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 | cmp %RDX_LP, %RCX_LP |
52 | jb HIDDEN_JUMPTARGET (__chk_fail) |
53 | END (MEMPCPY_CHK) |
54 | |
55 | ENTRY (MEMPCPY) |
56 | mov %RDI_LP, %RAX_LP |
57 | add %RDX_LP, %RAX_LP |
58 | jmp L(start) |
59 | END (MEMPCPY) |
60 | #endif |
61 | |
62 | #if !defined USE_AS_BCOPY |
63 | ENTRY (MEMCPY_CHK) |
64 | cmp %RDX_LP, %RCX_LP |
65 | jb HIDDEN_JUMPTARGET (__chk_fail) |
66 | END (MEMCPY_CHK) |
67 | #endif |
68 | |
69 | ENTRY (MEMCPY) |
70 | mov %RDI_LP, %RAX_LP |
71 | #ifdef USE_AS_MEMPCPY |
72 | add %RDX_LP, %RAX_LP |
73 | #endif |
74 | |
75 | #ifdef __ILP32__ |
76 | /* Clear the upper 32 bits. */ |
77 | mov %edx, %edx |
78 | #endif |
79 | |
80 | #ifdef USE_AS_MEMMOVE |
81 | cmp %rsi, %rdi |
82 | jb L(copy_forward) |
83 | je L(bwd_write_0bytes) |
84 | cmp $144, %rdx |
85 | jae L(copy_backward) |
86 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
87 | L(copy_forward): |
88 | #endif |
89 | L(start): |
90 | cmp $144, %rdx |
91 | jae L(144bytesormore) |
92 | |
93 | L(fwd_write_less32bytes): |
94 | #ifndef USE_AS_MEMMOVE |
95 | cmp %dil, %sil |
96 | jbe L(bk_write) |
97 | #endif |
98 | add %rdx, %rsi |
99 | add %rdx, %rdi |
100 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
101 | #ifndef USE_AS_MEMMOVE |
102 | L(bk_write): |
103 | |
104 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
105 | #endif |
106 | |
107 | .p2align 4 |
108 | L(144bytesormore): |
109 | |
110 | #ifndef USE_AS_MEMMOVE |
111 | cmp %dil, %sil |
112 | jle L(copy_backward) |
113 | #endif |
114 | movdqu (%rsi), %xmm0 |
115 | mov %rdi, %r8 |
116 | and $-16, %rdi |
117 | add $16, %rdi |
118 | mov %rdi, %r9 |
119 | sub %r8, %r9 |
120 | sub %r9, %rdx |
121 | add %r9, %rsi |
122 | mov %rsi, %r9 |
123 | and $0xf, %r9 |
124 | jz L(shl_0) |
125 | #ifdef DATA_CACHE_SIZE |
126 | mov $DATA_CACHE_SIZE, %RCX_LP |
127 | #else |
128 | mov __x86_data_cache_size(%rip), %RCX_LP |
129 | #endif |
130 | cmp %rcx, %rdx |
131 | jae L(gobble_mem_fwd) |
132 | lea L(shl_table_fwd)(%rip), %r11 |
133 | sub $0x80, %rdx |
134 | movslq (%r11, %r9, 4), %r9 |
135 | add %r11, %r9 |
136 | jmp *%r9 |
137 | ud2 |
138 | |
139 | .p2align 4 |
140 | L(copy_backward): |
141 | #ifdef DATA_CACHE_SIZE |
142 | mov $DATA_CACHE_SIZE, %RCX_LP |
143 | #else |
144 | mov __x86_data_cache_size(%rip), %RCX_LP |
145 | #endif |
146 | shl $1, %rcx |
147 | cmp %rcx, %rdx |
148 | ja L(gobble_mem_bwd) |
149 | |
150 | add %rdx, %rdi |
151 | add %rdx, %rsi |
152 | movdqu -16(%rsi), %xmm0 |
153 | lea -16(%rdi), %r8 |
154 | mov %rdi, %r9 |
155 | and $0xf, %r9 |
156 | xor %r9, %rdi |
157 | sub %r9, %rsi |
158 | sub %r9, %rdx |
159 | mov %rsi, %r9 |
160 | and $0xf, %r9 |
161 | jz L(shl_0_bwd) |
162 | lea L(shl_table_bwd)(%rip), %r11 |
163 | sub $0x80, %rdx |
164 | movslq (%r11, %r9, 4), %r9 |
165 | add %r11, %r9 |
166 | jmp *%r9 |
167 | ud2 |
168 | |
169 | .p2align 4 |
170 | L(shl_0): |
171 | |
172 | mov %rdx, %r9 |
173 | shr $8, %r9 |
174 | add %rdx, %r9 |
175 | #ifdef DATA_CACHE_SIZE |
176 | cmp $DATA_CACHE_SIZE_HALF, %R9_LP |
177 | #else |
178 | cmp __x86_data_cache_size_half(%rip), %R9_LP |
179 | #endif |
180 | jae L(gobble_mem_fwd) |
181 | sub $0x80, %rdx |
182 | .p2align 4 |
183 | L(shl_0_loop): |
184 | movdqa (%rsi), %xmm1 |
185 | movdqa %xmm1, (%rdi) |
186 | movaps 0x10(%rsi), %xmm2 |
187 | movaps %xmm2, 0x10(%rdi) |
188 | movaps 0x20(%rsi), %xmm3 |
189 | movaps %xmm3, 0x20(%rdi) |
190 | movaps 0x30(%rsi), %xmm4 |
191 | movaps %xmm4, 0x30(%rdi) |
192 | movaps 0x40(%rsi), %xmm1 |
193 | movaps %xmm1, 0x40(%rdi) |
194 | movaps 0x50(%rsi), %xmm2 |
195 | movaps %xmm2, 0x50(%rdi) |
196 | movaps 0x60(%rsi), %xmm3 |
197 | movaps %xmm3, 0x60(%rdi) |
198 | movaps 0x70(%rsi), %xmm4 |
199 | movaps %xmm4, 0x70(%rdi) |
200 | sub $0x80, %rdx |
201 | lea 0x80(%rsi), %rsi |
202 | lea 0x80(%rdi), %rdi |
203 | jae L(shl_0_loop) |
204 | movdqu %xmm0, (%r8) |
205 | add $0x80, %rdx |
206 | add %rdx, %rsi |
207 | add %rdx, %rdi |
208 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
209 | |
210 | .p2align 4 |
211 | L(shl_0_bwd): |
212 | sub $0x80, %rdx |
213 | L(copy_backward_loop): |
214 | movaps -0x10(%rsi), %xmm1 |
215 | movaps %xmm1, -0x10(%rdi) |
216 | movaps -0x20(%rsi), %xmm2 |
217 | movaps %xmm2, -0x20(%rdi) |
218 | movaps -0x30(%rsi), %xmm3 |
219 | movaps %xmm3, -0x30(%rdi) |
220 | movaps -0x40(%rsi), %xmm4 |
221 | movaps %xmm4, -0x40(%rdi) |
222 | movaps -0x50(%rsi), %xmm5 |
223 | movaps %xmm5, -0x50(%rdi) |
224 | movaps -0x60(%rsi), %xmm5 |
225 | movaps %xmm5, -0x60(%rdi) |
226 | movaps -0x70(%rsi), %xmm5 |
227 | movaps %xmm5, -0x70(%rdi) |
228 | movaps -0x80(%rsi), %xmm5 |
229 | movaps %xmm5, -0x80(%rdi) |
230 | sub $0x80, %rdx |
231 | lea -0x80(%rdi), %rdi |
232 | lea -0x80(%rsi), %rsi |
233 | jae L(copy_backward_loop) |
234 | |
235 | movdqu %xmm0, (%r8) |
236 | add $0x80, %rdx |
237 | sub %rdx, %rdi |
238 | sub %rdx, %rsi |
239 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
240 | |
241 | .p2align 4 |
242 | L(shl_1): |
243 | sub $0x80, %rdx |
244 | movaps -0x01(%rsi), %xmm1 |
245 | movaps 0x0f(%rsi), %xmm2 |
246 | movaps 0x1f(%rsi), %xmm3 |
247 | movaps 0x2f(%rsi), %xmm4 |
248 | movaps 0x3f(%rsi), %xmm5 |
249 | movaps 0x4f(%rsi), %xmm6 |
250 | movaps 0x5f(%rsi), %xmm7 |
251 | movaps 0x6f(%rsi), %xmm8 |
252 | movaps 0x7f(%rsi), %xmm9 |
253 | lea 0x80(%rsi), %rsi |
254 | palignr $1, %xmm8, %xmm9 |
255 | movaps %xmm9, 0x70(%rdi) |
256 | palignr $1, %xmm7, %xmm8 |
257 | movaps %xmm8, 0x60(%rdi) |
258 | palignr $1, %xmm6, %xmm7 |
259 | movaps %xmm7, 0x50(%rdi) |
260 | palignr $1, %xmm5, %xmm6 |
261 | movaps %xmm6, 0x40(%rdi) |
262 | palignr $1, %xmm4, %xmm5 |
263 | movaps %xmm5, 0x30(%rdi) |
264 | palignr $1, %xmm3, %xmm4 |
265 | movaps %xmm4, 0x20(%rdi) |
266 | palignr $1, %xmm2, %xmm3 |
267 | movaps %xmm3, 0x10(%rdi) |
268 | palignr $1, %xmm1, %xmm2 |
269 | movaps %xmm2, (%rdi) |
270 | lea 0x80(%rdi), %rdi |
271 | jae L(shl_1) |
272 | movdqu %xmm0, (%r8) |
273 | add $0x80, %rdx |
274 | add %rdx, %rdi |
275 | add %rdx, %rsi |
276 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
277 | |
278 | .p2align 4 |
279 | L(shl_1_bwd): |
280 | movaps -0x01(%rsi), %xmm1 |
281 | |
282 | movaps -0x11(%rsi), %xmm2 |
283 | palignr $1, %xmm2, %xmm1 |
284 | movaps %xmm1, -0x10(%rdi) |
285 | |
286 | movaps -0x21(%rsi), %xmm3 |
287 | palignr $1, %xmm3, %xmm2 |
288 | movaps %xmm2, -0x20(%rdi) |
289 | |
290 | movaps -0x31(%rsi), %xmm4 |
291 | palignr $1, %xmm4, %xmm3 |
292 | movaps %xmm3, -0x30(%rdi) |
293 | |
294 | movaps -0x41(%rsi), %xmm5 |
295 | palignr $1, %xmm5, %xmm4 |
296 | movaps %xmm4, -0x40(%rdi) |
297 | |
298 | movaps -0x51(%rsi), %xmm6 |
299 | palignr $1, %xmm6, %xmm5 |
300 | movaps %xmm5, -0x50(%rdi) |
301 | |
302 | movaps -0x61(%rsi), %xmm7 |
303 | palignr $1, %xmm7, %xmm6 |
304 | movaps %xmm6, -0x60(%rdi) |
305 | |
306 | movaps -0x71(%rsi), %xmm8 |
307 | palignr $1, %xmm8, %xmm7 |
308 | movaps %xmm7, -0x70(%rdi) |
309 | |
310 | movaps -0x81(%rsi), %xmm9 |
311 | palignr $1, %xmm9, %xmm8 |
312 | movaps %xmm8, -0x80(%rdi) |
313 | |
314 | sub $0x80, %rdx |
315 | lea -0x80(%rdi), %rdi |
316 | lea -0x80(%rsi), %rsi |
317 | jae L(shl_1_bwd) |
318 | movdqu %xmm0, (%r8) |
319 | add $0x80, %rdx |
320 | sub %rdx, %rdi |
321 | sub %rdx, %rsi |
322 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
323 | |
324 | .p2align 4 |
325 | L(shl_2): |
326 | sub $0x80, %rdx |
327 | movaps -0x02(%rsi), %xmm1 |
328 | movaps 0x0e(%rsi), %xmm2 |
329 | movaps 0x1e(%rsi), %xmm3 |
330 | movaps 0x2e(%rsi), %xmm4 |
331 | movaps 0x3e(%rsi), %xmm5 |
332 | movaps 0x4e(%rsi), %xmm6 |
333 | movaps 0x5e(%rsi), %xmm7 |
334 | movaps 0x6e(%rsi), %xmm8 |
335 | movaps 0x7e(%rsi), %xmm9 |
336 | lea 0x80(%rsi), %rsi |
337 | palignr $2, %xmm8, %xmm9 |
338 | movaps %xmm9, 0x70(%rdi) |
339 | palignr $2, %xmm7, %xmm8 |
340 | movaps %xmm8, 0x60(%rdi) |
341 | palignr $2, %xmm6, %xmm7 |
342 | movaps %xmm7, 0x50(%rdi) |
343 | palignr $2, %xmm5, %xmm6 |
344 | movaps %xmm6, 0x40(%rdi) |
345 | palignr $2, %xmm4, %xmm5 |
346 | movaps %xmm5, 0x30(%rdi) |
347 | palignr $2, %xmm3, %xmm4 |
348 | movaps %xmm4, 0x20(%rdi) |
349 | palignr $2, %xmm2, %xmm3 |
350 | movaps %xmm3, 0x10(%rdi) |
351 | palignr $2, %xmm1, %xmm2 |
352 | movaps %xmm2, (%rdi) |
353 | lea 0x80(%rdi), %rdi |
354 | jae L(shl_2) |
355 | movdqu %xmm0, (%r8) |
356 | add $0x80, %rdx |
357 | add %rdx, %rdi |
358 | add %rdx, %rsi |
359 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
360 | |
361 | .p2align 4 |
362 | L(shl_2_bwd): |
363 | movaps -0x02(%rsi), %xmm1 |
364 | |
365 | movaps -0x12(%rsi), %xmm2 |
366 | palignr $2, %xmm2, %xmm1 |
367 | movaps %xmm1, -0x10(%rdi) |
368 | |
369 | movaps -0x22(%rsi), %xmm3 |
370 | palignr $2, %xmm3, %xmm2 |
371 | movaps %xmm2, -0x20(%rdi) |
372 | |
373 | movaps -0x32(%rsi), %xmm4 |
374 | palignr $2, %xmm4, %xmm3 |
375 | movaps %xmm3, -0x30(%rdi) |
376 | |
377 | movaps -0x42(%rsi), %xmm5 |
378 | palignr $2, %xmm5, %xmm4 |
379 | movaps %xmm4, -0x40(%rdi) |
380 | |
381 | movaps -0x52(%rsi), %xmm6 |
382 | palignr $2, %xmm6, %xmm5 |
383 | movaps %xmm5, -0x50(%rdi) |
384 | |
385 | movaps -0x62(%rsi), %xmm7 |
386 | palignr $2, %xmm7, %xmm6 |
387 | movaps %xmm6, -0x60(%rdi) |
388 | |
389 | movaps -0x72(%rsi), %xmm8 |
390 | palignr $2, %xmm8, %xmm7 |
391 | movaps %xmm7, -0x70(%rdi) |
392 | |
393 | movaps -0x82(%rsi), %xmm9 |
394 | palignr $2, %xmm9, %xmm8 |
395 | movaps %xmm8, -0x80(%rdi) |
396 | |
397 | sub $0x80, %rdx |
398 | lea -0x80(%rdi), %rdi |
399 | lea -0x80(%rsi), %rsi |
400 | jae L(shl_2_bwd) |
401 | movdqu %xmm0, (%r8) |
402 | add $0x80, %rdx |
403 | sub %rdx, %rdi |
404 | sub %rdx, %rsi |
405 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
406 | |
407 | .p2align 4 |
408 | L(shl_3): |
409 | sub $0x80, %rdx |
410 | movaps -0x03(%rsi), %xmm1 |
411 | movaps 0x0d(%rsi), %xmm2 |
412 | movaps 0x1d(%rsi), %xmm3 |
413 | movaps 0x2d(%rsi), %xmm4 |
414 | movaps 0x3d(%rsi), %xmm5 |
415 | movaps 0x4d(%rsi), %xmm6 |
416 | movaps 0x5d(%rsi), %xmm7 |
417 | movaps 0x6d(%rsi), %xmm8 |
418 | movaps 0x7d(%rsi), %xmm9 |
419 | lea 0x80(%rsi), %rsi |
420 | palignr $3, %xmm8, %xmm9 |
421 | movaps %xmm9, 0x70(%rdi) |
422 | palignr $3, %xmm7, %xmm8 |
423 | movaps %xmm8, 0x60(%rdi) |
424 | palignr $3, %xmm6, %xmm7 |
425 | movaps %xmm7, 0x50(%rdi) |
426 | palignr $3, %xmm5, %xmm6 |
427 | movaps %xmm6, 0x40(%rdi) |
428 | palignr $3, %xmm4, %xmm5 |
429 | movaps %xmm5, 0x30(%rdi) |
430 | palignr $3, %xmm3, %xmm4 |
431 | movaps %xmm4, 0x20(%rdi) |
432 | palignr $3, %xmm2, %xmm3 |
433 | movaps %xmm3, 0x10(%rdi) |
434 | palignr $3, %xmm1, %xmm2 |
435 | movaps %xmm2, (%rdi) |
436 | lea 0x80(%rdi), %rdi |
437 | jae L(shl_3) |
438 | movdqu %xmm0, (%r8) |
439 | add $0x80, %rdx |
440 | add %rdx, %rdi |
441 | add %rdx, %rsi |
442 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
443 | |
444 | .p2align 4 |
445 | L(shl_3_bwd): |
446 | movaps -0x03(%rsi), %xmm1 |
447 | |
448 | movaps -0x13(%rsi), %xmm2 |
449 | palignr $3, %xmm2, %xmm1 |
450 | movaps %xmm1, -0x10(%rdi) |
451 | |
452 | movaps -0x23(%rsi), %xmm3 |
453 | palignr $3, %xmm3, %xmm2 |
454 | movaps %xmm2, -0x20(%rdi) |
455 | |
456 | movaps -0x33(%rsi), %xmm4 |
457 | palignr $3, %xmm4, %xmm3 |
458 | movaps %xmm3, -0x30(%rdi) |
459 | |
460 | movaps -0x43(%rsi), %xmm5 |
461 | palignr $3, %xmm5, %xmm4 |
462 | movaps %xmm4, -0x40(%rdi) |
463 | |
464 | movaps -0x53(%rsi), %xmm6 |
465 | palignr $3, %xmm6, %xmm5 |
466 | movaps %xmm5, -0x50(%rdi) |
467 | |
468 | movaps -0x63(%rsi), %xmm7 |
469 | palignr $3, %xmm7, %xmm6 |
470 | movaps %xmm6, -0x60(%rdi) |
471 | |
472 | movaps -0x73(%rsi), %xmm8 |
473 | palignr $3, %xmm8, %xmm7 |
474 | movaps %xmm7, -0x70(%rdi) |
475 | |
476 | movaps -0x83(%rsi), %xmm9 |
477 | palignr $3, %xmm9, %xmm8 |
478 | movaps %xmm8, -0x80(%rdi) |
479 | |
480 | sub $0x80, %rdx |
481 | lea -0x80(%rdi), %rdi |
482 | lea -0x80(%rsi), %rsi |
483 | jae L(shl_3_bwd) |
484 | movdqu %xmm0, (%r8) |
485 | add $0x80, %rdx |
486 | sub %rdx, %rdi |
487 | sub %rdx, %rsi |
488 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
489 | |
490 | .p2align 4 |
491 | L(shl_4): |
492 | sub $0x80, %rdx |
493 | movaps -0x04(%rsi), %xmm1 |
494 | movaps 0x0c(%rsi), %xmm2 |
495 | movaps 0x1c(%rsi), %xmm3 |
496 | movaps 0x2c(%rsi), %xmm4 |
497 | movaps 0x3c(%rsi), %xmm5 |
498 | movaps 0x4c(%rsi), %xmm6 |
499 | movaps 0x5c(%rsi), %xmm7 |
500 | movaps 0x6c(%rsi), %xmm8 |
501 | movaps 0x7c(%rsi), %xmm9 |
502 | lea 0x80(%rsi), %rsi |
503 | palignr $4, %xmm8, %xmm9 |
504 | movaps %xmm9, 0x70(%rdi) |
505 | palignr $4, %xmm7, %xmm8 |
506 | movaps %xmm8, 0x60(%rdi) |
507 | palignr $4, %xmm6, %xmm7 |
508 | movaps %xmm7, 0x50(%rdi) |
509 | palignr $4, %xmm5, %xmm6 |
510 | movaps %xmm6, 0x40(%rdi) |
511 | palignr $4, %xmm4, %xmm5 |
