1 | /* memcpy with SSSE3 and REP string |
2 | Copyright (C) 2010-2021 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 | <https://www.gnu.org/licenses/>. */ |
19 | |
20 | #include <sysdep.h> |
21 | |
22 | #if IS_IN (libc) |
23 | |
24 | #include "asm-syntax.h" |
25 | |
26 | #ifndef MEMCPY |
27 | # define MEMCPY __memcpy_ssse3_back |
28 | # define MEMCPY_CHK __memcpy_chk_ssse3_back |
29 | # define MEMPCPY __mempcpy_ssse3_back |
30 | # define MEMPCPY_CHK __mempcpy_chk_ssse3_back |
31 | #endif |
32 | |
33 | #define JMPTBL(I, B) I - B |
34 | |
35 | /* Branch to an entry in a jump table. TABLE is a jump table with |
36 | relative offsets. INDEX is a register contains the index into the |
37 | jump table. SCALE is the scale of INDEX. */ |
38 | #define BRANCH_TO_JMPTBL_ENTRY(TABLE, INDEX, SCALE) \ |
39 | lea TABLE(%rip), %r11; \ |
40 | movslq (%r11, INDEX, SCALE), INDEX; \ |
41 | lea (%r11, INDEX), INDEX; \ |
42 | _CET_NOTRACK jmp *INDEX; \ |
43 | ud2 |
44 | |
45 | .section .text.ssse3,"ax" ,@progbits |
46 | #if !defined USE_AS_MEMPCPY && !defined USE_AS_MEMMOVE |
47 | ENTRY (MEMPCPY_CHK) |
48 | cmp %RDX_LP, %RCX_LP |
49 | jb HIDDEN_JUMPTARGET (__chk_fail) |
50 | END (MEMPCPY_CHK) |
51 | |
52 | ENTRY (MEMPCPY) |
53 | mov %RDI_LP, %RAX_LP |
54 | add %RDX_LP, %RAX_LP |
55 | jmp L(start) |
56 | END (MEMPCPY) |
57 | #endif |
58 | |
59 | #if !defined USE_AS_BCOPY |
60 | ENTRY (MEMCPY_CHK) |
61 | cmp %RDX_LP, %RCX_LP |
62 | jb HIDDEN_JUMPTARGET (__chk_fail) |
63 | END (MEMCPY_CHK) |
64 | #endif |
65 | |
66 | ENTRY (MEMCPY) |
67 | mov %RDI_LP, %RAX_LP |
68 | #ifdef USE_AS_MEMPCPY |
69 | add %RDX_LP, %RAX_LP |
70 | #endif |
71 | |
72 | #ifdef __ILP32__ |
73 | /* Clear the upper 32 bits. */ |
74 | mov %edx, %edx |
75 | #endif |
76 | |
77 | #ifdef USE_AS_MEMMOVE |
78 | cmp %rsi, %rdi |
79 | jb L(copy_forward) |
80 | je L(bwd_write_0bytes) |
81 | cmp $144, %rdx |
82 | jae L(copy_backward) |
83 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
84 | L(copy_forward): |
85 | #endif |
86 | L(start): |
87 | cmp $144, %rdx |
88 | jae L(144bytesormore) |
89 | |
90 | L(fwd_write_less32bytes): |
91 | #ifndef USE_AS_MEMMOVE |
92 | cmp %dil, %sil |
93 | jbe L(bk_write) |
94 | #endif |
95 | add %rdx, %rsi |
96 | add %rdx, %rdi |
97 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
98 | #ifndef USE_AS_MEMMOVE |
99 | L(bk_write): |
100 | |
101 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
102 | #endif |
103 | |
104 | .p2align 4 |
105 | L(144bytesormore): |
106 | |
107 | #ifndef USE_AS_MEMMOVE |
108 | cmp %dil, %sil |
109 | jle L(copy_backward) |
110 | #endif |
111 | movdqu (%rsi), %xmm0 |
112 | mov %rdi, %r8 |
113 | and $-16, %rdi |
114 | add $16, %rdi |
115 | mov %rdi, %r9 |
116 | sub %r8, %r9 |
117 | sub %r9, %rdx |
118 | add %r9, %rsi |
119 | mov %rsi, %r9 |
120 | and $0xf, %r9 |
121 | jz L(shl_0) |
122 | #ifdef DATA_CACHE_SIZE |
123 | mov $DATA_CACHE_SIZE, %RCX_LP |
124 | #else |
125 | mov __x86_data_cache_size(%rip), %RCX_LP |
126 | #endif |
127 | cmp %rcx, %rdx |
128 | jae L(gobble_mem_fwd) |
129 | lea L(shl_table_fwd)(%rip), %r11 |
130 | sub $0x80, %rdx |
131 | movslq (%r11, %r9, 4), %r9 |
132 | add %r11, %r9 |
133 | _CET_NOTRACK jmp *%r9 |
134 | ud2 |
135 | |
136 | .p2align 4 |
137 | L(copy_backward): |
138 | #ifdef DATA_CACHE_SIZE |
139 | mov $DATA_CACHE_SIZE, %RCX_LP |
140 | #else |
141 | mov __x86_data_cache_size(%rip), %RCX_LP |
142 | #endif |
143 | shl $1, %rcx |
144 | cmp %rcx, %rdx |
145 | ja L(gobble_mem_bwd) |
146 | |
147 | add %rdx, %rdi |
148 | add %rdx, %rsi |
149 | movdqu -16(%rsi), %xmm0 |
150 | lea -16(%rdi), %r8 |
151 | mov %rdi, %r9 |
152 | and $0xf, %r9 |
153 | xor %r9, %rdi |
154 | sub %r9, %rsi |
155 | sub %r9, %rdx |
156 | mov %rsi, %r9 |
157 | and $0xf, %r9 |
158 | jz L(shl_0_bwd) |
159 | lea L(shl_table_bwd)(%rip), %r11 |
160 | sub $0x80, %rdx |
161 | movslq (%r11, %r9, 4), %r9 |
162 | add %r11, %r9 |
163 | _CET_NOTRACK jmp *%r9 |
164 | ud2 |
165 | |
166 | .p2align 4 |
167 | L(shl_0): |
168 | |
169 | mov %rdx, %r9 |
170 | shr $8, %r9 |
171 | add %rdx, %r9 |
172 | #ifdef DATA_CACHE_SIZE |
173 | cmp $DATA_CACHE_SIZE_HALF, %R9_LP |
174 | #else |
175 | cmp __x86_data_cache_size_half(%rip), %R9_LP |
176 | #endif |
177 | jae L(gobble_mem_fwd) |
178 | sub $0x80, %rdx |
179 | .p2align 4 |
180 | L(shl_0_loop): |
181 | movdqa (%rsi), %xmm1 |
182 | movdqa %xmm1, (%rdi) |
183 | movaps 0x10(%rsi), %xmm2 |
184 | movaps %xmm2, 0x10(%rdi) |
185 | movaps 0x20(%rsi), %xmm3 |
186 | movaps %xmm3, 0x20(%rdi) |
187 | movaps 0x30(%rsi), %xmm4 |
188 | movaps %xmm4, 0x30(%rdi) |
189 | movaps 0x40(%rsi), %xmm1 |
190 | movaps %xmm1, 0x40(%rdi) |
191 | movaps 0x50(%rsi), %xmm2 |
192 | movaps %xmm2, 0x50(%rdi) |
193 | movaps 0x60(%rsi), %xmm3 |
194 | movaps %xmm3, 0x60(%rdi) |
195 | movaps 0x70(%rsi), %xmm4 |
196 | movaps %xmm4, 0x70(%rdi) |
197 | sub $0x80, %rdx |
198 | lea 0x80(%rsi), %rsi |
199 | lea 0x80(%rdi), %rdi |
200 | jae L(shl_0_loop) |
201 | movdqu %xmm0, (%r8) |
202 | add $0x80, %rdx |
203 | add %rdx, %rsi |
204 | add %rdx, %rdi |
205 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
206 | |
207 | .p2align 4 |
208 | L(shl_0_bwd): |
209 | sub $0x80, %rdx |
210 | L(copy_backward_loop): |
211 | movaps -0x10(%rsi), %xmm1 |
212 | movaps %xmm1, -0x10(%rdi) |
213 | movaps -0x20(%rsi), %xmm2 |
214 | movaps %xmm2, -0x20(%rdi) |
215 | movaps -0x30(%rsi), %xmm3 |
216 | movaps %xmm3, -0x30(%rdi) |
217 | movaps -0x40(%rsi), %xmm4 |
218 | movaps %xmm4, -0x40(%rdi) |
219 | movaps -0x50(%rsi), %xmm5 |
220 | movaps %xmm5, -0x50(%rdi) |
221 | movaps -0x60(%rsi), %xmm5 |
222 | movaps %xmm5, -0x60(%rdi) |
223 | movaps -0x70(%rsi), %xmm5 |
224 | movaps %xmm5, -0x70(%rdi) |
225 | movaps -0x80(%rsi), %xmm5 |
226 | movaps %xmm5, -0x80(%rdi) |
227 | sub $0x80, %rdx |
228 | lea -0x80(%rdi), %rdi |
229 | lea -0x80(%rsi), %rsi |
230 | jae L(copy_backward_loop) |
231 | |
232 | movdqu %xmm0, (%r8) |
233 | add $0x80, %rdx |
234 | sub %rdx, %rdi |
235 | sub %rdx, %rsi |
236 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
237 | |
238 | .p2align 4 |
239 | L(shl_1): |
240 | sub $0x80, %rdx |
241 | movaps -0x01(%rsi), %xmm1 |
242 | movaps 0x0f(%rsi), %xmm2 |
243 | movaps 0x1f(%rsi), %xmm3 |
244 | movaps 0x2f(%rsi), %xmm4 |
245 | movaps 0x3f(%rsi), %xmm5 |
246 | movaps 0x4f(%rsi), %xmm6 |
247 | movaps 0x5f(%rsi), %xmm7 |
248 | movaps 0x6f(%rsi), %xmm8 |
249 | movaps 0x7f(%rsi), %xmm9 |
250 | lea 0x80(%rsi), %rsi |
251 | palignr $1, %xmm8, %xmm9 |
252 | movaps %xmm9, 0x70(%rdi) |
253 | palignr $1, %xmm7, %xmm8 |
254 | movaps %xmm8, 0x60(%rdi) |
255 | palignr $1, %xmm6, %xmm7 |
256 | movaps %xmm7, 0x50(%rdi) |
257 | palignr $1, %xmm5, %xmm6 |
258 | movaps %xmm6, 0x40(%rdi) |
259 | palignr $1, %xmm4, %xmm5 |
260 | movaps %xmm5, 0x30(%rdi) |
261 | palignr $1, %xmm3, %xmm4 |
262 | movaps %xmm4, 0x20(%rdi) |
263 | palignr $1, %xmm2, %xmm3 |
264 | movaps %xmm3, 0x10(%rdi) |
265 | palignr $1, %xmm1, %xmm2 |
266 | movaps %xmm2, (%rdi) |
267 | lea 0x80(%rdi), %rdi |
268 | jae L(shl_1) |
269 | movdqu %xmm0, (%r8) |
270 | add $0x80, %rdx |
271 | add %rdx, %rdi |
272 | add %rdx, %rsi |
273 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
274 | |
275 | .p2align 4 |
276 | L(shl_1_bwd): |
277 | movaps -0x01(%rsi), %xmm1 |
278 | |
279 | movaps -0x11(%rsi), %xmm2 |
280 | palignr $1, %xmm2, %xmm1 |
281 | movaps %xmm1, -0x10(%rdi) |
282 | |
283 | movaps -0x21(%rsi), %xmm3 |
284 | palignr $1, %xmm3, %xmm2 |
285 | movaps %xmm2, -0x20(%rdi) |
286 | |
287 | movaps -0x31(%rsi), %xmm4 |
288 | palignr $1, %xmm4, %xmm3 |
289 | movaps %xmm3, -0x30(%rdi) |
290 | |
291 | movaps -0x41(%rsi), %xmm5 |
292 | palignr $1, %xmm5, %xmm4 |
293 | movaps %xmm4, -0x40(%rdi) |
294 | |
295 | movaps -0x51(%rsi), %xmm6 |
296 | palignr $1, %xmm6, %xmm5 |
297 | movaps %xmm5, -0x50(%rdi) |
298 | |
299 | movaps -0x61(%rsi), %xmm7 |
300 | palignr $1, %xmm7, %xmm6 |
301 | movaps %xmm6, -0x60(%rdi) |
302 | |
303 | movaps -0x71(%rsi), %xmm8 |
304 | palignr $1, %xmm8, %xmm7 |
305 | movaps %xmm7, -0x70(%rdi) |
306 | |
307 | movaps -0x81(%rsi), %xmm9 |
308 | palignr $1, %xmm9, %xmm8 |
309 | movaps %xmm8, -0x80(%rdi) |
310 | |
311 | sub $0x80, %rdx |
312 | lea -0x80(%rdi), %rdi |
313 | lea -0x80(%rsi), %rsi |
314 | jae L(shl_1_bwd) |
315 | movdqu %xmm0, (%r8) |
316 | add $0x80, %rdx |
317 | sub %rdx, %rdi |
318 | sub %rdx, %rsi |
319 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
320 | |
321 | .p2align 4 |
322 | L(shl_2): |
323 | sub $0x80, %rdx |
324 | movaps -0x02(%rsi), %xmm1 |
325 | movaps 0x0e(%rsi), %xmm2 |
326 | movaps 0x1e(%rsi), %xmm3 |
327 | movaps 0x2e(%rsi), %xmm4 |
328 | movaps 0x3e(%rsi), %xmm5 |
329 | movaps 0x4e(%rsi), %xmm6 |
330 | movaps 0x5e(%rsi), %xmm7 |
331 | movaps 0x6e(%rsi), %xmm8 |
332 | movaps 0x7e(%rsi), %xmm9 |
333 | lea 0x80(%rsi), %rsi |
334 | palignr $2, %xmm8, %xmm9 |
335 | movaps %xmm9, 0x70(%rdi) |
336 | palignr $2, %xmm7, %xmm8 |
337 | movaps %xmm8, 0x60(%rdi) |
338 | palignr $2, %xmm6, %xmm7 |
339 | movaps %xmm7, 0x50(%rdi) |
340 | palignr $2, %xmm5, %xmm6 |
341 | movaps %xmm6, 0x40(%rdi) |
342 | palignr $2, %xmm4, %xmm5 |
343 | movaps %xmm5, 0x30(%rdi) |
344 | palignr $2, %xmm3, %xmm4 |
345 | movaps %xmm4, 0x20(%rdi) |
346 | palignr $2, %xmm2, %xmm3 |
347 | movaps %xmm3, 0x10(%rdi) |
348 | palignr $2, %xmm1, %xmm2 |
349 | movaps %xmm2, (%rdi) |
350 | lea 0x80(%rdi), %rdi |
351 | jae L(shl_2) |
352 | movdqu %xmm0, (%r8) |
353 | add $0x80, %rdx |
354 | add %rdx, %rdi |
355 | add %rdx, %rsi |
356 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
357 | |
358 | .p2align 4 |
359 | L(shl_2_bwd): |
360 | movaps -0x02(%rsi), %xmm1 |
361 | |
362 | movaps -0x12(%rsi), %xmm2 |
363 | palignr $2, %xmm2, %xmm1 |
364 | movaps %xmm1, -0x10(%rdi) |
365 | |
366 | movaps -0x22(%rsi), %xmm3 |
367 | palignr $2, %xmm3, %xmm2 |
368 | movaps %xmm2, -0x20(%rdi) |
369 | |
370 | movaps -0x32(%rsi), %xmm4 |
371 | palignr $2, %xmm4, %xmm3 |
372 | movaps %xmm3, -0x30(%rdi) |
373 | |
374 | movaps -0x42(%rsi), %xmm5 |
375 | palignr $2, %xmm5, %xmm4 |
376 | movaps %xmm4, -0x40(%rdi) |
377 | |
378 | movaps -0x52(%rsi), %xmm6 |
379 | palignr $2, %xmm6, %xmm5 |
380 | movaps %xmm5, -0x50(%rdi) |
381 | |
382 | movaps -0x62(%rsi), %xmm7 |
383 | palignr $2, %xmm7, %xmm6 |
384 | movaps %xmm6, -0x60(%rdi) |
385 | |
386 | movaps -0x72(%rsi), %xmm8 |
387 | palignr $2, %xmm8, %xmm7 |
388 | movaps %xmm7, -0x70(%rdi) |
389 | |
390 | movaps -0x82(%rsi), %xmm9 |
391 | palignr $2, %xmm9, %xmm8 |
392 | movaps %xmm8, -0x80(%rdi) |
393 | |
394 | sub $0x80, %rdx |
395 | lea -0x80(%rdi), %rdi |
396 | lea -0x80(%rsi), %rsi |
397 | jae L(shl_2_bwd) |
398 | movdqu %xmm0, (%r8) |
399 | add $0x80, %rdx |
400 | sub %rdx, %rdi |
401 | sub %rdx, %rsi |
402 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
403 | |
404 | .p2align 4 |
405 | L(shl_3): |
406 | sub $0x80, %rdx |
407 | movaps -0x03(%rsi), %xmm1 |
408 | movaps 0x0d(%rsi), %xmm2 |
409 | movaps 0x1d(%rsi), %xmm3 |
410 | movaps 0x2d(%rsi), %xmm4 |
411 | movaps 0x3d(%rsi), %xmm5 |
412 | movaps 0x4d(%rsi), %xmm6 |
413 | movaps 0x5d(%rsi), %xmm7 |
414 | movaps 0x6d(%rsi), %xmm8 |
415 | movaps 0x7d(%rsi), %xmm9 |
416 | lea 0x80(%rsi), %rsi |
417 | palignr $3, %xmm8, %xmm9 |
418 | movaps %xmm9, 0x70(%rdi) |
419 | palignr $3, %xmm7, %xmm8 |
420 | movaps %xmm8, 0x60(%rdi) |
421 | palignr $3, %xmm6, %xmm7 |
422 | movaps %xmm7, 0x50(%rdi) |
423 | palignr $3, %xmm5, %xmm6 |
424 | movaps %xmm6, 0x40(%rdi) |
425 | palignr $3, %xmm4, %xmm5 |
426 | movaps %xmm5, 0x30(%rdi) |
427 | palignr $3, %xmm3, %xmm4 |
428 | movaps %xmm4, 0x20(%rdi) |
429 | palignr $3, %xmm2, %xmm3 |
430 | movaps %xmm3, 0x10(%rdi) |
431 | palignr $3, %xmm1, %xmm2 |
432 | movaps %xmm2, (%rdi) |
433 | lea 0x80(%rdi), %rdi |
434 | jae L(shl_3) |
435 | movdqu %xmm0, (%r8) |
436 | add $0x80, %rdx |
437 | add %rdx, %rdi |
438 | add %rdx, %rsi |
439 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
440 | |
441 | .p2align 4 |
442 | L(shl_3_bwd): |
443 | movaps -0x03(%rsi), %xmm1 |
444 | |
445 | movaps -0x13(%rsi), %xmm2 |
446 | palignr $3, %xmm2, %xmm1 |
447 | movaps %xmm1, -0x10(%rdi) |
448 | |
449 | movaps -0x23(%rsi), %xmm3 |
450 | palignr $3, %xmm3, %xmm2 |
451 | movaps %xmm2, -0x20(%rdi) |
452 | |
453 | movaps -0x33(%rsi), %xmm4 |
454 | palignr $3, %xmm4, %xmm3 |
455 | movaps %xmm3, -0x30(%rdi) |
456 | |
457 | movaps -0x43(%rsi), %xmm5 |
458 | palignr $3, %xmm5, %xmm4 |
459 | movaps %xmm4, -0x40(%rdi) |
460 | |
461 | movaps -0x53(%rsi), %xmm6 |
462 | palignr $3, %xmm6, %xmm5 |
463 | movaps %xmm5, -0x50(%rdi) |
464 | |
465 | movaps -0x63(%rsi), %xmm7 |
466 | palignr $3, %xmm7, %xmm6 |
467 | movaps %xmm6, -0x60(%rdi) |
468 | |
469 | movaps -0x73(%rsi), %xmm8 |
470 | palignr $3, %xmm8, %xmm7 |
471 | movaps %xmm7, -0x70(%rdi) |
472 | |
473 | movaps -0x83(%rsi), %xmm9 |
474 | palignr $3, %xmm9, %xmm8 |
475 | movaps %xmm8, -0x80(%rdi) |
476 | |
477 | sub $0x80, %rdx |
