VirtualBox

source: vbox/trunk/src/libs/openssl-3.0.1/crypto/genasm-nasm/sha1-mb-x86_64.S@ 94083

Last change on this file since 94083 was 94083, checked in by vboxsync, 3 years ago

libs/openssl-3.0.1: Recreate asm files, bugref:10128
File size: 147.2 KB
Line 
1default rel
2%define XMMWORD
3%define YMMWORD
4%define ZMMWORD
5section .text code align=64
6
7
8EXTERN OPENSSL_ia32cap_P
9
10global sha1_multi_block
11
12ALIGN 32
13sha1_multi_block:
14 mov QWORD[8+rsp],rdi ;WIN64 prologue
15 mov QWORD[16+rsp],rsi
16 mov rax,rsp
17$L$SEH_begin_sha1_multi_block:
18 mov rdi,rcx
19 mov rsi,rdx
20 mov rdx,r8
21
22
23
24 mov rcx,QWORD[((OPENSSL_ia32cap_P+4))]
25 bt rcx,61
26 jc NEAR _shaext_shortcut
27 test ecx,268435456
28 jnz NEAR _avx_shortcut
29 mov rax,rsp
30
31 push rbx
32
33 push rbp
34
35 lea rsp,[((-168))+rsp]
36 movaps XMMWORD[rsp],xmm6
37 movaps XMMWORD[16+rsp],xmm7
38 movaps XMMWORD[32+rsp],xmm8
39 movaps XMMWORD[48+rsp],xmm9
40 movaps XMMWORD[(-120)+rax],xmm10
41 movaps XMMWORD[(-104)+rax],xmm11
42 movaps XMMWORD[(-88)+rax],xmm12
43 movaps XMMWORD[(-72)+rax],xmm13
44 movaps XMMWORD[(-56)+rax],xmm14
45 movaps XMMWORD[(-40)+rax],xmm15
46 sub rsp,288
47 and rsp,-256
48 mov QWORD[272+rsp],rax
49
50$L$body:
51 lea rbp,[K_XX_XX]
52 lea rbx,[256+rsp]
53
54$L$oop_grande:
55 mov DWORD[280+rsp],edx
56 xor edx,edx
57
58 mov r8,QWORD[rsi]
59
60 mov ecx,DWORD[8+rsi]
61 cmp ecx,edx
62 cmovg edx,ecx
63 test ecx,ecx
64 mov DWORD[rbx],ecx
65 cmovle r8,rbp
66
67 mov r9,QWORD[16+rsi]
68
69 mov ecx,DWORD[24+rsi]
70 cmp ecx,edx
71 cmovg edx,ecx
72 test ecx,ecx
73 mov DWORD[4+rbx],ecx
74 cmovle r9,rbp
75
76 mov r10,QWORD[32+rsi]
77
78 mov ecx,DWORD[40+rsi]
79 cmp ecx,edx
80 cmovg edx,ecx
81 test ecx,ecx
82 mov DWORD[8+rbx],ecx
83 cmovle r10,rbp
84
85 mov r11,QWORD[48+rsi]
86
87 mov ecx,DWORD[56+rsi]
88 cmp ecx,edx
89 cmovg edx,ecx
90 test ecx,ecx
91 mov DWORD[12+rbx],ecx
92 cmovle r11,rbp
93 test edx,edx
94 jz NEAR $L$done
95
96 movdqu xmm10,XMMWORD[rdi]
97 lea rax,[128+rsp]
98 movdqu xmm11,XMMWORD[32+rdi]
99 movdqu xmm12,XMMWORD[64+rdi]
100 movdqu xmm13,XMMWORD[96+rdi]
101 movdqu xmm14,XMMWORD[128+rdi]
102 movdqa xmm5,XMMWORD[96+rbp]
103 movdqa xmm15,XMMWORD[((-32))+rbp]
104 jmp NEAR $L$oop
105
106ALIGN 32
107$L$oop:
108 movd xmm0,DWORD[r8]
109 lea r8,[64+r8]
110 movd xmm2,DWORD[r9]
111 lea r9,[64+r9]
112 movd xmm3,DWORD[r10]
113 lea r10,[64+r10]
114 movd xmm4,DWORD[r11]
115 lea r11,[64+r11]
116 punpckldq xmm0,xmm3
117 movd xmm1,DWORD[((-60))+r8]
118 punpckldq xmm2,xmm4
119 movd xmm9,DWORD[((-60))+r9]
120 punpckldq xmm0,xmm2
121 movd xmm8,DWORD[((-60))+r10]
122DB 102,15,56,0,197
123 movd xmm7,DWORD[((-60))+r11]
124 punpckldq xmm1,xmm8
125 movdqa xmm8,xmm10
126 paddd xmm14,xmm15
127 punpckldq xmm9,xmm7
128 movdqa xmm7,xmm11
129 movdqa xmm6,xmm11
130 pslld xmm8,5
131 pandn xmm7,xmm13
132 pand xmm6,xmm12
133 punpckldq xmm1,xmm9
134 movdqa xmm9,xmm10
135
136 movdqa XMMWORD[(0-128)+rax],xmm0
137 paddd xmm14,xmm0
138 movd xmm2,DWORD[((-56))+r8]
139 psrld xmm9,27
140 pxor xmm6,xmm7
141 movdqa xmm7,xmm11
142
143 por xmm8,xmm9
144 movd xmm9,DWORD[((-56))+r9]
145 pslld xmm7,30
146 paddd xmm14,xmm6
147
148 psrld xmm11,2
149 paddd xmm14,xmm8
150DB 102,15,56,0,205
151 movd xmm8,DWORD[((-56))+r10]
152 por xmm11,xmm7
153 movd xmm7,DWORD[((-56))+r11]
154 punpckldq xmm2,xmm8
155 movdqa xmm8,xmm14
156 paddd xmm13,xmm15
157 punpckldq xmm9,xmm7
158 movdqa xmm7,xmm10
159 movdqa xmm6,xmm10
160 pslld xmm8,5
161 pandn xmm7,xmm12
162 pand xmm6,xmm11
163 punpckldq xmm2,xmm9
164 movdqa xmm9,xmm14
165
166 movdqa XMMWORD[(16-128)+rax],xmm1
167 paddd xmm13,xmm1
168 movd xmm3,DWORD[((-52))+r8]
169 psrld xmm9,27
170 pxor xmm6,xmm7
171 movdqa xmm7,xmm10
172
173 por xmm8,xmm9
174 movd xmm9,DWORD[((-52))+r9]
175 pslld xmm7,30
176 paddd xmm13,xmm6
177
178 psrld xmm10,2
179 paddd xmm13,xmm8
180DB 102,15,56,0,213
181 movd xmm8,DWORD[((-52))+r10]
182 por xmm10,xmm7
183 movd xmm7,DWORD[((-52))+r11]
184 punpckldq xmm3,xmm8
185 movdqa xmm8,xmm13
186 paddd xmm12,xmm15
187 punpckldq xmm9,xmm7
188 movdqa xmm7,xmm14
189 movdqa xmm6,xmm14
190 pslld xmm8,5
191 pandn xmm7,xmm11
192 pand xmm6,xmm10
193 punpckldq xmm3,xmm9
194 movdqa xmm9,xmm13
195
196 movdqa XMMWORD[(32-128)+rax],xmm2
197 paddd xmm12,xmm2
198 movd xmm4,DWORD[((-48))+r8]
199 psrld xmm9,27
200 pxor xmm6,xmm7
201 movdqa xmm7,xmm14
202
203 por xmm8,xmm9
204 movd xmm9,DWORD[((-48))+r9]
205 pslld xmm7,30
206 paddd xmm12,xmm6
207
208 psrld xmm14,2
209 paddd xmm12,xmm8
210DB 102,15,56,0,221
211 movd xmm8,DWORD[((-48))+r10]
212 por xmm14,xmm7
213 movd xmm7,DWORD[((-48))+r11]
214 punpckldq xmm4,xmm8
215 movdqa xmm8,xmm12
216 paddd xmm11,xmm15
217 punpckldq xmm9,xmm7
218 movdqa xmm7,xmm13
219 movdqa xmm6,xmm13
220 pslld xmm8,5
221 pandn xmm7,xmm10
222 pand xmm6,xmm14
223 punpckldq xmm4,xmm9
224 movdqa xmm9,xmm12
225
226 movdqa XMMWORD[(48-128)+rax],xmm3
227 paddd xmm11,xmm3
228 movd xmm0,DWORD[((-44))+r8]
229 psrld xmm9,27
230 pxor xmm6,xmm7
231 movdqa xmm7,xmm13
232
233 por xmm8,xmm9
234 movd xmm9,DWORD[((-44))+r9]
235 pslld xmm7,30
236 paddd xmm11,xmm6
237
238 psrld xmm13,2
239 paddd xmm11,xmm8
240DB 102,15,56,0,229
241 movd xmm8,DWORD[((-44))+r10]
242 por xmm13,xmm7
243 movd xmm7,DWORD[((-44))+r11]
244 punpckldq xmm0,xmm8
245 movdqa xmm8,xmm11
246 paddd xmm10,xmm15
247 punpckldq xmm9,xmm7
248 movdqa xmm7,xmm12
249 movdqa xmm6,xmm12
250 pslld xmm8,5
251 pandn xmm7,xmm14
252 pand xmm6,xmm13
253 punpckldq xmm0,xmm9
254 movdqa xmm9,xmm11
255
256 movdqa XMMWORD[(64-128)+rax],xmm4
257 paddd xmm10,xmm4
258 movd xmm1,DWORD[((-40))+r8]
259 psrld xmm9,27
260 pxor xmm6,xmm7
261 movdqa xmm7,xmm12
262
263 por xmm8,xmm9
264 movd xmm9,DWORD[((-40))+r9]
265 pslld xmm7,30
266 paddd xmm10,xmm6
267
268 psrld xmm12,2
269 paddd xmm10,xmm8
270DB 102,15,56,0,197
271 movd xmm8,DWORD[((-40))+r10]
272 por xmm12,xmm7
273 movd xmm7,DWORD[((-40))+r11]
274 punpckldq xmm1,xmm8
275 movdqa xmm8,xmm10
276 paddd xmm14,xmm15
277 punpckldq xmm9,xmm7
278 movdqa xmm7,xmm11
279 movdqa xmm6,xmm11
280 pslld xmm8,5
281 pandn xmm7,xmm13
282 pand xmm6,xmm12
283 punpckldq xmm1,xmm9
284 movdqa xmm9,xmm10
285
286 movdqa XMMWORD[(80-128)+rax],xmm0
287 paddd xmm14,xmm0
288 movd xmm2,DWORD[((-36))+r8]
289 psrld xmm9,27
290 pxor xmm6,xmm7
291 movdqa xmm7,xmm11
292
293 por xmm8,xmm9
294 movd xmm9,DWORD[((-36))+r9]
295 pslld xmm7,30
296 paddd xmm14,xmm6
297
298 psrld xmm11,2
299 paddd xmm14,xmm8
300DB 102,15,56,0,205
301 movd xmm8,DWORD[((-36))+r10]
302 por xmm11,xmm7
303 movd xmm7,DWORD[((-36))+r11]
304 punpckldq xmm2,xmm8
305 movdqa xmm8,xmm14
306 paddd xmm13,xmm15
307 punpckldq xmm9,xmm7
308 movdqa xmm7,xmm10
309 movdqa xmm6,xmm10
310 pslld xmm8,5
311 pandn xmm7,xmm12
312 pand xmm6,xmm11
313 punpckldq xmm2,xmm9
314 movdqa xmm9,xmm14
315
316 movdqa XMMWORD[(96-128)+rax],xmm1
317 paddd xmm13,xmm1
318 movd xmm3,DWORD[((-32))+r8]
319 psrld xmm9,27
320 pxor xmm6,xmm7
321 movdqa xmm7,xmm10
322
323 por xmm8,xmm9
324 movd xmm9,DWORD[((-32))+r9]
325 pslld xmm7,30
326 paddd xmm13,xmm6
327
328 psrld xmm10,2
329 paddd xmm13,xmm8
330DB 102,15,56,0,213
331 movd xmm8,DWORD[((-32))+r10]
332 por xmm10,xmm7
333 movd xmm7,DWORD[((-32))+r11]
334 punpckldq xmm3,xmm8
335 movdqa xmm8,xmm13
336 paddd xmm12,xmm15
337 punpckldq xmm9,xmm7
338 movdqa xmm7,xmm14
339 movdqa xmm6,xmm14
340 pslld xmm8,5
341 pandn xmm7,xmm11
342 pand xmm6,xmm10
343 punpckldq xmm3,xmm9
344 movdqa xmm9,xmm13
345
346 movdqa XMMWORD[(112-128)+rax],xmm2
347 paddd xmm12,xmm2
348 movd xmm4,DWORD[((-28))+r8]
349 psrld xmm9,27
350 pxor xmm6,xmm7
351 movdqa xmm7,xmm14
352
353 por xmm8,xmm9
354 movd xmm9,DWORD[((-28))+r9]
355 pslld xmm7,30
356 paddd xmm12,xmm6
357
358 psrld xmm14,2
359 paddd xmm12,xmm8
360DB 102,15,56,0,221
361 movd xmm8,DWORD[((-28))+r10]
362 por xmm14,xmm7
363 movd xmm7,DWORD[((-28))+r11]
364 punpckldq xmm4,xmm8
365 movdqa xmm8,xmm12
366 paddd xmm11,xmm15
367 punpckldq xmm9,xmm7
368 movdqa xmm7,xmm13
369 movdqa xmm6,xmm13
370 pslld xmm8,5
371 pandn xmm7,xmm10
372 pand xmm6,xmm14
373 punpckldq xmm4,xmm9
374 movdqa xmm9,xmm12
375
376 movdqa XMMWORD[(128-128)+rax],xmm3
377 paddd xmm11,xmm3
378 movd xmm0,DWORD[((-24))+r8]
379 psrld xmm9,27
380 pxor xmm6,xmm7
381 movdqa xmm7,xmm13
382
383 por xmm8,xmm9
384 movd xmm9,DWORD[((-24))+r9]
385 pslld xmm7,30
386 paddd xmm11,xmm6
387
388 psrld xmm13,2
389 paddd xmm11,xmm8
390DB 102,15,56,0,229
391 movd xmm8,DWORD[((-24))+r10]
392 por xmm13,xmm7
393 movd xmm7,DWORD[((-24))+r11]
394 punpckldq xmm0,xmm8
395 movdqa xmm8,xmm11
396 paddd xmm10,xmm15
397 punpckldq xmm9,xmm7
398 movdqa xmm7,xmm12
399 movdqa xmm6,xmm12
400 pslld xmm8,5
401 pandn xmm7,xmm14
402 pand xmm6,xmm13
403 punpckldq xmm0,xmm9
404 movdqa xmm9,xmm11
405
406 movdqa XMMWORD[(144-128)+rax],xmm4
407 paddd xmm10,xmm4
408 movd xmm1,DWORD[((-20))+r8]
409 psrld xmm9,27
410 pxor xmm6,xmm7
411 movdqa xmm7,xmm12
412
413 por xmm8,xmm9
414 movd xmm9,DWORD[((-20))+r9]
415 pslld xmm7,30
416 paddd xmm10,xmm6
417
418 psrld xmm12,2
419 paddd xmm10,xmm8
420DB 102,15,56,0,197
421 movd xmm8,DWORD[((-20))+r10]
422 por xmm12,xmm7
423 movd xmm7,DWORD[((-20))+r11]
424 punpckldq xmm1,xmm8
425 movdqa xmm8,xmm10
426 paddd xmm14,xmm15
427 punpckldq xmm9,xmm7
428 movdqa xmm7,xmm11
429 movdqa xmm6,xmm11
430 pslld xmm8,5
431 pandn xmm7,xmm13
432 pand xmm6,xmm12
433 punpckldq xmm1,xmm9
434 movdqa xmm9,xmm10
435
436 movdqa XMMWORD[(160-128)+rax],xmm0
437 paddd xmm14,xmm0
438 movd xmm2,DWORD[((-16))+r8]
439 psrld xmm9,27
440 pxor xmm6,xmm7
441 movdqa xmm7,xmm11
442
443 por xmm8,xmm9
444 movd xmm9,DWORD[((-16))+r9]
445 pslld xmm7,30
446 paddd xmm14,xmm6
447
448 psrld xmm11,2
449 paddd xmm14,xmm8
450DB 102,15,56,0,205
451 movd xmm8,DWORD[((-16))+r10]
452 por xmm11,xmm7
453 movd xmm7,DWORD[((-16))+r11]
454 punpckldq xmm2,xmm8
455 movdqa xmm8,xmm14
456 paddd xmm13,xmm15
457 punpckldq xmm9,xmm7
458 movdqa xmm7,xmm10
459 movdqa xmm6,xmm10
460 pslld xmm8,5
461 pandn xmm7,xmm12
462 pand xmm6,xmm11
463 punpckldq xmm2,xmm9
464 movdqa xmm9,xmm14
465
466 movdqa XMMWORD[(176-128)+rax],xmm1
467 paddd xmm13,xmm1
468 movd xmm3,DWORD[((-12))+r8]
469 psrld xmm9,27
470 pxor xmm6,xmm7
471 movdqa xmm7,xmm10
472
473 por xmm8,xmm9
474 movd xmm9,DWORD[((-12))+r9]
475 pslld xmm7,30
476 paddd xmm13,xmm6
477
478 psrld xmm10,2
479 paddd xmm13,xmm8
480DB 102,15,56,0,213
481 movd xmm8,DWORD[((-12))+r10]
482 por xmm10,xmm7
483 movd xmm7,DWORD[((-12))+r11]
484 punpckldq xmm3,xmm8
485 movdqa xmm8,xmm13
486 paddd xmm12,xmm15
487 punpckldq xmm9,xmm7
488 movdqa xmm7,xmm14
489 movdqa xmm6,xmm14
490 pslld xmm8,5
491 pandn xmm7,xmm11
492 pand xmm6,xmm10
493 punpckldq xmm3,xmm9
494 movdqa xmm9,xmm13
495
496 movdqa XMMWORD[(192-128)+rax],xmm2
497 paddd xmm12,xmm2
498 movd xmm4,DWORD[((-8))+r8]
499 psrld xmm9,27
500 pxor xmm6,xmm7
501 movdqa xmm7,xmm14
502
503 por xmm8,xmm9
504 movd xmm9,DWORD[((-8))+r9]
505 pslld xmm7,30
506 paddd xmm12,xmm6
507
508 psrld xmm14,2
509 paddd xmm12,xmm8
510DB 102,15,56,0,221
511 movd xmm8,DWORD[((-8))+r10]
512 por xmm14,xmm7
513 movd xmm7,DWORD[((-8))+r11]
514 punpckldq xmm4,xmm8
515 movdqa xmm8,xmm12
516 paddd xmm11,xmm15
517 punpckldq xmm9,xmm7
518 movdqa xmm7,xmm13
519 movdqa xmm6,xmm13
520 pslld xmm8,5
521 pandn xmm7,xmm10
522 pand xmm6,xmm14
523 punpckldq xmm4,xmm9
524 movdqa xmm9,xmm12
525
526 movdqa XMMWORD[(208-128)+rax],xmm3
527 paddd xmm11,xmm3
528 movd xmm0,DWORD[((-4))+r8]
529 psrld xmm9,27
530 pxor xmm6,xmm7
531 movdqa xmm7,xmm13
532
533 por xmm8,xmm9
534 movd xmm9,DWORD[((-4))+r9]
535 pslld xmm7,30
536 paddd xmm11,xmm6
537
538 psrld xmm13,2
539 paddd xmm11,xmm8
540DB 102,15,56,0,229
541 movd xmm8,DWORD[((-4))+r10]
542 por xmm13,xmm7
543 movdqa xmm1,XMMWORD[((0-128))+rax]
544 movd xmm7,DWORD[((-4))+r11]
545 punpckldq xmm0,xmm8
546 movdqa xmm8,xmm11
547 paddd xmm10,xmm15
548 punpckldq xmm9,xmm7
549 movdqa xmm7,xmm12
550 movdqa xmm6,xmm12
551 pslld xmm8,5
552 prefetcht0 [63+r8]
553 pandn xmm7,xmm14
554 pand xmm6,xmm13
555 punpckldq xmm0,xmm9
556 movdqa xmm9,xmm11
557
558 movdqa XMMWORD[(224-128)+rax],xmm4
559 paddd xmm10,xmm4
560 psrld xmm9,27
561 pxor xmm6,xmm7
562 movdqa xmm7,xmm12
563 prefetcht0 [63+r9]
564
565 por xmm8,xmm9
566 pslld xmm7,30
567 paddd xmm10,xmm6
568 prefetcht0 [63+r10]
569
570 psrld xmm12,2
571 paddd xmm10,xmm8
572DB 102,15,56,0,197
573 prefetcht0 [63+r11]
574 por xmm12,xmm7
575 movdqa xmm2,XMMWORD[((16-128))+rax]
576 pxor xmm1,xmm3
577 movdqa xmm3,XMMWORD[((32-128))+rax]
578
579 movdqa xmm8,xmm10
580 pxor xmm1,XMMWORD[((128-128))+rax]
581 paddd xmm14,xmm15
582 movdqa xmm7,xmm11
583 pslld xmm8,5
584 pxor xmm1,xmm3
585 movdqa xmm6,xmm11
586 pandn xmm7,xmm13
587 movdqa xmm5,xmm1
588 pand xmm6,xmm12
589 movdqa xmm9,xmm10
590 psrld xmm5,31
591 paddd xmm1,xmm1
592
593 movdqa XMMWORD[(240-128)+rax],xmm0
594 paddd xmm14,xmm0
595 psrld xmm9,27
596 pxor xmm6,xmm7
597
598 movdqa xmm7,xmm11
599 por xmm8,xmm9
600 pslld xmm7,30
601 paddd xmm14,xmm6
602
603 psrld xmm11,2
604 paddd xmm14,xmm8
605 por xmm1,xmm5
606 por xmm11,xmm7
607 pxor xmm2,xmm4
608 movdqa xmm4,XMMWORD[((48-128))+rax]
609
610 movdqa xmm8,xmm14
611 pxor xmm2,XMMWORD[((144-128))+rax]
612 paddd xmm13,xmm15
613 movdqa xmm7,xmm10
614 pslld xmm8,5
615 pxor xmm2,xmm4
616 movdqa xmm6,xmm10
617 pandn xmm7,xmm12
618 movdqa xmm5,xmm2
619 pand xmm6,xmm11
620 movdqa xmm9,xmm14
621 psrld xmm5,31
622 paddd xmm2,xmm2
623
624 movdqa XMMWORD[(0-128)+rax],xmm1
625 paddd xmm13,xmm1
626 psrld xmm9,27
627 pxor xmm6,xmm7
628
629 movdqa xmm7,xmm10
630 por xmm8,xmm9
631 pslld xmm7,30
632 paddd xmm13,xmm6
633
634 psrld xmm10,2
635 paddd xmm13,xmm8
636 por xmm2,xmm5
637 por xmm10,xmm7
638 pxor xmm3,xmm0
639 movdqa xmm0,XMMWORD[((64-128))+rax]
640
641 movdqa xmm8,xmm13
642 pxor xmm3,XMMWORD[((160-128))+rax]
643 paddd xmm12,xmm15
644 movdqa xmm7,xmm14
645 pslld xmm8,5
646 pxor xmm3,xmm0
647 movdqa xmm6,xmm14
648 pandn xmm7,xmm11
649 movdqa xmm5,xmm3
650 pand xmm6,xmm10
651 movdqa xmm9,xmm13
652 psrld xmm5,31
653 paddd xmm3,xmm3
654
655 movdqa XMMWORD[(16-128)+rax],xmm2
656 paddd xmm12,xmm2
657 psrld xmm9,27
658 pxor xmm6,xmm7
659
660 movdqa xmm7,xmm14
661 por xmm8,xmm9
662 pslld xmm7,30
