(%rsi), %xmm4 |
618 | palignr $5, %xmm4, %xmm3 |
619 | movaps %xmm3, -0x30(%rdi) |
620 | |
621 | movaps -0x45(%rsi), %xmm5 |
622 | palignr $5, %xmm5, %xmm4 |
623 | movaps %xmm4, -0x40(%rdi) |
624 | |
625 | movaps -0x55(%rsi), %xmm6 |
626 | palignr $5, %xmm6, %xmm5 |
627 | movaps %xmm5, -0x50(%rdi) |
628 | |
629 | movaps -0x65(%rsi), %xmm7 |
630 | palignr $5, %xmm7, %xmm6 |
631 | movaps %xmm6, -0x60(%rdi) |
632 | |
633 | movaps -0x75(%rsi), %xmm8 |
634 | palignr $5, %xmm8, %xmm7 |
635 | movaps %xmm7, -0x70(%rdi) |
636 | |
637 | movaps -0x85(%rsi), %xmm9 |
638 | palignr $5, %xmm9, %xmm8 |
639 | movaps %xmm8, -0x80(%rdi) |
640 | |
641 | sub $0x80, %rdx |
642 | lea -0x80(%rdi), %rdi |
643 | lea -0x80(%rsi), %rsi |
644 | jae L(shl_5_bwd) |
645 | movdqu %xmm0, (%r8) |
646 | add $0x80, %rdx |
647 | sub %rdx, %rdi |
648 | sub %rdx, %rsi |
649 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
650 | |
651 | .p2align 4 |
652 | L(shl_6): |
653 | sub $0x80, %rdx |
654 | movaps -0x06(%rsi), %xmm1 |
655 | movaps 0x0a(%rsi), %xmm2 |
656 | movaps 0x1a(%rsi), %xmm3 |
657 | movaps 0x2a(%rsi), %xmm4 |
658 | movaps 0x3a(%rsi), %xmm5 |
659 | movaps 0x4a(%rsi), %xmm6 |
660 | movaps 0x5a(%rsi), %xmm7 |
661 | movaps 0x6a(%rsi), %xmm8 |
662 | movaps 0x7a(%rsi), %xmm9 |
663 | lea 0x80(%rsi), %rsi |
664 | palignr $6, %xmm8, %xmm9 |
665 | movaps %xmm9, 0x70(%rdi) |
666 | palignr $6, %xmm7, %xmm8 |
667 | movaps %xmm8, 0x60(%rdi) |
668 | palignr $6, %xmm6, %xmm7 |
669 | movaps %xmm7, 0x50(%rdi) |
670 | palignr $6, %xmm5, %xmm6 |
671 | movaps %xmm6, 0x40(%rdi) |
672 | palignr $6, %xmm4, %xmm5 |
673 | movaps %xmm5, 0x30(%rdi) |
674 | palignr $6, %xmm3, %xmm4 |
675 | movaps %xmm4, 0x20(%rdi) |
676 | palignr $6, %xmm2, %xmm3 |
677 | movaps %xmm3, 0x10(%rdi) |
678 | palignr $6, %xmm1, %xmm2 |
679 | movaps %xmm2, (%rdi) |
680 | lea 0x80(%rdi), %rdi |
681 | jae L(shl_6) |
682 | movdqu %xmm0, (%r8) |
683 | add $0x80, %rdx |
684 | add %rdx, %rdi |
685 | add %rdx, %rsi |
686 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
687 | |
688 | .p2align 4 |
689 | L(shl_6_bwd): |
690 | movaps -0x06(%rsi), %xmm1 |
691 | |
692 | movaps -0x16(%rsi), %xmm2 |
693 | palignr $6, %xmm2, %xmm1 |
694 | movaps %xmm1, -0x10(%rdi) |
695 | |
696 | movaps -0x26(%rsi), %xmm3 |
697 | palignr $6, %xmm3, %xmm2 |
698 | movaps %xmm2, -0x20(%rdi) |
699 | |
700 | movaps -0x36(%rsi), %xmm4 |
701 | palignr $6, %xmm4, %xmm3 |
702 | movaps %xmm3, -0x30(%rdi) |
703 | |
704 | movaps -0x46(%rsi), %xmm5 |
705 | palignr $6, %xmm5, %xmm4 |
706 | movaps %xmm4, -0x40(%rdi) |
707 | |
708 | movaps -0x56(%rsi), %xmm6 |
709 | palignr $6, %xmm6, %xmm5 |
710 | movaps %xmm5, -0x50(%rdi) |
711 | |
712 | movaps -0x66(%rsi), %xmm7 |
713 | palignr $6, %xmm7, %xmm6 |
714 | movaps %xmm6, -0x60(%rdi) |
715 | |
716 | movaps -0x76(%rsi), %xmm8 |
717 | palignr $6, %xmm8, %xmm7 |
718 | movaps %xmm7, -0x70(%rdi) |
719 | |
720 | movaps -0x86(%rsi), %xmm9 |
721 | palignr $6, %xmm9, %xmm8 |
722 | movaps %xmm8, -0x80(%rdi) |
723 | |
724 | sub $0x80, %rdx |
725 | lea -0x80(%rdi), %rdi |
726 | lea -0x80(%rsi), %rsi |
727 | jae L(shl_6_bwd) |
728 | movdqu %xmm0, (%r8) |
729 | add $0x80, %rdx |
730 | sub %rdx, %rdi |
731 | sub %rdx, %rsi |
732 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
733 | |
734 | .p2align 4 |
735 | L(shl_7): |
736 | sub $0x80, %rdx |
737 | movaps -0x07(%rsi), %xmm1 |
738 | movaps 0x09(%rsi), %xmm2 |
739 | movaps 0x19(%rsi), %xmm3 |
740 | movaps 0x29(%rsi), %xmm4 |
741 | movaps 0x39(%rsi), %xmm5 |
742 | movaps 0x49(%rsi), %xmm6 |
743 | movaps 0x59(%rsi), %xmm7 |
744 | movaps 0x69(%rsi), %xmm8 |
745 | movaps 0x79(%rsi), %xmm9 |
746 | lea 0x80(%rsi), %rsi |
747 | palignr $7, %xmm8, %xmm9 |
748 | movaps %xmm9, 0x70(%rdi) |
749 | palignr $7, %xmm7, %xmm8 |
750 | movaps %xmm8, 0x60(%rdi) |
751 | palignr $7, %xmm6, %xmm7 |
752 | movaps %xmm7, 0x50(%rdi) |
753 | palignr $7, %xmm5, %xmm6 |
754 | movaps %xmm6, 0x40(%rdi) |
755 | palignr $7, %xmm4, %xmm5 |
756 | movaps %xmm5, 0x30(%rdi) |
757 | palignr $7, %xmm3, %xmm4 |
758 | movaps %xmm4, 0x20(%rdi) |
759 | palignr $7, %xmm2, %xmm3 |
760 | movaps %xmm3, 0x10(%rdi) |
761 | palignr $7, %xmm1, %xmm2 |
762 | movaps %xmm2, (%rdi) |
763 | lea 0x80(%rdi), %rdi |
764 | jae L(shl_7) |
765 | movdqu %xmm0, (%r8) |
766 | add $0x80, %rdx |
767 | add %rdx, %rdi |
768 | add %rdx, %rsi |
769 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
770 | |
771 | .p2align 4 |
772 | L(shl_7_bwd): |
773 | movaps -0x07(%rsi), %xmm1 |
774 | |
775 | movaps -0x17(%rsi), %xmm2 |
776 | palignr $7, %xmm2, %xmm1 |
777 | movaps %xmm1, -0x10(%rdi) |
778 | |
779 | movaps -0x27(%rsi), %xmm3 |
780 | palignr $7, %xmm3, %xmm2 |
781 | movaps %xmm2, -0x20(%rdi) |
782 | |
783 | movaps -0x37(%rsi), %xmm4 |
784 | palignr $7, %xmm4, %xmm3 |
785 | movaps %xmm3, -0x30(%rdi) |
786 | |
787 | movaps -0x47(%rsi), %xmm5 |
788 | palignr $7, %xmm5, %xmm4 |
789 | movaps %xmm4, -0x40(%rdi) |
790 | |
791 | movaps -0x57(%rsi), %xmm6 |
792 | palignr $7, %xmm6, %xmm5 |
793 | movaps %xmm5, -0x50(%rdi) |
794 | |
795 | movaps -0x67(%rsi), %xmm7 |
796 | palignr $7, %xmm7, %xmm6 |
797 | movaps %xmm6, -0x60(%rdi) |
798 | |
799 | movaps -0x77(%rsi), %xmm8 |
800 | palignr $7, %xmm8, %xmm7 |
801 | movaps %xmm7, -0x70(%rdi) |
802 | |
803 | movaps -0x87(%rsi), %xmm9 |
804 | palignr $7, %xmm9, %xmm8 |
805 | movaps %xmm8, -0x80(%rdi) |
806 | |
807 | sub $0x80, %rdx |
808 | lea -0x80(%rdi), %rdi |
809 | lea -0x80(%rsi), %rsi |
810 | jae L(shl_7_bwd) |
811 | movdqu %xmm0, (%r8) |
812 | add $0x80, %rdx |
813 | sub %rdx, %rdi |
814 | sub %rdx, %rsi |
815 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
816 | |
817 | .p2align 4 |
818 | L(shl_8): |
819 | sub $0x80, %rdx |
820 | movaps -0x08(%rsi), %xmm1 |
821 | movaps 0x08(%rsi), %xmm2 |
822 | movaps 0x18(%rsi), %xmm3 |
823 | movaps 0x28(%rsi), %xmm4 |
824 | movaps 0x38(%rsi), %xmm5 |
825 | movaps 0x48(%rsi), %xmm6 |
826 | movaps 0x58(%rsi), %xmm7 |
827 | movaps 0x68(%rsi), %xmm8 |
828 | movaps 0x78(%rsi), %xmm9 |
829 | lea 0x80(%rsi), %rsi |
830 | palignr $8, %xmm8, %xmm9 |
831 | movaps %xmm9, 0x70(%rdi) |
832 | palignr $8, %xmm7, %xmm8 |
833 | movaps %xmm8, 0x60(%rdi) |
834 | palignr $8, %xmm6, %xmm7 |
835 | movaps %xmm7, 0x50(%rdi) |
836 | palignr $8, %xmm5, %xmm6 |
837 | movaps %xmm6, 0x40(%rdi) |
838 | palignr $8, %xmm4, %xmm5 |
839 | movaps %xmm5, 0x30(%rdi) |
840 | palignr $8, %xmm3, %xmm4 |
841 | movaps %xmm4, 0x20(%rdi) |
842 | palignr $8, %xmm2, %xmm3 |
843 | movaps %xmm3, 0x10(%rdi) |
844 | palignr $8, %xmm1, %xmm2 |
845 | movaps %xmm2, (%rdi) |
846 | lea 0x80(%rdi), %rdi |
847 | jae L(shl_8) |
848 | movdqu %xmm0, (%r8) |
849 | add $0x80, %rdx |
850 | add %rdx, %rdi |
851 | add %rdx, %rsi |
852 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
853 | |
854 | .p2align 4 |
855 | L(shl_8_bwd): |
856 | movaps -0x08(%rsi), %xmm1 |
857 | |
858 | movaps -0x18(%rsi), %xmm2 |
859 | palignr $8, %xmm2, %xmm1 |
860 | movaps %xmm1, -0x10(%rdi) |
861 | |
862 | movaps -0x28(%rsi), %xmm3 |
863 | palignr $8, %xmm3, %xmm2 |
864 | movaps %xmm2, -0x20(%rdi) |
865 | |
866 | movaps -0x38(%rsi), %xmm4 |
867 | palignr $8, %xmm4, %xmm3 |
868 | movaps %xmm3, -0x30(%rdi) |
869 | |
870 | movaps -0x48(%rsi), %xmm5 |
871 | palignr $8, %xmm5, %xmm4 |
872 | movaps %xmm4, -0x40(%rdi) |
873 | |
874 | movaps -0x58(%rsi), %xmm6 |
875 | palignr $8, %xmm6, %xmm5 |
876 | movaps %xmm5, -0x50(%rdi) |
877 | |
878 | movaps -0x68(%rsi), %xmm7 |
879 | palignr $8, %xmm7, %xmm6 |
880 | movaps %xmm6, -0x60(%rdi) |
881 | |
882 | movaps -0x78(%rsi), %xmm8 |
883 | palignr $8, %xmm8, %xmm7 |
884 | movaps %xmm7, -0x70(%rdi) |
885 | |
886 | movaps -0x88(%rsi), %xmm9 |
887 | palignr $8, %xmm9, %xmm8 |
888 | movaps %xmm8, -0x80(%rdi) |
889 | |
890 | sub $0x80, %rdx |
891 | lea -0x80(%rdi), %rdi |
892 | lea -0x80(%rsi), %rsi |
893 | jae L(shl_8_bwd) |
894 | L(shl_8_end_bwd): |
895 | movdqu %xmm0, (%r8) |
896 | add $0x80, %rdx |
897 | sub %rdx, %rdi |
898 | sub %rdx, %rsi |
899 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
900 | |
901 | .p2align 4 |
902 | L(shl_9): |
903 | sub $0x80, %rdx |
904 | movaps -0x09(%rsi), %xmm1 |
905 | movaps 0x07(%rsi), %xmm2 |
906 | movaps 0x17(%rsi), %xmm3 |
907 | movaps 0x27(%rsi), %xmm4 |
908 | movaps 0x37(%rsi), %xmm5 |
909 | movaps 0x47(%rsi), %xmm6 |
910 | movaps 0x57(%rsi), %xmm7 |
911 | movaps 0x67(%rsi), %xmm8 |
912 | movaps 0x77(%rsi), %xmm9 |
913 | lea 0x80(%rsi), %rsi |
914 | palignr $9, %xmm8, %xmm9 |
915 | movaps %xmm9, 0x70(%rdi) |
916 | palignr $9, %xmm7, %xmm8 |
917 | movaps %xmm8, 0x60(%rdi) |
918 | palignr $9, %xmm6, %xmm7 |
919 | movaps %xmm7, 0x50(%rdi) |
920 | palignr $9, %xmm5, %xmm6 |
921 | movaps %xmm6, 0x40(%rdi) |
922 | palignr $9, %xmm4, %xmm5 |
923 | movaps %xmm5, 0x30(%rdi) |
924 | palignr $9, %xmm3, %xmm4 |
925 | movaps %xmm4, 0x20(%rdi) |
926 | palignr $9, %xmm2, %xmm3 |
927 | movaps %xmm3, 0x10(%rdi) |
928 | palignr $9, %xmm1, %xmm2 |
929 | movaps %xmm2, (%rdi) |
930 | lea 0x80(%rdi), %rdi |
931 | jae L(shl_9) |
932 | movdqu %xmm0, (%r8) |
933 | add $0x80, %rdx |
934 | add %rdx, %rdi |
935 | add %rdx, %rsi |
936 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
937 | |
938 | .p2align 4 |
939 | L(shl_9_bwd): |
940 | movaps -0x09(%rsi), %xmm1 |
941 | |
942 | movaps -0x19(%rsi), %xmm2 |
943 | palignr $9, %xmm2, %xmm1 |
944 | movaps %xmm1, -0x10(%rdi) |
945 | |
946 | movaps -0x29(%rsi), %xmm3 |
947 | palignr $9, %xmm3, %xmm2 |
948 | movaps %xmm2, -0x20(%rdi) |
949 | |
950 | movaps -0x39(%rsi), %xmm4 |
951 | palignr $9, %xmm4, %xmm3 |
952 | movaps %xmm3, -0x30(%rdi) |
953 | |
954 | movaps -0x49(%rsi), %xmm5 |
955 | palignr $9, %xmm5, %xmm4 |
956 | movaps %xmm4, -0x40(%rdi) |
957 | |
958 | movaps -0x59(%rsi), %xmm6 |
959 | palignr $9, %xmm6, %xmm5 |
960 | movaps %xmm5, -0x50(%rdi) |
961 | |
962 | movaps -0x69(%rsi), %xmm7 |
963 | palignr $9, %xmm7, %xmm6 |
964 | movaps %xmm6, -0x60(%rdi) |
965 | |
966 | movaps -0x79(%rsi), %xmm8 |
967 | palignr $9, %xmm8, %xmm7 |
968 | movaps %xmm7, -0x70(%rdi) |
969 | |
970 | movaps -0x89(%rsi), %xmm9 |
971 | palignr $9, %xmm9, %xmm8 |
972 | movaps %xmm8, -0x80(%rdi) |
973 | |
974 | sub $0x80, %rdx |
975 | lea -0x80(%rdi), %rdi |
976 | lea -0x80(%rsi), %rsi |
977 | jae L(shl_9_bwd) |
978 | movdqu %xmm0, (%r8) |
979 | add $0x80, %rdx |
980 | sub %rdx, %rdi |
981 | sub %rdx, %rsi |
982 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
983 | |
984 | .p2align 4 |
985 | L(shl_10): |
986 | sub $0x80, %rdx |
987 | movaps -0x0a(%rsi), %xmm1 |
988 | movaps 0x06(%rsi), %xmm2 |
989 | movaps 0x16(%rsi), %xmm3 |
990 | movaps 0x26(%rsi), %xmm4 |
991 | movaps 0x36(%rsi), %xmm5 |
992 | movaps 0x46(%rsi), %xmm6 |
993 | movaps 0x56(%rsi), %xmm7 |
994 | movaps 0x66(%rsi), %xmm8 |
995 | movaps 0x76(%rsi), %xmm9 |