512 | movaps %xmm5, 0x30(%rdi) |
513 | palignr $4, %xmm3, %xmm4 |
514 | movaps %xmm4, 0x20(%rdi) |
515 | palignr $4, %xmm2, %xmm3 |
516 | movaps %xmm3, 0x10(%rdi) |
517 | palignr $4, %xmm1, %xmm2 |
518 | movaps %xmm2, (%rdi) |
519 | lea 0x80(%rdi), %rdi |
520 | jae L(shl_4) |
521 | movdqu %xmm0, (%r8) |
522 | add $0x80, %rdx |
523 | add %rdx, %rdi |
524 | add %rdx, %rsi |
525 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
526 | |
527 | .p2align 4 |
528 | L(shl_4_bwd): |
529 | movaps -0x04(%rsi), %xmm1 |
530 | |
531 | movaps -0x14(%rsi), %xmm2 |
532 | palignr $4, %xmm2, %xmm1 |
533 | movaps %xmm1, -0x10(%rdi) |
534 | |
535 | movaps -0x24(%rsi), %xmm3 |
536 | palignr $4, %xmm3, %xmm2 |
537 | movaps %xmm2, -0x20(%rdi) |
538 | |
539 | movaps -0x34(%rsi), %xmm4 |
540 | palignr $4, %xmm4, %xmm3 |
541 | movaps %xmm3, -0x30(%rdi) |
542 | |
543 | movaps -0x44(%rsi), %xmm5 |
544 | palignr $4, %xmm5, %xmm4 |
545 | movaps %xmm4, -0x40(%rdi) |
546 | |
547 | movaps -0x54(%rsi), %xmm6 |
548 | palignr $4, %xmm6, %xmm5 |
549 | movaps %xmm5, -0x50(%rdi) |
550 | |
551 | movaps -0x64(%rsi), %xmm7 |
552 | palignr $4, %xmm7, %xmm6 |
553 | movaps %xmm6, -0x60(%rdi) |
554 | |
555 | movaps -0x74(%rsi), %xmm8 |
556 | palignr $4, %xmm8, %xmm7 |
557 | movaps %xmm7, -0x70(%rdi) |
558 | |
559 | movaps -0x84(%rsi), %xmm9 |
560 | palignr $4, %xmm9, %xmm8 |
561 | movaps %xmm8, -0x80(%rdi) |
562 | |
563 | sub $0x80, %rdx |
564 | lea -0x80(%rdi), %rdi |
565 | lea -0x80(%rsi), %rsi |
566 | jae L(shl_4_bwd) |
567 | movdqu %xmm0, (%r8) |
568 | add $0x80, %rdx |
569 | sub %rdx, %rdi |
570 | sub %rdx, %rsi |
571 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
572 | |
573 | .p2align 4 |
574 | L(shl_5): |
575 | sub $0x80, %rdx |
576 | movaps -0x05(%rsi), %xmm1 |
577 | movaps 0x0b(%rsi), %xmm2 |
578 | movaps 0x1b(%rsi), %xmm3 |
579 | movaps 0x2b(%rsi), %xmm4 |
580 | movaps 0x3b(%rsi), %xmm5 |
581 | movaps 0x4b(%rsi), %xmm6 |
582 | movaps 0x5b(%rsi), %xmm7 |
583 | movaps 0x6b(%rsi), %xmm8 |
584 | movaps 0x7b(%rsi), %xmm9 |
585 | lea 0x80(%rsi), %rsi |
586 | palignr $5, %xmm8, %xmm9 |
587 | movaps %xmm9, 0x70(%rdi) |
588 | palignr $5, %xmm7, %xmm8 |
589 | movaps %xmm8, 0x60(%rdi) |
590 | palignr $5, %xmm6, %xmm7 |
591 | movaps %xmm7, 0x50(%rdi) |
592 | palignr $5, %xmm5, %xmm6 |
593 | movaps %xmm6, 0x40(%rdi) |
594 | palignr $5, %xmm4, %xmm5 |
595 | movaps %xmm5, 0x30(%rdi) |
596 | palignr $5, %xmm3, %xmm4 |
597 | movaps %xmm4, 0x20(%rdi) |
598 | palignr $5, %xmm2, %xmm3 |
599 | movaps %xmm3, 0x10(%rdi) |
600 | palignr $5, %xmm1, %xmm2 |
601 | movaps %xmm2, (%rdi) |
602 | lea 0x80(%rdi), %rdi |
603 | jae L(shl_5) |
604 | movdqu %xmm0, (%r8) |
605 | add $0x80, %rdx |
606 | add %rdx, %rdi |
607 | add %rdx, %rsi |
608 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
609 | |
610 | .p2align 4 |
611 | L(shl_5_bwd): |
612 | movaps -0x05(%rsi), %xmm1 |
613 | |
614 | movaps -0x15(%rsi), %xmm2 |
615 | palignr $5, %xmm2, %xmm1 |
616 | movaps %xmm1, -0x10(%rdi) |
617 | |
618 | movaps -0x25(%rsi), %xmm3 |
619 | palignr $5, %xmm3, %xmm2 |
620 | movaps %xmm2, -0x20(%rdi) |
621 | |
622 | movaps -0x35(%rsi), %xmm4 |
623 | palignr $5, %xmm4, %xmm3 |
624 | movaps %xmm3, -0x30(%rdi) |
625 | |
626 | movaps -0x45(%rsi), %xmm5 |
627 | palignr $5, %xmm5, %xmm4 |
628 | movaps %xmm4, -0x40(%rdi) |
629 | |
630 | movaps -0x55(%rsi), %xmm6 |
631 | palignr $5, %xmm6, %xmm5 |
632 | movaps %xmm5, -0x50(%rdi) |
633 | |
634 | movaps -0x65(%rsi), %xmm7 |
635 | palignr $5, %xmm7, %xmm6 |
636 | movaps %xmm6, -0x60(%rdi) |
637 | |
638 | movaps -0x75(%rsi), %xmm8 |
639 | palignr $5, %xmm8, %xmm7 |
640 | movaps %xmm7, -0x70(%rdi) |
641 | |
642 | movaps -0x85(%rsi), %xmm9 |
643 | palignr $5, %xmm9, %xmm8 |
644 | movaps %xmm8, -0x80(%rdi) |
645 | |
646 | sub $0x80, %rdx |
647 | lea -0x80(%rdi), %rdi |
648 | lea -0x80(%rsi), %rsi |
649 | jae L(shl_5_bwd) |
650 | movdqu %xmm0, (%r8) |
651 | add $0x80, %rdx |
652 | sub %rdx, %rdi |
653 | sub %rdx, %rsi |
654 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
655 | |
656 | .p2align 4 |
657 | L(shl_6): |
658 | sub $0x80, %rdx |
659 | movaps -0x06(%rsi), %xmm1 |
660 | movaps 0x0a(%rsi), %xmm2 |
661 | movaps 0x1a(%rsi), %xmm3 |
662 | movaps 0x2a(%rsi), %xmm4 |
663 | movaps 0x3a(%rsi), %xmm5 |
664 | movaps 0x4a(%rsi), %xmm6 |
665 | movaps 0x5a(%rsi), %xmm7 |
666 | movaps 0x6a(%rsi), %xmm8 |
667 | movaps 0x7a(%rsi), %xmm9 |
668 | lea 0x80(%rsi), %rsi |
669 | palignr $6, %xmm8, %xmm9 |
670 | movaps %xmm9, 0x70(%rdi) |
671 | palignr $6, %xmm7, %xmm8 |
672 | movaps %xmm8, 0x60(%rdi) |
673 | palignr $6, %xmm6, %xmm7 |
674 | movaps %xmm7, 0x50(%rdi) |
675 | palignr $6, %xmm5, %xmm6 |
676 | movaps %xmm6, 0x40(%rdi) |
677 | palignr $6, %xmm4, %xmm5 |
678 | movaps %xmm5, 0x30(%rdi) |
679 | palignr $6, %xmm3, %xmm4 |
680 | movaps %xmm4, 0x20(%rdi) |
681 | palignr $6, %xmm2, %xmm3 |
682 | movaps %xmm3, 0x10(%rdi) |
683 | palignr $6, %xmm1, %xmm2 |
684 | movaps %xmm2, (%rdi) |
685 | lea 0x80(%rdi), %rdi |
686 | jae L(shl_6) |
687 | movdqu %xmm0, (%r8) |
688 | add $0x80, %rdx |
689 | add %rdx, %rdi |
690 | add %rdx, %rsi |
691 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
692 | |
693 | .p2align 4 |
694 | L(shl_6_bwd): |
695 | movaps -0x06(%rsi), %xmm1 |
696 | |
697 | movaps -0x16(%rsi), %xmm2 |
698 | palignr $6, %xmm2, %xmm1 |
699 | movaps %xmm1, -0x10(%rdi) |
700 | |
701 | movaps -0x26(%rsi), %xmm3 |
702 | palignr $6, %xmm3, %xmm2 |
703 | movaps %xmm2, -0x20(%rdi) |
704 | |
705 | movaps -0x36(%rsi), %xmm4 |
706 | palignr $6, %xmm4, %xmm3 |
707 | movaps %xmm3, -0x30(%rdi) |
708 | |
709 | movaps -0x46(%rsi), %xmm5 |
710 | palignr $6, %xmm5, %xmm4 |
711 | movaps %xmm4, -0x40(%rdi) |
712 | |
713 | movaps -0x56(%rsi), %xmm6 |
714 | palignr $6, %xmm6, %xmm5 |
715 | movaps %xmm5, -0x50(%rdi) |
716 | |
717 | movaps -0x66(%rsi), %xmm7 |
718 | palignr $6, %xmm7, %xmm6 |
719 | movaps %xmm6, -0x60(%rdi) |
720 | |
721 | movaps -0x76(%rsi), %xmm8 |
722 | palignr $6, %xmm8, %xmm7 |
723 | movaps %xmm7, -0x70(%rdi) |
724 | |
725 | movaps -0x86(%rsi), %xmm9 |
726 | palignr $6, %xmm9, %xmm8 |
727 | movaps %xmm8, -0x80(%rdi) |
728 | |
729 | sub $0x80, %rdx |
730 | lea -0x80(%rdi), %rdi |
731 | lea -0x80(%rsi), %rsi |
732 | jae L(shl_6_bwd) |
733 | movdqu %xmm0, (%r8) |
734 | add $0x80, %rdx |
735 | sub %rdx, %rdi |
736 | sub %rdx, %rsi |
737 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
738 | |
739 | .p2align 4 |
740 | L(shl_7): |
741 | sub $0x80, %rdx |
742 | movaps -0x07(%rsi), %xmm1 |
743 | movaps 0x09(%rsi), %xmm2 |
744 | movaps 0x19(%rsi), %xmm3 |
745 | movaps 0x29(%rsi), %xmm4 |
746 | movaps 0x39(%rsi), %xmm5 |
747 | movaps 0x49(%rsi), %xmm6 |
748 | movaps 0x59(%rsi), %xmm7 |
749 | movaps 0x69(%rsi), %xmm8 |
750 | movaps 0x79(%rsi), %xmm9 |
751 | lea 0x80(%rsi), %rsi |
752 | palignr $7, %xmm8, %xmm9 |
753 | movaps %xmm9, 0x70(%rdi) |
754 | palignr $7, %xmm7, %xmm8 |
755 | movaps %xmm8, 0x60(%rdi) |
756 | palignr $7, %xmm6, %xmm7 |
757 | movaps %xmm7, 0x50(%rdi) |
758 | palignr $7, %xmm5, %xmm6 |
759 | movaps %xmm6, 0x40(%rdi) |
760 | palignr $7, %xmm4, %xmm5 |
761 | movaps %xmm5, 0x30(%rdi) |
762 | palignr $7, %xmm3, %xmm4 |
763 | movaps %xmm4, 0x20(%rdi) |
764 | palignr $7, %xmm2, %xmm3 |
765 | movaps %xmm3, 0x10(%rdi) |
766 | palignr $7, %xmm1, %xmm2 |
767 | movaps %xmm2, (%rdi) |
768 | lea 0x80(%rdi), %rdi |
769 | jae L(shl_7) |
770 | movdqu %xmm0, (%r8) |
771 | add $0x80, %rdx |
772 | add %rdx, %rdi |
773 | add %rdx, %rsi |
774 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
775 | |
776 | .p2align 4 |
777 | L(shl_7_bwd): |
778 | movaps -0x07(%rsi), %xmm1 |
779 | |
780 | movaps -0x17(%rsi), %xmm2 |
781 | palignr $7, %xmm2, %xmm1 |
782 | movaps %xmm1, -0x10(%rdi) |
783 | |
784 | movaps -0x27(%rsi), %xmm3 |
785 | palignr $7, %xmm3, %xmm2 |
786 | movaps %xmm2, -0x20(%rdi) |
787 | |
788 | movaps -0x37(%rsi), %xmm4 |
789 | palignr $7, %xmm4, %xmm3 |
790 | movaps %xmm3, -0x30(%rdi) |
791 | |
792 | movaps -0x47(%rsi), %xmm5 |
793 | palignr $7, %xmm5, %xmm4 |
794 | movaps %xmm4, -0x40(%rdi) |
795 | |
796 | movaps -0x57(%rsi), %xmm6 |
797 | palignr $7, %xmm6, %xmm5 |
798 | movaps %xmm5, -0x50(%rdi) |
799 | |
800 | movaps -0x67(%rsi), %xmm7 |
801 | palignr $7, %xmm7, %xmm6 |
802 | movaps %xmm6, -0x60(%rdi) |
803 | |
804 | movaps -0x77(%rsi), %xmm8 |
805 | palignr $7, %xmm8, %xmm7 |
806 | movaps %xmm7, -0x70(%rdi) |
807 | |
808 | movaps -0x87(%rsi), %xmm9 |
809 | palignr $7, %xmm9, %xmm8 |
810 | movaps %xmm8, -0x80(%rdi) |
811 | |
812 | sub $0x80, %rdx |
813 | lea -0x80(%rdi), %rdi |
814 | lea -0x80(%rsi), %rsi |
815 | jae L(shl_7_bwd) |
816 | movdqu %xmm0, (%r8) |
817 | add $0x80, %rdx |
818 | sub %rdx, %rdi |
819 | sub %rdx, %rsi |
820 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
821 | |
822 | .p2align 4 |
823 | L(shl_8): |
824 | sub $0x80, %rdx |
825 | movaps -0x08(%rsi), %xmm1 |
826 | movaps 0x08(%rsi), %xmm2 |
827 | movaps 0x18(%rsi), %xmm3 |
828 | movaps 0x28(%rsi), %xmm4 |
829 | movaps 0x38(%rsi), %xmm5 |
830 | movaps 0x48(%rsi), %xmm6 |
831 | movaps 0x58(%rsi), %xmm7 |
832 | movaps 0x68(%rsi), %xmm8 |
833 | movaps 0x78(%rsi), %xmm9 |
834 | lea 0x80(%rsi), %rsi |
835 | palignr $8, %xmm8, %xmm9 |
836 | movaps %xmm9, 0x70(%rdi) |
837 | palignr $8, %xmm7, %xmm8 |
838 | movaps %xmm8, 0x60(%rdi) |
839 | palignr $8, %xmm6, %xmm7 |
840 | movaps %xmm7, 0x50(%rdi) |
841 | palignr $8, %xmm5, %xmm6 |
842 | movaps %xmm6, 0x40(%rdi) |
843 | palignr $8, %xmm4, %xmm5 |
844 | movaps %xmm5, 0x30(%rdi) |
845 | palignr $8, %xmm3, %xmm4 |
846 | movaps %xmm4, 0x20(%rdi) |
847 | palignr $8, %xmm2, %xmm3 |
848 | movaps %xmm3, 0x10(%rdi) |
849 | palignr $8, %xmm1, %xmm2 |
850 | movaps %xmm2, (%rdi) |
851 | lea 0x80(%rdi), %rdi |
852 | jae L(shl_8) |
853 | movdqu %xmm0, (%r8) |
854 | add $0x80, %rdx |
855 | add %rdx, %rdi |
856 | add %rdx, %rsi |
857 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
858 | |
859 | .p2align 4 |
860 | L(shl_8_bwd): |
861 | movaps -0x08(%rsi), %xmm1 |
862 | |
863 | movaps -0x18(%rsi), %xmm2 |
864 | palignr $8, %xmm2, %xmm1 |
865 | movaps %xmm1, -0x10(%rdi) |
866 | |
867 | movaps -0x28(%rsi), %xmm3 |
868 | palignr $8, %xmm3, %xmm2 |
869 | movaps %xmm2, -0x20(%rdi) |
870 | |
871 | movaps -0x38(%rsi), %xmm4 |
872 | palignr $8, %xmm4, %xmm3 |
873 | movaps %xmm3, -0x30(%rdi) |
874 | |
875 | movaps -0x48(%rsi), %xmm5 |
876 | palignr $8, %xmm5, %xmm4 |
877 | movaps %xmm4, -0x40(%rdi) |
878 | |
879 | movaps -0x58(%rsi), %xmm6 |
880 | palignr $8, %xmm6, %xmm5 |
881 | movaps %xmm5, -0x50(%rdi) |
882 | |
883 | movaps -0x68(%rsi), %xmm7 |
884 | palignr $8, %xmm7, %xmm6 |
885 | movaps %xmm6, -0x60(%rdi) |
886 | |
887 | movaps -0x78(%rsi), %xmm8 |
888 | palignr $8, %xmm8, %xmm7 |
889 | movaps %xmm7, -0x70(%rdi) |
890 | |
891 | movaps -0x88(%rsi), %xmm9 |
892 | palignr $8, %xmm9, %xmm8 |
893 | movaps %xmm8, -0x80(%rdi) |
894 | |
895 | sub $0x80, %rdx |
896 | lea -0x80(%rdi), %rdi |
897 | lea -0x80(%rsi), %rsi |
898 | jae L(shl_8_bwd) |
899 | L(shl_8_end_bwd): |
900 | movdqu %xmm0, (%r8) |
901 | add $0x80, %rdx |
902 | sub %rdx, %rdi |
903 | sub %rdx, %rsi |
904 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
905 | |
906 | .p2align 4 |
907 | L(shl_9): |
908 | sub $0x80, %rdx |
909 | movaps -0x09(%rsi), %xmm1 |
910 | movaps 0x07(%rsi), %xmm2 |
911 | movaps 0x17(%rsi), %xmm3 |
912 | movaps 0x27(%rsi), %xmm4 |
913 | movaps 0x37(%rsi), %xmm5 |
914 | movaps 0x47(%rsi), %xmm6 |
915 | movaps 0x57(%rsi), %xmm7 |
916 | movaps 0x67(%rsi), %xmm8 |
917 | movaps 0x77(%rsi), %xmm9 |
918 | lea 0x80(%rsi), %rsi |
919 | palignr $9, %xmm8, %xmm9 |
920 | movaps %xmm9, 0x70(%rdi) |
921 | palignr $9, %xmm7, %xmm8 |
922 | movaps %xmm8, 0x60(%rdi) |
923 | palignr $9, %xmm6, %xmm7 |
924 | movaps %xmm7, 0x50(%rdi) |
925 | palignr $9, %xmm5, %xmm6 |
926 | movaps %xmm6, 0x40(%rdi) |
927 | palignr $9, %xmm4, %xmm5 |
928 | movaps %xmm5, 0x30(%rdi) |
929 | palignr $9, %xmm3, %xmm4 |
930 | movaps %xmm4, 0x20(%rdi) |
931 | palignr $9, %xmm2, %xmm3 |
932 | movaps %xmm3, 0x10(%rdi) |
933 | palignr $9, %xmm1, %xmm2 |
934 | movaps %xmm2, (%rdi) |
935 | lea 0x80(%rdi), %rdi |
936 | jae L(shl_9) |
937 | movdqu %xmm0, (%r8) |
938 | add $0x80, %rdx |
939 | add %rdx, %rdi |
940 | add %rdx, %rsi |
941 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
942 | |
943 | .p2align 4 |
944 | L(shl_9_bwd): |
945 | movaps -0x09(%rsi), %xmm1 |
946 | |
947 | movaps -0x19(%rsi), %xmm2 |
948 | palignr $9, %xmm2, %xmm1 |
949 | movaps %xmm1, -0x10(%rdi) |
950 | |
951 | movaps -0x29(%rsi), %xmm3 |
952 | palignr $9, %xmm3, %xmm2 |
953 | movaps %xmm2, -0x20(%rdi) |
954 | |
955 | movaps -0x39(%rsi), %xmm4 |
956 | palignr $9, %xmm4, %xmm3 |
957 | movaps %xmm3, -0x30(%rdi) |
958 | |
959 | movaps -0x49(%rsi), %xmm5 |
960 | palignr $9, %xmm5, %xmm4 |