663 paddd xmm12,xmm6
664
665 psrld xmm14,2
666 paddd xmm12,xmm8
667 por xmm3,xmm5
668 por xmm14,xmm7
669 pxor xmm4,xmm1
670 movdqa xmm1,XMMWORD[((80-128))+rax]
671
672 movdqa xmm8,xmm12
673 pxor xmm4,XMMWORD[((176-128))+rax]
674 paddd xmm11,xmm15
675 movdqa xmm7,xmm13
676 pslld xmm8,5
677 pxor xmm4,xmm1
678 movdqa xmm6,xmm13
679 pandn xmm7,xmm10
680 movdqa xmm5,xmm4
681 pand xmm6,xmm14
682 movdqa xmm9,xmm12
683 psrld xmm5,31
684 paddd xmm4,xmm4
685
686 movdqa XMMWORD[(32-128)+rax],xmm3
687 paddd xmm11,xmm3
688 psrld xmm9,27
689 pxor xmm6,xmm7
690
691 movdqa xmm7,xmm13
692 por xmm8,xmm9
693 pslld xmm7,30
694 paddd xmm11,xmm6
695
696 psrld xmm13,2
697 paddd xmm11,xmm8
698 por xmm4,xmm5
699 por xmm13,xmm7
700 pxor xmm0,xmm2
701 movdqa xmm2,XMMWORD[((96-128))+rax]
702
703 movdqa xmm8,xmm11
704 pxor xmm0,XMMWORD[((192-128))+rax]
705 paddd xmm10,xmm15
706 movdqa xmm7,xmm12
707 pslld xmm8,5
708 pxor xmm0,xmm2
709 movdqa xmm6,xmm12
710 pandn xmm7,xmm14
711 movdqa xmm5,xmm0
712 pand xmm6,xmm13
713 movdqa xmm9,xmm11
714 psrld xmm5,31
715 paddd xmm0,xmm0
716
717 movdqa XMMWORD[(48-128)+rax],xmm4
718 paddd xmm10,xmm4
719 psrld xmm9,27
720 pxor xmm6,xmm7
721
722 movdqa xmm7,xmm12
723 por xmm8,xmm9
724 pslld xmm7,30
725 paddd xmm10,xmm6
726
727 psrld xmm12,2
728 paddd xmm10,xmm8
729 por xmm0,xmm5
730 por xmm12,xmm7
731 movdqa xmm15,XMMWORD[rbp]
732 pxor xmm1,xmm3
733 movdqa xmm3,XMMWORD[((112-128))+rax]
734
735 movdqa xmm8,xmm10
736 movdqa xmm6,xmm13
737 pxor xmm1,XMMWORD[((208-128))+rax]
738 paddd xmm14,xmm15
739 pslld xmm8,5
740 pxor xmm6,xmm11
741
742 movdqa xmm9,xmm10
743 movdqa XMMWORD[(64-128)+rax],xmm0
744 paddd xmm14,xmm0
745 pxor xmm1,xmm3
746 psrld xmm9,27
747 pxor xmm6,xmm12
748 movdqa xmm7,xmm11
749
750 pslld xmm7,30
751 movdqa xmm5,xmm1
752 por xmm8,xmm9
753 psrld xmm5,31
754 paddd xmm14,xmm6
755 paddd xmm1,xmm1
756
757 psrld xmm11,2
758 paddd xmm14,xmm8
759 por xmm1,xmm5
760 por xmm11,xmm7
761 pxor xmm2,xmm4
762 movdqa xmm4,XMMWORD[((128-128))+rax]
763
764 movdqa xmm8,xmm14
765 movdqa xmm6,xmm12
766 pxor xmm2,XMMWORD[((224-128))+rax]
767 paddd xmm13,xmm15
768 pslld xmm8,5
769 pxor xmm6,xmm10
770
771 movdqa xmm9,xmm14
772 movdqa XMMWORD[(80-128)+rax],xmm1
773 paddd xmm13,xmm1
774 pxor xmm2,xmm4
775 psrld xmm9,27
776 pxor xmm6,xmm11
777 movdqa xmm7,xmm10
778
779 pslld xmm7,30
780 movdqa xmm5,xmm2
781 por xmm8,xmm9
782 psrld xmm5,31
783 paddd xmm13,xmm6
784 paddd xmm2,xmm2
785
786 psrld xmm10,2
787 paddd xmm13,xmm8
788 por xmm2,xmm5
789 por xmm10,xmm7
790 pxor xmm3,xmm0
791 movdqa xmm0,XMMWORD[((144-128))+rax]
792
793 movdqa xmm8,xmm13
794 movdqa xmm6,xmm11
795 pxor xmm3,XMMWORD[((240-128))+rax]
796 paddd xmm12,xmm15
797 pslld xmm8,5
798 pxor xmm6,xmm14
799
800 movdqa xmm9,xmm13
801 movdqa XMMWORD[(96-128)+rax],xmm2
802 paddd xmm12,xmm2
803 pxor xmm3,xmm0
804 psrld xmm9,27
805 pxor xmm6,xmm10
806 movdqa xmm7,xmm14
807
808 pslld xmm7,30
809 movdqa xmm5,xmm3
810 por xmm8,xmm9
811 psrld xmm5,31
812 paddd xmm12,xmm6
813 paddd xmm3,xmm3
814
815 psrld xmm14,2
816 paddd xmm12,xmm8
817 por xmm3,xmm5
818 por xmm14,xmm7
819 pxor xmm4,xmm1
820 movdqa xmm1,XMMWORD[((160-128))+rax]
821
822 movdqa xmm8,xmm12
823 movdqa xmm6,xmm10
824 pxor xmm4,XMMWORD[((0-128))+rax]
825 paddd xmm11,xmm15
826 pslld xmm8,5
827 pxor xmm6,xmm13
828
829 movdqa xmm9,xmm12
830 movdqa XMMWORD[(112-128)+rax],xmm3
831 paddd xmm11,xmm3
832 pxor xmm4,xmm1
833 psrld xmm9,27
834 pxor xmm6,xmm14
835 movdqa xmm7,xmm13
836
837 pslld xmm7,30
838 movdqa xmm5,xmm4
839 por xmm8,xmm9
840 psrld xmm5,31
841 paddd xmm11,xmm6
842 paddd xmm4,xmm4
843
844 psrld xmm13,2
845 paddd xmm11,xmm8
846 por xmm4,xmm5
847 por xmm13,xmm7
848 pxor xmm0,xmm2
849 movdqa xmm2,XMMWORD[((176-128))+rax]
850
851 movdqa xmm8,xmm11
852 movdqa xmm6,xmm14
853 pxor xmm0,XMMWORD[((16-128))+rax]
854 paddd xmm10,xmm15
855 pslld xmm8,5
856 pxor xmm6,xmm12
857
858 movdqa xmm9,xmm11
859 movdqa XMMWORD[(128-128)+rax],xmm4
860 paddd xmm10,xmm4
861 pxor xmm0,xmm2
862 psrld xmm9,27
863 pxor xmm6,xmm13
864 movdqa xmm7,xmm12
865
866 pslld xmm7,30
867 movdqa xmm5,xmm0
868 por xmm8,xmm9
869 psrld xmm5,31
870 paddd xmm10,xmm6
871 paddd xmm0,xmm0
872
873 psrld xmm12,2
874 paddd xmm10,xmm8
875 por xmm0,xmm5
876 por xmm12,xmm7
877 pxor xmm1,xmm3
878 movdqa xmm3,XMMWORD[((192-128))+rax]
879
880 movdqa xmm8,xmm10
881 movdqa xmm6,xmm13
882 pxor xmm1,XMMWORD[((32-128))+rax]
883 paddd xmm14,xmm15
884 pslld xmm8,5
885 pxor xmm6,xmm11
886
887 movdqa xmm9,xmm10
888 movdqa XMMWORD[(144-128)+rax],xmm0
889 paddd xmm14,xmm0
890 pxor xmm1,xmm3
891 psrld xmm9,27
892 pxor xmm6,xmm12
893 movdqa xmm7,xmm11
894
895 pslld xmm7,30
896 movdqa xmm5,xmm1
897 por xmm8,xmm9
898 psrld xmm5,31
899 paddd xmm14,xmm6
900 paddd xmm1,xmm1
901
902 psrld xmm11,2
903 paddd xmm14,xmm8
904 por xmm1,xmm5
905 por xmm11,xmm7
906 pxor xmm2,xmm4
907 movdqa xmm4,XMMWORD[((208-128))+rax]
908
909 movdqa xmm8,xmm14
910 movdqa xmm6,xmm12
911 pxor xmm2,XMMWORD[((48-128))+rax]
912 paddd xmm13,xmm15
913 pslld xmm8,5
914 pxor xmm6,xmm10
915
916 movdqa xmm9,xmm14
917 movdqa XMMWORD[(160-128)+rax],xmm1
918 paddd xmm13,xmm1
919 pxor xmm2,xmm4
920 psrld xmm9,27
921 pxor xmm6,xmm11
922 movdqa xmm7,xmm10
923
924 pslld xmm7,30
925 movdqa xmm5,xmm2
926 por xmm8,xmm9
927 psrld xmm5,31
928 paddd xmm13,xmm6
929 paddd xmm2,xmm2
930
931 psrld xmm10,2
932 paddd xmm13,xmm8
933 por xmm2,xmm5
934 por xmm10,xmm7
935 pxor xmm3,xmm0
936 movdqa xmm0,XMMWORD[((224-128))+rax]
937
938 movdqa xmm8,xmm13
939 movdqa xmm6,xmm11
940 pxor xmm3,XMMWORD[((64-128))+rax]
941 paddd xmm12,xmm15
942 pslld xmm8,5
943 pxor xmm6,xmm14
944
945 movdqa xmm9,xmm13
946 movdqa XMMWORD[(176-128)+rax],xmm2
947 paddd xmm12,xmm2
948 pxor xmm3,xmm0
949 psrld xmm9,27
950 pxor xmm6,xmm10
951 movdqa xmm7,xmm14
952
953 pslld xmm7,30
954 movdqa xmm5,xmm3
955 por xmm8,xmm9
956 psrld xmm5,31
957 paddd xmm12,xmm6
958 paddd xmm3,xmm3
959
960 psrld xmm14,2
961 paddd xmm12,xmm8
962 por xmm3,xmm5
963 por xmm14,xmm7
964 pxor xmm4,xmm1
965 movdqa xmm1,XMMWORD[((240-128))+rax]
966
967 movdqa xmm8,xmm12
968 movdqa xmm6,xmm10
969 pxor xmm4,XMMWORD[((80-128))+rax]
970 paddd xmm11,xmm15
971 pslld xmm8,5
972 pxor xmm6,xmm13
973
974 movdqa xmm9,xmm12
975 movdqa XMMWORD[(192-128)+rax],xmm3
976 paddd xmm11,xmm3
977 pxor xmm4,xmm1
978 psrld xmm9,27
979 pxor xmm6,xmm14
980 movdqa xmm7,xmm13
981
982 pslld xmm7,30
983 movdqa xmm5,xmm4
984 por xmm8,xmm9
985 psrld xmm5,31
986 paddd xmm11,xmm6
987 paddd xmm4,xmm4
988
989 psrld xmm13,2
990 paddd xmm11,xmm8
991 por xmm4,xmm5
992 por xmm13,xmm7
993 pxor xmm0,xmm2
994 movdqa xmm2,XMMWORD[((0-128))+rax]
995
996 movdqa xmm8,xmm11
997 movdqa xmm6,xmm14
998 pxor xmm0,XMMWORD[((96-128))+rax]
999 paddd xmm10,xmm15
1000 pslld xmm8,5
1001 pxor xmm6,xmm12
1002
1003 movdqa xmm9,xmm11
1004 movdqa XMMWORD[(208-128)+rax],xmm4
1005 paddd xmm10,xmm4
1006 pxor xmm0,xmm2
1007 psrld xmm9,27
1008 pxor xmm6,xmm13
1009 movdqa xmm7,xmm12
1010
1011 pslld xmm7,30
1012 movdqa xmm5,xmm0
1013 por xmm8,xmm9
1014 psrld xmm5,31
1015 paddd xmm10,xmm6
1016 paddd xmm0,xmm0
1017
1018 psrld xmm12,2
1019 paddd xmm10,xmm8
1020 por xmm0,xmm5
1021 por xmm12,xmm7
1022 pxor xmm1,xmm3
1023 movdqa xmm3,XMMWORD[((16-128))+rax]
1024
1025 movdqa xmm8,xmm10
1026 movdqa xmm6,xmm13
1027 pxor xmm1,XMMWORD[((112-128))+rax]
1028 paddd xmm14,xmm15
1029 pslld xmm8,5
1030 pxor xmm6,xmm11
1031
1032 movdqa xmm9,xmm10
1033 movdqa XMMWORD[(224-128)+rax],xmm0
1034 paddd xmm14,xmm0
1035 pxor xmm1,xmm3
1036 psrld xmm9,27
1037 pxor xmm6,xmm12
1038 movdqa xmm7,xmm11
1039
1040 pslld xmm7,30
1041 movdqa xmm5,xmm1
1042 por xmm8,xmm9
1043 psrld xmm5,31
1044 paddd xmm14,xmm6
1045 paddd xmm1,xmm1
1046
1047 psrld xmm11,2
1048 paddd xmm14,xmm8
1049 por xmm1,xmm5
1050 por xmm11,xmm7
1051 pxor xmm2,xmm4
1052 movdqa xmm4,XMMWORD[((32-128))+rax]
1053
1054 movdqa xmm8,xmm14
1055 movdqa xmm6,xmm12
1056 pxor xmm2,XMMWORD[((128-128))+rax]
1057 paddd xmm13,xmm15
1058 pslld xmm8,5
1059 pxor xmm6,xmm10
1060
1061 movdqa xmm9,xmm14
1062 movdqa XMMWORD[(240-128)+rax],xmm1
1063 paddd xmm13,xmm1
1064 pxor xmm2,xmm4
1065 psrld xmm9,27
1066 pxor xmm6,xmm11
1067 movdqa xmm7,xmm10
1068
1069 pslld xmm7,30
1070 movdqa xmm5,xmm2
1071 por xmm8,xmm9
1072 psrld xmm5,31
1073 paddd xmm13,xmm6
1074 paddd xmm2,xmm2
1075
1076 psrld xmm10,2
1077 paddd xmm13,xmm8
1078 por xmm2,xmm5
1079 por xmm10,xmm7
1080 pxor xmm3,xmm0
1081 movdqa xmm0,XMMWORD[((48-128))+rax]
1082
1083 movdqa xmm8,xmm13
1084 movdqa xmm6,xmm11
1085 pxor xmm3,XMMWORD[((144-128))+rax]
1086 paddd xmm12,xmm15
1087 pslld xmm8,5
1088 pxor xmm6,xmm14
1089
1090 movdqa xmm9,xmm13
1091 movdqa XMMWORD[(0-128)+rax],xmm2
1092 paddd xmm12,xmm2
1093 pxor xmm3,xmm0
1094 psrld xmm9,27
1095 pxor xmm6,xmm10
1096 movdqa xmm7,xmm14
1097
1098 pslld xmm7,30
1099 movdqa xmm5,xmm3
1100 por xmm8,xmm9
1101 psrld xmm5,31
1102 paddd xmm12,xmm6
1103 paddd xmm3,xmm3
1104
1105 psrld xmm14,2
1106 paddd xmm12,xmm8
1107 por xmm3,xmm5
1108 por xmm14,xmm7
1109 pxor xmm4,xmm1
1110 movdqa xmm1,XMMWORD[((64-128))+rax]
1111
1112 movdqa xmm8,xmm12
1113 movdqa xmm6,xmm10
1114 pxor xmm4,XMMWORD[((160-128))+rax]
1115 paddd xmm11,xmm15
1116 pslld xmm8,5
1117 pxor xmm6,xmm13
1118
1119 movdqa xmm9,xmm12
1120 movdqa XMMWORD[(16-128)+rax],xmm3
1121 paddd xmm11,xmm3
1122 pxor xmm4,xmm1
1123 psrld xmm9,27
1124 pxor xmm6,xmm14
1125 movdqa xmm7,xmm13
1126
1127 pslld xmm7,30
1128 movdqa xmm5,xmm4
1129 por xmm8,xmm9
1130 psrld xmm5,31
1131 paddd xmm11,xmm6
1132 paddd xmm4,xmm4
1133
1134 psrld xmm13,2
1135 paddd xmm11,xmm8
1136 por xmm4,xmm5
1137 por xmm13,xmm7
1138 pxor xmm0,xmm2
1139 movdqa xmm2,XMMWORD[((80-128))+rax]
1140
1141 movdqa xmm8,xmm11
1142 movdqa xmm6,xmm14
1143 pxor xmm0,XMMWORD[((176-128))+rax]
1144 paddd xmm10,xmm15
1145 pslld xmm8,5
1146 pxor xmm6,xmm12
1147
1148 movdqa xmm9,xmm11
1149 movdqa XMMWORD[(32-128)+rax],xmm4
1150 paddd xmm10,xmm4
1151 pxor xmm0,xmm2
1152 psrld xmm9,27
1153 pxor xmm6,xmm13
1154 movdqa xmm7,xmm12
1155
1156 pslld xmm7,30
1157 movdqa xmm5,xmm0
1158 por xmm8,xmm9
1159 psrld xmm5,31
1160 paddd xmm10,xmm6
1161 paddd xmm0,xmm0
1162
1163 psrld xmm12,2
1164 paddd xmm10,xmm8
1165 por xmm0,xmm5
1166 por xmm12,xmm7
1167 pxor xmm1,xmm3
1168 movdqa xmm3,XMMWORD[((96-128))+rax]
1169
1170 movdqa xmm8,xmm10
1171 movdqa xmm6,xmm13
1172 pxor xmm1,XMMWORD[((192-128))+rax]
1173 paddd xmm14,xmm15
1174 pslld xmm8,5
1175 pxor xmm6,xmm11
1176
1177 movdqa xmm9,xmm10
1178 movdqa XMMWORD[(48-128)+rax],xmm0
1179 paddd xmm14,xmm0
1180 pxor xmm1,xmm3
1181 psrld xmm9,27
1182 pxor xmm6,xmm12
1183 movdqa xmm7,xmm11
1184
1185 pslld xmm7,30
1186 movdqa xmm5,xmm1
1187 por xmm8,xmm9
1188 psrld xmm5,31
1189 paddd xmm14,xmm6
1190 paddd xmm1,xmm1
1191
1192 psrld xmm11,2
1193 paddd xmm14,xmm8
1194 por xmm1,xmm5
1195 por xmm11,xmm7
1196 pxor xmm2,xmm4
1197 movdqa xmm4,XMMWORD[((112-128))+rax]
1198
1199 movdqa xmm8,xmm14
1200 movdqa xmm6,xmm12
1201 pxor xmm2,XMMWORD[((208-128))+rax]
1202 paddd xmm13,xmm15
1203 pslld xmm8,5
1204 pxor xmm6,xmm10
1205
1206 movdqa xmm9,xmm14
1207 movdqa XMMWORD[(64-128)+rax],xmm1
1208 paddd xmm13,xmm1
1209 pxor xmm2,xmm4
1210 psrld xmm9,27
1211 pxor xmm6,xmm11
1212 movdqa xmm7,xmm10
1213
1214 pslld xmm7,30
1215 movdqa xmm5,xmm2
1216 por xmm8,xmm9
1217 psrld xmm5,31
1218 paddd xmm13,xmm6
1219 paddd xmm2,xmm2
1220
1221 psrld xmm10,2
1222 paddd xmm13,xmm8
1223 por xmm2,xmm5
1224 por xmm10,xmm7
1225 pxor xmm3,xmm0
1226 movdqa xmm0,XMMWORD[((128-128))+rax]
1227
1228 movdqa xmm8,xmm13
1229 movdqa xmm6,xmm11
1230 pxor xmm3,XMMWORD[((224-128))+rax]
1231 paddd xmm12,xmm15
1232 pslld xmm8,5
1233 pxor xmm6,xmm14
1234
1235 movdqa xmm9,xmm13
1236 movdqa XMMWORD[(80-128)+rax],xmm2
1237 paddd xmm12,xmm2
1238 pxor xmm3,xmm0
1239 psrld xmm9,27
1240 pxor xmm6,xmm10
1241 movdqa xmm7,xmm14
1242
1243 pslld xmm7,30
1244 movdqa xmm5,xmm3
1245 por xmm8,xmm9
1246 psrld xmm5,31
1247 paddd xmm12,xmm6
1248 paddd xmm3,xmm3
1249
1250 psrld xmm14,2
1251 paddd xmm12,xmm8
1252 por xmm3,xmm5
1253 por xmm14,xmm7
1254 pxor xmm4,xmm1
1255 movdqa xmm1,XMMWORD[((144-128))+rax]
1256
1257 movdqa xmm8,xmm12
1258 movdqa xmm6,xmm10
1259 pxor xmm4,XMMWORD[((240-128))+rax]
1260 paddd xmm11,xmm15
1261 pslld xmm8,5
1262 pxor xmm6,xmm13
1263
1264 movdqa xmm9,xmm12
1265 movdqa XMMWORD[(96-128)+rax],xmm3
1266 paddd xmm11,xmm3
1267 pxor xmm4,xmm1
1268 psrld xmm9,27
1269 pxor xmm6,xmm14
1270 movdqa xmm7,xmm13
1271
1272 pslld xmm7,30
1273 movdqa xmm5,xmm4
1274 por xmm8,xmm9
1275 psrld xmm5,31
1276 paddd xmm11,xmm6
1277 paddd xmm4,xmm4
1278
1279 psrld xmm13,2
1280 paddd xmm11,xmm8
1281 por xmm4,xmm5
1282 por xmm13,xmm7
1283 pxor xmm0,xmm2
1284 movdqa xmm2,XMMWORD[((160-128))+rax]
1285
1286 movdqa xmm8,xmm11
1287 movdqa xmm6,xmm14
1288 pxor xmm0,XMMWORD[((0-128))+rax]
1289 paddd xmm10,xmm15
1290 pslld xmm8,5
1291 pxor xmm6,xmm12
1292
1293 movdqa xmm9,xmm11
1294 movdqa XMMWORD[(112-128)+rax],xmm4
1295 paddd xmm10,xmm4
1296 pxor xmm0,xmm2
1297 psrld xmm9,27
1298 pxor xmm6,xmm13
1299 movdqa xmm7,xmm12
1300
1301 pslld xmm7,30
1302 movdqa xmm5,xmm0
1303 por xmm8,xmm9
1304 psrld xmm5,31
1305 paddd xmm10,xmm6
1306 paddd xmm0,xmm0
1307
1308 psrld xmm12,2
1309 paddd xmm10,xmm8
1310 por xmm0,xmm5
1311 por xmm12,xmm7
1312 movdqa xmm15,XMMWORD[32+rbp]
1313 pxor xmm1,xmm3
1314 movdqa xmm3,XMMWORD[((176-128))+rax]
1315
1316 movdqa xmm8,xmm10
1317 movdqa xmm7,xmm13
1318 pxor xmm1,XMMWORD[((16-128))+rax]
1319 pxor xmm1,xmm3
1320 paddd xmm14,xmm15
1321 pslld xmm8,5
1322 movdqa xmm9,xmm10
1323 pand xmm7,xmm12
1324
1325 movdqa xmm6,xmm13
1326 movdqa xmm5,xmm1
1327 psrld xmm9,27
1328 paddd xmm14,xmm7
1329 pxor xmm6,xmm12
1330
1331 movdqa XMMWORD[(128-128)+rax],xmm0
1332 paddd xmm14,xmm0
1333 por xmm8,xmm9
1334 psrld xmm5,31
1335 pand xmm6,xmm11
1336 movdqa xmm7,xmm11
1337
1338 pslld xmm7,30
1339 paddd xmm1,xmm1
1340 paddd xmm14,xmm6
1341
1342 psrld xmm11,2
1343 paddd xmm14,xmm8
1344 por xmm1,xmm5
1345 por xmm11,xmm7
1346 pxor xmm2,xmm4
1347 movdqa xmm4,XMMWORD[((192-128))+rax]
1348
1349 movdqa xmm8,xmm14
1350 movdqa xmm7,xmm12
1351 pxor xmm2,XMMWORD[((32-128))+rax]
1352 pxor xmm2,xmm4
1353 paddd xmm13,xmm15
1354 pslld xmm8,5