996 | lea 0x80(%rsi), %rsi |
997 | palignr $10, %xmm8, %xmm9 |
998 | movaps %xmm9, 0x70(%rdi) |
999 | palignr $10, %xmm7, %xmm8 |
1000 | movaps %xmm8, 0x60(%rdi) |
1001 | palignr $10, %xmm6, %xmm7 |
1002 | movaps %xmm7, 0x50(%rdi) |
1003 | palignr $10, %xmm5, %xmm6 |
1004 | movaps %xmm6, 0x40(%rdi) |
1005 | palignr $10, %xmm4, %xmm5 |
1006 | movaps %xmm5, 0x30(%rdi) |
1007 | palignr $10, %xmm3, %xmm4 |
1008 | movaps %xmm4, 0x20(%rdi) |
1009 | palignr $10, %xmm2, %xmm3 |
1010 | movaps %xmm3, 0x10(%rdi) |
1011 | palignr $10, %xmm1, %xmm2 |
1012 | movaps %xmm2, (%rdi) |
1013 | lea 0x80(%rdi), %rdi |
1014 | jae L(shl_10) |
1015 | movdqu %xmm0, (%r8) |
1016 | add $0x80, %rdx |
1017 | add %rdx, %rdi |
1018 | add %rdx, %rsi |
1019 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1020 | |
1021 | .p2align 4 |
1022 | L(shl_10_bwd): |
1023 | movaps -0x0a(%rsi), %xmm1 |
1024 | |
1025 | movaps -0x1a(%rsi), %xmm2 |
1026 | palignr $10, %xmm2, %xmm1 |
1027 | movaps %xmm1, -0x10(%rdi) |
1028 | |
1029 | movaps -0x2a(%rsi), %xmm3 |
1030 | palignr $10, %xmm3, %xmm2 |
1031 | movaps %xmm2, -0x20(%rdi) |
1032 | |
1033 | movaps -0x3a(%rsi), %xmm4 |
1034 | palignr $10, %xmm4, %xmm3 |
1035 | movaps %xmm3, -0x30(%rdi) |
1036 | |
1037 | movaps -0x4a(%rsi), %xmm5 |
1038 | palignr $10, %xmm5, %xmm4 |
1039 | movaps %xmm4, -0x40(%rdi) |
1040 | |
1041 | movaps -0x5a(%rsi), %xmm6 |
1042 | palignr $10, %xmm6, %xmm5 |
1043 | movaps %xmm5, -0x50(%rdi) |
1044 | |
1045 | movaps -0x6a(%rsi), %xmm7 |
1046 | palignr $10, %xmm7, %xmm6 |
1047 | movaps %xmm6, -0x60(%rdi) |
1048 | |
1049 | movaps -0x7a(%rsi), %xmm8 |
1050 | palignr $10, %xmm8, %xmm7 |
1051 | movaps %xmm7, -0x70(%rdi) |
1052 | |
1053 | movaps -0x8a(%rsi), %xmm9 |
1054 | palignr $10, %xmm9, %xmm8 |
1055 | movaps %xmm8, -0x80(%rdi) |
1056 | |
1057 | sub $0x80, %rdx |
1058 | lea -0x80(%rdi), %rdi |
1059 | lea -0x80(%rsi), %rsi |
1060 | jae L(shl_10_bwd) |
1061 | movdqu %xmm0, (%r8) |
1062 | add $0x80, %rdx |
1063 | sub %rdx, %rdi |
1064 | sub %rdx, %rsi |
1065 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1066 | |
1067 | .p2align 4 |
1068 | L(shl_11): |
1069 | sub $0x80, %rdx |
1070 | movaps -0x0b(%rsi), %xmm1 |
1071 | movaps 0x05(%rsi), %xmm2 |
1072 | movaps 0x15(%rsi), %xmm3 |
1073 | movaps 0x25(%rsi), %xmm4 |
1074 | movaps 0x35(%rsi), %xmm5 |
1075 | movaps 0x45(%rsi), %xmm6 |
1076 | movaps 0x55(%rsi), %xmm7 |
1077 | movaps 0x65(%rsi), %xmm8 |
1078 | movaps 0x75(%rsi), %xmm9 |
1079 | lea 0x80(%rsi), %rsi |
1080 | palignr $11, %xmm8, %xmm9 |
1081 | movaps %xmm9, 0x70(%rdi) |
1082 | palignr $11, %xmm7, %xmm8 |
1083 | movaps %xmm8, 0x60(%rdi) |
1084 | palignr $11, %xmm6, %xmm7 |
1085 | movaps %xmm7, 0x50(%rdi) |
1086 | palignr $11, %xmm5, %xmm6 |
1087 | movaps %xmm6, 0x40(%rdi) |
1088 | palignr $11, %xmm4, %xmm5 |
1089 | movaps %xmm5, 0x30(%rdi) |
1090 | palignr $11, %xmm3, %xmm4 |
1091 | movaps %xmm4, 0x20(%rdi) |
1092 | palignr $11, %xmm2, %xmm3 |
1093 | movaps %xmm3, 0x10(%rdi) |
1094 | palignr $11, %xmm1, %xmm2 |
1095 | movaps %xmm2, (%rdi) |
1096 | lea 0x80(%rdi), %rdi |
1097 | jae L(shl_11) |
1098 | movdqu %xmm0, (%r8) |
1099 | add $0x80, %rdx |
1100 | add %rdx, %rdi |
1101 | add %rdx, %rsi |
1102 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1103 | |
1104 | .p2align 4 |
1105 | L(shl_11_bwd): |
1106 | movaps -0x0b(%rsi), %xmm1 |
1107 | |
1108 | movaps -0x1b(%rsi), %xmm2 |
1109 | palignr $11, %xmm2, %xmm1 |
1110 | movaps %xmm1, -0x10(%rdi) |
1111 | |
1112 | movaps -0x2b(%rsi), %xmm3 |
1113 | palignr $11, %xmm3, %xmm2 |
1114 | movaps %xmm2, -0x20(%rdi) |
1115 | |
1116 | movaps -0x3b(%rsi), %xmm4 |
1117 | palignr $11, %xmm4, %xmm3 |
1118 | movaps %xmm3, -0x30(%rdi) |
1119 | |
1120 | movaps -0x4b(%rsi), %xmm5 |
1121 | palignr $11, %xmm5, %xmm4 |
1122 | movaps %xmm4, -0x40(%rdi) |
1123 | |
1124 | movaps -0x5b(%rsi), %xmm6 |
1125 | palignr $11, %xmm6, %xmm5 |
1126 | movaps %xmm5, -0x50(%rdi) |
1127 | |
1128 | movaps -0x6b(%rsi), %xmm7 |
1129 | palignr $11, %xmm7, %xmm6 |
1130 | movaps %xmm6, -0x60(%rdi) |
1131 | |
1132 | movaps -0x7b(%rsi), %xmm8 |
1133 | palignr $11, %xmm8, %xmm7 |
1134 | movaps %xmm7, -0x70(%rdi) |
1135 | |
1136 | movaps -0x8b(%rsi), %xmm9 |
1137 | palignr $11, %xmm9, %xmm8 |
1138 | movaps %xmm8, -0x80(%rdi) |
1139 | |
1140 | sub $0x80, %rdx |
1141 | lea -0x80(%rdi), %rdi |
1142 | lea -0x80(%rsi), %rsi |
1143 | jae L(shl_11_bwd) |
1144 | movdqu %xmm0, (%r8) |
1145 | add $0x80, %rdx |
1146 | sub %rdx, %rdi |
1147 | sub %rdx, %rsi |
1148 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1149 | |
1150 | .p2align 4 |
1151 | L(shl_12): |
1152 | sub $0x80, %rdx |
1153 | movdqa -0x0c(%rsi), %xmm1 |
1154 | movaps 0x04(%rsi), %xmm2 |
1155 | movaps 0x14(%rsi), %xmm3 |
1156 | movaps 0x24(%rsi), %xmm4 |
1157 | movaps 0x34(%rsi), %xmm5 |
1158 | movaps 0x44(%rsi), %xmm6 |
1159 | movaps 0x54(%rsi), %xmm7 |
1160 | movaps 0x64(%rsi), %xmm8 |
1161 | movaps 0x74(%rsi), %xmm9 |
1162 | lea 0x80(%rsi), %rsi |
1163 | palignr $12, %xmm8, %xmm9 |
1164 | movaps %xmm9, 0x70(%rdi) |
1165 | palignr $12, %xmm7, %xmm8 |
1166 | movaps %xmm8, 0x60(%rdi) |
1167 | palignr $12, %xmm6, %xmm7 |
1168 | movaps %xmm7, 0x50(%rdi) |
1169 | palignr $12, %xmm5, %xmm6 |
1170 | movaps %xmm6, 0x40(%rdi) |
1171 | palignr $12, %xmm4, %xmm5 |
1172 | movaps %xmm5, 0x30(%rdi) |
1173 | palignr $12, %xmm3, %xmm4 |
1174 | movaps %xmm4, 0x20(%rdi) |
1175 | palignr $12, %xmm2, %xmm3 |
1176 | movaps %xmm3, 0x10(%rdi) |
1177 | palignr $12, %xmm1, %xmm2 |
1178 | movaps %xmm2, (%rdi) |
1179 | |
1180 | lea 0x80(%rdi), %rdi |
1181 | jae L(shl_12) |
1182 | movdqu %xmm0, (%r8) |
1183 | add $0x80, %rdx |
1184 | add %rdx, %rdi |
1185 | add %rdx, %rsi |
1186 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1187 | |
1188 | .p2align 4 |
1189 | L(shl_12_bwd): |
1190 | movaps -0x0c(%rsi), %xmm1 |
1191 | |
1192 | movaps -0x1c(%rsi), %xmm2 |
1193 | palignr $12, %xmm2, %xmm1 |
1194 | movaps %xmm1, -0x10(%rdi) |
1195 | |
1196 | movaps -0x2c(%rsi), %xmm3 |
1197 | palignr $12, %xmm3, %xmm2 |
1198 | movaps %xmm2, -0x20(%rdi) |
1199 | |
1200 | movaps -0x3c(%rsi), %xmm4 |
1201 | palignr $12, %xmm4, %xmm3 |
1202 | movaps %xmm3, -0x30(%rdi) |
1203 | |
1204 | movaps -0x4c(%rsi), %xmm5 |
1205 | palignr $12, %xmm5, %xmm4 |
1206 | movaps %xmm4, -0x40(%rdi) |
1207 | |
1208 | movaps -0x5c(%rsi), %xmm6 |
1209 | palignr $12, %xmm6, %xmm5 |
1210 | movaps %xmm5, -0x50(%rdi) |
1211 | |
1212 | movaps -0x6c(%rsi), %xmm7 |
1213 | palignr $12, %xmm7, %xmm6 |
1214 | movaps %xmm6, -0x60(%rdi) |
1215 | |
1216 | movaps -0x7c(%rsi), %xmm8 |
1217 | palignr $12, %xmm8, %xmm7 |
1218 | movaps %xmm7, -0x70(%rdi) |
1219 | |
1220 | movaps -0x8c(%rsi), %xmm9 |
1221 | palignr $12, %xmm9, %xmm8 |
1222 | movaps %xmm8, -0x80(%rdi) |
1223 | |
1224 | sub $0x80, %rdx |
1225 | lea -0x80(%rdi), %rdi |
1226 | lea -0x80(%rsi), %rsi |
1227 | jae L(shl_12_bwd) |
1228 | movdqu %xmm0, (%r8) |
1229 | add $0x80, %rdx |
1230 | sub %rdx, %rdi |
1231 | sub %rdx, %rsi |
1232 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1233 | |
1234 | .p2align 4 |
1235 | L(shl_13): |
1236 | sub $0x80, %rdx |
1237 | movaps -0x0d(%rsi), %xmm1 |
1238 | movaps 0x03(%rsi), %xmm2 |
1239 | movaps 0x13(%rsi), %xmm3 |
1240 | movaps 0x23(%rsi), %xmm4 |
1241 | movaps 0x33(%rsi), %xmm5 |
1242 | movaps 0x43(%rsi), %xmm6 |
1243 | movaps 0x53(%rsi), %xmm7 |
1244 | movaps 0x63(%rsi), %xmm8 |
1245 | movaps 0x73(%rsi), %xmm9 |
1246 | lea 0x80(%rsi), %rsi |
1247 | palignr $13, %xmm8, %xmm9 |
1248 | movaps %xmm9, 0x70(%rdi) |
1249 | palignr $13, %xmm7, %xmm8 |
1250 | movaps %xmm8, 0x60(%rdi) |
1251 | palignr $13, %xmm6, %xmm7 |
1252 | movaps %xmm7, 0x50(%rdi) |
1253 | palignr $13, %xmm5, %xmm6 |
1254 | movaps %xmm6, 0x40(%rdi) |
1255 | palignr $13, %xmm4, %xmm5 |
1256 | movaps %xmm5, 0x30(%rdi) |
1257 | palignr $13, %xmm3, %xmm4 |
1258 | movaps %xmm4, 0x20(%rdi) |
1259 | palignr $13, %xmm2, %xmm3 |
1260 | movaps %xmm3, 0x10(%rdi) |
1261 | palignr $13, %xmm1, %xmm2 |
1262 | movaps %xmm2, (%rdi) |
1263 | lea 0x80(%rdi), %rdi |
1264 | jae L(shl_13) |
1265 | movdqu %xmm0, (%r8) |
1266 | add $0x80, %rdx |
1267 | add %rdx, %rdi |
1268 | add %rdx, %rsi |
1269 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1270 | |
1271 | .p2align 4 |
1272 | L(shl_13_bwd): |
1273 | movaps -0x0d(%rsi), %xmm1 |
1274 | |
1275 | movaps -0x1d(%rsi), %xmm2 |
1276 | palignr $13, %xmm2, %xmm1 |
1277 | movaps %xmm1, -0x10(%rdi) |
1278 | |
1279 | movaps -0x2d(%rsi), %xmm3 |
1280 | palignr $13, %xmm3, %xmm2 |
1281 | movaps %xmm2, -0x20(%rdi) |
1282 | |
1283 | movaps -0x3d(%rsi), %xmm4 |
1284 | palignr $13, %xmm4, %xmm3 |
1285 | movaps %xmm3, -0x30(%rdi) |
1286 | |
1287 | movaps -0x4d(%rsi), %xmm5 |
1288 | palignr $13, %xmm5, %xmm4 |
1289 | movaps %xmm4, -0x40(%rdi) |
1290 | |
1291 | movaps -0x5d(%rsi), %xmm6 |
1292 | palignr $13, %xmm6, %xmm5 |
1293 | movaps %xmm5, -0x50(%rdi) |
1294 | |
1295 | movaps -0x6d(%rsi), %xmm7 |
1296 | palignr $13, %xmm7, %xmm6 |
1297 | movaps %xmm6, -0x60(%rdi) |
1298 | |
1299 | movaps -0x7d(%rsi), %xmm8 |
1300 | palignr $13, %xmm8, %xmm7 |
1301 | movaps %xmm7, -0x70(%rdi) |
1302 | |
1303 | movaps -0x8d(%rsi), %xmm9 |
1304 | palignr $13, %xmm9, %xmm8 |
1305 | movaps %xmm8, -0x80(%rdi) |
1306 | |
1307 | sub $0x80, %rdx |
1308 | lea -0x80(%rdi), %rdi |
1309 | lea -0x80(%rsi), %rsi |
1310 | jae L(shl_13_bwd) |
1311 | movdqu %xmm0, (%r8) |
1312 | add $0x80, %rdx |
1313 | sub %rdx, %rdi |
1314 | sub %rdx, %rsi |
1315 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1316 | |
1317 | .p2align 4 |
1318 | L(shl_14): |
1319 | sub $0x80, %rdx |
1320 | movaps -0x0e(%rsi), %xmm1 |
1321 | movaps 0x02(%rsi), %xmm2 |
1322 | movaps 0x12(%rsi), %xmm3 |
1323 | movaps 0x22(%rsi), %xmm4 |
1324 | movaps 0x32(%rsi), %xmm5 |
1325 | movaps 0x42(%rsi), %xmm6 |
1326 | movaps 0x52(%rsi), %xmm7 |
1327 | movaps 0x62(%rsi), %xmm8 |
1328 | movaps 0x72(%rsi), %xmm9 |
1329 | lea 0x80(%rsi), %rsi |
1330 | palignr $14, %xmm8, %xmm9 |
1331 | movaps %xmm9, 0x70(%rdi) |
1332 | palignr $14, %xmm7, %xmm8 |
1333 | movaps %xmm8, 0x60(%rdi) |
1334 | palignr $14, %xmm6, %xmm7 |
1335 | movaps %xmm7, 0x50(%rdi) |
1336 | palignr $14, %xmm5, %xmm6 |
1337 | movaps %xmm6, 0x40(%rdi) |
1338 | palignr $14, %xmm4, %xmm5 |
1339 | movaps %xmm5, 0x30(%rdi) |
1340 | palignr $14, %xmm3, %xmm4 |
1341 | movaps %xmm4, 0x20(%rdi) |
1342 | palignr $14, %xmm2, %xmm3 |
1343 | movaps %xmm3, 0x10(%rdi) |
1344 | palignr $14, %xmm1, %xmm2 |
1345 | movaps %xmm2, (%rdi) |
1346 | lea 0x80(%rdi), %rdi |
1347 | jae L(shl_14) |
1348 | movdqu %xmm0, (%r8) |
1349 | add $0x80, %rdx |
1350 | add %rdx, %rdi |
1351 | add %rdx, %rsi |
1352 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1353 | |
1354 | .p2align 4 |
1355 | L(shl_14_bwd): |
1356 | movaps -0x0e(%rsi), %xmm1 |
1357 | |
1358 | movaps -0x1e(%rsi), %xmm2 |