478 | lea -0x80(%rdi), %rdi |
479 | lea -0x80(%rsi), %rsi |
480 | jae L(shl_3_bwd) |
481 | movdqu %xmm0, (%r8) |
482 | add $0x80, %rdx |
483 | sub %rdx, %rdi |
484 | sub %rdx, %rsi |
485 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
486 | |
487 | .p2align 4 |
488 | L(shl_4): |
489 | sub $0x80, %rdx |
490 | movaps -0x04(%rsi), %xmm1 |
491 | movaps 0x0c(%rsi), %xmm2 |
492 | movaps 0x1c(%rsi), %xmm3 |
493 | movaps 0x2c(%rsi), %xmm4 |
494 | movaps 0x3c(%rsi), %xmm5 |
495 | movaps 0x4c(%rsi), %xmm6 |
496 | movaps 0x5c(%rsi), %xmm7 |
497 | movaps 0x6c(%rsi), %xmm8 |
498 | movaps 0x7c(%rsi), %xmm9 |
499 | lea 0x80(%rsi), %rsi |
500 | palignr $4, %xmm8, %xmm9 |
501 | movaps %xmm9, 0x70(%rdi) |
502 | palignr $4, %xmm7, %xmm8 |
503 | movaps %xmm8, 0x60(%rdi) |
504 | palignr $4, %xmm6, %xmm7 |
505 | movaps %xmm7, 0x50(%rdi) |
506 | palignr $4, %xmm5, %xmm6 |
507 | movaps %xmm6, 0x40(%rdi) |
508 | palignr $4, %xmm4, %xmm5 |
509 | movaps %xmm5, 0x30(%rdi) |
510 | palignr $4, %xmm3, %xmm4 |
511 | movaps %xmm4, 0x20(%rdi) |
512 | palignr $4, %xmm2, %xmm3 |
513 | movaps %xmm3, 0x10(%rdi) |
514 | palignr $4, %xmm1, %xmm2 |
515 | movaps %xmm2, (%rdi) |
516 | lea 0x80(%rdi), %rdi |
517 | jae L(shl_4) |
518 | movdqu %xmm0, (%r8) |
519 | add $0x80, %rdx |
520 | add %rdx, %rdi |
521 | add %rdx, %rsi |
522 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
523 | |
524 | .p2align 4 |
525 | L(shl_4_bwd): |
526 | movaps -0x04(%rsi), %xmm1 |
527 | |
528 | movaps -0x14(%rsi), %xmm2 |
529 | palignr $4, %xmm2, %xmm1 |
530 | movaps %xmm1, -0x10(%rdi) |
531 | |
532 | movaps -0x24(%rsi), %xmm3 |
533 | palignr $4, %xmm3, %xmm2 |
534 | movaps %xmm2, -0x20(%rdi) |
535 | |
536 | movaps -0x34(%rsi), %xmm4 |
537 | palignr $4, %xmm4, %xmm3 |
538 | movaps %xmm3, -0x30(%rdi) |
539 | |
540 | movaps -0x44(%rsi), %xmm5 |
541 | palignr $4, %xmm5, %xmm4 |
542 | movaps %xmm4, -0x40(%rdi) |
543 | |
544 | movaps -0x54(%rsi), %xmm6 |
545 | palignr $4, %xmm6, %xmm5 |
546 | movaps %xmm5, -0x50(%rdi) |
547 | |
548 | movaps -0x64(%rsi), %xmm7 |
549 | palignr $4, %xmm7, %xmm6 |
550 | movaps %xmm6, -0x60(%rdi) |
551 | |
552 | movaps -0x74(%rsi), %xmm8 |
553 | palignr $4, %xmm8, %xmm7 |
554 | movaps %xmm7, -0x70(%rdi) |
555 | |
556 | movaps -0x84(%rsi), %xmm9 |
557 | palignr $4, %xmm9, %xmm8 |
558 | movaps %xmm8, -0x80(%rdi) |
559 | |
560 | sub $0x80, %rdx |
561 | lea -0x80(%rdi), %rdi |
562 | lea -0x80(%rsi), %rsi |
563 | jae L(shl_4_bwd) |
564 | movdqu %xmm0, (%r8) |
565 | add $0x80, %rdx |
566 | sub %rdx, %rdi |
567 | sub %rdx, %rsi |
568 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
569 | |
570 | .p2align 4 |
571 | L(shl_5): |
572 | sub $0x80, %rdx |
573 | movaps -0x05(%rsi), %xmm1 |
574 | movaps 0x0b(%rsi), %xmm2 |
575 | movaps 0x1b(%rsi), %xmm3 |
576 | movaps 0x2b(%rsi), %xmm4 |
577 | movaps 0x3b(%rsi), %xmm5 |
578 | movaps 0x4b(%rsi), %xmm6 |
579 | movaps 0x5b(%rsi), %xmm7 |
580 | movaps 0x6b(%rsi), %xmm8 |
581 | movaps 0x7b(%rsi), %xmm9 |
582 | lea 0x80(%rsi), %rsi |
583 | palignr $5, %xmm8, %xmm9 |
584 | movaps %xmm9, 0x70(%rdi) |
585 | palignr $5, %xmm7, %xmm8 |
586 | movaps %xmm8, 0x60(%rdi) |
587 | palignr $5, %xmm6, %xmm7 |
588 | movaps %xmm7, 0x50(%rdi) |
589 | palignr $5, %xmm5, %xmm6 |
590 | movaps %xmm6, 0x40(%rdi) |
591 | palignr $5, %xmm4, %xmm5 |
592 | movaps %xmm5, 0x30(%rdi) |
593 | palignr $5, %xmm3, %xmm4 |
594 | movaps %xmm4, 0x20(%rdi) |
595 | palignr $5, %xmm2, %xmm3 |
596 | movaps %xmm3, 0x10(%rdi) |
597 | palignr $5, %xmm1, %xmm2 |
598 | movaps %xmm2, (%rdi) |
599 | lea 0x80(%rdi), %rdi |
600 | jae L(shl_5) |
601 | movdqu %xmm0, (%r8) |
602 | add $0x80, %rdx |
603 | add %rdx, %rdi |
604 | add %rdx, %rsi |
605 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
606 | |
607 | .p2align 4 |
608 | L(shl_5_bwd): |
609 | movaps -0x05(%rsi), %xmm1 |
610 | |
611 | movaps -0x15(%rsi), %xmm2 |
612 | palignr $5, %xmm2, %xmm1 |
613 | movaps %xmm1, -0x10(%rdi) |
614 | |
615 | movaps -0x25(%rsi), %xmm3 |
616 | palignr $5, %xmm3, %xmm2 |
617 | movaps %xmm2, -0x20(%rdi) |
618 | |
619 | movaps -0x35(%rsi), %xmm4 |
620 | palignr $5, %xmm4, %xmm3 |
621 | movaps %xmm3, -0x30(%rdi) |
622 | |
623 | movaps -0x45(%rsi), %xmm5 |
624 | palignr $5, %xmm5, %xmm4 |
625 | movaps %xmm4, -0x40(%rdi) |
626 | |
627 | movaps -0x55(%rsi), %xmm6 |
628 | palignr $5, %xmm6, %xmm5 |
629 | movaps %xmm5, -0x50(%rdi) |
630 | |
631 | movaps -0x65(%rsi), %xmm7 |
632 | palignr $5, %xmm7, %xmm6 |
633 | movaps %xmm6, -0x60(%rdi) |
634 | |
635 | movaps -0x75(%rsi), %xmm8 |
636 | palignr $5, %xmm8, %xmm7 |
637 | movaps %xmm7, -0x70(%rdi) |
638 | |
639 | movaps -0x85(%rsi), %xmm9 |
640 | palignr $5, %xmm9, %xmm8 |
641 | movaps %xmm8, -0x80(%rdi) |
642 | |
643 | sub $0x80, %rdx |
644 | lea -0x80(%rdi), %rdi |
645 | lea -0x80(%rsi), %rsi |
646 | jae L(shl_5_bwd) |
647 | movdqu %xmm0, (%r8) |
648 | add $0x80, %rdx |
649 | sub %rdx, %rdi |
650 | sub %rdx, %rsi |
651 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
652 | |
653 | .p2align 4 |
654 | L(shl_6): |
655 | sub $0x80, %rdx |
656 | movaps -0x06(%rsi), %xmm1 |
657 | movaps 0x0a(%rsi), %xmm2 |
658 | movaps 0x1a(%rsi), %xmm3 |
659 | movaps 0x2a(%rsi), %xmm4 |
660 | movaps 0x3a(%rsi), %xmm5 |
661 | movaps 0x4a(%rsi), %xmm6 |
662 | movaps 0x5a(%rsi), %xmm7 |
663 | movaps 0x6a(%rsi), %xmm8 |
664 | movaps 0x7a(%rsi), %xmm9 |
665 | lea 0x80(%rsi), %rsi |
666 | palignr $6, %xmm8, %xmm9 |
667 | movaps %xmm9, 0x70(%rdi) |
668 | palignr $6, %xmm7, %xmm8 |
669 | movaps %xmm8, 0x60(%rdi) |
670 | palignr $6, %xmm6, %xmm7 |
671 | movaps %xmm7, 0x50(%rdi) |
672 | palignr $6, %xmm5, %xmm6 |
673 | movaps %xmm6, 0x40(%rdi) |
674 | palignr $6, %xmm4, %xmm5 |
675 | movaps %xmm5, 0x30(%rdi) |
676 | palignr $6, %xmm3, %xmm4 |
677 | movaps %xmm4, 0x20(%rdi) |
678 | palignr $6, %xmm2, %xmm3 |
679 | movaps %xmm3, 0x10(%rdi) |
680 | palignr $6, %xmm1, %xmm2 |
681 | movaps %xmm2, (%rdi) |
682 | lea 0x80(%rdi), %rdi |
683 | jae L(shl_6) |
684 | movdqu %xmm0, (%r8) |
685 | add $0x80, %rdx |
686 | add %rdx, %rdi |
687 | add %rdx, %rsi |
688 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
689 | |
690 | .p2align 4 |
691 | L(shl_6_bwd): |
692 | movaps -0x06(%rsi), %xmm1 |
693 | |
694 | movaps -0x16(%rsi), %xmm2 |
695 | palignr $6, %xmm2, %xmm1 |
696 | movaps %xmm1, -0x10(%rdi) |
697 | |
698 | movaps -0x26(%rsi), %xmm3 |
699 | palignr $6, %xmm3, %xmm2 |
700 | movaps %xmm2, -0x20(%rdi) |
701 | |
702 | movaps -0x36(%rsi), %xmm4 |
703 | palignr $6, %xmm4, %xmm3 |
704 | movaps %xmm3, -0x30(%rdi) |
705 | |
706 | movaps -0x46(%rsi), %xmm5 |
707 | palignr $6, %xmm5, %xmm4 |
708 | movaps %xmm4, -0x40(%rdi) |
709 | |
710 | movaps -0x56(%rsi), %xmm6 |
711 | palignr $6, %xmm6, %xmm5 |
712 | movaps %xmm5, -0x50(%rdi) |
713 | |
714 | movaps -0x66(%rsi), %xmm7 |
715 | palignr $6, %xmm7, %xmm6 |
716 | movaps %xmm6, -0x60(%rdi) |
717 | |
718 | movaps -0x76(%rsi), %xmm8 |
719 | palignr $6, %xmm8, %xmm7 |
720 | movaps %xmm7, -0x70(%rdi) |
721 | |
722 | movaps -0x86(%rsi), %xmm9 |
723 | palignr $6, %xmm9, %xmm8 |
724 | movaps %xmm8, -0x80(%rdi) |
725 | |
726 | sub $0x80, %rdx |
727 | lea -0x80(%rdi), %rdi |
728 | lea -0x80(%rsi), %rsi |
729 | jae L(shl_6_bwd) |
730 | movdqu %xmm0, (%r8) |
731 | add $0x80, %rdx |
732 | sub %rdx, %rdi |
733 | sub %rdx, %rsi |
734 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
735 | |
736 | .p2align 4 |
737 | L(shl_7): |
738 | sub $0x80, %rdx |
739 | movaps -0x07(%rsi), %xmm1 |
740 | movaps 0x09(%rsi), %xmm2 |
741 | movaps 0x19(%rsi), %xmm3 |
742 | movaps 0x29(%rsi), %xmm4 |
743 | movaps 0x39(%rsi), %xmm5 |
744 | movaps 0x49(%rsi), %xmm6 |
745 | movaps 0x59(%rsi), %xmm7 |
746 | movaps 0x69(%rsi), %xmm8 |
747 | movaps 0x79(%rsi), %xmm9 |
748 | lea 0x80(%rsi), %rsi |
749 | palignr $7, %xmm8, %xmm9 |
750 | movaps %xmm9, 0x70(%rdi) |
751 | palignr $7, %xmm7, %xmm8 |
752 | movaps %xmm8, 0x60(%rdi) |
753 | palignr $7, %xmm6, %xmm7 |
754 | movaps %xmm7, 0x50(%rdi) |
755 | palignr $7, %xmm5, %xmm6 |
756 | movaps %xmm6, 0x40(%rdi) |
757 | palignr $7, %xmm4, %xmm5 |
758 | movaps %xmm5, 0x30(%rdi) |
759 | palignr $7, %xmm3, %xmm4 |
760 | movaps %xmm4, 0x20(%rdi) |
761 | palignr $7, %xmm2, %xmm3 |
762 | movaps %xmm3, 0x10(%rdi) |
763 | palignr $7, %xmm1, %xmm2 |
764 | movaps %xmm2, (%rdi) |
765 | lea 0x80(%rdi), %rdi |
766 | jae L(shl_7) |
767 | movdqu %xmm0, (%r8) |
768 | add $0x80, %rdx |
769 | add %rdx, %rdi |
770 | add %rdx, %rsi |
771 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
772 | |
773 | .p2align 4 |
774 | L(shl_7_bwd): |
775 | movaps -0x07(%rsi), %xmm1 |
776 | |
777 | movaps -0x17(%rsi), %xmm2 |
778 | palignr $7, %xmm2, %xmm1 |
779 | movaps %xmm1, -0x10(%rdi) |
780 | |
781 | movaps -0x27(%rsi), %xmm3 |
782 | palignr $7, %xmm3, %xmm2 |
783 | movaps %xmm2, -0x20(%rdi) |
784 | |
785 | movaps -0x37(%rsi), %xmm4 |
786 | palignr $7, %xmm4, %xmm3 |
787 | movaps %xmm3, -0x30(%rdi) |
788 | |
789 | movaps -0x47(%rsi), %xmm5 |
790 | palignr $7, %xmm5, %xmm4 |
791 | movaps %xmm4, -0x40(%rdi) |
792 | |
793 | movaps -0x57(%rsi), %xmm6 |
794 | palignr $7, %xmm6, %xmm5 |
795 | movaps %xmm5, -0x50(%rdi) |
796 | |
797 | movaps -0x67(%rsi), %xmm7 |
798 | palignr $7, %xmm7, %xmm6 |
799 | movaps %xmm6, -0x60(%rdi) |
800 | |
801 | movaps -0x77(%rsi), %xmm8 |
802 | palignr $7, %xmm8, %xmm7 |
803 | movaps %xmm7, -0x70(%rdi) |
804 | |
805 | movaps -0x87(%rsi), %xmm9 |
806 | palignr $7, %xmm9, %xmm8 |
807 | movaps %xmm8, -0x80(%rdi) |
808 | |
809 | sub $0x80, %rdx |
810 | lea -0x80(%rdi), %rdi |
811 | lea -0x80(%rsi), %rsi |
812 | jae L(shl_7_bwd) |
813 | movdqu %xmm0, (%r8) |
814 | add $0x80, %rdx |
815 | sub %rdx, %rdi |
816 | sub %rdx, %rsi |
817 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
818 | |
819 | .p2align 4 |
820 | L(shl_8): |
821 | sub $0x80, %rdx |
822 | movaps -0x08(%rsi), %xmm1 |
823 | movaps 0x08(%rsi), %xmm2 |
824 | movaps 0x18(%rsi), %xmm3 |
825 | movaps 0x28(%rsi), %xmm4 |
826 | movaps 0x38(%rsi), %xmm5 |
827 | movaps 0x48(%rsi), %xmm6 |
828 | movaps 0x58(%rsi), %xmm7 |
829 | movaps 0x68(%rsi), %xmm8 |
830 | movaps 0x78(%rsi), %xmm9 |
831 | lea 0x80(%rsi), %rsi |
832 | palignr $8, %xmm8, %xmm9 |
833 | movaps %xmm9, 0x70(%rdi) |
834 | palignr $8, %xmm7, %xmm8 |
835 | movaps %xmm8, 0x60(%rdi) |
836 | palignr $8, %xmm6, %xmm7 |
837 | movaps %xmm7, 0x50(%rdi) |
838 | palignr $8, %xmm5, %xmm6 |
839 | movaps %xmm6, 0x40(%rdi) |
840 | palignr $8, %xmm4, %xmm5 |
841 | movaps %xmm5, 0x30(%rdi) |
842 | palignr $8, %xmm3, %xmm4 |
843 | movaps %xmm4, 0x20(%rdi) |
844 | palignr $8, %xmm2, %xmm3 |
845 | movaps %xmm3, 0x10(%rdi) |
846 | palignr $8, %xmm1, %xmm2 |
847 | movaps %xmm2, (%rdi) |
848 | lea 0x80(%rdi), %rdi |
849 | jae L(shl_8) |
850 | movdqu %xmm0, (%r8) |
851 | add $0x80, %rdx |
852 | add %rdx, %rdi |
853 | add %rdx, %rsi |
854 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
855 | |
856 | .p2align 4 |
857 | L(shl_8_bwd): |
858 | movaps -0x08(%rsi), %xmm1 |
859 | |
860 | movaps -0x18(%rsi), %xmm2 |
861 | palignr $8, %xmm2, %xmm1 |
862 | movaps %xmm1, -0x10(%rdi) |
863 | |
864 | movaps -0x28(%rsi), %xmm3 |
865 | palignr $8, %xmm3, %xmm2 |
866 | movaps %xmm2, -0x20(%rdi) |
867 | |
868 | movaps -0x38(%rsi), %xmm4 |
869 | palignr $8, %xmm4, %xmm3 |
870 | movaps %xmm3, -0x30(%rdi) |
871 | |
872 | movaps -0x48(%rsi), %xmm5 |
873 | palignr $8, %xmm5, %xmm4 |
874 | movaps %xmm4, -0x40(%rdi) |
875 | |
876 | movaps -0x58(%rsi), %xmm6 |
877 | palignr $8, %xmm6, %xmm5 |
878 | movaps %xmm5, -0x50(%rdi) |
879 | |
880 | movaps -0x68(%rsi), %xmm7 |
881 | palignr $8, %xmm7, %xmm6 |
882 | movaps %xmm6, -0x60(%rdi) |
883 | |
884 | movaps -0x78(%rsi), %xmm8 |
885 | palignr $8, %xmm8, %xmm7 |
886 | movaps %xmm7, -0x70(%rdi) |
887 | |
888 | movaps -0x88(%rsi), %xmm9 |
889 | palignr $8, %xmm9, %xmm8 |
890 | movaps %xmm8, -0x80(%rdi) |
891 | |
892 | sub $0x80, %rdx |
893 | lea -0x80(%rdi), %rdi |
894 | lea -0x80(%rsi), %rsi |
895 | jae L(shl_8_bwd) |
896 | L(shl_8_end_bwd): |
897 | movdqu %xmm0, (%r8) |
898 | add $0x80, %rdx |
899 | sub %rdx, %rdi |
900 | sub %rdx, %rsi |
901 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
902 | |
903 | .p2align 4 |
904 | L(shl_9): |
905 | sub $0x80, %rdx |
906 | movaps -0x09(%rsi), %xmm1 |
907 | movaps 0x07(%rsi), %xmm2 |
908 | movaps 0x17(%rsi), %xmm3 |
909 | movaps 0x27(%rsi), %xmm4 |
910 | movaps 0x37(%rsi), %xmm5 |
911 | movaps 0x47(%rsi), %xmm6 |
912 | movaps 0x57(%rsi), %xmm7 |
913 | movaps 0x67(%rsi), %xmm8 |
914 | movaps 0x77(%rsi), %xmm9 |
915 | lea 0x80(%rsi), %rsi |
916 | palignr $9, %xmm8, %xmm9 |
917 | movaps %xmm9, 0x70(%rdi) |
918 | palignr $9, %xmm7, %xmm8 |
919 | movaps %xmm8, 0x60(%rdi) |
920 | palignr $9, %xmm6, %xmm7 |
921 | movaps %xmm7, 0x50(%rdi) |
922 | palignr $9, %xmm5, %xmm6 |
923 | movaps %xmm6, 0x40(%rdi) |
924 | palignr $9, %xmm4, %xmm5 |
925 | movaps %xmm5, 0x30(%rdi) |
926 | palignr $9, %xmm3, %xmm4 |
927 | movaps %xmm4, 0x20(%rdi) |