1355 movdqa xmm9,xmm14
1356 pand xmm7,xmm11
1357
1358 movdqa xmm6,xmm12
1359 movdqa xmm5,xmm2
1360 psrld xmm9,27
1361 paddd xmm13,xmm7
1362 pxor xmm6,xmm11
1363
1364 movdqa XMMWORD[(144-128)+rax],xmm1
1365 paddd xmm13,xmm1
1366 por xmm8,xmm9
1367 psrld xmm5,31
1368 pand xmm6,xmm10
1369 movdqa xmm7,xmm10
1370
1371 pslld xmm7,30
1372 paddd xmm2,xmm2
1373 paddd xmm13,xmm6
1374
1375 psrld xmm10,2
1376 paddd xmm13,xmm8
1377 por xmm2,xmm5
1378 por xmm10,xmm7
1379 pxor xmm3,xmm0
1380 movdqa xmm0,XMMWORD[((208-128))+rax]
1381
1382 movdqa xmm8,xmm13
1383 movdqa xmm7,xmm11
1384 pxor xmm3,XMMWORD[((48-128))+rax]
1385 pxor xmm3,xmm0
1386 paddd xmm12,xmm15
1387 pslld xmm8,5
1388 movdqa xmm9,xmm13
1389 pand xmm7,xmm10
1390
1391 movdqa xmm6,xmm11
1392 movdqa xmm5,xmm3
1393 psrld xmm9,27
1394 paddd xmm12,xmm7
1395 pxor xmm6,xmm10
1396
1397 movdqa XMMWORD[(160-128)+rax],xmm2
1398 paddd xmm12,xmm2
1399 por xmm8,xmm9
1400 psrld xmm5,31
1401 pand xmm6,xmm14
1402 movdqa xmm7,xmm14
1403
1404 pslld xmm7,30
1405 paddd xmm3,xmm3
1406 paddd xmm12,xmm6
1407
1408 psrld xmm14,2
1409 paddd xmm12,xmm8
1410 por xmm3,xmm5
1411 por xmm14,xmm7
1412 pxor xmm4,xmm1
1413 movdqa xmm1,XMMWORD[((224-128))+rax]
1414
1415 movdqa xmm8,xmm12
1416 movdqa xmm7,xmm10
1417 pxor xmm4,XMMWORD[((64-128))+rax]
1418 pxor xmm4,xmm1
1419 paddd xmm11,xmm15
1420 pslld xmm8,5
1421 movdqa xmm9,xmm12
1422 pand xmm7,xmm14
1423
1424 movdqa xmm6,xmm10
1425 movdqa xmm5,xmm4
1426 psrld xmm9,27
1427 paddd xmm11,xmm7
1428 pxor xmm6,xmm14
1429
1430 movdqa XMMWORD[(176-128)+rax],xmm3
1431 paddd xmm11,xmm3
1432 por xmm8,xmm9
1433 psrld xmm5,31
1434 pand xmm6,xmm13
1435 movdqa xmm7,xmm13
1436
1437 pslld xmm7,30
1438 paddd xmm4,xmm4
1439 paddd xmm11,xmm6
1440
1441 psrld xmm13,2
1442 paddd xmm11,xmm8
1443 por xmm4,xmm5
1444 por xmm13,xmm7
1445 pxor xmm0,xmm2
1446 movdqa xmm2,XMMWORD[((240-128))+rax]
1447
1448 movdqa xmm8,xmm11
1449 movdqa xmm7,xmm14
1450 pxor xmm0,XMMWORD[((80-128))+rax]
1451 pxor xmm0,xmm2
1452 paddd xmm10,xmm15
1453 pslld xmm8,5
1454 movdqa xmm9,xmm11
1455 pand xmm7,xmm13
1456
1457 movdqa xmm6,xmm14
1458 movdqa xmm5,xmm0
1459 psrld xmm9,27
1460 paddd xmm10,xmm7
1461 pxor xmm6,xmm13
1462
1463 movdqa XMMWORD[(192-128)+rax],xmm4
1464 paddd xmm10,xmm4
1465 por xmm8,xmm9
1466 psrld xmm5,31
1467 pand xmm6,xmm12
1468 movdqa xmm7,xmm12
1469
1470 pslld xmm7,30
1471 paddd xmm0,xmm0
1472 paddd xmm10,xmm6
1473
1474 psrld xmm12,2
1475 paddd xmm10,xmm8
1476 por xmm0,xmm5
1477 por xmm12,xmm7
1478 pxor xmm1,xmm3
1479 movdqa xmm3,XMMWORD[((0-128))+rax]
1480
1481 movdqa xmm8,xmm10
1482 movdqa xmm7,xmm13
1483 pxor xmm1,XMMWORD[((96-128))+rax]
1484 pxor xmm1,xmm3
1485 paddd xmm14,xmm15
1486 pslld xmm8,5
1487 movdqa xmm9,xmm10
1488 pand xmm7,xmm12
1489
1490 movdqa xmm6,xmm13
1491 movdqa xmm5,xmm1
1492 psrld xmm9,27
1493 paddd xmm14,xmm7
1494 pxor xmm6,xmm12
1495
1496 movdqa XMMWORD[(208-128)+rax],xmm0
1497 paddd xmm14,xmm0
1498 por xmm8,xmm9
1499 psrld xmm5,31
1500 pand xmm6,xmm11
1501 movdqa xmm7,xmm11
1502
1503 pslld xmm7,30
1504 paddd xmm1,xmm1
1505 paddd xmm14,xmm6
1506
1507 psrld xmm11,2
1508 paddd xmm14,xmm8
1509 por xmm1,xmm5
1510 por xmm11,xmm7
1511 pxor xmm2,xmm4
1512 movdqa xmm4,XMMWORD[((16-128))+rax]
1513
1514 movdqa xmm8,xmm14
1515 movdqa xmm7,xmm12
1516 pxor xmm2,XMMWORD[((112-128))+rax]
1517 pxor xmm2,xmm4
1518 paddd xmm13,xmm15
1519 pslld xmm8,5
1520 movdqa xmm9,xmm14
1521 pand xmm7,xmm11
1522
1523 movdqa xmm6,xmm12
1524 movdqa xmm5,xmm2
1525 psrld xmm9,27
1526 paddd xmm13,xmm7
1527 pxor xmm6,xmm11
1528
1529 movdqa XMMWORD[(224-128)+rax],xmm1
1530 paddd xmm13,xmm1
1531 por xmm8,xmm9
1532 psrld xmm5,31
1533 pand xmm6,xmm10
1534 movdqa xmm7,xmm10
1535
1536 pslld xmm7,30
1537 paddd xmm2,xmm2
1538 paddd xmm13,xmm6
1539
1540 psrld xmm10,2
1541 paddd xmm13,xmm8
1542 por xmm2,xmm5
1543 por xmm10,xmm7
1544 pxor xmm3,xmm0
1545 movdqa xmm0,XMMWORD[((32-128))+rax]
1546
1547 movdqa xmm8,xmm13
1548 movdqa xmm7,xmm11
1549 pxor xmm3,XMMWORD[((128-128))+rax]
1550 pxor xmm3,xmm0
1551 paddd xmm12,xmm15
1552 pslld xmm8,5
1553 movdqa xmm9,xmm13
1554 pand xmm7,xmm10
1555
1556 movdqa xmm6,xmm11
1557 movdqa xmm5,xmm3
1558 psrld xmm9,27
1559 paddd xmm12,xmm7
1560 pxor xmm6,xmm10
1561
1562 movdqa XMMWORD[(240-128)+rax],xmm2
1563 paddd xmm12,xmm2
1564 por xmm8,xmm9
1565 psrld xmm5,31
1566 pand xmm6,xmm14
1567 movdqa xmm7,xmm14
1568
1569 pslld xmm7,30
1570 paddd xmm3,xmm3
1571 paddd xmm12,xmm6
1572
1573 psrld xmm14,2
1574 paddd xmm12,xmm8
1575 por xmm3,xmm5
1576 por xmm14,xmm7
1577 pxor xmm4,xmm1
1578 movdqa xmm1,XMMWORD[((48-128))+rax]
1579
1580 movdqa xmm8,xmm12
1581 movdqa xmm7,xmm10
1582 pxor xmm4,XMMWORD[((144-128))+rax]
1583 pxor xmm4,xmm1
1584 paddd xmm11,xmm15
1585 pslld xmm8,5
1586 movdqa xmm9,xmm12
1587 pand xmm7,xmm14
1588
1589 movdqa xmm6,xmm10
1590 movdqa xmm5,xmm4
1591 psrld xmm9,27
1592 paddd xmm11,xmm7
1593 pxor xmm6,xmm14
1594
1595 movdqa XMMWORD[(0-128)+rax],xmm3
1596 paddd xmm11,xmm3
1597 por xmm8,xmm9
1598 psrld xmm5,31
1599 pand xmm6,xmm13
1600 movdqa xmm7,xmm13
1601
1602 pslld xmm7,30
1603 paddd xmm4,xmm4
1604 paddd xmm11,xmm6
1605
1606 psrld xmm13,2
1607 paddd xmm11,xmm8
1608 por xmm4,xmm5
1609 por xmm13,xmm7
1610 pxor xmm0,xmm2
1611 movdqa xmm2,XMMWORD[((64-128))+rax]
1612
1613 movdqa xmm8,xmm11
1614 movdqa xmm7,xmm14
1615 pxor xmm0,XMMWORD[((160-128))+rax]
1616 pxor xmm0,xmm2
1617 paddd xmm10,xmm15
1618 pslld xmm8,5
1619 movdqa xmm9,xmm11
1620 pand xmm7,xmm13
1621
1622 movdqa xmm6,xmm14
1623 movdqa xmm5,xmm0
1624 psrld xmm9,27
1625 paddd xmm10,xmm7
1626 pxor xmm6,xmm13
1627
1628 movdqa XMMWORD[(16-128)+rax],xmm4
1629 paddd xmm10,xmm4
1630 por xmm8,xmm9
1631 psrld xmm5,31
1632 pand xmm6,xmm12
1633 movdqa xmm7,xmm12
1634
1635 pslld xmm7,30
1636 paddd xmm0,xmm0
1637 paddd xmm10,xmm6
1638
1639 psrld xmm12,2
1640 paddd xmm10,xmm8
1641 por xmm0,xmm5
1642 por xmm12,xmm7
1643 pxor xmm1,xmm3
1644 movdqa xmm3,XMMWORD[((80-128))+rax]
1645
1646 movdqa xmm8,xmm10
1647 movdqa xmm7,xmm13
1648 pxor xmm1,XMMWORD[((176-128))+rax]
1649 pxor xmm1,xmm3
1650 paddd xmm14,xmm15
1651 pslld xmm8,5
1652 movdqa xmm9,xmm10
1653 pand xmm7,xmm12
1654
1655 movdqa xmm6,xmm13
1656 movdqa xmm5,xmm1
1657 psrld xmm9,27
1658 paddd xmm14,xmm7
1659 pxor xmm6,xmm12
1660
1661 movdqa XMMWORD[(32-128)+rax],xmm0
1662 paddd xmm14,xmm0
1663 por xmm8,xmm9
1664 psrld xmm5,31
1665 pand xmm6,xmm11
1666 movdqa xmm7,xmm11
1667
1668 pslld xmm7,30
1669 paddd xmm1,xmm1
1670 paddd xmm14,xmm6
1671
1672 psrld xmm11,2
1673 paddd xmm14,xmm8
1674 por xmm1,xmm5
1675 por xmm11,xmm7
1676 pxor xmm2,xmm4
1677 movdqa xmm4,XMMWORD[((96-128))+rax]
1678
1679 movdqa xmm8,xmm14
1680 movdqa xmm7,xmm12
1681 pxor xmm2,XMMWORD[((192-128))+rax]
1682 pxor xmm2,xmm4
1683 paddd xmm13,xmm15
1684 pslld xmm8,5
1685 movdqa xmm9,xmm14
1686 pand xmm7,xmm11
1687
1688 movdqa xmm6,xmm12
1689 movdqa xmm5,xmm2
1690 psrld xmm9,27
1691 paddd xmm13,xmm7
1692 pxor xmm6,xmm11
1693
1694 movdqa XMMWORD[(48-128)+rax],xmm1
1695 paddd xmm13,xmm1
1696 por xmm8,xmm9
1697 psrld xmm5,31
1698 pand xmm6,xmm10
1699 movdqa xmm7,xmm10
1700
1701 pslld xmm7,30
1702 paddd xmm2,xmm2
1703 paddd xmm13,xmm6
1704
1705 psrld xmm10,2
1706 paddd xmm13,xmm8
1707 por xmm2,xmm5
1708 por xmm10,xmm7
1709 pxor xmm3,xmm0
1710 movdqa xmm0,XMMWORD[((112-128))+rax]
1711
1712 movdqa xmm8,xmm13
1713 movdqa xmm7,xmm11
1714 pxor xmm3,XMMWORD[((208-128))+rax]
1715 pxor xmm3,xmm0
1716 paddd xmm12,xmm15
1717 pslld xmm8,5
1718 movdqa xmm9,xmm13
1719 pand xmm7,xmm10
1720
1721 movdqa xmm6,xmm11
1722 movdqa xmm5,xmm3
1723 psrld xmm9,27
1724 paddd xmm12,xmm7
1725 pxor xmm6,xmm10
1726
1727 movdqa XMMWORD[(64-128)+rax],xmm2
1728 paddd xmm12,xmm2
1729 por xmm8,xmm9
1730 psrld xmm5,31
1731 pand xmm6,xmm14
1732 movdqa xmm7,xmm14
1733
1734 pslld xmm7,30
1735 paddd xmm3,xmm3
1736 paddd xmm12,xmm6
1737
1738 psrld xmm14,2
1739 paddd xmm12,xmm8
1740 por xmm3,xmm5
1741 por xmm14,xmm7
1742 pxor xmm4,xmm1
1743 movdqa xmm1,XMMWORD[((128-128))+rax]
1744
1745 movdqa xmm8,xmm12
1746 movdqa xmm7,xmm10
1747 pxor xmm4,XMMWORD[((224-128))+rax]
1748 pxor xmm4,xmm1
1749 paddd xmm11,xmm15
1750 pslld xmm8,5
1751 movdqa xmm9,xmm12
1752 pand xmm7,xmm14
1753
1754 movdqa xmm6,xmm10
1755 movdqa xmm5,xmm4
1756 psrld xmm9,27
1757 paddd xmm11,xmm7
1758 pxor xmm6,xmm14
1759
1760 movdqa XMMWORD[(80-128)+rax],xmm3
1761 paddd xmm11,xmm3
1762 por xmm8,xmm9
1763 psrld xmm5,31
1764 pand xmm6,xmm13
1765 movdqa xmm7,xmm13
1766
1767 pslld xmm7,30
1768 paddd xmm4,xmm4
1769 paddd xmm11,xmm6
1770
1771 psrld xmm13,2
1772 paddd xmm11,xmm8
1773 por xmm4,xmm5
1774 por xmm13,xmm7
1775 pxor xmm0,xmm2
1776 movdqa xmm2,XMMWORD[((144-128))+rax]
1777
1778 movdqa xmm8,xmm11
1779 movdqa xmm7,xmm14
1780 pxor xmm0,XMMWORD[((240-128))+rax]
1781 pxor xmm0,xmm2
1782 paddd xmm10,xmm15
1783 pslld xmm8,5
1784 movdqa xmm9,xmm11
1785 pand xmm7,xmm13
1786
1787 movdqa xmm6,xmm14
1788 movdqa xmm5,xmm0
1789 psrld xmm9,27
1790 paddd xmm10,xmm7
1791 pxor xmm6,xmm13
1792
1793 movdqa XMMWORD[(96-128)+rax],xmm4
1794 paddd xmm10,xmm4
1795 por xmm8,xmm9
1796 psrld xmm5,31
1797 pand xmm6,xmm12
1798 movdqa xmm7,xmm12
1799
1800 pslld xmm7,30
1801 paddd xmm0,xmm0
1802 paddd xmm10,xmm6
1803
1804 psrld xmm12,2
1805 paddd xmm10,xmm8
1806 por xmm0,xmm5
1807 por xmm12,xmm7
1808 pxor xmm1,xmm3
1809 movdqa xmm3,XMMWORD[((160-128))+rax]
1810
1811 movdqa xmm8,xmm10
1812 movdqa xmm7,xmm13
1813 pxor xmm1,XMMWORD[((0-128))+rax]
1814 pxor xmm1,xmm3
1815 paddd xmm14,xmm15
1816 pslld xmm8,5
1817 movdqa xmm9,xmm10
1818 pand xmm7,xmm12
1819
1820 movdqa xmm6,xmm13
1821 movdqa xmm5,xmm1
1822 psrld xmm9,27
1823 paddd xmm14,xmm7
1824 pxor xmm6,xmm12
1825
1826 movdqa XMMWORD[(112-128)+rax],xmm0
1827 paddd xmm14,xmm0
1828 por xmm8,xmm9
1829 psrld xmm5,31
1830 pand xmm6,xmm11
1831 movdqa xmm7,xmm11
1832
1833 pslld xmm7,30
1834 paddd xmm1,xmm1
1835 paddd xmm14,xmm6
1836
1837 psrld xmm11,2
1838 paddd xmm14,xmm8
1839 por xmm1,xmm5
1840 por xmm11,xmm7
1841 pxor xmm2,xmm4
1842 movdqa xmm4,XMMWORD[((176-128))+rax]
1843
1844 movdqa xmm8,xmm14
1845 movdqa xmm7,xmm12
1846 pxor xmm2,XMMWORD[((16-128))+rax]
1847 pxor xmm2,xmm4
1848 paddd xmm13,xmm15
1849 pslld xmm8,5
1850 movdqa xmm9,xmm14
1851 pand xmm7,xmm11
1852
1853 movdqa xmm6,xmm12
1854 movdqa xmm5,xmm2
1855 psrld xmm9,27
1856 paddd xmm13,xmm7
1857 pxor xmm6,xmm11
1858
1859 movdqa XMMWORD[(128-128)+rax],xmm1
1860 paddd xmm13,xmm1
1861 por xmm8,xmm9
1862 psrld xmm5,31
1863 pand xmm6,xmm10
1864 movdqa xmm7,xmm10
1865
1866 pslld xmm7,30
1867 paddd xmm2,xmm2
1868 paddd xmm13,xmm6
1869
1870 psrld xmm10,2
1871 paddd xmm13,xmm8
1872 por xmm2,xmm5
1873 por xmm10,xmm7
1874 pxor xmm3,xmm0
1875 movdqa xmm0,XMMWORD[((192-128))+rax]
1876
1877 movdqa xmm8,xmm13
1878 movdqa xmm7,xmm11
1879 pxor xmm3,XMMWORD[((32-128))+rax]
1880 pxor xmm3,xmm0
1881 paddd xmm12,xmm15
1882 pslld xmm8,5
1883 movdqa xmm9,xmm13
1884 pand xmm7,xmm10
1885
1886 movdqa xmm6,xmm11
1887 movdqa xmm5,xmm3
1888 psrld xmm9,27
1889 paddd xmm12,xmm7
1890 pxor xmm6,xmm10
1891
1892 movdqa XMMWORD[(144-128)+rax],xmm2
1893 paddd xmm12,xmm2
1894 por xmm8,xmm9
1895 psrld xmm5,31
1896 pand xmm6,xmm14
1897 movdqa xmm7,xmm14
1898
1899 pslld xmm7,30
1900 paddd xmm3,xmm3
1901 paddd xmm12,xmm6
1902
1903 psrld xmm14,2
1904 paddd xmm12,xmm8
1905 por xmm3,xmm5
1906 por xmm14,xmm7
1907 pxor xmm4,xmm1
1908 movdqa xmm1,XMMWORD[((208-128))+rax]
1909
1910 movdqa xmm8,xmm12
1911 movdqa xmm7,xmm10
1912 pxor xmm4,XMMWORD[((48-128))+rax]
1913 pxor xmm4,xmm1
1914 paddd xmm11,xmm15
1915 pslld xmm8,5
1916 movdqa xmm9,xmm12
1917 pand xmm7,xmm14
1918
1919 movdqa xmm6,xmm10
1920 movdqa xmm5,xmm4
1921 psrld xmm9,27
1922 paddd xmm11,xmm7
1923 pxor xmm6,xmm14
1924
1925 movdqa XMMWORD[(160-128)+rax],xmm3
1926 paddd xmm11,xmm3
1927 por xmm8,xmm9
1928 psrld xmm5,31
1929 pand xmm6,xmm13
1930 movdqa xmm7,xmm13
1931
1932 pslld xmm7,30
1933 paddd xmm4,xmm4
1934 paddd xmm11,xmm6
1935
1936 psrld xmm13,2
1937 paddd xmm11,xmm8
1938 por xmm4,xmm5
1939 por xmm13,xmm7
1940 pxor xmm0,xmm2
1941 movdqa xmm2,XMMWORD[((224-128))+rax]
1942
1943 movdqa xmm8,xmm11
1944 movdqa xmm7,xmm14
1945 pxor xmm0,XMMWORD[((64-128))+rax]
1946 pxor xmm0,xmm2
1947 paddd xmm10,xmm15
1948 pslld xmm8,5
1949 movdqa xmm9,xmm11
1950 pand xmm7,xmm13
1951
1952 movdqa xmm6,xmm14
1953 movdqa xmm5,xmm0
1954 psrld xmm9,27
1955 paddd xmm10,xmm7
1956 pxor xmm6,xmm13
1957
1958 movdqa XMMWORD[(176-128)+rax],xmm4
1959 paddd xmm10,xmm4
1960 por xmm8,xmm9
1961 psrld xmm5,31
1962 pand xmm6,xmm12
1963 movdqa xmm7,xmm12
1964
1965 pslld xmm7,30
1966 paddd xmm0,xmm0
1967 paddd xmm10,xmm6
1968
1969 psrld xmm12,2
1970 paddd xmm10,xmm8
1971 por xmm0,xmm5
1972 por xmm12,xmm7
1973 movdqa xmm15,XMMWORD[64+rbp]
1974 pxor xmm1,xmm3
1975 movdqa xmm3,XMMWORD[((240-128))+rax]
1976
1977 movdqa xmm8,xmm10
1978 movdqa xmm6,xmm13
1979 pxor xmm1,XMMWORD[((80-128))+rax]
1980 paddd xmm14,xmm15
1981 pslld xmm8,5
1982 pxor xmm6,xmm11
1983
1984 movdqa xmm9,xmm10
1985 movdqa XMMWORD[(192-128)+rax],xmm0
1986 paddd xmm14,xmm0
1987 pxor xmm1,xmm3
1988 psrld xmm9,27
1989 pxor xmm6,xmm12
1990 movdqa xmm7,xmm11
1991
1992 pslld xmm7,30
1993 movdqa xmm5,xmm1
1994 por xmm8,xmm9
1995 psrld xmm5,31
1996 paddd xmm14,xmm6
1997 paddd xmm1,xmm1
1998
1999 psrld xmm11,2
2000 paddd xmm14,xmm8
2001 por xmm1,xmm5
2002 por xmm11,xmm7
2003 pxor xmm2,xmm4
2004 movdqa xmm4,XMMWORD[((0-128))+rax]
2005
2006 movdqa xmm8,xmm14
2007 movdqa xmm6,xmm12
2008 pxor xmm2,XMMWORD[((96-128))+rax]
2009 paddd xmm13,xmm15
2010 pslld xmm8,5
2011 pxor xmm6,xmm10
2012
2013 movdqa xmm9,xmm14
2014 movdqa XMMWORD[(208-128)+rax],xmm1
2015 paddd xmm13,xmm1
2016 pxor xmm2,xmm4
2017 psrld xmm9,27
2018 pxor xmm6,xmm11
2019 movdqa xmm7,xmm10
2020
2021 pslld xmm7,30
2022 movdqa xmm5,xmm2
2023 por xmm8,xmm9
2024 psrld xmm5,31
2025 paddd xmm13,xmm6
2026 paddd xmm2,xmm2
2027
2028 psrld xmm10,2
2029 paddd xmm13,xmm8
2030 por xmm2,xmm5
2031 por xmm10,xmm7
2032 pxor xmm3,xmm0
2033 movdqa xmm0,XMMWORD[((16-128))+rax]
2034
2035 movdqa xmm8,xmm13
2036 movdqa xmm6,xmm11
2037 pxor xmm3,XMMWORD[((112-128))+rax]
2038 paddd xmm12,xmm15
2039 pslld xmm8,5
2040 pxor xmm6,xmm14
2041
2042 movdqa xmm9,xmm13