1359 | palignr $14, %xmm2, %xmm1 |
1360 | movaps %xmm1, -0x10(%rdi) |
1361 | |
1362 | movaps -0x2e(%rsi), %xmm3 |
1363 | palignr $14, %xmm3, %xmm2 |
1364 | movaps %xmm2, -0x20(%rdi) |
1365 | |
1366 | movaps -0x3e(%rsi), %xmm4 |
1367 | palignr $14, %xmm4, %xmm3 |
1368 | movaps %xmm3, -0x30(%rdi) |
1369 | |
1370 | movaps -0x4e(%rsi), %xmm5 |
1371 | palignr $14, %xmm5, %xmm4 |
1372 | movaps %xmm4, -0x40(%rdi) |
1373 | |
1374 | movaps -0x5e(%rsi), %xmm6 |
1375 | palignr $14, %xmm6, %xmm5 |
1376 | movaps %xmm5, -0x50(%rdi) |
1377 | |
1378 | movaps -0x6e(%rsi), %xmm7 |
1379 | palignr $14, %xmm7, %xmm6 |
1380 | movaps %xmm6, -0x60(%rdi) |
1381 | |
1382 | movaps -0x7e(%rsi), %xmm8 |
1383 | palignr $14, %xmm8, %xmm7 |
1384 | movaps %xmm7, -0x70(%rdi) |
1385 | |
1386 | movaps -0x8e(%rsi), %xmm9 |
1387 | palignr $14, %xmm9, %xmm8 |
1388 | movaps %xmm8, -0x80(%rdi) |
1389 | |
1390 | sub $0x80, %rdx |
1391 | lea -0x80(%rdi), %rdi |
1392 | lea -0x80(%rsi), %rsi |
1393 | jae L(shl_14_bwd) |
1394 | movdqu %xmm0, (%r8) |
1395 | add $0x80, %rdx |
1396 | sub %rdx, %rdi |
1397 | sub %rdx, %rsi |
1398 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1399 | |
1400 | .p2align 4 |
1401 | L(shl_15): |
1402 | sub $0x80, %rdx |
1403 | movaps -0x0f(%rsi), %xmm1 |
1404 | movaps 0x01(%rsi), %xmm2 |
1405 | movaps 0x11(%rsi), %xmm3 |
1406 | movaps 0x21(%rsi), %xmm4 |
1407 | movaps 0x31(%rsi), %xmm5 |
1408 | movaps 0x41(%rsi), %xmm6 |
1409 | movaps 0x51(%rsi), %xmm7 |
1410 | movaps 0x61(%rsi), %xmm8 |
1411 | movaps 0x71(%rsi), %xmm9 |
1412 | lea 0x80(%rsi), %rsi |
1413 | palignr $15, %xmm8, %xmm9 |
1414 | movaps %xmm9, 0x70(%rdi) |
1415 | palignr $15, %xmm7, %xmm8 |
1416 | movaps %xmm8, 0x60(%rdi) |
1417 | palignr $15, %xmm6, %xmm7 |
1418 | movaps %xmm7, 0x50(%rdi) |
1419 | palignr $15, %xmm5, %xmm6 |
1420 | movaps %xmm6, 0x40(%rdi) |
1421 | palignr $15, %xmm4, %xmm5 |
1422 | movaps %xmm5, 0x30(%rdi) |
1423 | palignr $15, %xmm3, %xmm4 |
1424 | movaps %xmm4, 0x20(%rdi) |
1425 | palignr $15, %xmm2, %xmm3 |
1426 | movaps %xmm3, 0x10(%rdi) |
1427 | palignr $15, %xmm1, %xmm2 |
1428 | movaps %xmm2, (%rdi) |
1429 | lea 0x80(%rdi), %rdi |
1430 | jae L(shl_15) |
1431 | movdqu %xmm0, (%r8) |
1432 | add $0x80, %rdx |
1433 | add %rdx, %rdi |
1434 | add %rdx, %rsi |
1435 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1436 | |
1437 | .p2align 4 |
1438 | L(shl_15_bwd): |
1439 | movaps -0x0f(%rsi), %xmm1 |
1440 | |
1441 | movaps -0x1f(%rsi), %xmm2 |
1442 | palignr $15, %xmm2, %xmm1 |
1443 | movaps %xmm1, -0x10(%rdi) |
1444 | |
1445 | movaps -0x2f(%rsi), %xmm3 |
1446 | palignr $15, %xmm3, %xmm2 |
1447 | movaps %xmm2, -0x20(%rdi) |
1448 | |
1449 | movaps -0x3f(%rsi), %xmm4 |
1450 | palignr $15, %xmm4, %xmm3 |
1451 | movaps %xmm3, -0x30(%rdi) |
1452 | |
1453 | movaps -0x4f(%rsi), %xmm5 |
1454 | palignr $15, %xmm5, %xmm4 |
1455 | movaps %xmm4, -0x40(%rdi) |
1456 | |
1457 | movaps -0x5f(%rsi), %xmm6 |
1458 | palignr $15, %xmm6, %xmm5 |
1459 | movaps %xmm5, -0x50(%rdi) |
1460 | |
1461 | movaps -0x6f(%rsi), %xmm7 |
1462 | palignr $15, %xmm7, %xmm6 |
1463 | movaps %xmm6, -0x60(%rdi) |
1464 | |
1465 | movaps -0x7f(%rsi), %xmm8 |
1466 | palignr $15, %xmm8, %xmm7 |
1467 | movaps %xmm7, -0x70(%rdi) |
1468 | |
1469 | movaps -0x8f(%rsi), %xmm9 |
1470 | palignr $15, %xmm9, %xmm8 |
1471 | movaps %xmm8, -0x80(%rdi) |
1472 | |
1473 | sub $0x80, %rdx |
1474 | lea -0x80(%rdi), %rdi |
1475 | lea -0x80(%rsi), %rsi |
1476 | jae L(shl_15_bwd) |
1477 | movdqu %xmm0, (%r8) |
1478 | add $0x80, %rdx |
1479 | sub %rdx, %rdi |
1480 | sub %rdx, %rsi |
1481 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1482 | |
1483 | .p2align 4 |
1484 | L(gobble_mem_fwd): |
1485 | movdqu (%rsi), %xmm1 |
1486 | movdqu %xmm0, (%r8) |
1487 | movdqa %xmm1, (%rdi) |
1488 | sub $16, %rdx |
1489 | add $16, %rsi |
1490 | add $16, %rdi |
1491 | |
1492 | #ifdef SHARED_CACHE_SIZE_HALF |
1493 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1494 | #else |
1495 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1496 | #endif |
1497 | #ifdef USE_AS_MEMMOVE |
1498 | mov %rsi, %r9 |
1499 | sub %rdi, %r9 |
1500 | cmp %rdx, %r9 |
1501 | jae L(memmove_is_memcpy_fwd) |
1502 | cmp %rcx, %r9 |
1503 | jbe L(ll_cache_copy_fwd_start) |
1504 | L(memmove_is_memcpy_fwd): |
1505 | #endif |
1506 | cmp %rcx, %rdx |
1507 | ja L(bigger_in_fwd) |
1508 | mov %rdx, %rcx |
1509 | L(bigger_in_fwd): |
1510 | sub %rcx, %rdx |
1511 | cmp $0x1000, %rdx |
1512 | jbe L(ll_cache_copy_fwd) |
1513 | |
1514 | mov %rcx, %r9 |
1515 | shl $3, %r9 |
1516 | cmp %r9, %rdx |
1517 | jbe L(2steps_copy_fwd) |
1518 | add %rcx, %rdx |
1519 | xor %rcx, %rcx |
1520 | L(2steps_copy_fwd): |
1521 | sub $0x80, %rdx |
1522 | L(gobble_mem_fwd_loop): |
1523 | sub $0x80, %rdx |
1524 | prefetcht0 0x200(%rsi) |
1525 | prefetcht0 0x300(%rsi) |
1526 | movdqu (%rsi), %xmm0 |
1527 | movdqu 0x10(%rsi), %xmm1 |
1528 | movdqu 0x20(%rsi), %xmm2 |
1529 | movdqu 0x30(%rsi), %xmm3 |
1530 | movdqu 0x40(%rsi), %xmm4 |
1531 | movdqu 0x50(%rsi), %xmm5 |
1532 | movdqu 0x60(%rsi), %xmm6 |
1533 | movdqu 0x70(%rsi), %xmm7 |
1534 | lfence |
1535 | movntdq %xmm0, (%rdi) |
1536 | movntdq %xmm1, 0x10(%rdi) |
1537 | movntdq %xmm2, 0x20(%rdi) |
1538 | movntdq %xmm3, 0x30(%rdi) |
1539 | movntdq %xmm4, 0x40(%rdi) |
1540 | movntdq %xmm5, 0x50(%rdi) |
1541 | movntdq %xmm6, 0x60(%rdi) |
1542 | movntdq %xmm7, 0x70(%rdi) |
1543 | lea 0x80(%rsi), %rsi |
1544 | lea 0x80(%rdi), %rdi |
1545 | jae L(gobble_mem_fwd_loop) |
1546 | sfence |
1547 | cmp $0x80, %rcx |
1548 | jb L(gobble_mem_fwd_end) |
1549 | add $0x80, %rdx |
1550 | L(ll_cache_copy_fwd): |
1551 | add %rcx, %rdx |
1552 | L(ll_cache_copy_fwd_start): |
1553 | sub $0x80, %rdx |
1554 | L(gobble_ll_loop_fwd): |
1555 | prefetchnta 0x1c0(%rsi) |
1556 | prefetchnta 0x280(%rsi) |
1557 | prefetchnta 0x1c0(%rdi) |
1558 | prefetchnta 0x280(%rdi) |
1559 | sub $0x80, %rdx |
1560 | movdqu (%rsi), %xmm0 |
1561 | movdqu 0x10(%rsi), %xmm1 |
1562 | movdqu 0x20(%rsi), %xmm2 |
1563 | movdqu 0x30(%rsi), %xmm3 |
1564 | movdqu 0x40(%rsi), %xmm4 |
1565 | movdqu 0x50(%rsi), %xmm5 |
1566 | movdqu 0x60(%rsi), %xmm6 |
1567 | movdqu 0x70(%rsi), %xmm7 |
1568 | movdqa %xmm0, (%rdi) |
1569 | movdqa %xmm1, 0x10(%rdi) |
1570 | movdqa %xmm2, 0x20(%rdi) |
1571 | movdqa %xmm3, 0x30(%rdi) |
1572 | movdqa %xmm4, 0x40(%rdi) |
1573 | movdqa %xmm5, 0x50(%rdi) |
1574 | movdqa %xmm6, 0x60(%rdi) |
1575 | movdqa %xmm7, 0x70(%rdi) |
1576 | lea 0x80(%rsi), %rsi |
1577 | lea 0x80(%rdi), %rdi |
1578 | jae L(gobble_ll_loop_fwd) |
1579 | L(gobble_mem_fwd_end): |
1580 | add $0x80, %rdx |
1581 | add %rdx, %rsi |
1582 | add %rdx, %rdi |
1583 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1584 | |
1585 | .p2align 4 |
1586 | L(gobble_mem_bwd): |
1587 | add %rdx, %rsi |
1588 | add %rdx, %rdi |
1589 | |
1590 | movdqu -16(%rsi), %xmm0 |
1591 | lea -16(%rdi), %r8 |
1592 | mov %rdi, %r9 |
1593 | and $-16, %rdi |
1594 | sub %rdi, %r9 |
1595 | sub %r9, %rsi |
1596 | sub %r9, %rdx |
1597 | |
1598 | |
1599 | #ifdef SHARED_CACHE_SIZE_HALF |
1600 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1601 | #else |
1602 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1603 | #endif |
1604 | #ifdef USE_AS_MEMMOVE |
1605 | mov %rdi, %r9 |
1606 | sub %rsi, %r9 |
1607 | cmp %rdx, %r9 |
1608 | jae L(memmove_is_memcpy_bwd) |
1609 | cmp %rcx, %r9 |
1610 | jbe L(ll_cache_copy_bwd_start) |
1611 | L(memmove_is_memcpy_bwd): |
1612 | #endif |
1613 | cmp %rcx, %rdx |
1614 | ja L(bigger) |
1615 | mov %rdx, %rcx |
1616 | L(bigger): |
1617 | sub %rcx, %rdx |
1618 | cmp $0x1000, %rdx |
1619 | jbe L(ll_cache_copy) |
1620 | |
1621 | mov %rcx, %r9 |
1622 | shl $3, %r9 |
1623 | cmp %r9, %rdx |
1624 | jbe L(2steps_copy) |
1625 | add %rcx, %rdx |
1626 | xor %rcx, %rcx |
1627 | L(2steps_copy): |
1628 | sub $0x80, %rdx |
1629 | L(gobble_mem_bwd_loop): |
1630 | sub $0x80, %rdx |
1631 | prefetcht0 -0x200(%rsi) |
1632 | prefetcht0 -0x300(%rsi) |
1633 | movdqu -0x10(%rsi), %xmm1 |
1634 | movdqu -0x20(%rsi), %xmm2 |
1635 | movdqu -0x30(%rsi), %xmm3 |
1636 | movdqu -0x40(%rsi), %xmm4 |
1637 | movdqu -0x50(%rsi), %xmm5 |
1638 | movdqu -0x60(%rsi), %xmm6 |
1639 | movdqu -0x70(%rsi), %xmm7 |
1640 | movdqu -0x80(%rsi), %xmm8 |
1641 | lfence |
1642 | movntdq %xmm1, -0x10(%rdi) |
1643 | movntdq %xmm2, -0x20(%rdi) |
1644 | movntdq %xmm3, -0x30(%rdi) |
1645 | movntdq %xmm4, -0x40(%rdi) |
1646 | movntdq %xmm5, -0x50(%rdi) |
1647 | movntdq %xmm6, -0x60(%rdi) |
1648 | movntdq %xmm7, -0x70(%rdi) |
1649 | movntdq %xmm8, -0x80(%rdi) |
1650 | lea -0x80(%rsi), %rsi |
1651 | lea -0x80(%rdi), %rdi |
1652 | jae L(gobble_mem_bwd_loop) |
1653 | sfence |
1654 | cmp $0x80, %rcx |
1655 | jb L(gobble_mem_bwd_end) |
1656 | add $0x80, %rdx |
1657 | L(ll_cache_copy): |
1658 | add %rcx, %rdx |
1659 | L(ll_cache_copy_bwd_start): |
1660 | sub $0x80, %rdx |
1661 | L(gobble_ll_loop): |
1662 | prefetchnta -0x1c0(%rsi) |
1663 | prefetchnta -0x280(%rsi) |
1664 | prefetchnta -0x1c0(%rdi) |
1665 | prefetchnta -0x280(%rdi) |
1666 | sub $0x80, %rdx |
1667 | movdqu -0x10(%rsi), %xmm1 |
1668 | movdqu -0x20(%rsi), %xmm2 |
1669 | movdqu -0x30(%rsi), %xmm3 |
1670 | movdqu -0x40(%rsi), %xmm4 |
1671 | movdqu -0x50(%rsi), %xmm5 |
1672 | movdqu -0x60(%rsi), %xmm6 |
1673 | movdqu -0x70(%rsi), %xmm7 |
1674 | movdqu -0x80(%rsi), %xmm8 |
1675 | movdqa %xmm1, -0x10(%rdi) |
1676 | movdqa %xmm2, -0x20(%rdi) |
1677 | movdqa %xmm3, -0x30(%rdi) |
1678 | movdqa %xmm4, -0x40(%rdi) |
1679 | movdqa %xmm5, -0x50(%rdi) |
1680 | movdqa %xmm6, -0x60(%rdi) |
1681 | movdqa %xmm7, -0x70(%rdi) |
1682 | movdqa %xmm8, -0x80(%rdi) |
1683 | lea -0x80(%rsi), %rsi |
1684 | lea -0x80(%rdi), %rdi |
1685 | jae L(gobble_ll_loop) |
1686 | L(gobble_mem_bwd_end): |
1687 | movdqu %xmm0, (%r8) |
1688 | add $0x80, %rdx |
1689 | sub %rdx, %rsi |
1690 | sub %rdx, %rdi |
1691 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1692 | |
1693 | .p2align 4 |
1694 | L(fwd_write_128bytes): |
1695 | lddqu -128(%rsi), %xmm0 |
1696 | movdqu %xmm0, -128(%rdi) |
1697 | L(fwd_write_112bytes): |
1698 | lddqu -112(%rsi), %xmm0 |
1699 | movdqu %xmm0, -112(%rdi) |
1700 | L(fwd_write_96bytes): |
1701 | lddqu -96(%rsi), %xmm0 |
1702 | movdqu %xmm0, -96(%rdi) |
1703 | L(fwd_write_80bytes): |
1704 | lddqu -80(%rsi), %xmm0 |
1705 | movdqu %xmm0, -80(%rdi) |
1706 | L(fwd_write_64bytes): |
1707 | lddqu -64(%rsi), %xmm0 |
1708 | movdqu %xmm0, -64(%rdi) |
1709 | L(fwd_write_48bytes): |
1710 | lddqu -48(%rsi), %xmm0 |
1711 | movdqu %xmm0, -48(%rdi) |
1712 | L(fwd_write_32bytes): |
1713 | lddqu -32(%rsi), %xmm0 |
1714 | movdqu %xmm0, -32(%rdi) |
1715 | L(fwd_write_16bytes): |
1716 | lddqu -16(%rsi), %xmm0 |
1717 | movdqu %xmm0, -16(%rdi) |
1718 | L(fwd_write_0bytes): |
1719 | ret |
1720 | |
1721 | |
1722 | .p2align 4 |