961 | movaps %xmm4, -0x40(%rdi) |
962 | |
963 | movaps -0x59(%rsi), %xmm6 |
964 | palignr $9, %xmm6, %xmm5 |
965 | movaps %xmm5, -0x50(%rdi) |
966 | |
967 | movaps -0x69(%rsi), %xmm7 |
968 | palignr $9, %xmm7, %xmm6 |
969 | movaps %xmm6, -0x60(%rdi) |
970 | |
971 | movaps -0x79(%rsi), %xmm8 |
972 | palignr $9, %xmm8, %xmm7 |
973 | movaps %xmm7, -0x70(%rdi) |
974 | |
975 | movaps -0x89(%rsi), %xmm9 |
976 | palignr $9, %xmm9, %xmm8 |
977 | movaps %xmm8, -0x80(%rdi) |
978 | |
979 | sub $0x80, %rdx |
980 | lea -0x80(%rdi), %rdi |
981 | lea -0x80(%rsi), %rsi |
982 | jae L(shl_9_bwd) |
983 | movdqu %xmm0, (%r8) |
984 | add $0x80, %rdx |
985 | sub %rdx, %rdi |
986 | sub %rdx, %rsi |
987 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
988 | |
989 | .p2align 4 |
990 | L(shl_10): |
991 | sub $0x80, %rdx |
992 | movaps -0x0a(%rsi), %xmm1 |
993 | movaps 0x06(%rsi), %xmm2 |
994 | movaps 0x16(%rsi), %xmm3 |
995 | movaps 0x26(%rsi), %xmm4 |
996 | movaps 0x36(%rsi), %xmm5 |
997 | movaps 0x46(%rsi), %xmm6 |
998 | movaps 0x56(%rsi), %xmm7 |
999 | movaps 0x66(%rsi), %xmm8 |
1000 | movaps 0x76(%rsi), %xmm9 |
1001 | lea 0x80(%rsi), %rsi |
1002 | palignr $10, %xmm8, %xmm9 |
1003 | movaps %xmm9, 0x70(%rdi) |
1004 | palignr $10, %xmm7, %xmm8 |
1005 | movaps %xmm8, 0x60(%rdi) |
1006 | palignr $10, %xmm6, %xmm7 |
1007 | movaps %xmm7, 0x50(%rdi) |
1008 | palignr $10, %xmm5, %xmm6 |
1009 | movaps %xmm6, 0x40(%rdi) |
1010 | palignr $10, %xmm4, %xmm5 |
1011 | movaps %xmm5, 0x30(%rdi) |
1012 | palignr $10, %xmm3, %xmm4 |
1013 | movaps %xmm4, 0x20(%rdi) |
1014 | palignr $10, %xmm2, %xmm3 |
1015 | movaps %xmm3, 0x10(%rdi) |
1016 | palignr $10, %xmm1, %xmm2 |
1017 | movaps %xmm2, (%rdi) |
1018 | lea 0x80(%rdi), %rdi |
1019 | jae L(shl_10) |
1020 | movdqu %xmm0, (%r8) |
1021 | add $0x80, %rdx |
1022 | add %rdx, %rdi |
1023 | add %rdx, %rsi |
1024 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1025 | |
1026 | .p2align 4 |
1027 | L(shl_10_bwd): |
1028 | movaps -0x0a(%rsi), %xmm1 |
1029 | |
1030 | movaps -0x1a(%rsi), %xmm2 |
1031 | palignr $10, %xmm2, %xmm1 |
1032 | movaps %xmm1, -0x10(%rdi) |
1033 | |
1034 | movaps -0x2a(%rsi), %xmm3 |
1035 | palignr $10, %xmm3, %xmm2 |
1036 | movaps %xmm2, -0x20(%rdi) |
1037 | |
1038 | movaps -0x3a(%rsi), %xmm4 |
1039 | palignr $10, %xmm4, %xmm3 |
1040 | movaps %xmm3, -0x30(%rdi) |
1041 | |
1042 | movaps -0x4a(%rsi), %xmm5 |
1043 | palignr $10, %xmm5, %xmm4 |
1044 | movaps %xmm4, -0x40(%rdi) |
1045 | |
1046 | movaps -0x5a(%rsi), %xmm6 |
1047 | palignr $10, %xmm6, %xmm5 |
1048 | movaps %xmm5, -0x50(%rdi) |
1049 | |
1050 | movaps -0x6a(%rsi), %xmm7 |
1051 | palignr $10, %xmm7, %xmm6 |
1052 | movaps %xmm6, -0x60(%rdi) |
1053 | |
1054 | movaps -0x7a(%rsi), %xmm8 |
1055 | palignr $10, %xmm8, %xmm7 |
1056 | movaps %xmm7, -0x70(%rdi) |
1057 | |
1058 | movaps -0x8a(%rsi), %xmm9 |
1059 | palignr $10, %xmm9, %xmm8 |
1060 | movaps %xmm8, -0x80(%rdi) |
1061 | |
1062 | sub $0x80, %rdx |
1063 | lea -0x80(%rdi), %rdi |
1064 | lea -0x80(%rsi), %rsi |
1065 | jae L(shl_10_bwd) |
1066 | movdqu %xmm0, (%r8) |
1067 | add $0x80, %rdx |
1068 | sub %rdx, %rdi |
1069 | sub %rdx, %rsi |
1070 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1071 | |
1072 | .p2align 4 |
1073 | L(shl_11): |
1074 | sub $0x80, %rdx |
1075 | movaps -0x0b(%rsi), %xmm1 |
1076 | movaps 0x05(%rsi), %xmm2 |
1077 | movaps 0x15(%rsi), %xmm3 |
1078 | movaps 0x25(%rsi), %xmm4 |
1079 | movaps 0x35(%rsi), %xmm5 |
1080 | movaps 0x45(%rsi), %xmm6 |
1081 | movaps 0x55(%rsi), %xmm7 |
1082 | movaps 0x65(%rsi), %xmm8 |
1083 | movaps 0x75(%rsi), %xmm9 |
1084 | lea 0x80(%rsi), %rsi |
1085 | palignr $11, %xmm8, %xmm9 |
1086 | movaps %xmm9, 0x70(%rdi) |
1087 | palignr $11, %xmm7, %xmm8 |
1088 | movaps %xmm8, 0x60(%rdi) |
1089 | palignr $11, %xmm6, %xmm7 |
1090 | movaps %xmm7, 0x50(%rdi) |
1091 | palignr $11, %xmm5, %xmm6 |
1092 | movaps %xmm6, 0x40(%rdi) |
1093 | palignr $11, %xmm4, %xmm5 |
1094 | movaps %xmm5, 0x30(%rdi) |
1095 | palignr $11, %xmm3, %xmm4 |
1096 | movaps %xmm4, 0x20(%rdi) |
1097 | palignr $11, %xmm2, %xmm3 |
1098 | movaps %xmm3, 0x10(%rdi) |
1099 | palignr $11, %xmm1, %xmm2 |
1100 | movaps %xmm2, (%rdi) |
1101 | lea 0x80(%rdi), %rdi |
1102 | jae L(shl_11) |
1103 | movdqu %xmm0, (%r8) |
1104 | add $0x80, %rdx |
1105 | add %rdx, %rdi |
1106 | add %rdx, %rsi |
1107 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1108 | |
1109 | .p2align 4 |
1110 | L(shl_11_bwd): |
1111 | movaps -0x0b(%rsi), %xmm1 |
1112 | |
1113 | movaps -0x1b(%rsi), %xmm2 |
1114 | palignr $11, %xmm2, %xmm1 |
1115 | movaps %xmm1, -0x10(%rdi) |
1116 | |
1117 | movaps -0x2b(%rsi), %xmm3 |
1118 | palignr $11, %xmm3, %xmm2 |
1119 | movaps %xmm2, -0x20(%rdi) |
1120 | |
1121 | movaps -0x3b(%rsi), %xmm4 |
1122 | palignr $11, %xmm4, %xmm3 |
1123 | movaps %xmm3, -0x30(%rdi) |
1124 | |
1125 | movaps -0x4b(%rsi), %xmm5 |
1126 | palignr $11, %xmm5, %xmm4 |
1127 | movaps %xmm4, -0x40(%rdi) |
1128 | |
1129 | movaps -0x5b(%rsi), %xmm6 |
1130 | palignr $11, %xmm6, %xmm5 |
1131 | movaps %xmm5, -0x50(%rdi) |
1132 | |
1133 | movaps -0x6b(%rsi), %xmm7 |
1134 | palignr $11, %xmm7, %xmm6 |
1135 | movaps %xmm6, -0x60(%rdi) |
1136 | |
1137 | movaps -0x7b(%rsi), %xmm8 |
1138 | palignr $11, %xmm8, %xmm7 |
1139 | movaps %xmm7, -0x70(%rdi) |
1140 | |
1141 | movaps -0x8b(%rsi), %xmm9 |
1142 | palignr $11, %xmm9, %xmm8 |
1143 | movaps %xmm8, -0x80(%rdi) |
1144 | |
1145 | sub $0x80, %rdx |
1146 | lea -0x80(%rdi), %rdi |
1147 | lea -0x80(%rsi), %rsi |
1148 | jae L(shl_11_bwd) |
1149 | movdqu %xmm0, (%r8) |
1150 | add $0x80, %rdx |
1151 | sub %rdx, %rdi |
1152 | sub %rdx, %rsi |
1153 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1154 | |
1155 | .p2align 4 |
1156 | L(shl_12): |
1157 | sub $0x80, %rdx |
1158 | movdqa -0x0c(%rsi), %xmm1 |
1159 | movaps 0x04(%rsi), %xmm2 |
1160 | movaps 0x14(%rsi), %xmm3 |
1161 | movaps 0x24(%rsi), %xmm4 |
1162 | movaps 0x34(%rsi), %xmm5 |
1163 | movaps 0x44(%rsi), %xmm6 |
1164 | movaps 0x54(%rsi), %xmm7 |
1165 | movaps 0x64(%rsi), %xmm8 |
1166 | movaps 0x74(%rsi), %xmm9 |
1167 | lea 0x80(%rsi), %rsi |
1168 | palignr $12, %xmm8, %xmm9 |
1169 | movaps %xmm9, 0x70(%rdi) |
1170 | palignr $12, %xmm7, %xmm8 |
1171 | movaps %xmm8, 0x60(%rdi) |
1172 | palignr $12, %xmm6, %xmm7 |
1173 | movaps %xmm7, 0x50(%rdi) |
1174 | palignr $12, %xmm5, %xmm6 |
1175 | movaps %xmm6, 0x40(%rdi) |
1176 | palignr $12, %xmm4, %xmm5 |
1177 | movaps %xmm5, 0x30(%rdi) |
1178 | palignr $12, %xmm3, %xmm4 |
1179 | movaps %xmm4, 0x20(%rdi) |
1180 | palignr $12, %xmm2, %xmm3 |
1181 | movaps %xmm3, 0x10(%rdi) |
1182 | palignr $12, %xmm1, %xmm2 |
1183 | movaps %xmm2, (%rdi) |
1184 | |
1185 | lea 0x80(%rdi), %rdi |
1186 | jae L(shl_12) |
1187 | movdqu %xmm0, (%r8) |
1188 | add $0x80, %rdx |
1189 | add %rdx, %rdi |
1190 | add %rdx, %rsi |
1191 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1192 | |
1193 | .p2align 4 |
1194 | L(shl_12_bwd): |
1195 | movaps -0x0c(%rsi), %xmm1 |
1196 | |
1197 | movaps -0x1c(%rsi), %xmm2 |
1198 | palignr $12, %xmm2, %xmm1 |
1199 | movaps %xmm1, -0x10(%rdi) |
1200 | |
1201 | movaps -0x2c(%rsi), %xmm3 |
1202 | palignr $12, %xmm3, %xmm2 |
1203 | movaps %xmm2, -0x20(%rdi) |
1204 | |
1205 | movaps -0x3c(%rsi), %xmm4 |
1206 | palignr $12, %xmm4, %xmm3 |
1207 | movaps %xmm3, -0x30(%rdi) |
1208 | |
1209 | movaps -0x4c(%rsi), %xmm5 |
1210 | palignr $12, %xmm5, %xmm4 |
1211 | movaps %xmm4, -0x40(%rdi) |
1212 | |
1213 | movaps -0x5c(%rsi), %xmm6 |
1214 | palignr $12, %xmm6, %xmm5 |
1215 | movaps %xmm5, -0x50(%rdi) |
1216 | |
1217 | movaps -0x6c(%rsi), %xmm7 |
1218 | palignr $12, %xmm7, %xmm6 |
1219 | movaps %xmm6, -0x60(%rdi) |
1220 | |
1221 | movaps -0x7c(%rsi), %xmm8 |
1222 | palignr $12, %xmm8, %xmm7 |
1223 | movaps %xmm7, -0x70(%rdi) |
1224 | |
1225 | movaps -0x8c(%rsi), %xmm9 |
1226 | palignr $12, %xmm9, %xmm8 |
1227 | movaps %xmm8, -0x80(%rdi) |
1228 | |
1229 | sub $0x80, %rdx |
1230 | lea -0x80(%rdi), %rdi |
1231 | lea -0x80(%rsi), %rsi |
1232 | jae L(shl_12_bwd) |
1233 | movdqu %xmm0, (%r8) |
1234 | add $0x80, %rdx |
1235 | sub %rdx, %rdi |
1236 | sub %rdx, %rsi |
1237 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1238 | |
1239 | .p2align 4 |
1240 | L(shl_13): |
1241 | sub $0x80, %rdx |
1242 | movaps -0x0d(%rsi), %xmm1 |
1243 | movaps 0x03(%rsi), %xmm2 |
1244 | movaps 0x13(%rsi), %xmm3 |
1245 | movaps 0x23(%rsi), %xmm4 |
1246 | movaps 0x33(%rsi), %xmm5 |
1247 | movaps 0x43(%rsi), %xmm6 |
1248 | movaps 0x53(%rsi), %xmm7 |
1249 | movaps 0x63(%rsi), %xmm8 |
1250 | movaps 0x73(%rsi), %xmm9 |
1251 | lea 0x80(%rsi), %rsi |
1252 | palignr $13, %xmm8, %xmm9 |
1253 | movaps %xmm9, 0x70(%rdi) |
1254 | palignr $13, %xmm7, %xmm8 |
1255 | movaps %xmm8, 0x60(%rdi) |
1256 | palignr $13, %xmm6, %xmm7 |
1257 | movaps %xmm7, 0x50(%rdi) |
1258 | palignr $13, %xmm5, %xmm6 |
1259 | movaps %xmm6, 0x40(%rdi) |
1260 | palignr $13, %xmm4, %xmm5 |
1261 | movaps %xmm5, 0x30(%rdi) |
1262 | palignr $13, %xmm3, %xmm4 |
1263 | movaps %xmm4, 0x20(%rdi) |
1264 | palignr $13, %xmm2, %xmm3 |
1265 | movaps %xmm3, 0x10(%rdi) |
1266 | palignr $13, %xmm1, %xmm2 |
1267 | movaps %xmm2, (%rdi) |
1268 | lea 0x80(%rdi), %rdi |
1269 | jae L(shl_13) |
1270 | movdqu %xmm0, (%r8) |
1271 | add $0x80, %rdx |
1272 | add %rdx, %rdi |
1273 | add %rdx, %rsi |
1274 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1275 | |
1276 | .p2align 4 |
1277 | L(shl_13_bwd): |
1278 | movaps -0x0d(%rsi), %xmm1 |
1279 | |
1280 | movaps -0x1d(%rsi), %xmm2 |
1281 | palignr $13, %xmm2, %xmm1 |
1282 | movaps %xmm1, -0x10(%rdi) |
1283 | |
1284 | movaps -0x2d(%rsi), %xmm3 |
1285 | palignr $13, %xmm3, %xmm2 |
1286 | movaps %xmm2, -0x20(%rdi) |
1287 | |
1288 | movaps -0x3d(%rsi), %xmm4 |
1289 | palignr $13, %xmm4, %xmm3 |
1290 | movaps %xmm3, -0x30(%rdi) |
1291 | |
1292 | movaps -0x4d(%rsi), %xmm5 |
1293 | palignr $13, %xmm5, %xmm4 |
1294 | movaps %xmm4, -0x40(%rdi) |
1295 | |
1296 | movaps -0x5d(%rsi), %xmm6 |
1297 | palignr $13, %xmm6, %xmm5 |
1298 | movaps %xmm5, -0x50(%rdi) |
1299 | |
1300 | movaps -0x6d(%rsi), %xmm7 |
1301 | palignr $13, %xmm7, %xmm6 |
1302 | movaps %xmm6, -0x60(%rdi) |
1303 | |
1304 | movaps -0x7d(%rsi), %xmm8 |
1305 | palignr $13, %xmm8, %xmm7 |
1306 | movaps %xmm7, -0x70(%rdi) |
1307 | |
1308 | movaps -0x8d(%rsi), %xmm9 |
1309 | palignr $13, %xmm9, %xmm8 |
1310 | movaps %xmm8, -0x80(%rdi) |
1311 | |
1312 | sub $0x80, %rdx |
1313 | lea -0x80(%rdi), %rdi |
1314 | lea -0x80(%rsi), %rsi |
1315 | jae L(shl_13_bwd) |
1316 | movdqu %xmm0, (%r8) |
1317 | add $0x80, %rdx |
1318 | sub %rdx, %rdi |
1319 | sub %rdx, %rsi |
1320 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1321 | |
1322 | .p2align 4 |
1323 | L(shl_14): |
1324 | sub $0x80, %rdx |
1325 | movaps -0x0e(%rsi), %xmm1 |
1326 | movaps 0x02(%rsi), %xmm2 |
1327 | movaps 0x12(%rsi), %xmm3 |
1328 | movaps 0x22(%rsi), %xmm4 |
1329 | movaps 0x32(%rsi), %xmm5 |
1330 | movaps 0x42(%rsi), %xmm6 |
1331 | movaps 0x52(%rsi), %xmm7 |
1332 | movaps 0x62(%rsi), %xmm8 |
1333 | movaps 0x72(%rsi), %xmm9 |
1334 | lea 0x80(%rsi), %rsi |
1335 | palignr $14, %xmm8, %xmm9 |
1336 | movaps %xmm9, 0x70(%rdi) |
1337 | palignr $14, %xmm7, %xmm8 |
1338 | movaps %xmm8, 0x60(%rdi) |
1339 | palignr $14, %xmm6, %xmm7 |
1340 | movaps %xmm7, 0x50(%rdi) |
1341 | palignr $14, %xmm5, %xmm6 |
1342 | movaps %xmm6, 0x40(%rdi) |
1343 | palignr $14, %xmm4, %xmm5 |
1344 | movaps %xmm5, 0x30(%rdi) |
1345 | palignr $14, %xmm3, %xmm4 |
1346 | movaps %xmm4, 0x20(%rdi) |
1347 | palignr $14, %xmm2, %xmm3 |
1348 | movaps %xmm3, 0x10(%rdi) |
1349 | palignr $14, %xmm1, %xmm2 |
1350 | movaps %xmm2, (%rdi) |
1351 | lea 0x80(%rdi), %rdi |
1352 | jae L(shl_14) |
1353 | movdqu %xmm0, (%r8) |
1354 | add $0x80, %rdx |
1355 | add %rdx, %rdi |
1356 | add %rdx, %rsi |
1357 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1358 | |
1359 | .p2align 4 |
1360 | L(shl_14_bwd): |
1361 | movaps -0x0e(%rsi), %xmm1 |
1362 | |
1363 | movaps -0x1e(%rsi), %xmm2 |
1364 | palignr $14, %xmm2, %xmm1 |
1365 | movaps %xmm1, -0x10(%rdi) |
1366 | |
1367 | movaps -0x2e(%rsi), %xmm3 |
1368 | palignr $14, %xmm3, %xmm2 |
1369 | movaps %xmm2, -0x20(%rdi) |
1370 | |
1371 | movaps -0x3e(%rsi), %xmm4 |
1372 | palignr $14, %xmm4, %xmm3 |
1373 | movaps %xmm3, -0x30(%rdi) |
1374 | |
1375 | movaps -0x4e(%rsi), %xmm5 |
1376 | palignr $14, %xmm5, %xmm4 |
1377 | movaps %xmm4, -0x40(%rdi) |
1378 | |
1379 | movaps -0x5e(%rsi), %xmm6 |
1380 | palignr $14, %xmm6, %xmm5 |
1381 | movaps %xmm5, -0x50(%rdi) |
1382 | |
1383 | movaps -0x6e(%rsi), %xmm7 |
1384 | palignr $14, %xmm7, %xmm6 |
1385 | movaps %xmm6, -0x60(%rdi) |
1386 | |
1387 | movaps -0x7e(%rsi), %xmm8 |
1388 | palignr $14, %xmm8, %xmm7 |
1389 | movaps %xmm7, -0x70(%rdi) |
1390 | |
1391 | movaps -0x8e(%rsi), %xmm9 |
1392 | palignr $14, %xmm9, %xmm8 |