928 | palignr $9, %xmm2, %xmm3 |
929 | movaps %xmm3, 0x10(%rdi) |
930 | palignr $9, %xmm1, %xmm2 |
931 | movaps %xmm2, (%rdi) |
932 | lea 0x80(%rdi), %rdi |
933 | jae L(shl_9) |
934 | movdqu %xmm0, (%r8) |
935 | add $0x80, %rdx |
936 | add %rdx, %rdi |
937 | add %rdx, %rsi |
938 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
939 | |
940 | .p2align 4 |
941 | L(shl_9_bwd): |
942 | movaps -0x09(%rsi), %xmm1 |
943 | |
944 | movaps -0x19(%rsi), %xmm2 |
945 | palignr $9, %xmm2, %xmm1 |
946 | movaps %xmm1, -0x10(%rdi) |
947 | |
948 | movaps -0x29(%rsi), %xmm3 |
949 | palignr $9, %xmm3, %xmm2 |
950 | movaps %xmm2, -0x20(%rdi) |
951 | |
952 | movaps -0x39(%rsi), %xmm4 |
953 | palignr $9, %xmm4, %xmm3 |
954 | movaps %xmm3, -0x30(%rdi) |
955 | |
956 | movaps -0x49(%rsi), %xmm5 |
957 | palignr $9, %xmm5, %xmm4 |
958 | movaps %xmm4, -0x40(%rdi) |
959 | |
960 | movaps -0x59(%rsi), %xmm6 |
961 | palignr $9, %xmm6, %xmm5 |
962 | movaps %xmm5, -0x50(%rdi) |
963 | |
964 | movaps -0x69(%rsi), %xmm7 |
965 | palignr $9, %xmm7, %xmm6 |
966 | movaps %xmm6, -0x60(%rdi) |
967 | |
968 | movaps -0x79(%rsi), %xmm8 |
969 | palignr $9, %xmm8, %xmm7 |
970 | movaps %xmm7, -0x70(%rdi) |
971 | |
972 | movaps -0x89(%rsi), %xmm9 |
973 | palignr $9, %xmm9, %xmm8 |
974 | movaps %xmm8, -0x80(%rdi) |
975 | |
976 | sub $0x80, %rdx |
977 | lea -0x80(%rdi), %rdi |
978 | lea -0x80(%rsi), %rsi |
979 | jae L(shl_9_bwd) |
980 | movdqu %xmm0, (%r8) |
981 | add $0x80, %rdx |
982 | sub %rdx, %rdi |
983 | sub %rdx, %rsi |
984 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
985 | |
986 | .p2align 4 |
987 | L(shl_10): |
988 | sub $0x80, %rdx |
989 | movaps -0x0a(%rsi), %xmm1 |
990 | movaps 0x06(%rsi), %xmm2 |
991 | movaps 0x16(%rsi), %xmm3 |
992 | movaps 0x26(%rsi), %xmm4 |
993 | movaps 0x36(%rsi), %xmm5 |
994 | movaps 0x46(%rsi), %xmm6 |
995 | movaps 0x56(%rsi), %xmm7 |
996 | movaps 0x66(%rsi), %xmm8 |
997 | movaps 0x76(%rsi), %xmm9 |
998 | lea 0x80(%rsi), %rsi |
999 | palignr $10, %xmm8, %xmm9 |
1000 | movaps %xmm9, 0x70(%rdi) |
1001 | palignr $10, %xmm7, %xmm8 |
1002 | movaps %xmm8, 0x60(%rdi) |
1003 | palignr $10, %xmm6, %xmm7 |
1004 | movaps %xmm7, 0x50(%rdi) |
1005 | palignr $10, %xmm5, %xmm6 |
1006 | movaps %xmm6, 0x40(%rdi) |
1007 | palignr $10, %xmm4, %xmm5 |
1008 | movaps %xmm5, 0x30(%rdi) |
1009 | palignr $10, %xmm3, %xmm4 |
1010 | movaps %xmm4, 0x20(%rdi) |
1011 | palignr $10, %xmm2, %xmm3 |
1012 | movaps %xmm3, 0x10(%rdi) |
1013 | palignr $10, %xmm1, %xmm2 |
1014 | movaps %xmm2, (%rdi) |
1015 | lea 0x80(%rdi), %rdi |
1016 | jae L(shl_10) |
1017 | movdqu %xmm0, (%r8) |
1018 | add $0x80, %rdx |
1019 | add %rdx, %rdi |
1020 | add %rdx, %rsi |
1021 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1022 | |
1023 | .p2align 4 |
1024 | L(shl_10_bwd): |
1025 | movaps -0x0a(%rsi), %xmm1 |
1026 | |
1027 | movaps -0x1a(%rsi), %xmm2 |
1028 | palignr $10, %xmm2, %xmm1 |
1029 | movaps %xmm1, -0x10(%rdi) |
1030 | |
1031 | movaps -0x2a(%rsi), %xmm3 |
1032 | palignr $10, %xmm3, %xmm2 |
1033 | movaps %xmm2, -0x20(%rdi) |
1034 | |
1035 | movaps -0x3a(%rsi), %xmm4 |
1036 | palignr $10, %xmm4, %xmm3 |
1037 | movaps %xmm3, -0x30(%rdi) |
1038 | |
1039 | movaps -0x4a(%rsi), %xmm5 |
1040 | palignr $10, %xmm5, %xmm4 |
1041 | movaps %xmm4, -0x40(%rdi) |
1042 | |
1043 | movaps -0x5a(%rsi), %xmm6 |
1044 | palignr $10, %xmm6, %xmm5 |
1045 | movaps %xmm5, -0x50(%rdi) |
1046 | |
1047 | movaps -0x6a(%rsi), %xmm7 |
1048 | palignr $10, %xmm7, %xmm6 |
1049 | movaps %xmm6, -0x60(%rdi) |
1050 | |
1051 | movaps -0x7a(%rsi), %xmm8 |
1052 | palignr $10, %xmm8, %xmm7 |
1053 | movaps %xmm7, -0x70(%rdi) |
1054 | |
1055 | movaps -0x8a(%rsi), %xmm9 |
1056 | palignr $10, %xmm9, %xmm8 |
1057 | movaps %xmm8, -0x80(%rdi) |
1058 | |
1059 | sub $0x80, %rdx |
1060 | lea -0x80(%rdi), %rdi |
1061 | lea -0x80(%rsi), %rsi |
1062 | jae L(shl_10_bwd) |
1063 | movdqu %xmm0, (%r8) |
1064 | add $0x80, %rdx |
1065 | sub %rdx, %rdi |
1066 | sub %rdx, %rsi |
1067 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1068 | |
1069 | .p2align 4 |
1070 | L(shl_11): |
1071 | sub $0x80, %rdx |
1072 | movaps -0x0b(%rsi), %xmm1 |
1073 | movaps 0x05(%rsi), %xmm2 |
1074 | movaps 0x15(%rsi), %xmm3 |
1075 | movaps 0x25(%rsi), %xmm4 |
1076 | movaps 0x35(%rsi), %xmm5 |
1077 | movaps 0x45(%rsi), %xmm6 |
1078 | movaps 0x55(%rsi), %xmm7 |
1079 | movaps 0x65(%rsi), %xmm8 |
1080 | movaps 0x75(%rsi), %xmm9 |
1081 | lea 0x80(%rsi), %rsi |
1082 | palignr $11, %xmm8, %xmm9 |
1083 | movaps %xmm9, 0x70(%rdi) |
1084 | palignr $11, %xmm7, %xmm8 |
1085 | movaps %xmm8, 0x60(%rdi) |
1086 | palignr $11, %xmm6, %xmm7 |
1087 | movaps %xmm7, 0x50(%rdi) |
1088 | palignr $11, %xmm5, %xmm6 |
1089 | movaps %xmm6, 0x40(%rdi) |
1090 | palignr $11, %xmm4, %xmm5 |
1091 | movaps %xmm5, 0x30(%rdi) |
1092 | palignr $11, %xmm3, %xmm4 |
1093 | movaps %xmm4, 0x20(%rdi) |
1094 | palignr $11, %xmm2, %xmm3 |
1095 | movaps %xmm3, 0x10(%rdi) |
1096 | palignr $11, %xmm1, %xmm2 |
1097 | movaps %xmm2, (%rdi) |
1098 | lea 0x80(%rdi), %rdi |
1099 | jae L(shl_11) |
1100 | movdqu %xmm0, (%r8) |
1101 | add $0x80, %rdx |
1102 | add %rdx, %rdi |
1103 | add %rdx, %rsi |
1104 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1105 | |
1106 | .p2align 4 |
1107 | L(shl_11_bwd): |
1108 | movaps -0x0b(%rsi), %xmm1 |
1109 | |
1110 | movaps -0x1b(%rsi), %xmm2 |
1111 | palignr $11, %xmm2, %xmm1 |
1112 | movaps %xmm1, -0x10(%rdi) |
1113 | |
1114 | movaps -0x2b(%rsi), %xmm3 |
1115 | palignr $11, %xmm3, %xmm2 |
1116 | movaps %xmm2, -0x20(%rdi) |
1117 | |
1118 | movaps -0x3b(%rsi), %xmm4 |
1119 | palignr $11, %xmm4, %xmm3 |
1120 | movaps %xmm3, -0x30(%rdi) |
1121 | |
1122 | movaps -0x4b(%rsi), %xmm5 |
1123 | palignr $11, %xmm5, %xmm4 |
1124 | movaps %xmm4, -0x40(%rdi) |
1125 | |
1126 | movaps -0x5b(%rsi), %xmm6 |
1127 | palignr $11, %xmm6, %xmm5 |
1128 | movaps %xmm5, -0x50(%rdi) |
1129 | |
1130 | movaps -0x6b(%rsi), %xmm7 |
1131 | palignr $11, %xmm7, %xmm6 |
1132 | movaps %xmm6, -0x60(%rdi) |
1133 | |
1134 | movaps -0x7b(%rsi), %xmm8 |
1135 | palignr $11, %xmm8, %xmm7 |
1136 | movaps %xmm7, -0x70(%rdi) |
1137 | |
1138 | movaps -0x8b(%rsi), %xmm9 |
1139 | palignr $11, %xmm9, %xmm8 |
1140 | movaps %xmm8, -0x80(%rdi) |
1141 | |
1142 | sub $0x80, %rdx |
1143 | lea -0x80(%rdi), %rdi |
1144 | lea -0x80(%rsi), %rsi |
1145 | jae L(shl_11_bwd) |
1146 | movdqu %xmm0, (%r8) |
1147 | add $0x80, %rdx |
1148 | sub %rdx, %rdi |
1149 | sub %rdx, %rsi |
1150 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1151 | |
1152 | .p2align 4 |
1153 | L(shl_12): |
1154 | sub $0x80, %rdx |
1155 | movdqa -0x0c(%rsi), %xmm1 |
1156 | movaps 0x04(%rsi), %xmm2 |
1157 | movaps 0x14(%rsi), %xmm3 |
1158 | movaps 0x24(%rsi), %xmm4 |
1159 | movaps 0x34(%rsi), %xmm5 |
1160 | movaps 0x44(%rsi), %xmm6 |
1161 | movaps 0x54(%rsi), %xmm7 |
1162 | movaps 0x64(%rsi), %xmm8 |
1163 | movaps 0x74(%rsi), %xmm9 |
1164 | lea 0x80(%rsi), %rsi |
1165 | palignr $12, %xmm8, %xmm9 |
1166 | movaps %xmm9, 0x70(%rdi) |
1167 | palignr $12, %xmm7, %xmm8 |
1168 | movaps %xmm8, 0x60(%rdi) |
1169 | palignr $12, %xmm6, %xmm7 |
1170 | movaps %xmm7, 0x50(%rdi) |
1171 | palignr $12, %xmm5, %xmm6 |
1172 | movaps %xmm6, 0x40(%rdi) |
1173 | palignr $12, %xmm4, %xmm5 |
1174 | movaps %xmm5, 0x30(%rdi) |
1175 | palignr $12, %xmm3, %xmm4 |
1176 | movaps %xmm4, 0x20(%rdi) |
1177 | palignr $12, %xmm2, %xmm3 |
1178 | movaps %xmm3, 0x10(%rdi) |
1179 | palignr $12, %xmm1, %xmm2 |
1180 | movaps %xmm2, (%rdi) |
1181 | |
1182 | lea 0x80(%rdi), %rdi |
1183 | jae L(shl_12) |
1184 | movdqu %xmm0, (%r8) |
1185 | add $0x80, %rdx |
1186 | add %rdx, %rdi |
1187 | add %rdx, %rsi |
1188 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1189 | |
1190 | .p2align 4 |
1191 | L(shl_12_bwd): |
1192 | movaps -0x0c(%rsi), %xmm1 |
1193 | |
1194 | movaps -0x1c(%rsi), %xmm2 |
1195 | palignr $12, %xmm2, %xmm1 |
1196 | movaps %xmm1, -0x10(%rdi) |
1197 | |
1198 | movaps -0x2c(%rsi), %xmm3 |
1199 | palignr $12, %xmm3, %xmm2 |
1200 | movaps %xmm2, -0x20(%rdi) |
1201 | |
1202 | movaps -0x3c(%rsi), %xmm4 |
1203 | palignr $12, %xmm4, %xmm3 |
1204 | movaps %xmm3, -0x30(%rdi) |
1205 | |
1206 | movaps -0x4c(%rsi), %xmm5 |
1207 | palignr $12, %xmm5, %xmm4 |
1208 | movaps %xmm4, -0x40(%rdi) |
1209 | |
1210 | movaps -0x5c(%rsi), %xmm6 |
1211 | palignr $12, %xmm6, %xmm5 |
1212 | movaps %xmm5, -0x50(%rdi) |
1213 | |
1214 | movaps -0x6c(%rsi), %xmm7 |
1215 | palignr $12, %xmm7, %xmm6 |
1216 | movaps %xmm6, -0x60(%rdi) |
1217 | |
1218 | movaps -0x7c(%rsi), %xmm8 |
1219 | palignr $12, %xmm8, %xmm7 |
1220 | movaps %xmm7, -0x70(%rdi) |
1221 | |
1222 | movaps -0x8c(%rsi), %xmm9 |
1223 | palignr $12, %xmm9, %xmm8 |
1224 | movaps %xmm8, -0x80(%rdi) |
1225 | |
1226 | sub $0x80, %rdx |
1227 | lea -0x80(%rdi), %rdi |
1228 | lea -0x80(%rsi), %rsi |
1229 | jae L(shl_12_bwd) |
1230 | movdqu %xmm0, (%r8) |
1231 | add $0x80, %rdx |
1232 | sub %rdx, %rdi |
1233 | sub %rdx, %rsi |
1234 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1235 | |
1236 | .p2align 4 |
1237 | L(shl_13): |
1238 | sub $0x80, %rdx |
1239 | movaps -0x0d(%rsi), %xmm1 |
1240 | movaps 0x03(%rsi), %xmm2 |
1241 | movaps 0x13(%rsi), %xmm3 |
1242 | movaps 0x23(%rsi), %xmm4 |
1243 | movaps 0x33(%rsi), %xmm5 |
1244 | movaps 0x43(%rsi), %xmm6 |
1245 | movaps 0x53(%rsi), %xmm7 |
1246 | movaps 0x63(%rsi), %xmm8 |
1247 | movaps 0x73(%rsi), %xmm9 |
1248 | lea 0x80(%rsi), %rsi |
1249 | palignr $13, %xmm8, %xmm9 |
1250 | movaps %xmm9, 0x70(%rdi) |
1251 | palignr $13, %xmm7, %xmm8 |
1252 | movaps %xmm8, 0x60(%rdi) |
1253 | palignr $13, %xmm6, %xmm7 |
1254 | movaps %xmm7, 0x50(%rdi) |
1255 | palignr $13, %xmm5, %xmm6 |
1256 | movaps %xmm6, 0x40(%rdi) |
1257 | palignr $13, %xmm4, %xmm5 |
1258 | movaps %xmm5, 0x30(%rdi) |
1259 | palignr $13, %xmm3, %xmm4 |
1260 | movaps %xmm4, 0x20(%rdi) |
1261 | palignr $13, %xmm2, %xmm3 |
1262 | movaps %xmm3, 0x10(%rdi) |
1263 | palignr $13, %xmm1, %xmm2 |
1264 | movaps %xmm2, (%rdi) |
1265 | lea 0x80(%rdi), %rdi |
1266 | jae L(shl_13) |
1267 | movdqu %xmm0, (%r8) |
1268 | add $0x80, %rdx |
1269 | add %rdx, %rdi |
1270 | add %rdx, %rsi |
1271 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1272 | |
1273 | .p2align 4 |
1274 | L(shl_13_bwd): |
1275 | movaps -0x0d(%rsi), %xmm1 |
1276 | |
1277 | movaps -0x1d(%rsi), %xmm2 |
1278 | palignr $13, %xmm2, %xmm1 |
1279 | movaps %xmm1, -0x10(%rdi) |
1280 | |
1281 | movaps -0x2d(%rsi), %xmm3 |
1282 | palignr $13, %xmm3, %xmm2 |
1283 | movaps %xmm2, -0x20(%rdi) |
1284 | |
1285 | movaps -0x3d(%rsi), %xmm4 |
1286 | palignr $13, %xmm4, %xmm3 |
1287 | movaps %xmm3, -0x30(%rdi) |
1288 | |
1289 | movaps -0x4d(%rsi), %xmm5 |
1290 | palignr $13, %xmm5, %xmm4 |
1291 | movaps %xmm4, -0x40(%rdi) |
1292 | |
1293 | movaps -0x5d(%rsi), %xmm6 |
1294 | palignr $13, %xmm6, %xmm5 |
1295 | movaps %xmm5, -0x50(%rdi) |
1296 | |
1297 | movaps -0x6d(%rsi), %xmm7 |
1298 | palignr $13, %xmm7, %xmm6 |
1299 | movaps %xmm6, -0x60(%rdi) |
1300 | |
1301 | movaps -0x7d(%rsi), %xmm8 |
1302 | palignr $13, %xmm8, %xmm7 |
1303 | movaps %xmm7, -0x70(%rdi) |
1304 | |
1305 | movaps -0x8d(%rsi), %xmm9 |
1306 | palignr $13, %xmm9, %xmm8 |
1307 | movaps %xmm8, -0x80(%rdi) |
1308 | |
1309 | sub $0x80, %rdx |
1310 | lea -0x80(%rdi), %rdi |
1311 | lea -0x80(%rsi), %rsi |
1312 | jae L(shl_13_bwd) |
1313 | movdqu %xmm0, (%r8) |
1314 | add $0x80, %rdx |
1315 | sub %rdx, %rdi |
1316 | sub %rdx, %rsi |
1317 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1318 | |
1319 | .p2align 4 |
1320 | L(shl_14): |
1321 | sub $0x80, %rdx |
1322 | movaps -0x0e(%rsi), %xmm1 |
1323 | movaps 0x02(%rsi), %xmm2 |
1324 | movaps 0x12(%rsi), %xmm3 |
1325 | movaps 0x22(%rsi), %xmm4 |
1326 | movaps 0x32(%rsi), %xmm5 |
1327 | movaps 0x42(%rsi), %xmm6 |
1328 | movaps 0x52(%rsi), %xmm7 |
1329 | movaps 0x62(%rsi), %xmm8 |
1330 | movaps 0x72(%rsi), %xmm9 |
1331 | lea 0x80(%rsi), %rsi |
1332 | palignr $14, %xmm8, %xmm9 |
1333 | movaps %xmm9, 0x70(%rdi) |
1334 | palignr $14, %xmm7, %xmm8 |
1335 | movaps %xmm8, 0x60(%rdi) |
1336 | palignr $14, %xmm6, %xmm7 |
1337 | movaps %xmm7, 0x50(%rdi) |
1338 | palignr $14, %xmm5, %xmm6 |
1339 | movaps %xmm6, 0x40(%rdi) |
1340 | palignr $14, %xmm4, %xmm5 |
1341 | movaps %xmm5, 0x30(%rdi) |
1342 | palignr $14, %xmm3, %xmm4 |
1343 | movaps %xmm4, 0x20(%rdi) |
1344 | palignr $14, %xmm2, %xmm3 |
1345 | movaps %xmm3, 0x10(%rdi) |
1346 | palignr $14, %xmm1, %xmm2 |
1347 | movaps %xmm2, (%rdi) |
1348 | lea 0x80(%rdi), %rdi |
1349 | jae L(shl_14) |
1350 | movdqu %xmm0, (%r8) |
1351 | add $0x80, %rdx |
1352 | add %rdx, %rdi |
1353 | add %rdx, %rsi |
1354 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1355 | |
1356 | .p2align 4 |
1357 | L(shl_14_bwd): |
1358 | movaps -0x0e(%rsi), %xmm1 |
1359 | |
1360 | movaps -0x1e(%rsi), %xmm2 |
1361 | palignr $14, %xmm2, %xmm1 |
1362 | movaps %xmm1, -0x10(%rdi) |
1363 | |
1364 | movaps -0x2e(%rsi), %xmm3 |
1365 | palignr $14, %xmm3, %xmm2 |