1723 | L(fwd_write_143bytes): |
1724 | lddqu -143(%rsi), %xmm0 |
1725 | movdqu %xmm0, -143(%rdi) |
1726 | L(fwd_write_127bytes): |
1727 | lddqu -127(%rsi), %xmm0 |
1728 | movdqu %xmm0, -127(%rdi) |
1729 | L(fwd_write_111bytes): |
1730 | lddqu -111(%rsi), %xmm0 |
1731 | movdqu %xmm0, -111(%rdi) |
1732 | L(fwd_write_95bytes): |
1733 | lddqu -95(%rsi), %xmm0 |
1734 | movdqu %xmm0, -95(%rdi) |
1735 | L(fwd_write_79bytes): |
1736 | lddqu -79(%rsi), %xmm0 |
1737 | movdqu %xmm0, -79(%rdi) |
1738 | L(fwd_write_63bytes): |
1739 | lddqu -63(%rsi), %xmm0 |
1740 | movdqu %xmm0, -63(%rdi) |
1741 | L(fwd_write_47bytes): |
1742 | lddqu -47(%rsi), %xmm0 |
1743 | movdqu %xmm0, -47(%rdi) |
1744 | L(fwd_write_31bytes): |
1745 | lddqu -31(%rsi), %xmm0 |
1746 | lddqu -16(%rsi), %xmm1 |
1747 | movdqu %xmm0, -31(%rdi) |
1748 | movdqu %xmm1, -16(%rdi) |
1749 | ret |
1750 | |
1751 | .p2align 4 |
1752 | L(fwd_write_15bytes): |
1753 | mov -15(%rsi), %rdx |
1754 | mov -8(%rsi), %rcx |
1755 | mov %rdx, -15(%rdi) |
1756 | mov %rcx, -8(%rdi) |
1757 | ret |
1758 | |
1759 | .p2align 4 |
1760 | L(fwd_write_142bytes): |
1761 | lddqu -142(%rsi), %xmm0 |
1762 | movdqu %xmm0, -142(%rdi) |
1763 | L(fwd_write_126bytes): |
1764 | lddqu -126(%rsi), %xmm0 |
1765 | movdqu %xmm0, -126(%rdi) |
1766 | L(fwd_write_110bytes): |
1767 | lddqu -110(%rsi), %xmm0 |
1768 | movdqu %xmm0, -110(%rdi) |
1769 | L(fwd_write_94bytes): |
1770 | lddqu -94(%rsi), %xmm0 |
1771 | movdqu %xmm0, -94(%rdi) |
1772 | L(fwd_write_78bytes): |
1773 | lddqu -78(%rsi), %xmm0 |
1774 | movdqu %xmm0, -78(%rdi) |
1775 | L(fwd_write_62bytes): |
1776 | lddqu -62(%rsi), %xmm0 |
1777 | movdqu %xmm0, -62(%rdi) |
1778 | L(fwd_write_46bytes): |
1779 | lddqu -46(%rsi), %xmm0 |
1780 | movdqu %xmm0, -46(%rdi) |
1781 | L(fwd_write_30bytes): |
1782 | lddqu -30(%rsi), %xmm0 |
1783 | lddqu -16(%rsi), %xmm1 |
1784 | movdqu %xmm0, -30(%rdi) |
1785 | movdqu %xmm1, -16(%rdi) |
1786 | ret |
1787 | |
1788 | .p2align 4 |
1789 | L(fwd_write_14bytes): |
1790 | mov -14(%rsi), %rdx |
1791 | mov -8(%rsi), %rcx |
1792 | mov %rdx, -14(%rdi) |
1793 | mov %rcx, -8(%rdi) |
1794 | ret |
1795 | |
1796 | .p2align 4 |
1797 | L(fwd_write_141bytes): |
1798 | lddqu -141(%rsi), %xmm0 |
1799 | movdqu %xmm0, -141(%rdi) |
1800 | L(fwd_write_125bytes): |
1801 | lddqu -125(%rsi), %xmm0 |
1802 | movdqu %xmm0, -125(%rdi) |
1803 | L(fwd_write_109bytes): |
1804 | lddqu -109(%rsi), %xmm0 |
1805 | movdqu %xmm0, -109(%rdi) |
1806 | L(fwd_write_93bytes): |
1807 | lddqu -93(%rsi), %xmm0 |
1808 | movdqu %xmm0, -93(%rdi) |
1809 | L(fwd_write_77bytes): |
1810 | lddqu -77(%rsi), %xmm0 |
1811 | movdqu %xmm0, -77(%rdi) |
1812 | L(fwd_write_61bytes): |
1813 | lddqu -61(%rsi), %xmm0 |
1814 | movdqu %xmm0, -61(%rdi) |
1815 | L(fwd_write_45bytes): |
1816 | lddqu -45(%rsi), %xmm0 |
1817 | movdqu %xmm0, -45(%rdi) |
1818 | L(fwd_write_29bytes): |
1819 | lddqu -29(%rsi), %xmm0 |
1820 | lddqu -16(%rsi), %xmm1 |
1821 | movdqu %xmm0, -29(%rdi) |
1822 | movdqu %xmm1, -16(%rdi) |
1823 | ret |
1824 | |
1825 | .p2align 4 |
1826 | L(fwd_write_13bytes): |
1827 | mov -13(%rsi), %rdx |
1828 | mov -8(%rsi), %rcx |
1829 | mov %rdx, -13(%rdi) |
1830 | mov %rcx, -8(%rdi) |
1831 | ret |
1832 | |
1833 | .p2align 4 |
1834 | L(fwd_write_140bytes): |
1835 | lddqu -140(%rsi), %xmm0 |
1836 | movdqu %xmm0, -140(%rdi) |
1837 | L(fwd_write_124bytes): |
1838 | lddqu -124(%rsi), %xmm0 |
1839 | movdqu %xmm0, -124(%rdi) |
1840 | L(fwd_write_108bytes): |
1841 | lddqu -108(%rsi), %xmm0 |
1842 | movdqu %xmm0, -108(%rdi) |
1843 | L(fwd_write_92bytes): |
1844 | lddqu -92(%rsi), %xmm0 |
1845 | movdqu %xmm0, -92(%rdi) |
1846 | L(fwd_write_76bytes): |
1847 | lddqu -76(%rsi), %xmm0 |
1848 | movdqu %xmm0, -76(%rdi) |
1849 | L(fwd_write_60bytes): |
1850 | lddqu -60(%rsi), %xmm0 |
1851 | movdqu %xmm0, -60(%rdi) |
1852 | L(fwd_write_44bytes): |
1853 | lddqu -44(%rsi), %xmm0 |
1854 | movdqu %xmm0, -44(%rdi) |
1855 | L(fwd_write_28bytes): |
1856 | lddqu -28(%rsi), %xmm0 |
1857 | lddqu -16(%rsi), %xmm1 |
1858 | movdqu %xmm0, -28(%rdi) |
1859 | movdqu %xmm1, -16(%rdi) |
1860 | ret |
1861 | |
1862 | .p2align 4 |
1863 | L(fwd_write_12bytes): |
1864 | mov -12(%rsi), %rdx |
1865 | mov -4(%rsi), %ecx |
1866 | mov %rdx, -12(%rdi) |
1867 | mov %ecx, -4(%rdi) |
1868 | ret |
1869 | |
1870 | .p2align 4 |
1871 | L(fwd_write_139bytes): |
1872 | lddqu -139(%rsi), %xmm0 |
1873 | movdqu %xmm0, -139(%rdi) |
1874 | L(fwd_write_123bytes): |
1875 | lddqu -123(%rsi), %xmm0 |
1876 | movdqu %xmm0, -123(%rdi) |
1877 | L(fwd_write_107bytes): |
1878 | lddqu -107(%rsi), %xmm0 |
1879 | movdqu %xmm0, -107(%rdi) |
1880 | L(fwd_write_91bytes): |
1881 | lddqu -91(%rsi), %xmm0 |
1882 | movdqu %xmm0, -91(%rdi) |
1883 | L(fwd_write_75bytes): |
1884 | lddqu -75(%rsi), %xmm0 |
1885 | movdqu %xmm0, -75(%rdi) |
1886 | L(fwd_write_59bytes): |
1887 | lddqu -59(%rsi), %xmm0 |
1888 | movdqu %xmm0, -59(%rdi) |
1889 | L(fwd_write_43bytes): |
1890 | lddqu -43(%rsi), %xmm0 |
1891 | movdqu %xmm0, -43(%rdi) |
1892 | L(fwd_write_27bytes): |
1893 | lddqu -27(%rsi), %xmm0 |
1894 | lddqu -16(%rsi), %xmm1 |
1895 | movdqu %xmm0, -27(%rdi) |
1896 | movdqu %xmm1, -16(%rdi) |
1897 | ret |
1898 | |
1899 | .p2align 4 |
1900 | L(fwd_write_11bytes): |
1901 | mov -11(%rsi), %rdx |
1902 | mov -4(%rsi), %ecx |
1903 | mov %rdx, -11(%rdi) |
1904 | mov %ecx, -4(%rdi) |
1905 | ret |
1906 | |
1907 | .p2align 4 |
1908 | L(fwd_write_138bytes): |
1909 | lddqu -138(%rsi), %xmm0 |
1910 | movdqu %xmm0, -138(%rdi) |
1911 | L(fwd_write_122bytes): |
1912 | lddqu -122(%rsi), %xmm0 |
1913 | movdqu %xmm0, -122(%rdi) |
1914 | L(fwd_write_106bytes): |
1915 | lddqu -106(%rsi), %xmm0 |
1916 | movdqu %xmm0, -106(%rdi) |
1917 | L(fwd_write_90bytes): |
1918 | lddqu -90(%rsi), %xmm0 |
1919 | movdqu %xmm0, -90(%rdi) |
1920 | L(fwd_write_74bytes): |
1921 | lddqu -74(%rsi), %xmm0 |
1922 | movdqu %xmm0, -74(%rdi) |
1923 | L(fwd_write_58bytes): |
1924 | lddqu -58(%rsi), %xmm0 |
1925 | movdqu %xmm0, -58(%rdi) |
1926 | L(fwd_write_42bytes): |
1927 | lddqu -42(%rsi), %xmm0 |
1928 | movdqu %xmm0, -42(%rdi) |
1929 | L(fwd_write_26bytes): |
1930 | lddqu -26(%rsi), %xmm0 |
1931 | lddqu -16(%rsi), %xmm1 |
1932 | movdqu %xmm0, -26(%rdi) |
1933 | movdqu %xmm1, -16(%rdi) |
1934 | ret |
1935 | |
1936 | .p2align 4 |
1937 | L(fwd_write_10bytes): |
1938 | mov -10(%rsi), %rdx |
1939 | mov -4(%rsi), %ecx |
1940 | mov %rdx, -10(%rdi) |
1941 | mov %ecx, -4(%rdi) |
1942 | ret |
1943 | |
1944 | .p2align 4 |
1945 | L(fwd_write_137bytes): |
1946 | lddqu -137(%rsi), %xmm0 |
1947 | movdqu %xmm0, -137(%rdi) |
1948 | L(fwd_write_121bytes): |
1949 | lddqu -121(%rsi), %xmm0 |
1950 | movdqu %xmm0, -121(%rdi) |
1951 | L(fwd_write_105bytes): |
1952 | lddqu -105(%rsi), %xmm0 |
1953 | movdqu %xmm0, -105(%rdi) |
1954 | L(fwd_write_89bytes): |
1955 | lddqu -89(%rsi), %xmm0 |
1956 | movdqu %xmm0, -89(%rdi) |
1957 | L(fwd_write_73bytes): |
1958 | lddqu -73(%rsi), %xmm0 |
1959 | movdqu %xmm0, -73(%rdi) |
1960 | L(fwd_write_57bytes): |
1961 | lddqu -57(%rsi), %xmm0 |
1962 | movdqu %xmm0, -57(%rdi) |
1963 | L(fwd_write_41bytes): |
1964 | lddqu -41(%rsi), %xmm0 |
1965 | movdqu %xmm0, -41(%rdi) |
1966 | L(fwd_write_25bytes): |
1967 | lddqu -25(%rsi), %xmm0 |
1968 | lddqu -16(%rsi), %xmm1 |
1969 | movdqu %xmm0, -25(%rdi) |
1970 | movdqu %xmm1, -16(%rdi) |
1971 | ret |
1972 | |
1973 | .p2align 4 |
1974 | L(fwd_write_9bytes): |
1975 | mov -9(%rsi), %rdx |
1976 | mov -4(%rsi), %ecx |
1977 | mov %rdx, -9(%rdi) |
1978 | mov %ecx, -4(%rdi) |
1979 | ret |
1980 | |
1981 | .p2align 4 |
1982 | L(fwd_write_136bytes): |
1983 | lddqu -136(%rsi), %xmm0 |
1984 | movdqu %xmm0, -136(%rdi) |
1985 | L(fwd_write_120bytes): |
1986 | lddqu -120(%rsi), %xmm0 |
1987 | movdqu %xmm0, -120(%rdi) |
1988 | L(fwd_write_104bytes): |
1989 | lddqu -104(%rsi), %xmm0 |
1990 | movdqu %xmm0, -104(%rdi) |
1991 | L(fwd_write_88bytes): |
1992 | lddqu -88(%rsi), %xmm0 |
1993 | movdqu %xmm0, -88(%rdi) |
1994 | L(fwd_write_72bytes): |
1995 | lddqu -72(%rsi), %xmm0 |
1996 | movdqu %xmm0, -72(%rdi) |
1997 | L(fwd_write_56bytes): |
1998 | lddqu -56(%rsi), %xmm0 |
1999 | movdqu %xmm0, -56(%rdi) |
2000 | L(fwd_write_40bytes): |
2001 | lddqu -40(%rsi), %xmm0 |
2002 | movdqu %xmm0, -40(%rdi) |
2003 | L(fwd_write_24bytes): |
2004 | lddqu -24(%rsi), %xmm0 |
2005 | lddqu -16(%rsi), %xmm1 |
2006 | movdqu %xmm0, -24(%rdi) |
2007 | movdqu %xmm1, -16(%rdi) |
2008 | ret |
2009 | |
2010 | .p2align 4 |
2011 | L(fwd_write_8bytes): |
2012 | mov -8(%rsi), %rdx |
2013 | mov %rdx, -8(%rdi) |
2014 | ret |
2015 | |
2016 | .p2align 4 |
2017 | L(fwd_write_135bytes): |
2018 | lddqu -135(%rsi), %xmm0 |
2019 | movdqu %xmm0, -135(%rdi) |
2020 | L(fwd_write_119bytes): |
2021 | lddqu -119(%rsi), %xmm0 |
2022 | movdqu %xmm0, -119(%rdi) |
2023 | L(fwd_write_103bytes): |
2024 | lddqu -103(%rsi), %xmm0 |
2025 | movdqu %xmm0, -103(%rdi) |
2026 | L(fwd_write_87bytes): |
2027 | lddqu -87(%rsi), %xmm0 |
2028 | movdqu %xmm0, -87(%rdi) |
2029 | L(fwd_write_71bytes): |
2030 | lddqu -71(%rsi), %xmm0 |
2031 | movdqu %xmm0, -71(%rdi) |
2032 | L(fwd_write_55bytes): |
2033 | lddqu -55(%rsi), %xmm0 |
2034 | movdqu %xmm0, -55(%rdi) |
2035 | L(fwd_write_39bytes): |
2036 | lddqu -39(%rsi), %xmm0 |
2037 | movdqu %xmm0, -39(%rdi) |
2038 | L(fwd_write_23bytes): |
2039 | lddqu -23(%rsi), %xmm0 |
2040 | lddqu -16(%rsi), %xmm1 |
2041 | movdqu %xmm0, -23(%rdi) |
2042 | movdqu %xmm1, -16(%rdi) |
2043 | ret |
2044 | |
2045 | .p2align 4 |
2046 | L(fwd_write_7bytes): |
2047 | mov -7(%rsi), %edx |
2048 | mov -4(%rsi), %ecx |
2049 | mov %edx, -7(%rdi) |
2050 | mov %ecx, -4(%rdi) |
2051 | ret |
2052 | |
2053 | .p2align 4 |
2054 | L(fwd_write_134bytes): |
2055 | lddqu -134(%rsi), %xmm0 |
2056 | movdqu %xmm0, -134(%rdi) |
2057 | L(fwd_write_118bytes): |
2058 | lddqu -118(%rsi), %xmm0 |
2059 | movdqu %xmm0, -118(%rdi) |
2060 | L(fwd_write_102bytes): |
2061 | lddqu -102(%rsi), %xmm0 |
2062 | movdqu %xmm0, -102(%rdi) |
2063 | L(fwd_write_86bytes): |
2064 | lddqu -86(%rsi), %xmm0 |
2065 | movdqu %xmm0, -86(%rdi) |
2066 | L(fwd_write_70bytes): |
2067 | lddqu -70(%rsi), %xmm0 |
2068 | movdqu %xmm0, -70(%rdi) |
2069 | L(fwd_write_54bytes): |
2070 | lddqu -54(%rsi), %xmm0 |
2071 | movdqu %xmm0, -54(%rdi) |
2072 | L(fwd_write_38bytes): |
2073 | lddqu -38(%rsi), %xmm0 |
2074 | movdqu %xmm0, -38(%rdi) |
2075 | L(fwd_write_22bytes): |
2076 | lddqu -22(%rsi), %xmm0 |
2077 | lddqu -16(%rsi), %xmm1 |
2078 | movdqu %xmm0, -22(%rdi) |
2079 | movdqu %xmm1, -16(%rdi) |
2080 | ret |
2081 | |
2082 | .p2align 4 |
2083 | L(fwd_write_6bytes): |
2084 | mov -6(%rsi), %edx |
2085 | mov -4(%rsi), %ecx |
2086 | mov %edx, -6(%rdi) |
2087 | mov %ecx, -4(%rdi) |
2088 | ret |
2089 | |
2090 | .p2align 4 |
2091 | L(fwd_write_133bytes): |
2092 | lddqu -133(%rsi), %xmm0 |
2093 | movdqu %xmm0, -133(%rdi) |
