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