1366 | movaps %xmm2, -0x20(%rdi) |
1367 | |
1368 | movaps -0x3e(%rsi), %xmm4 |
1369 | palignr $14, %xmm4, %xmm3 |
1370 | movaps %xmm3, -0x30(%rdi) |
1371 | |
1372 | movaps -0x4e(%rsi), %xmm5 |
1373 | palignr $14, %xmm5, %xmm4 |
1374 | movaps %xmm4, -0x40(%rdi) |
1375 | |
1376 | movaps -0x5e(%rsi), %xmm6 |
1377 | palignr $14, %xmm6, %xmm5 |
1378 | movaps %xmm5, -0x50(%rdi) |
1379 | |
1380 | movaps -0x6e(%rsi), %xmm7 |
1381 | palignr $14, %xmm7, %xmm6 |
1382 | movaps %xmm6, -0x60(%rdi) |
1383 | |
1384 | movaps -0x7e(%rsi), %xmm8 |
1385 | palignr $14, %xmm8, %xmm7 |
1386 | movaps %xmm7, -0x70(%rdi) |
1387 | |
1388 | movaps -0x8e(%rsi), %xmm9 |
1389 | palignr $14, %xmm9, %xmm8 |
1390 | movaps %xmm8, -0x80(%rdi) |
1391 | |
1392 | sub $0x80, %rdx |
1393 | lea -0x80(%rdi), %rdi |
1394 | lea -0x80(%rsi), %rsi |
1395 | jae L(shl_14_bwd) |
1396 | movdqu %xmm0, (%r8) |
1397 | add $0x80, %rdx |
1398 | sub %rdx, %rdi |
1399 | sub %rdx, %rsi |
1400 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1401 | |
1402 | .p2align 4 |
1403 | L(shl_15): |
1404 | sub $0x80, %rdx |
1405 | movaps -0x0f(%rsi), %xmm1 |
1406 | movaps 0x01(%rsi), %xmm2 |
1407 | movaps 0x11(%rsi), %xmm3 |
1408 | movaps 0x21(%rsi), %xmm4 |
1409 | movaps 0x31(%rsi), %xmm5 |
1410 | movaps 0x41(%rsi), %xmm6 |
1411 | movaps 0x51(%rsi), %xmm7 |
1412 | movaps 0x61(%rsi), %xmm8 |
1413 | movaps 0x71(%rsi), %xmm9 |
1414 | lea 0x80(%rsi), %rsi |
1415 | palignr $15, %xmm8, %xmm9 |
1416 | movaps %xmm9, 0x70(%rdi) |
1417 | palignr $15, %xmm7, %xmm8 |
1418 | movaps %xmm8, 0x60(%rdi) |
1419 | palignr $15, %xmm6, %xmm7 |
1420 | movaps %xmm7, 0x50(%rdi) |
1421 | palignr $15, %xmm5, %xmm6 |
1422 | movaps %xmm6, 0x40(%rdi) |
1423 | palignr $15, %xmm4, %xmm5 |
1424 | movaps %xmm5, 0x30(%rdi) |
1425 | palignr $15, %xmm3, %xmm4 |
1426 | movaps %xmm4, 0x20(%rdi) |
1427 | palignr $15, %xmm2, %xmm3 |
1428 | movaps %xmm3, 0x10(%rdi) |
1429 | palignr $15, %xmm1, %xmm2 |
1430 | movaps %xmm2, (%rdi) |
1431 | lea 0x80(%rdi), %rdi |
1432 | jae L(shl_15) |
1433 | movdqu %xmm0, (%r8) |
1434 | add $0x80, %rdx |
1435 | add %rdx, %rdi |
1436 | add %rdx, %rsi |
1437 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1438 | |
1439 | .p2align 4 |
1440 | L(shl_15_bwd): |
1441 | movaps -0x0f(%rsi), %xmm1 |
1442 | |
1443 | movaps -0x1f(%rsi), %xmm2 |
1444 | palignr $15, %xmm2, %xmm1 |
1445 | movaps %xmm1, -0x10(%rdi) |
1446 | |
1447 | movaps -0x2f(%rsi), %xmm3 |
1448 | palignr $15, %xmm3, %xmm2 |
1449 | movaps %xmm2, -0x20(%rdi) |
1450 | |
1451 | movaps -0x3f(%rsi), %xmm4 |
1452 | palignr $15, %xmm4, %xmm3 |
1453 | movaps %xmm3, -0x30(%rdi) |
1454 | |
1455 | movaps -0x4f(%rsi), %xmm5 |
1456 | palignr $15, %xmm5, %xmm4 |
1457 | movaps %xmm4, -0x40(%rdi) |
1458 | |
1459 | movaps -0x5f(%rsi), %xmm6 |
1460 | palignr $15, %xmm6, %xmm5 |
1461 | movaps %xmm5, -0x50(%rdi) |
1462 | |
1463 | movaps -0x6f(%rsi), %xmm7 |
1464 | palignr $15, %xmm7, %xmm6 |
1465 | movaps %xmm6, -0x60(%rdi) |
1466 | |
1467 | movaps -0x7f(%rsi), %xmm8 |
1468 | palignr $15, %xmm8, %xmm7 |
1469 | movaps %xmm7, -0x70(%rdi) |
1470 | |
1471 | movaps -0x8f(%rsi), %xmm9 |
1472 | palignr $15, %xmm9, %xmm8 |
1473 | movaps %xmm8, -0x80(%rdi) |
1474 | |
1475 | sub $0x80, %rdx |
1476 | lea -0x80(%rdi), %rdi |
1477 | lea -0x80(%rsi), %rsi |
1478 | jae L(shl_15_bwd) |
1479 | movdqu %xmm0, (%r8) |
1480 | add $0x80, %rdx |
1481 | sub %rdx, %rdi |
1482 | sub %rdx, %rsi |
1483 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1484 | |
1485 | .p2align 4 |
1486 | L(gobble_mem_fwd): |
1487 | movdqu (%rsi), %xmm1 |
1488 | movdqu %xmm0, (%r8) |
1489 | movdqa %xmm1, (%rdi) |
1490 | sub $16, %rdx |
1491 | add $16, %rsi |
1492 | add $16, %rdi |
1493 | |
1494 | #ifdef SHARED_CACHE_SIZE_HALF |
1495 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1496 | #else |
1497 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1498 | #endif |
1499 | #ifdef USE_AS_MEMMOVE |
1500 | mov %rsi, %r9 |
1501 | sub %rdi, %r9 |
1502 | cmp %rdx, %r9 |
1503 | jae L(memmove_is_memcpy_fwd) |
1504 | cmp %rcx, %r9 |
1505 | jbe L(ll_cache_copy_fwd_start) |
1506 | L(memmove_is_memcpy_fwd): |
1507 | #endif |
1508 | cmp %rcx, %rdx |
1509 | ja L(bigger_in_fwd) |
1510 | mov %rdx, %rcx |
1511 | L(bigger_in_fwd): |
1512 | sub %rcx, %rdx |
1513 | cmp $0x1000, %rdx |
1514 | jbe L(ll_cache_copy_fwd) |
1515 | |
1516 | mov %rcx, %r9 |
1517 | shl $3, %r9 |
1518 | cmp %r9, %rdx |
1519 | jbe L(2steps_copy_fwd) |
1520 | add %rcx, %rdx |
1521 | xor %rcx, %rcx |
1522 | L(2steps_copy_fwd): |
1523 | sub $0x80, %rdx |
1524 | L(gobble_mem_fwd_loop): |
1525 | sub $0x80, %rdx |
1526 | prefetcht0 0x200(%rsi) |
1527 | prefetcht0 0x300(%rsi) |
1528 | movdqu (%rsi), %xmm0 |
1529 | movdqu 0x10(%rsi), %xmm1 |
1530 | movdqu 0x20(%rsi), %xmm2 |
1531 | movdqu 0x30(%rsi), %xmm3 |
1532 | movdqu 0x40(%rsi), %xmm4 |
1533 | movdqu 0x50(%rsi), %xmm5 |
1534 | movdqu 0x60(%rsi), %xmm6 |
1535 | movdqu 0x70(%rsi), %xmm7 |
1536 | lfence |
1537 | movntdq %xmm0, (%rdi) |
1538 | movntdq %xmm1, 0x10(%rdi) |
1539 | movntdq %xmm2, 0x20(%rdi) |
1540 | movntdq %xmm3, 0x30(%rdi) |
1541 | movntdq %xmm4, 0x40(%rdi) |
1542 | movntdq %xmm5, 0x50(%rdi) |
1543 | movntdq %xmm6, 0x60(%rdi) |
1544 | movntdq %xmm7, 0x70(%rdi) |
1545 | lea 0x80(%rsi), %rsi |
1546 | lea 0x80(%rdi), %rdi |
1547 | jae L(gobble_mem_fwd_loop) |
1548 | sfence |
1549 | cmp $0x80, %rcx |
1550 | jb L(gobble_mem_fwd_end) |
1551 | add $0x80, %rdx |
1552 | L(ll_cache_copy_fwd): |
1553 | add %rcx, %rdx |
1554 | L(ll_cache_copy_fwd_start): |
1555 | sub $0x80, %rdx |
1556 | L(gobble_ll_loop_fwd): |
1557 | prefetchnta 0x1c0(%rsi) |
1558 | prefetchnta 0x280(%rsi) |
1559 | prefetchnta 0x1c0(%rdi) |
1560 | prefetchnta 0x280(%rdi) |
1561 | sub $0x80, %rdx |
1562 | movdqu (%rsi), %xmm0 |
1563 | movdqu 0x10(%rsi), %xmm1 |
1564 | movdqu 0x20(%rsi), %xmm2 |
1565 | movdqu 0x30(%rsi), %xmm3 |
1566 | movdqu 0x40(%rsi), %xmm4 |
1567 | movdqu 0x50(%rsi), %xmm5 |
1568 | movdqu 0x60(%rsi), %xmm6 |
1569 | movdqu 0x70(%rsi), %xmm7 |
1570 | movdqa %xmm0, (%rdi) |
1571 | movdqa %xmm1, 0x10(%rdi) |
1572 | movdqa %xmm2, 0x20(%rdi) |
1573 | movdqa %xmm3, 0x30(%rdi) |
1574 | movdqa %xmm4, 0x40(%rdi) |
1575 | movdqa %xmm5, 0x50(%rdi) |
1576 | movdqa %xmm6, 0x60(%rdi) |
1577 | movdqa %xmm7, 0x70(%rdi) |
1578 | lea 0x80(%rsi), %rsi |
1579 | lea 0x80(%rdi), %rdi |
1580 | jae L(gobble_ll_loop_fwd) |
1581 | L(gobble_mem_fwd_end): |
1582 | add $0x80, %rdx |
1583 | add %rdx, %rsi |
1584 | add %rdx, %rdi |
1585 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_fwd), %rdx, 4) |
1586 | |
1587 | .p2align 4 |
1588 | L(gobble_mem_bwd): |
1589 | add %rdx, %rsi |
1590 | add %rdx, %rdi |
1591 | |
1592 | movdqu -16(%rsi), %xmm0 |
1593 | lea -16(%rdi), %r8 |
1594 | mov %rdi, %r9 |
1595 | and $-16, %rdi |
1596 | sub %rdi, %r9 |
1597 | sub %r9, %rsi |
1598 | sub %r9, %rdx |
1599 | |
1600 | |
1601 | #ifdef SHARED_CACHE_SIZE_HALF |
1602 | mov $SHARED_CACHE_SIZE_HALF, %RCX_LP |
1603 | #else |
1604 | mov __x86_shared_cache_size_half(%rip), %RCX_LP |
1605 | #endif |
1606 | #ifdef USE_AS_MEMMOVE |
1607 | mov %rdi, %r9 |
1608 | sub %rsi, %r9 |
1609 | cmp %rdx, %r9 |
1610 | jae L(memmove_is_memcpy_bwd) |
1611 | cmp %rcx, %r9 |
1612 | jbe L(ll_cache_copy_bwd_start) |
1613 | L(memmove_is_memcpy_bwd): |
1614 | #endif |
1615 | cmp %rcx, %rdx |
1616 | ja L(bigger) |
1617 | mov %rdx, %rcx |
1618 | L(bigger): |
1619 | sub %rcx, %rdx |
1620 | cmp $0x1000, %rdx |
1621 | jbe L(ll_cache_copy) |
1622 | |
1623 | mov %rcx, %r9 |
1624 | shl $3, %r9 |
1625 | cmp %r9, %rdx |
1626 | jbe L(2steps_copy) |
1627 | add %rcx, %rdx |
1628 | xor %rcx, %rcx |
1629 | L(2steps_copy): |
1630 | sub $0x80, %rdx |
1631 | L(gobble_mem_bwd_loop): |
1632 | sub $0x80, %rdx |
1633 | prefetcht0 -0x200(%rsi) |
1634 | prefetcht0 -0x300(%rsi) |
1635 | movdqu -0x10(%rsi), %xmm1 |
1636 | movdqu -0x20(%rsi), %xmm2 |
1637 | movdqu -0x30(%rsi), %xmm3 |
1638 | movdqu -0x40(%rsi), %xmm4 |
1639 | movdqu -0x50(%rsi), %xmm5 |
1640 | movdqu -0x60(%rsi), %xmm6 |
1641 | movdqu -0x70(%rsi), %xmm7 |
1642 | movdqu -0x80(%rsi), %xmm8 |
1643 | lfence |
1644 | movntdq %xmm1, -0x10(%rdi) |
1645 | movntdq %xmm2, -0x20(%rdi) |
1646 | movntdq %xmm3, -0x30(%rdi) |
1647 | movntdq %xmm4, -0x40(%rdi) |
1648 | movntdq %xmm5, -0x50(%rdi) |
1649 | movntdq %xmm6, -0x60(%rdi) |
1650 | movntdq %xmm7, -0x70(%rdi) |
1651 | movntdq %xmm8, -0x80(%rdi) |
1652 | lea -0x80(%rsi), %rsi |
1653 | lea -0x80(%rdi), %rdi |
1654 | jae L(gobble_mem_bwd_loop) |
1655 | sfence |
1656 | cmp $0x80, %rcx |
1657 | jb L(gobble_mem_bwd_end) |
1658 | add $0x80, %rdx |
1659 | L(ll_cache_copy): |
1660 | add %rcx, %rdx |
1661 | L(ll_cache_copy_bwd_start): |
1662 | sub $0x80, %rdx |
1663 | L(gobble_ll_loop): |
1664 | prefetchnta -0x1c0(%rsi) |
1665 | prefetchnta -0x280(%rsi) |
1666 | prefetchnta -0x1c0(%rdi) |
1667 | prefetchnta -0x280(%rdi) |
1668 | sub $0x80, %rdx |
1669 | movdqu -0x10(%rsi), %xmm1 |
1670 | movdqu -0x20(%rsi), %xmm2 |
1671 | movdqu -0x30(%rsi), %xmm3 |
1672 | movdqu -0x40(%rsi), %xmm4 |
1673 | movdqu -0x50(%rsi), %xmm5 |
1674 | movdqu -0x60(%rsi), %xmm6 |
1675 | movdqu -0x70(%rsi), %xmm7 |
1676 | movdqu -0x80(%rsi), %xmm8 |
1677 | movdqa %xmm1, -0x10(%rdi) |
1678 | movdqa %xmm2, -0x20(%rdi) |
1679 | movdqa %xmm3, -0x30(%rdi) |
1680 | movdqa %xmm4, -0x40(%rdi) |
1681 | movdqa %xmm5, -0x50(%rdi) |
1682 | movdqa %xmm6, -0x60(%rdi) |
1683 | movdqa %xmm7, -0x70(%rdi) |
1684 | movdqa %xmm8, -0x80(%rdi) |
1685 | lea -0x80(%rsi), %rsi |
1686 | lea -0x80(%rdi), %rdi |
1687 | jae L(gobble_ll_loop) |
1688 | L(gobble_mem_bwd_end): |
1689 | movdqu %xmm0, (%r8) |
1690 | add $0x80, %rdx |
1691 | sub %rdx, %rsi |
1692 | sub %rdx, %rdi |
1693 | BRANCH_TO_JMPTBL_ENTRY (L(table_144_bytes_bwd), %rdx, 4) |
1694 | |
1695 | .p2align 4 |
1696 | L(fwd_write_128bytes): |
1697 | lddqu -128(%rsi), %xmm0 |
1698 | movdqu %xmm0, -128(%rdi) |
1699 | L(fwd_write_112bytes): |
1700 | lddqu -112(%rsi), %xmm0 |
1701 | movdqu %xmm0, -112(%rdi) |
1702 | L(fwd_write_96bytes): |
1703 | lddqu -96(%rsi), %xmm0 |
1704 | movdqu %xmm0, -96(%rdi) |
1705 | L(fwd_write_80bytes): |
1706 | lddqu -80(%rsi), %xmm0 |
1707 | movdqu %xmm0, -80(%rdi) |
1708 | L(fwd_write_64bytes): |
1709 | lddqu -64(%rsi), %xmm0 |
1710 | movdqu %xmm0, -64(%rdi) |
1711 | L(fwd_write_48bytes): |
1712 | lddqu -48(%rsi), %xmm0 |
1713 | movdqu %xmm0, -48(%rdi) |
1714 | L(fwd_write_32bytes): |
1715 | lddqu -32(%rsi), %xmm0 |
1716 | movdqu %xmm0, -32(%rdi) |
1717 | L(fwd_write_16bytes): |
1718 | lddqu -16(%rsi), %xmm0 |
1719 | movdqu %xmm0, -16(%rdi) |
1720 | L(fwd_write_0bytes): |
1721 | ret |
1722 | |
1723 | |
1724 | .p2align 4 |
1725 | L(fwd_write_143bytes): |
1726 | lddqu -143(%rsi), %xmm0 |
1727 | movdqu %xmm0, -143(%rdi) |
1728 | L(fwd_write_127bytes): |
1729 | lddqu -127(%rsi), %xmm0 |
1730 | movdqu %xmm0, -127(%rdi) |
1731 | L(fwd_write_111bytes): |
1732 | lddqu -111(%rsi), %xmm0 |
1733 | movdqu %xmm0, -111(%rdi) |
1734 | L(fwd_write_95bytes): |
1735 | lddqu -95(%rsi), %xmm0 |
1736 | movdqu %xmm0, -95(%rdi) |
1737 | L(fwd_write_79bytes): |
1738 | lddqu -79(%rsi), %xmm0 |
1739 | movdqu %xmm0, -79(%rdi) |
1740 | L(fwd_write_63bytes): |
1741 | lddqu -63(%rsi), %xmm0 |
1742 | movdqu %xmm0, -63(%rdi) |
1743 | L(fwd_write_47bytes): |
1744 | lddqu -47(%rsi), %xmm0 |
1745 | movdqu %xmm0, -47(%rdi) |
1746 | L(fwd_write_31bytes): |
1747 | lddqu -31(%rsi), %xmm0 |
1748 | lddqu -16(%rsi), %xmm1 |
1749 | movdqu %xmm0, -31(%rdi) |
1750 | movdqu %xmm1, -16(%rdi) |
1751 | ret |
1752 | |
1753 | .p2align 4 |
1754 | L(fwd_write_15bytes): |
1755 | mov -15(%rsi), %rdx |
1756 | mov -8(%rsi), %rcx |
1757 | mov %rdx, -15(%rdi) |
1758 | mov %rcx, -8(%rdi) |
1759 | ret |
1760 | |
1761 | .p2align 4 |
1762 | L(fwd_write_142bytes): |
1763 | lddqu -142(%rsi), %xmm0 |
1764 | movdqu %xmm0, -142(%rdi) |
1765 | L(fwd_write_126bytes): |
1766 | lddqu -126(%rsi), %xmm0 |
1767 | movdqu %xmm0, -126(%rdi) |
1768 | L(fwd_write_110bytes): |
1769 | lddqu -110(%rsi), %xmm0 |
1770 | movdqu %xmm0, -110(%rdi) |
1771 | L(fwd_write_94bytes): |
1772 | lddqu -94(%rsi), %xmm0 |
1773 | movdqu %xmm0, -94(%rdi) |
1774 | L(fwd_write_78bytes): |
1775 | lddqu -78(%rsi), %xmm0 |
1776 | movdqu %xmm0, -78(%rdi) |
1777 | L(fwd_write_62bytes): |
1778 | lddqu -62(%rsi), %xmm0 |
1779 | movdqu %xmm0, -62(%rdi) |
1780 | L(fwd_write_46bytes): |
1781 | lddqu -46(%rsi), %xmm0 |
1782 | movdqu %xmm0, -46(%rdi) |
1783 | L(fwd_write_30bytes): |
1784 | lddqu -30(%rsi), %xmm0 |
1785 | lddqu -16(%rsi), %xmm1 |
1786 | movdqu %xmm0, -30(%rdi) |
1787 | movdqu %xmm1, -16(%rdi) |
1788 | ret |
1789 | |
1790 | .p2align 4 |
1791 | L(fwd_write_14bytes): |
1792 | mov -14(%rsi), %rdx |
1793 | mov -8(%rsi), %rcx |
1794 | mov %rdx, -14(%rdi) |
1795 | mov %rcx, -8(%rdi) |
1796 | ret |
1797 | |
1798 | .p2align 4 |
1799 | L(fwd_write_141bytes): |
1800 | lddqu -141(%rsi), %xmm0 |
1801 | movdqu %xmm0, -141(%rdi) |
1802 | L(fwd_write_125bytes): |