2094 | L(fwd_write_117bytes): |
2095 | lddqu -117(%rsi), %xmm0 |
2096 | movdqu %xmm0, -117(%rdi) |
2097 | L(fwd_write_101bytes): |
2098 | lddqu -101(%rsi), %xmm0 |
2099 | movdqu %xmm0, -101(%rdi) |
2100 | L(fwd_write_85bytes): |
2101 | lddqu -85(%rsi), %xmm0 |
2102 | movdqu %xmm0, -85(%rdi) |
2103 | L(fwd_write_69bytes): |
2104 | lddqu -69(%rsi), %xmm0 |
2105 | movdqu %xmm0, -69(%rdi) |
2106 | L(fwd_write_53bytes): |
2107 | lddqu -53(%rsi), %xmm0 |
2108 | movdqu %xmm0, -53(%rdi) |
2109 | L(fwd_write_37bytes): |
2110 | lddqu -37(%rsi), %xmm0 |
2111 | movdqu %xmm0, -37(%rdi) |
2112 | L(fwd_write_21bytes): |
2113 | lddqu -21(%rsi), %xmm0 |
2114 | lddqu -16(%rsi), %xmm1 |
2115 | movdqu %xmm0, -21(%rdi) |
2116 | movdqu %xmm1, -16(%rdi) |
2117 | ret |
2118 | |
2119 | .p2align 4 |
2120 | L(fwd_write_5bytes): |
2121 | mov -5(%rsi), %edx |
2122 | mov -4(%rsi), %ecx |
2123 | mov %edx, -5(%rdi) |
2124 | mov %ecx, -4(%rdi) |
2125 | ret |
2126 | |
2127 | .p2align 4 |
2128 | L(fwd_write_132bytes): |
2129 | lddqu -132(%rsi), %xmm0 |
2130 | movdqu %xmm0, -132(%rdi) |
2131 | L(fwd_write_116bytes): |
2132 | lddqu -116(%rsi), %xmm0 |
2133 | movdqu %xmm0, -116(%rdi) |
2134 | L(fwd_write_100bytes): |
2135 | lddqu -100(%rsi), %xmm0 |
2136 | movdqu %xmm0, -100(%rdi) |
2137 | L(fwd_write_84bytes): |
2138 | lddqu -84(%rsi), %xmm0 |
2139 | movdqu %xmm0, -84(%rdi) |
2140 | L(fwd_write_68bytes): |
2141 | lddqu -68(%rsi), %xmm0 |
2142 | movdqu %xmm0, -68(%rdi) |
2143 | L(fwd_write_52bytes): |
2144 | lddqu -52(%rsi), %xmm0 |
2145 | movdqu %xmm0, -52(%rdi) |
2146 | L(fwd_write_36bytes): |
2147 | lddqu -36(%rsi), %xmm0 |
2148 | movdqu %xmm0, -36(%rdi) |
2149 | L(fwd_write_20bytes): |
2150 | lddqu -20(%rsi), %xmm0 |
2151 | lddqu -16(%rsi), %xmm1 |
2152 | movdqu %xmm0, -20(%rdi) |
2153 | movdqu %xmm1, -16(%rdi) |
2154 | ret |
2155 | |
2156 | .p2align 4 |
2157 | L(fwd_write_4bytes): |
2158 | mov -4(%rsi), %edx |
2159 | mov %edx, -4(%rdi) |
2160 | ret |
2161 | |
2162 | .p2align 4 |
2163 | L(fwd_write_131bytes): |
2164 | lddqu -131(%rsi), %xmm0 |
2165 | movdqu %xmm0, -131(%rdi) |
2166 | L(fwd_write_115bytes): |
2167 | lddqu -115(%rsi), %xmm0 |
2168 | movdqu %xmm0, -115(%rdi) |
2169 | L(fwd_write_99bytes): |
2170 | lddqu -99(%rsi), %xmm0 |
2171 | movdqu %xmm0, -99(%rdi) |
2172 | L(fwd_write_83bytes): |
2173 | lddqu -83(%rsi), %xmm0 |
2174 | movdqu %xmm0, -83(%rdi) |
2175 | L(fwd_write_67bytes): |
2176 | lddqu -67(%rsi), %xmm0 |
2177 | movdqu %xmm0, -67(%rdi) |
2178 | L(fwd_write_51bytes): |
2179 | lddqu -51(%rsi), %xmm0 |
2180 | movdqu %xmm0, -51(%rdi) |
2181 | L(fwd_write_35bytes): |
2182 | lddqu -35(%rsi), %xmm0 |
2183 | movdqu %xmm0, -35(%rdi) |
2184 | L(fwd_write_19bytes): |
2185 | lddqu -19(%rsi), %xmm0 |
2186 | lddqu -16(%rsi), %xmm1 |
2187 | movdqu %xmm0, -19(%rdi) |
2188 | movdqu %xmm1, -16(%rdi) |
2189 | ret |
2190 | |
2191 | .p2align 4 |
2192 | L(fwd_write_3bytes): |
2193 | mov -3(%rsi), %dx |
2194 | mov -2(%rsi), %cx |
2195 | mov %dx, -3(%rdi) |
2196 | mov %cx, -2(%rdi) |
2197 | ret |
2198 | |
2199 | .p2align 4 |
2200 | L(fwd_write_130bytes): |
2201 | lddqu -130(%rsi), %xmm0 |
2202 | movdqu %xmm0, -130(%rdi) |
2203 | L(fwd_write_114bytes): |
2204 | lddqu -114(%rsi), %xmm0 |
2205 | movdqu %xmm0, -114(%rdi) |
2206 | L(fwd_write_98bytes): |
2207 | lddqu -98(%rsi), %xmm0 |
2208 | movdqu %xmm0, -98(%rdi) |
2209 | L(fwd_write_82bytes): |
2210 | lddqu -82(%rsi), %xmm0 |
2211 | movdqu %xmm0, -82(%rdi) |
2212 | L(fwd_write_66bytes): |
2213 | lddqu -66(%rsi), %xmm0 |
2214 | movdqu %xmm0, -66(%rdi) |
2215 | L(fwd_write_50bytes): |
2216 | lddqu -50(%rsi), %xmm0 |
2217 | movdqu %xmm0, -50(%rdi) |
2218 | L(fwd_write_34bytes): |
2219 | lddqu -34(%rsi), %xmm0 |
2220 | movdqu %xmm0, -34(%rdi) |
2221 | L(fwd_write_18bytes): |
2222 | lddqu -18(%rsi), %xmm0 |
2223 | lddqu -16(%rsi), %xmm1 |
2224 | movdqu %xmm0, -18(%rdi) |
2225 | movdqu %xmm1, -16(%rdi) |
2226 | ret |
2227 | |
2228 | .p2align 4 |
2229 | L(fwd_write_2bytes): |
2230 | movzwl -2(%rsi), %edx |
2231 | mov %dx, -2(%rdi) |
2232 | ret |
2233 | |
2234 | .p2align 4 |
2235 | L(fwd_write_129bytes): |
2236 | lddqu -129(%rsi), %xmm0 |
2237 | movdqu %xmm0, -129(%rdi) |
2238 | L(fwd_write_113bytes): |
2239 | lddqu -113(%rsi), %xmm0 |
2240 | movdqu %xmm0, -113(%rdi) |
2241 | L(fwd_write_97bytes): |
2242 | lddqu -97(%rsi), %xmm0 |
2243 | movdqu %xmm0, -97(%rdi) |
2244 | L(fwd_write_81bytes): |
2245 | lddqu -81(%rsi), %xmm0 |
2246 | movdqu %xmm0, -81(%rdi) |
2247 | L(fwd_write_65bytes): |
2248 | lddqu -65(%rsi), %xmm0 |
2249 | movdqu %xmm0, -65(%rdi) |
2250 | L(fwd_write_49bytes): |
2251 | lddqu -49(%rsi), %xmm0 |
2252 | movdqu %xmm0, -49(%rdi) |
2253 | L(fwd_write_33bytes): |
2254 | lddqu -33(%rsi), %xmm0 |
2255 | movdqu %xmm0, -33(%rdi) |
2256 | L(fwd_write_17bytes): |
2257 | lddqu -17(%rsi), %xmm0 |
2258 | lddqu -16(%rsi), %xmm1 |
2259 | movdqu %xmm0, -17(%rdi) |
2260 | movdqu %xmm1, -16(%rdi) |
2261 | ret |
2262 | |
2263 | .p2align 4 |
2264 | L(fwd_write_1bytes): |
2265 | movzbl -1(%rsi), %edx |
2266 | mov %dl, -1(%rdi) |
2267 | ret |
2268 | |
2269 | .p2align 4 |
2270 | L(bwd_write_128bytes): |
2271 | lddqu 112(%rsi), %xmm0 |
2272 | movdqu %xmm0, 112(%rdi) |
2273 | L(bwd_write_112bytes): |
2274 | lddqu 96(%rsi), %xmm0 |
2275 | movdqu %xmm0, 96(%rdi) |
2276 | L(bwd_write_96bytes): |
2277 | lddqu 80(%rsi), %xmm0 |
2278 | movdqu %xmm0, 80(%rdi) |
2279 | L(bwd_write_80bytes): |
2280 | lddqu 64(%rsi), %xmm0 |
2281 | movdqu %xmm0, 64(%rdi) |
2282 | L(bwd_write_64bytes): |
2283 | lddqu 48(%rsi), %xmm0 |
2284 | movdqu %xmm0, 48(%rdi) |
2285 | L(bwd_write_48bytes): |
2286 | lddqu 32(%rsi), %xmm0 |
2287 | movdqu %xmm0, 32(%rdi) |
2288 | L(bwd_write_32bytes): |
2289 | lddqu 16(%rsi), %xmm0 |
2290 | movdqu %xmm0, 16(%rdi) |
2291 | L(bwd_write_16bytes): |
2292 | lddqu (%rsi), %xmm0 |
2293 | movdqu %xmm0, (%rdi) |
2294 | L(bwd_write_0bytes): |
2295 | ret |
2296 | |
2297 | .p2align 4 |
2298 | L(bwd_write_143bytes): |
2299 | lddqu 127(%rsi), %xmm0 |
2300 | movdqu %xmm0, 127(%rdi) |
2301 | L(bwd_write_127bytes): |
2302 | lddqu 111(%rsi), %xmm0 |
2303 | movdqu %xmm0, 111(%rdi) |
2304 | L(bwd_write_111bytes): |
2305 | lddqu 95(%rsi), %xmm0 |
2306 | movdqu %xmm0, 95(%rdi) |
2307 | L(bwd_write_95bytes): |
2308 | lddqu 79(%rsi), %xmm0 |
2309 | movdqu %xmm0, 79(%rdi) |
2310 | L(bwd_write_79bytes): |
2311 | lddqu 63(%rsi), %xmm0 |
2312 | movdqu %xmm0, 63(%rdi) |
2313 | L(bwd_write_63bytes): |
2314 | lddqu 47(%rsi), %xmm0 |
2315 | movdqu %xmm0, 47(%rdi) |
2316 | L(bwd_write_47bytes): |
2317 | lddqu 31(%rsi), %xmm0 |
2318 | movdqu %xmm0, 31(%rdi) |
2319 | L(bwd_write_31bytes): |
2320 | lddqu 15(%rsi), %xmm0 |
2321 | lddqu (%rsi), %xmm1 |
2322 | movdqu %xmm0, 15(%rdi) |
2323 | movdqu %xmm1, (%rdi) |
2324 | ret |
2325 | |
2326 | |
2327 | .p2align 4 |
2328 | L(bwd_write_15bytes): |
2329 | mov 7(%rsi), %rdx |
2330 | mov (%rsi), %rcx |
2331 | mov %rdx, 7(%rdi) |
2332 | mov %rcx, (%rdi) |
2333 | ret |
2334 | |
2335 | .p2align 4 |
2336 | L(bwd_write_142bytes): |
2337 | lddqu 126(%rsi), %xmm0 |
2338 | movdqu %xmm0, 126(%rdi) |
2339 | L(bwd_write_126bytes): |
2340 | lddqu 110(%rsi), %xmm0 |
2341 | movdqu %xmm0, 110(%rdi) |
2342 | L(bwd_write_110bytes): |
2343 | lddqu 94(%rsi), %xmm0 |
2344 | movdqu %xmm0, 94(%rdi) |
2345 | L(bwd_write_94bytes): |
2346 | lddqu 78(%rsi), %xmm0 |
2347 | movdqu %xmm0, 78(%rdi) |
2348 | L(bwd_write_78bytes): |
2349 | lddqu 62(%rsi), %xmm0 |
2350 | movdqu %xmm0, 62(%rdi) |
2351 | L(bwd_write_62bytes): |
2352 | lddqu 46(%rsi), %xmm0 |
2353 | movdqu %xmm0, 46(%rdi) |
2354 | L(bwd_write_46bytes): |
2355 | lddqu 30(%rsi), %xmm0 |
2356 | movdqu %xmm0, 30(%rdi) |
2357 | L(bwd_write_30bytes): |
2358 | lddqu 14(%rsi), %xmm0 |
2359 | lddqu (%rsi), %xmm1 |
2360 | movdqu %xmm0, 14(%rdi) |
2361 | movdqu %xmm1, (%rdi) |
2362 | ret |
2363 | |
2364 | .p2align 4 |
2365 | L(bwd_write_14bytes): |
2366 | mov 6(%rsi), %rdx |
2367 | mov (%rsi), %rcx |
2368 | mov %rdx, 6(%rdi) |
2369 | mov %rcx, (%rdi) |
2370 | ret |
2371 | |
2372 | .p2align 4 |
2373 | L(bwd_write_141bytes): |
2374 | lddqu 125(%rsi), %xmm0 |
2375 | movdqu %xmm0, 125(%rdi) |
2376 | L(bwd_write_125bytes): |
2377 | lddqu 109(%rsi), %xmm0 |
2378 | movdqu %xmm0, 109(%rdi) |
2379 | L(bwd_write_109bytes): |
2380 | lddqu 93(%rsi), %xmm0 |
2381 | movdqu %xmm0, 93(%rdi) |
2382 | L(bwd_write_93bytes): |
2383 | lddqu 77(%rsi), %xmm0 |
2384 | movdqu %xmm0, 77(%rdi) |
2385 | L(bwd_write_77bytes): |
2386 | lddqu 61(%rsi), %xmm0 |
2387 | movdqu %xmm0, 61(%rdi) |
2388 | L(bwd_write_61bytes): |
2389 | lddqu 45(%rsi), %xmm0 |
2390 | movdqu %xmm0, 45(%rdi) |
2391 | L(bwd_write_45bytes): |
2392 | lddqu 29(%rsi), %xmm0 |
2393 | movdqu %xmm0, 29(%rdi) |
2394 | L(bwd_write_29bytes): |
2395 | lddqu 13(%rsi), %xmm0 |
2396 | lddqu (%rsi), %xmm1 |
2397 | movdqu %xmm0, 13(%rdi) |
2398 | movdqu %xmm1, (%rdi) |
2399 | ret |
2400 | |
2401 | .p2align 4 |
2402 | L(bwd_write_13bytes): |
2403 | mov 5(%rsi), %rdx |
2404 | mov (%rsi), %rcx |
2405 | mov %rdx, 5(%rdi) |
2406 | mov %rcx, (%rdi) |
2407 | ret |
2408 | |
2409 | .p2align 4 |
2410 | L(bwd_write_140bytes): |
2411 | lddqu 124(%rsi), %xmm0 |
2412 | movdqu %xmm0, 124(%rdi) |
2413 | L(bwd_write_124bytes): |
2414 | lddqu 108(%rsi), %xmm0 |
2415 | movdqu %xmm0, 108(%rdi) |
2416 | L(bwd_write_108bytes): |
2417 | lddqu 92(%rsi), %xmm0 |
2418 | movdqu %xmm0, 92(%rdi) |
2419 | L(bwd_write_92bytes): |
2420 | lddqu 76(%rsi), %xmm0 |
2421 | movdqu %xmm0, 76(%rdi) |
2422 | L(bwd_write_76bytes): |
2423 | lddqu 60(%rsi), %xmm0 |
2424 | movdqu %xmm0, 60(%rdi) |
2425 | L(bwd_write_60bytes): |
2426 | lddqu 44(%rsi), %xmm0 |
2427 | movdqu %xmm0, 44(%rdi) |
2428 | L(bwd_write_44bytes): |
2429 | lddqu 28(%rsi), %xmm0 |
2430 | movdqu %xmm0, 28(%rdi) |
2431 | L(bwd_write_28bytes): |
2432 | lddqu 12(%rsi), %xmm0 |
2433 | lddqu (%rsi), %xmm1 |
2434 | movdqu %xmm0, 12(%rdi) |
2435 | movdqu %xmm1, (%rdi) |
2436 | ret |
2437 | |
2438 | .p2align 4 |
2439 | L(bwd_write_12bytes): |
2440 | mov 4(%rsi), %rdx |
2441 | mov (%rsi), %rcx |
2442 | mov %rdx, 4(%rdi) |
2443 | mov %rcx, (%rdi) |
2444 | ret |
2445 | |
2446 | .p2align 4 |
2447 | L(bwd_write_139bytes): |
2448 | lddqu 123(%rsi), %xmm0 |
2449 | movdqu %xmm0, 123(%rdi) |
2450 | L(bwd_write_123bytes): |
2451 | lddqu 107(%rsi), %xmm0 |
2452 | movdqu %xmm0, 107(%rdi) |
2453 | L(bwd_write_107bytes): |
2454 | lddqu 91(%rsi), %xmm0 |
2455 | movdqu %xmm0, 91(%rdi) |
2456 | L(bwd_write_91bytes): |
2457 | lddqu 75(%rsi), %xmm0 |
2458 | movdqu %xmm0, 75(%rdi) |
2459 | L(bwd_write_75bytes): |
2460 | lddqu 59(%rsi), %xmm0 |
2461 | movdqu %xmm0, 59(%rdi) |
2462 | L(bwd_write_59bytes): |
2463 | lddqu 43(%rsi), %xmm0 |
2464 | movdqu %xmm0, 43(%rdi) |
2465 | L(bwd_write_43bytes): |
2466 | lddqu 27(%rsi), %xmm0 |
2467 | movdqu %xmm0, 27(%rdi) |