1393 | movaps %xmm8, -0x80(%rdi) |
1394 | |
1395 | sub $0x80, %rdx |
1396 | lea -0x80(%rdi), %rdi |
1397 | lea -0x80(%rsi), %rsi |
1398 | jae L(shl_14_bwd) |
1399 | movdqu %xmm0, (%r8) |
1400 | add $0x80, %rdx |
1401 | sub %rdx, %rdi |
1402 | sub %rdx, %rsi |
1403 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1404 | |
1405 | .p2align 4 |
1406 | L(shl_15): |
1407 | sub $0x80, %rdx |
1408 | movaps -0x0f(%rsi), %xmm1 |
1409 | movaps 0x01(%rsi), %xmm2 |
1410 | movaps 0x11(%rsi), %xmm3 |
1411 | movaps 0x21(%rsi), %xmm4 |
1412 | movaps 0x31(%rsi), %xmm5 |
1413 | movaps 0x41(%rsi), %xmm6 |
1414 | movaps 0x51(%rsi), %xmm7 |
1415 | movaps 0x61(%rsi), %xmm8 |
1416 | movaps 0x71(%rsi), %xmm9 |
1417 | lea 0x80(%rsi), %rsi |
1418 | palignr $15, %xmm8, %xmm9 |
1419 | movaps %xmm9, 0x70(%rdi) |
1420 | palignr $15, %xmm7, %xmm8 |
1421 | movaps %xmm8, 0x60(%rdi) |
1422 | palignr $15, %xmm6, %xmm7 |
1423 | movaps %xmm7, 0x50(%rdi) |
1424 | palignr $15, %xmm5, %xmm6 |
1425 | movaps %xmm6, 0x40(%rdi) |
1426 | palignr $15, %xmm4, %xmm5 |
1427 | movaps %xmm5, 0x30(%rdi) |
1428 | palignr $15, %xmm3, %xmm4 |
1429 | movaps %xmm4, 0x20(%rdi) |
1430 | palignr $15, %xmm2, %xmm3 |
1431 | movaps %xmm3, 0x10(%rdi) |
1432 | palignr $15, %xmm1, %xmm2 |
1433 | movaps %xmm2, (%rdi) |
1434 | lea 0x80(%rdi), %rdi |
1435 | jae L(shl_15) |
1436 | movdqu %xmm0, (%r8) |
1437 | add $0x80, %rdx |
1438 | add %rdx, %rdi |
1439 | add %rdx, %rsi |
1440 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1441 | |
1442 | .p2align 4 |
1443 | L(shl_15_bwd): |
1444 | movaps -0x0f(%rsi), %xmm1 |
1445 | |
1446 | movaps -0x1f(%rsi), %xmm2 |
1447 | palignr $15, %xmm2, %xmm1 |
1448 | movaps %xmm1, -0x10(%rdi) |
1449 | |
1450 | movaps -0x2f(%rsi), %xmm3 |
1451 | palignr $15, %xmm3, %xmm2 |
1452 | movaps %xmm2, -0x20(%rdi) |
1453 | |
1454 | movaps -0x3f(%rsi), %xmm4 |
1455 | palignr $15, %xmm4, %xmm3 |
1456 | movaps %xmm3, -0x30(%rdi) |
1457 | |
1458 | movaps -0x4f(%rsi), %xmm5 |
1459 | palignr $15, %xmm5, %xmm4 |
1460 | movaps %xmm4, -0x40(%rdi) |
1461 | |
1462 | movaps -0x5f(%rsi), %xmm6 |
1463 | palignr $15, %xmm6, %xmm5 |
1464 | movaps %xmm5, -0x50(%rdi) |
1465 | |
1466 | movaps -0x6f(%rsi), %xmm7 |
1467 | palignr $15, %xmm7, %xmm6 |
1468 | movaps %xmm6, -0x60(%rdi) |
1469 | |
1470 | movaps -0x7f(%rsi), %xmm8 |
1471 | palignr $15, %xmm8, %xmm7 |
1472 | movaps %xmm7, -0x70(%rdi) |
1473 | |
1474 | movaps -0x8f(%rsi), %xmm9 |
1475 | palignr $15, %xmm9, %xmm8 |
1476 | movaps %xmm8, -0x80(%rdi) |
1477 | |
1478 | sub $0x80, %rdx |
1479 | lea -0x80(%rdi), %rdi |
1480 | lea -0x80(%rsi), %rsi |
1481 | jae L(shl_15_bwd) |
1482 | movdqu %xmm0, (%r8) |
1483 | add $0x80, %rdx |
1484 | sub %rdx, %rdi |
1485 | sub %rdx, %rsi |
1486 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1487 | |
1488 | .p2align 4 |
1489 | L(gobble_mem_fwd): |
1490 | movdqu (%rsi), %xmm1 |
1491 | movdqu %xmm0, (%r8) |
1492 | movdqa %xmm1, (%rdi) |
1493 | sub $16, %rdx |
1494 | add $16, %rsi |
1495 | add $16, %rdi |
1496 | |
1497 | #ifdef SHARED_CACHE_SIZE_HALF |
1498 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1499 | #else |
1500 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1501 | #endif |
1502 | #ifdef USE_AS_MEMMOVE |
1503 | mov %rsi, %r9 |
1504 | sub %rdi, %r9 |
1505 | cmp %rdx, %r9 |
1506 | jae L(memmove_is_memcpy_fwd) |
1507 | cmp %rcx, %r9 |
1508 | jbe L(ll_cache_copy_fwd_start) |
1509 | L(memmove_is_memcpy_fwd): |
1510 | #endif |
1511 | cmp %rcx, %rdx |
1512 | ja L(bigger_in_fwd) |
1513 | mov %rdx, %rcx |
1514 | L(bigger_in_fwd): |
1515 | sub %rcx, %rdx |
1516 | cmp $0x1000, %rdx |
1517 | jbe L(ll_cache_copy_fwd) |
1518 | |
1519 | mov %rcx, %r9 |
1520 | shl $3, %r9 |
1521 | cmp %r9, %rdx |
1522 | jbe L(2steps_copy_fwd) |
1523 | add %rcx, %rdx |
1524 | xor %rcx, %rcx |
1525 | L(2steps_copy_fwd): |
1526 | sub $0x80, %rdx |
1527 | L(gobble_mem_fwd_loop): |
1528 | sub $0x80, %rdx |
1529 | prefetcht0 0x200(%rsi) |
1530 | prefetcht0 0x300(%rsi) |
1531 | movdqu (%rsi), %xmm0 |
1532 | movdqu 0x10(%rsi), %xmm1 |
1533 | movdqu 0x20(%rsi), %xmm2 |
1534 | movdqu 0x30(%rsi), %xmm3 |
1535 | movdqu 0x40(%rsi), %xmm4 |
1536 | movdqu 0x50(%rsi), %xmm5 |
1537 | movdqu 0x60(%rsi), %xmm6 |
1538 | movdqu 0x70(%rsi), %xmm7 |
1539 | lfence |
1540 | movntdq %xmm0, (%rdi) |
1541 | movntdq %xmm1, 0x10(%rdi) |
1542 | movntdq %xmm2, 0x20(%rdi) |
1543 | movntdq %xmm3, 0x30(%rdi) |
1544 | movntdq %xmm4, 0x40(%rdi) |
1545 | movntdq %xmm5, 0x50(%rdi) |
1546 | movntdq %xmm6, 0x60(%rdi) |
1547 | movntdq %xmm7, 0x70(%rdi) |
1548 | lea 0x80(%rsi), %rsi |
1549 | lea 0x80(%rdi), %rdi |
1550 | jae L(gobble_mem_fwd_loop) |
1551 | sfence |
1552 | cmp $0x80, %rcx |
1553 | jb L(gobble_mem_fwd_end) |
1554 | add $0x80, %rdx |
1555 | L(ll_cache_copy_fwd): |
1556 | add %rcx, %rdx |
1557 | L(ll_cache_copy_fwd_start): |
1558 | sub $0x80, %rdx |
1559 | L(gobble_ll_loop_fwd): |
1560 | prefetchnta 0x1c0(%rsi) |
1561 | prefetchnta 0x280(%rsi) |
1562 | prefetchnta 0x1c0(%rdi) |
1563 | prefetchnta 0x280(%rdi) |
1564 | sub $0x80, %rdx |
1565 | movdqu (%rsi), %xmm0 |
1566 | movdqu 0x10(%rsi), %xmm1 |
1567 | movdqu 0x20(%rsi), %xmm2 |
1568 | movdqu 0x30(%rsi), %xmm3 |
1569 | movdqu 0x40(%rsi), %xmm4 |
1570 | movdqu 0x50(%rsi), %xmm5 |
1571 | movdqu 0x60(%rsi), %xmm6 |
1572 | movdqu 0x70(%rsi), %xmm7 |
1573 | movdqa %xmm0, (%rdi) |
1574 | movdqa %xmm1, 0x10(%rdi) |
1575 | movdqa %xmm2, 0x20(%rdi) |
1576 | movdqa %xmm3, 0x30(%rdi) |
1577 | movdqa %xmm4, 0x40(%rdi) |
1578 | movdqa %xmm5, 0x50(%rdi) |
1579 | movdqa %xmm6, 0x60(%rdi) |
1580 | movdqa %xmm7, 0x70(%rdi) |
1581 | lea 0x80(%rsi), %rsi |
1582 | lea 0x80(%rdi), %rdi |
1583 | jae L(gobble_ll_loop_fwd) |
1584 | L(gobble_mem_fwd_end): |
1585 | add $0x80, %rdx |
1586 | add %rdx, %rsi |
1587 | add %rdx, %rdi |
1588 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1589 | |
1590 | .p2align 4 |
1591 | L(gobble_mem_bwd): |
1592 | add %rdx, %rsi |
1593 | add %rdx, %rdi |
1594 | |
1595 | movdqu -16(%rsi), %xmm0 |
1596 | lea -16(%rdi), %r8 |
1597 | mov %rdi, %r9 |
1598 | and $-16, %rdi |
1599 | sub %rdi, %r9 |
1600 | sub %r9, %rsi |
1601 | sub %r9, %rdx |
1602 | |
1603 | |
1604 | #ifdef SHARED_CACHE_SIZE_HALF |
1605 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1606 | #else |
1607 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1608 | #endif |
1609 | #ifdef USE_AS_MEMMOVE |
1610 | mov %rdi, %r9 |
1611 | sub %rsi, %r9 |
1612 | cmp %rdx, %r9 |
1613 | jae L(memmove_is_memcpy_bwd) |
1614 | cmp %rcx, %r9 |
1615 | jbe L(ll_cache_copy_bwd_start) |
1616 | L(memmove_is_memcpy_bwd): |
1617 | #endif |
1618 | cmp %rcx, %rdx |
1619 | ja L(bigger) |
1620 | mov %rdx, %rcx |
1621 | L(bigger): |
1622 | sub %rcx, %rdx |
1623 | cmp $0x1000, %rdx |
1624 | jbe L(ll_cache_copy) |
1625 | |
1626 | mov %rcx, %r9 |
1627 | shl $3, %r9 |
1628 | cmp %r9, %rdx |
1629 | jbe L(2steps_copy) |
1630 | add %rcx, %rdx |
1631 | xor %rcx, %rcx |
1632 | L(2steps_copy): |
1633 | sub $0x80, %rdx |
1634 | L(gobble_mem_bwd_loop): |
1635 | sub $0x80, %rdx |
1636 | prefetcht0 -0x200(%rsi) |
1637 | prefetcht0 -0x300(%rsi) |
1638 | movdqu -0x10(%rsi), %xmm1 |
1639 | movdqu -0x20(%rsi), %xmm2 |
1640 | movdqu -0x30(%rsi), %xmm3 |
1641 | movdqu -0x40(%rsi), %xmm4 |
1642 | movdqu -0x50(%rsi), %xmm5 |
1643 | movdqu -0x60(%rsi), %xmm6 |
1644 | movdqu -0x70(%rsi), %xmm7 |
1645 | movdqu -0x80(%rsi), %xmm8 |
1646 | lfence |
1647 | movntdq %xmm1, -0x10(%rdi) |
1648 | movntdq %xmm2, -0x20(%rdi) |
1649 | movntdq %xmm3, -0x30(%rdi) |
1650 | movntdq %xmm4, -0x40(%rdi) |
1651 | movntdq %xmm5, -0x50(%rdi) |
1652 | movntdq %xmm6, -0x60(%rdi) |
1653 | movntdq %xmm7, -0x70(%rdi) |
1654 | movntdq %xmm8, -0x80(%rdi) |
1655 | lea -0x80(%rsi), %rsi |
1656 | lea -0x80(%rdi), %rdi |
1657 | jae L(gobble_mem_bwd_loop) |
1658 | sfence |
1659 | cmp $0x80, %rcx |
1660 | jb L(gobble_mem_bwd_end) |
1661 | add $0x80, %rdx |
1662 | L(ll_cache_copy): |
1663 | add %rcx, %rdx |
1664 | L(ll_cache_copy_bwd_start): |
1665 | sub $0x80, %rdx |
1666 | L(gobble_ll_loop): |
1667 | prefetchnta -0x1c0(%rsi) |
1668 | prefetchnta -0x280(%rsi) |
1669 | prefetchnta -0x1c0(%rdi) |
1670 | prefetchnta -0x280(%rdi) |
1671 | sub $0x80, %rdx |
1672 | movdqu -0x10(%rsi), %xmm1 |
1673 | movdqu -0x20(%rsi), %xmm2 |
1674 | movdqu -0x30(%rsi), %xmm3 |
1675 | movdqu -0x40(%rsi), %xmm4 |
1676 | movdqu -0x50(%rsi), %xmm5 |
1677 | movdqu -0x60(%rsi), %xmm6 |
1678 | movdqu -0x70(%rsi), %xmm7 |
1679 | movdqu -0x80(%rsi), %xmm8 |
1680 | movdqa %xmm1, -0x10(%rdi) |
1681 | movdqa %xmm2, -0x20(%rdi) |
1682 | movdqa %xmm3, -0x30(%rdi) |
1683 | movdqa %xmm4, -0x40(%rdi) |
1684 | movdqa %xmm5, -0x50(%rdi) |
1685 | movdqa %xmm6, -0x60(%rdi) |
1686 | movdqa %xmm7, -0x70(%rdi) |
1687 | movdqa %xmm8, -0x80(%rdi) |
1688 | lea -0x80(%rsi), %rsi |
1689 | lea -0x80(%rdi), %rdi |
1690 | jae L(gobble_ll_loop) |
1691 | L(gobble_mem_bwd_end): |
1692 | movdqu %xmm0, (%r8) |
1693 | add $0x80, %rdx |
1694 | sub %rdx, %rsi |
1695 | sub %rdx, %rdi |
1696 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1697 | |
1698 | .p2align 4 |
1699 | L(fwd_write_128bytes): |
1700 | lddqu -128(%rsi), %xmm0 |
1701 | movdqu %xmm0, -128(%rdi) |
1702 | L(fwd_write_112bytes): |
1703 | lddqu -112(%rsi), %xmm0 |
1704 | movdqu %xmm0, -112(%rdi) |
1705 | L(fwd_write_96bytes): |
1706 | lddqu -96(%rsi), %xmm0 |
1707 | movdqu %xmm0, -96(%rdi) |
1708 | L(fwd_write_80bytes): |
1709 | lddqu -80(%rsi), %xmm0 |
1710 | movdqu %xmm0, -80(%rdi) |
1711 | L(fwd_write_64bytes): |
1712 | lddqu -64(%rsi), %xmm0 |
1713 | movdqu %xmm0, -64(%rdi) |
1714 | L(fwd_write_48bytes): |
1715 | lddqu -48(%rsi), %xmm0 |
1716 | movdqu %xmm0, -48(%rdi) |
1717 | L(fwd_write_32bytes): |
1718 | lddqu -32(%rsi), %xmm0 |
1719 | movdqu %xmm0, -32(%rdi) |
1720 | L(fwd_write_16bytes): |
1721 | lddqu -16(%rsi), %xmm0 |
1722 | movdqu %xmm0, -16(%rdi) |
1723 | L(fwd_write_0bytes): |
1724 | ret |
1725 | |
1726 | |
1727 | .p2align 4 |
1728 | L(fwd_write_143bytes): |
1729 | lddqu -143(%rsi), %xmm0 |
1730 | movdqu %xmm0, -143(%rdi) |
1731 | L(fwd_write_127bytes): |
1732 | lddqu -127(%rsi), %xmm0 |
1733 | movdqu %xmm0, -127(%rdi) |
1734 | L(fwd_write_111bytes): |
1735 | lddqu -111(%rsi), %xmm0 |
1736 | movdqu %xmm0, -111(%rdi) |
1737 | L(fwd_write_95bytes): |
1738 | lddqu -95(%rsi), %xmm0 |
1739 | movdqu %xmm0, -95(%rdi) |
1740 | L(fwd_write_79bytes): |
1741 | lddqu -79(%rsi), %xmm0 |
1742 | movdqu %xmm0, -79(%rdi) |
1743 | L(fwd_write_63bytes): |
1744 | lddqu -63(%rsi), %xmm0 |
1745 | movdqu %xmm0, -63(%rdi) |
1746 | L(fwd_write_47bytes): |
1747 | lddqu -47(%rsi), %xmm0 |
1748 | movdqu %xmm0, -47(%rdi) |
1749 | L(fwd_write_31bytes): |
1750 | lddqu -31(%rsi), %xmm0 |
1751 | lddqu -16(%rsi), %xmm1 |
1752 | movdqu %xmm0, -31(%rdi) |
1753 | movdqu %xmm1, -16(%rdi) |
1754 | ret |
1755 | |
1756 | .p2align 4 |
1757 | L(fwd_write_15bytes): |
1758 | mov -15(%rsi), %rdx |
1759 | mov -8(%rsi), %rcx |
1760 | mov %rdx, -15(%rdi) |
1761 | mov %rcx, -8(%rdi) |
1762 | ret |
1763 | |
1764 | .p2align 4 |
1765 | L(fwd_write_142bytes): |
1766 | lddqu -142(%rsi), %xmm0 |
1767 | movdqu %xmm0, -142(%rdi) |
1768 | L(fwd_write_126bytes): |
1769 | lddqu -126(%rsi), %xmm0 |
1770 | movdqu %xmm0, -126(%rdi) |
1771 | L(fwd_write_110bytes): |
1772 | lddqu -110(%rsi), %xmm0 |
1773 | movdqu %xmm0, -110(%rdi) |
1774 | L(fwd_write_94bytes): |
1775 | lddqu -94(%rsi), %xmm0 |
1776 | movdqu %xmm0, -94(%rdi) |
1777 | L(fwd_write_78bytes): |
1778 | lddqu -78(%rsi), %xmm0 |
1779 | movdqu %xmm0, -78(%rdi) |
1780 | L(fwd_write_62bytes): |
1781 | lddqu -62(%rsi), %xmm0 |
1782 | movdqu %xmm0, -62(%rdi) |
1783 | L(fwd_write_46bytes): |
1784 | lddqu -46(%rsi), %xmm0 |
1785 | movdqu %xmm0, -46(%rdi) |
1786 | L(fwd_write_30bytes): |
1787 | lddqu -30(%rsi), %xmm0 |
1788 | lddqu -16(%rsi), %xmm1 |
1789 | movdqu %xmm0, -30(%rdi) |
1790 | movdqu %xmm1, -16(%rdi) |
1791 | ret |
1792 | |
1793 | .p2align 4 |
1794 | L(fwd_write_14bytes): |
1795 | mov -14(%rsi), %rdx |
1796 | mov -8(%rsi), %rcx |
1797 | mov %rdx, -14(%rdi) |
1798 | mov %rcx, -8(%rdi) |
1799 | ret |
1800 | |
1801 | .p2align 4 |
1802 | L(fwd_write_141bytes): |
1803 | lddqu -141(%rsi), %xmm0 |
1804 | movdqu %xmm0, -141(%rdi) |
1805 | L(fwd_write_125bytes): |
1806 | lddqu -125(%rsi), %xmm0 |
1807 | movdqu %xmm0, -125(%rdi) |
1808 | L(fwd_write_109bytes): |
1809 | lddqu -109(%rsi), %xmm0 |
1810 | movdqu %xmm0, -109(%rdi) |
1811 | L(fwd_write_93bytes): |
1812 | lddqu -93(%rsi), %xmm0 |
1813 | movdqu %xmm0, -93(%rdi) |
1814 | L(fwd_write_77bytes): |
1815 | lddqu -77(%rsi), %xmm0 |
1816 | movdqu %xmm0, -77(%rdi) |
1817 | L(fwd_write_61bytes): |
1818 | lddqu -61(%rsi), %xmm0 |
1819 | movdqu %xmm0, -61(%rdi) |
1820 | L(fwd_write_45bytes): |
1821 | lddqu -45(%rsi), %xmm0 |
1822 | movdqu %xmm0, -45(%rdi) |