1803 | lddqu -125(%rsi), %xmm0 |
1804 | movdqu %xmm0, -125(%rdi) |
1805 | L(fwd_write_109bytes): |
1806 | lddqu -109(%rsi), %xmm0 |
1807 | movdqu %xmm0, -109(%rdi) |
1808 | L(fwd_write_93bytes): |
1809 | lddqu -93(%rsi), %xmm0 |
1810 | movdqu %xmm0, -93(%rdi) |
1811 | L(fwd_write_77bytes): |
1812 | lddqu -77(%rsi), %xmm0 |
1813 | movdqu %xmm0, -77(%rdi) |
1814 | L(fwd_write_61bytes): |
1815 | lddqu -61(%rsi), %xmm0 |
1816 | movdqu %xmm0, -61(%rdi) |
1817 | L(fwd_write_45bytes): |
1818 | lddqu -45(%rsi), %xmm0 |
1819 | movdqu %xmm0, -45(%rdi) |
1820 | L(fwd_write_29bytes): |
1821 | lddqu -29(%rsi), %xmm0 |
1822 | lddqu -16(%rsi), %xmm1 |
1823 | movdqu %xmm0, -29(%rdi) |
1824 | movdqu %xmm1, -16(%rdi) |
1825 | ret |
1826 | |
1827 | .p2align 4 |
1828 | L(fwd_write_13bytes): |
1829 | mov -13(%rsi), %rdx |
1830 | mov -8(%rsi), %rcx |
1831 | mov %rdx, -13(%rdi) |
1832 | mov %rcx, -8(%rdi) |
1833 | ret |
1834 | |
1835 | .p2align 4 |
1836 | L(fwd_write_140bytes): |
1837 | lddqu -140(%rsi), %xmm0 |
1838 | movdqu %xmm0, -140(%rdi) |
1839 | L(fwd_write_124bytes): |
1840 | lddqu -124(%rsi), %xmm0 |
1841 | movdqu %xmm0, -124(%rdi) |
1842 | L(fwd_write_108bytes): |
1843 | lddqu -108(%rsi), %xmm0 |
1844 | movdqu %xmm0, -108(%rdi) |
1845 | L(fwd_write_92bytes): |
1846 | lddqu -92(%rsi), %xmm0 |
1847 | movdqu %xmm0, -92(%rdi) |
1848 | L(fwd_write_76bytes): |
1849 | lddqu -76(%rsi), %xmm0 |
1850 | movdqu %xmm0, -76(%rdi) |
1851 | L(fwd_write_60bytes): |
1852 | lddqu -60(%rsi), %xmm0 |
1853 | movdqu %xmm0, -60(%rdi) |
1854 | L(fwd_write_44bytes): |
1855 | lddqu -44(%rsi), %xmm0 |
1856 | movdqu %xmm0, -44(%rdi) |
1857 | L(fwd_write_28bytes): |
1858 | lddqu -28(%rsi), %xmm0 |
1859 | lddqu -16(%rsi), %xmm1 |
1860 | movdqu %xmm0, -28(%rdi) |
1861 | movdqu %xmm1, -16(%rdi) |
1862 | ret |
1863 | |
1864 | .p2align 4 |
1865 | L(fwd_write_12bytes): |
1866 | mov -12(%rsi), %rdx |
1867 | mov -4(%rsi), %ecx |
1868 | mov %rdx, -12(%rdi) |
1869 | mov %ecx, -4(%rdi) |
1870 | ret |
1871 | |
1872 | .p2align 4 |
1873 | L(fwd_write_139bytes): |
1874 | lddqu -139(%rsi), %xmm0 |
1875 | movdqu %xmm0, -139(%rdi) |
1876 | L(fwd_write_123bytes): |
1877 | lddqu -123(%rsi), %xmm0 |
1878 | movdqu %xmm0, -123(%rdi) |
1879 | L(fwd_write_107bytes): |
1880 | lddqu -107(%rsi), %xmm0 |
1881 | movdqu %xmm0, -107(%rdi) |
1882 | L(fwd_write_91bytes): |
1883 | lddqu -91(%rsi), %xmm0 |
1884 | movdqu %xmm0, -91(%rdi) |
1885 | L(fwd_write_75bytes): |
1886 | lddqu -75(%rsi), %xmm0 |
1887 | movdqu %xmm0, -75(%rdi) |
1888 | L(fwd_write_59bytes): |
1889 | lddqu -59(%rsi), %xmm0 |
1890 | movdqu %xmm0, -59(%rdi) |
1891 | L(fwd_write_43bytes): |
1892 | lddqu -43(%rsi), %xmm0 |
1893 | movdqu %xmm0, -43(%rdi) |
1894 | L(fwd_write_27bytes): |
1895 | lddqu -27(%rsi), %xmm0 |
1896 | lddqu -16(%rsi), %xmm1 |
1897 | movdqu %xmm0, -27(%rdi) |
1898 | movdqu %xmm1, -16(%rdi) |
1899 | ret |
1900 | |
1901 | .p2align 4 |
1902 | L(fwd_write_11bytes): |
1903 | mov -11(%rsi), %rdx |
1904 | mov -4(%rsi), %ecx |
1905 | mov %rdx, -11(%rdi) |
1906 | mov %ecx, -4(%rdi) |
1907 | ret |
1908 | |
1909 | .p2align 4 |
1910 | L(fwd_write_138bytes): |
1911 | lddqu -138(%rsi), %xmm0 |
1912 | movdqu %xmm0, -138(%rdi) |
1913 | L(fwd_write_122bytes): |
1914 | lddqu -122(%rsi), %xmm0 |
1915 | movdqu %xmm0, -122(%rdi) |
1916 | L(fwd_write_106bytes): |
1917 | lddqu -106(%rsi), %xmm0 |
1918 | movdqu %xmm0, -106(%rdi) |
1919 | L(fwd_write_90bytes): |
1920 | lddqu -90(%rsi), %xmm0 |
1921 | movdqu %xmm0, -90(%rdi) |
1922 | L(fwd_write_74bytes): |
1923 | lddqu -74(%rsi), %xmm0 |
1924 | movdqu %xmm0, -74(%rdi) |
1925 | L(fwd_write_58bytes): |
1926 | lddqu -58(%rsi), %xmm0 |
1927 | movdqu %xmm0, -58(%rdi) |
1928 | L(fwd_write_42bytes): |
1929 | lddqu -42(%rsi), %xmm0 |
1930 | movdqu %xmm0, -42(%rdi) |
1931 | L(fwd_write_26bytes): |
1932 | lddqu -26(%rsi), %xmm0 |
1933 | lddqu -16(%rsi), %xmm1 |
1934 | movdqu %xmm0, -26(%rdi) |
1935 | movdqu %xmm1, -16(%rdi) |
1936 | ret |
1937 | |
1938 | .p2align 4 |
1939 | L(fwd_write_10bytes): |
1940 | mov -10(%rsi), %rdx |
1941 | mov -4(%rsi), %ecx |
1942 | mov %rdx, -10(%rdi) |
1943 | mov %ecx, -4(%rdi) |
1944 | ret |
1945 | |
1946 | .p2align 4 |
1947 | L(fwd_write_137bytes): |
1948 | lddqu -137(%rsi), %xmm0 |
1949 | movdqu %xmm0, -137(%rdi) |
1950 | L(fwd_write_121bytes): |
1951 | lddqu -121(%rsi), %xmm0 |
1952 | movdqu %xmm0, -121(%rdi) |
1953 | L(fwd_write_105bytes): |
1954 | lddqu -105(%rsi), %xmm0 |
1955 | movdqu %xmm0, -105(%rdi) |
1956 | L(fwd_write_89bytes): |
1957 | lddqu -89(%rsi), %xmm0 |
1958 | movdqu %xmm0, -89(%rdi) |
1959 | L(fwd_write_73bytes): |
1960 | lddqu -73(%rsi), %xmm0 |
1961 | movdqu %xmm0, -73(%rdi) |
1962 | L(fwd_write_57bytes): |
1963 | lddqu -57(%rsi), %xmm0 |
1964 | movdqu %xmm0, -57(%rdi) |
1965 | L(fwd_write_41bytes): |
1966 | lddqu -41(%rsi), %xmm0 |
1967 | movdqu %xmm0, -41(%rdi) |
1968 | L(fwd_write_25bytes): |
1969 | lddqu -25(%rsi), %xmm0 |
1970 | lddqu -16(%rsi), %xmm1 |
1971 | movdqu %xmm0, -25(%rdi) |
1972 | movdqu %xmm1, -16(%rdi) |
1973 | ret |
1974 | |
1975 | .p2align 4 |
1976 | L(fwd_write_9bytes): |
1977 | mov -9(%rsi), %rdx |
1978 | mov -4(%rsi), %ecx |
1979 | mov %rdx, -9(%rdi) |
1980 | mov %ecx, -4(%rdi) |
1981 | ret |
1982 | |
1983 | .p2align 4 |
1984 | L(fwd_write_136bytes): |
1985 | lddqu -136(%rsi), %xmm0 |
1986 | movdqu %xmm0, -136(%rdi) |
1987 | L(fwd_write_120bytes): |
1988 | lddqu -120(%rsi), %xmm0 |
1989 | movdqu %xmm0, -120(%rdi) |
1990 | L(fwd_write_104bytes): |
1991 | lddqu -104(%rsi), %xmm0 |
1992 | movdqu %xmm0, -104(%rdi) |
1993 | L(fwd_write_88bytes): |
1994 | lddqu -88(%rsi), %xmm0 |
1995 | movdqu %xmm0, -88(%rdi) |
1996 | L(fwd_write_72bytes): |
1997 | lddqu -72(%rsi), %xmm0 |
1998 | movdqu %xmm0, -72(%rdi) |
1999 | L(fwd_write_56bytes): |
2000 | lddqu -56(%rsi), %xmm0 |
2001 | movdqu %xmm0, -56(%rdi) |
2002 | L(fwd_write_40bytes): |
2003 | lddqu -40(%rsi), %xmm0 |
2004 | movdqu %xmm0, -40(%rdi) |
2005 | L(fwd_write_24bytes): |
2006 | lddqu -24(%rsi), %xmm0 |
2007 | lddqu -16(%rsi), %xmm1 |
2008 | movdqu %xmm0, -24(%rdi) |
2009 | movdqu %xmm1, -16(%rdi) |
2010 | ret |
2011 | |
2012 | .p2align 4 |
2013 | L(fwd_write_8bytes): |
2014 | mov -8(%rsi), %rdx |
2015 | mov %rdx, -8(%rdi) |
2016 | ret |
2017 | |
2018 | .p2align 4 |
2019 | L(fwd_write_135bytes): |
2020 | lddqu -135(%rsi), %xmm0 |
2021 | movdqu %xmm0, -135(%rdi) |
2022 | L(fwd_write_119bytes): |
2023 | lddqu -119(%rsi), %xmm0 |
2024 | movdqu %xmm0, -119(%rdi) |
2025 | L(fwd_write_103bytes): |
2026 | lddqu -103(%rsi), %xmm0 |
2027 | movdqu %xmm0, -103(%rdi) |
2028 | L(fwd_write_87bytes): |
2029 | lddqu -87(%rsi), %xmm0 |
2030 | movdqu %xmm0, -87(%rdi) |
2031 | L(fwd_write_71bytes): |
2032 | lddqu -71(%rsi), %xmm0 |
2033 | movdqu %xmm0, -71(%rdi) |
2034 | L(fwd_write_55bytes): |
2035 | lddqu -55(%rsi), %xmm0 |
2036 | movdqu %xmm0, -55(%rdi) |
2037 | L(fwd_write_39bytes): |
2038 | lddqu -39(%rsi), %xmm0 |
2039 | movdqu %xmm0, -39(%rdi) |
2040 | L(fwd_write_23bytes): |
2041 | lddqu -23(%rsi), %xmm0 |
2042 | lddqu -16(%rsi), %xmm1 |
2043 | movdqu %xmm0, -23(%rdi) |
2044 | movdqu %xmm1, -16(%rdi) |
2045 | ret |
2046 | |
2047 | .p2align 4 |
2048 | L(fwd_write_7bytes): |
2049 | mov -7(%rsi), %edx |
2050 | mov -4(%rsi), %ecx |
2051 | mov %edx, -7(%rdi) |
2052 | mov %ecx, -4(%rdi) |
2053 | ret |
2054 | |
2055 | .p2align 4 |
2056 | L(fwd_write_134bytes): |
2057 | lddqu -134(%rsi), %xmm0 |
2058 | movdqu %xmm0, -134(%rdi) |
2059 | L(fwd_write_118bytes): |
2060 | lddqu -118(%rsi), %xmm0 |
2061 | movdqu %xmm0, -118(%rdi) |
2062 | L(fwd_write_102bytes): |
2063 | lddqu -102(%rsi), %xmm0 |
2064 | movdqu %xmm0, -102(%rdi) |
2065 | L(fwd_write_86bytes): |
2066 | lddqu -86(%rsi), %xmm0 |
2067 | movdqu %xmm0, -86(%rdi) |
2068 | L(fwd_write_70bytes): |
2069 | lddqu -70(%rsi), %xmm0 |
2070 | movdqu %xmm0, -70(%rdi) |
2071 | L(fwd_write_54bytes): |
2072 | lddqu -54(%rsi), %xmm0 |
2073 | movdqu %xmm0, -54(%rdi) |
2074 | L(fwd_write_38bytes): |
2075 | lddqu -38(%rsi), %xmm0 |
2076 | movdqu %xmm0, -38(%rdi) |
2077 | L(fwd_write_22bytes): |
2078 | lddqu -22(%rsi), %xmm0 |
2079 | lddqu -16(%rsi), %xmm1 |
2080 | movdqu %xmm0, -22(%rdi) |
2081 | movdqu %xmm1, -16(%rdi) |
2082 | ret |
2083 | |
2084 | .p2align 4 |
2085 | L(fwd_write_6bytes): |
2086 | mov -6(%rsi), %edx |
2087 | mov -4(%rsi), %ecx |
2088 | mov %edx, -6(%rdi) |
2089 | mov %ecx, -4(%rdi) |
2090 | ret |
2091 | |
2092 | .p2align 4 |
2093 | L(fwd_write_133bytes): |
2094 | lddqu -133(%rsi), %xmm0 |
2095 | movdqu %xmm0, -133(%rdi) |
2096 | L(fwd_write_117bytes): |
2097 | lddqu -117(%rsi), %xmm0 |
2098 | movdqu %xmm0, -117(%rdi) |
2099 | L(fwd_write_101bytes): |
2100 | lddqu -101(%rsi), %xmm0 |
2101 | movdqu %xmm0, -101(%rdi) |
2102 | L(fwd_write_85bytes): |
2103 | lddqu -85(%rsi), %xmm0 |
2104 | movdqu %xmm0, -85(%rdi) |
2105 | L(fwd_write_69bytes): |
2106 | lddqu -69(%rsi), %xmm0 |
2107 | movdqu %xmm0, -69(%rdi) |
2108 | L(fwd_write_53bytes): |
2109 | lddqu -53(%rsi), %xmm0 |
2110 | movdqu %xmm0, -53(%rdi) |
2111 | L(fwd_write_37bytes): |
2112 | lddqu -37(%rsi), %xmm0 |
2113 | movdqu %xmm0, -37(%rdi) |
2114 | L(fwd_write_21bytes): |
2115 | lddqu -21(%rsi), %xmm0 |
2116 | lddqu -16(%rsi), %xmm1 |
2117 | movdqu %xmm0, -21(%rdi) |
2118 | movdqu %xmm1, -16(%rdi) |
2119 | ret |
2120 | |
2121 | .p2align 4 |
2122 | L(fwd_write_5bytes): |
2123 | mov -5(%rsi), %edx |
2124 | mov -4(%rsi), %ecx |
2125 | mov %edx, -5(%rdi) |
2126 | mov %ecx, -4(%rdi) |
2127 | ret |
2128 | |
2129 | .p2align 4 |
2130 | L(fwd_write_132bytes): |
2131 | lddqu -132(%rsi), %xmm0 |
2132 | movdqu %xmm0, -132(%rdi) |
2133 | L(fwd_write_116bytes): |
2134 | lddqu -116(%rsi), %xmm0 |
2135 | movdqu %xmm0, -116(%rdi) |
2136 | L(fwd_write_100bytes): |
2137 | lddqu -100(%rsi), %xmm0 |
2138 | movdqu %xmm0, -100(%rdi) |
2139 | L(fwd_write_84bytes): |
2140 | lddqu -84(%rsi), %xmm0 |
2141 | movdqu %xmm0, -84(%rdi) |
2142 | L(fwd_write_68bytes): |
2143 | lddqu -68(%rsi), %xmm0 |
2144 | movdqu %xmm0, -68(%rdi) |
2145 | L(fwd_write_52bytes): |
2146 | lddqu -52(%rsi), %xmm0 |
2147 | movdqu %xmm0, -52(%rdi) |
2148 | L(fwd_write_36bytes): |
2149 | lddqu -36(%rsi), %xmm0 |
2150 | movdqu %xmm0, -36(%rdi) |
2151 | L(fwd_write_20bytes): |
2152 | lddqu -20(%rsi), %xmm0 |
2153 | lddqu -16(%rsi), %xmm1 |
2154 | movdqu %xmm0, -20(%rdi) |
2155 | movdqu %xmm1, -16(%rdi) |
2156 | ret |
2157 | |
2158 | .p2align 4 |
2159 | L(fwd_write_4bytes): |
2160 | mov -4(%rsi), %edx |
2161 | mov %edx, -4(%rdi) |
2162 | ret |
2163 | |
2164 | .p2align 4 |
2165 | L(fwd_write_131bytes): |
2166 | lddqu -131(%rsi), %xmm0 |
2167 | movdqu %xmm0, -131(%rdi) |
2168 | L(fwd_write_115bytes): |
2169 | lddqu -115(%rsi), %xmm0 |
2170 | movdqu %xmm0, -115(%rdi) |
2171 | L(fwd_write_99bytes): |
2172 | lddqu -99(%rsi), %xmm0 |
2173 | movdqu %xmm0, -99(%rdi) |
2174 | L(fwd_write_83bytes): |
2175 | lddqu -83(%rsi), %xmm0 |
2176 | movdqu %xmm0, -83(%rdi) |
2177 | L(fwd_write_67bytes): |
2178 | lddqu -67(%rsi), %xmm0 |
2179 | movdqu %xmm0, -67(%rdi) |
2180 | L(fwd_write_51bytes): |
2181 | lddqu -51(%rsi), %xmm0 |
2182 | movdqu %xmm0, -51(%rdi) |
2183 | L(fwd_write_35bytes): |
2184 | lddqu -35(%rsi), %xmm0 |
2185 | movdqu %xmm0, -35(%rdi) |
2186 | L(fwd_write_19bytes): |
2187 | lddqu -19(%rsi), %xmm0 |
2188 | lddqu -16(%rsi), %xmm1 |
2189 | movdqu %xmm0, -19(%rdi) |
2190 | movdqu %xmm1, -16(%rdi) |
2191 | ret |
2192 | |
2193 | .p2align 4 |
2194 | L(fwd_write_3bytes): |
2195 | mov -3(%rsi), %dx |
2196 | mov -2(%rsi), %cx |
2197 | mov %dx, -3(%rdi) |
2198 | mov %cx, -2(%rdi) |
2199 | ret |
2200 | |
2201 | .p2align 4 |
2202 | L(fwd_write_130bytes): |
2203 | lddqu -130(%rsi), %xmm0 |
2204 | movdqu %xmm0, -130(%rdi) |
2205 | L(fwd_write_114bytes): |
2206 | lddqu -114(%rsi), %xmm0 |
2207 | movdqu %xmm0, -114(%rdi) |
2208 | L(fwd_write_98bytes): |
2209 | lddqu -98(%rsi), %xmm0 |
2210 | movdqu %xmm0, -98(%rdi) |
2211 | L(fwd_write_82bytes): |
2212 | lddqu -82(%rsi), %xmm0 |
2213 | movdqu %xmm0, -82(%rdi) |
2214 | L(fwd_write_66bytes): |
2215 | lddqu -66(%rsi), %xmm0 |
2216 | movdqu %xmm0, -66(%rdi) |
2217 | L(fwd_write_50bytes): |
2218 | lddqu -50(%rsi), %xmm0 |
2219 | movdqu %xmm0, -50(%rdi) |
2220 | L(fwd_write_34bytes): |
2221 | lddqu -34(%rsi), %xmm0 |
2222 | movdqu %xmm0, -34(%rdi) |
2223 | L(fwd_write_18bytes): |
2224 | lddqu -18(%rsi), %xmm0 |
2225 | lddqu -16(%rsi), %xmm1 |
2226 | movdqu %xmm0, -18(%rdi) |
2227 | movdqu %xmm1, -16(%rdi) |
2228 | ret |
2229 | |
2230 | .p2align 4 |
2231 | L(fwd_write_2bytes): |
2232 | movzwl -2(%rsi), %edx |
2233 | mov %dx, -2(%rdi) |
2234 | ret |
2235 | |
2236 | .p2align 4 |
2237 | L(fwd_write_129bytes): |
2238 | lddqu -129(%rsi), %xmm0 |
2239 | movdqu %xmm0, -129(%rdi) |
2240 | L(fwd_write_113bytes): |
2241 | lddqu -113(%rsi), %xmm0 |
2242 | movdqu %xmm0, -113(%rdi) |
2243 | L(fwd_write_97bytes): |
2244 | lddqu -97(%rsi), %xmm0 |
2245 | movdqu %xmm0, -97(%rdi) |
2246 | L(fwd_write_81bytes): |