2468 | L(bwd_write_27bytes): |
2469 | lddqu 11(%rsi), %xmm0 |
2470 | lddqu (%rsi), %xmm1 |
2471 | movdqu %xmm0, 11(%rdi) |
2472 | movdqu %xmm1, (%rdi) |
2473 | ret |
2474 | |
2475 | .p2align 4 |
2476 | L(bwd_write_11bytes): |
2477 | mov 3(%rsi), %rdx |
2478 | mov (%rsi), %rcx |
2479 | mov %rdx, 3(%rdi) |
2480 | mov %rcx, (%rdi) |
2481 | ret |
2482 | |
2483 | .p2align 4 |
2484 | L(bwd_write_138bytes): |
2485 | lddqu 122(%rsi), %xmm0 |
2486 | movdqu %xmm0, 122(%rdi) |
2487 | L(bwd_write_122bytes): |
2488 | lddqu 106(%rsi), %xmm0 |
2489 | movdqu %xmm0, 106(%rdi) |
2490 | L(bwd_write_106bytes): |
2491 | lddqu 90(%rsi), %xmm0 |
2492 | movdqu %xmm0, 90(%rdi) |
2493 | L(bwd_write_90bytes): |
2494 | lddqu 74(%rsi), %xmm0 |
2495 | movdqu %xmm0, 74(%rdi) |
2496 | L(bwd_write_74bytes): |
2497 | lddqu 58(%rsi), %xmm0 |
2498 | movdqu %xmm0, 58(%rdi) |
2499 | L(bwd_write_58bytes): |
2500 | lddqu 42(%rsi), %xmm0 |
2501 | movdqu %xmm0, 42(%rdi) |
2502 | L(bwd_write_42bytes): |
2503 | lddqu 26(%rsi), %xmm0 |
2504 | movdqu %xmm0, 26(%rdi) |
2505 | L(bwd_write_26bytes): |
2506 | lddqu 10(%rsi), %xmm0 |
2507 | lddqu (%rsi), %xmm1 |
2508 | movdqu %xmm0, 10(%rdi) |
2509 | movdqu %xmm1, (%rdi) |
2510 | ret |
2511 | |
2512 | .p2align 4 |
2513 | L(bwd_write_10bytes): |
2514 | mov 2(%rsi), %rdx |
2515 | mov (%rsi), %rcx |
2516 | mov %rdx, 2(%rdi) |
2517 | mov %rcx, (%rdi) |
2518 | ret |
2519 | |
2520 | .p2align 4 |
2521 | L(bwd_write_137bytes): |
2522 | lddqu 121(%rsi), %xmm0 |
2523 | movdqu %xmm0, 121(%rdi) |
2524 | L(bwd_write_121bytes): |
2525 | lddqu 105(%rsi), %xmm0 |
2526 | movdqu %xmm0, 105(%rdi) |
2527 | L(bwd_write_105bytes): |
2528 | lddqu 89(%rsi), %xmm0 |
2529 | movdqu %xmm0, 89(%rdi) |
2530 | L(bwd_write_89bytes): |
2531 | lddqu 73(%rsi), %xmm0 |
2532 | movdqu %xmm0, 73(%rdi) |
2533 | L(bwd_write_73bytes): |
2534 | lddqu 57(%rsi), %xmm0 |
2535 | movdqu %xmm0, 57(%rdi) |
2536 | L(bwd_write_57bytes): |
2537 | lddqu 41(%rsi), %xmm0 |
2538 | movdqu %xmm0, 41(%rdi) |
2539 | L(bwd_write_41bytes): |
2540 | lddqu 25(%rsi), %xmm0 |
2541 | movdqu %xmm0, 25(%rdi) |
2542 | L(bwd_write_25bytes): |
2543 | lddqu 9(%rsi), %xmm0 |
2544 | lddqu (%rsi), %xmm1 |
2545 | movdqu %xmm0, 9(%rdi) |
2546 | movdqu %xmm1, (%rdi) |
2547 | ret |
2548 | |
2549 | .p2align 4 |
2550 | L(bwd_write_9bytes): |
2551 | mov 1(%rsi), %rdx |
2552 | mov (%rsi), %rcx |
2553 | mov %rdx, 1(%rdi) |
2554 | mov %rcx, (%rdi) |
2555 | ret |
2556 | |
2557 | .p2align 4 |
2558 | L(bwd_write_136bytes): |
2559 | lddqu 120(%rsi), %xmm0 |
2560 | movdqu %xmm0, 120(%rdi) |
2561 | L(bwd_write_120bytes): |
2562 | lddqu 104(%rsi), %xmm0 |
2563 | movdqu %xmm0, 104(%rdi) |
2564 | L(bwd_write_104bytes): |
2565 | lddqu 88(%rsi), %xmm0 |
2566 | movdqu %xmm0, 88(%rdi) |
2567 | L(bwd_write_88bytes): |
2568 | lddqu 72(%rsi), %xmm0 |
2569 | movdqu %xmm0, 72(%rdi) |
2570 | L(bwd_write_72bytes): |
2571 | lddqu 56(%rsi), %xmm0 |
2572 | movdqu %xmm0, 56(%rdi) |
2573 | L(bwd_write_56bytes): |
2574 | lddqu 40(%rsi), %xmm0 |
2575 | movdqu %xmm0, 40(%rdi) |
2576 | L(bwd_write_40bytes): |
2577 | lddqu 24(%rsi), %xmm0 |
2578 | movdqu %xmm0, 24(%rdi) |
2579 | L(bwd_write_24bytes): |
2580 | lddqu 8(%rsi), %xmm0 |
2581 | lddqu (%rsi), %xmm1 |
2582 | movdqu %xmm0, 8(%rdi) |
2583 | movdqu %xmm1, (%rdi) |
2584 | ret |
2585 | |
2586 | .p2align 4 |
2587 | L(bwd_write_8bytes): |
2588 | mov (%rsi), %rdx |
2589 | mov %rdx, (%rdi) |
2590 | ret |
2591 | |
2592 | .p2align 4 |
2593 | L(bwd_write_135bytes): |
2594 | lddqu 119(%rsi), %xmm0 |
2595 | movdqu %xmm0, 119(%rdi) |
2596 | L(bwd_write_119bytes): |
2597 | lddqu 103(%rsi), %xmm0 |
2598 | movdqu %xmm0, 103(%rdi) |
2599 | L(bwd_write_103bytes): |
2600 | lddqu 87(%rsi), %xmm0 |
2601 | movdqu %xmm0, 87(%rdi) |
2602 | L(bwd_write_87bytes): |
2603 | lddqu 71(%rsi), %xmm0 |
2604 | movdqu %xmm0, 71(%rdi) |
2605 | L(bwd_write_71bytes): |
2606 | lddqu 55(%rsi), %xmm0 |
2607 | movdqu %xmm0, 55(%rdi) |
2608 | L(bwd_write_55bytes): |
2609 | lddqu 39(%rsi), %xmm0 |
2610 | movdqu %xmm0, 39(%rdi) |
2611 | L(bwd_write_39bytes): |
2612 | lddqu 23(%rsi), %xmm0 |
2613 | movdqu %xmm0, 23(%rdi) |
2614 | L(bwd_write_23bytes): |
2615 | lddqu 7(%rsi), %xmm0 |
2616 | lddqu (%rsi), %xmm1 |
2617 | movdqu %xmm0, 7(%rdi) |
2618 | movdqu %xmm1, (%rdi) |
2619 | ret |
2620 | |
2621 | .p2align 4 |
2622 | L(bwd_write_7bytes): |
2623 | mov 3(%rsi), %edx |
2624 | mov (%rsi), %ecx |
2625 | mov %edx, 3(%rdi) |
2626 | mov %ecx, (%rdi) |
2627 | ret |
2628 | |
2629 | .p2align 4 |
2630 | L(bwd_write_134bytes): |
2631 | lddqu 118(%rsi), %xmm0 |
2632 | movdqu %xmm0, 118(%rdi) |
2633 | L(bwd_write_118bytes): |
2634 | lddqu 102(%rsi), %xmm0 |
2635 | movdqu %xmm0, 102(%rdi) |
2636 | L(bwd_write_102bytes): |
2637 | lddqu 86(%rsi), %xmm0 |
2638 | movdqu %xmm0, 86(%rdi) |
2639 | L(bwd_write_86bytes): |
2640 | lddqu 70(%rsi), %xmm0 |
2641 | movdqu %xmm0, 70(%rdi) |
2642 | L(bwd_write_70bytes): |
2643 | lddqu 54(%rsi), %xmm0 |
2644 | movdqu %xmm0, 54(%rdi) |
2645 | L(bwd_write_54bytes): |
2646 | lddqu 38(%rsi), %xmm0 |
2647 | movdqu %xmm0, 38(%rdi) |
2648 | L(bwd_write_38bytes): |
2649 | lddqu 22(%rsi), %xmm0 |
2650 | movdqu %xmm0, 22(%rdi) |
2651 | L(bwd_write_22bytes): |
2652 | lddqu 6(%rsi), %xmm0 |
2653 | lddqu (%rsi), %xmm1 |
2654 | movdqu %xmm0, 6(%rdi) |
2655 | movdqu %xmm1, (%rdi) |
2656 | ret |
2657 | |
2658 | .p2align 4 |
2659 | L(bwd_write_6bytes): |
2660 | mov 2(%rsi), %edx |
2661 | mov (%rsi), %ecx |
2662 | mov %edx, 2(%rdi) |
2663 | mov %ecx, (%rdi) |
2664 | ret |
2665 | |
2666 | .p2align 4 |
2667 | L(bwd_write_133bytes): |
2668 | lddqu 117(%rsi), %xmm0 |
2669 | movdqu %xmm0, 117(%rdi) |
2670 | L(bwd_write_117bytes): |
2671 | lddqu 101(%rsi), %xmm0 |
2672 | movdqu %xmm0, 101(%rdi) |
2673 | L(bwd_write_101bytes): |
2674 | lddqu 85(%rsi), %xmm0 |
2675 | movdqu %xmm0, 85(%rdi) |
2676 | L(bwd_write_85bytes): |
2677 | lddqu 69(%rsi), %xmm0 |
2678 | movdqu %xmm0, 69(%rdi) |
2679 | L(bwd_write_69bytes): |
2680 | lddqu 53(%rsi), %xmm0 |
2681 | movdqu %xmm0, 53(%rdi) |
2682 | L(bwd_write_53bytes): |
2683 | lddqu 37(%rsi), %xmm0 |
2684 | movdqu %xmm0, 37(%rdi) |
2685 | L(bwd_write_37bytes): |
2686 | lddqu 21(%rsi), %xmm0 |
2687 | movdqu %xmm0, 21(%rdi) |
2688 | L(bwd_write_21bytes): |
2689 | lddqu 5(%rsi), %xmm0 |
2690 | lddqu (%rsi), %xmm1 |
2691 | movdqu %xmm0, 5(%rdi) |
2692 | movdqu %xmm1, (%rdi) |
2693 | ret |
2694 | |
2695 | .p2align 4 |
2696 | L(bwd_write_5bytes): |
2697 | mov 1(%rsi), %edx |
2698 | mov (%rsi), %ecx |
2699 | mov %edx, 1(%rdi) |
2700 | mov %ecx, (%rdi) |
2701 | ret |
2702 | |
2703 | .p2align 4 |
2704 | L(bwd_write_132bytes): |
2705 | lddqu 116(%rsi), %xmm0 |
2706 | movdqu %xmm0, 116(%rdi) |
2707 | L(bwd_write_116bytes): |
2708 | lddqu 100(%rsi), %xmm0 |
2709 | movdqu %xmm0, 100(%rdi) |
2710 | L(bwd_write_100bytes): |
2711 | lddqu 84(%rsi), %xmm0 |
2712 | movdqu %xmm0, 84(%rdi) |
2713 | L(bwd_write_84bytes): |
2714 | lddqu 68(%rsi), %xmm0 |
2715 | movdqu %xmm0, 68(%rdi) |
2716 | L(bwd_write_68bytes): |
2717 | lddqu 52(%rsi), %xmm0 |
2718 | movdqu %xmm0, 52(%rdi) |
2719 | L(bwd_write_52bytes): |
2720 | lddqu 36(%rsi), %xmm0 |
2721 | movdqu %xmm0, 36(%rdi) |
2722 | L(bwd_write_36bytes): |
2723 | lddqu 20(%rsi), %xmm0 |
2724 | movdqu %xmm0, 20(%rdi) |
2725 | L(bwd_write_20bytes): |
2726 | lddqu 4(%rsi), %xmm0 |
2727 | lddqu (%rsi), %xmm1 |
2728 | movdqu %xmm0, 4(%rdi) |
2729 | movdqu %xmm1, (%rdi) |
2730 | ret |
2731 | |
2732 | .p2align 4 |
2733 | L(bwd_write_4bytes): |
2734 | mov (%rsi), %edx |
2735 | mov %edx, (%rdi) |
2736 | ret |
2737 | |
2738 | .p2align 4 |
2739 | L(bwd_write_131bytes): |
2740 | lddqu 115(%rsi), %xmm0 |
2741 | movdqu %xmm0, 115(%rdi) |
2742 | L(bwd_write_115bytes): |
2743 | lddqu 99(%rsi), %xmm0 |
2744 | movdqu %xmm0, 99(%rdi) |
2745 | L(bwd_write_99bytes): |
2746 | lddqu 83(%rsi), %xmm0 |
2747 | movdqu %xmm0, 83(%rdi) |
2748 | L(bwd_write_83bytes): |
2749 | lddqu 67(%rsi), %xmm0 |
2750 | movdqu %xmm0, 67(%rdi) |
2751 | L(bwd_write_67bytes): |
2752 | lddqu 51(%rsi), %xmm0 |
2753 | movdqu %xmm0, 51(%rdi) |
2754 | L(bwd_write_51bytes): |
2755 | lddqu 35(%rsi), %xmm0 |
2756 | movdqu %xmm0, 35(%rdi) |
2757 | L(bwd_write_35bytes): |
2758 | lddqu 19(%rsi), %xmm0 |
2759 | movdqu %xmm0, 19(%rdi) |
2760 | L(bwd_write_19bytes): |
2761 | lddqu 3(%rsi), %xmm0 |
2762 | lddqu (%rsi), %xmm1 |
2763 | movdqu %xmm0, 3(%rdi) |
2764 | movdqu %xmm1, (%rdi) |
2765 | ret |
2766 | |
2767 | .p2align 4 |
2768 | L(bwd_write_3bytes): |
2769 | mov 1(%rsi), %dx |
2770 | mov (%rsi), %cx |
2771 | mov %dx, 1(%rdi) |
2772 | mov %cx, (%rdi) |
2773 | ret |
2774 | |
2775 | .p2align 4 |
2776 | L(bwd_write_130bytes): |
2777 | lddqu 114(%rsi), %xmm0 |
2778 | movdqu %xmm0, 114(%rdi) |
2779 | L(bwd_write_114bytes): |
2780 | lddqu 98(%rsi), %xmm0 |
2781 | movdqu %xmm0, 98(%rdi) |
2782 | L(bwd_write_98bytes): |
2783 | lddqu 82(%rsi), %xmm0 |
2784 | movdqu %xmm0, 82(%rdi) |
2785 | L(bwd_write_82bytes): |
2786 | lddqu 66(%rsi), %xmm0 |
2787 | movdqu %xmm0, 66(%rdi) |
2788 | L(bwd_write_66bytes): |
2789 | lddqu 50(%rsi), %xmm0 |
2790 | movdqu %xmm0, 50(%rdi) |
2791 | L(bwd_write_50bytes): |
2792 | lddqu 34(%rsi), %xmm0 |
2793 | movdqu %xmm0, 34(%rdi) |
2794 | L(bwd_write_34bytes): |
2795 | lddqu 18(%rsi), %xmm0 |
2796 | movdqu %xmm0, 18(%rdi) |
2797 | L(bwd_write_18bytes): |
2798 | lddqu 2(%rsi), %xmm0 |
2799 | lddqu (%rsi), %xmm1 |
2800 | movdqu %xmm0, 2(%rdi) |
2801 | movdqu %xmm1, (%rdi) |
2802 | ret |
2803 | |
2804 | .p2align 4 |
2805 | L(bwd_write_2bytes): |
2806 | movzwl (%rsi), %edx |
2807 | mov %dx, (%rdi) |
2808 | ret |
2809 | |
2810 | .p2align 4 |
2811 | L(bwd_write_129bytes): |
2812 | lddqu 113(%rsi), %xmm0 |
2813 | movdqu %xmm0, 113(%rdi) |
2814 | L(bwd_write_113bytes): |
2815 | lddqu 97(%rsi), %xmm0 |
2816 | movdqu %xmm0, 97(%rdi) |
2817 | L(bwd_write_97bytes): |
2818 | lddqu 81(%rsi), %xmm0 |
2819 | movdqu %xmm0, 81(%rdi) |
2820 | L(bwd_write_81bytes): |
2821 | lddqu 65(%rsi), %xmm0 |
2822 | movdqu %xmm0, 65(%rdi) |
2823 | L(bwd_write_65bytes): |
2824 | lddqu 49(%rsi), %xmm0 |
2825 | movdqu %xmm0, 49(%rdi) |
2826 | L(bwd_write_49bytes): |
2827 | lddqu 33(%rsi), %xmm0 |
2828 | movdqu %xmm0, 33(%rdi) |
2829 | L(bwd_write_33bytes): |
2830 | lddqu 17(%rsi), %xmm0 |
2831 | movdqu %xmm0, 17(%rdi) |
2832 | L(bwd_write_17bytes): |
2833 | lddqu 1(%rsi), %xmm0 |
2834 | lddqu (%rsi), %xmm1 |
2835 | movdqu %xmm0, 1(%rdi) |
2836 | movdqu %xmm1, (%rdi) |
2837 | ret |
2838 | |
2839 | .p2align 4 |
2840 | L(bwd_write_1bytes): |
2841 | movzbl (%rsi), %edx |
2842 | mov %dl, (%rdi) |
2843 | ret |
2844 | |
2845 | END (MEMCPY) |
2846 | |
2847 | .section .rodata.ssse3,"a" ,@progbits |
2848 | .p2align 3 |
2849 | L(table_144_bytes_bwd): |
2850 | .int JMPTBL (L(bwd_write_0bytes), L(table_144_bytes_bwd)) |