2043 movdqa XMMWORD[(224-128)+rax],xmm2
2044 paddd xmm12,xmm2
2045 pxor xmm3,xmm0
2046 psrld xmm9,27
2047 pxor xmm6,xmm10
2048 movdqa xmm7,xmm14
2049
2050 pslld xmm7,30
2051 movdqa xmm5,xmm3
2052 por xmm8,xmm9
2053 psrld xmm5,31
2054 paddd xmm12,xmm6
2055 paddd xmm3,xmm3
2056
2057 psrld xmm14,2
2058 paddd xmm12,xmm8
2059 por xmm3,xmm5
2060 por xmm14,xmm7
2061 pxor xmm4,xmm1
2062 movdqa xmm1,XMMWORD[((32-128))+rax]
2063
2064 movdqa xmm8,xmm12
2065 movdqa xmm6,xmm10
2066 pxor xmm4,XMMWORD[((128-128))+rax]
2067 paddd xmm11,xmm15
2068 pslld xmm8,5
2069 pxor xmm6,xmm13
2070
2071 movdqa xmm9,xmm12
2072 movdqa XMMWORD[(240-128)+rax],xmm3
2073 paddd xmm11,xmm3
2074 pxor xmm4,xmm1
2075 psrld xmm9,27
2076 pxor xmm6,xmm14
2077 movdqa xmm7,xmm13
2078
2079 pslld xmm7,30
2080 movdqa xmm5,xmm4
2081 por xmm8,xmm9
2082 psrld xmm5,31
2083 paddd xmm11,xmm6
2084 paddd xmm4,xmm4
2085
2086 psrld xmm13,2
2087 paddd xmm11,xmm8
2088 por xmm4,xmm5
2089 por xmm13,xmm7
2090 pxor xmm0,xmm2
2091 movdqa xmm2,XMMWORD[((48-128))+rax]
2092
2093 movdqa xmm8,xmm11
2094 movdqa xmm6,xmm14
2095 pxor xmm0,XMMWORD[((144-128))+rax]
2096 paddd xmm10,xmm15
2097 pslld xmm8,5
2098 pxor xmm6,xmm12
2099
2100 movdqa xmm9,xmm11
2101 movdqa XMMWORD[(0-128)+rax],xmm4
2102 paddd xmm10,xmm4
2103 pxor xmm0,xmm2
2104 psrld xmm9,27
2105 pxor xmm6,xmm13
2106 movdqa xmm7,xmm12
2107
2108 pslld xmm7,30
2109 movdqa xmm5,xmm0
2110 por xmm8,xmm9
2111 psrld xmm5,31
2112 paddd xmm10,xmm6
2113 paddd xmm0,xmm0
2114
2115 psrld xmm12,2
2116 paddd xmm10,xmm8
2117 por xmm0,xmm5
2118 por xmm12,xmm7
2119 pxor xmm1,xmm3
2120 movdqa xmm3,XMMWORD[((64-128))+rax]
2121
2122 movdqa xmm8,xmm10
2123 movdqa xmm6,xmm13
2124 pxor xmm1,XMMWORD[((160-128))+rax]
2125 paddd xmm14,xmm15
2126 pslld xmm8,5
2127 pxor xmm6,xmm11
2128
2129 movdqa xmm9,xmm10
2130 movdqa XMMWORD[(16-128)+rax],xmm0
2131 paddd xmm14,xmm0
2132 pxor xmm1,xmm3
2133 psrld xmm9,27
2134 pxor xmm6,xmm12
2135 movdqa xmm7,xmm11
2136
2137 pslld xmm7,30
2138 movdqa xmm5,xmm1
2139 por xmm8,xmm9
2140 psrld xmm5,31
2141 paddd xmm14,xmm6
2142 paddd xmm1,xmm1
2143
2144 psrld xmm11,2
2145 paddd xmm14,xmm8
2146 por xmm1,xmm5
2147 por xmm11,xmm7
2148 pxor xmm2,xmm4
2149 movdqa xmm4,XMMWORD[((80-128))+rax]
2150
2151 movdqa xmm8,xmm14
2152 movdqa xmm6,xmm12
2153 pxor xmm2,XMMWORD[((176-128))+rax]
2154 paddd xmm13,xmm15
2155 pslld xmm8,5
2156 pxor xmm6,xmm10
2157
2158 movdqa xmm9,xmm14
2159 movdqa XMMWORD[(32-128)+rax],xmm1
2160 paddd xmm13,xmm1
2161 pxor xmm2,xmm4
2162 psrld xmm9,27
2163 pxor xmm6,xmm11
2164 movdqa xmm7,xmm10
2165
2166 pslld xmm7,30
2167 movdqa xmm5,xmm2
2168 por xmm8,xmm9
2169 psrld xmm5,31
2170 paddd xmm13,xmm6
2171 paddd xmm2,xmm2
2172
2173 psrld xmm10,2
2174 paddd xmm13,xmm8
2175 por xmm2,xmm5
2176 por xmm10,xmm7
2177 pxor xmm3,xmm0
2178 movdqa xmm0,XMMWORD[((96-128))+rax]
2179
2180 movdqa xmm8,xmm13
2181 movdqa xmm6,xmm11
2182 pxor xmm3,XMMWORD[((192-128))+rax]
2183 paddd xmm12,xmm15
2184 pslld xmm8,5
2185 pxor xmm6,xmm14
2186
2187 movdqa xmm9,xmm13
2188 movdqa XMMWORD[(48-128)+rax],xmm2
2189 paddd xmm12,xmm2
2190 pxor xmm3,xmm0
2191 psrld xmm9,27
2192 pxor xmm6,xmm10
2193 movdqa xmm7,xmm14
2194
2195 pslld xmm7,30
2196 movdqa xmm5,xmm3
2197 por xmm8,xmm9
2198 psrld xmm5,31
2199 paddd xmm12,xmm6
2200 paddd xmm3,xmm3
2201
2202 psrld xmm14,2
2203 paddd xmm12,xmm8
2204 por xmm3,xmm5
2205 por xmm14,xmm7
2206 pxor xmm4,xmm1
2207 movdqa xmm1,XMMWORD[((112-128))+rax]
2208
2209 movdqa xmm8,xmm12
2210 movdqa xmm6,xmm10
2211 pxor xmm4,XMMWORD[((208-128))+rax]
2212 paddd xmm11,xmm15
2213 pslld xmm8,5
2214 pxor xmm6,xmm13
2215
2216 movdqa xmm9,xmm12
2217 movdqa XMMWORD[(64-128)+rax],xmm3
2218 paddd xmm11,xmm3
2219 pxor xmm4,xmm1
2220 psrld xmm9,27
2221 pxor xmm6,xmm14
2222 movdqa xmm7,xmm13
2223
2224 pslld xmm7,30
2225 movdqa xmm5,xmm4
2226 por xmm8,xmm9
2227 psrld xmm5,31
2228 paddd xmm11,xmm6
2229 paddd xmm4,xmm4
2230
2231 psrld xmm13,2
2232 paddd xmm11,xmm8
2233 por xmm4,xmm5
2234 por xmm13,xmm7
2235 pxor xmm0,xmm2
2236 movdqa xmm2,XMMWORD[((128-128))+rax]
2237
2238 movdqa xmm8,xmm11
2239 movdqa xmm6,xmm14
2240 pxor xmm0,XMMWORD[((224-128))+rax]
2241 paddd xmm10,xmm15
2242 pslld xmm8,5
2243 pxor xmm6,xmm12
2244
2245 movdqa xmm9,xmm11
2246 movdqa XMMWORD[(80-128)+rax],xmm4
2247 paddd xmm10,xmm4
2248 pxor xmm0,xmm2
2249 psrld xmm9,27
2250 pxor xmm6,xmm13
2251 movdqa xmm7,xmm12
2252
2253 pslld xmm7,30
2254 movdqa xmm5,xmm0
2255 por xmm8,xmm9
2256 psrld xmm5,31
2257 paddd xmm10,xmm6
2258 paddd xmm0,xmm0
2259
2260 psrld xmm12,2
2261 paddd xmm10,xmm8
2262 por xmm0,xmm5
2263 por xmm12,xmm7
2264 pxor xmm1,xmm3
2265 movdqa xmm3,XMMWORD[((144-128))+rax]
2266
2267 movdqa xmm8,xmm10
2268 movdqa xmm6,xmm13
2269 pxor xmm1,XMMWORD[((240-128))+rax]
2270 paddd xmm14,xmm15
2271 pslld xmm8,5
2272 pxor xmm6,xmm11
2273
2274 movdqa xmm9,xmm10
2275 movdqa XMMWORD[(96-128)+rax],xmm0
2276 paddd xmm14,xmm0
2277 pxor xmm1,xmm3
2278 psrld xmm9,27
2279 pxor xmm6,xmm12
2280 movdqa xmm7,xmm11
2281
2282 pslld xmm7,30
2283 movdqa xmm5,xmm1
2284 por xmm8,xmm9
2285 psrld xmm5,31
2286 paddd xmm14,xmm6
2287 paddd xmm1,xmm1
2288
2289 psrld xmm11,2
2290 paddd xmm14,xmm8
2291 por xmm1,xmm5
2292 por xmm11,xmm7
2293 pxor xmm2,xmm4
2294 movdqa xmm4,XMMWORD[((160-128))+rax]
2295
2296 movdqa xmm8,xmm14
2297 movdqa xmm6,xmm12
2298 pxor xmm2,XMMWORD[((0-128))+rax]
2299 paddd xmm13,xmm15
2300 pslld xmm8,5
2301 pxor xmm6,xmm10
2302
2303 movdqa xmm9,xmm14
2304 movdqa XMMWORD[(112-128)+rax],xmm1
2305 paddd xmm13,xmm1
2306 pxor xmm2,xmm4
2307 psrld xmm9,27
2308 pxor xmm6,xmm11
2309 movdqa xmm7,xmm10
2310
2311 pslld xmm7,30
2312 movdqa xmm5,xmm2
2313 por xmm8,xmm9
2314 psrld xmm5,31
2315 paddd xmm13,xmm6
2316 paddd xmm2,xmm2
2317
2318 psrld xmm10,2
2319 paddd xmm13,xmm8
2320 por xmm2,xmm5
2321 por xmm10,xmm7
2322 pxor xmm3,xmm0
2323 movdqa xmm0,XMMWORD[((176-128))+rax]
2324
2325 movdqa xmm8,xmm13
2326 movdqa xmm6,xmm11
2327 pxor xmm3,XMMWORD[((16-128))+rax]
2328 paddd xmm12,xmm15
2329 pslld xmm8,5
2330 pxor xmm6,xmm14
2331
2332 movdqa xmm9,xmm13
2333 paddd xmm12,xmm2
2334 pxor xmm3,xmm0
2335 psrld xmm9,27
2336 pxor xmm6,xmm10
2337 movdqa xmm7,xmm14
2338
2339 pslld xmm7,30
2340 movdqa xmm5,xmm3
2341 por xmm8,xmm9
2342 psrld xmm5,31
2343 paddd xmm12,xmm6
2344 paddd xmm3,xmm3
2345
2346 psrld xmm14,2
2347 paddd xmm12,xmm8
2348 por xmm3,xmm5
2349 por xmm14,xmm7
2350 pxor xmm4,xmm1
2351 movdqa xmm1,XMMWORD[((192-128))+rax]
2352
2353 movdqa xmm8,xmm12
2354 movdqa xmm6,xmm10
2355 pxor xmm4,XMMWORD[((32-128))+rax]
2356 paddd xmm11,xmm15
2357 pslld xmm8,5
2358 pxor xmm6,xmm13
2359
2360 movdqa xmm9,xmm12
2361 paddd xmm11,xmm3
2362 pxor xmm4,xmm1
2363 psrld xmm9,27
2364 pxor xmm6,xmm14
2365 movdqa xmm7,xmm13
2366
2367 pslld xmm7,30
2368 movdqa xmm5,xmm4
2369 por xmm8,xmm9
2370 psrld xmm5,31
2371 paddd xmm11,xmm6
2372 paddd xmm4,xmm4
2373
2374 psrld xmm13,2
2375 paddd xmm11,xmm8
2376 por xmm4,xmm5
2377 por xmm13,xmm7
2378 pxor xmm0,xmm2
2379 movdqa xmm2,XMMWORD[((208-128))+rax]
2380
2381 movdqa xmm8,xmm11
2382 movdqa xmm6,xmm14
2383 pxor xmm0,XMMWORD[((48-128))+rax]
2384 paddd xmm10,xmm15
2385 pslld xmm8,5
2386 pxor xmm6,xmm12
2387
2388 movdqa xmm9,xmm11
2389 paddd xmm10,xmm4
2390 pxor xmm0,xmm2
2391 psrld xmm9,27
2392 pxor xmm6,xmm13
2393 movdqa xmm7,xmm12
2394
2395 pslld xmm7,30
2396 movdqa xmm5,xmm0
2397 por xmm8,xmm9
2398 psrld xmm5,31
2399 paddd xmm10,xmm6
2400 paddd xmm0,xmm0
2401
2402 psrld xmm12,2
2403 paddd xmm10,xmm8
2404 por xmm0,xmm5
2405 por xmm12,xmm7
2406 pxor xmm1,xmm3
2407 movdqa xmm3,XMMWORD[((224-128))+rax]
2408
2409 movdqa xmm8,xmm10
2410 movdqa xmm6,xmm13
2411 pxor xmm1,XMMWORD[((64-128))+rax]
2412 paddd xmm14,xmm15
2413 pslld xmm8,5
2414 pxor xmm6,xmm11
2415
2416 movdqa xmm9,xmm10
2417 paddd xmm14,xmm0
2418 pxor xmm1,xmm3
2419 psrld xmm9,27
2420 pxor xmm6,xmm12
2421 movdqa xmm7,xmm11
2422
2423 pslld xmm7,30
2424 movdqa xmm5,xmm1
2425 por xmm8,xmm9
2426 psrld xmm5,31
2427 paddd xmm14,xmm6
2428 paddd xmm1,xmm1
2429
2430 psrld xmm11,2
2431 paddd xmm14,xmm8
2432 por xmm1,xmm5
2433 por xmm11,xmm7
2434 pxor xmm2,xmm4
2435 movdqa xmm4,XMMWORD[((240-128))+rax]
2436
2437 movdqa xmm8,xmm14
2438 movdqa xmm6,xmm12
2439 pxor xmm2,XMMWORD[((80-128))+rax]
2440 paddd xmm13,xmm15
2441 pslld xmm8,5
2442 pxor xmm6,xmm10
2443
2444 movdqa xmm9,xmm14
2445 paddd xmm13,xmm1
2446 pxor xmm2,xmm4
2447 psrld xmm9,27
2448 pxor xmm6,xmm11
2449 movdqa xmm7,xmm10
2450
2451 pslld xmm7,30
2452 movdqa xmm5,xmm2
2453 por xmm8,xmm9
2454 psrld xmm5,31
2455 paddd xmm13,xmm6
2456 paddd xmm2,xmm2
2457
2458 psrld xmm10,2
2459 paddd xmm13,xmm8
2460 por xmm2,xmm5
2461 por xmm10,xmm7
2462 pxor xmm3,xmm0
2463 movdqa xmm0,XMMWORD[((0-128))+rax]
2464
2465 movdqa xmm8,xmm13
2466 movdqa xmm6,xmm11
2467 pxor xmm3,XMMWORD[((96-128))+rax]
2468 paddd xmm12,xmm15
2469 pslld xmm8,5
2470 pxor xmm6,xmm14
2471
2472 movdqa xmm9,xmm13
2473 paddd xmm12,xmm2
2474 pxor xmm3,xmm0
2475 psrld xmm9,27
2476 pxor xmm6,xmm10
2477 movdqa xmm7,xmm14
2478
2479 pslld xmm7,30
2480 movdqa xmm5,xmm3
2481 por xmm8,xmm9
2482 psrld xmm5,31
2483 paddd xmm12,xmm6
2484 paddd xmm3,xmm3
2485
2486 psrld xmm14,2
2487 paddd xmm12,xmm8
2488 por xmm3,xmm5
2489 por xmm14,xmm7
2490 pxor xmm4,xmm1
2491 movdqa xmm1,XMMWORD[((16-128))+rax]
2492
2493 movdqa xmm8,xmm12
2494 movdqa xmm6,xmm10
2495 pxor xmm4,XMMWORD[((112-128))+rax]
2496 paddd xmm11,xmm15
2497 pslld xmm8,5
2498 pxor xmm6,xmm13
2499
2500 movdqa xmm9,xmm12
2501 paddd xmm11,xmm3
2502 pxor xmm4,xmm1
2503 psrld xmm9,27
2504 pxor xmm6,xmm14
2505 movdqa xmm7,xmm13
2506
2507 pslld xmm7,30
2508 movdqa xmm5,xmm4
2509 por xmm8,xmm9
2510 psrld xmm5,31
2511 paddd xmm11,xmm6
2512 paddd xmm4,xmm4
2513
2514 psrld xmm13,2
2515 paddd xmm11,xmm8
2516 por xmm4,xmm5
2517 por xmm13,xmm7
2518 movdqa xmm8,xmm11
2519 paddd xmm10,xmm15
2520 movdqa xmm6,xmm14
2521 pslld xmm8,5
2522 pxor xmm6,xmm12
2523
2524 movdqa xmm9,xmm11
2525 paddd xmm10,xmm4
2526 psrld xmm9,27
2527 movdqa xmm7,xmm12
2528 pxor xmm6,xmm13
2529
2530 pslld xmm7,30
2531 por xmm8,xmm9
2532 paddd xmm10,xmm6
2533
2534 psrld xmm12,2
2535 paddd xmm10,xmm8
2536 por xmm12,xmm7
2537 movdqa xmm0,XMMWORD[rbx]
2538 mov ecx,1
2539 cmp ecx,DWORD[rbx]
2540 pxor xmm8,xmm8
2541 cmovge r8,rbp
2542 cmp ecx,DWORD[4+rbx]
2543 movdqa xmm1,xmm0
2544 cmovge r9,rbp
2545 cmp ecx,DWORD[8+rbx]
2546 pcmpgtd xmm1,xmm8
2547 cmovge r10,rbp
2548 cmp ecx,DWORD[12+rbx]
2549 paddd xmm0,xmm1
2550 cmovge r11,rbp
2551
2552 movdqu xmm6,XMMWORD[rdi]
2553 pand xmm10,xmm1
2554 movdqu xmm7,XMMWORD[32+rdi]
2555 pand xmm11,xmm1
2556 paddd xmm10,xmm6
2557 movdqu xmm8,XMMWORD[64+rdi]
2558 pand xmm12,xmm1
2559 paddd xmm11,xmm7
2560 movdqu xmm9,XMMWORD[96+rdi]
2561 pand xmm13,xmm1
2562 paddd xmm12,xmm8
2563 movdqu xmm5,XMMWORD[128+rdi]
2564 pand xmm14,xmm1
2565 movdqu XMMWORD[rdi],xmm10
2566 paddd xmm13,xmm9
2567 movdqu XMMWORD[32+rdi],xmm11
2568 paddd xmm14,xmm5
2569 movdqu XMMWORD[64+rdi],xmm12
2570 movdqu XMMWORD[96+rdi],xmm13
2571 movdqu XMMWORD[128+rdi],xmm14
2572
2573 movdqa XMMWORD[rbx],xmm0
2574 movdqa xmm5,XMMWORD[96+rbp]
2575 movdqa xmm15,XMMWORD[((-32))+rbp]
2576 dec edx
2577 jnz NEAR $L$oop
2578
2579 mov edx,DWORD[280+rsp]
2580 lea rdi,[16+rdi]
2581 lea rsi,[64+rsi]
2582 dec edx
2583 jnz NEAR $L$oop_grande
2584
2585$L$done:
2586 mov rax,QWORD[272+rsp]
2587
2588 movaps xmm6,XMMWORD[((-184))+rax]
2589 movaps xmm7,XMMWORD[((-168))+rax]
2590 movaps xmm8,XMMWORD[((-152))+rax]
2591 movaps xmm9,XMMWORD[((-136))+rax]
2592 movaps xmm10,XMMWORD[((-120))+rax]
2593 movaps xmm11,XMMWORD[((-104))+rax]
2594 movaps xmm12,XMMWORD[((-88))+rax]
2595 movaps xmm13,XMMWORD[((-72))+rax]
2596 movaps xmm14,XMMWORD[((-56))+rax]
2597 movaps xmm15,XMMWORD[((-40))+rax]
2598 mov rbp,QWORD[((-16))+rax]
2599
2600 mov rbx,QWORD[((-8))+rax]
2601
2602 lea rsp,[rax]
2603
2604$L$epilogue:
2605 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2606 mov rsi,QWORD[16+rsp]
2607 DB 0F3h,0C3h ;repret
2608
2609$L$SEH_end_sha1_multi_block:
2610
2611ALIGN 32
2612sha1_multi_block_shaext:
2613 mov QWORD[8+rsp],rdi ;WIN64 prologue
2614 mov QWORD[16+rsp],rsi
2615 mov rax,rsp
2616$L$SEH_begin_sha1_multi_block_shaext:
2617 mov rdi,rcx
2618 mov rsi,rdx
2619 mov rdx,r8
2620
2621
2622
2623_shaext_shortcut:
2624 mov rax,rsp
2625
2626 push rbx
2627
2628 push rbp
2629
2630 lea rsp,[((-168))+rsp]
2631 movaps XMMWORD[rsp],xmm6
2632 movaps XMMWORD[16+rsp],xmm7
2633 movaps XMMWORD[32+rsp],xmm8
2634 movaps XMMWORD[48+rsp],xmm9
2635 movaps XMMWORD[(-120)+rax],xmm10
2636 movaps XMMWORD[(-104)+rax],xmm11
2637 movaps XMMWORD[(-88)+rax],xmm12
2638 movaps XMMWORD[(-72)+rax],xmm13
2639 movaps XMMWORD[(-56)+rax],xmm14
2640 movaps XMMWORD[(-40)+rax],xmm15
2641 sub rsp,288
2642 shl edx,1
2643 and rsp,-256
2644 lea rdi,[64+rdi]
2645 mov QWORD[272+rsp],rax
2646$L$body_shaext:
2647 lea rbx,[256+rsp]
2648 movdqa xmm3,XMMWORD[((K_XX_XX+128))]
2649
2650$L$oop_grande_shaext:
2651 mov DWORD[280+rsp],edx
2652 xor edx,edx
2653
2654 mov r8,QWORD[rsi]
2655
2656 mov ecx,DWORD[8+rsi]
2657 cmp ecx,edx
2658 cmovg edx,ecx
2659 test ecx,ecx
2660 mov DWORD[rbx],ecx
2661 cmovle r8,rsp
2662
2663 mov r9,QWORD[16+rsi]
2664
2665 mov ecx,DWORD[24+rsi]
2666 cmp ecx,edx
2667 cmovg edx,ecx
2668 test ecx,ecx
2669 mov DWORD[4+rbx],ecx
2670 cmovle r9,rsp
2671 test edx,edx
2672 jz NEAR $L$done_shaext
2673
2674 movq xmm0,QWORD[((0-64))+rdi]
2675 movq xmm4,QWORD[((32-64))+rdi]
2676 movq xmm5,QWORD[((64-64))+rdi]
2677 movq xmm6,QWORD[((96-64))+rdi]
2678 movq xmm7,QWORD[((128-64))+rdi]
2679
2680 punpckldq xmm0,xmm4
2681 punpckldq xmm5,xmm6
2682
2683 movdqa xmm8,xmm0
2684 punpcklqdq xmm0,xmm5
2685 punpckhqdq xmm8,xmm5
2686
2687 pshufd xmm1,xmm7,63
2688 pshufd xmm9,xmm7,127
2689 pshufd xmm0,xmm0,27
2690 pshufd xmm8,xmm8,27
2691 jmp NEAR $L$oop_shaext
2692
2693ALIGN 32
2694$L$oop_shaext:
2695 movdqu xmm4,XMMWORD[r8]
2696 movdqu xmm11,XMMWORD[r9]
2697 movdqu xmm5,XMMWORD[16+r8]
2698 movdqu xmm12,XMMWORD[16+r9]
2699 movdqu xmm6,XMMWORD[32+r8]
2700DB 102,15,56,0,227
2701 movdqu xmm13,XMMWORD[32+r9]
2702DB 102,68,15,56,0,219
2703 movdqu xmm7,XMMWORD[48+r8]
2704 lea r8,[64+r8]
2705DB 102,15,56,0,235