1823 | L(fwd_write_29bytes): |
1824 | lddqu -29(%rsi), %xmm0 |
1825 | lddqu -16(%rsi), %xmm1 |
1826 | movdqu %xmm0, -29(%rdi) |
1827 | movdqu %xmm1, -16(%rdi) |
1828 | ret |
1829 | |
1830 | .p2align 4 |
1831 | L(fwd_write_13bytes): |
1832 | mov -13(%rsi), %rdx |
1833 | mov -8(%rsi), %rcx |
1834 | mov %rdx, -13(%rdi) |
1835 | mov %rcx, -8(%rdi) |
1836 | ret |
1837 | |
1838 | .p2align 4 |
1839 | L(fwd_write_140bytes): |
1840 | lddqu -140(%rsi), %xmm0 |
1841 | movdqu %xmm0, -140(%rdi) |
1842 | L(fwd_write_124bytes): |
1843 | lddqu -124(%rsi), %xmm0 |
1844 | movdqu %xmm0, -124(%rdi) |
1845 | L(fwd_write_108bytes): |
1846 | lddqu -108(%rsi), %xmm0 |
1847 | movdqu %xmm0, -108(%rdi) |
1848 | L(fwd_write_92bytes): |
1849 | lddqu -92(%rsi), %xmm0 |
1850 | movdqu %xmm0, -92(%rdi) |
1851 | L(fwd_write_76bytes): |
1852 | lddqu -76(%rsi), %xmm0 |
1853 | movdqu %xmm0, -76(%rdi) |
1854 | L(fwd_write_60bytes): |
1855 | lddqu -60(%rsi), %xmm0 |
1856 | movdqu %xmm0, -60(%rdi) |
1857 | L(fwd_write_44bytes): |
1858 | lddqu -44(%rsi), %xmm0 |
1859 | movdqu %xmm0, -44(%rdi) |
1860 | L(fwd_write_28bytes): |
1861 | lddqu -28(%rsi), %xmm0 |
1862 | lddqu -16(%rsi), %xmm1 |
1863 | movdqu %xmm0, -28(%rdi) |
1864 | movdqu %xmm1, -16(%rdi) |
1865 | ret |
1866 | |
1867 | .p2align 4 |
1868 | L(fwd_write_12bytes): |
1869 | mov -12(%rsi), %rdx |
1870 | mov -4(%rsi), %ecx |
1871 | mov %rdx, -12(%rdi) |
1872 | mov %ecx, -4(%rdi) |
1873 | ret |
1874 | |
1875 | .p2align 4 |
1876 | L(fwd_write_139bytes): |
1877 | lddqu -139(%rsi), %xmm0 |
1878 | movdqu %xmm0, -139(%rdi) |
1879 | L(fwd_write_123bytes): |
1880 | lddqu -123(%rsi), %xmm0 |
1881 | movdqu %xmm0, -123(%rdi) |
1882 | L(fwd_write_107bytes): |
1883 | lddqu -107(%rsi), %xmm0 |
1884 | movdqu %xmm0, -107(%rdi) |
1885 | L(fwd_write_91bytes): |
1886 | lddqu -91(%rsi), %xmm0 |
1887 | movdqu %xmm0, -91(%rdi) |
1888 | L(fwd_write_75bytes): |
1889 | lddqu -75(%rsi), %xmm0 |
1890 | movdqu %xmm0, -75(%rdi) |
1891 | L(fwd_write_59bytes): |
1892 | lddqu -59(%rsi), %xmm0 |
1893 | movdqu %xmm0, -59(%rdi) |
1894 | L(fwd_write_43bytes): |
1895 | lddqu -43(%rsi), %xmm0 |
1896 | movdqu %xmm0, -43(%rdi) |
1897 | L(fwd_write_27bytes): |
1898 | lddqu -27(%rsi), %xmm0 |
1899 | lddqu -16(%rsi), %xmm1 |
1900 | movdqu %xmm0, -27(%rdi) |
1901 | movdqu %xmm1, -16(%rdi) |
1902 | ret |
1903 | |
1904 | .p2align 4 |
1905 | L(fwd_write_11bytes): |
1906 | mov -11(%rsi), %rdx |
1907 | mov -4(%rsi), %ecx |
1908 | mov %rdx, -11(%rdi) |
1909 | mov %ecx, -4(%rdi) |
1910 | ret |
1911 | |
1912 | .p2align 4 |
1913 | L(fwd_write_138bytes): |
1914 | lddqu -138(%rsi), %xmm0 |
1915 | movdqu %xmm0, -138(%rdi) |
1916 | L(fwd_write_122bytes): |
1917 | lddqu -122(%rsi), %xmm0 |
1918 | movdqu %xmm0, -122(%rdi) |
1919 | L(fwd_write_106bytes): |
1920 | lddqu -106(%rsi), %xmm0 |
1921 | movdqu %xmm0, -106(%rdi) |
1922 | L(fwd_write_90bytes): |
1923 | lddqu -90(%rsi), %xmm0 |
1924 | movdqu %xmm0, -90(%rdi) |
1925 | L(fwd_write_74bytes): |
1926 | lddqu -74(%rsi), %xmm0 |
1927 | movdqu %xmm0, -74(%rdi) |
1928 | L(fwd_write_58bytes): |
1929 | lddqu -58(%rsi), %xmm0 |
1930 | movdqu %xmm0, -58(%rdi) |
1931 | L(fwd_write_42bytes): |
1932 | lddqu -42(%rsi), %xmm0 |
1933 | movdqu %xmm0, -42(%rdi) |
1934 | L(fwd_write_26bytes): |
1935 | lddqu -26(%rsi), %xmm0 |
1936 | lddqu -16(%rsi), %xmm1 |
1937 | movdqu %xmm0, -26(%rdi) |
1938 | movdqu %xmm1, -16(%rdi) |
1939 | ret |
1940 | |
1941 | .p2align 4 |
1942 | L(fwd_write_10bytes): |
1943 | mov -10(%rsi), %rdx |
1944 | mov -4(%rsi), %ecx |
1945 | mov %rdx, -10(%rdi) |
1946 | mov %ecx, -4(%rdi) |
1947 | ret |
1948 | |
1949 | .p2align 4 |
1950 | L(fwd_write_137bytes): |
1951 | lddqu -137(%rsi), %xmm0 |
1952 | movdqu %xmm0, -137(%rdi) |
1953 | L(fwd_write_121bytes): |
1954 | lddqu -121(%rsi), %xmm0 |
1955 | movdqu %xmm0, -121(%rdi) |
1956 | L(fwd_write_105bytes): |
1957 | lddqu -105(%rsi), %xmm0 |
1958 | movdqu %xmm0, -105(%rdi) |
1959 | L(fwd_write_89bytes): |
1960 | lddqu -89(%rsi), %xmm0 |
1961 | movdqu %xmm0, -89(%rdi) |
1962 | L(fwd_write_73bytes): |
1963 | lddqu -73(%rsi), %xmm0 |
1964 | movdqu %xmm0, -73(%rdi) |
1965 | L(fwd_write_57bytes): |
1966 | lddqu -57(%rsi), %xmm0 |
1967 | movdqu %xmm0, -57(%rdi) |
1968 | L(fwd_write_41bytes): |
1969 | lddqu -41(%rsi), %xmm0 |
1970 | movdqu %xmm0, -41(%rdi) |
1971 | L(fwd_write_25bytes): |
1972 | lddqu -25(%rsi), %xmm0 |
1973 | lddqu -16(%rsi), %xmm1 |
1974 | movdqu %xmm0, -25(%rdi) |
1975 | movdqu %xmm1, -16(%rdi) |
1976 | ret |
1977 | |
1978 | .p2align 4 |
1979 | L(fwd_write_9bytes): |
1980 | mov -9(%rsi), %rdx |
1981 | mov -4(%rsi), %ecx |
1982 | mov %rdx, -9(%rdi) |
1983 | mov %ecx, -4(%rdi) |
1984 | ret |
1985 | |
1986 | .p2align 4 |
1987 | L(fwd_write_136bytes): |
1988 | lddqu -136(%rsi), %xmm0 |
1989 | movdqu %xmm0, -136(%rdi) |
1990 | L(fwd_write_120bytes): |
1991 | lddqu -120(%rsi), %xmm0 |
1992 | movdqu %xmm0, -120(%rdi) |
1993 | L(fwd_write_104bytes): |
1994 | lddqu -104(%rsi), %xmm0 |
1995 | movdqu %xmm0, -104(%rdi) |
1996 | L(fwd_write_88bytes): |
1997 | lddqu -88(%rsi), %xmm0 |
1998 | movdqu %xmm0, -88(%rdi) |
1999 | L(fwd_write_72bytes): |
2000 | lddqu -72(%rsi), %xmm0 |
2001 | movdqu %xmm0, -72(%rdi) |
2002 | L(fwd_write_56bytes): |
2003 | lddqu -56(%rsi), %xmm0 |
2004 | movdqu %xmm0, -56(%rdi) |
2005 | L(fwd_write_40bytes): |
2006 | lddqu -40(%rsi), %xmm0 |
2007 | movdqu %xmm0, -40(%rdi) |
2008 | L(fwd_write_24bytes): |
2009 | lddqu -24(%rsi), %xmm0 |
2010 | lddqu -16(%rsi), %xmm1 |
2011 | movdqu %xmm0, -24(%rdi) |
2012 | movdqu %xmm1, -16(%rdi) |
2013 | ret |
2014 | |
2015 | .p2align 4 |
2016 | L(fwd_write_8bytes): |
2017 | mov -8(%rsi), %rdx |
2018 | mov %rdx, -8(%rdi) |
2019 | ret |
2020 | |
2021 | .p2align 4 |
2022 | L(fwd_write_135bytes): |
2023 | lddqu -135(%rsi), %xmm0 |
2024 | movdqu %xmm0, -135(%rdi) |
2025 | L(fwd_write_119bytes): |
2026 | lddqu -119(%rsi), %xmm0 |
2027 | movdqu %xmm0, -119(%rdi) |
2028 | L(fwd_write_103bytes): |
2029 | lddqu -103(%rsi), %xmm0 |
2030 | movdqu %xmm0, -103(%rdi) |
2031 | L(fwd_write_87bytes): |
2032 | lddqu -87(%rsi), %xmm0 |
2033 | movdqu %xmm0, -87(%rdi) |
2034 | L(fwd_write_71bytes): |
2035 | lddqu -71(%rsi), %xmm0 |
2036 | movdqu %xmm0, -71(%rdi) |
2037 | L(fwd_write_55bytes): |
2038 | lddqu -55(%rsi), %xmm0 |
2039 | movdqu %xmm0, -55(%rdi) |
2040 | L(fwd_write_39bytes): |
2041 | lddqu -39(%rsi), %xmm0 |
2042 | movdqu %xmm0, -39(%rdi) |
2043 | L(fwd_write_23bytes): |
2044 | lddqu -23(%rsi), %xmm0 |
2045 | lddqu -16(%rsi), %xmm1 |
2046 | movdqu %xmm0, -23(%rdi) |
2047 | movdqu %xmm1, -16(%rdi) |
2048 | ret |
2049 | |
2050 | .p2align 4 |
2051 | L(fwd_write_7bytes): |
2052 | mov -7(%rsi), %edx |
2053 | mov -4(%rsi), %ecx |
2054 | mov %edx, -7(%rdi) |
2055 | mov %ecx, -4(%rdi) |
2056 | ret |
2057 | |
2058 | .p2align 4 |
2059 | L(fwd_write_134bytes): |
2060 | lddqu -134(%rsi), %xmm0 |
2061 | movdqu %xmm0, -134(%rdi) |
2062 | L(fwd_write_118bytes): |
2063 | lddqu -118(%rsi), %xmm0 |
2064 | movdqu %xmm0, -118(%rdi) |
2065 | L(fwd_write_102bytes): |
2066 | lddqu -102(%rsi), %xmm0 |
2067 | movdqu %xmm0, -102(%rdi) |
2068 | L(fwd_write_86bytes): |
2069 | lddqu -86(%rsi), %xmm0 |
2070 | movdqu %xmm0, -86(%rdi) |
2071 | L(fwd_write_70bytes): |
2072 | lddqu -70(%rsi), %xmm0 |
2073 | movdqu %xmm0, -70(%rdi) |
2074 | L(fwd_write_54bytes): |
2075 | lddqu -54(%rsi), %xmm0 |
2076 | movdqu %xmm0, -54(%rdi) |
2077 | L(fwd_write_38bytes): |
2078 | lddqu -38(%rsi), %xmm0 |
2079 | movdqu %xmm0, -38(%rdi) |
2080 | L(fwd_write_22bytes): |
2081 | lddqu -22(%rsi), %xmm0 |
2082 | lddqu -16(%rsi), %xmm1 |
2083 | movdqu %xmm0, -22(%rdi) |
2084 | movdqu %xmm1, -16(%rdi) |
2085 | ret |
2086 | |
2087 | .p2align 4 |
2088 | L(fwd_write_6bytes): |
2089 | mov -6(%rsi), %edx |
2090 | mov -4(%rsi), %ecx |
2091 | mov %edx, -6(%rdi) |
2092 | mov %ecx, -4(%rdi) |
2093 | ret |
2094 | |
2095 | .p2align 4 |
2096 | L(fwd_write_133bytes): |
2097 | lddqu -133(%rsi), %xmm0 |
2098 | movdqu %xmm0, -133(%rdi) |
2099 | L(fwd_write_117bytes): |
2100 | lddqu -117(%rsi), %xmm0 |
2101 | movdqu %xmm0, -117(%rdi) |
2102 | L(fwd_write_101bytes): |
2103 | lddqu -101(%rsi), %xmm0 |
2104 | movdqu %xmm0, -101(%rdi) |
2105 | L(fwd_write_85bytes): |
2106 | lddqu -85(%rsi), %xmm0 |
2107 | movdqu %xmm0, -85(%rdi) |
2108 | L(fwd_write_69bytes): |
2109 | lddqu -69(%rsi), %xmm0 |
2110 | movdqu %xmm0, -69(%rdi) |
2111 | L(fwd_write_53bytes): |
2112 | lddqu -53(%rsi), %xmm0 |
2113 | movdqu %xmm0, -53(%rdi) |
2114 | L(fwd_write_37bytes): |
2115 | lddqu -37(%rsi), %xmm0 |
2116 | movdqu %xmm0, -37(%rdi) |
2117 | L(fwd_write_21bytes): |
2118 | lddqu -21(%rsi), %xmm0 |
2119 | lddqu -16(%rsi), %xmm1 |
2120 | movdqu %xmm0, -21(%rdi) |
2121 | movdqu %xmm1, -16(%rdi) |
2122 | ret |
2123 | |
2124 | .p2align 4 |
2125 | L(fwd_write_5bytes): |
2126 | mov -5(%rsi), %edx |
2127 | mov -4(%rsi), %ecx |
2128 | mov %edx, -5(%rdi) |
2129 | mov %ecx, -4(%rdi) |
2130 | ret |
2131 | |
2132 | .p2align 4 |
2133 | L(fwd_write_132bytes): |
2134 | lddqu -132(%rsi), %xmm0 |
2135 | movdqu %xmm0, -132(%rdi) |
2136 | L(fwd_write_116bytes): |
2137 | lddqu -116(%rsi), %xmm0 |
2138 | movdqu %xmm0, -116(%rdi) |
2139 | L(fwd_write_100bytes): |
2140 | lddqu -100(%rsi), %xmm0 |
2141 | movdqu %xmm0, -100(%rdi) |
2142 | L(fwd_write_84bytes): |
2143 | lddqu -84(%rsi), %xmm0 |
2144 | movdqu %xmm0, -84(%rdi) |
2145 | L(fwd_write_68bytes): |
2146 | lddqu -68(%rsi), %xmm0 |
2147 | movdqu %xmm0, -68(%rdi) |
2148 | L(fwd_write_52bytes): |
2149 | lddqu -52(%rsi), %xmm0 |
2150 | movdqu %xmm0, -52(%rdi) |
2151 | L(fwd_write_36bytes): |
2152 | lddqu -36(%rsi), %xmm0 |
2153 | movdqu %xmm0, -36(%rdi) |
2154 | L(fwd_write_20bytes): |
2155 | lddqu -20(%rsi), %xmm0 |
2156 | lddqu -16(%rsi), %xmm1 |
2157 | movdqu %xmm0, -20(%rdi) |
2158 | movdqu %xmm1, -16(%rdi) |
2159 | ret |
2160 | |
2161 | .p2align 4 |
2162 | L(fwd_write_4bytes): |
2163 | mov -4(%rsi), %edx |
2164 | mov %edx, -4(%rdi) |
2165 | ret |
2166 | |
2167 | .p2align 4 |
2168 | L(fwd_write_131bytes): |
2169 | lddqu -131(%rsi), %xmm0 |
2170 | movdqu %xmm0, -131(%rdi) |
2171 | L(fwd_write_115bytes): |
2172 | lddqu -115(%rsi), %xmm0 |
2173 | movdqu %xmm0, -115(%rdi) |
2174 | L(fwd_write_99bytes): |
2175 | lddqu -99(%rsi), %xmm0 |
2176 | movdqu %xmm0, -99(%rdi) |
2177 | L(fwd_write_83bytes): |
2178 | lddqu -83(%rsi), %xmm0 |
2179 | movdqu %xmm0, -83(%rdi) |
2180 | L(fwd_write_67bytes): |
2181 | lddqu -67(%rsi), %xmm0 |
2182 | movdqu %xmm0, -67(%rdi) |
2183 | L(fwd_write_51bytes): |
2184 | lddqu -51(%rsi), %xmm0 |
2185 | movdqu %xmm0, -51(%rdi) |
2186 | L(fwd_write_35bytes): |
2187 | lddqu -35(%rsi), %xmm0 |
2188 | movdqu %xmm0, -35(%rdi) |
2189 | L(fwd_write_19bytes): |
2190 | lddqu -19(%rsi), %xmm0 |
2191 | lddqu -16(%rsi), %xmm1 |
2192 | movdqu %xmm0, -19(%rdi) |
2193 | movdqu %xmm1, -16(%rdi) |
2194 | ret |
2195 | |
2196 | .p2align 4 |
2197 | L(fwd_write_3bytes): |
2198 | mov -3(%rsi), %dx |
2199 | mov -2(%rsi), %cx |
2200 | mov %dx, -3(%rdi) |
2201 | mov %cx, -2(%rdi) |
2202 | ret |
2203 | |
2204 | .p2align 4 |
2205 | L(fwd_write_130bytes): |
2206 | lddqu -130(%rsi), %xmm0 |
2207 | movdqu %xmm0, -130(%rdi) |
2208 | L(fwd_write_114bytes): |
2209 | lddqu -114(%rsi), %xmm0 |
2210 | movdqu %xmm0, -114(%rdi) |
2211 | L(fwd_write_98bytes): |
2212 | lddqu -98(%rsi), %xmm0 |
2213 | movdqu %xmm0, -98(%rdi) |
2214 | L(fwd_write_82bytes): |
2215 | lddqu -82(%rsi), %xmm0 |
2216 | movdqu %xmm0, -82(%rdi) |
2217 | L(fwd_write_66bytes): |
2218 | lddqu -66(%rsi), %xmm0 |
2219 | movdqu %xmm0, -66(%rdi) |
2220 | L(fwd_write_50bytes): |
2221 | lddqu -50(%rsi), %xmm0 |
2222 | movdqu %xmm0, -50(%rdi) |
2223 | L(fwd_write_34bytes): |
2224 | lddqu -34(%rsi), %xmm0 |
2225 | movdqu %xmm0, -34(%rdi) |
2226 | L(fwd_write_18bytes): |
2227 | lddqu -18(%rsi), %xmm0 |
2228 | lddqu -16(%rsi), %xmm1 |
2229 | movdqu %xmm0, -18(%rdi) |
2230 | movdqu %xmm1, -16(%rdi) |
2231 | ret |
2232 | |
2233 | .p2align 4 |
2234 | L(fwd_write_2bytes): |
2235 | movzwl -2(%rsi), %edx |
2236 | mov %dx, -2(%rdi) |
2237 | ret |
2238 | |
2239 | .p2align 4 |
2240 | L(fwd_write_129bytes): |
2241 | lddqu -129(%rsi), %xmm0 |
2242 | movdqu %xmm0, -129(%rdi) |
2243 | L(fwd_write_113bytes): |
2244 | lddqu -113(%rsi), %xmm0 |
2245 | movdqu %xmm0, -113(%rdi) |
2246 | L(fwd_write_97bytes): |
2247 | lddqu -97(%rsi), %xmm0 |
2248 | movdqu %xmm0, -97(%rdi) |
2249 | L(fwd_write_81bytes): |
2250 | lddqu -81(%rsi), %xmm0 |
2251 | movdqu %xmm0, -81(%rdi) |
2252 | L(fwd_write_65bytes): |
2253 | lddqu -65(%rsi), %xmm0 |
2254 | movdqu %xmm0, -65(%rdi) |
2255 | L(fwd_write_49bytes): |
2256 | lddqu -49(%rsi), %xmm0 |
2257 | movdqu %xmm0, -49(%rdi) |
2258 | L(fwd_write_33bytes): |
2259 | lddqu -33(%rsi), %xmm0 |
2260 | movdqu %xmm0, -33(%rdi) |
2261 | L(fwd_write_17bytes): |
2262 | lddqu -17(%rsi), %xmm0 |