2851 | .int JMPTBL (L(bwd_write_1bytes), L(table_144_bytes_bwd)) |
2852 | .int JMPTBL (L(bwd_write_2bytes), L(table_144_bytes_bwd)) |
2853 | .int JMPTBL (L(bwd_write_3bytes), L(table_144_bytes_bwd)) |
2854 | .int JMPTBL (L(bwd_write_4bytes), L(table_144_bytes_bwd)) |
2855 | .int JMPTBL (L(bwd_write_5bytes), L(table_144_bytes_bwd)) |
2856 | .int JMPTBL (L(bwd_write_6bytes), L(table_144_bytes_bwd)) |
2857 | .int JMPTBL (L(bwd_write_7bytes), L(table_144_bytes_bwd)) |
2858 | .int JMPTBL (L(bwd_write_8bytes), L(table_144_bytes_bwd)) |
2859 | .int JMPTBL (L(bwd_write_9bytes), L(table_144_bytes_bwd)) |
2860 | .int JMPTBL (L(bwd_write_10bytes), L(table_144_bytes_bwd)) |
2861 | .int JMPTBL (L(bwd_write_11bytes), L(table_144_bytes_bwd)) |
2862 | .int JMPTBL (L(bwd_write_12bytes), L(table_144_bytes_bwd)) |
2863 | .int JMPTBL (L(bwd_write_13bytes), L(table_144_bytes_bwd)) |
2864 | .int JMPTBL (L(bwd_write_14bytes), L(table_144_bytes_bwd)) |
2865 | .int JMPTBL (L(bwd_write_15bytes), L(table_144_bytes_bwd)) |
2866 | .int JMPTBL (L(bwd_write_16bytes), L(table_144_bytes_bwd)) |
2867 | .int JMPTBL (L(bwd_write_17bytes), L(table_144_bytes_bwd)) |
2868 | .int JMPTBL (L(bwd_write_18bytes), L(table_144_bytes_bwd)) |
2869 | .int JMPTBL (L(bwd_write_19bytes), L(table_144_bytes_bwd)) |
2870 | .int JMPTBL (L(bwd_write_20bytes), L(table_144_bytes_bwd)) |
2871 | .int JMPTBL (L(bwd_write_21bytes), L(table_144_bytes_bwd)) |
2872 | .int JMPTBL (L(bwd_write_22bytes), L(table_144_bytes_bwd)) |
2873 | .int JMPTBL (L(bwd_write_23bytes), L(table_144_bytes_bwd)) |
2874 | .int JMPTBL (L(bwd_write_24bytes), L(table_144_bytes_bwd)) |
2875 | .int JMPTBL (L(bwd_write_25bytes), L(table_144_bytes_bwd)) |
2876 | .int JMPTBL (L(bwd_write_26bytes), L(table_144_bytes_bwd)) |
2877 | .int JMPTBL (L(bwd_write_27bytes), L(table_144_bytes_bwd)) |
2878 | .int JMPTBL (L(bwd_write_28bytes), L(table_144_bytes_bwd)) |
2879 | .int JMPTBL (L(bwd_write_29bytes), L(table_144_bytes_bwd)) |
2880 | .int JMPTBL (L(bwd_write_30bytes), L(table_144_bytes_bwd)) |
2881 | .int JMPTBL (L(bwd_write_31bytes), L(table_144_bytes_bwd)) |
2882 | .int JMPTBL (L(bwd_write_32bytes), L(table_144_bytes_bwd)) |
2883 | .int JMPTBL (L(bwd_write_33bytes), L(table_144_bytes_bwd)) |
2884 | .int JMPTBL (L(bwd_write_34bytes), L(table_144_bytes_bwd)) |
2885 | .int JMPTBL (L(bwd_write_35bytes), L(table_144_bytes_bwd)) |
2886 | .int JMPTBL (L(bwd_write_36bytes), L(table_144_bytes_bwd)) |
2887 | .int JMPTBL (L(bwd_write_37bytes), L(table_144_bytes_bwd)) |
2888 | .int JMPTBL (L(bwd_write_38bytes), L(table_144_bytes_bwd)) |
2889 | .int JMPTBL (L(bwd_write_39bytes), L(table_144_bytes_bwd)) |
2890 | .int JMPTBL (L(bwd_write_40bytes), L(table_144_bytes_bwd)) |
2891 | .int JMPTBL (L(bwd_write_41bytes), L(table_144_bytes_bwd)) |
2892 | .int JMPTBL (L(bwd_write_42bytes), L(table_144_bytes_bwd)) |
2893 | .int JMPTBL (L(bwd_write_43bytes), L(table_144_bytes_bwd)) |
2894 | .int JMPTBL (L(bwd_write_44bytes), L(table_144_bytes_bwd)) |
2895 | .int JMPTBL (L(bwd_write_45bytes), L(table_144_bytes_bwd)) |
2896 | .int JMPTBL (L(bwd_write_46bytes), L(table_144_bytes_bwd)) |
2897 | .int JMPTBL (L(bwd_write_47bytes), L(table_144_bytes_bwd)) |
2898 | .int JMPTBL (L(bwd_write_48bytes), L(table_144_bytes_bwd)) |
2899 | .int JMPTBL (L(bwd_write_49bytes), L(table_144_bytes_bwd)) |
2900 | .int JMPTBL (L(bwd_write_50bytes), L(table_144_bytes_bwd)) |
2901 | .int JMPTBL (L(bwd_write_51bytes), L(table_144_bytes_bwd)) |
2902 | .int JMPTBL (L(bwd_write_52bytes), L(table_144_bytes_bwd)) |
2903 | .int JMPTBL (L(bwd_write_53bytes), L(table_144_bytes_bwd)) |
2904 | .int JMPTBL (L(bwd_write_54bytes), L(table_144_bytes_bwd)) |
2905 | .int JMPTBL (L(bwd_write_55bytes), L(table_144_bytes_bwd)) |
2906 | .int JMPTBL (L(bwd_write_56bytes), L(table_144_bytes_bwd)) |
2907 | .int JMPTBL (L(bwd_write_57bytes), L(table_144_bytes_bwd)) |
2908 | .int JMPTBL (L(bwd_write_58bytes), L(table_144_bytes_bwd)) |
2909 | .int JMPTBL (L(bwd_write_59bytes), L(table_144_bytes_bwd)) |
2910 | .int JMPTBL (L(bwd_write_60bytes), L(table_144_bytes_bwd)) |
2911 | .int JMPTBL (L(bwd_write_61bytes), L(table_144_bytes_bwd)) |
2912 | .int JMPTBL (L(bwd_write_62bytes), L(table_144_bytes_bwd)) |
2913 | .int JMPTBL (L(bwd_write_63bytes), L(table_144_bytes_bwd)) |
2914 | .int JMPTBL (L(bwd_write_64bytes), L(table_144_bytes_bwd)) |
2915 | .int JMPTBL (L(bwd_write_65bytes), L(table_144_bytes_bwd)) |
2916 | .int JMPTBL (L(bwd_write_66bytes), L(table_144_bytes_bwd)) |
2917 | .int JMPTBL (L(bwd_write_67bytes), L(table_144_bytes_bwd)) |
2918 | .int JMPTBL (L(bwd_write_68bytes), L(table_144_bytes_bwd)) |
2919 | .int JMPTBL (L(bwd_write_69bytes), L(table_144_bytes_bwd)) |
2920 | .int JMPTBL (L(bwd_write_70bytes), L(table_144_bytes_bwd)) |
2921 | .int JMPTBL (L(bwd_write_71bytes), L(table_144_bytes_bwd)) |
2922 | .int JMPTBL (L(bwd_write_72bytes), L(table_144_bytes_bwd)) |
2923 | .int JMPTBL (L(bwd_write_73bytes), L(table_144_bytes_bwd)) |
2924 | .int JMPTBL (L(bwd_write_74bytes), L(table_144_bytes_bwd)) |
2925 | .int JMPTBL (L(bwd_write_75bytes), L(table_144_bytes_bwd)) |
2926 | .int JMPTBL (L(bwd_write_76bytes), L(table_144_bytes_bwd)) |
2927 | .int JMPTBL (L(bwd_write_77bytes), L(table_144_bytes_bwd)) |
2928 | .int JMPTBL (L(bwd_write_78bytes), L(table_144_bytes_bwd)) |
2929 | .int JMPTBL (L(bwd_write_79bytes), L(table_144_bytes_bwd)) |
2930 | .int JMPTBL (L(bwd_write_80bytes), L(table_144_bytes_bwd)) |
2931 | .int JMPTBL (L(bwd_write_81bytes), L(table_144_bytes_bwd)) |
2932 | .int JMPTBL (L(bwd_write_82bytes), L(table_144_bytes_bwd)) |
2933 | .int JMPTBL (L(bwd_write_83bytes), L(table_144_bytes_bwd)) |
2934 | .int JMPTBL (L(bwd_write_84bytes), L(table_144_bytes_bwd)) |
2935 | .int JMPTBL (L(bwd_write_85bytes), L(table_144_bytes_bwd)) |
2936 | .int JMPTBL (L(bwd_write_86bytes), L(table_144_bytes_bwd)) |
2937 | .int JMPTBL (L(bwd_write_87bytes), L(table_144_bytes_bwd)) |
2938 | .int JMPTBL (L(bwd_write_88bytes), L(table_144_bytes_bwd)) |
2939 | .int JMPTBL (L(bwd_write_89bytes), L(table_144_bytes_bwd)) |
2940 | .int JMPTBL (L(bwd_write_90bytes), L(table_144_bytes_bwd)) |
2941 | .int JMPTBL (L(bwd_write_91bytes), L(table_144_bytes_bwd)) |
2942 | .int JMPTBL (L(bwd_write_92bytes), L(table_144_bytes_bwd)) |
2943 | .int JMPTBL (L(bwd_write_93bytes), L(table_144_bytes_bwd)) |
2944 | .int JMPTBL (L(bwd_write_94bytes), L(table_144_bytes_bwd)) |
2945 | .int JMPTBL (L(bwd_write_95bytes), L(table_144_bytes_bwd)) |
2946 | .int JMPTBL (L(bwd_write_96bytes), L(table_144_bytes_bwd)) |
2947 | .int JMPTBL (L(bwd_write_97bytes), L(table_144_bytes_bwd)) |
2948 | .int JMPTBL (L(bwd_write_98bytes), L(table_144_bytes_bwd)) |
2949 | .int JMPTBL (L(bwd_write_99bytes), L(table_144_bytes_bwd)) |
2950 | .int JMPTBL (L(bwd_write_100bytes), L(table_144_bytes_bwd)) |
2951 | .int JMPTBL (L(bwd_write_101bytes), L(table_144_bytes_bwd)) |
2952 | .int JMPTBL (L(bwd_write_102bytes), L(table_144_bytes_bwd)) |
2953 | .int JMPTBL (L(bwd_write_103bytes), L(table_144_bytes_bwd)) |
2954 | .int JMPTBL (L(bwd_write_104bytes), L(table_144_bytes_bwd)) |
2955 | .int JMPTBL (L(bwd_write_105bytes), L(table_144_bytes_bwd)) |
2956 | .int JMPTBL (L(bwd_write_106bytes), L(table_144_bytes_bwd)) |
2957 | .int JMPTBL (L(bwd_write_107bytes), L(table_144_bytes_bwd)) |
2958 | .int JMPTBL (L(bwd_write_108bytes), L(table_144_bytes_bwd)) |
2959 | .int JMPTBL (L(bwd_write_109bytes), L(table_144_bytes_bwd)) |
2960 | .int JMPTBL (L(bwd_write_110bytes), L(table_144_bytes_bwd)) |
2961 | .int JMPTBL (L(bwd_write_111bytes), L(table_144_bytes_bwd)) |
2962 | .int JMPTBL (L(bwd_write_112bytes), L(table_144_bytes_bwd)) |
2963 | .int JMPTBL (L(bwd_write_113bytes), L(table_144_bytes_bwd)) |
2964 | .int JMPTBL (L(bwd_write_114bytes), L(table_144_bytes_bwd)) |
2965 | .int JMPTBL (L(bwd_write_115bytes), L(table_144_bytes_bwd)) |
2966 | .int JMPTBL (L(bwd_write_116bytes), L(table_144_bytes_bwd)) |
2967 | .int JMPTBL (L(bwd_write_117bytes), L(table_144_bytes_bwd)) |
2968 | .int JMPTBL (L(bwd_write_118bytes), L(table_144_bytes_bwd)) |
2969 | .int JMPTBL (L(bwd_write_119bytes), L(table_144_bytes_bwd)) |
2970 | .int JMPTBL (L(bwd_write_120bytes), L(table_144_bytes_bwd)) |
2971 | .int JMPTBL (L(bwd_write_121bytes), L(table_144_bytes_bwd)) |
2972 | .int JMPTBL (L(bwd_write_122bytes), L(table_144_bytes_bwd)) |
2973 | .int JMPTBL (L(bwd_write_123bytes), L(table_144_bytes_bwd)) |
2974 | .int JMPTBL (L(bwd_write_124bytes), L(table_144_bytes_bwd)) |
2975 | .int JMPTBL (L(bwd_write_125bytes), L(table_144_bytes_bwd)) |
2976 | .int JMPTBL (L(bwd_write_126bytes), L(table_144_bytes_bwd)) |
2977 | .int JMPTBL (L(bwd_write_127bytes), L(table_144_bytes_bwd)) |
2978 | .int JMPTBL (L(bwd_write_128bytes), L(table_144_bytes_bwd)) |
2979 | .int JMPTBL (L(bwd_write_129bytes), L(table_144_bytes_bwd)) |
2980 | .int JMPTBL (L(bwd_write_130bytes), L(table_144_bytes_bwd)) |
2981 | .int JMPTBL (L(bwd_write_131bytes), L(table_144_bytes_bwd)) |
2982 | .int JMPTBL (L(bwd_write_132bytes), L(table_144_bytes_bwd)) |
2983 | .int JMPTBL (L(bwd_write_133bytes), L(table_144_bytes_bwd)) |
2984 | .int JMPTBL (L(bwd_write_134bytes), L(table_144_bytes_bwd)) |
2985 | .int JMPTBL (L(bwd_write_135bytes), L(table_144_bytes_bwd)) |
2986 | .int JMPTBL (L(bwd_write_136bytes), L(table_144_bytes_bwd)) |
2987 | .int JMPTBL (L(bwd_write_137bytes), L(table_144_bytes_bwd)) |
2988 | .int JMPTBL (L(bwd_write_138bytes), L(table_144_bytes_bwd)) |
2989 | .int JMPTBL (L(bwd_write_139bytes), L(table_144_bytes_bwd)) |
2990 | .int JMPTBL (L(bwd_write_140bytes), L(table_144_bytes_bwd)) |
2991 | .int JMPTBL (L(bwd_write_141bytes), L(table_144_bytes_bwd)) |
2992 | .int JMPTBL (L(bwd_write_142bytes), L(table_144_bytes_bwd)) |
2993 | .int JMPTBL (L(bwd_write_143bytes), L(table_144_bytes_bwd)) |
2994 | |
2995 | .p2align 3 |
2996 | L(table_144_bytes_fwd): |
2997 | .int JMPTBL (L(fwd_write_0bytes), L(table_144_bytes_fwd)) |
2998 | .int JMPTBL (L(fwd_write_1bytes), L(table_144_bytes_fwd)) |
2999 | .int JMPTBL (L(fwd_write_2bytes), L(table_144_bytes_fwd)) |
3000 | .int JMPTBL (L(fwd_write_3bytes), L(table_144_bytes_fwd)) |
3001 | .int JMPTBL (L(fwd_write_4bytes), L(table_144_bytes_fwd)) |
3002 | .int JMPTBL (L(fwd_write_5bytes), L(table_144_bytes_fwd)) |
3003 | .int JMPTBL (L(fwd_write_6bytes), L(table_144_bytes_fwd)) |
3004 | .int JMPTBL (L(fwd_write_7bytes), L(table_144_bytes_fwd)) |
3005 | .int JMPTBL (L(fwd_write_8bytes), L(table_144_bytes_fwd)) |
3006 | .int JMPTBL (L(fwd_write_9bytes), L(table_144_bytes_fwd)) |
3007 | .int JMPTBL (L(fwd_write_10bytes), L(table_144_bytes_fwd)) |
3008 | .int JMPTBL (L(fwd_write_11bytes), L(table_144_bytes_fwd)) |
3009 | .int JMPTBL (L(fwd_write_12bytes), L(table_144_bytes_fwd)) |
3010 | .int JMPTBL (L(fwd_write_13bytes), L(table_144_bytes_fwd)) |