2247 | lddqu -81(%rsi), %xmm0 |
2248 | movdqu %xmm0, -81(%rdi) |
2249 | L(fwd_write_65bytes): |
2250 | lddqu -65(%rsi), %xmm0 |
2251 | movdqu %xmm0, -65(%rdi) |
2252 | L(fwd_write_49bytes): |
2253 | lddqu -49(%rsi), %xmm0 |
2254 | movdqu %xmm0, -49(%rdi) |
2255 | L(fwd_write_33bytes): |
2256 | lddqu -33(%rsi), %xmm0 |
2257 | movdqu %xmm0, -33(%rdi) |
2258 | L(fwd_write_17bytes): |
2259 | lddqu -17(%rsi), %xmm0 |
2260 | lddqu -16(%rsi), %xmm1 |
2261 | movdqu %xmm0, -17(%rdi) |
2262 | movdqu %xmm1, -16(%rdi) |
2263 | ret |
2264 | |
2265 | .p2align 4 |
2266 | L(fwd_write_1bytes): |
2267 | movzbl -1(%rsi), %edx |
2268 | mov %dl, -1(%rdi) |
2269 | ret |
2270 | |
2271 | .p2align 4 |
2272 | L(bwd_write_128bytes): |
2273 | lddqu 112(%rsi), %xmm0 |
2274 | movdqu %xmm0, 112(%rdi) |
2275 | L(bwd_write_112bytes): |
2276 | lddqu 96(%rsi), %xmm0 |
2277 | movdqu %xmm0, 96(%rdi) |
2278 | L(bwd_write_96bytes): |
2279 | lddqu 80(%rsi), %xmm0 |
2280 | movdqu %xmm0, 80(%rdi) |
2281 | L(bwd_write_80bytes): |
2282 | lddqu 64(%rsi), %xmm0 |
2283 | movdqu %xmm0, 64(%rdi) |
2284 | L(bwd_write_64bytes): |
2285 | lddqu 48(%rsi), %xmm0 |
2286 | movdqu %xmm0, 48(%rdi) |
2287 | L(bwd_write_48bytes): |
2288 | lddqu 32(%rsi), %xmm0 |
2289 | movdqu %xmm0, 32(%rdi) |
2290 | L(bwd_write_32bytes): |
2291 | lddqu 16(%rsi), %xmm0 |
2292 | movdqu %xmm0, 16(%rdi) |
2293 | L(bwd_write_16bytes): |
2294 | lddqu (%rsi), %xmm0 |
2295 | movdqu %xmm0, (%rdi) |
2296 | L(bwd_write_0bytes): |
2297 | ret |
2298 | |
2299 | .p2align 4 |
2300 | L(bwd_write_143bytes): |
2301 | lddqu 127(%rsi), %xmm0 |
2302 | movdqu %xmm0, 127(%rdi) |
2303 | L(bwd_write_127bytes): |
2304 | lddqu 111(%rsi), %xmm0 |
2305 | movdqu %xmm0, 111(%rdi) |
2306 | L(bwd_write_111bytes): |
2307 | lddqu 95(%rsi), %xmm0 |
2308 | movdqu %xmm0, 95(%rdi) |
2309 | L(bwd_write_95bytes): |
2310 | lddqu 79(%rsi), %xmm0 |
2311 | movdqu %xmm0, 79(%rdi) |
2312 | L(bwd_write_79bytes): |
2313 | lddqu 63(%rsi), %xmm0 |
2314 | movdqu %xmm0, 63(%rdi) |
2315 | L(bwd_write_63bytes): |
2316 | lddqu 47(%rsi), %xmm0 |
2317 | movdqu %xmm0, 47(%rdi) |
2318 | L(bwd_write_47bytes): |
2319 | lddqu 31(%rsi), %xmm0 |
2320 | movdqu %xmm0, 31(%rdi) |
2321 | L(bwd_write_31bytes): |
2322 | lddqu 15(%rsi), %xmm0 |
2323 | lddqu (%rsi), %xmm1 |
2324 | movdqu %xmm0, 15(%rdi) |
2325 | movdqu %xmm1, (%rdi) |
2326 | ret |
2327 | |
2328 | |
2329 | .p2align 4 |
2330 | L(bwd_write_15bytes): |
2331 | mov 7(%rsi), %rdx |
2332 | mov (%rsi), %rcx |
2333 | mov %rdx, 7(%rdi) |
2334 | mov %rcx, (%rdi) |
2335 | ret |
2336 | |
2337 | .p2align 4 |
2338 | L(bwd_write_142bytes): |
2339 | lddqu 126(%rsi), %xmm0 |
2340 | movdqu %xmm0, 126(%rdi) |
2341 | L(bwd_write_126bytes): |
2342 | lddqu 110(%rsi), %xmm0 |
2343 | movdqu %xmm0, 110(%rdi) |
2344 | L(bwd_write_110bytes): |
2345 | lddqu 94(%rsi), %xmm0 |
2346 | movdqu %xmm0, 94(%rdi) |
2347 | L(bwd_write_94bytes): |
2348 | lddqu 78(%rsi), %xmm0 |
2349 | movdqu %xmm0, 78(%rdi) |
2350 | L(bwd_write_78bytes): |
2351 | lddqu 62(%rsi), %xmm0 |
2352 | movdqu %xmm0, 62(%rdi) |
2353 | L(bwd_write_62bytes): |
2354 | lddqu 46(%rsi), %xmm0 |
2355 | movdqu %xmm0, 46(%rdi) |
2356 | L(bwd_write_46bytes): |
2357 | lddqu 30(%rsi), %xmm0 |
2358 | movdqu %xmm0, 30(%rdi) |
2359 | L(bwd_write_30bytes): |
2360 | lddqu 14(%rsi), %xmm0 |
2361 | lddqu (%rsi), %xmm1 |
2362 | movdqu %xmm0, 14(%rdi) |
2363 | movdqu %xmm1, (%rdi) |
2364 | ret |
2365 | |
2366 | .p2align 4 |
2367 | L(bwd_write_14bytes): |
2368 | mov 6(%rsi), %rdx |
2369 | mov (%rsi), %rcx |
2370 | mov %rdx, 6(%rdi) |
2371 | mov %rcx, (%rdi) |
2372 | ret |
2373 | |
2374 | .p2align 4 |
2375 | L(bwd_write_141bytes): |
2376 | lddqu 125(%rsi), %xmm0 |
2377 | movdqu %xmm0, 125(%rdi) |
2378 | L(bwd_write_125bytes): |
2379 | lddqu 109(%rsi), %xmm0 |
2380 | movdqu %xmm0, 109(%rdi) |
2381 | L(bwd_write_109bytes): |
2382 | lddqu 93(%rsi), %xmm0 |
2383 | movdqu %xmm0, 93(%rdi) |
2384 | L(bwd_write_93bytes): |
2385 | lddqu 77(%rsi), %xmm0 |
2386 | movdqu %xmm0, 77(%rdi) |
2387 | L(bwd_write_77bytes): |
2388 | lddqu 61(%rsi), %xmm0 |
2389 | movdqu %xmm0, 61(%rdi) |
2390 | L(bwd_write_61bytes): |
2391 | lddqu 45(%rsi), %xmm0 |
2392 | movdqu %xmm0, 45(%rdi) |
2393 | L(bwd_write_45bytes): |
2394 | lddqu 29(%rsi), %xmm0 |
2395 | movdqu %xmm0, 29(%rdi) |
2396 | L(bwd_write_29bytes): |
2397 | lddqu 13(%rsi), %xmm0 |
2398 | lddqu (%rsi), %xmm1 |
2399 | movdqu %xmm0, 13(%rdi) |
2400 | movdqu %xmm1, (%rdi) |
2401 | ret |
2402 | |
2403 | .p2align 4 |
2404 | L(bwd_write_13bytes): |
2405 | mov 5(%rsi), %rdx |
2406 | mov (%rsi), %rcx |
2407 | mov %rdx, 5(%rdi) |
2408 | mov %rcx, (%rdi) |
2409 | ret |
2410 | |
2411 | .p2align 4 |
2412 | L(bwd_write_140bytes): |
2413 | lddqu 124(%rsi), %xmm0 |
2414 | movdqu %xmm0, 124(%rdi) |
2415 | L(bwd_write_124bytes): |
2416 | lddqu 108(%rsi), %xmm0 |
2417 | movdqu %xmm0, 108(%rdi) |
2418 | L(bwd_write_108bytes): |
2419 | lddqu 92(%rsi), %xmm0 |
2420 | movdqu %xmm0, 92(%rdi) |
2421 | L(bwd_write_92bytes): |
2422 | lddqu 76(%rsi), %xmm0 |
2423 | movdqu %xmm0, 76(%rdi) |
2424 | L(bwd_write_76bytes): |
2425 | lddqu 60(%rsi), %xmm0 |
2426 | movdqu %xmm0, 60(%rdi) |
2427 | L(bwd_write_60bytes): |
2428 | lddqu 44(%rsi), %xmm0 |
2429 | movdqu %xmm0, 44(%rdi) |
2430 | L(bwd_write_44bytes): |
2431 | lddqu 28(%rsi), %xmm0 |
2432 | movdqu %xmm0, 28(%rdi) |
2433 | L(bwd_write_28bytes): |
2434 | lddqu 12(%rsi), %xmm0 |
2435 | lddqu (%rsi), %xmm1 |
2436 | movdqu %xmm0, 12(%rdi) |
2437 | movdqu %xmm1, (%rdi) |
2438 | ret |
2439 | |
2440 | .p2align 4 |
2441 | L(bwd_write_12bytes): |
2442 | mov 4(%rsi), %rdx |
2443 | mov (%rsi), %rcx |
2444 | mov %rdx, 4(%rdi) |
2445 | mov %rcx, (%rdi) |
2446 | ret |
2447 | |
2448 | .p2align 4 |
2449 | L(bwd_write_139bytes): |
2450 | lddqu 123(%rsi), %xmm0 |
2451 | movdqu %xmm0, 123(%rdi) |
2452 | L(bwd_write_123bytes): |
2453 | lddqu 107(%rsi), %xmm0 |
2454 | movdqu %xmm0, 107(%rdi) |
2455 | L(bwd_write_107bytes): |
2456 | lddqu 91(%rsi), %xmm0 |
2457 | movdqu %xmm0, 91(%rdi) |
2458 | L(bwd_write_91bytes): |
2459 | lddqu 75(%rsi), %xmm0 |
2460 | movdqu %xmm0, 75(%rdi) |
2461 | L(bwd_write_75bytes): |
2462 | lddqu 59(%rsi), %xmm0 |
2463 | movdqu %xmm0, 59(%rdi) |
2464 | L(bwd_write_59bytes): |
2465 | lddqu 43(%rsi), %xmm0 |
2466 | movdqu %xmm0, 43(%rdi) |
2467 | L(bwd_write_43bytes): |
2468 | lddqu 27(%rsi), %xmm0 |
2469 | movdqu %xmm0, 27(%rdi) |
2470 | L(bwd_write_27bytes): |
2471 | lddqu 11(%rsi), %xmm0 |
2472 | lddqu (%rsi), %xmm1 |
2473 | movdqu %xmm0, 11(%rdi) |
2474 | movdqu %xmm1, (%rdi) |
2475 | ret |
2476 | |
2477 | .p2align 4 |
2478 | L(bwd_write_11bytes): |
2479 | mov 3(%rsi), %rdx |
2480 | mov (%rsi), %rcx |
2481 | mov %rdx, 3(%rdi) |
2482 | mov %rcx, (%rdi) |
2483 | ret |
2484 | |
2485 | .p2align 4 |
2486 | L(bwd_write_138bytes): |
2487 | lddqu 122(%rsi), %xmm0 |
2488 | movdqu %xmm0, 122(%rdi) |
2489 | L(bwd_write_122bytes): |
2490 | lddqu 106(%rsi), %xmm0 |
2491 | movdqu %xmm0, 106(%rdi) |
2492 | L(bwd_write_106bytes): |
2493 | lddqu 90(%rsi), %xmm0 |
2494 | movdqu %xmm0, 90(%rdi) |
2495 | L(bwd_write_90bytes): |
2496 | lddqu 74(%rsi), %xmm0 |
2497 | movdqu %xmm0, 74(%rdi) |
2498 | L(bwd_write_74bytes): |
2499 | lddqu 58(%rsi), %xmm0 |
2500 | movdqu %xmm0, 58(%rdi) |
2501 | L(bwd_write_58bytes): |
2502 | lddqu 42(%rsi), %xmm0 |
2503 | movdqu %xmm0, 42(%rdi) |
2504 | L(bwd_write_42bytes): |
2505 | lddqu 26(%rsi), %xmm0 |
2506 | movdqu %xmm0, 26(%rdi) |
2507 | L(bwd_write_26bytes): |
2508 | lddqu 10(%rsi), %xmm0 |
2509 | lddqu (%rsi), %xmm1 |
2510 | movdqu %xmm0, 10(%rdi) |
2511 | movdqu %xmm1, (%rdi) |
2512 | ret |
2513 | |
2514 | .p2align 4 |
2515 | L(bwd_write_10bytes): |
2516 | mov 2(%rsi), %rdx |
2517 | mov (%rsi), %rcx |
2518 | mov %rdx, 2(%rdi) |
2519 | mov %rcx, (%rdi) |
2520 | ret |
2521 | |
2522 | .p2align 4 |
2523 | L(bwd_write_137bytes): |
2524 | lddqu 121(%rsi), %xmm0 |
2525 | movdqu %xmm0, 121(%rdi) |
2526 | L(bwd_write_121bytes): |
2527 | lddqu 105(%rsi), %xmm0 |
2528 | movdqu %xmm0, 105(%rdi) |
2529 | L(bwd_write_105bytes): |
2530 | lddqu 89(%rsi), %xmm0 |
2531 | movdqu %xmm0, 89(%rdi) |
2532 | L(bwd_write_89bytes): |
2533 | lddqu 73(%rsi), %xmm0 |
2534 | movdqu %xmm0, 73(%rdi) |
2535 | L(bwd_write_73bytes): |
2536 | lddqu 57(%rsi), %xmm0 |
2537 | movdqu %xmm0, 57(%rdi) |
2538 | L(bwd_write_57bytes): |
2539 | lddqu 41(%rsi), %xmm0 |
2540 | movdqu %xmm0, 41(%rdi) |
2541 | L(bwd_write_41bytes): |
2542 | lddqu 25(%rsi), %xmm0 |
2543 | movdqu %xmm0, 25(%rdi) |
2544 | L(bwd_write_25bytes): |
2545 | lddqu 9(%rsi), %xmm0 |
2546 | lddqu (%rsi), %xmm1 |
2547 | movdqu %xmm0, 9(%rdi) |
2548 | movdqu %xmm1, (%rdi) |
2549 | ret |
2550 | |
2551 | .p2align 4 |
2552 | L(bwd_write_9bytes): |
2553 | mov 1(%rsi), %rdx |
2554 | mov (%rsi), %rcx |
2555 | mov %rdx, 1(%rdi) |
2556 | mov %rcx, (%rdi) |
2557 | ret |
2558 | |
2559 | .p2align 4 |
2560 | L(bwd_write_136bytes): |
2561 | lddqu 120(%rsi), %xmm0 |
2562 | movdqu %xmm0, 120(%rdi) |
2563 | L(bwd_write_120bytes): |
2564 | lddqu 104(%rsi), %xmm0 |
2565 | movdqu %xmm0, 104(%rdi) |
2566 | L(bwd_write_104bytes): |
2567 | lddqu 88(%rsi), %xmm0 |
2568 | movdqu %xmm0, 88(%rdi) |
2569 | L(bwd_write_88bytes): |
2570 | lddqu 72(%rsi), %xmm0 |
2571 | movdqu %xmm0, 72(%rdi) |
2572 | L(bwd_write_72bytes): |
2573 | lddqu 56(%rsi), %xmm0 |
2574 | movdqu %xmm0, 56(%rdi) |
2575 | L(bwd_write_56bytes): |
2576 | lddqu 40(%rsi), %xmm0 |
2577 | movdqu %xmm0, 40(%rdi) |
2578 | L(bwd_write_40bytes): |
2579 | lddqu 24(%rsi), %xmm0 |
2580 | movdqu %xmm0, 24(%rdi) |
2581 | L(bwd_write_24bytes): |
2582 | lddqu 8(%rsi), %xmm0 |
2583 | lddqu (%rsi), %xmm1 |
2584 | movdqu %xmm0, 8(%rdi) |
2585 | movdqu %xmm1, (%rdi) |
2586 | ret |
2587 | |
2588 | .p2align 4 |
2589 | L(bwd_write_8bytes): |
2590 | mov (%rsi), %rdx |
2591 | mov %rdx, (%rdi) |
2592 | ret |
2593 | |
2594 | .p2align 4 |
2595 | L(bwd_write_135bytes): |
2596 | lddqu 119(%rsi), %xmm0 |
2597 | movdqu %xmm0, 119(%rdi) |
2598 | L(bwd_write_119bytes): |
2599 | lddqu 103(%rsi), %xmm0 |
2600 | movdqu %xmm0, 103(%rdi) |
2601 | L(bwd_write_103bytes): |
2602 | lddqu 87(%rsi), %xmm0 |
2603 | movdqu %xmm0, 87(%rdi) |
2604 | L(bwd_write_87bytes): |
2605 | lddqu 71(%rsi), %xmm0 |
2606 | movdqu %xmm0, 71(%rdi) |
2607 | L(bwd_write_71bytes): |
2608 | lddqu 55(%rsi), %xmm0 |
2609 | movdqu %xmm0, 55(%rdi) |
2610 | L(bwd_write_55bytes): |
2611 | lddqu 39(%rsi), %xmm0 |
2612 | movdqu %xmm0, 39(%rdi) |
2613 | L(bwd_write_39bytes): |
2614 | lddqu 23(%rsi), %xmm0 |
2615 | movdqu %xmm0, 23(%rdi) |
2616 | L(bwd_write_23bytes): |
2617 | lddqu 7(%rsi), %xmm0 |
2618 | lddqu (%rsi), %xmm1 |
2619 | movdqu %xmm0, 7(%rdi) |
2620 | movdqu %xmm1, (%rdi) |
2621 | ret |
2622 | |
2623 | .p2align 4 |
2624 | L(bwd_write_7bytes): |
2625 | mov 3(%rsi), %edx |
2626 | mov (%rsi), %ecx |
2627 | mov %edx, 3(%rdi) |
2628 | mov %ecx, (%rdi) |
2629 | ret |
2630 | |
2631 | .p2align 4 |
2632 | L(bwd_write_134bytes): |
2633 | lddqu 118(%rsi), %xmm0 |
2634 | movdqu %xmm0, 118(%rdi) |
2635 | L(bwd_write_118bytes): |
2636 | lddqu 102(%rsi), %xmm0 |
2637 | movdqu %xmm0, 102(%rdi) |
2638 | L(bwd_write_102bytes): |
2639 | lddqu 86(%rsi), %xmm0 |
2640 | movdqu %xmm0, 86(%rdi) |
2641 | L(bwd_write_86bytes): |
2642 | lddqu 70(%rsi), %xmm0 |
2643 | movdqu %xmm0, 70(%rdi) |
2644 | L(bwd_write_70bytes): |
2645 | lddqu 54(%rsi), %xmm0 |
2646 | movdqu %xmm0, 54(%rdi) |
2647 | L(bwd_write_54bytes): |
2648 | lddqu 38(%rsi), %xmm0 |
2649 | movdqu %xmm0, 38(%rdi) |
2650 | L(bwd_write_38bytes): |
2651 | lddqu 22(%rsi), %xmm0 |
2652 | movdqu %xmm0, 22(%rdi) |
2653 | L(bwd_write_22bytes): |
2654 | lddqu 6(%rsi), %xmm0 |
2655 | lddqu (%rsi), %xmm1 |
2656 | movdqu %xmm0, 6(%rdi) |
2657 | movdqu %xmm1, (%rdi) |
2658 | ret |
2659 | |
2660 | .p2align 4 |
2661 | L(bwd_write_6bytes): |
2662 | mov 2(%rsi), %edx |
2663 | mov (%rsi), %ecx |
2664 | mov %edx, 2(%rdi) |
2665 | mov %ecx, (%rdi) |
2666 | ret |
2667 | |
2668 | .p2align 4 |
2669 | L(bwd_write_133bytes): |
2670 | lddqu 117(%rsi), %xmm0 |
2671 | movdqu %xmm0, 117(%rdi) |
2672 | L(bwd_write_117bytes): |
2673 | lddqu 101(%rsi), %xmm0 |
2674 | movdqu %xmm0, 101(%rdi) |
2675 | L(bwd_write_101bytes): |
2676 | lddqu 85(%rsi), %xmm0 |
2677 | movdqu %xmm0, 85(%rdi) |
2678 | L(bwd_write_85bytes): |
2679 | lddqu 69(%rsi), %xmm0 |
2680 | movdqu %xmm0, 69(%rdi) |
2681 | L(bwd_write_69bytes): |
2682 | lddqu 53(%rsi), %xmm0 |
2683 | movdqu %xmm0, 53(%rdi) |
2684 | L(bwd_write_53bytes): |
2685 | lddqu 37(%rsi), %xmm0 |
2686 | movdqu %xmm0, 37(%rdi) |
2687 | L(bwd_write_37bytes): |
2688 | lddqu 21(%rsi), %xmm0 |
2689 | movdqu %xmm0, 21(%rdi) |
2690 | L(bwd_write_21bytes): |
2691 | lddqu 5(%rsi), %xmm0 |
2692 | lddqu (%rsi), %xmm1 |
2693 | movdqu %xmm0, 5(%rdi) |
2694 | movdqu %xmm1, (%rdi) |
2695 | ret |
2696 | |
2697 | .p2align 4 |
2698 | L(bwd_write_5bytes): |
2699 | mov 1(%rsi), %edx |
2700 | mov (%rsi), %ecx |
2701 | mov %edx, 1(%rdi) |
2702 | mov %ecx, (%rdi) |