2263 | lddqu -16(%rsi), %xmm1 |
2264 | movdqu %xmm0, -17(%rdi) |
2265 | movdqu %xmm1, -16(%rdi) |
2266 | ret |
2267 | |
2268 | .p2align 4 |
2269 | L(fwd_write_1bytes): |
2270 | movzbl -1(%rsi), %edx |
2271 | mov %dl, -1(%rdi) |
2272 | ret |
2273 | |
2274 | .p2align 4 |
2275 | L(bwd_write_128bytes): |
2276 | lddqu 112(%rsi), %xmm0 |
2277 | movdqu %xmm0, 112(%rdi) |
2278 | L(bwd_write_112bytes): |
2279 | lddqu 96(%rsi), %xmm0 |
2280 | movdqu %xmm0, 96(%rdi) |
2281 | L(bwd_write_96bytes): |
2282 | lddqu 80(%rsi), %xmm0 |
2283 | movdqu %xmm0, 80(%rdi) |
2284 | L(bwd_write_80bytes): |
2285 | lddqu 64(%rsi), %xmm0 |
2286 | movdqu %xmm0, 64(%rdi) |
2287 | L(bwd_write_64bytes): |
2288 | lddqu 48(%rsi), %xmm0 |
2289 | movdqu %xmm0, 48(%rdi) |
2290 | L(bwd_write_48bytes): |
2291 | lddqu 32(%rsi), %xmm0 |
2292 | movdqu %xmm0, 32(%rdi) |
2293 | L(bwd_write_32bytes): |
2294 | lddqu 16(%rsi), %xmm0 |
2295 | movdqu %xmm0, 16(%rdi) |
2296 | L(bwd_write_16bytes): |
2297 | lddqu (%rsi), %xmm0 |
2298 | movdqu %xmm0, (%rdi) |
2299 | L(bwd_write_0bytes): |
2300 | ret |
2301 | |
2302 | .p2align 4 |
2303 | L(bwd_write_143bytes): |
2304 | lddqu 127(%rsi), %xmm0 |
2305 | movdqu %xmm0, 127(%rdi) |
2306 | L(bwd_write_127bytes): |
2307 | lddqu 111(%rsi), %xmm0 |
2308 | movdqu %xmm0, 111(%rdi) |
2309 | L(bwd_write_111bytes): |
2310 | lddqu 95(%rsi), %xmm0 |
2311 | movdqu %xmm0, 95(%rdi) |
2312 | L(bwd_write_95bytes): |
2313 | lddqu 79(%rsi), %xmm0 |
2314 | movdqu %xmm0, 79(%rdi) |
2315 | L(bwd_write_79bytes): |
2316 | lddqu 63(%rsi), %xmm0 |
2317 | movdqu %xmm0, 63(%rdi) |
2318 | L(bwd_write_63bytes): |
2319 | lddqu 47(%rsi), %xmm0 |
2320 | movdqu %xmm0, 47(%rdi) |
2321 | L(bwd_write_47bytes): |
2322 | lddqu 31(%rsi), %xmm0 |
2323 | movdqu %xmm0, 31(%rdi) |
2324 | L(bwd_write_31bytes): |
2325 | lddqu 15(%rsi), %xmm0 |
2326 | lddqu (%rsi), %xmm1 |
2327 | movdqu %xmm0, 15(%rdi) |
2328 | movdqu %xmm1, (%rdi) |
2329 | ret |
2330 | |
2331 | |
2332 | .p2align 4 |
2333 | L(bwd_write_15bytes): |
2334 | mov 7(%rsi), %rdx |
2335 | mov (%rsi), %rcx |
2336 | mov %rdx, 7(%rdi) |
2337 | mov %rcx, (%rdi) |
2338 | ret |
2339 | |
2340 | .p2align 4 |
2341 | L(bwd_write_142bytes): |
2342 | lddqu 126(%rsi), %xmm0 |
2343 | movdqu %xmm0, 126(%rdi) |
2344 | L(bwd_write_126bytes): |
2345 | lddqu 110(%rsi), %xmm0 |
2346 | movdqu %xmm0, 110(%rdi) |
2347 | L(bwd_write_110bytes): |
2348 | lddqu 94(%rsi), %xmm0 |
2349 | movdqu %xmm0, 94(%rdi) |
2350 | L(bwd_write_94bytes): |
2351 | lddqu 78(%rsi), %xmm0 |
2352 | movdqu %xmm0, 78(%rdi) |
2353 | L(bwd_write_78bytes): |
2354 | lddqu 62(%rsi), %xmm0 |
2355 | movdqu %xmm0, 62(%rdi) |
2356 | L(bwd_write_62bytes): |
2357 | lddqu 46(%rsi), %xmm0 |
2358 | movdqu %xmm0, 46(%rdi) |
2359 | L(bwd_write_46bytes): |
2360 | lddqu 30(%rsi), %xmm0 |
2361 | movdqu %xmm0, 30(%rdi) |
2362 | L(bwd_write_30bytes): |
2363 | lddqu 14(%rsi), %xmm0 |
2364 | lddqu (%rsi), %xmm1 |
2365 | movdqu %xmm0, 14(%rdi) |
2366 | movdqu %xmm1, (%rdi) |
2367 | ret |
2368 | |
2369 | .p2align 4 |
2370 | L(bwd_write_14bytes): |
2371 | mov 6(%rsi), %rdx |
2372 | mov (%rsi), %rcx |
2373 | mov %rdx, 6(%rdi) |
2374 | mov %rcx, (%rdi) |
2375 | ret |
2376 | |
2377 | .p2align 4 |
2378 | L(bwd_write_141bytes): |
2379 | lddqu 125(%rsi), %xmm0 |
2380 | movdqu %xmm0, 125(%rdi) |
2381 | L(bwd_write_125bytes): |
2382 | lddqu 109(%rsi), %xmm0 |
2383 | movdqu %xmm0, 109(%rdi) |
2384 | L(bwd_write_109bytes): |
2385 | lddqu 93(%rsi), %xmm0 |
2386 | movdqu %xmm0, 93(%rdi) |
2387 | L(bwd_write_93bytes): |
2388 | lddqu 77(%rsi), %xmm0 |
2389 | movdqu %xmm0, 77(%rdi) |
2390 | L(bwd_write_77bytes): |
2391 | lddqu 61(%rsi), %xmm0 |
2392 | movdqu %xmm0, 61(%rdi) |
2393 | L(bwd_write_61bytes): |
2394 | lddqu 45(%rsi), %xmm0 |
2395 | movdqu %xmm0, 45(%rdi) |
2396 | L(bwd_write_45bytes): |
2397 | lddqu 29(%rsi), %xmm0 |
2398 | movdqu %xmm0, 29(%rdi) |
2399 | L(bwd_write_29bytes): |
2400 | lddqu 13(%rsi), %xmm0 |
2401 | lddqu (%rsi), %xmm1 |
2402 | movdqu %xmm0, 13(%rdi) |
2403 | movdqu %xmm1, (%rdi) |
2404 | ret |
2405 | |
2406 | .p2align 4 |
2407 | L(bwd_write_13bytes): |
2408 | mov 5(%rsi), %rdx |
2409 | mov (%rsi), %rcx |
2410 | mov %rdx, 5(%rdi) |
2411 | mov %rcx, (%rdi) |
2412 | ret |
2413 | |
2414 | .p2align 4 |
2415 | L(bwd_write_140bytes): |
2416 | lddqu 124(%rsi), %xmm0 |
2417 | movdqu %xmm0, 124(%rdi) |
2418 | L(bwd_write_124bytes): |
2419 | lddqu 108(%rsi), %xmm0 |
2420 | movdqu %xmm0, 108(%rdi) |
2421 | L(bwd_write_108bytes): |
2422 | lddqu 92(%rsi), %xmm0 |
2423 | movdqu %xmm0, 92(%rdi) |
2424 | L(bwd_write_92bytes): |
2425 | lddqu 76(%rsi), %xmm0 |
2426 | movdqu %xmm0, 76(%rdi) |
2427 | L(bwd_write_76bytes): |
2428 | lddqu 60(%rsi), %xmm0 |
2429 | movdqu %xmm0, 60(%rdi) |
2430 | L(bwd_write_60bytes): |
2431 | lddqu 44(%rsi), %xmm0 |
2432 | movdqu %xmm0, 44(%rdi) |
2433 | L(bwd_write_44bytes): |
2434 | lddqu 28(%rsi), %xmm0 |
2435 | movdqu %xmm0, 28(%rdi) |
2436 | L(bwd_write_28bytes): |
2437 | lddqu 12(%rsi), %xmm0 |
2438 | lddqu (%rsi), %xmm1 |
2439 | movdqu %xmm0, 12(%rdi) |
2440 | movdqu %xmm1, (%rdi) |
2441 | ret |
2442 | |
2443 | .p2align 4 |
2444 | L(bwd_write_12bytes): |
2445 | mov 4(%rsi), %rdx |
2446 | mov (%rsi), %rcx |
2447 | mov %rdx, 4(%rdi) |
2448 | mov %rcx, (%rdi) |
2449 | ret |
2450 | |
2451 | .p2align 4 |
2452 | L(bwd_write_139bytes): |
2453 | lddqu 123(%rsi), %xmm0 |
2454 | movdqu %xmm0, 123(%rdi) |
2455 | L(bwd_write_123bytes): |
2456 | lddqu 107(%rsi), %xmm0 |
2457 | movdqu %xmm0, 107(%rdi) |
2458 | L(bwd_write_107bytes): |
2459 | lddqu 91(%rsi), %xmm0 |
2460 | movdqu %xmm0, 91(%rdi) |
2461 | L(bwd_write_91bytes): |
2462 | lddqu 75(%rsi), %xmm0 |
2463 | movdqu %xmm0, 75(%rdi) |
2464 | L(bwd_write_75bytes): |
2465 | lddqu 59(%rsi), %xmm0 |
2466 | movdqu %xmm0, 59(%rdi) |
2467 | L(bwd_write_59bytes): |
2468 | lddqu 43(%rsi), %xmm0 |
2469 | movdqu %xmm0, 43(%rdi) |
2470 | L(bwd_write_43bytes): |
2471 | lddqu 27(%rsi), %xmm0 |
2472 | movdqu %xmm0, 27(%rdi) |
2473 | L(bwd_write_27bytes): |
2474 | lddqu 11(%rsi), %xmm0 |
2475 | lddqu (%rsi), %xmm1 |
2476 | movdqu %xmm0, 11(%rdi) |
2477 | movdqu %xmm1, (%rdi) |
2478 | ret |
2479 | |
2480 | .p2align 4 |
2481 | L(bwd_write_11bytes): |
2482 | mov 3(%rsi), %rdx |
2483 | mov (%rsi), %rcx |
2484 | mov %rdx, 3(%rdi) |
2485 | mov %rcx, (%rdi) |
2486 | ret |
2487 | |
2488 | .p2align 4 |
2489 | L(bwd_write_138bytes): |
2490 | lddqu 122(%rsi), %xmm0 |
2491 | movdqu %xmm0, 122(%rdi) |
2492 | L(bwd_write_122bytes): |
2493 | lddqu 106(%rsi), %xmm0 |
2494 | movdqu %xmm0, 106(%rdi) |
2495 | L(bwd_write_106bytes): |
2496 | lddqu 90(%rsi), %xmm0 |
2497 | movdqu %xmm0, 90(%rdi) |
2498 | L(bwd_write_90bytes): |
2499 | lddqu 74(%rsi), %xmm0 |
2500 | movdqu %xmm0, 74(%rdi) |
2501 | L(bwd_write_74bytes): |
2502 | lddqu 58(%rsi), %xmm0 |
2503 | movdqu %xmm0, 58(%rdi) |
2504 | L(bwd_write_58bytes): |
2505 | lddqu 42(%rsi), %xmm0 |
2506 | movdqu %xmm0, 42(%rdi) |
2507 | L(bwd_write_42bytes): |
2508 | lddqu 26(%rsi), %xmm0 |
2509 | movdqu %xmm0, 26(%rdi) |
2510 | L(bwd_write_26bytes): |
2511 | lddqu 10(%rsi), %xmm0 |
2512 | lddqu (%rsi), %xmm1 |
2513 | movdqu %xmm0, 10(%rdi) |
2514 | movdqu %xmm1, (%rdi) |
2515 | ret |
2516 | |
2517 | .p2align 4 |
2518 | L(bwd_write_10bytes): |
2519 | mov 2(%rsi), %rdx |
2520 | mov (%rsi), %rcx |
2521 | mov %rdx, 2(%rdi) |
2522 | mov %rcx, (%rdi) |
2523 | ret |
2524 | |
2525 | .p2align 4 |
2526 | L(bwd_write_137bytes): |
2527 | lddqu 121(%rsi), %xmm0 |
2528 | movdqu %xmm0, 121(%rdi) |
2529 | L(bwd_write_121bytes): |
2530 | lddqu 105(%rsi), %xmm0 |
2531 | movdqu %xmm0, 105(%rdi) |
2532 | L(bwd_write_105bytes): |
2533 | lddqu 89(%rsi), %xmm0 |
2534 | movdqu %xmm0, 89(%rdi) |
2535 | L(bwd_write_89bytes): |
2536 | lddqu 73(%rsi), %xmm0 |
2537 | movdqu %xmm0, 73(%rdi) |
2538 | L(bwd_write_73bytes): |
2539 | lddqu 57(%rsi), %xmm0 |
2540 | movdqu %xmm0, 57(%rdi) |
2541 | L(bwd_write_57bytes): |
2542 | lddqu 41(%rsi), %xmm0 |
2543 | movdqu %xmm0, 41(%rdi) |
2544 | L(bwd_write_41bytes): |
2545 | lddqu 25(%rsi), %xmm0 |
2546 | movdqu %xmm0, 25(%rdi) |
2547 | L(bwd_write_25bytes): |
2548 | lddqu 9(%rsi), %xmm0 |
2549 | lddqu (%rsi), %xmm1 |
2550 | movdqu %xmm0, 9(%rdi) |
2551 | movdqu %xmm1, (%rdi) |
2552 | ret |
2553 | |
2554 | .p2align 4 |
2555 | L(bwd_write_9bytes): |
2556 | mov 1(%rsi), %rdx |
2557 | mov (%rsi), %rcx |
2558 | mov %rdx, 1(%rdi) |
2559 | mov %rcx, (%rdi) |
2560 | ret |
2561 | |
2562 | .p2align 4 |
2563 | L(bwd_write_136bytes): |
2564 | lddqu 120(%rsi), %xmm0 |
2565 | movdqu %xmm0, 120(%rdi) |
2566 | L(bwd_write_120bytes): |
2567 | lddqu 104(%rsi), %xmm0 |
2568 | movdqu %xmm0, 104(%rdi) |
2569 | L(bwd_write_104bytes): |
2570 | lddqu 88(%rsi), %xmm0 |
2571 | movdqu %xmm0, 88(%rdi) |
2572 | L(bwd_write_88bytes): |
2573 | lddqu 72(%rsi), %xmm0 |
2574 | movdqu %xmm0, 72(%rdi) |
2575 | L(bwd_write_72bytes): |
2576 | lddqu 56(%rsi), %xmm0 |
2577 | movdqu %xmm0, 56(%rdi) |
2578 | L(bwd_write_56bytes): |
2579 | lddqu 40(%rsi), %xmm0 |
2580 | movdqu %xmm0, 40(%rdi) |
2581 | L(bwd_write_40bytes): |
2582 | lddqu 24(%rsi), %xmm0 |
2583 | movdqu %xmm0, 24(%rdi) |
2584 | L(bwd_write_24bytes): |
2585 | lddqu 8(%rsi), %xmm0 |
2586 | lddqu (%rsi), %xmm1 |
2587 | movdqu %xmm0, 8(%rdi) |
2588 | movdqu %xmm1, (%rdi) |
2589 | ret |
2590 | |
2591 | .p2align 4 |
2592 | L(bwd_write_8bytes): |
2593 | mov (%rsi), %rdx |
2594 | mov %rdx, (%rdi) |
2595 | ret |
2596 | |
2597 | .p2align 4 |
2598 | L(bwd_write_135bytes): |
2599 | lddqu 119(%rsi), %xmm0 |
2600 | movdqu %xmm0, 119(%rdi) |
2601 | L(bwd_write_119bytes): |
2602 | lddqu 103(%rsi), %xmm0 |
2603 | movdqu %xmm0, 103(%rdi) |
2604 | L(bwd_write_103bytes): |
2605 | lddqu 87(%rsi), %xmm0 |
2606 | movdqu %xmm0, 87(%rdi) |
2607 | L(bwd_write_87bytes): |
2608 | lddqu 71(%rsi), %xmm0 |
2609 | movdqu %xmm0, 71(%rdi) |
2610 | L(bwd_write_71bytes): |
2611 | lddqu 55(%rsi), %xmm0 |
2612 | movdqu %xmm0, 55(%rdi) |
2613 | L(bwd_write_55bytes): |
2614 | lddqu 39(%rsi), %xmm0 |
2615 | movdqu %xmm0, 39(%rdi) |
2616 | L(bwd_write_39bytes): |
2617 | lddqu 23(%rsi), %xmm0 |
2618 | movdqu %xmm0, 23(%rdi) |
2619 | L(bwd_write_23bytes): |
2620 | lddqu 7(%rsi), %xmm0 |
2621 | lddqu (%rsi), %xmm1 |
2622 | movdqu %xmm0, 7(%rdi) |
2623 | movdqu %xmm1, (%rdi) |
2624 | ret |
2625 | |
2626 | .p2align 4 |
2627 | L(bwd_write_7bytes): |
2628 | mov 3(%rsi), %edx |
2629 | mov (%rsi), %ecx |
2630 | mov %edx, 3(%rdi) |
2631 | mov %ecx, (%rdi) |
2632 | ret |
2633 | |
2634 | .p2align 4 |
2635 | L(bwd_write_134bytes): |
2636 | lddqu 118(%rsi), %xmm0 |
2637 | movdqu %xmm0, 118(%rdi) |
2638 | L(bwd_write_118bytes): |
2639 | lddqu 102(%rsi), %xmm0 |
2640 | movdqu %xmm0, 102(%rdi) |
2641 | L(bwd_write_102bytes): |
2642 | lddqu 86(%rsi), %xmm0 |
2643 | movdqu %xmm0, 86(%rdi) |
2644 | L(bwd_write_86bytes): |
2645 | lddqu 70(%rsi), %xmm0 |
2646 | movdqu %xmm0, 70(%rdi) |
2647 | L(bwd_write_70bytes): |
2648 | lddqu 54(%rsi), %xmm0 |
2649 | movdqu %xmm0, 54(%rdi) |
2650 | L(bwd_write_54bytes): |
2651 | lddqu 38(%rsi), %xmm0 |
2652 | movdqu %xmm0, 38(%rdi) |
2653 | L(bwd_write_38bytes): |
2654 | lddqu 22(%rsi), %xmm0 |
2655 | movdqu %xmm0, 22(%rdi) |
2656 | L(bwd_write_22bytes): |
2657 | lddqu 6(%rsi), %xmm0 |
2658 | lddqu (%rsi), %xmm1 |
2659 | movdqu %xmm0, 6(%rdi) |
2660 | movdqu %xmm1, (%rdi) |
2661 | ret |
2662 | |
2663 | .p2align 4 |
2664 | L(bwd_write_6bytes): |
2665 | mov 2(%rsi), %edx |
2666 | mov (%rsi), %ecx |
2667 | mov %edx, 2(%rdi) |
2668 | mov %ecx, (%rdi) |
2669 | ret |
2670 | |
2671 | .p2align 4 |
2672 | L(bwd_write_133bytes): |
2673 | lddqu 117(%rsi), %xmm0 |
2674 | movdqu %xmm0, 117(%rdi) |
2675 | L(bwd_write_117bytes): |
2676 | lddqu 101(%rsi), %xmm0 |
2677 | movdqu %xmm0, 101(%rdi) |
2678 | L(bwd_write_101bytes): |
2679 | lddqu 85(%rsi), %xmm0 |
2680 | movdqu %xmm0, 85(%rdi) |
2681 | L(bwd_write_85bytes): |
2682 | lddqu 69(%rsi), %xmm0 |
2683 | movdqu %xmm0, 69(%rdi) |
2684 | L(bwd_write_69bytes): |
2685 | lddqu 53(%rsi), %xmm0 |
2686 | movdqu %xmm0, 53(%rdi) |
2687 | L(bwd_write_53bytes): |
2688 | lddqu 37(%rsi), %xmm0 |
2689 | movdqu %xmm0, 37(%rdi) |
2690 | L(bwd_write_37bytes): |
2691 | lddqu 21(%rsi), %xmm0 |
2692 | movdqu %xmm0, 21(%rdi) |
2693 | L(bwd_write_21bytes): |
2694 | lddqu 5(%rsi), %xmm0 |
2695 | lddqu (%rsi), %xmm1 |
2696 | movdqu %xmm0, 5(%rdi) |
2697 | movdqu %xmm1, (%rdi) |
2698 | ret |
2699 | |
2700 | .p2align 4 |
2701 | L(bwd_write_5bytes): |
2702 | mov 1(%rsi), %edx |
2703 | mov (%rsi), %ecx |
2704 | mov %edx, 1(%rdi) |
2705 | mov %ecx, (%rdi) |
2706 | ret |
2707 | |
2708 | .p2align 4 |
2709 | L(bwd_write_132bytes): |
2710 | lddqu 116(%rsi), %xmm0 |
2711 | movdqu %xmm0, 116(%rdi) |
2712 | L(bwd_write_116bytes): |
2713 | lddqu 100(%rsi), %xmm0 |
2714 | movdqu %xmm0, 100(%rdi) |
2715 | L(bwd_write_100bytes): |