2706 movdqu xmm14,XMMWORD[48+r9]
2707 lea r9,[64+r9]
2708DB 102,68,15,56,0,227
2709
2710 movdqa XMMWORD[80+rsp],xmm1
2711 paddd xmm1,xmm4
2712 movdqa XMMWORD[112+rsp],xmm9
2713 paddd xmm9,xmm11
2714 movdqa XMMWORD[64+rsp],xmm0
2715 movdqa xmm2,xmm0
2716 movdqa XMMWORD[96+rsp],xmm8
2717 movdqa xmm10,xmm8
2718DB 15,58,204,193,0
2719DB 15,56,200,213
2720DB 69,15,58,204,193,0
2721DB 69,15,56,200,212
2722DB 102,15,56,0,243
2723 prefetcht0 [127+r8]
2724DB 15,56,201,229
2725DB 102,68,15,56,0,235
2726 prefetcht0 [127+r9]
2727DB 69,15,56,201,220
2728
2729DB 102,15,56,0,251
2730 movdqa xmm1,xmm0
2731DB 102,68,15,56,0,243
2732 movdqa xmm9,xmm8
2733DB 15,58,204,194,0
2734DB 15,56,200,206
2735DB 69,15,58,204,194,0
2736DB 69,15,56,200,205
2737 pxor xmm4,xmm6
2738DB 15,56,201,238
2739 pxor xmm11,xmm13
2740DB 69,15,56,201,229
2741 movdqa xmm2,xmm0
2742 movdqa xmm10,xmm8
2743DB 15,58,204,193,0
2744DB 15,56,200,215
2745DB 69,15,58,204,193,0
2746DB 69,15,56,200,214
2747DB 15,56,202,231
2748DB 69,15,56,202,222
2749 pxor xmm5,xmm7
2750DB 15,56,201,247
2751 pxor xmm12,xmm14
2752DB 69,15,56,201,238
2753 movdqa xmm1,xmm0
2754 movdqa xmm9,xmm8
2755DB 15,58,204,194,0
2756DB 15,56,200,204
2757DB 69,15,58,204,194,0
2758DB 69,15,56,200,203
2759DB 15,56,202,236
2760DB 69,15,56,202,227
2761 pxor xmm6,xmm4
2762DB 15,56,201,252
2763 pxor xmm13,xmm11
2764DB 69,15,56,201,243
2765 movdqa xmm2,xmm0
2766 movdqa xmm10,xmm8
2767DB 15,58,204,193,0
2768DB 15,56,200,213
2769DB 69,15,58,204,193,0
2770DB 69,15,56,200,212
2771DB 15,56,202,245
2772DB 69,15,56,202,236
2773 pxor xmm7,xmm5
2774DB 15,56,201,229
2775 pxor xmm14,xmm12
2776DB 69,15,56,201,220
2777 movdqa xmm1,xmm0
2778 movdqa xmm9,xmm8
2779DB 15,58,204,194,1
2780DB 15,56,200,206
2781DB 69,15,58,204,194,1
2782DB 69,15,56,200,205
2783DB 15,56,202,254
2784DB 69,15,56,202,245
2785 pxor xmm4,xmm6
2786DB 15,56,201,238
2787 pxor xmm11,xmm13
2788DB 69,15,56,201,229
2789 movdqa xmm2,xmm0
2790 movdqa xmm10,xmm8
2791DB 15,58,204,193,1
2792DB 15,56,200,215
2793DB 69,15,58,204,193,1
2794DB 69,15,56,200,214
2795DB 15,56,202,231
2796DB 69,15,56,202,222
2797 pxor xmm5,xmm7
2798DB 15,56,201,247
2799 pxor xmm12,xmm14
2800DB 69,15,56,201,238
2801 movdqa xmm1,xmm0
2802 movdqa xmm9,xmm8
2803DB 15,58,204,194,1
2804DB 15,56,200,204
2805DB 69,15,58,204,194,1
2806DB 69,15,56,200,203
2807DB 15,56,202,236
2808DB 69,15,56,202,227
2809 pxor xmm6,xmm4
2810DB 15,56,201,252
2811 pxor xmm13,xmm11
2812DB 69,15,56,201,243
2813 movdqa xmm2,xmm0
2814 movdqa xmm10,xmm8
2815DB 15,58,204,193,1
2816DB 15,56,200,213
2817DB 69,15,58,204,193,1
2818DB 69,15,56,200,212
2819DB 15,56,202,245
2820DB 69,15,56,202,236
2821 pxor xmm7,xmm5
2822DB 15,56,201,229
2823 pxor xmm14,xmm12
2824DB 69,15,56,201,220
2825 movdqa xmm1,xmm0
2826 movdqa xmm9,xmm8
2827DB 15,58,204,194,1
2828DB 15,56,200,206
2829DB 69,15,58,204,194,1
2830DB 69,15,56,200,205
2831DB 15,56,202,254
2832DB 69,15,56,202,245
2833 pxor xmm4,xmm6
2834DB 15,56,201,238
2835 pxor xmm11,xmm13
2836DB 69,15,56,201,229
2837 movdqa xmm2,xmm0
2838 movdqa xmm10,xmm8
2839DB 15,58,204,193,2
2840DB 15,56,200,215
2841DB 69,15,58,204,193,2
2842DB 69,15,56,200,214
2843DB 15,56,202,231
2844DB 69,15,56,202,222
2845 pxor xmm5,xmm7
2846DB 15,56,201,247
2847 pxor xmm12,xmm14
2848DB 69,15,56,201,238
2849 movdqa xmm1,xmm0
2850 movdqa xmm9,xmm8
2851DB 15,58,204,194,2
2852DB 15,56,200,204
2853DB 69,15,58,204,194,2
2854DB 69,15,56,200,203
2855DB 15,56,202,236
2856DB 69,15,56,202,227
2857 pxor xmm6,xmm4
2858DB 15,56,201,252
2859 pxor xmm13,xmm11
2860DB 69,15,56,201,243
2861 movdqa xmm2,xmm0
2862 movdqa xmm10,xmm8
2863DB 15,58,204,193,2
2864DB 15,56,200,213
2865DB 69,15,58,204,193,2
2866DB 69,15,56,200,212
2867DB 15,56,202,245
2868DB 69,15,56,202,236
2869 pxor xmm7,xmm5
2870DB 15,56,201,229
2871 pxor xmm14,xmm12
2872DB 69,15,56,201,220
2873 movdqa xmm1,xmm0
2874 movdqa xmm9,xmm8
2875DB 15,58,204,194,2
2876DB 15,56,200,206
2877DB 69,15,58,204,194,2
2878DB 69,15,56,200,205
2879DB 15,56,202,254
2880DB 69,15,56,202,245
2881 pxor xmm4,xmm6
2882DB 15,56,201,238
2883 pxor xmm11,xmm13
2884DB 69,15,56,201,229
2885 movdqa xmm2,xmm0
2886 movdqa xmm10,xmm8
2887DB 15,58,204,193,2
2888DB 15,56,200,215
2889DB 69,15,58,204,193,2
2890DB 69,15,56,200,214
2891DB 15,56,202,231
2892DB 69,15,56,202,222
2893 pxor xmm5,xmm7
2894DB 15,56,201,247
2895 pxor xmm12,xmm14
2896DB 69,15,56,201,238
2897 movdqa xmm1,xmm0
2898 movdqa xmm9,xmm8
2899DB 15,58,204,194,3
2900DB 15,56,200,204
2901DB 69,15,58,204,194,3
2902DB 69,15,56,200,203
2903DB 15,56,202,236
2904DB 69,15,56,202,227
2905 pxor xmm6,xmm4
2906DB 15,56,201,252
2907 pxor xmm13,xmm11
2908DB 69,15,56,201,243
2909 movdqa xmm2,xmm0
2910 movdqa xmm10,xmm8
2911DB 15,58,204,193,3
2912DB 15,56,200,213
2913DB 69,15,58,204,193,3
2914DB 69,15,56,200,212
2915DB 15,56,202,245
2916DB 69,15,56,202,236
2917 pxor xmm7,xmm5
2918 pxor xmm14,xmm12
2919
2920 mov ecx,1
2921 pxor xmm4,xmm4
2922 cmp ecx,DWORD[rbx]
2923 cmovge r8,rsp
2924
2925 movdqa xmm1,xmm0
2926 movdqa xmm9,xmm8
2927DB 15,58,204,194,3
2928DB 15,56,200,206
2929DB 69,15,58,204,194,3
2930DB 69,15,56,200,205
2931DB 15,56,202,254
2932DB 69,15,56,202,245
2933
2934 cmp ecx,DWORD[4+rbx]
2935 cmovge r9,rsp
2936 movq xmm6,QWORD[rbx]
2937
2938 movdqa xmm2,xmm0
2939 movdqa xmm10,xmm8
2940DB 15,58,204,193,3
2941DB 15,56,200,215
2942DB 69,15,58,204,193,3
2943DB 69,15,56,200,214
2944
2945 pshufd xmm11,xmm6,0x00
2946 pshufd xmm12,xmm6,0x55
2947 movdqa xmm7,xmm6
2948 pcmpgtd xmm11,xmm4
2949 pcmpgtd xmm12,xmm4
2950
2951 movdqa xmm1,xmm0
2952 movdqa xmm9,xmm8
2953DB 15,58,204,194,3
2954DB 15,56,200,204
2955DB 69,15,58,204,194,3
2956DB 68,15,56,200,204
2957
2958 pcmpgtd xmm7,xmm4
2959 pand xmm0,xmm11
2960 pand xmm1,xmm11
2961 pand xmm8,xmm12
2962 pand xmm9,xmm12
2963 paddd xmm6,xmm7
2964
2965 paddd xmm0,XMMWORD[64+rsp]
2966 paddd xmm1,XMMWORD[80+rsp]
2967 paddd xmm8,XMMWORD[96+rsp]
2968 paddd xmm9,XMMWORD[112+rsp]
2969
2970 movq QWORD[rbx],xmm6
2971 dec edx
2972 jnz NEAR $L$oop_shaext
2973
2974 mov edx,DWORD[280+rsp]
2975
2976 pshufd xmm0,xmm0,27
2977 pshufd xmm8,xmm8,27
2978
2979 movdqa xmm6,xmm0
2980 punpckldq xmm0,xmm8
2981 punpckhdq xmm6,xmm8
2982 punpckhdq xmm1,xmm9
2983 movq QWORD[(0-64)+rdi],xmm0
2984 psrldq xmm0,8
2985 movq QWORD[(64-64)+rdi],xmm6
2986 psrldq xmm6,8
2987 movq QWORD[(32-64)+rdi],xmm0
2988 psrldq xmm1,8
2989 movq QWORD[(96-64)+rdi],xmm6
2990 movq QWORD[(128-64)+rdi],xmm1
2991
2992 lea rdi,[8+rdi]
2993 lea rsi,[32+rsi]
2994 dec edx
2995 jnz NEAR $L$oop_grande_shaext
2996
2997$L$done_shaext:
2998
2999 movaps xmm6,XMMWORD[((-184))+rax]
3000 movaps xmm7,XMMWORD[((-168))+rax]
3001 movaps xmm8,XMMWORD[((-152))+rax]
3002 movaps xmm9,XMMWORD[((-136))+rax]
3003 movaps xmm10,XMMWORD[((-120))+rax]
3004 movaps xmm11,XMMWORD[((-104))+rax]
3005 movaps xmm12,XMMWORD[((-88))+rax]
3006 movaps xmm13,XMMWORD[((-72))+rax]
3007 movaps xmm14,XMMWORD[((-56))+rax]
3008 movaps xmm15,XMMWORD[((-40))+rax]
3009 mov rbp,QWORD[((-16))+rax]
3010
3011 mov rbx,QWORD[((-8))+rax]
3012
3013 lea rsp,[rax]
3014
3015$L$epilogue_shaext:
3016 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
3017 mov rsi,QWORD[16+rsp]
3018 DB 0F3h,0C3h ;repret
3019
3020$L$SEH_end_sha1_multi_block_shaext:
3021
3022ALIGN 32
3023sha1_multi_block_avx:
3024 mov QWORD[8+rsp],rdi ;WIN64 prologue
3025 mov QWORD[16+rsp],rsi
3026 mov rax,rsp
3027$L$SEH_begin_sha1_multi_block_avx:
3028 mov rdi,rcx
3029 mov rsi,rdx
3030 mov rdx,r8
3031
3032
3033
3034_avx_shortcut:
3035 shr rcx,32
3036 cmp edx,2
3037 jb NEAR $L$avx
3038 test ecx,32
3039 jnz NEAR _avx2_shortcut
3040 jmp NEAR $L$avx
3041ALIGN 32
3042$L$avx:
3043 mov rax,rsp
3044
3045 push rbx
3046
3047 push rbp
3048
3049 lea rsp,[((-168))+rsp]
3050 movaps XMMWORD[rsp],xmm6
3051 movaps XMMWORD[16+rsp],xmm7
3052 movaps XMMWORD[32+rsp],xmm8
3053 movaps XMMWORD[48+rsp],xmm9
3054 movaps XMMWORD[(-120)+rax],xmm10
3055 movaps XMMWORD[(-104)+rax],xmm11
3056 movaps XMMWORD[(-88)+rax],xmm12
3057 movaps XMMWORD[(-72)+rax],xmm13
3058 movaps XMMWORD[(-56)+rax],xmm14
3059 movaps XMMWORD[(-40)+rax],xmm15
3060 sub rsp,288
3061 and rsp,-256
3062 mov QWORD[272+rsp],rax
3063
3064$L$body_avx:
3065 lea rbp,[K_XX_XX]
3066 lea rbx,[256+rsp]
3067
3068 vzeroupper
3069$L$oop_grande_avx:
3070 mov DWORD[280+rsp],edx
3071 xor edx,edx
3072
3073 mov r8,QWORD[rsi]
3074
3075 mov ecx,DWORD[8+rsi]
3076 cmp ecx,edx
3077 cmovg edx,ecx
3078 test ecx,ecx
3079 mov DWORD[rbx],ecx
3080 cmovle r8,rbp
3081
3082 mov r9,QWORD[16+rsi]
3083
3084 mov ecx,DWORD[24+rsi]
3085 cmp ecx,edx
3086 cmovg edx,ecx
3087 test ecx,ecx
3088 mov DWORD[4+rbx],ecx
3089 cmovle r9,rbp
3090
3091 mov r10,QWORD[32+rsi]
3092
3093 mov ecx,DWORD[40+rsi]
3094 cmp ecx,edx
3095 cmovg edx,ecx
3096 test ecx,ecx
3097 mov DWORD[8+rbx],ecx
3098 cmovle r10,rbp
3099
3100 mov r11,QWORD[48+rsi]
3101
3102 mov ecx,DWORD[56+rsi]
3103 cmp ecx,edx
3104 cmovg edx,ecx
3105 test ecx,ecx
3106 mov DWORD[12+rbx],ecx
3107 cmovle r11,rbp
3108 test edx,edx
3109 jz NEAR $L$done_avx
3110
3111 vmovdqu xmm10,XMMWORD[rdi]
3112 lea rax,[128+rsp]
3113 vmovdqu xmm11,XMMWORD[32+rdi]
3114 vmovdqu xmm12,XMMWORD[64+rdi]
3115 vmovdqu xmm13,XMMWORD[96+rdi]
3116 vmovdqu xmm14,XMMWORD[128+rdi]
3117 vmovdqu xmm5,XMMWORD[96+rbp]
3118 jmp NEAR $L$oop_avx
3119
3120ALIGN 32
3121$L$oop_avx:
3122 vmovdqa xmm15,XMMWORD[((-32))+rbp]
3123 vmovd xmm0,DWORD[r8]
3124 lea r8,[64+r8]
3125 vmovd xmm2,DWORD[r9]
3126 lea r9,[64+r9]
3127 vpinsrd xmm0,xmm0,DWORD[r10],1
3128 lea r10,[64+r10]
3129 vpinsrd xmm2,xmm2,DWORD[r11],1
3130 lea r11,[64+r11]
3131 vmovd xmm1,DWORD[((-60))+r8]
3132 vpunpckldq xmm0,xmm0,xmm2
3133 vmovd xmm9,DWORD[((-60))+r9]
3134 vpshufb xmm0,xmm0,xmm5
3135 vpinsrd xmm1,xmm1,DWORD[((-60))+r10],1
3136 vpinsrd xmm9,xmm9,DWORD[((-60))+r11],1
3137 vpaddd xmm14,xmm14,xmm15
3138 vpslld xmm8,xmm10,5
3139 vpandn xmm7,xmm11,xmm13
3140 vpand xmm6,xmm11,xmm12
3141
3142 vmovdqa XMMWORD[(0-128)+rax],xmm0
3143 vpaddd xmm14,xmm14,xmm0
3144 vpunpckldq xmm1,xmm1,xmm9
3145 vpsrld xmm9,xmm10,27
3146 vpxor xmm6,xmm6,xmm7
3147 vmovd xmm2,DWORD[((-56))+r8]
3148
3149 vpslld xmm7,xmm11,30
3150 vpor xmm8,xmm8,xmm9
3151 vmovd xmm9,DWORD[((-56))+r9]
3152 vpaddd xmm14,xmm14,xmm6
3153
3154 vpsrld xmm11,xmm11,2
3155 vpaddd xmm14,xmm14,xmm8
3156 vpshufb xmm1,xmm1,xmm5
3157 vpor xmm11,xmm11,xmm7
3158 vpinsrd xmm2,xmm2,DWORD[((-56))+r10],1
3159 vpinsrd xmm9,xmm9,DWORD[((-56))+r11],1
3160 vpaddd xmm13,xmm13,xmm15
3161 vpslld xmm8,xmm14,5
3162 vpandn xmm7,xmm10,xmm12
3163 vpand xmm6,xmm10,xmm11
3164
3165 vmovdqa XMMWORD[(16-128)+rax],xmm1
3166 vpaddd xmm13,xmm13,xmm1
3167 vpunpckldq xmm2,xmm2,xmm9
3168 vpsrld xmm9,xmm14,27
3169 vpxor xmm6,xmm6,xmm7
3170 vmovd xmm3,DWORD[((-52))+r8]
3171
3172 vpslld xmm7,xmm10,30
3173 vpor xmm8,xmm8,xmm9
3174 vmovd xmm9,DWORD[((-52))+r9]
3175 vpaddd xmm13,xmm13,xmm6
3176
3177 vpsrld xmm10,xmm10,2
3178 vpaddd xmm13,xmm13,xmm8
3179 vpshufb xmm2,xmm2,xmm5
3180 vpor xmm10,xmm10,xmm7
3181 vpinsrd xmm3,xmm3,DWORD[((-52))+r10],1
3182 vpinsrd xmm9,xmm9,DWORD[((-52))+r11],1
3183 vpaddd xmm12,xmm12,xmm15
3184 vpslld xmm8,xmm13,5
3185 vpandn xmm7,xmm14,xmm11
3186 vpand xmm6,xmm14,xmm10
3187
3188 vmovdqa XMMWORD[(32-128)+rax],xmm2
3189 vpaddd xmm12,xmm12,xmm2
3190 vpunpckldq xmm3,xmm3,xmm9
3191 vpsrld xmm9,xmm13,27
3192 vpxor xmm6,xmm6,xmm7
3193 vmovd xmm4,DWORD[((-48))+r8]
3194
3195 vpslld xmm7,xmm14,30
3196 vpor xmm8,xmm8,xmm9
3197 vmovd xmm9,DWORD[((-48))+r9]
3198 vpaddd xmm12,xmm12,xmm6
3199
3200 vpsrld xmm14,xmm14,2
3201 vpaddd xmm12,xmm12,xmm8
3202 vpshufb xmm3,xmm3,xmm5
3203 vpor xmm14,xmm14,xmm7
3204 vpinsrd xmm4,xmm4,DWORD[((-48))+r10],1
3205 vpinsrd xmm9,xmm9,DWORD[((-48))+r11],1
3206 vpaddd xmm11,xmm11,xmm15
3207 vpslld xmm8,xmm12,5
3208 vpandn xmm7,xmm13,xmm10
3209 vpand xmm6,xmm13,xmm14
3210
3211 vmovdqa XMMWORD[(48-128)+rax],xmm3
3212 vpaddd xmm11,xmm11,xmm3
3213 vpunpckldq xmm4,xmm4,xmm9
3214 vpsrld xmm9,xmm12,27
3215 vpxor xmm6,xmm6,xmm7
3216 vmovd xmm0,DWORD[((-44))+r8]
3217
3218 vpslld xmm7,xmm13,30
3219 vpor xmm8,xmm8,xmm9
3220 vmovd xmm9,DWORD[((-44))+r9]
3221 vpaddd xmm11,xmm11,xmm6
3222
3223 vpsrld xmm13,xmm13,2
3224 vpaddd xmm11,xmm11,xmm8
3225 vpshufb xmm4,xmm4,xmm5
3226 vpor xmm13,xmm13,xmm7
3227 vpinsrd xmm0,xmm0,DWORD[((-44))+r10],1
3228 vpinsrd xmm9,xmm9,DWORD[((-44))+r11],1
3229 vpaddd xmm10,xmm10,xmm15
3230 vpslld xmm8,xmm11,5
3231 vpandn xmm7,xmm12,xmm14
3232 vpand xmm6,xmm12,xmm13
3233
3234 vmovdqa XMMWORD[(64-128)+rax],xmm4
3235 vpaddd xmm10,xmm10,xmm4
3236 vpunpckldq xmm0,xmm0,xmm9
3237 vpsrld xmm9,xmm11,27
3238 vpxor xmm6,xmm6,xmm7
3239 vmovd xmm1,DWORD[((-40))+r8]
3240
3241 vpslld xmm7,xmm12,30
3242 vpor xmm8,xmm8,xmm9
3243 vmovd xmm9,DWORD[((-40))+r9]
3244 vpaddd xmm10,xmm10,xmm6
3245
3246 vpsrld xmm12,xmm12,2
3247 vpaddd xmm10,xmm10,xmm8
3248 vpshufb xmm0,xmm0,xmm5
3249 vpor xmm12,xmm12,xmm7
3250 vpinsrd xmm1,xmm1,DWORD[((-40))+r10],1
3251 vpinsrd xmm9,xmm9,DWORD[((-40))+r11],1
3252 vpaddd xmm14,xmm14,xmm15
3253 vpslld xmm8,xmm10,5
3254 vpandn xmm7,xmm11,xmm13
3255 vpand xmm6,xmm11,xmm12
3256
3257 vmovdqa XMMWORD[(80-128)+rax],xmm0
3258 vpaddd xmm14,xmm14,xmm0
3259 vpunpckldq xmm1,xmm1,xmm9
3260 vpsrld xmm9,xmm10,27
3261 vpxor xmm6,xmm6,xmm7
3262 vmovd xmm2,DWORD[((-36))+r8]
3263
3264 vpslld xmm7,xmm11,30
3265 vpor xmm8,xmm8,xmm9
3266 vmovd xmm9,DWORD[((-36))+r9]
3267 vpaddd xmm14,xmm14,xmm6
3268
3269 vpsrld xmm11,xmm11,2
3270 vpaddd xmm14,xmm14,xmm8
3271 vpshufb xmm1,xmm1,xmm5
3272 vpor xmm11,xmm11,xmm7
3273 vpinsrd xmm2,xmm2,DWORD[((-36))+r10],1
3274 vpinsrd xmm9,xmm9,DWORD[((-36))+r11],1
3275 vpaddd xmm13,xmm13,xmm15
3276 vpslld xmm8,xmm14,5
3277 vpandn xmm7,xmm10,xmm12
3278 vpand xmm6,xmm10,xmm11
3279
3280 vmovdqa XMMWORD[(96-128)+rax],xmm1
3281 vpaddd xmm13,xmm13,xmm1
3282 vpunpckldq xmm2,xmm2,xmm9
3283 vpsrld xmm9,xmm14,27
3284 vpxor xmm6,xmm6,xmm7
3285 vmovd xmm3,DWORD[((-32))+r8]
3286
3287 vpslld xmm7,xmm10,30
3288 vpor xmm8,xmm8,xmm9
3289 vmovd xmm9,DWORD[((-32))+r9]
3290 vpaddd xmm13,xmm13,xmm6
3291
3292 vpsrld xmm10,xmm10,2
3293 vpaddd xmm13,xmm13,xmm8
3294 vpshufb xmm2,xmm2,xmm5
3295 vpor xmm10,xmm10,xmm7
3296 vpinsrd xmm3,xmm3,DWORD[((-32))+r10],1