3011 | .int JMPTBL (L(fwd_write_14bytes), L(table_144_bytes_fwd)) |
3012 | .int JMPTBL (L(fwd_write_15bytes), L(table_144_bytes_fwd)) |
3013 | .int JMPTBL (L(fwd_write_16bytes), L(table_144_bytes_fwd)) |
3014 | .int JMPTBL (L(fwd_write_17bytes), L(table_144_bytes_fwd)) |
3015 | .int JMPTBL (L(fwd_write_18bytes), L(table_144_bytes_fwd)) |
3016 | .int JMPTBL (L(fwd_write_19bytes), L(table_144_bytes_fwd)) |
3017 | .int JMPTBL (L(fwd_write_20bytes), L(table_144_bytes_fwd)) |
3018 | .int JMPTBL (L(fwd_write_21bytes), L(table_144_bytes_fwd)) |
3019 | .int JMPTBL (L(fwd_write_22bytes), L(table_144_bytes_fwd)) |
3020 | .int JMPTBL (L(fwd_write_23bytes), L(table_144_bytes_fwd)) |
3021 | .int JMPTBL (L(fwd_write_24bytes), L(table_144_bytes_fwd)) |
3022 | .int JMPTBL (L(fwd_write_25bytes), L(table_144_bytes_fwd)) |
3023 | .int JMPTBL (L(fwd_write_26bytes), L(table_144_bytes_fwd)) |
3024 | .int JMPTBL (L(fwd_write_27bytes), L(table_144_bytes_fwd)) |
3025 | .int JMPTBL (L(fwd_write_28bytes), L(table_144_bytes_fwd)) |
3026 | .int JMPTBL (L(fwd_write_29bytes), L(table_144_bytes_fwd)) |
3027 | .int JMPTBL (L(fwd_write_30bytes), L(table_144_bytes_fwd)) |
3028 | .int JMPTBL (L(fwd_write_31bytes), L(table_144_bytes_fwd)) |
3029 | .int JMPTBL (L(fwd_write_32bytes), L(table_144_bytes_fwd)) |
3030 | .int JMPTBL (L(fwd_write_33bytes), L(table_144_bytes_fwd)) |
3031 | .int JMPTBL (L(fwd_write_34bytes), L(table_144_bytes_fwd)) |
3032 | .int JMPTBL (L(fwd_write_35bytes), L(table_144_bytes_fwd)) |
3033 | .int JMPTBL (L(fwd_write_36bytes), L(table_144_bytes_fwd)) |
3034 | .int JMPTBL (L(fwd_write_37bytes), L(table_144_bytes_fwd)) |
3035 | .int JMPTBL (L(fwd_write_38bytes), L(table_144_bytes_fwd)) |
3036 | .int JMPTBL (L(fwd_write_39bytes), L(table_144_bytes_fwd)) |
3037 | .int JMPTBL (L(fwd_write_40bytes), L(table_144_bytes_fwd)) |
3038 | .int JMPTBL (L(fwd_write_41bytes), L(table_144_bytes_fwd)) |
3039 | .int JMPTBL (L(fwd_write_42bytes), L(table_144_bytes_fwd)) |
3040 | .int JMPTBL (L(fwd_write_43bytes), L(table_144_bytes_fwd)) |
3041 | .int JMPTBL (L(fwd_write_44bytes), L(table_144_bytes_fwd)) |
3042 | .int JMPTBL (L(fwd_write_45bytes), L(table_144_bytes_fwd)) |
3043 | .int JMPTBL (L(fwd_write_46bytes), L(table_144_bytes_fwd)) |
3044 | .int JMPTBL (L(fwd_write_47bytes), L(table_144_bytes_fwd)) |
3045 | .int JMPTBL (L(fwd_write_48bytes), L(table_144_bytes_fwd)) |
3046 | .int JMPTBL (L(fwd_write_49bytes), L(table_144_bytes_fwd)) |
3047 | .int JMPTBL (L(fwd_write_50bytes), L(table_144_bytes_fwd)) |
3048 | .int JMPTBL (L(fwd_write_51bytes), L(table_144_bytes_fwd)) |
3049 | .int JMPTBL (L(fwd_write_52bytes), L(table_144_bytes_fwd)) |
3050 | .int JMPTBL (L(fwd_write_53bytes), L(table_144_bytes_fwd)) |
3051 | .int JMPTBL (L(fwd_write_54bytes), L(table_144_bytes_fwd)) |
3052 | .int JMPTBL (L(fwd_write_55bytes), L(table_144_bytes_fwd)) |
3053 | .int JMPTBL (L(fwd_write_56bytes), L(table_144_bytes_fwd)) |
3054 | .int JMPTBL (L(fwd_write_57bytes), L(table_144_bytes_fwd)) |
3055 | .int JMPTBL (L(fwd_write_58bytes), L(table_144_bytes_fwd)) |
3056 | .int JMPTBL (L(fwd_write_59bytes), L(table_144_bytes_fwd)) |
3057 | .int JMPTBL (L(fwd_write_60bytes), L(table_144_bytes_fwd)) |
3058 | .int JMPTBL (L(fwd_write_61bytes), L(table_144_bytes_fwd)) |
3059 | .int JMPTBL (L(fwd_write_62bytes), L(table_144_bytes_fwd)) |
3060 | .int JMPTBL (L(fwd_write_63bytes), L(table_144_bytes_fwd)) |
3061 | .int JMPTBL (L(fwd_write_64bytes), L(table_144_bytes_fwd)) |
3062 | .int JMPTBL (L(fwd_write_65bytes), L(table_144_bytes_fwd)) |
3063 | .int JMPTBL (L(fwd_write_66bytes), L(table_144_bytes_fwd)) |
3064 | .int JMPTBL (L(fwd_write_67bytes), L(table_144_bytes_fwd)) |
3065 | .int JMPTBL (L(fwd_write_68bytes), L(table_144_bytes_fwd)) |
3066 | .int JMPTBL (L(fwd_write_69bytes), L(table_144_bytes_fwd)) |
3067 | .int JMPTBL (L(fwd_write_70bytes), L(table_144_bytes_fwd)) |
3068 | .int JMPTBL (L(fwd_write_71bytes), L(table_144_bytes_fwd)) |
3069 | .int JMPTBL (L(fwd_write_72bytes), L(table_144_bytes_fwd)) |
3070 | .int JMPTBL (L(fwd_write_73bytes), L(table_144_bytes_fwd)) |
3071 | .int JMPTBL (L(fwd_write_74bytes), L(table_144_bytes_fwd)) |
3072 | .int JMPTBL (L(fwd_write_75bytes), L(table_144_bytes_fwd)) |
3073 | .int JMPTBL (L(fwd_write_76bytes), L(table_144_bytes_fwd)) |
3074 | .int JMPTBL (L(fwd_write_77bytes), L(table_144_bytes_fwd)) |
3075 | .int JMPTBL (L(fwd_write_78bytes), L(table_144_bytes_fwd)) |
3076 | .int JMPTBL (L(fwd_write_79bytes), L(table_144_bytes_fwd)) |
3077 | .int JMPTBL (L(fwd_write_80bytes), L(table_144_bytes_fwd)) |
3078 | .int JMPTBL (L(fwd_write_81bytes), L(table_144_bytes_fwd)) |
3079 | .int JMPTBL (L(fwd_write_82bytes), L(table_144_bytes_fwd)) |
3080 | .int JMPTBL (L(fwd_write_83bytes), L(table_144_bytes_fwd)) |
3081 | .int JMPTBL (L(fwd_write_84bytes), L(table_144_bytes_fwd)) |
3082 | .int JMPTBL (L(fwd_write_85bytes), L(table_144_bytes_fwd)) |
3083 | .int JMPTBL (L(fwd_write_86bytes), L(table_144_bytes_fwd)) |
3084 | .int JMPTBL (L(fwd_write_87bytes), L(table_144_bytes_fwd)) |
3085 | .int JMPTBL (L(fwd_write_88bytes), L(table_144_bytes_fwd)) |
3086 | .int JMPTBL (L(fwd_write_89bytes), L(table_144_bytes_fwd)) |
3087 | .int JMPTBL (L(fwd_write_90bytes), L(table_144_bytes_fwd)) |
3088 | .int JMPTBL (L(fwd_write_91bytes), L(table_144_bytes_fwd)) |
3089 | .int JMPTBL (L(fwd_write_92bytes), L(table_144_bytes_fwd)) |
3090 | .int JMPTBL (L(fwd_write_93bytes), L(table_144_bytes_fwd)) |
3091 | .int JMPTBL (L(fwd_write_94bytes), L(table_144_bytes_fwd)) |
3092 | .int JMPTBL (L(fwd_write_95bytes), L(table_144_bytes_fwd)) |
3093 | .int JMPTBL (L(fwd_write_96bytes), L(table_144_bytes_fwd)) |
3094 | .int JMPTBL (L(fwd_write_97bytes), L(table_144_bytes_fwd)) |
3095 | .int JMPTBL (L(fwd_write_98bytes), L(table_144_bytes_fwd)) |
3096 | .int JMPTBL (L(fwd_write_99bytes), L(table_144_bytes_fwd)) |
3097 | .int JMPTBL (L(fwd_write_100bytes), L(table_144_bytes_fwd)) |
3098 | .int JMPTBL (L(fwd_write_101bytes), L(table_144_bytes_fwd)) |
3099 | .int JMPTBL (L(fwd_write_102bytes), L(table_144_bytes_fwd)) |
3100 | .int JMPTBL (L(fwd_write_103bytes), L(table_144_bytes_fwd)) |
3101 | .int JMPTBL (L(fwd_write_104bytes), L(table_144_bytes_fwd)) |
3102 | .int JMPTBL (L(fwd_write_105bytes), L(table_144_bytes_fwd)) |
3103 | .int JMPTBL (L(fwd_write_106bytes), L(table_144_bytes_fwd)) |
3104 | .int JMPTBL (L(fwd_write_107bytes), L(table_144_bytes_fwd)) |
3105 | .int JMPTBL (L(fwd_write_108bytes), L(table_144_bytes_fwd)) |
3106 | .int JMPTBL (L(fwd_write_109bytes), L(table_144_bytes_fwd)) |
3107 | .int JMPTBL (L(fwd_write_110bytes), L(table_144_bytes_fwd)) |
3108 | .int JMPTBL (L(fwd_write_111bytes), L(table_144_bytes_fwd)) |
3109 | .int JMPTBL (L(fwd_write_112bytes), L(table_144_bytes_fwd)) |
3110 | .int JMPTBL (L(fwd_write_113bytes), L(table_144_bytes_fwd)) |
3111 | .int JMPTBL (L(fwd_write_114bytes), L(table_144_bytes_fwd)) |
3112 | .int JMPTBL (L(fwd_write_115bytes), L(table_144_bytes_fwd)) |
3113 | .int JMPTBL (L(fwd_write_116bytes), L(table_144_bytes_fwd)) |
3114 | .int JMPTBL (L(fwd_write_117bytes), L(table_144_bytes_fwd)) |
3115 | .int JMPTBL (L(fwd_write_118bytes), L(table_144_bytes_fwd)) |
3116 | .int JMPTBL (L(fwd_write_119bytes), L(table_144_bytes_fwd)) |
3117 | .int JMPTBL (L(fwd_write_120bytes), L(table_144_bytes_fwd)) |
3118 | .int JMPTBL (L(fwd_write_121bytes), L(table_144_bytes_fwd)) |
3119 | .int JMPTBL (L(fwd_write_122bytes), L(table_144_bytes_fwd)) |
3120 | .int JMPTBL (L(fwd_write_123bytes), L(table_144_bytes_fwd)) |
3121 | .int JMPTBL (L(fwd_write_124bytes), L(table_144_bytes_fwd)) |
3122 | .int JMPTBL (L(fwd_write_125bytes), L(table_144_bytes_fwd)) |
3123 | .int JMPTBL (L(fwd_write_126bytes), L(table_144_bytes_fwd)) |
3124 | .int JMPTBL (L(fwd_write_127bytes), L(table_144_bytes_fwd)) |
3125 | .int JMPTBL (L(fwd_write_128bytes), L(table_144_bytes_fwd)) |
3126 | .int JMPTBL (L(fwd_write_129bytes), L(table_144_bytes_fwd)) |
3127 | .int JMPTBL (L(fwd_write_130bytes), L(table_144_bytes_fwd)) |
3128 | .int JMPTBL (L(fwd_write_131bytes), L(table_144_bytes_fwd)) |
3129 | .int JMPTBL (L(fwd_write_132bytes), L(table_144_bytes_fwd)) |
3130 | .int JMPTBL (L(fwd_write_133bytes), L(table_144_bytes_fwd)) |
3131 | .int JMPTBL (L(fwd_write_134bytes), L(table_144_bytes_fwd)) |
3132 | .int JMPTBL (L(fwd_write_135bytes), L(table_144_bytes_fwd)) |
3133 | .int JMPTBL (L(fwd_write_136bytes), L(table_144_bytes_fwd)) |
3134 | .int JMPTBL (L(fwd_write_137bytes), L(table_144_bytes_fwd)) |
3135 | .int JMPTBL (L(fwd_write_138bytes), L(table_144_bytes_fwd)) |
3136 | .int JMPTBL (L(fwd_write_139bytes), L(table_144_bytes_fwd)) |
3137 | .int JMPTBL (L(fwd_write_140bytes), L(table_144_bytes_fwd)) |
3138 | .int JMPTBL (L(fwd_write_141bytes), L(table_144_bytes_fwd)) |
3139 | .int JMPTBL (L(fwd_write_142bytes), L(table_144_bytes_fwd)) |
3140 | .int JMPTBL (L(fwd_write_143bytes), L(table_144_bytes_fwd)) |
3141 | |
3142 | .p2align 3 |
3143 | L(shl_table_fwd): |
3144 | .int JMPTBL (L(shl_0), L(shl_table_fwd)) |
3145 | .int JMPTBL (L(shl_1), L(shl_table_fwd)) |
3146 | .int JMPTBL (L(shl_2), L(shl_table_fwd)) |
3147 | .int JMPTBL (L(shl_3), L(shl_table_fwd)) |
3148 | .int JMPTBL (L(shl_4), L(shl_table_fwd)) |
3149 | .int JMPTBL (L(shl_5), L(shl_table_fwd)) |
3150 | .int JMPTBL (L(shl_6), L(shl_table_fwd)) |
3151 | .int JMPTBL (L(shl_7), L(shl_table_fwd)) |
3152 | .int JMPTBL (L(shl_8), L(shl_table_fwd)) |
3153 | .int JMPTBL (L(shl_9), L(shl_table_fwd)) |
3154 | .int JMPTBL (L(shl_10), L(shl_table_fwd)) |
3155 | .int JMPTBL (L(shl_11), L(shl_table_fwd)) |
3156 | .int JMPTBL (L(shl_12), L(shl_table_fwd)) |
3157 | .int JMPTBL (L(shl_13), L(shl_table_fwd)) |
3158 | .int JMPTBL (L(shl_14), L(shl_table_fwd)) |
3159 | .int JMPTBL (L(shl_15), L(shl_table_fwd)) |
3160 | |
3161 | .p2align 3 |
3162 | L(shl_table_bwd): |
3163 | .int JMPTBL (L(shl_0_bwd), L(shl_table_bwd)) |
3164 | .int JMPTBL (L(shl_1_bwd), L(shl_table_bwd)) |
3165 | .int JMPTBL (L(shl_2_bwd), L(shl_table_bwd)) |
3166 | .int JMPTBL (L(shl_3_bwd), L(shl_table_bwd)) |
3167 | .int JMPTBL (L(shl_4_bwd), L(shl_table_bwd)) |
3168 | .int JMPTBL (L(shl_5_bwd), L(shl_table_bwd)) |
3169 | .int JMPTBL (L(shl_6_bwd), L(shl_table_bwd)) |
3170 | .int JMPTBL (L(shl_7_bwd), L(shl_table_bwd)) |
3171 | .int JMPTBL (L(shl_8_bwd), L(shl_table_bwd)) |
3172 | .int JMPTBL (L(shl_9_bwd), L(shl_table_bwd)) |
3173 | .int JMPTBL (L(shl_10_bwd), L(shl_table_bwd)) |
3174 | .int JMPTBL (L(shl_11_bwd), L(shl_table_bwd)) |
3175 | .int JMPTBL (L(shl_12_bwd), L(shl_table_bwd)) |
3176 | .int JMPTBL (L(shl_13_bwd), L(shl_table_bwd)) |
3177 | .int JMPTBL (L(shl_14_bwd), L(shl_table_bwd)) |
3178 | .int JMPTBL (L(shl_15_bwd), L(shl_table_bwd)) |
3179 | |
3180 | #endif |
3181 | |