2716 | lddqu 84(%rsi), %xmm0 |
2717 | movdqu %xmm0, 84(%rdi) |
2718 | L(bwd_write_84bytes): |
2719 | lddqu 68(%rsi), %xmm0 |
2720 | movdqu %xmm0, 68(%rdi) |
2721 | L(bwd_write_68bytes): |
2722 | lddqu 52(%rsi), %xmm0 |
2723 | movdqu %xmm0, 52(%rdi) |
2724 | L(bwd_write_52bytes): |
2725 | lddqu 36(%rsi), %xmm0 |
2726 | movdqu %xmm0, 36(%rdi) |
2727 | L(bwd_write_36bytes): |
2728 | lddqu 20(%rsi), %xmm0 |
2729 | movdqu %xmm0, 20(%rdi) |
2730 | L(bwd_write_20bytes): |
2731 | lddqu 4(%rsi), %xmm0 |
2732 | lddqu (%rsi), %xmm1 |
2733 | movdqu %xmm0, 4(%rdi) |
2734 | movdqu %xmm1, (%rdi) |
2735 | ret |
2736 | |
2737 | .p2align 4 |
2738 | L(bwd_write_4bytes): |
2739 | mov (%rsi), %edx |
2740 | mov %edx, (%rdi) |
2741 | ret |
2742 | |
2743 | .p2align 4 |
2744 | L(bwd_write_131bytes): |
2745 | lddqu 115(%rsi), %xmm0 |
2746 | movdqu %xmm0, 115(%rdi) |
2747 | L(bwd_write_115bytes): |
2748 | lddqu 99(%rsi), %xmm0 |
2749 | movdqu %xmm0, 99(%rdi) |
2750 | L(bwd_write_99bytes): |
2751 | lddqu 83(%rsi), %xmm0 |
2752 | movdqu %xmm0, 83(%rdi) |
2753 | L(bwd_write_83bytes): |
2754 | lddqu 67(%rsi), %xmm0 |
2755 | movdqu %xmm0, 67(%rdi) |
2756 | L(bwd_write_67bytes): |
2757 | lddqu 51(%rsi), %xmm0 |
2758 | movdqu %xmm0, 51(%rdi) |
2759 | L(bwd_write_51bytes): |
2760 | lddqu 35(%rsi), %xmm0 |
2761 | movdqu %xmm0, 35(%rdi) |
2762 | L(bwd_write_35bytes): |
2763 | lddqu 19(%rsi), %xmm0 |
2764 | movdqu %xmm0, 19(%rdi) |
2765 | L(bwd_write_19bytes): |
2766 | lddqu 3(%rsi), %xmm0 |
2767 | lddqu (%rsi), %xmm1 |
2768 | movdqu %xmm0, 3(%rdi) |
2769 | movdqu %xmm1, (%rdi) |
2770 | ret |
2771 | |
2772 | .p2align 4 |
2773 | L(bwd_write_3bytes): |
2774 | mov 1(%rsi), %dx |
2775 | mov (%rsi), %cx |
2776 | mov %dx, 1(%rdi) |
2777 | mov %cx, (%rdi) |
2778 | ret |
2779 | |
2780 | .p2align 4 |
2781 | L(bwd_write_130bytes): |
2782 | lddqu 114(%rsi), %xmm0 |
2783 | movdqu %xmm0, 114(%rdi) |
2784 | L(bwd_write_114bytes): |
2785 | lddqu 98(%rsi), %xmm0 |
2786 | movdqu %xmm0, 98(%rdi) |
2787 | L(bwd_write_98bytes): |
2788 | lddqu 82(%rsi), %xmm0 |
2789 | movdqu %xmm0, 82(%rdi) |
2790 | L(bwd_write_82bytes): |
2791 | lddqu 66(%rsi), %xmm0 |
2792 | movdqu %xmm0, 66(%rdi) |
2793 | L(bwd_write_66bytes): |
2794 | lddqu 50(%rsi), %xmm0 |
2795 | movdqu %xmm0, 50(%rdi) |
2796 | L(bwd_write_50bytes): |
2797 | lddqu 34(%rsi), %xmm0 |
2798 | movdqu %xmm0, 34(%rdi) |
2799 | L(bwd_write_34bytes): |
2800 | lddqu 18(%rsi), %xmm0 |
2801 | movdqu %xmm0, 18(%rdi) |
2802 | L(bwd_write_18bytes): |
2803 | lddqu 2(%rsi), %xmm0 |
2804 | lddqu (%rsi), %xmm1 |
2805 | movdqu %xmm0, 2(%rdi) |
2806 | movdqu %xmm1, (%rdi) |
2807 | ret |
2808 | |
2809 | .p2align 4 |
2810 | L(bwd_write_2bytes): |
2811 | movzwl (%rsi), %edx |
2812 | mov %dx, (%rdi) |
2813 | ret |
2814 | |
2815 | .p2align 4 |
2816 | L(bwd_write_129bytes): |
2817 | lddqu 113(%rsi), %xmm0 |
2818 | movdqu %xmm0, 113(%rdi) |
2819 | L(bwd_write_113bytes): |
2820 | lddqu 97(%rsi), %xmm0 |
2821 | movdqu %xmm0, 97(%rdi) |
2822 | L(bwd_write_97bytes): |
2823 | lddqu 81(%rsi), %xmm0 |
2824 | movdqu %xmm0, 81(%rdi) |
2825 | L(bwd_write_81bytes): |
2826 | lddqu 65(%rsi), %xmm0 |
2827 | movdqu %xmm0, 65(%rdi) |
2828 | L(bwd_write_65bytes): |
2829 | lddqu 49(%rsi), %xmm0 |
2830 | movdqu %xmm0, 49(%rdi) |
2831 | L(bwd_write_49bytes): |
2832 | lddqu 33(%rsi), %xmm0 |
2833 | movdqu %xmm0, 33(%rdi) |
2834 | L(bwd_write_33bytes): |
2835 | lddqu 17(%rsi), %xmm0 |
2836 | movdqu %xmm0, 17(%rdi) |
2837 | L(bwd_write_17bytes): |
2838 | lddqu 1(%rsi), %xmm0 |
2839 | lddqu (%rsi), %xmm1 |
2840 | movdqu %xmm0, 1(%rdi) |
2841 | movdqu %xmm1, (%rdi) |
2842 | ret |
2843 | |
2844 | .p2align 4 |
2845 | L(bwd_write_1bytes): |
2846 | movzbl (%rsi), %edx |
2847 | mov %dl, (%rdi) |
2848 | ret |
2849 | |
2850 | END (MEMCPY) |
2851 | |
2852 | .section .rodata.ssse3,"a" ,@progbits |
2853 | .p2align 3 |
2854 | L(table_144_bytes_bwd): |
2855 | .int JMPTBL (L(bwd_write_0bytes), L(table_144_bytes_bwd)) |
2856 | .int JMPTBL (L(bwd_write_1bytes), L(table_144_bytes_bwd)) |
2857 | .int JMPTBL (L(bwd_write_2bytes), L(table_144_bytes_bwd)) |
2858 | .int JMPTBL (L(bwd_write_3bytes), L(table_144_bytes_bwd)) |
2859 | .int JMPTBL (L(bwd_write_4bytes), L(table_144_bytes_bwd)) |
2860 | .int JMPTBL (L(bwd_write_5bytes), L(table_144_bytes_bwd)) |
2861 | .int JMPTBL (L(bwd_write_6bytes), L(table_144_bytes_bwd)) |
2862 | .int JMPTBL (L(bwd_write_7bytes), L(table_144_bytes_bwd)) |
2863 | .int JMPTBL (L(bwd_write_8bytes), L(table_144_bytes_bwd)) |
2864 | .int JMPTBL (L(bwd_write_9bytes), L(table_144_bytes_bwd)) |
2865 | .int JMPTBL (L(bwd_write_10bytes), L(table_144_bytes_bwd)) |
2866 | .int JMPTBL (L(bwd_write_11bytes), L(table_144_bytes_bwd)) |
2867 | .int JMPTBL (L(bwd_write_12bytes), L(table_144_bytes_bwd)) |
2868 | .int JMPTBL (L(bwd_write_13bytes), L(table_144_bytes_bwd)) |
2869 | .int JMPTBL (L(bwd_write_14bytes), L(table_144_bytes_bwd)) |
2870 | .int JMPTBL (L(bwd_write_15bytes), L(table_144_bytes_bwd)) |
2871 | .int JMPTBL (L(bwd_write_16bytes), L(table_144_bytes_bwd)) |
2872 | .int JMPTBL (L(bwd_write_17bytes), L(table_144_bytes_bwd)) |
2873 | .int JMPTBL (L(bwd_write_18bytes), L(table_144_bytes_bwd)) |
2874 | .int JMPTBL (L(bwd_write_19bytes), L(table_144_bytes_bwd)) |
2875 | .int JMPTBL (L(bwd_write_20bytes), L(table_144_bytes_bwd)) |
2876 | .int JMPTBL (L(bwd_write_21bytes), L(table_144_bytes_bwd)) |
2877 | .int JMPTBL (L(bwd_write_22bytes), L(table_144_bytes_bwd)) |
2878 | .int JMPTBL (L(bwd_write_23bytes), L(table_144_bytes_bwd)) |
2879 | .int JMPTBL (L(bwd_write_24bytes), L(table_144_bytes_bwd)) |
2880 | .int JMPTBL (L(bwd_write_25bytes), L(table_144_bytes_bwd)) |
2881 | .int JMPTBL (L(bwd_write_26bytes), L(table_144_bytes_bwd)) |
2882 | .int JMPTBL (L(bwd_write_27bytes), L(table_144_bytes_bwd)) |
2883 | .int JMPTBL (L(bwd_write_28bytes), L(table_144_bytes_bwd)) |
2884 | .int JMPTBL (L(bwd_write_29bytes), L(table_144_bytes_bwd)) |
2885 | .int JMPTBL (L(bwd_write_30bytes), L(table_144_bytes_bwd)) |
2886 | .int JMPTBL (L(bwd_write_31bytes), L(table_144_bytes_bwd)) |
2887 | .int JMPTBL (L(bwd_write_32bytes), L(table_144_bytes_bwd)) |
2888 | .int JMPTBL (L(bwd_write_33bytes), L(table_144_bytes_bwd)) |
2889 | .int JMPTBL (L(bwd_write_34bytes), L(table_144_bytes_bwd)) |
2890 | .int JMPTBL (L(bwd_write_35bytes), L(table_144_bytes_bwd)) |
2891 | .int JMPTBL (L(bwd_write_36bytes), L(table_144_bytes_bwd)) |
2892 | .int JMPTBL (L(bwd_write_37bytes), L(table_144_bytes_bwd)) |
2893 | .int JMPTBL (L(bwd_write_38bytes), L(table_144_bytes_bwd)) |
2894 | .int JMPTBL (L(bwd_write_39bytes), L(table_144_bytes_bwd)) |
2895 | .int JMPTBL (L(bwd_write_40bytes), L(table_144_bytes_bwd)) |
2896 | .int JMPTBL (L(bwd_write_41bytes), L(table_144_bytes_bwd)) |
2897 | .int JMPTBL (L(bwd_write_42bytes), L(table_144_bytes_bwd)) |
2898 | .int JMPTBL (L(bwd_write_43bytes), L(table_144_bytes_bwd)) |
2899 | .int JMPTBL (L(bwd_write_44bytes), L(table_144_bytes_bwd)) |
2900 | .int JMPTBL (L(bwd_write_45bytes), L(table_144_bytes_bwd)) |
2901 | .int JMPTBL (L(bwd_write_46bytes), L(table_144_bytes_bwd)) |
2902 | .int JMPTBL (L(bwd_write_47bytes), L(table_144_bytes_bwd)) |
2903 | .int JMPTBL (L(bwd_write_48bytes), L(table_144_bytes_bwd)) |
2904 | .int JMPTBL (L(bwd_write_49bytes), L(table_144_bytes_bwd)) |
2905 | .int JMPTBL (L(bwd_write_50bytes), L(table_144_bytes_bwd)) |
2906 | .int JMPTBL (L(bwd_write_51bytes), L(table_144_bytes_bwd)) |
2907 | .int JMPTBL (L(bwd_write_52bytes), L(table_144_bytes_bwd)) |
2908 | .int JMPTBL (L(bwd_write_53bytes), L(table_144_bytes_bwd)) |
2909 | .int JMPTBL (L(bwd_write_54bytes), L(table_144_bytes_bwd)) |
2910 | .int JMPTBL (L(bwd_write_55bytes), L(table_144_bytes_bwd)) |
2911 | .int JMPTBL (L(bwd_write_56bytes), L(table_144_bytes_bwd)) |
2912 | .int JMPTBL (L(bwd_write_57bytes), L(table_144_bytes_bwd)) |
2913 | .int JMPTBL (L(bwd_write_58bytes), L(table_144_bytes_bwd)) |
2914 | .int JMPTBL (L(bwd_write_59bytes), L(table_144_bytes_bwd)) |
2915 | .int JMPTBL (L(bwd_write_60bytes), L(table_144_bytes_bwd)) |
2916 | .int JMPTBL (L(bwd_write_61bytes), L(table_144_bytes_bwd)) |
2917 | .int JMPTBL (L(bwd_write_62bytes), L(table_144_bytes_bwd)) |
2918 | .int JMPTBL (L(bwd_write_63bytes), L(table_144_bytes_bwd)) |
2919 | .int JMPTBL (L(bwd_write_64bytes), L(table_144_bytes_bwd)) |
2920 | .int JMPTBL (L(bwd_write_65bytes), L(table_144_bytes_bwd)) |
2921 | .int JMPTBL (L(bwd_write_66bytes), L(table_144_bytes_bwd)) |
2922 | .int JMPTBL (L(bwd_write_67bytes), L(table_144_bytes_bwd)) |
2923 | .int JMPTBL (L(bwd_write_68bytes), L(table_144_bytes_bwd)) |
2924 | .int JMPTBL (L(bwd_write_69bytes), L(table_144_bytes_bwd)) |
2925 | .int JMPTBL (L(bwd_write_70bytes), L(table_144_bytes_bwd)) |
2926 | .int JMPTBL (L(bwd_write_71bytes), L(table_144_bytes_bwd)) |
2927 | .int JMPTBL (L(bwd_write_72bytes), L(table_144_bytes_bwd)) |
2928 | .int JMPTBL (L(bwd_write_73bytes), L(table_144_bytes_bwd)) |
2929 | .int JMPTBL (L(bwd_write_74bytes), L(table_144_bytes_bwd)) |
2930 | .int JMPTBL (L(bwd_write_75bytes), L(table_144_bytes_bwd)) |
2931 | .int JMPTBL (L(bwd_write_76bytes), L(table_144_bytes_bwd)) |
2932 | .int JMPTBL (L(bwd_write_77bytes), L(table_144_bytes_bwd)) |
2933 | .int JMPTBL (L(bwd_write_78bytes), L(table_144_bytes_bwd)) |
2934 | .int JMPTBL (L(bwd_write_79bytes), L(table_144_bytes_bwd)) |
2935 | .int JMPTBL (L(bwd_write_80bytes), L(table_144_bytes_bwd)) |
2936 | .int JMPTBL (L(bwd_write_81bytes), L(table_144_bytes_bwd)) |
2937 | .int JMPTBL (L(bwd_write_82bytes), L(table_144_bytes_bwd)) |
2938 | .int JMPTBL (L(bwd_write_83bytes), L(table_144_bytes_bwd)) |
2939 | .int JMPTBL (L(bwd_write_84bytes), L(table_144_bytes_bwd)) |
2940 | .int JMPTBL (L(bwd_write_85bytes), L(table_144_bytes_bwd)) |
2941 | .int JMPTBL (L(bwd_write_86bytes), L(table_144_bytes_bwd)) |
2942 | .int JMPTBL (L(bwd_write_87bytes), L(table_144_bytes_bwd)) |
2943 | .int JMPTBL (L(bwd_write_88bytes), L(table_144_bytes_bwd)) |
2944 | .int JMPTBL (L(bwd_write_89bytes), L(table_144_bytes_bwd)) |
2945 | .int JMPTBL (L(bwd_write_90bytes), L(table_144_bytes_bwd)) |
2946 | .int JMPTBL (L(bwd_write_91bytes), L(table_144_bytes_bwd)) |
2947 | .int JMPTBL (L(bwd_write_92bytes), L(table_144_bytes_bwd)) |
2948 | .int JMPTBL (L(bwd_write_93bytes), L(table_144_bytes_bwd)) |
2949 | .int JMPTBL (L(bwd_write_94bytes), L(table_144_bytes_bwd)) |
2950 | .int JMPTBL (L(bwd_write_95bytes), L(table_144_bytes_bwd)) |
2951 | .int JMPTBL (L(bwd_write_96bytes), L(table_144_bytes_bwd)) |
2952 | .int JMPTBL (L(bwd_write_97bytes), L(table_144_bytes_bwd)) |
2953 | .int JMPTBL (L(bwd_write_98bytes), L(table_144_bytes_bwd)) |
2954 | .int JMPTBL (L(bwd_write_99bytes), L(table_144_bytes_bwd)) |
2955 | .int JMPTBL (L(bwd_write_100bytes), L(table_144_bytes_bwd)) |
2956 | .int JMPTBL (L(bwd_write_101bytes), L(table_144_bytes_bwd)) |
2957 | .int JMPTBL (L(bwd_write_102bytes), L(table_144_bytes_bwd)) |
2958 | .int JMPTBL (L(bwd_write_103bytes), L(table_144_bytes_bwd)) |
2959 | .int JMPTBL (L(bwd_write_104bytes), L(table_144_bytes_bwd)) |
2960 | .int JMPTBL (L(bwd_write_105bytes), L(table_144_bytes_bwd)) |
2961 | .int JMPTBL (L(bwd_write_106bytes), L(table_144_bytes_bwd)) |
2962 | .int JMPTBL (L(bwd_write_107bytes), L(table_144_bytes_bwd)) |
2963 | .int JMPTBL (L(bwd_write_108bytes), L(table_144_bytes_bwd)) |
2964 | .int JMPTBL (L(bwd_write_109bytes), L(table_144_bytes_bwd)) |
2965 | .int JMPTBL (L(bwd_write_110bytes), L(table_144_bytes_bwd)) |
2966 | .int JMPTBL (L(bwd_write_111bytes), L(table_144_bytes_bwd)) |
2967 | .int JMPTBL (L(bwd_write_112bytes), L(table_144_bytes_bwd)) |
2968 | .int JMPTBL (L(bwd_write_113bytes), L(table_144_bytes_bwd)) |
2969 | .int JMPTBL (L(bwd_write_114bytes), L(table_144_bytes_bwd)) |
2970 | .int JMPTBL (L(bwd_write_115bytes), L(table_144_bytes_bwd)) |
2971 | .int JMPTBL (L(bwd_write_116bytes), L(table_144_bytes_bwd)) |
2972 | .int JMPTBL (L(bwd_write_117bytes), L(table_144_bytes_bwd)) |
2973 | .int JMPTBL (L(bwd_write_118bytes), L(table_144_bytes_bwd)) |
2974 | .int JMPTBL (L(bwd_write_119bytes), L(table_144_bytes_bwd)) |
2975 | .int JMPTBL (L(bwd_write_120bytes), L(table_144_bytes_bwd)) |
2976 | .int JMPTBL (L(bwd_write_121bytes), L(table_144_bytes_bwd)) |
2977 | .int JMPTBL (L(bwd_write_122bytes), L(table_144_bytes_bwd)) |
2978 | .int JMPTBL (L(bwd_write_123bytes), L(table_144_bytes_bwd)) |
2979 | .int JMPTBL (L(bwd_write_124bytes), L(table_144_bytes_bwd)) |
2980 | .int JMPTBL (L(bwd_write_125bytes), L(table_144_bytes_bwd)) |
2981 | .int JMPTBL (L(bwd_write_126bytes), L(table_144_bytes_bwd)) |
2982 | .int JMPTBL (L(bwd_write_127bytes), L(table_144_bytes_bwd)) |
2983 | .int JMPTBL (L(bwd_write_128bytes), L(table_144_bytes_bwd)) |
2984 | .int JMPTBL (L(bwd_write_129bytes), L(table_144_bytes_bwd)) |