3297 vpinsrd xmm9,xmm9,DWORD[((-32))+r11],1
3298 vpaddd xmm12,xmm12,xmm15
3299 vpslld xmm8,xmm13,5
3300 vpandn xmm7,xmm14,xmm11
3301 vpand xmm6,xmm14,xmm10
3302
3303 vmovdqa XMMWORD[(112-128)+rax],xmm2
3304 vpaddd xmm12,xmm12,xmm2
3305 vpunpckldq xmm3,xmm3,xmm9
3306 vpsrld xmm9,xmm13,27
3307 vpxor xmm6,xmm6,xmm7
3308 vmovd xmm4,DWORD[((-28))+r8]
3309
3310 vpslld xmm7,xmm14,30
3311 vpor xmm8,xmm8,xmm9
3312 vmovd xmm9,DWORD[((-28))+r9]
3313 vpaddd xmm12,xmm12,xmm6
3314
3315 vpsrld xmm14,xmm14,2
3316 vpaddd xmm12,xmm12,xmm8
3317 vpshufb xmm3,xmm3,xmm5
3318 vpor xmm14,xmm14,xmm7
3319 vpinsrd xmm4,xmm4,DWORD[((-28))+r10],1
3320 vpinsrd xmm9,xmm9,DWORD[((-28))+r11],1
3321 vpaddd xmm11,xmm11,xmm15
3322 vpslld xmm8,xmm12,5
3323 vpandn xmm7,xmm13,xmm10
3324 vpand xmm6,xmm13,xmm14
3325
3326 vmovdqa XMMWORD[(128-128)+rax],xmm3
3327 vpaddd xmm11,xmm11,xmm3
3328 vpunpckldq xmm4,xmm4,xmm9
3329 vpsrld xmm9,xmm12,27
3330 vpxor xmm6,xmm6,xmm7
3331 vmovd xmm0,DWORD[((-24))+r8]
3332
3333 vpslld xmm7,xmm13,30
3334 vpor xmm8,xmm8,xmm9
3335 vmovd xmm9,DWORD[((-24))+r9]
3336 vpaddd xmm11,xmm11,xmm6
3337
3338 vpsrld xmm13,xmm13,2
3339 vpaddd xmm11,xmm11,xmm8
3340 vpshufb xmm4,xmm4,xmm5
3341 vpor xmm13,xmm13,xmm7
3342 vpinsrd xmm0,xmm0,DWORD[((-24))+r10],1
3343 vpinsrd xmm9,xmm9,DWORD[((-24))+r11],1
3344 vpaddd xmm10,xmm10,xmm15
3345 vpslld xmm8,xmm11,5
3346 vpandn xmm7,xmm12,xmm14
3347 vpand xmm6,xmm12,xmm13
3348
3349 vmovdqa XMMWORD[(144-128)+rax],xmm4
3350 vpaddd xmm10,xmm10,xmm4
3351 vpunpckldq xmm0,xmm0,xmm9
3352 vpsrld xmm9,xmm11,27
3353 vpxor xmm6,xmm6,xmm7
3354 vmovd xmm1,DWORD[((-20))+r8]
3355
3356 vpslld xmm7,xmm12,30
3357 vpor xmm8,xmm8,xmm9
3358 vmovd xmm9,DWORD[((-20))+r9]
3359 vpaddd xmm10,xmm10,xmm6
3360
3361 vpsrld xmm12,xmm12,2
3362 vpaddd xmm10,xmm10,xmm8
3363 vpshufb xmm0,xmm0,xmm5
3364 vpor xmm12,xmm12,xmm7
3365 vpinsrd xmm1,xmm1,DWORD[((-20))+r10],1
3366 vpinsrd xmm9,xmm9,DWORD[((-20))+r11],1
3367 vpaddd xmm14,xmm14,xmm15
3368 vpslld xmm8,xmm10,5
3369 vpandn xmm7,xmm11,xmm13
3370 vpand xmm6,xmm11,xmm12
3371
3372 vmovdqa XMMWORD[(160-128)+rax],xmm0
3373 vpaddd xmm14,xmm14,xmm0
3374 vpunpckldq xmm1,xmm1,xmm9
3375 vpsrld xmm9,xmm10,27
3376 vpxor xmm6,xmm6,xmm7
3377 vmovd xmm2,DWORD[((-16))+r8]
3378
3379 vpslld xmm7,xmm11,30
3380 vpor xmm8,xmm8,xmm9
3381 vmovd xmm9,DWORD[((-16))+r9]
3382 vpaddd xmm14,xmm14,xmm6
3383
3384 vpsrld xmm11,xmm11,2
3385 vpaddd xmm14,xmm14,xmm8
3386 vpshufb xmm1,xmm1,xmm5
3387 vpor xmm11,xmm11,xmm7
3388 vpinsrd xmm2,xmm2,DWORD[((-16))+r10],1
3389 vpinsrd xmm9,xmm9,DWORD[((-16))+r11],1
3390 vpaddd xmm13,xmm13,xmm15
3391 vpslld xmm8,xmm14,5
3392 vpandn xmm7,xmm10,xmm12
3393 vpand xmm6,xmm10,xmm11
3394
3395 vmovdqa XMMWORD[(176-128)+rax],xmm1
3396 vpaddd xmm13,xmm13,xmm1
3397 vpunpckldq xmm2,xmm2,xmm9
3398 vpsrld xmm9,xmm14,27
3399 vpxor xmm6,xmm6,xmm7
3400 vmovd xmm3,DWORD[((-12))+r8]
3401
3402 vpslld xmm7,xmm10,30
3403 vpor xmm8,xmm8,xmm9
3404 vmovd xmm9,DWORD[((-12))+r9]
3405 vpaddd xmm13,xmm13,xmm6
3406
3407 vpsrld xmm10,xmm10,2
3408 vpaddd xmm13,xmm13,xmm8
3409 vpshufb xmm2,xmm2,xmm5
3410 vpor xmm10,xmm10,xmm7
3411 vpinsrd xmm3,xmm3,DWORD[((-12))+r10],1
3412 vpinsrd xmm9,xmm9,DWORD[((-12))+r11],1
3413 vpaddd xmm12,xmm12,xmm15
3414 vpslld xmm8,xmm13,5
3415 vpandn xmm7,xmm14,xmm11
3416 vpand xmm6,xmm14,xmm10
3417
3418 vmovdqa XMMWORD[(192-128)+rax],xmm2
3419 vpaddd xmm12,xmm12,xmm2
3420 vpunpckldq xmm3,xmm3,xmm9
3421 vpsrld xmm9,xmm13,27
3422 vpxor xmm6,xmm6,xmm7
3423 vmovd xmm4,DWORD[((-8))+r8]
3424
3425 vpslld xmm7,xmm14,30
3426 vpor xmm8,xmm8,xmm9
3427 vmovd xmm9,DWORD[((-8))+r9]
3428 vpaddd xmm12,xmm12,xmm6
3429
3430 vpsrld xmm14,xmm14,2
3431 vpaddd xmm12,xmm12,xmm8
3432 vpshufb xmm3,xmm3,xmm5
3433 vpor xmm14,xmm14,xmm7
3434 vpinsrd xmm4,xmm4,DWORD[((-8))+r10],1
3435 vpinsrd xmm9,xmm9,DWORD[((-8))+r11],1
3436 vpaddd xmm11,xmm11,xmm15
3437 vpslld xmm8,xmm12,5
3438 vpandn xmm7,xmm13,xmm10
3439 vpand xmm6,xmm13,xmm14
3440
3441 vmovdqa XMMWORD[(208-128)+rax],xmm3
3442 vpaddd xmm11,xmm11,xmm3
3443 vpunpckldq xmm4,xmm4,xmm9
3444 vpsrld xmm9,xmm12,27
3445 vpxor xmm6,xmm6,xmm7
3446 vmovd xmm0,DWORD[((-4))+r8]
3447
3448 vpslld xmm7,xmm13,30
3449 vpor xmm8,xmm8,xmm9
3450 vmovd xmm9,DWORD[((-4))+r9]
3451 vpaddd xmm11,xmm11,xmm6
3452
3453 vpsrld xmm13,xmm13,2
3454 vpaddd xmm11,xmm11,xmm8
3455 vpshufb xmm4,xmm4,xmm5
3456 vpor xmm13,xmm13,xmm7
3457 vmovdqa xmm1,XMMWORD[((0-128))+rax]
3458 vpinsrd xmm0,xmm0,DWORD[((-4))+r10],1
3459 vpinsrd xmm9,xmm9,DWORD[((-4))+r11],1
3460 vpaddd xmm10,xmm10,xmm15
3461 prefetcht0 [63+r8]
3462 vpslld xmm8,xmm11,5
3463 vpandn xmm7,xmm12,xmm14
3464 vpand xmm6,xmm12,xmm13
3465
3466 vmovdqa XMMWORD[(224-128)+rax],xmm4
3467 vpaddd xmm10,xmm10,xmm4
3468 vpunpckldq xmm0,xmm0,xmm9
3469 vpsrld xmm9,xmm11,27
3470 prefetcht0 [63+r9]
3471 vpxor xmm6,xmm6,xmm7
3472
3473 vpslld xmm7,xmm12,30
3474 vpor xmm8,xmm8,xmm9
3475 prefetcht0 [63+r10]
3476 vpaddd xmm10,xmm10,xmm6
3477
3478 vpsrld xmm12,xmm12,2
3479 vpaddd xmm10,xmm10,xmm8
3480 prefetcht0 [63+r11]
3481 vpshufb xmm0,xmm0,xmm5
3482 vpor xmm12,xmm12,xmm7
3483 vmovdqa xmm2,XMMWORD[((16-128))+rax]
3484 vpxor xmm1,xmm1,xmm3
3485 vmovdqa xmm3,XMMWORD[((32-128))+rax]
3486
3487 vpaddd xmm14,xmm14,xmm15
3488 vpslld xmm8,xmm10,5
3489 vpandn xmm7,xmm11,xmm13
3490
3491 vpand xmm6,xmm11,xmm12
3492
3493 vmovdqa XMMWORD[(240-128)+rax],xmm0
3494 vpaddd xmm14,xmm14,xmm0
3495 vpxor xmm1,xmm1,XMMWORD[((128-128))+rax]
3496 vpsrld xmm9,xmm10,27
3497 vpxor xmm6,xmm6,xmm7
3498 vpxor xmm1,xmm1,xmm3
3499
3500
3501 vpslld xmm7,xmm11,30
3502 vpor xmm8,xmm8,xmm9
3503 vpaddd xmm14,xmm14,xmm6
3504
3505 vpsrld xmm5,xmm1,31
3506 vpaddd xmm1,xmm1,xmm1
3507
3508 vpsrld xmm11,xmm11,2
3509
3510 vpaddd xmm14,xmm14,xmm8
3511 vpor xmm1,xmm1,xmm5
3512 vpor xmm11,xmm11,xmm7
3513 vpxor xmm2,xmm2,xmm4
3514 vmovdqa xmm4,XMMWORD[((48-128))+rax]
3515
3516 vpaddd xmm13,xmm13,xmm15
3517 vpslld xmm8,xmm14,5
3518 vpandn xmm7,xmm10,xmm12
3519
3520 vpand xmm6,xmm10,xmm11
3521
3522 vmovdqa XMMWORD[(0-128)+rax],xmm1
3523 vpaddd xmm13,xmm13,xmm1
3524 vpxor xmm2,xmm2,XMMWORD[((144-128))+rax]
3525 vpsrld xmm9,xmm14,27
3526 vpxor xmm6,xmm6,xmm7
3527 vpxor xmm2,xmm2,xmm4
3528
3529
3530 vpslld xmm7,xmm10,30
3531 vpor xmm8,xmm8,xmm9
3532 vpaddd xmm13,xmm13,xmm6
3533
3534 vpsrld xmm5,xmm2,31
3535 vpaddd xmm2,xmm2,xmm2
3536
3537 vpsrld xmm10,xmm10,2
3538
3539 vpaddd xmm13,xmm13,xmm8
3540 vpor xmm2,xmm2,xmm5
3541 vpor xmm10,xmm10,xmm7
3542 vpxor xmm3,xmm3,xmm0
3543 vmovdqa xmm0,XMMWORD[((64-128))+rax]
3544
3545 vpaddd xmm12,xmm12,xmm15
3546 vpslld xmm8,xmm13,5
3547 vpandn xmm7,xmm14,xmm11
3548
3549 vpand xmm6,xmm14,xmm10
3550
3551 vmovdqa XMMWORD[(16-128)+rax],xmm2
3552 vpaddd xmm12,xmm12,xmm2
3553 vpxor xmm3,xmm3,XMMWORD[((160-128))+rax]
3554 vpsrld xmm9,xmm13,27
3555 vpxor xmm6,xmm6,xmm7
3556 vpxor xmm3,xmm3,xmm0
3557
3558
3559 vpslld xmm7,xmm14,30
3560 vpor xmm8,xmm8,xmm9
3561 vpaddd xmm12,xmm12,xmm6
3562
3563 vpsrld xmm5,xmm3,31
3564 vpaddd xmm3,xmm3,xmm3
3565
3566 vpsrld xmm14,xmm14,2
3567
3568 vpaddd xmm12,xmm12,xmm8
3569 vpor xmm3,xmm3,xmm5
3570 vpor xmm14,xmm14,xmm7
3571 vpxor xmm4,xmm4,xmm1
3572 vmovdqa xmm1,XMMWORD[((80-128))+rax]
3573
3574 vpaddd xmm11,xmm11,xmm15
3575 vpslld xmm8,xmm12,5
3576 vpandn xmm7,xmm13,xmm10
3577
3578 vpand xmm6,xmm13,xmm14
3579
3580 vmovdqa XMMWORD[(32-128)+rax],xmm3
3581 vpaddd xmm11,xmm11,xmm3
3582 vpxor xmm4,xmm4,XMMWORD[((176-128))+rax]
3583 vpsrld xmm9,xmm12,27
3584 vpxor xmm6,xmm6,xmm7
3585 vpxor xmm4,xmm4,xmm1
3586
3587
3588 vpslld xmm7,xmm13,30
3589 vpor xmm8,xmm8,xmm9
3590 vpaddd xmm11,xmm11,xmm6
3591
3592 vpsrld xmm5,xmm4,31
3593 vpaddd xmm4,xmm4,xmm4
3594
3595 vpsrld xmm13,xmm13,2
3596
3597 vpaddd xmm11,xmm11,xmm8
3598 vpor xmm4,xmm4,xmm5
3599 vpor xmm13,xmm13,xmm7
3600 vpxor xmm0,xmm0,xmm2
3601 vmovdqa xmm2,XMMWORD[((96-128))+rax]
3602
3603 vpaddd xmm10,xmm10,xmm15
3604 vpslld xmm8,xmm11,5
3605 vpandn xmm7,xmm12,xmm14
3606
3607 vpand xmm6,xmm12,xmm13
3608
3609 vmovdqa XMMWORD[(48-128)+rax],xmm4
3610 vpaddd xmm10,xmm10,xmm4
3611 vpxor xmm0,xmm0,XMMWORD[((192-128))+rax]
3612 vpsrld xmm9,xmm11,27
3613 vpxor xmm6,xmm6,xmm7
3614 vpxor xmm0,xmm0,xmm2
3615
3616
3617 vpslld xmm7,xmm12,30
3618 vpor xmm8,xmm8,xmm9
3619 vpaddd xmm10,xmm10,xmm6
3620
3621 vpsrld xmm5,xmm0,31
3622 vpaddd xmm0,xmm0,xmm0
3623
3624 vpsrld xmm12,xmm12,2
3625
3626 vpaddd xmm10,xmm10,xmm8
3627 vpor xmm0,xmm0,xmm5
3628 vpor xmm12,xmm12,xmm7
3629 vmovdqa xmm15,XMMWORD[rbp]
3630 vpxor xmm1,xmm1,xmm3
3631 vmovdqa xmm3,XMMWORD[((112-128))+rax]
3632
3633 vpslld xmm8,xmm10,5
3634 vpaddd xmm14,xmm14,xmm15
3635 vpxor xmm6,xmm13,xmm11
3636 vmovdqa XMMWORD[(64-128)+rax],xmm0
3637 vpaddd xmm14,xmm14,xmm0
3638 vpxor xmm1,xmm1,XMMWORD[((208-128))+rax]
3639 vpsrld xmm9,xmm10,27
3640 vpxor xmm6,xmm6,xmm12
3641 vpxor xmm1,xmm1,xmm3
3642
3643 vpslld xmm7,xmm11,30
3644 vpor xmm8,xmm8,xmm9
3645 vpaddd xmm14,xmm14,xmm6
3646 vpsrld xmm5,xmm1,31
3647 vpaddd xmm1,xmm1,xmm1
3648
3649 vpsrld xmm11,xmm11,2
3650 vpaddd xmm14,xmm14,xmm8
3651 vpor xmm1,xmm1,xmm5
3652 vpor xmm11,xmm11,xmm7
3653 vpxor xmm2,xmm2,xmm4
3654 vmovdqa xmm4,XMMWORD[((128-128))+rax]
3655
3656 vpslld xmm8,xmm14,5
3657 vpaddd xmm13,xmm13,xmm15
3658 vpxor xmm6,xmm12,xmm10
3659 vmovdqa XMMWORD[(80-128)+rax],xmm1
3660 vpaddd xmm13,xmm13,xmm1
3661 vpxor xmm2,xmm2,XMMWORD[((224-128))+rax]
3662 vpsrld xmm9,xmm14,27
3663 vpxor xmm6,xmm6,xmm11
3664 vpxor xmm2,xmm2,xmm4
3665
3666 vpslld xmm7,xmm10,30
3667 vpor xmm8,xmm8,xmm9
3668 vpaddd xmm13,xmm13,xmm6
3669 vpsrld xmm5,xmm2,31
3670 vpaddd xmm2,xmm2,xmm2
3671
3672 vpsrld xmm10,xmm10,2
3673 vpaddd xmm13,xmm13,xmm8
3674 vpor xmm2,xmm2,xmm5
3675 vpor xmm10,xmm10,xmm7
3676 vpxor xmm3,xmm3,xmm0
3677 vmovdqa xmm0,XMMWORD[((144-128))+rax]
3678
3679 vpslld xmm8,xmm13,5
3680 vpaddd xmm12,xmm12,xmm15
3681 vpxor xmm6,xmm11,xmm14
3682 vmovdqa XMMWORD[(96-128)+rax],xmm2
3683 vpaddd xmm12,xmm12,xmm2
3684 vpxor xmm3,xmm3,XMMWORD[((240-128))+rax]
3685 vpsrld xmm9,xmm13,27
3686 vpxor xmm6,xmm6,xmm10
3687 vpxor xmm3,xmm3,xmm0
3688
3689 vpslld xmm7,xmm14,30
3690 vpor xmm8,xmm8,xmm9
3691 vpaddd xmm12,xmm12,xmm6
3692 vpsrld xmm5,xmm3,31
3693 vpaddd xmm3,xmm3,xmm3
3694
3695 vpsrld xmm14,xmm14,2
3696 vpaddd xmm12,xmm12,xmm8
3697 vpor xmm3,xmm3,xmm5
3698 vpor xmm14,xmm14,xmm7
3699 vpxor xmm4,xmm4,xmm1
3700 vmovdqa xmm1,XMMWORD[((160-128))+rax]
3701
3702 vpslld xmm8,xmm12,5
3703 vpaddd xmm11,xmm11,xmm15
3704 vpxor xmm6,xmm10,xmm13
3705 vmovdqa XMMWORD[(112-128)+rax],xmm3
3706 vpaddd xmm11,xmm11,xmm3
3707 vpxor xmm4,xmm4,XMMWORD[((0-128))+rax]
3708 vpsrld xmm9,xmm12,27
3709 vpxor xmm6,xmm6,xmm14
3710 vpxor xmm4,xmm4,xmm1
3711
3712 vpslld xmm7,xmm13,30
3713 vpor xmm8,xmm8,xmm9
3714 vpaddd xmm11,xmm11,xmm6
3715 vpsrld xmm5,xmm4,31
3716 vpaddd xmm4,xmm4,xmm4
3717
3718 vpsrld xmm13,xmm13,2
3719 vpaddd xmm11,xmm11,xmm8
3720 vpor xmm4,xmm4,xmm5
3721 vpor xmm13,xmm13,xmm7
3722 vpxor xmm0,xmm0,xmm2
3723 vmovdqa xmm2,XMMWORD[((176-128))+rax]
3724
3725 vpslld xmm8,xmm11,5
3726 vpaddd xmm10,xmm10,xmm15
3727 vpxor xmm6,xmm14,xmm12
3728 vmovdqa XMMWORD[(128-128)+rax],xmm4
3729 vpaddd xmm10,xmm10,xmm4
3730 vpxor xmm0,xmm0,XMMWORD[((16-128))+rax]
3731 vpsrld xmm9,xmm11,27
3732 vpxor xmm6,xmm6,xmm13
3733 vpxor xmm0,xmm0,xmm2
3734
3735 vpslld xmm7,xmm12,30
3736 vpor xmm8,xmm8,xmm9
3737 vpaddd xmm10,xmm10,xmm6
3738 vpsrld xmm5,xmm0,31
3739 vpaddd xmm0,xmm0,xmm0
3740
3741 vpsrld xmm12,xmm12,2
3742 vpaddd xmm10,xmm10,xmm8
3743 vpor xmm0,xmm0,xmm5
3744 vpor xmm12,xmm12,xmm7
3745 vpxor xmm1,xmm1,xmm3
3746 vmovdqa xmm3,XMMWORD[((192-128))+rax]
3747
3748 vpslld xmm8,xmm10,5
3749 vpaddd xmm14,xmm14,xmm15
3750 vpxor xmm6,xmm13,xmm11
3751 vmovdqa XMMWORD[(144-128)+rax],xmm0
3752 vpaddd xmm14,xmm14,xmm0
3753 vpxor xmm1,xmm1,XMMWORD[((32-128))+rax]
3754 vpsrld xmm9,xmm10,27
3755 vpxor xmm6,xmm6,xmm12
3756 vpxor xmm1,xmm1,xmm3
3757
3758 vpslld xmm7,xmm11,30
3759 vpor xmm8,xmm8,xmm9
3760 vpaddd xmm14,xmm14,xmm6
3761 vpsrld xmm5,xmm1,31
3762 vpaddd xmm1,xmm1,xmm1
3763
3764 vpsrld xmm11,xmm11,2
3765 vpaddd xmm14,xmm14,xmm8
3766 vpor xmm1,xmm1,xmm5
3767 vpor xmm11,xmm11,xmm7
3768 vpxor xmm2,xmm2,xmm4
3769 vmovdqa xmm4,XMMWORD[((208-128))+rax]
3770
3771 vpslld xmm8,xmm14,5
3772 vpaddd xmm13,xmm13,xmm15
3773 vpxor xmm6,xmm12,xmm10
3774 vmovdqa XMMWORD[(160-128)+rax],xmm1
3775 vpaddd xmm13,xmm13,xmm1
3776 vpxor xmm2,xmm2,XMMWORD[((48-128))+rax]
3777 vpsrld xmm9,xmm14,27
3778 vpxor xmm6,xmm6,xmm11
3779 vpxor xmm2,xmm2,xmm4
3780
3781 vpslld xmm7,xmm10,30
3782 vpor xmm8,xmm8,xmm9
3783 vpaddd xmm13,xmm13,xmm6
3784 vpsrld xmm5,xmm2,31
3785 vpaddd xmm2,xmm2,xmm2
3786
3787 vpsrld xmm10,xmm10,2
3788 vpaddd xmm13,xmm13,xmm8
3789 vpor xmm2,xmm2,xmm5
3790 vpor xmm10,xmm10,xmm7
3791 vpxor xmm3,xmm3,xmm0
3792 vmovdqa xmm0,XMMWORD[((224-128))+rax]
3793
3794 vpslld xmm8,xmm13,5
3795 vpaddd xmm12,xmm12,xmm15
3796 vpxor xmm6,xmm11,xmm14
3797 vmovdqa XMMWORD[(176-128)+rax],xmm2
3798 vpaddd xmm12,xmm12,xmm2
3799 vpxor xmm3,xmm3,XMMWORD[((64-128))+rax]
3800 vpsrld xmm9,xmm13,27
3801 vpxor xmm6,xmm6,xmm10
3802 vpxor xmm3,xmm3,xmm0
3803
3804 vpslld xmm7,xmm14,30
3805 vpor xmm8,xmm8,xmm9
3806 vpaddd xmm12,xmm12,xmm6
3807 vpsrld xmm5,xmm3,31
3808 vpaddd xmm3,xmm3,xmm3
3809
3810 vpsrld xmm14,xmm14,2
3811 vpaddd xmm12,xmm12,xmm8
3812 vpor xmm3,xmm3,xmm5
3813 vpor xmm14,xmm14,xmm7
3814 vpxor xmm4,xmm4,xmm1
3815 vmovdqa xmm1,XMMWORD[((240-128))+rax]
3816
3817 vpslld xmm8,xmm12,5
3818 vpaddd xmm11,xmm11,xmm15
3819 vpxor xmm6,xmm10,xmm13
3820 vmovdqa XMMWORD[(192-128)+rax],xmm3
3821 vpaddd xmm11,xmm11,xmm3
3822 vpxor xmm4,xmm4,XMMWORD[((80-128))+rax]
3823 vpsrld xmm9,xmm12,27
3824 vpxor xmm6,xmm6,xmm14
3825 vpxor xmm4,xmm4,xmm1
3826
3827 vpslld xmm7,xmm13,30
3828 vpor xmm8,xmm8,xmm9
3829 vpaddd xmm11,xmm11,xmm6
3830 vpsrld xmm5,xmm4,31
3831 vpaddd xmm4,xmm4,xmm4
3832
3833 vpsrld xmm13,xmm13,2
3834 vpaddd xmm11,xmm11,xmm8
3835 vpor xmm4,xmm4,xmm5
3836 vpor xmm13,xmm13,xmm7
3837 vpxor xmm0,xmm0,xmm2
3838 vmovdqa xmm2,XMMWORD[((0-128))+rax]
3839
3840 vpslld xmm8,xmm11,5
3841 vpaddd xmm10,xmm10,xmm15
3842 vpxor xmm6,xmm14,xmm12
3843 vmovdqa XMMWORD[(208-128)+rax],xmm4
3844 vpaddd xmm10,xmm10,xmm4