2703 | ret |
2704 | |
2705 | .p2align 4 |
2706 | L(bwd_write_132bytes): |
2707 | lddqu 116(%rsi), %xmm0 |
2708 | movdqu %xmm0, 116(%rdi) |
2709 | L(bwd_write_116bytes): |
2710 | lddqu 100(%rsi), %xmm0 |
2711 | movdqu %xmm0, 100(%rdi) |
2712 | L(bwd_write_100bytes): |
2713 | lddqu 84(%rsi), %xmm0 |
2714 | movdqu %xmm0, 84(%rdi) |
2715 | L(bwd_write_84bytes): |
2716 | lddqu 68(%rsi), %xmm0 |
2717 | movdqu %xmm0, 68(%rdi) |
2718 | L(bwd_write_68bytes): |
2719 | lddqu 52(%rsi), %xmm0 |
2720 | movdqu %xmm0, 52(%rdi) |
2721 | L(bwd_write_52bytes): |
2722 | lddqu 36(%rsi), %xmm0 |
2723 | movdqu %xmm0, 36(%rdi) |
2724 | L(bwd_write_36bytes): |
2725 | lddqu 20(%rsi), %xmm0 |
2726 | movdqu %xmm0, 20(%rdi) |
2727 | L(bwd_write_20bytes): |
2728 | lddqu 4(%rsi), %xmm0 |
2729 | lddqu (%rsi), %xmm1 |
2730 | movdqu %xmm0, 4(%rdi) |
2731 | movdqu %xmm1, (%rdi) |
2732 | ret |
2733 | |
2734 | .p2align 4 |
2735 | L(bwd_write_4bytes): |
2736 | mov (%rsi), %edx |
2737 | mov %edx, (%rdi) |
2738 | ret |
2739 | |
2740 | .p2align 4 |
2741 | L(bwd_write_131bytes): |
2742 | lddqu 115(%rsi), %xmm0 |
2743 | movdqu %xmm0, 115(%rdi) |
2744 | L(bwd_write_115bytes): |
2745 | lddqu 99(%rsi), %xmm0 |
2746 | movdqu %xmm0, 99(%rdi) |
2747 | L(bwd_write_99bytes): |
2748 | lddqu 83(%rsi), %xmm0 |
2749 | movdqu %xmm0, 83(%rdi) |
2750 | L(bwd_write_83bytes): |
2751 | lddqu 67(%rsi), %xmm0 |
2752 | movdqu %xmm0, 67(%rdi) |
2753 | L(bwd_write_67bytes): |
2754 | lddqu 51(%rsi), %xmm0 |
2755 | movdqu %xmm0, 51(%rdi) |
2756 | L(bwd_write_51bytes): |
2757 | lddqu 35(%rsi), %xmm0 |
2758 | movdqu %xmm0, 35(%rdi) |
2759 | L(bwd_write_35bytes): |
2760 | lddqu 19(%rsi), %xmm0 |
2761 | movdqu %xmm0, 19(%rdi) |
2762 | L(bwd_write_19bytes): |
2763 | lddqu 3(%rsi), %xmm0 |
2764 | lddqu (%rsi), %xmm1 |
2765 | movdqu %xmm0, 3(%rdi) |
2766 | movdqu %xmm1, (%rdi) |
2767 | ret |
2768 | |
2769 | .p2align 4 |
2770 | L(bwd_write_3bytes): |
2771 | mov 1(%rsi), %dx |
2772 | mov (%rsi), %cx |
2773 | mov %dx, 1(%rdi) |
2774 | mov %cx, (%rdi) |
2775 | ret |
2776 | |
2777 | .p2align 4 |
2778 | L(bwd_write_130bytes): |
2779 | lddqu 114(%rsi), %xmm0 |
2780 | movdqu %xmm0, 114(%rdi) |
2781 | L(bwd_write_114bytes): |
2782 | lddqu 98(%rsi), %xmm0 |
2783 | movdqu %xmm0, 98(%rdi) |
2784 | L(bwd_write_98bytes): |
2785 | lddqu 82(%rsi), %xmm0 |
2786 | movdqu %xmm0, 82(%rdi) |
2787 | L(bwd_write_82bytes): |
2788 | lddqu 66(%rsi), %xmm0 |
2789 | movdqu %xmm0, 66(%rdi) |
2790 | L(bwd_write_66bytes): |
2791 | lddqu 50(%rsi), %xmm0 |
2792 | movdqu %xmm0, 50(%rdi) |
2793 | L(bwd_write_50bytes): |
2794 | lddqu 34(%rsi), %xmm0 |
2795 | movdqu %xmm0, 34(%rdi) |
2796 | L(bwd_write_34bytes): |
2797 | lddqu 18(%rsi), %xmm0 |
2798 | movdqu %xmm0, 18(%rdi) |
2799 | L(bwd_write_18bytes): |
2800 | lddqu 2(%rsi), %xmm0 |
2801 | lddqu (%rsi), %xmm1 |
2802 | movdqu %xmm0, 2(%rdi) |
2803 | movdqu %xmm1, (%rdi) |
2804 | ret |
2805 | |
2806 | .p2align 4 |
2807 | L(bwd_write_2bytes): |
2808 | movzwl (%rsi), %edx |
2809 | mov %dx, (%rdi) |
2810 | ret |
2811 | |
2812 | .p2align 4 |
2813 | L(bwd_write_129bytes): |
2814 | lddqu 113(%rsi), %xmm0 |
2815 | movdqu %xmm0, 113(%rdi) |
2816 | L(bwd_write_113bytes): |
2817 | lddqu 97(%rsi), %xmm0 |
2818 | movdqu %xmm0, 97(%rdi) |
2819 | L(bwd_write_97bytes): |
2820 | lddqu 81(%rsi), %xmm0 |
2821 | movdqu %xmm0, 81(%rdi) |
2822 | L(bwd_write_81bytes): |
2823 | lddqu 65(%rsi), %xmm0 |
2824 | movdqu %xmm0, 65(%rdi) |
2825 | L(bwd_write_65bytes): |
2826 | lddqu 49(%rsi), %xmm0 |
2827 | movdqu %xmm0, 49(%rdi) |
2828 | L(bwd_write_49bytes): |
2829 | lddqu 33(%rsi), %xmm0 |
2830 | movdqu %xmm0, 33(%rdi) |
2831 | L(bwd_write_33bytes): |
2832 | lddqu 17(%rsi), %xmm0 |
2833 | movdqu %xmm0, 17(%rdi) |
2834 | L(bwd_write_17bytes): |
2835 | lddqu 1(%rsi), %xmm0 |
2836 | lddqu (%rsi), %xmm1 |
2837 | movdqu %xmm0, 1(%rdi) |
2838 | movdqu %xmm1, (%rdi) |
2839 | ret |
2840 | |
2841 | .p2align 4 |
2842 | L(bwd_write_1bytes): |
2843 | movzbl (%rsi), %edx |
2844 | mov %dl, (%rdi) |
2845 | ret |
2846 | |
2847 | END (MEMCPY) |
2848 | |
2849 | .section .rodata.ssse3,"a" ,@progbits |
2850 | .p2align 3 |
2851 | L(table_144_bytes_bwd): |
2852 | .int JMPTBL (L(bwd_write_0bytes), L(table_144_bytes_bwd)) |
2853 | .int JMPTBL (L(bwd_write_1bytes), L(table_144_bytes_bwd)) |
2854 | .int JMPTBL (L(bwd_write_2bytes), L(table_144_bytes_bwd)) |
2855 | .int JMPTBL (L(bwd_write_3bytes), L(table_144_bytes_bwd)) |
2856 | .int JMPTBL (L(bwd_write_4bytes), L(table_144_bytes_bwd)) |
2857 | .int JMPTBL (L(bwd_write_5bytes), L(table_144_bytes_bwd)) |
2858 | .int JMPTBL (L(bwd_write_6bytes), L(table_144_bytes_bwd)) |
2859 | .int JMPTBL (L(bwd_write_7bytes), L(table_144_bytes_bwd)) |
2860 | .int JMPTBL (L(bwd_write_8bytes), L(table_144_bytes_bwd)) |
2861 | .int JMPTBL (L(bwd_write_9bytes), L(table_144_bytes_bwd)) |
2862 | .int JMPTBL (L(bwd_write_10bytes), L(table_144_bytes_bwd)) |
2863 | .int JMPTBL (L(bwd_write_11bytes), L(table_144_bytes_bwd)) |
2864 | .int JMPTBL (L(bwd_write_12bytes), L(table_144_bytes_bwd)) |
2865 | .int JMPTBL (L(bwd_write_13bytes), L(table_144_bytes_bwd)) |
2866 | .int JMPTBL (L(bwd_write_14bytes), L(table_144_bytes_bwd)) |
2867 | .int JMPTBL (L(bwd_write_15bytes), L(table_144_bytes_bwd)) |
2868 | .int JMPTBL (L(bwd_write_16bytes), L(table_144_bytes_bwd)) |
2869 | .int JMPTBL (L(bwd_write_17bytes), L(table_144_bytes_bwd)) |
2870 | .int JMPTBL (L(bwd_write_18bytes), L(table_144_bytes_bwd)) |
2871 | .int JMPTBL (L(bwd_write_19bytes), L(table_144_bytes_bwd)) |
2872 | .int JMPTBL (L(bwd_write_20bytes), L(table_144_bytes_bwd)) |
2873 | .int JMPTBL (L(bwd_write_21bytes), L(table_144_bytes_bwd)) |
2874 | .int JMPTBL (L(bwd_write_22bytes), L(table_144_bytes_bwd)) |
2875 | .int JMPTBL (L(bwd_write_23bytes), L(table_144_bytes_bwd)) |
2876 | .int JMPTBL (L(bwd_write_24bytes), L(table_144_bytes_bwd)) |
2877 | .int JMPTBL (L(bwd_write_25bytes), L(table_144_bytes_bwd)) |
2878 | .int JMPTBL (L(bwd_write_26bytes), L(table_144_bytes_bwd)) |
2879 | .int JMPTBL (L(bwd_write_27bytes), L(table_144_bytes_bwd)) |
2880 | .int JMPTBL (L(bwd_write_28bytes), L(table_144_bytes_bwd)) |
2881 | .int JMPTBL (L(bwd_write_29bytes), L(table_144_bytes_bwd)) |
2882 | .int JMPTBL (L(bwd_write_30bytes), L(table_144_bytes_bwd)) |
2883 | .int JMPTBL (L(bwd_write_31bytes), L(table_144_bytes_bwd)) |
2884 | .int JMPTBL (L(bwd_write_32bytes), L(table_144_bytes_bwd)) |
2885 | .int JMPTBL (L(bwd_write_33bytes), L(table_144_bytes_bwd)) |
2886 | .int JMPTBL (L(bwd_write_34bytes), L(table_144_bytes_bwd)) |
2887 | .int JMPTBL (L(bwd_write_35bytes), L(table_144_bytes_bwd)) |
2888 | .int JMPTBL (L(bwd_write_36bytes), L(table_144_bytes_bwd)) |
2889 | .int JMPTBL (L(bwd_write_37bytes), L(table_144_bytes_bwd)) |
2890 | .int JMPTBL (L(bwd_write_38bytes), L(table_144_bytes_bwd)) |
2891 | .int JMPTBL (L(bwd_write_39bytes), L(table_144_bytes_bwd)) |
2892 | .int JMPTBL (L(bwd_write_40bytes), L(table_144_bytes_bwd)) |
2893 | .int JMPTBL (L(bwd_write_41bytes), L(table_144_bytes_bwd)) |
2894 | .int JMPTBL (L(bwd_write_42bytes), L(table_144_bytes_bwd)) |
2895 | .int JMPTBL (L(bwd_write_43bytes), L(table_144_bytes_bwd)) |
2896 | .int JMPTBL (L(bwd_write_44bytes), L(table_144_bytes_bwd)) |
2897 | .int JMPTBL (L(bwd_write_45bytes), L(table_144_bytes_bwd)) |
2898 | .int JMPTBL (L(bwd_write_46bytes), L(table_144_bytes_bwd)) |
2899 | .int JMPTBL (L(bwd_write_47bytes), L(table_144_bytes_bwd)) |
2900 | .int JMPTBL (L(bwd_write_48bytes), L(table_144_bytes_bwd)) |
2901 | .int JMPTBL (L(bwd_write_49bytes), L(table_144_bytes_bwd)) |
2902 | .int JMPTBL (L(bwd_write_50bytes), L(table_144_bytes_bwd)) |
2903 | .int JMPTBL (L(bwd_write_51bytes), L(table_144_bytes_bwd)) |
2904 | .int JMPTBL (L(bwd_write_52bytes), L(table_144_bytes_bwd)) |
2905 | .int JMPTBL (L(bwd_write_53bytes), L(table_144_bytes_bwd)) |
2906 | .int JMPTBL (L(bwd_write_54bytes), L(table_144_bytes_bwd)) |
2907 | .int JMPTBL (L(bwd_write_55bytes), L(table_144_bytes_bwd)) |
2908 | .int JMPTBL (L(bwd_write_56bytes), L(table_144_bytes_bwd)) |
2909 | .int JMPTBL (L(bwd_write_57bytes), L(table_144_bytes_bwd)) |
2910 | .int JMPTBL (L(bwd_write_58bytes), L(table_144_bytes_bwd)) |
2911 | .int JMPTBL (L(bwd_write_59bytes), L(table_144_bytes_bwd)) |
2912 | .int JMPTBL (L(bwd_write_60bytes), L(table_144_bytes_bwd)) |
2913 | .int JMPTBL (L(bwd_write_61bytes), L(table_144_bytes_bwd)) |
2914 | .int JMPTBL (L(bwd_write_62bytes), L(table_144_bytes_bwd)) |
2915 | .int JMPTBL (L(bwd_write_63bytes), L(table_144_bytes_bwd)) |
2916 | .int JMPTBL (L(bwd_write_64bytes), L(table_144_bytes_bwd)) |
2917 | .int JMPTBL (L(bwd_write_65bytes), L(table_144_bytes_bwd)) |
2918 | .int JMPTBL (L(bwd_write_66bytes), L(table_144_bytes_bwd)) |
2919 | .int JMPTBL (L(bwd_write_67bytes), L(table_144_bytes_bwd)) |
2920 | .int JMPTBL (L(bwd_write_68bytes), L(table_144_bytes_bwd)) |
2921 | .int JMPTBL (L(bwd_write_69bytes), L(table_144_bytes_bwd)) |
2922 | .int JMPTBL (L(bwd_write_70bytes), L(table_144_bytes_bwd)) |
2923 | .int JMPTBL (L(bwd_write_71bytes), L(table_144_bytes_bwd)) |
2924 | .int JMPTBL (L(bwd_write_72bytes), L(table_144_bytes_bwd)) |
2925 | .int JMPTBL (L(bwd_write_73bytes), L(table_144_bytes_bwd)) |
2926 | .int JMPTBL (L(bwd_write_74bytes), L(table_144_bytes_bwd)) |
2927 | .int JMPTBL (L(bwd_write_75bytes), L(table_144_bytes_bwd)) |
2928 | .int JMPTBL (L(bwd_write_76bytes), L(table_144_bytes_bwd)) |
2929 | .int JMPTBL (L(bwd_write_77bytes), L(table_144_bytes_bwd)) |
2930 | .int JMPTBL (L(bwd_write_78bytes), L(table_144_bytes_bwd)) |
2931 | .int JMPTBL (L(bwd_write_79bytes), L(table_144_bytes_bwd)) |
2932 | .int JMPTBL (L(bwd_write_80bytes), L(table_144_bytes_bwd)) |
2933 | .int JMPTBL (L(bwd_write_81bytes), L(table_144_bytes_bwd)) |
2934 | .int JMPTBL (L(bwd_write_82bytes), L(table_144_bytes_bwd)) |
2935 | .int JMPTBL (L(bwd_write_83bytes), L(table_144_bytes_bwd)) |
2936 | .int JMPTBL (L(bwd_write_84bytes), L(table_144_bytes_bwd)) |
2937 | .int JMPTBL (L(bwd_write_85bytes), L(table_144_bytes_bwd)) |
2938 | .int JMPTBL (L(bwd_write_86bytes), L(table_144_bytes_bwd)) |
2939 | .int JMPTBL (L(bwd_write_87bytes), L(table_144_bytes_bwd)) |
2940 | .int JMPTBL (L(bwd_write_88bytes), L(table_144_bytes_bwd)) |
2941 | .int JMPTBL (L(bwd_write_89bytes), L(table_144_bytes_bwd)) |
2942 | .int JMPTBL (L(bwd_write_90bytes), L(table_144_bytes_bwd)) |
2943 | .int JMPTBL (L(bwd_write_91bytes), L(table_144_bytes_bwd)) |
2944 | .int JMPTBL (L(bwd_write_92bytes), L(table_144_bytes_bwd)) |
2945 | .int JMPTBL (L(bwd_write_93bytes), L(table_144_bytes_bwd)) |
2946 | .int JMPTBL (L(bwd_write_94bytes), L(table_144_bytes_bwd)) |
2947 | .int JMPTBL (L(bwd_write_95bytes), L(table_144_bytes_bwd)) |
2948 | .int JMPTBL (L(bwd_write_96bytes), L(table_144_bytes_bwd)) |
2949 | .int JMPTBL (L(bwd_write_97bytes), L(table_144_bytes_bwd)) |
2950 | .int JMPTBL (L(bwd_write_98bytes), L(table_144_bytes_bwd)) |
2951 | .int JMPTBL (L(bwd_write_99bytes), L(table_144_bytes_bwd)) |
2952 | .int JMPTBL (L(bwd_write_100bytes), L(table_144_bytes_bwd)) |
2953 | .int JMPTBL (L(bwd_write_101bytes), L(table_144_bytes_bwd)) |
2954 | .int JMPTBL (L(bwd_write_102bytes), L(table_144_bytes_bwd)) |
2955 | .int JMPTBL (L(bwd_write_103bytes), L(table_144_bytes_bwd)) |
2956 | .int JMPTBL (L(bwd_write_104bytes), L(table_144_bytes_bwd)) |
2957 | .int JMPTBL (L(bwd_write_105bytes), L(table_144_bytes_bwd)) |
2958 | .int JMPTBL (L(bwd_write_106bytes), L(table_144_bytes_bwd)) |
2959 | .int JMPTBL (L(bwd_write_107bytes), L(table_144_bytes_bwd)) |
2960 | .int JMPTBL (L(bwd_write_108bytes), L(table_144_bytes_bwd)) |
2961 | .int JMPTBL (L(bwd_write_109bytes), L(table_144_bytes_bwd)) |
2962 | .int JMPTBL (L(bwd_write_110bytes), L(table_144_bytes_bwd)) |
2963 | .int JMPTBL (L(bwd_write_111bytes), L(table_144_bytes_bwd)) |
2964 | .int JMPTBL (L(bwd_write_112bytes), L(table_144_bytes_bwd)) |
2965 | .int JMPTBL (L(bwd_write_113bytes), L(table_144_bytes_bwd)) |
2966 | .int JMPTBL (L(bwd_write_114bytes), L(table_144_bytes_bwd)) |
2967 | .int JMPTBL (L(bwd_write_115bytes), L(table_144_bytes_bwd)) |
2968 | .int JMPTBL (L(bwd_write_116bytes), L(table_144_bytes_bwd)) |
2969 | .int JMPTBL (L(bwd_write_117bytes), L(table_144_bytes_bwd)) |
2970 | .int JMPTBL (L(bwd_write_118bytes), L(table_144_bytes_bwd)) |
2971 | .int JMPTBL (L(bwd_write_119bytes), L(table_144_bytes_bwd)) |
2972 | .int JMPTBL (L(bwd_write_120bytes), L(table_144_bytes_bwd)) |
2973 | .int JMPTBL (L(bwd_write_121bytes), L(table_144_bytes_bwd)) |
2974 | .int JMPTBL (L(bwd_write_122bytes), L(table_144_bytes_bwd)) |
2975 | .int JMPTBL (L(bwd_write_123bytes), L(table_144_bytes_bwd)) |
2976 | .int JMPTBL (L(bwd_write_124bytes), L(table_144_bytes_bwd)) |
2977 | .int JMPTBL (L(bwd_write_125bytes), L(table_144_bytes_bwd)) |
2978 | .int JMPTBL (L(bwd_write_126bytes), L(table_144_bytes_bwd)) |
2979 | .int JMPTBL (L(bwd_write_127bytes), L(table_144_bytes_bwd)) |