2985 | .int JMPTBL (L(bwd_write_130bytes), L(table_144_bytes_bwd)) |
2986 | .int JMPTBL (L(bwd_write_131bytes), L(table_144_bytes_bwd)) |
2987 | .int JMPTBL (L(bwd_write_132bytes), L(table_144_bytes_bwd)) |
2988 | .int JMPTBL (L(bwd_write_133bytes), L(table_144_bytes_bwd)) |
2989 | .int JMPTBL (L(bwd_write_134bytes), L(table_144_bytes_bwd)) |
2990 | .int JMPTBL (L(bwd_write_135bytes), L(table_144_bytes_bwd)) |
2991 | .int JMPTBL (L(bwd_write_136bytes), L(table_144_bytes_bwd)) |
2992 | .int JMPTBL (L(bwd_write_137bytes), L(table_144_bytes_bwd)) |
2993 | .int JMPTBL (L(bwd_write_138bytes), L(table_144_bytes_bwd)) |
2994 | .int JMPTBL (L(bwd_write_139bytes), L(table_144_bytes_bwd)) |
2995 | .int JMPTBL (L(bwd_write_140bytes), L(table_144_bytes_bwd)) |
2996 | .int JMPTBL (L(bwd_write_141bytes), L(table_144_bytes_bwd)) |
2997 | .int JMPTBL (L(bwd_write_142bytes), L(table_144_bytes_bwd)) |
2998 | .int JMPTBL (L(bwd_write_143bytes), L(table_144_bytes_bwd)) |
2999 | |
3000 | .p2align 3 |
3001 | L(table_144_bytes_fwd): |
3002 | .int JMPTBL (L(fwd_write_0bytes), L(table_144_bytes_fwd)) |
3003 | .int JMPTBL (L(fwd_write_1bytes), L(table_144_bytes_fwd)) |
3004 | .int JMPTBL (L(fwd_write_2bytes), L(table_144_bytes_fwd)) |
3005 | .int JMPTBL (L(fwd_write_3bytes), L(table_144_bytes_fwd)) |
3006 | .int JMPTBL (L(fwd_write_4bytes), L(table_144_bytes_fwd)) |
3007 | .int JMPTBL (L(fwd_write_5bytes), L(table_144_bytes_fwd)) |
3008 | .int JMPTBL (L(fwd_write_6bytes), L(table_144_bytes_fwd)) |
3009 | .int JMPTBL (L(fwd_write_7bytes), L(table_144_bytes_fwd)) |
3010 | .int JMPTBL (L(fwd_write_8bytes), L(table_144_bytes_fwd)) |
3011 | .int JMPTBL (L(fwd_write_9bytes), L(table_144_bytes_fwd)) |
3012 | .int JMPTBL (L(fwd_write_10bytes), L(table_144_bytes_fwd)) |
3013 | .int JMPTBL (L(fwd_write_11bytes), L(table_144_bytes_fwd)) |
3014 | .int JMPTBL (L(fwd_write_12bytes), L(table_144_bytes_fwd)) |
3015 | .int JMPTBL (L(fwd_write_13bytes), L(table_144_bytes_fwd)) |
3016 | .int JMPTBL (L(fwd_write_14bytes), L(table_144_bytes_fwd)) |
3017 | .int JMPTBL (L(fwd_write_15bytes), L(table_144_bytes_fwd)) |
3018 | .int JMPTBL (L(fwd_write_16bytes), L(table_144_bytes_fwd)) |
3019 | .int JMPTBL (L(fwd_write_17bytes), L(table_144_bytes_fwd)) |
3020 | .int JMPTBL (L(fwd_write_18bytes), L(table_144_bytes_fwd)) |
3021 | .int JMPTBL (L(fwd_write_19bytes), L(table_144_bytes_fwd)) |
3022 | .int JMPTBL (L(fwd_write_20bytes), L(table_144_bytes_fwd)) |
3023 | .int JMPTBL (L(fwd_write_21bytes), L(table_144_bytes_fwd)) |
3024 | .int JMPTBL (L(fwd_write_22bytes), L(table_144_bytes_fwd)) |
3025 | .int JMPTBL (L(fwd_write_23bytes), L(table_144_bytes_fwd)) |
3026 | .int JMPTBL (L(fwd_write_24bytes), L(table_144_bytes_fwd)) |
3027 | .int JMPTBL (L(fwd_write_25bytes), L(table_144_bytes_fwd)) |
3028 | .int JMPTBL (L(fwd_write_26bytes), L(table_144_bytes_fwd)) |
3029 | .int JMPTBL (L(fwd_write_27bytes), L(table_144_bytes_fwd)) |
3030 | .int JMPTBL (L(fwd_write_28bytes), L(table_144_bytes_fwd)) |
3031 | .int JMPTBL (L(fwd_write_29bytes), L(table_144_bytes_fwd)) |
3032 | .int JMPTBL (L(fwd_write_30bytes), L(table_144_bytes_fwd)) |
3033 | .int JMPTBL (L(fwd_write_31bytes), L(table_144_bytes_fwd)) |
3034 | .int JMPTBL (L(fwd_write_32bytes), L(table_144_bytes_fwd)) |
3035 | .int JMPTBL (L(fwd_write_33bytes), L(table_144_bytes_fwd)) |
3036 | .int JMPTBL (L(fwd_write_34bytes), L(table_144_bytes_fwd)) |
3037 | .int JMPTBL (L(fwd_write_35bytes), L(table_144_bytes_fwd)) |
3038 | .int JMPTBL (L(fwd_write_36bytes), L(table_144_bytes_fwd)) |
3039 | .int JMPTBL (L(fwd_write_37bytes), L(table_144_bytes_fwd)) |
3040 | .int JMPTBL (L(fwd_write_38bytes), L(table_144_bytes_fwd)) |
3041 | .int JMPTBL (L(fwd_write_39bytes), L(table_144_bytes_fwd)) |
3042 | .int JMPTBL (L(fwd_write_40bytes), L(table_144_bytes_fwd)) |
3043 | .int JMPTBL (L(fwd_write_41bytes), L(table_144_bytes_fwd)) |
3044 | .int JMPTBL (L(fwd_write_42bytes), L(table_144_bytes_fwd)) |
3045 | .int JMPTBL (L(fwd_write_43bytes), L(table_144_bytes_fwd)) |
3046 | .int JMPTBL (L(fwd_write_44bytes), L(table_144_bytes_fwd)) |
3047 | .int JMPTBL (L(fwd_write_45bytes), L(table_144_bytes_fwd)) |
3048 | .int JMPTBL (L(fwd_write_46bytes), L(table_144_bytes_fwd)) |
3049 | .int JMPTBL (L(fwd_write_47bytes), L(table_144_bytes_fwd)) |
3050 | .int JMPTBL (L(fwd_write_48bytes), L(table_144_bytes_fwd)) |
3051 | .int JMPTBL (L(fwd_write_49bytes), L(table_144_bytes_fwd)) |
3052 | .int JMPTBL (L(fwd_write_50bytes), L(table_144_bytes_fwd)) |
3053 | .int JMPTBL (L(fwd_write_51bytes), L(table_144_bytes_fwd)) |
3054 | .int JMPTBL (L(fwd_write_52bytes), L(table_144_bytes_fwd)) |
3055 | .int JMPTBL (L(fwd_write_53bytes), L(table_144_bytes_fwd)) |
3056 | .int JMPTBL (L(fwd_write_54bytes), L(table_144_bytes_fwd)) |
3057 | .int JMPTBL (L(fwd_write_55bytes), L(table_144_bytes_fwd)) |
3058 | .int JMPTBL (L(fwd_write_56bytes), L(table_144_bytes_fwd)) |
3059 | .int JMPTBL (L(fwd_write_57bytes), L(table_144_bytes_fwd)) |
3060 | .int JMPTBL (L(fwd_write_58bytes), L(table_144_bytes_fwd)) |
3061 | .int JMPTBL (L(fwd_write_59bytes), L(table_144_bytes_fwd)) |
3062 | .int JMPTBL (L(fwd_write_60bytes), L(table_144_bytes_fwd)) |
3063 | .int JMPTBL (L(fwd_write_61bytes), L(table_144_bytes_fwd)) |
3064 | .int JMPTBL (L(fwd_write_62bytes), L(table_144_bytes_fwd)) |
3065 | .int JMPTBL (L(fwd_write_63bytes), L(table_144_bytes_fwd)) |
3066 | .int JMPTBL (L(fwd_write_64bytes), L(table_144_bytes_fwd)) |
3067 | .int JMPTBL (L(fwd_write_65bytes), L(table_144_bytes_fwd)) |
3068 | .int JMPTBL (L(fwd_write_66bytes), L(table_144_bytes_fwd)) |
3069 | .int JMPTBL (L(fwd_write_67bytes), L(table_144_bytes_fwd)) |
3070 | .int JMPTBL (L(fwd_write_68bytes), L(table_144_bytes_fwd)) |
3071 | .int JMPTBL (L(fwd_write_69bytes), L(table_144_bytes_fwd)) |
3072 | .int JMPTBL (L(fwd_write_70bytes), L(table_144_bytes_fwd)) |
3073 | .int JMPTBL (L(fwd_write_71bytes), L(table_144_bytes_fwd)) |
3074 | .int JMPTBL (L(fwd_write_72bytes), L(table_144_bytes_fwd)) |
3075 | .int JMPTBL (L(fwd_write_73bytes), L(table_144_bytes_fwd)) |
3076 | .int JMPTBL (L(fwd_write_74bytes), L(table_144_bytes_fwd)) |
3077 | .int JMPTBL (L(fwd_write_75bytes), L(table_144_bytes_fwd)) |
3078 | .int JMPTBL (L(fwd_write_76bytes), L(table_144_bytes_fwd)) |
3079 | .int JMPTBL (L(fwd_write_77bytes), L(table_144_bytes_fwd)) |
3080 | .int JMPTBL (L(fwd_write_78bytes), L(table_144_bytes_fwd)) |
3081 | .int JMPTBL (L(fwd_write_79bytes), L(table_144_bytes_fwd)) |
3082 | .int JMPTBL (L(fwd_write_80bytes), L(table_144_bytes_fwd)) |
3083 | .int JMPTBL (L(fwd_write_81bytes), L(table_144_bytes_fwd)) |
3084 | .int JMPTBL (L(fwd_write_82bytes), L(table_144_bytes_fwd)) |
3085 | .int JMPTBL (L(fwd_write_83bytes), L(table_144_bytes_fwd)) |
3086 | .int JMPTBL (L(fwd_write_84bytes), L(table_144_bytes_fwd)) |
3087 | .int JMPTBL (L(fwd_write_85bytes), L(table_144_bytes_fwd)) |
3088 | .int JMPTBL (L(fwd_write_86bytes), L(table_144_bytes_fwd)) |
3089 | .int JMPTBL (L(fwd_write_87bytes), L(table_144_bytes_fwd)) |
3090 | .int JMPTBL (L(fwd_write_88bytes), L(table_144_bytes_fwd)) |
3091 | .int JMPTBL (L(fwd_write_89bytes), L(table_144_bytes_fwd)) |
3092 | .int JMPTBL (L(fwd_write_90bytes), L(table_144_bytes_fwd)) |
3093 | .int JMPTBL (L(fwd_write_91bytes), L(table_144_bytes_fwd)) |
3094 | .int JMPTBL (L(fwd_write_92bytes), L(table_144_bytes_fwd)) |
3095 | .int JMPTBL (L(fwd_write_93bytes), L(table_144_bytes_fwd)) |
3096 | .int JMPTBL (L(fwd_write_94bytes), L(table_144_bytes_fwd)) |
3097 | .int JMPTBL (L(fwd_write_95bytes), L(table_144_bytes_fwd)) |
3098 | .int JMPTBL (L(fwd_write_96bytes), L(table_144_bytes_fwd)) |
3099 | .int JMPTBL (L(fwd_write_97bytes), L(table_144_bytes_fwd)) |
3100 | .int JMPTBL (L(fwd_write_98bytes), L(table_144_bytes_fwd)) |
3101 | .int JMPTBL (L(fwd_write_99bytes), L(table_144_bytes_fwd)) |
3102 | .int JMPTBL (L(fwd_write_100bytes), L(table_144_bytes_fwd)) |
3103 | .int JMPTBL (L(fwd_write_101bytes), L(table_144_bytes_fwd)) |
3104 | .int JMPTBL (L(fwd_write_102bytes), L(table_144_bytes_fwd)) |
3105 | .int JMPTBL (L(fwd_write_103bytes), L(table_144_bytes_fwd)) |
3106 | .int JMPTBL (L(fwd_write_104bytes), L(table_144_bytes_fwd)) |
3107 | .int JMPTBL (L(fwd_write_105bytes), L(table_144_bytes_fwd)) |
3108 | .int JMPTBL (L(fwd_write_106bytes), L(table_144_bytes_fwd)) |
3109 | .int JMPTBL (L(fwd_write_107bytes), L(table_144_bytes_fwd)) |
3110 | .int JMPTBL (L(fwd_write_108bytes), L(table_144_bytes_fwd)) |
3111 | .int JMPTBL (L(fwd_write_109bytes), L(table_144_bytes_fwd)) |
3112 | .int JMPTBL (L(fwd_write_110bytes), L(table_144_bytes_fwd)) |
3113 | .int JMPTBL (L(fwd_write_111bytes), L(table_144_bytes_fwd)) |
3114 | .int JMPTBL (L(fwd_write_112bytes), L(table_144_bytes_fwd)) |
3115 | .int JMPTBL (L(fwd_write_113bytes), L(table_144_bytes_fwd)) |
3116 | .int JMPTBL (L(fwd_write_114bytes), L(table_144_bytes_fwd)) |
3117 | .int JMPTBL (L(fwd_write_115bytes), L(table_144_bytes_fwd)) |
3118 | .int JMPTBL (L(fwd_write_116bytes), L(table_144_bytes_fwd)) |
3119 | .int JMPTBL (L(fwd_write_117bytes), L(table_144_bytes_fwd)) |
3120 | .int JMPTBL (L(fwd_write_118bytes), L(table_144_bytes_fwd)) |
3121 | .int JMPTBL (L(fwd_write_119bytes), L(table_144_bytes_fwd)) |
3122 | .int JMPTBL (L(fwd_write_120bytes), L(table_144_bytes_fwd)) |
3123 | .int JMPTBL (L(fwd_write_121bytes), L(table_144_bytes_fwd)) |
3124 | .int JMPTBL (L(fwd_write_122bytes), L(table_144_bytes_fwd)) |
3125 | .int JMPTBL (L(fwd_write_123bytes), L(table_144_bytes_fwd)) |
3126 | .int JMPTBL (L(fwd_write_124bytes), L(table_144_bytes_fwd)) |
3127 | .int JMPTBL (L(fwd_write_125bytes), L(table_144_bytes_fwd)) |
3128 | .int JMPTBL (L(fwd_write_126bytes), L(table_144_bytes_fwd)) |
3129 | .int JMPTBL (L(fwd_write_127bytes), L(table_144_bytes_fwd)) |
3130 | .int JMPTBL (L(fwd_write_128bytes), L(table_144_bytes_fwd)) |
3131 | .int JMPTBL (L(fwd_write_129bytes), L(table_144_bytes_fwd)) |
3132 | .int JMPTBL (L(fwd_write_130bytes), L(table_144_bytes_fwd)) |
3133 | .int JMPTBL (L(fwd_write_131bytes), L(table_144_bytes_fwd)) |
3134 | .int JMPTBL (L(fwd_write_132bytes), L(table_144_bytes_fwd)) |
3135 | .int JMPTBL (L(fwd_write_133bytes), L(table_144_bytes_fwd)) |
3136 | .int JMPTBL (L(fwd_write_134bytes), L(table_144_bytes_fwd)) |
3137 | .int JMPTBL (L(fwd_write_135bytes), L(table_144_bytes_fwd)) |
3138 | .int JMPTBL (L(fwd_write_136bytes), L(table_144_bytes_fwd)) |
3139 | .int JMPTBL (L(fwd_write_137bytes), L(table_144_bytes_fwd)) |
3140 | .int JMPTBL (L(fwd_write_138bytes), L(table_144_bytes_fwd)) |
3141 | .int JMPTBL (L(fwd_write_139bytes), L(table_144_bytes_fwd)) |
3142 | .int JMPTBL (L(fwd_write_140bytes), L(table_144_bytes_fwd)) |
3143 | .int JMPTBL (L(fwd_write_141bytes), L(table_144_bytes_fwd)) |
3144 | .int JMPTBL (L(fwd_write_142bytes), L(table_144_bytes_fwd)) |
3145 | .int JMPTBL (L(fwd_write_143bytes), L(table_144_bytes_fwd)) |
3146 | |
3147 | .p2align 3 |
3148 | L(shl_table_fwd): |
3149 | .int JMPTBL (L(shl_0), L(shl_table_fwd)) |
3150 | .int JMPTBL (L(shl_1), L(shl_table_fwd)) |
3151 | .int JMPTBL (L(shl_2), L(shl_table_fwd)) |
3152 | .int JMPTBL (L(shl_3), L(shl_table_fwd)) |
3153 | .int JMPTBL (L(shl_4), L(shl_table_fwd)) |
3154 | .int JMPTBL (L(shl_5), L(shl_table_fwd)) |
3155 | .int JMPTBL (L(shl_6), L(shl_table_fwd)) |
3156 | .int JMPTBL (L(shl_7), L(shl_table_fwd)) |
3157 | .int JMPTBL (L(shl_8), L(shl_table_fwd)) |
3158 | .int JMPTBL (L(shl_9), L(shl_table_fwd)) |
3159 | .int JMPTBL (L(shl_10), L(shl_table_fwd)) |
3160 | .int JMPTBL (L(shl_11), L(shl_table_fwd)) |
3161 | .int JMPTBL (L(shl_12), L(shl_table_fwd)) |
3162 | .int JMPTBL (L(shl_13), L(shl_table_fwd)) |
3163 | .int JMPTBL (L(shl_14), L(shl_table_fwd)) |
3164 | .int JMPTBL (L(shl_15), L(shl_table_fwd)) |
3165 | |
3166 | .p2align 3 |
3167 | L(shl_table_bwd): |
3168 | .int JMPTBL (L(shl_0_bwd), L(shl_table_bwd)) |
3169 | .int JMPTBL (L(shl_1_bwd), L(shl_table_bwd)) |
3170 | .int JMPTBL (L(shl_2_bwd), L(shl_table_bwd)) |
3171 | .int JMPTBL (L(shl_3_bwd), L(shl_table_bwd)) |
3172 | .int JMPTBL (L(shl_4_bwd), L(shl_table_bwd)) |
3173 | .int JMPTBL (L(shl_5_bwd), L(shl_table_bwd)) |
3174 | .int JMPTBL (L(shl_6_bwd), L(shl_table_bwd)) |
3175 | .int JMPTBL (L(shl_7_bwd), L(shl_table_bwd)) |
3176 | .int JMPTBL (L(shl_8_bwd), L(shl_table_bwd)) |
3177 | .int JMPTBL (L(shl_9_bwd), L(shl_table_bwd)) |
3178 | .int JMPTBL (L(shl_10_bwd), L(shl_table_bwd)) |
3179 | .int JMPTBL (L(shl_11_bwd), L(shl_table_bwd)) |
3180 | .int JMPTBL (L(shl_12_bwd), L(shl_table_bwd)) |
3181 | .int JMPTBL (L(shl_13_bwd), L(shl_table_bwd)) |
3182 | .int JMPTBL (L(shl_14_bwd), L(shl_table_bwd)) |
3183 | .int JMPTBL (L(shl_15_bwd), L(shl_table_bwd)) |
3184 | |
3185 | #endif |
3186 | |