3845 vpxor xmm0,xmm0,XMMWORD[((96-128))+rax]
3846 vpsrld xmm9,xmm11,27
3847 vpxor xmm6,xmm6,xmm13
3848 vpxor xmm0,xmm0,xmm2
3849
3850 vpslld xmm7,xmm12,30
3851 vpor xmm8,xmm8,xmm9
3852 vpaddd xmm10,xmm10,xmm6
3853 vpsrld xmm5,xmm0,31
3854 vpaddd xmm0,xmm0,xmm0
3855
3856 vpsrld xmm12,xmm12,2
3857 vpaddd xmm10,xmm10,xmm8
3858 vpor xmm0,xmm0,xmm5
3859 vpor xmm12,xmm12,xmm7
3860 vpxor xmm1,xmm1,xmm3
3861 vmovdqa xmm3,XMMWORD[((16-128))+rax]
3862
3863 vpslld xmm8,xmm10,5
3864 vpaddd xmm14,xmm14,xmm15
3865 vpxor xmm6,xmm13,xmm11
3866 vmovdqa XMMWORD[(224-128)+rax],xmm0
3867 vpaddd xmm14,xmm14,xmm0
3868 vpxor xmm1,xmm1,XMMWORD[((112-128))+rax]
3869 vpsrld xmm9,xmm10,27
3870 vpxor xmm6,xmm6,xmm12
3871 vpxor xmm1,xmm1,xmm3
3872
3873 vpslld xmm7,xmm11,30
3874 vpor xmm8,xmm8,xmm9
3875 vpaddd xmm14,xmm14,xmm6
3876 vpsrld xmm5,xmm1,31
3877 vpaddd xmm1,xmm1,xmm1
3878
3879 vpsrld xmm11,xmm11,2
3880 vpaddd xmm14,xmm14,xmm8
3881 vpor xmm1,xmm1,xmm5
3882 vpor xmm11,xmm11,xmm7
3883 vpxor xmm2,xmm2,xmm4
3884 vmovdqa xmm4,XMMWORD[((32-128))+rax]
3885
3886 vpslld xmm8,xmm14,5
3887 vpaddd xmm13,xmm13,xmm15
3888 vpxor xmm6,xmm12,xmm10
3889 vmovdqa XMMWORD[(240-128)+rax],xmm1
3890 vpaddd xmm13,xmm13,xmm1
3891 vpxor xmm2,xmm2,XMMWORD[((128-128))+rax]
3892 vpsrld xmm9,xmm14,27
3893 vpxor xmm6,xmm6,xmm11
3894 vpxor xmm2,xmm2,xmm4
3895
3896 vpslld xmm7,xmm10,30
3897 vpor xmm8,xmm8,xmm9
3898 vpaddd xmm13,xmm13,xmm6
3899 vpsrld xmm5,xmm2,31
3900 vpaddd xmm2,xmm2,xmm2
3901
3902 vpsrld xmm10,xmm10,2
3903 vpaddd xmm13,xmm13,xmm8
3904 vpor xmm2,xmm2,xmm5
3905 vpor xmm10,xmm10,xmm7
3906 vpxor xmm3,xmm3,xmm0
3907 vmovdqa xmm0,XMMWORD[((48-128))+rax]
3908
3909 vpslld xmm8,xmm13,5
3910 vpaddd xmm12,xmm12,xmm15
3911 vpxor xmm6,xmm11,xmm14
3912 vmovdqa XMMWORD[(0-128)+rax],xmm2
3913 vpaddd xmm12,xmm12,xmm2
3914 vpxor xmm3,xmm3,XMMWORD[((144-128))+rax]
3915 vpsrld xmm9,xmm13,27
3916 vpxor xmm6,xmm6,xmm10
3917 vpxor xmm3,xmm3,xmm0
3918
3919 vpslld xmm7,xmm14,30
3920 vpor xmm8,xmm8,xmm9
3921 vpaddd xmm12,xmm12,xmm6
3922 vpsrld xmm5,xmm3,31
3923 vpaddd xmm3,xmm3,xmm3
3924
3925 vpsrld xmm14,xmm14,2
3926 vpaddd xmm12,xmm12,xmm8
3927 vpor xmm3,xmm3,xmm5
3928 vpor xmm14,xmm14,xmm7
3929 vpxor xmm4,xmm4,xmm1
3930 vmovdqa xmm1,XMMWORD[((64-128))+rax]
3931
3932 vpslld xmm8,xmm12,5
3933 vpaddd xmm11,xmm11,xmm15
3934 vpxor xmm6,xmm10,xmm13
3935 vmovdqa XMMWORD[(16-128)+rax],xmm3
3936 vpaddd xmm11,xmm11,xmm3
3937 vpxor xmm4,xmm4,XMMWORD[((160-128))+rax]
3938 vpsrld xmm9,xmm12,27
3939 vpxor xmm6,xmm6,xmm14
3940 vpxor xmm4,xmm4,xmm1
3941
3942 vpslld xmm7,xmm13,30
3943 vpor xmm8,xmm8,xmm9
3944 vpaddd xmm11,xmm11,xmm6
3945 vpsrld xmm5,xmm4,31
3946 vpaddd xmm4,xmm4,xmm4
3947
3948 vpsrld xmm13,xmm13,2
3949 vpaddd xmm11,xmm11,xmm8
3950 vpor xmm4,xmm4,xmm5
3951 vpor xmm13,xmm13,xmm7
3952 vpxor xmm0,xmm0,xmm2
3953 vmovdqa xmm2,XMMWORD[((80-128))+rax]
3954
3955 vpslld xmm8,xmm11,5
3956 vpaddd xmm10,xmm10,xmm15
3957 vpxor xmm6,xmm14,xmm12
3958 vmovdqa XMMWORD[(32-128)+rax],xmm4
3959 vpaddd xmm10,xmm10,xmm4
3960 vpxor xmm0,xmm0,XMMWORD[((176-128))+rax]
3961 vpsrld xmm9,xmm11,27
3962 vpxor xmm6,xmm6,xmm13
3963 vpxor xmm0,xmm0,xmm2
3964
3965 vpslld xmm7,xmm12,30
3966 vpor xmm8,xmm8,xmm9
3967 vpaddd xmm10,xmm10,xmm6
3968 vpsrld xmm5,xmm0,31
3969 vpaddd xmm0,xmm0,xmm0
3970
3971 vpsrld xmm12,xmm12,2
3972 vpaddd xmm10,xmm10,xmm8
3973 vpor xmm0,xmm0,xmm5
3974 vpor xmm12,xmm12,xmm7
3975 vpxor xmm1,xmm1,xmm3
3976 vmovdqa xmm3,XMMWORD[((96-128))+rax]
3977
3978 vpslld xmm8,xmm10,5
3979 vpaddd xmm14,xmm14,xmm15
3980 vpxor xmm6,xmm13,xmm11
3981 vmovdqa XMMWORD[(48-128)+rax],xmm0
3982 vpaddd xmm14,xmm14,xmm0
3983 vpxor xmm1,xmm1,XMMWORD[((192-128))+rax]
3984 vpsrld xmm9,xmm10,27
3985 vpxor xmm6,xmm6,xmm12
3986 vpxor xmm1,xmm1,xmm3
3987
3988 vpslld xmm7,xmm11,30
3989 vpor xmm8,xmm8,xmm9
3990 vpaddd xmm14,xmm14,xmm6
3991 vpsrld xmm5,xmm1,31
3992 vpaddd xmm1,xmm1,xmm1
3993
3994 vpsrld xmm11,xmm11,2
3995 vpaddd xmm14,xmm14,xmm8
3996 vpor xmm1,xmm1,xmm5
3997 vpor xmm11,xmm11,xmm7
3998 vpxor xmm2,xmm2,xmm4
3999 vmovdqa xmm4,XMMWORD[((112-128))+rax]
4000
4001 vpslld xmm8,xmm14,5
4002 vpaddd xmm13,xmm13,xmm15
4003 vpxor xmm6,xmm12,xmm10
4004 vmovdqa XMMWORD[(64-128)+rax],xmm1
4005 vpaddd xmm13,xmm13,xmm1
4006 vpxor xmm2,xmm2,XMMWORD[((208-128))+rax]
4007 vpsrld xmm9,xmm14,27
4008 vpxor xmm6,xmm6,xmm11
4009 vpxor xmm2,xmm2,xmm4
4010
4011 vpslld xmm7,xmm10,30
4012 vpor xmm8,xmm8,xmm9
4013 vpaddd xmm13,xmm13,xmm6
4014 vpsrld xmm5,xmm2,31
4015 vpaddd xmm2,xmm2,xmm2
4016
4017 vpsrld xmm10,xmm10,2
4018 vpaddd xmm13,xmm13,xmm8
4019 vpor xmm2,xmm2,xmm5
4020 vpor xmm10,xmm10,xmm7
4021 vpxor xmm3,xmm3,xmm0
4022 vmovdqa xmm0,XMMWORD[((128-128))+rax]
4023
4024 vpslld xmm8,xmm13,5
4025 vpaddd xmm12,xmm12,xmm15
4026 vpxor xmm6,xmm11,xmm14
4027 vmovdqa XMMWORD[(80-128)+rax],xmm2
4028 vpaddd xmm12,xmm12,xmm2
4029 vpxor xmm3,xmm3,XMMWORD[((224-128))+rax]
4030 vpsrld xmm9,xmm13,27
4031 vpxor xmm6,xmm6,xmm10
4032 vpxor xmm3,xmm3,xmm0
4033
4034 vpslld xmm7,xmm14,30
4035 vpor xmm8,xmm8,xmm9
4036 vpaddd xmm12,xmm12,xmm6
4037 vpsrld xmm5,xmm3,31
4038 vpaddd xmm3,xmm3,xmm3
4039
4040 vpsrld xmm14,xmm14,2
4041 vpaddd xmm12,xmm12,xmm8
4042 vpor xmm3,xmm3,xmm5
4043 vpor xmm14,xmm14,xmm7
4044 vpxor xmm4,xmm4,xmm1
4045 vmovdqa xmm1,XMMWORD[((144-128))+rax]
4046
4047 vpslld xmm8,xmm12,5
4048 vpaddd xmm11,xmm11,xmm15
4049 vpxor xmm6,xmm10,xmm13
4050 vmovdqa XMMWORD[(96-128)+rax],xmm3
4051 vpaddd xmm11,xmm11,xmm3
4052 vpxor xmm4,xmm4,XMMWORD[((240-128))+rax]
4053 vpsrld xmm9,xmm12,27
4054 vpxor xmm6,xmm6,xmm14
4055 vpxor xmm4,xmm4,xmm1
4056
4057 vpslld xmm7,xmm13,30
4058 vpor xmm8,xmm8,xmm9
4059 vpaddd xmm11,xmm11,xmm6
4060 vpsrld xmm5,xmm4,31
4061 vpaddd xmm4,xmm4,xmm4
4062
4063 vpsrld xmm13,xmm13,2
4064 vpaddd xmm11,xmm11,xmm8
4065 vpor xmm4,xmm4,xmm5
4066 vpor xmm13,xmm13,xmm7
4067 vpxor xmm0,xmm0,xmm2
4068 vmovdqa xmm2,XMMWORD[((160-128))+rax]
4069
4070 vpslld xmm8,xmm11,5
4071 vpaddd xmm10,xmm10,xmm15
4072 vpxor xmm6,xmm14,xmm12
4073 vmovdqa XMMWORD[(112-128)+rax],xmm4
4074 vpaddd xmm10,xmm10,xmm4
4075 vpxor xmm0,xmm0,XMMWORD[((0-128))+rax]
4076 vpsrld xmm9,xmm11,27
4077 vpxor xmm6,xmm6,xmm13
4078 vpxor xmm0,xmm0,xmm2
4079
4080 vpslld xmm7,xmm12,30
4081 vpor xmm8,xmm8,xmm9
4082 vpaddd xmm10,xmm10,xmm6
4083 vpsrld xmm5,xmm0,31
4084 vpaddd xmm0,xmm0,xmm0
4085
4086 vpsrld xmm12,xmm12,2
4087 vpaddd xmm10,xmm10,xmm8
4088 vpor xmm0,xmm0,xmm5
4089 vpor xmm12,xmm12,xmm7
4090 vmovdqa xmm15,XMMWORD[32+rbp]
4091 vpxor xmm1,xmm1,xmm3
4092 vmovdqa xmm3,XMMWORD[((176-128))+rax]
4093
4094 vpaddd xmm14,xmm14,xmm15
4095 vpslld xmm8,xmm10,5
4096 vpand xmm7,xmm13,xmm12
4097 vpxor xmm1,xmm1,XMMWORD[((16-128))+rax]
4098
4099 vpaddd xmm14,xmm14,xmm7
4100 vpsrld xmm9,xmm10,27
4101 vpxor xmm6,xmm13,xmm12
4102 vpxor xmm1,xmm1,xmm3
4103
4104 vmovdqu XMMWORD[(128-128)+rax],xmm0
4105 vpaddd xmm14,xmm14,xmm0
4106 vpor xmm8,xmm8,xmm9
4107 vpsrld xmm5,xmm1,31
4108 vpand xmm6,xmm6,xmm11
4109 vpaddd xmm1,xmm1,xmm1
4110
4111 vpslld xmm7,xmm11,30
4112 vpaddd xmm14,xmm14,xmm6
4113
4114 vpsrld xmm11,xmm11,2
4115 vpaddd xmm14,xmm14,xmm8
4116 vpor xmm1,xmm1,xmm5
4117 vpor xmm11,xmm11,xmm7
4118 vpxor xmm2,xmm2,xmm4
4119 vmovdqa xmm4,XMMWORD[((192-128))+rax]
4120
4121 vpaddd xmm13,xmm13,xmm15
4122 vpslld xmm8,xmm14,5
4123 vpand xmm7,xmm12,xmm11
4124 vpxor xmm2,xmm2,XMMWORD[((32-128))+rax]
4125
4126 vpaddd xmm13,xmm13,xmm7
4127 vpsrld xmm9,xmm14,27
4128 vpxor xmm6,xmm12,xmm11
4129 vpxor xmm2,xmm2,xmm4
4130
4131 vmovdqu XMMWORD[(144-128)+rax],xmm1
4132 vpaddd xmm13,xmm13,xmm1
4133 vpor xmm8,xmm8,xmm9
4134 vpsrld xmm5,xmm2,31
4135 vpand xmm6,xmm6,xmm10
4136 vpaddd xmm2,xmm2,xmm2
4137
4138 vpslld xmm7,xmm10,30
4139 vpaddd xmm13,xmm13,xmm6
4140
4141 vpsrld xmm10,xmm10,2
4142 vpaddd xmm13,xmm13,xmm8
4143 vpor xmm2,xmm2,xmm5
4144 vpor xmm10,xmm10,xmm7
4145 vpxor xmm3,xmm3,xmm0
4146 vmovdqa xmm0,XMMWORD[((208-128))+rax]
4147
4148 vpaddd xmm12,xmm12,xmm15
4149 vpslld xmm8,xmm13,5
4150 vpand xmm7,xmm11,xmm10
4151 vpxor xmm3,xmm3,XMMWORD[((48-128))+rax]
4152
4153 vpaddd xmm12,xmm12,xmm7
4154 vpsrld xmm9,xmm13,27
4155 vpxor xmm6,xmm11,xmm10
4156 vpxor xmm3,xmm3,xmm0
4157
4158 vmovdqu XMMWORD[(160-128)+rax],xmm2
4159 vpaddd xmm12,xmm12,xmm2
4160 vpor xmm8,xmm8,xmm9
4161 vpsrld xmm5,xmm3,31
4162 vpand xmm6,xmm6,xmm14
4163 vpaddd xmm3,xmm3,xmm3
4164
4165 vpslld xmm7,xmm14,30
4166 vpaddd xmm12,xmm12,xmm6
4167
4168 vpsrld xmm14,xmm14,2
4169 vpaddd xmm12,xmm12,xmm8
4170 vpor xmm3,xmm3,xmm5
4171 vpor xmm14,xmm14,xmm7
4172 vpxor xmm4,xmm4,xmm1
4173 vmovdqa xmm1,XMMWORD[((224-128))+rax]
4174
4175 vpaddd xmm11,xmm11,xmm15
4176 vpslld xmm8,xmm12,5
4177 vpand xmm7,xmm10,xmm14
4178 vpxor xmm4,xmm4,XMMWORD[((64-128))+rax]
4179
4180 vpaddd xmm11,xmm11,xmm7
4181 vpsrld xmm9,xmm12,27
4182 vpxor xmm6,xmm10,xmm14
4183 vpxor xmm4,xmm4,xmm1
4184
4185 vmovdqu XMMWORD[(176-128)+rax],xmm3
4186 vpaddd xmm11,xmm11,xmm3
4187 vpor xmm8,xmm8,xmm9
4188 vpsrld xmm5,xmm4,31
4189 vpand xmm6,xmm6,xmm13
4190 vpaddd xmm4,xmm4,xmm4
4191
4192 vpslld xmm7,xmm13,30
4193 vpaddd xmm11,xmm11,xmm6
4194
4195 vpsrld xmm13,xmm13,2
4196 vpaddd xmm11,xmm11,xmm8
4197 vpor xmm4,xmm4,xmm5
4198 vpor xmm13,xmm13,xmm7
4199 vpxor xmm0,xmm0,xmm2
4200 vmovdqa xmm2,XMMWORD[((240-128))+rax]
4201
4202 vpaddd xmm10,xmm10,xmm15
4203 vpslld xmm8,xmm11,5
4204 vpand xmm7,xmm14,xmm13
4205 vpxor xmm0,xmm0,XMMWORD[((80-128))+rax]
4206
4207 vpaddd xmm10,xmm10,xmm7
4208 vpsrld xmm9,xmm11,27
4209 vpxor xmm6,xmm14,xmm13
4210 vpxor xmm0,xmm0,xmm2
4211
4212 vmovdqu XMMWORD[(192-128)+rax],xmm4
4213 vpaddd xmm10,xmm10,xmm4
4214 vpor xmm8,xmm8,xmm9
4215 vpsrld xmm5,xmm0,31
4216 vpand xmm6,xmm6,xmm12
4217 vpaddd xmm0,xmm0,xmm0
4218
4219 vpslld xmm7,xmm12,30
4220 vpaddd xmm10,xmm10,xmm6
4221
4222 vpsrld xmm12,xmm12,2
4223 vpaddd xmm10,xmm10,xmm8
4224 vpor xmm0,xmm0,xmm5
4225 vpor xmm12,xmm12,xmm7
4226 vpxor xmm1,xmm1,xmm3
4227 vmovdqa xmm3,XMMWORD[((0-128))+rax]
4228
4229 vpaddd xmm14,xmm14,xmm15
4230 vpslld xmm8,xmm10,5
4231 vpand xmm7,xmm13,xmm12
4232 vpxor xmm1,xmm1,XMMWORD[((96-128))+rax]
4233
4234 vpaddd xmm14,xmm14,xmm7
4235 vpsrld xmm9,xmm10,27
4236 vpxor xmm6,xmm13,xmm12
4237 vpxor xmm1,xmm1,xmm3
4238
4239 vmovdqu XMMWORD[(208-128)+rax],xmm0
4240 vpaddd xmm14,xmm14,xmm0
4241 vpor xmm8,xmm8,xmm9
4242 vpsrld xmm5,xmm1,31
4243 vpand xmm6,xmm6,xmm11
4244 vpaddd xmm1,xmm1,xmm1
4245
4246 vpslld xmm7,xmm11,30
4247 vpaddd xmm14,xmm14,xmm6
4248
4249 vpsrld xmm11,xmm11,2
4250 vpaddd xmm14,xmm14,xmm8
4251 vpor xmm1,xmm1,xmm5
4252 vpor xmm11,xmm11,xmm7
4253 vpxor xmm2,xmm2,xmm4
4254 vmovdqa xmm4,XMMWORD[((16-128))+rax]
4255
4256 vpaddd xmm13,xmm13,xmm15
4257 vpslld xmm8,xmm14,5
4258 vpand xmm7,xmm12,xmm11
4259 vpxor xmm2,xmm2,XMMWORD[((112-128))+rax]
4260
4261 vpaddd xmm13,xmm13,xmm7
4262 vpsrld xmm9,xmm14,27
4263 vpxor xmm6,xmm12,xmm11
4264 vpxor xmm2,xmm2,xmm4
4265
4266 vmovdqu XMMWORD[(224-128)+rax],xmm1
4267 vpaddd xmm13,xmm13,xmm1
4268 vpor xmm8,xmm8,xmm9
4269 vpsrld xmm5,xmm2,31
4270 vpand xmm6,xmm6,xmm10
4271 vpaddd xmm2,xmm2,xmm2
4272
4273 vpslld xmm7,xmm10,30
4274 vpaddd xmm13,xmm13,xmm6
4275
4276 vpsrld xmm10,xmm10,2
4277 vpaddd xmm13,xmm13,xmm8
4278 vpor xmm2,xmm2,xmm5
4279 vpor xmm10,xmm10,xmm7
4280 vpxor xmm3,xmm3,xmm0
4281 vmovdqa xmm0,XMMWORD[((32-128))+rax]
4282
4283 vpaddd xmm12,xmm12,xmm15
4284 vpslld xmm8,xmm13,5
4285 vpand xmm7,xmm11,xmm10
4286 vpxor xmm3,xmm3,XMMWORD[((128-128))+rax]
4287
4288 vpaddd xmm12,xmm12,xmm7
4289 vpsrld xmm9,xmm13,27
4290 vpxor xmm6,xmm11,xmm10
4291 vpxor xmm3,xmm3,xmm0
4292
4293 vmovdqu XMMWORD[(240-128)+rax],xmm2
4294 vpaddd xmm12,xmm12,xmm2
4295 vpor xmm8,xmm8,xmm9
4296 vpsrld xmm5,xmm3,31
4297 vpand xmm6,xmm6,xmm14
4298 vpaddd xmm3,xmm3,xmm3
4299
4300 vpslld xmm7,xmm14,30
4301 vpaddd xmm12,xmm12,xmm6
4302
4303 vpsrld xmm14,xmm14,2
4304 vpaddd xmm12,xmm12,xmm8
4305 vpor xmm3,xmm3,xmm5
4306 vpor xmm14,xmm14,xmm7
4307 vpxor xmm4,xmm4,xmm1
4308 vmovdqa xmm1,XMMWORD[((48-128))+rax]
4309
4310 vpaddd xmm11,xmm11,xmm15
4311 vpslld xmm8,xmm12,5
4312 vpand xmm7,xmm10,xmm14
4313 vpxor xmm4,xmm4,XMMWORD[((144-128))+rax]
4314
4315 vpaddd xmm11,xmm11,xmm7
4316 vpsrld xmm9,xmm12,27
4317 vpxor xmm6,xmm10,xmm14
4318 vpxor xmm4,xmm4,xmm1
4319
4320 vmovdqu XMMWORD[(0-128)+rax],xmm3
4321 vpaddd xmm11,xmm11,xmm3
4322 vpor xmm8,xmm8,xmm9
4323 vpsrld xmm5,xmm4,31
4324 vpand xmm6,xmm6,xmm13
4325 vpaddd xmm4,xmm4,xmm4
4326
4327 vpslld xmm7,xmm13,30
4328 vpaddd xmm11,xmm11,xmm6
4329
4330 vpsrld xmm13,xmm13,2
4331 vpaddd xmm11,xmm11,xmm8
4332 vpor xmm4,xmm4,xmm5
4333 vpor xmm13,xmm13,xmm7
4334 vpxor xmm0,xmm0,xmm2
4335 vmovdqa xmm2,XMMWORD[((64-128))+rax]
4336
4337 vpaddd xmm10,xmm10,xmm15
4338 vpslld xmm8,xmm11,5
4339 vpand xmm7,xmm14,xmm13
4340 vpxor xmm0,xmm0,XMMWORD[((160-128))+rax]
4341
4342 vpaddd xmm10,xmm10,xmm7
4343 vpsrld xmm9,xmm11,27
4344 vpxor xmm6,xmm14,xmm13
4345 vpxor xmm0,xmm0,xmm2
4346
4347 vmovdqu XMMWORD[(16-128)+rax],xmm4
4348 vpaddd xmm10,xmm10,xmm4
4349 vpor xmm8,xmm8,xmm9
4350 vpsrld xmm5,xmm0,31
4351 vpand xmm6,xmm6,xmm12
4352 vpaddd xmm0,xmm0,xmm0
4353
4354 vpslld xmm7,xmm12,30
4355 vpaddd xmm10,xmm10,xmm6
4356
4357 vpsrld xmm12,xmm12,2
4358 vpaddd xmm10,xmm10,xmm8
4359 vpor xmm0,xmm0,xmm5
4360 vpor xmm12,xmm12,xmm7
4361 vpxor xmm1,xmm1,xmm3
4362 vmovdqa xmm3,XMMWORD[((80-128))+rax]
4363
4364 vpaddd xmm14,xmm14,xmm15
4365 vpslld xmm8,xmm10,5
4366 vpand xmm7,xmm13,xmm12
4367 vpxor xmm1,xmm1,XMMWORD[((176-128))+rax]
4368
4369 vpaddd xmm14,xmm14,xmm7
4370 vpsrld xmm9,xmm10,27
4371 vpxor xmm6,xmm13,xmm12
4372 vpxor xmm1,xmm1,xmm3
4373
4374 vmovdqu XMMWORD[(32-128)+rax],xmm0
4375 vpaddd xmm14,xmm14,xmm0
4376 vpor xmm8,xmm8,xmm9
4377 vpsrld xmm5,xmm1,31
4378 vpand xmm6,xmm6,xmm11
4379 vpaddd xmm1,xmm1,xmm1
4380
4381 vpslld xmm7,xmm11,30
4382 vpaddd xmm14,xmm14,xmm6
4383
4384 vpsrld xmm11,xmm11,2
4385 vpaddd xmm14,xmm14,xmm8
4386 vpor xmm1,xmm1,xmm5
4387 vpor xmm11,xmm11,xmm7
4388 vpxor xmm2,xmm2,xmm4
4389 vmovdqa xmm4,XMMWORD[((96-128))+rax]
4390
4391 vpaddd xmm13,xmm13,xmm15
4392 vpslld xmm8,xmm14,5
4393 vpand xmm7,xmm12,xmm11
4394 vpxor xmm2,xmm2,XMMWORD[((192-128))+rax]
4395
4396 vpaddd xmm13,xmm13,xmm7