2980 | .int JMPTBL (L(bwd_write_128bytes), L(table_144_bytes_bwd)) |
2981 | .int JMPTBL (L(bwd_write_129bytes), L(table_144_bytes_bwd)) |
2982 | .int JMPTBL (L(bwd_write_130bytes), L(table_144_bytes_bwd)) |
2983 | .int JMPTBL (L(bwd_write_131bytes), L(table_144_bytes_bwd)) |
2984 | .int JMPTBL (L(bwd_write_132bytes), L(table_144_bytes_bwd)) |
2985 | .int JMPTBL (L(bwd_write_133bytes), L(table_144_bytes_bwd)) |
2986 | .int JMPTBL (L(bwd_write_134bytes), L(table_144_bytes_bwd)) |
2987 | .int JMPTBL (L(bwd_write_135bytes), L(table_144_bytes_bwd)) |
2988 | .int JMPTBL (L(bwd_write_136bytes), L(table_144_bytes_bwd)) |
2989 | .int JMPTBL (L(bwd_write_137bytes), L(table_144_bytes_bwd)) |
2990 | .int JMPTBL (L(bwd_write_138bytes), L(table_144_bytes_bwd)) |
2991 | .int JMPTBL (L(bwd_write_139bytes), L(table_144_bytes_bwd)) |
2992 | .int JMPTBL (L(bwd_write_140bytes), L(table_144_bytes_bwd)) |
2993 | .int JMPTBL (L(bwd_write_141bytes), L(table_144_bytes_bwd)) |
2994 | .int JMPTBL (L(bwd_write_142bytes), L(table_144_bytes_bwd)) |
2995 | .int JMPTBL (L(bwd_write_143bytes), L(table_144_bytes_bwd)) |
2996 | |
2997 | .p2align 3 |
2998 | L(table_144_bytes_fwd): |
2999 | .int JMPTBL (L(fwd_write_0bytes), L(table_144_bytes_fwd)) |
3000 | .int JMPTBL (L(fwd_write_1bytes), L(table_144_bytes_fwd)) |
3001 | .int JMPTBL (L(fwd_write_2bytes), L(table_144_bytes_fwd)) |
3002 | .int JMPTBL (L(fwd_write_3bytes), L(table_144_bytes_fwd)) |
3003 | .int JMPTBL (L(fwd_write_4bytes), L(table_144_bytes_fwd)) |
3004 | .int JMPTBL (L(fwd_write_5bytes), L(table_144_bytes_fwd)) |
3005 | .int JMPTBL (L(fwd_write_6bytes), L(table_144_bytes_fwd)) |
3006 | .int JMPTBL (L(fwd_write_7bytes), L(table_144_bytes_fwd)) |
3007 | .int JMPTBL (L(fwd_write_8bytes), L(table_144_bytes_fwd)) |
3008 | .int JMPTBL (L(fwd_write_9bytes), L(table_144_bytes_fwd)) |
3009 | .int JMPTBL (L(fwd_write_10bytes), L(table_144_bytes_fwd)) |
3010 | .int JMPTBL (L(fwd_write_11bytes), L(table_144_bytes_fwd)) |
3011 | .int JMPTBL (L(fwd_write_12bytes), L(table_144_bytes_fwd)) |
3012 | .int JMPTBL (L(fwd_write_13bytes), L(table_144_bytes_fwd)) |
3013 | .int JMPTBL (L(fwd_write_14bytes), L(table_144_bytes_fwd)) |
3014 | .int JMPTBL (L(fwd_write_15bytes), L(table_144_bytes_fwd)) |
3015 | .int JMPTBL (L(fwd_write_16bytes), L(table_144_bytes_fwd)) |
3016 | .int JMPTBL (L(fwd_write_17bytes), L(table_144_bytes_fwd)) |
3017 | .int JMPTBL (L(fwd_write_18bytes), L(table_144_bytes_fwd)) |
3018 | .int JMPTBL (L(fwd_write_19bytes), L(table_144_bytes_fwd)) |
3019 | .int JMPTBL (L(fwd_write_20bytes), L(table_144_bytes_fwd)) |
3020 | .int JMPTBL (L(fwd_write_21bytes), L(table_144_bytes_fwd)) |
3021 | .int JMPTBL (L(fwd_write_22bytes), L(table_144_bytes_fwd)) |
3022 | .int JMPTBL (L(fwd_write_23bytes), L(table_144_bytes_fwd)) |
3023 | .int JMPTBL (L(fwd_write_24bytes), L(table_144_bytes_fwd)) |
3024 | .int JMPTBL (L(fwd_write_25bytes), L(table_144_bytes_fwd)) |
3025 | .int JMPTBL (L(fwd_write_26bytes), L(table_144_bytes_fwd)) |
3026 | .int JMPTBL (L(fwd_write_27bytes), L(table_144_bytes_fwd)) |
3027 | .int JMPTBL (L(fwd_write_28bytes), L(table_144_bytes_fwd)) |
3028 | .int JMPTBL (L(fwd_write_29bytes), L(table_144_bytes_fwd)) |
3029 | .int JMPTBL (L(fwd_write_30bytes), L(table_144_bytes_fwd)) |
3030 | .int JMPTBL (L(fwd_write_31bytes), L(table_144_bytes_fwd)) |
3031 | .int JMPTBL (L(fwd_write_32bytes), L(table_144_bytes_fwd)) |
3032 | .int JMPTBL (L(fwd_write_33bytes), L(table_144_bytes_fwd)) |
3033 | .int JMPTBL (L(fwd_write_34bytes), L(table_144_bytes_fwd)) |
3034 | .int JMPTBL (L(fwd_write_35bytes), L(table_144_bytes_fwd)) |
3035 | .int JMPTBL (L(fwd_write_36bytes), L(table_144_bytes_fwd)) |
3036 | .int JMPTBL (L(fwd_write_37bytes), L(table_144_bytes_fwd)) |
3037 | .int JMPTBL (L(fwd_write_38bytes), L(table_144_bytes_fwd)) |
3038 | .int JMPTBL (L(fwd_write_39bytes), L(table_144_bytes_fwd)) |
3039 | .int JMPTBL (L(fwd_write_40bytes), L(table_144_bytes_fwd)) |
3040 | .int JMPTBL (L(fwd_write_41bytes), L(table_144_bytes_fwd)) |
3041 | .int JMPTBL (L(fwd_write_42bytes), L(table_144_bytes_fwd)) |
3042 | .int JMPTBL (L(fwd_write_43bytes), L(table_144_bytes_fwd)) |
3043 | .int JMPTBL (L(fwd_write_44bytes), L(table_144_bytes_fwd)) |
3044 | .int JMPTBL (L(fwd_write_45bytes), L(table_144_bytes_fwd)) |
3045 | .int JMPTBL (L(fwd_write_46bytes), L(table_144_bytes_fwd)) |
3046 | .int JMPTBL (L(fwd_write_47bytes), L(table_144_bytes_fwd)) |
3047 | .int JMPTBL (L(fwd_write_48bytes), L(table_144_bytes_fwd)) |
3048 | .int JMPTBL (L(fwd_write_49bytes), L(table_144_bytes_fwd)) |
3049 | .int JMPTBL (L(fwd_write_50bytes), L(table_144_bytes_fwd)) |
3050 | .int JMPTBL (L(fwd_write_51bytes), L(table_144_bytes_fwd)) |
3051 | .int JMPTBL (L(fwd_write_52bytes), L(table_144_bytes_fwd)) |
3052 | .int JMPTBL (L(fwd_write_53bytes), L(table_144_bytes_fwd)) |
3053 | .int JMPTBL (L(fwd_write_54bytes), L(table_144_bytes_fwd)) |
3054 | .int JMPTBL (L(fwd_write_55bytes), L(table_144_bytes_fwd)) |
3055 | .int JMPTBL (L(fwd_write_56bytes), L(table_144_bytes_fwd)) |
3056 | .int JMPTBL (L(fwd_write_57bytes), L(table_144_bytes_fwd)) |
3057 | .int JMPTBL (L(fwd_write_58bytes), L(table_144_bytes_fwd)) |
3058 | .int JMPTBL (L(fwd_write_59bytes), L(table_144_bytes_fwd)) |
3059 | .int JMPTBL (L(fwd_write_60bytes), L(table_144_bytes_fwd)) |
3060 | .int JMPTBL (L(fwd_write_61bytes), L(table_144_bytes_fwd)) |
3061 | .int JMPTBL (L(fwd_write_62bytes), L(table_144_bytes_fwd)) |
3062 | .int JMPTBL (L(fwd_write_63bytes), L(table_144_bytes_fwd)) |
3063 | .int JMPTBL (L(fwd_write_64bytes), L(table_144_bytes_fwd)) |
3064 | .int JMPTBL (L(fwd_write_65bytes), L(table_144_bytes_fwd)) |
3065 | .int JMPTBL (L(fwd_write_66bytes), L(table_144_bytes_fwd)) |
3066 | .int JMPTBL (L(fwd_write_67bytes), L(table_144_bytes_fwd)) |
3067 | .int JMPTBL (L(fwd_write_68bytes), L(table_144_bytes_fwd)) |
3068 | .int JMPTBL (L(fwd_write_69bytes), L(table_144_bytes_fwd)) |
3069 | .int JMPTBL (L(fwd_write_70bytes), L(table_144_bytes_fwd)) |
3070 | .int JMPTBL (L(fwd_write_71bytes), L(table_144_bytes_fwd)) |
3071 | .int JMPTBL (L(fwd_write_72bytes), L(table_144_bytes_fwd)) |
3072 | .int JMPTBL (L(fwd_write_73bytes), L(table_144_bytes_fwd)) |
3073 | .int JMPTBL (L(fwd_write_74bytes), L(table_144_bytes_fwd)) |
3074 | .int JMPTBL (L(fwd_write_75bytes), L(table_144_bytes_fwd)) |
3075 | .int JMPTBL (L(fwd_write_76bytes), L(table_144_bytes_fwd)) |
3076 | .int JMPTBL (L(fwd_write_77bytes), L(table_144_bytes_fwd)) |
3077 | .int JMPTBL (L(fwd_write_78bytes), L(table_144_bytes_fwd)) |
3078 | .int JMPTBL (L(fwd_write_79bytes), L(table_144_bytes_fwd)) |
3079 | .int JMPTBL (L(fwd_write_80bytes), L(table_144_bytes_fwd)) |
3080 | .int JMPTBL (L(fwd_write_81bytes), L(table_144_bytes_fwd)) |
3081 | .int JMPTBL (L(fwd_write_82bytes), L(table_144_bytes_fwd)) |
3082 | .int JMPTBL (L(fwd_write_83bytes), L(table_144_bytes_fwd)) |
3083 | .int JMPTBL (L(fwd_write_84bytes), L(table_144_bytes_fwd)) |
3084 | .int JMPTBL (L(fwd_write_85bytes), L(table_144_bytes_fwd)) |
3085 | .int JMPTBL (L(fwd_write_86bytes), L(table_144_bytes_fwd)) |
3086 | .int JMPTBL (L(fwd_write_87bytes), L(table_144_bytes_fwd)) |
3087 | .int JMPTBL (L(fwd_write_88bytes), L(table_144_bytes_fwd)) |
3088 | .int JMPTBL (L(fwd_write_89bytes), L(table_144_bytes_fwd)) |
3089 | .int JMPTBL (L(fwd_write_90bytes), L(table_144_bytes_fwd)) |
3090 | .int JMPTBL (L(fwd_write_91bytes), L(table_144_bytes_fwd)) |
3091 | .int JMPTBL (L(fwd_write_92bytes), L(table_144_bytes_fwd)) |
3092 | .int JMPTBL (L(fwd_write_93bytes), L(table_144_bytes_fwd)) |
3093 | .int JMPTBL (L(fwd_write_94bytes), L(table_144_bytes_fwd)) |
3094 | .int JMPTBL (L(fwd_write_95bytes), L(table_144_bytes_fwd)) |
3095 | .int JMPTBL (L(fwd_write_96bytes), L(table_144_bytes_fwd)) |
3096 | .int JMPTBL (L(fwd_write_97bytes), L(table_144_bytes_fwd)) |
3097 | .int JMPTBL (L(fwd_write_98bytes), L(table_144_bytes_fwd)) |
3098 | .int JMPTBL (L(fwd_write_99bytes), L(table_144_bytes_fwd)) |
3099 | .int JMPTBL (L(fwd_write_100bytes), L(table_144_bytes_fwd)) |
3100 | .int JMPTBL (L(fwd_write_101bytes), L(table_144_bytes_fwd)) |
3101 | .int JMPTBL (L(fwd_write_102bytes), L(table_144_bytes_fwd)) |
3102 | .int JMPTBL (L(fwd_write_103bytes), L(table_144_bytes_fwd)) |
3103 | .int JMPTBL (L(fwd_write_104bytes), L(table_144_bytes_fwd)) |
3104 | .int JMPTBL (L(fwd_write_105bytes), L(table_144_bytes_fwd)) |
3105 | .int JMPTBL (L(fwd_write_106bytes), L(table_144_bytes_fwd)) |
3106 | .int JMPTBL (L(fwd_write_107bytes), L(table_144_bytes_fwd)) |
3107 | .int JMPTBL (L(fwd_write_108bytes), L(table_144_bytes_fwd)) |
3108 | .int JMPTBL (L(fwd_write_109bytes), L(table_144_bytes_fwd)) |
3109 | .int JMPTBL (L(fwd_write_110bytes), L(table_144_bytes_fwd)) |
3110 | .int JMPTBL (L(fwd_write_111bytes), L(table_144_bytes_fwd)) |
3111 | .int JMPTBL (L(fwd_write_112bytes), L(table_144_bytes_fwd)) |
3112 | .int JMPTBL (L(fwd_write_113bytes), L(table_144_bytes_fwd)) |
3113 | .int JMPTBL (L(fwd_write_114bytes), L(table_144_bytes_fwd)) |
3114 | .int JMPTBL (L(fwd_write_115bytes), L(table_144_bytes_fwd)) |
3115 | .int JMPTBL (L(fwd_write_116bytes), L(table_144_bytes_fwd)) |
3116 | .int JMPTBL (L(fwd_write_117bytes), L(table_144_bytes_fwd)) |
3117 | .int JMPTBL (L(fwd_write_118bytes), L(table_144_bytes_fwd)) |
3118 | .int JMPTBL (L(fwd_write_119bytes), L(table_144_bytes_fwd)) |
3119 | .int JMPTBL (L(fwd_write_120bytes), L(table_144_bytes_fwd)) |
3120 | .int JMPTBL (L(fwd_write_121bytes), L(table_144_bytes_fwd)) |
3121 | .int JMPTBL (L(fwd_write_122bytes), L(table_144_bytes_fwd)) |
3122 | .int JMPTBL (L(fwd_write_123bytes), L(table_144_bytes_fwd)) |
3123 | .int JMPTBL (L(fwd_write_124bytes), L(table_144_bytes_fwd)) |
3124 | .int JMPTBL (L(fwd_write_125bytes), L(table_144_bytes_fwd)) |
3125 | .int JMPTBL (L(fwd_write_126bytes), L(table_144_bytes_fwd)) |
3126 | .int JMPTBL (L(fwd_write_127bytes), L(table_144_bytes_fwd)) |
3127 | .int JMPTBL (L(fwd_write_128bytes), L(table_144_bytes_fwd)) |
3128 | .int JMPTBL (L(fwd_write_129bytes), L(table_144_bytes_fwd)) |
3129 | .int JMPTBL (L(fwd_write_130bytes), L(table_144_bytes_fwd)) |
3130 | .int JMPTBL (L(fwd_write_131bytes), L(table_144_bytes_fwd)) |
3131 | .int JMPTBL (L(fwd_write_132bytes), L(table_144_bytes_fwd)) |
3132 | .int JMPTBL (L(fwd_write_133bytes), L(table_144_bytes_fwd)) |
3133 | .int JMPTBL (L(fwd_write_134bytes), L(table_144_bytes_fwd)) |
3134 | .int JMPTBL (L(fwd_write_135bytes), L(table_144_bytes_fwd)) |
3135 | .int JMPTBL (L(fwd_write_136bytes), L(table_144_bytes_fwd)) |
3136 | .int JMPTBL (L(fwd_write_137bytes), L(table_144_bytes_fwd)) |
3137 | .int JMPTBL (L(fwd_write_138bytes), L(table_144_bytes_fwd)) |
3138 | .int JMPTBL (L(fwd_write_139bytes), L(table_144_bytes_fwd)) |
3139 | .int JMPTBL (L(fwd_write_140bytes), L(table_144_bytes_fwd)) |
3140 | .int JMPTBL (L(fwd_write_141bytes), L(table_144_bytes_fwd)) |
3141 | .int JMPTBL (L(fwd_write_142bytes), L(table_144_bytes_fwd)) |
3142 | .int JMPTBL (L(fwd_write_143bytes), L(table_144_bytes_fwd)) |
3143 | |
3144 | .p2align 3 |
3145 | L(shl_table_fwd): |
3146 | .int JMPTBL (L(shl_0), L(shl_table_fwd)) |
3147 | .int JMPTBL (L(shl_1), L(shl_table_fwd)) |
3148 | .int JMPTBL (L(shl_2), L(shl_table_fwd)) |
3149 | .int JMPTBL (L(shl_3), L(shl_table_fwd)) |
3150 | .int JMPTBL (L(shl_4), L(shl_table_fwd)) |
3151 | .int JMPTBL (L(shl_5), L(shl_table_fwd)) |
3152 | .int JMPTBL (L(shl_6), L(shl_table_fwd)) |
3153 | .int JMPTBL (L(shl_7), L(shl_table_fwd)) |
3154 | .int JMPTBL (L(shl_8), L(shl_table_fwd)) |
3155 | .int JMPTBL (L(shl_9), L(shl_table_fwd)) |
3156 | .int JMPTBL (L(shl_10), L(shl_table_fwd)) |
3157 | .int JMPTBL (L(shl_11), L(shl_table_fwd)) |
3158 | .int JMPTBL (L(shl_12), L(shl_table_fwd)) |
3159 | .int JMPTBL (L(shl_13), L(shl_table_fwd)) |
3160 | .int JMPTBL (L(shl_14), L(shl_table_fwd)) |
3161 | .int JMPTBL (L(shl_15), L(shl_table_fwd)) |
3162 | |
3163 | .p2align 3 |
3164 | L(shl_table_bwd): |
3165 | .int JMPTBL (L(shl_0_bwd), L(shl_table_bwd)) |
3166 | .int JMPTBL (L(shl_1_bwd), L(shl_table_bwd)) |
3167 | .int JMPTBL (L(shl_2_bwd), L(shl_table_bwd)) |
3168 | .int JMPTBL (L(shl_3_bwd), L(shl_table_bwd)) |
3169 | .int JMPTBL (L(shl_4_bwd), L(shl_table_bwd)) |
3170 | .int JMPTBL (L(shl_5_bwd), L(shl_table_bwd)) |
3171 | .int JMPTBL (L(shl_6_bwd), L(shl_table_bwd)) |
3172 | .int JMPTBL (L(shl_7_bwd), L(shl_table_bwd)) |
3173 | .int JMPTBL (L(shl_8_bwd), L(shl_table_bwd)) |
3174 | .int JMPTBL (L(shl_9_bwd), L(shl_table_bwd)) |
3175 | .int JMPTBL (L(shl_10_bwd), L(shl_table_bwd)) |
3176 | .int JMPTBL (L(shl_11_bwd), L(shl_table_bwd)) |
3177 | .int JMPTBL (L(shl_12_bwd), L(shl_table_bwd)) |
3178 | .int JMPTBL (L(shl_13_bwd), L(shl_table_bwd)) |
3179 | .int JMPTBL (L(shl_14_bwd), L(shl_table_bwd)) |
3180 | .int JMPTBL (L(shl_15_bwd), L(shl_table_bwd)) |
3181 | |
3182 | #endif |
3183 | |