4397 vpsrld xmm9,xmm14,27
4398 vpxor xmm6,xmm12,xmm11
4399 vpxor xmm2,xmm2,xmm4
4400
4401 vmovdqu XMMWORD[(48-128)+rax],xmm1
4402 vpaddd xmm13,xmm13,xmm1
4403 vpor xmm8,xmm8,xmm9
4404 vpsrld xmm5,xmm2,31
4405 vpand xmm6,xmm6,xmm10
4406 vpaddd xmm2,xmm2,xmm2
4407
4408 vpslld xmm7,xmm10,30
4409 vpaddd xmm13,xmm13,xmm6
4410
4411 vpsrld xmm10,xmm10,2
4412 vpaddd xmm13,xmm13,xmm8
4413 vpor xmm2,xmm2,xmm5
4414 vpor xmm10,xmm10,xmm7
4415 vpxor xmm3,xmm3,xmm0
4416 vmovdqa xmm0,XMMWORD[((112-128))+rax]
4417
4418 vpaddd xmm12,xmm12,xmm15
4419 vpslld xmm8,xmm13,5
4420 vpand xmm7,xmm11,xmm10
4421 vpxor xmm3,xmm3,XMMWORD[((208-128))+rax]
4422
4423 vpaddd xmm12,xmm12,xmm7
4424 vpsrld xmm9,xmm13,27
4425 vpxor xmm6,xmm11,xmm10
4426 vpxor xmm3,xmm3,xmm0
4427
4428 vmovdqu XMMWORD[(64-128)+rax],xmm2
4429 vpaddd xmm12,xmm12,xmm2
4430 vpor xmm8,xmm8,xmm9
4431 vpsrld xmm5,xmm3,31
4432 vpand xmm6,xmm6,xmm14
4433 vpaddd xmm3,xmm3,xmm3
4434
4435 vpslld xmm7,xmm14,30
4436 vpaddd xmm12,xmm12,xmm6
4437
4438 vpsrld xmm14,xmm14,2
4439 vpaddd xmm12,xmm12,xmm8
4440 vpor xmm3,xmm3,xmm5
4441 vpor xmm14,xmm14,xmm7
4442 vpxor xmm4,xmm4,xmm1
4443 vmovdqa xmm1,XMMWORD[((128-128))+rax]
4444
4445 vpaddd xmm11,xmm11,xmm15
4446 vpslld xmm8,xmm12,5
4447 vpand xmm7,xmm10,xmm14
4448 vpxor xmm4,xmm4,XMMWORD[((224-128))+rax]
4449
4450 vpaddd xmm11,xmm11,xmm7
4451 vpsrld xmm9,xmm12,27
4452 vpxor xmm6,xmm10,xmm14
4453 vpxor xmm4,xmm4,xmm1
4454
4455 vmovdqu XMMWORD[(80-128)+rax],xmm3
4456 vpaddd xmm11,xmm11,xmm3
4457 vpor xmm8,xmm8,xmm9
4458 vpsrld xmm5,xmm4,31
4459 vpand xmm6,xmm6,xmm13
4460 vpaddd xmm4,xmm4,xmm4
4461
4462 vpslld xmm7,xmm13,30
4463 vpaddd xmm11,xmm11,xmm6
4464
4465 vpsrld xmm13,xmm13,2
4466 vpaddd xmm11,xmm11,xmm8
4467 vpor xmm4,xmm4,xmm5
4468 vpor xmm13,xmm13,xmm7
4469 vpxor xmm0,xmm0,xmm2
4470 vmovdqa xmm2,XMMWORD[((144-128))+rax]
4471
4472 vpaddd xmm10,xmm10,xmm15
4473 vpslld xmm8,xmm11,5
4474 vpand xmm7,xmm14,xmm13
4475 vpxor xmm0,xmm0,XMMWORD[((240-128))+rax]
4476
4477 vpaddd xmm10,xmm10,xmm7
4478 vpsrld xmm9,xmm11,27
4479 vpxor xmm6,xmm14,xmm13
4480 vpxor xmm0,xmm0,xmm2
4481
4482 vmovdqu XMMWORD[(96-128)+rax],xmm4
4483 vpaddd xmm10,xmm10,xmm4
4484 vpor xmm8,xmm8,xmm9
4485 vpsrld xmm5,xmm0,31
4486 vpand xmm6,xmm6,xmm12
4487 vpaddd xmm0,xmm0,xmm0
4488
4489 vpslld xmm7,xmm12,30
4490 vpaddd xmm10,xmm10,xmm6
4491
4492 vpsrld xmm12,xmm12,2
4493 vpaddd xmm10,xmm10,xmm8
4494 vpor xmm0,xmm0,xmm5
4495 vpor xmm12,xmm12,xmm7
4496 vpxor xmm1,xmm1,xmm3
4497 vmovdqa xmm3,XMMWORD[((160-128))+rax]
4498
4499 vpaddd xmm14,xmm14,xmm15
4500 vpslld xmm8,xmm10,5
4501 vpand xmm7,xmm13,xmm12
4502 vpxor xmm1,xmm1,XMMWORD[((0-128))+rax]
4503
4504 vpaddd xmm14,xmm14,xmm7
4505 vpsrld xmm9,xmm10,27
4506 vpxor xmm6,xmm13,xmm12
4507 vpxor xmm1,xmm1,xmm3
4508
4509 vmovdqu XMMWORD[(112-128)+rax],xmm0
4510 vpaddd xmm14,xmm14,xmm0
4511 vpor xmm8,xmm8,xmm9
4512 vpsrld xmm5,xmm1,31
4513 vpand xmm6,xmm6,xmm11
4514 vpaddd xmm1,xmm1,xmm1
4515
4516 vpslld xmm7,xmm11,30
4517 vpaddd xmm14,xmm14,xmm6
4518
4519 vpsrld xmm11,xmm11,2
4520 vpaddd xmm14,xmm14,xmm8
4521 vpor xmm1,xmm1,xmm5
4522 vpor xmm11,xmm11,xmm7
4523 vpxor xmm2,xmm2,xmm4
4524 vmovdqa xmm4,XMMWORD[((176-128))+rax]
4525
4526 vpaddd xmm13,xmm13,xmm15
4527 vpslld xmm8,xmm14,5
4528 vpand xmm7,xmm12,xmm11
4529 vpxor xmm2,xmm2,XMMWORD[((16-128))+rax]
4530
4531 vpaddd xmm13,xmm13,xmm7
4532 vpsrld xmm9,xmm14,27
4533 vpxor xmm6,xmm12,xmm11
4534 vpxor xmm2,xmm2,xmm4
4535
4536 vmovdqu XMMWORD[(128-128)+rax],xmm1
4537 vpaddd xmm13,xmm13,xmm1
4538 vpor xmm8,xmm8,xmm9
4539 vpsrld xmm5,xmm2,31
4540 vpand xmm6,xmm6,xmm10
4541 vpaddd xmm2,xmm2,xmm2
4542
4543 vpslld xmm7,xmm10,30
4544 vpaddd xmm13,xmm13,xmm6
4545
4546 vpsrld xmm10,xmm10,2
4547 vpaddd xmm13,xmm13,xmm8
4548 vpor xmm2,xmm2,xmm5
4549 vpor xmm10,xmm10,xmm7
4550 vpxor xmm3,xmm3,xmm0
4551 vmovdqa xmm0,XMMWORD[((192-128))+rax]
4552
4553 vpaddd xmm12,xmm12,xmm15
4554 vpslld xmm8,xmm13,5
4555 vpand xmm7,xmm11,xmm10
4556 vpxor xmm3,xmm3,XMMWORD[((32-128))+rax]
4557
4558 vpaddd xmm12,xmm12,xmm7
4559 vpsrld xmm9,xmm13,27
4560 vpxor xmm6,xmm11,xmm10
4561 vpxor xmm3,xmm3,xmm0
4562
4563 vmovdqu XMMWORD[(144-128)+rax],xmm2
4564 vpaddd xmm12,xmm12,xmm2
4565 vpor xmm8,xmm8,xmm9
4566 vpsrld xmm5,xmm3,31
4567 vpand xmm6,xmm6,xmm14
4568 vpaddd xmm3,xmm3,xmm3
4569
4570 vpslld xmm7,xmm14,30
4571 vpaddd xmm12,xmm12,xmm6
4572
4573 vpsrld xmm14,xmm14,2
4574 vpaddd xmm12,xmm12,xmm8
4575 vpor xmm3,xmm3,xmm5
4576 vpor xmm14,xmm14,xmm7
4577 vpxor xmm4,xmm4,xmm1
4578 vmovdqa xmm1,XMMWORD[((208-128))+rax]
4579
4580 vpaddd xmm11,xmm11,xmm15
4581 vpslld xmm8,xmm12,5
4582 vpand xmm7,xmm10,xmm14
4583 vpxor xmm4,xmm4,XMMWORD[((48-128))+rax]
4584
4585 vpaddd xmm11,xmm11,xmm7
4586 vpsrld xmm9,xmm12,27
4587 vpxor xmm6,xmm10,xmm14
4588 vpxor xmm4,xmm4,xmm1
4589
4590 vmovdqu XMMWORD[(160-128)+rax],xmm3
4591 vpaddd xmm11,xmm11,xmm3
4592 vpor xmm8,xmm8,xmm9
4593 vpsrld xmm5,xmm4,31
4594 vpand xmm6,xmm6,xmm13
4595 vpaddd xmm4,xmm4,xmm4
4596
4597 vpslld xmm7,xmm13,30
4598 vpaddd xmm11,xmm11,xmm6
4599
4600 vpsrld xmm13,xmm13,2
4601 vpaddd xmm11,xmm11,xmm8
4602 vpor xmm4,xmm4,xmm5
4603 vpor xmm13,xmm13,xmm7
4604 vpxor xmm0,xmm0,xmm2
4605 vmovdqa xmm2,XMMWORD[((224-128))+rax]
4606
4607 vpaddd xmm10,xmm10,xmm15
4608 vpslld xmm8,xmm11,5
4609 vpand xmm7,xmm14,xmm13
4610 vpxor xmm0,xmm0,XMMWORD[((64-128))+rax]
4611
4612 vpaddd xmm10,xmm10,xmm7
4613 vpsrld xmm9,xmm11,27
4614 vpxor xmm6,xmm14,xmm13
4615 vpxor xmm0,xmm0,xmm2
4616
4617 vmovdqu XMMWORD[(176-128)+rax],xmm4
4618 vpaddd xmm10,xmm10,xmm4
4619 vpor xmm8,xmm8,xmm9
4620 vpsrld xmm5,xmm0,31
4621 vpand xmm6,xmm6,xmm12
4622 vpaddd xmm0,xmm0,xmm0
4623
4624 vpslld xmm7,xmm12,30
4625 vpaddd xmm10,xmm10,xmm6
4626
4627 vpsrld xmm12,xmm12,2
4628 vpaddd xmm10,xmm10,xmm8
4629 vpor xmm0,xmm0,xmm5
4630 vpor xmm12,xmm12,xmm7
4631 vmovdqa xmm15,XMMWORD[64+rbp]
4632 vpxor xmm1,xmm1,xmm3
4633 vmovdqa xmm3,XMMWORD[((240-128))+rax]
4634
4635 vpslld xmm8,xmm10,5
4636 vpaddd xmm14,xmm14,xmm15
4637 vpxor xmm6,xmm13,xmm11
4638 vmovdqa XMMWORD[(192-128)+rax],xmm0
4639 vpaddd xmm14,xmm14,xmm0
4640 vpxor xmm1,xmm1,XMMWORD[((80-128))+rax]
4641 vpsrld xmm9,xmm10,27
4642 vpxor xmm6,xmm6,xmm12
4643 vpxor xmm1,xmm1,xmm3
4644
4645 vpslld xmm7,xmm11,30
4646 vpor xmm8,xmm8,xmm9
4647 vpaddd xmm14,xmm14,xmm6
4648 vpsrld xmm5,xmm1,31
4649 vpaddd xmm1,xmm1,xmm1
4650
4651 vpsrld xmm11,xmm11,2
4652 vpaddd xmm14,xmm14,xmm8
4653 vpor xmm1,xmm1,xmm5
4654 vpor xmm11,xmm11,xmm7
4655 vpxor xmm2,xmm2,xmm4
4656 vmovdqa xmm4,XMMWORD[((0-128))+rax]
4657
4658 vpslld xmm8,xmm14,5
4659 vpaddd xmm13,xmm13,xmm15
4660 vpxor xmm6,xmm12,xmm10
4661 vmovdqa XMMWORD[(208-128)+rax],xmm1
4662 vpaddd xmm13,xmm13,xmm1
4663 vpxor xmm2,xmm2,XMMWORD[((96-128))+rax]
4664 vpsrld xmm9,xmm14,27
4665 vpxor xmm6,xmm6,xmm11
4666 vpxor xmm2,xmm2,xmm4
4667
4668 vpslld xmm7,xmm10,30
4669 vpor xmm8,xmm8,xmm9
4670 vpaddd xmm13,xmm13,xmm6
4671 vpsrld xmm5,xmm2,31
4672 vpaddd xmm2,xmm2,xmm2
4673
4674 vpsrld xmm10,xmm10,2
4675 vpaddd xmm13,xmm13,xmm8
4676 vpor xmm2,xmm2,xmm5
4677 vpor xmm10,xmm10,xmm7
4678 vpxor xmm3,xmm3,xmm0
4679 vmovdqa xmm0,XMMWORD[((16-128))+rax]
4680
4681 vpslld xmm8,xmm13,5
4682 vpaddd xmm12,xmm12,xmm15
4683 vpxor xmm6,xmm11,xmm14
4684 vmovdqa XMMWORD[(224-128)+rax],xmm2
4685 vpaddd xmm12,xmm12,xmm2
4686 vpxor xmm3,xmm3,XMMWORD[((112-128))+rax]
4687 vpsrld xmm9,xmm13,27
4688 vpxor xmm6,xmm6,xmm10
4689 vpxor xmm3,xmm3,xmm0
4690
4691 vpslld xmm7,xmm14,30
4692 vpor xmm8,xmm8,xmm9
4693 vpaddd xmm12,xmm12,xmm6
4694 vpsrld xmm5,xmm3,31
4695 vpaddd xmm3,xmm3,xmm3
4696
4697 vpsrld xmm14,xmm14,2
4698 vpaddd xmm12,xmm12,xmm8
4699 vpor xmm3,xmm3,xmm5
4700 vpor xmm14,xmm14,xmm7
4701 vpxor xmm4,xmm4,xmm1
4702 vmovdqa xmm1,XMMWORD[((32-128))+rax]
4703
4704 vpslld xmm8,xmm12,5
4705 vpaddd xmm11,xmm11,xmm15
4706 vpxor xmm6,xmm10,xmm13
4707 vmovdqa XMMWORD[(240-128)+rax],xmm3
4708 vpaddd xmm11,xmm11,xmm3
4709 vpxor xmm4,xmm4,XMMWORD[((128-128))+rax]
4710 vpsrld xmm9,xmm12,27
4711 vpxor xmm6,xmm6,xmm14
4712 vpxor xmm4,xmm4,xmm1
4713
4714 vpslld xmm7,xmm13,30
4715 vpor xmm8,xmm8,xmm9
4716 vpaddd xmm11,xmm11,xmm6
4717 vpsrld xmm5,xmm4,31
4718 vpaddd xmm4,xmm4,xmm4
4719
4720 vpsrld xmm13,xmm13,2
4721 vpaddd xmm11,xmm11,xmm8
4722 vpor xmm4,xmm4,xmm5
4723 vpor xmm13,xmm13,xmm7
4724 vpxor xmm0,xmm0,xmm2
4725 vmovdqa xmm2,XMMWORD[((48-128))+rax]
4726
4727 vpslld xmm8,xmm11,5
4728 vpaddd xmm10,xmm10,xmm15
4729 vpxor xmm6,xmm14,xmm12
4730 vmovdqa XMMWORD[(0-128)+rax],xmm4
4731 vpaddd xmm10,xmm10,xmm4
4732 vpxor xmm0,xmm0,XMMWORD[((144-128))+rax]
4733 vpsrld xmm9,xmm11,27
4734 vpxor xmm6,xmm6,xmm13
4735 vpxor xmm0,xmm0,xmm2
4736
4737 vpslld xmm7,xmm12,30
4738 vpor xmm8,xmm8,xmm9
4739 vpaddd xmm10,xmm10,xmm6
4740 vpsrld xmm5,xmm0,31
4741 vpaddd xmm0,xmm0,xmm0
4742
4743 vpsrld xmm12,xmm12,2
4744 vpaddd xmm10,xmm10,xmm8
4745 vpor xmm0,xmm0,xmm5
4746 vpor xmm12,xmm12,xmm7
4747 vpxor xmm1,xmm1,xmm3
4748 vmovdqa xmm3,XMMWORD[((64-128))+rax]
4749
4750 vpslld xmm8,xmm10,5
4751 vpaddd xmm14,xmm14,xmm15
4752 vpxor xmm6,xmm13,xmm11
4753 vmovdqa XMMWORD[(16-128)+rax],xmm0
4754 vpaddd xmm14,xmm14,xmm0
4755 vpxor xmm1,xmm1,XMMWORD[((160-128))+rax]
4756 vpsrld xmm9,xmm10,27
4757 vpxor xmm6,xmm6,xmm12
4758 vpxor xmm1,xmm1,xmm3
4759
4760 vpslld xmm7,xmm11,30
4761 vpor xmm8,xmm8,xmm9
4762 vpaddd xmm14,xmm14,xmm6
4763 vpsrld xmm5,xmm1,31
4764 vpaddd xmm1,xmm1,xmm1
4765
4766 vpsrld xmm11,xmm11,2
4767 vpaddd xmm14,xmm14,xmm8
4768 vpor xmm1,xmm1,xmm5
4769 vpor xmm11,xmm11,xmm7
4770 vpxor xmm2,xmm2,xmm4
4771 vmovdqa xmm4,XMMWORD[((80-128))+rax]
4772
4773 vpslld xmm8,xmm14,5
4774 vpaddd xmm13,xmm13,xmm15
4775 vpxor xmm6,xmm12,xmm10
4776 vmovdqa XMMWORD[(32-128)+rax],xmm1
4777 vpaddd xmm13,xmm13,xmm1
4778 vpxor xmm2,xmm2,XMMWORD[((176-128))+rax]
4779 vpsrld xmm9,xmm14,27
4780 vpxor xmm6,xmm6,xmm11
4781 vpxor xmm2,xmm2,xmm4
4782
4783 vpslld xmm7,xmm10,30
4784 vpor xmm8,xmm8,xmm9
4785 vpaddd xmm13,xmm13,xmm6
4786 vpsrld xmm5,xmm2,31
4787 vpaddd xmm2,xmm2,xmm2
4788
4789 vpsrld xmm10,xmm10,2
4790 vpaddd xmm13,xmm13,xmm8
4791 vpor xmm2,xmm2,xmm5
4792 vpor xmm10,xmm10,xmm7
4793 vpxor xmm3,xmm3,xmm0
4794 vmovdqa xmm0,XMMWORD[((96-128))+rax]
4795
4796 vpslld xmm8,xmm13,5
4797 vpaddd xmm12,xmm12,xmm15
4798 vpxor xmm6,xmm11,xmm14
4799 vmovdqa XMMWORD[(48-128)+rax],xmm2
4800 vpaddd xmm12,xmm12,xmm2
4801 vpxor xmm3,xmm3,XMMWORD[((192-128))+rax]
4802 vpsrld xmm9,xmm13,27
4803 vpxor xmm6,xmm6,xmm10
4804 vpxor xmm3,xmm3,xmm0
4805
4806 vpslld xmm7,xmm14,30
4807 vpor xmm8,xmm8,xmm9
4808 vpaddd xmm12,xmm12,xmm6
4809 vpsrld xmm5,xmm3,31
4810 vpaddd xmm3,xmm3,xmm3
4811
4812 vpsrld xmm14,xmm14,2
4813 vpaddd xmm12,xmm12,xmm8
4814 vpor xmm3,xmm3,xmm5
4815 vpor xmm14,xmm14,xmm7
4816 vpxor xmm4,xmm4,xmm1
4817 vmovdqa xmm1,XMMWORD[((112-128))+rax]
4818
4819 vpslld xmm8,xmm12,5
4820 vpaddd xmm11,xmm11,xmm15
4821 vpxor xmm6,xmm10,xmm13
4822 vmovdqa XMMWORD[(64-128)+rax],xmm3
4823 vpaddd xmm11,xmm11,xmm3
4824 vpxor xmm4,xmm4,XMMWORD[((208-128))+rax]
4825 vpsrld xmm9,xmm12,27
4826 vpxor xmm6,xmm6,xmm14
4827 vpxor xmm4,xmm4,xmm1
4828
4829 vpslld xmm7,xmm13,30
4830 vpor xmm8,xmm8,xmm9
4831 vpaddd xmm11,xmm11,xmm6
4832 vpsrld xmm5,xmm4,31
4833 vpaddd xmm4,xmm4,xmm4
4834
4835 vpsrld xmm13,xmm13,2
4836 vpaddd xmm11,xmm11,xmm8
4837 vpor xmm4,xmm4,xmm5
4838 vpor xmm13,xmm13,xmm7
4839 vpxor xmm0,xmm0,xmm2
4840 vmovdqa xmm2,XMMWORD[((128-128))+rax]
4841
4842 vpslld xmm8,xmm11,5
4843 vpaddd xmm10,xmm10,xmm15
4844 vpxor xmm6,xmm14,xmm12
4845 vmovdqa XMMWORD[(80-128)+rax],xmm4
4846 vpaddd xmm10,xmm10,xmm4
4847 vpxor xmm0,xmm0,XMMWORD[((224-128))+rax]
4848 vpsrld xmm9,xmm11,27
4849 vpxor xmm6,xmm6,xmm13
4850 vpxor xmm0,xmm0,xmm2
4851
4852 vpslld xmm7,xmm12,30
4853 vpor xmm8,xmm8,xmm9
4854 vpaddd xmm10,xmm10,xmm6
4855 vpsrld xmm5,xmm0,31
4856 vpaddd xmm0,xmm0,xmm0
4857
4858 vpsrld xmm12,xmm12,2
4859 vpaddd xmm10,xmm10,xmm8
4860 vpor xmm0,xmm0,xmm5
4861 vpor xmm12,xmm12,xmm7
4862 vpxor xmm1,xmm1,xmm3
4863 vmovdqa xmm3,XMMWORD[((144-128))+rax]
4864
4865 vpslld xmm8,xmm10,5
4866 vpaddd xmm14,xmm14,xmm15
4867 vpxor xmm6,xmm13,xmm11
4868 vmovdqa XMMWORD[(96-128)+rax],xmm0
4869 vpaddd xmm14,xmm14,xmm0
4870 vpxor xmm1,xmm1,XMMWORD[((240-128))+rax]
4871 vpsrld xmm9,xmm10,27
4872 vpxor xmm6,xmm6,xmm12
4873 vpxor xmm1,xmm1,xmm3
4874
4875 vpslld xmm7,xmm11,30
4876 vpor xmm8,xmm8,xmm9
4877 vpaddd xmm14,xmm14,xmm6
4878 vpsrld xmm5,xmm1,31
4879 vpaddd xmm1,xmm1,xmm1
4880
4881 vpsrld xmm11,xmm11,2
4882 vpaddd xmm14,xmm14,xmm8
4883 vpor xmm1,xmm1,xmm5
4884 vpor xmm11,xmm11,xmm7
4885 vpxor xmm2,xmm2,xmm4
4886 vmovdqa xmm4,XMMWORD[((160-128))+rax]
4887
4888 vpslld xmm8,xmm14,5
4889 vpaddd xmm13,xmm13,xmm15
4890 vpxor xmm6,xmm12,xmm10
4891 vmovdqa XMMWORD[(112-128)+rax],xmm1
4892 vpaddd xmm13,xmm13,xmm1
4893 vpxor xmm2,xmm2,XMMWORD[((0-128))+rax]
4894 vpsrld xmm9,xmm14,27
4895 vpxor xmm6,xmm6,xmm11
4896 vpxor xmm2,xmm2,xmm4
4897
4898 vpslld xmm7,xmm10,30
4899 vpor xmm8,xmm8,xmm9
4900 vpaddd xmm13,xmm13,xmm6
4901 vpsrld xmm5,xmm2,31
4902 vpaddd xmm2,xmm2,xmm2
4903
4904 vpsrld xmm10,xmm10,2
4905 vpaddd xmm13,xmm13,xmm8
4906 vpor xmm2,xmm2,xmm5
4907 vpor xmm10,xmm10,xmm7
4908 vpxor xmm3,xmm3,xmm0
4909 vmovdqa xmm0,XMMWORD[((176-128))+rax]
4910
4911 vpslld xmm8,xmm13,5
4912 vpaddd xmm12,xmm12,xmm15
4913 vpxor xmm6,xmm11,xmm14
4914 vpaddd xmm12,xmm12,xmm2
4915 vpxor xmm3,xmm3,XMMWORD[((16-128))+rax]
4916 vpsrld xmm9,xmm13,27
4917 vpxor xmm6,xmm6,xmm10
4918 vpxor xmm3,xmm3,xmm0
4919
4920 vpslld xmm7,xmm14,30
4921 vpor xmm8,xmm8,xmm9
4922 vpaddd xmm12,xmm12,xmm6
4923 vpsrld xmm5,xmm3,31
4924 vpaddd xmm3,xmm3,xmm3
4925
4926 vpsrld xmm14,xmm14,2
4927 vpaddd xmm12,xmm12,xmm8
4928 vpor xmm3,xmm3,xmm5
4929 vpor xmm14,xmm14,xmm7
4930 vpxor xmm4,xmm4,xmm1
4931 vmovdqa xmm1,XMMWORD[((192-128))+rax]
4932
4933 vpslld xmm8,xmm12,5
4934 vpaddd xmm11,xmm11,xmm15
4935 vpxor xmm6,xmm10,xmm13
4936 vpaddd xmm11,xmm11,xmm3
4937 vpxor xmm4,xmm4,XMMWORD[((32-128))+rax]
4938 vpsrld xmm9,xmm12,27
4939 vpxor xmm6,xmm6,xmm14
4940 vpxor xmm4,xmm4,xmm1
4941
4942 vpslld xmm7,xmm13,30
4943 vpor xmm8,xmm8,xmm9
4944 vpaddd xmm11,xmm11,xmm6
4945 vpsrld xmm5,xmm4,31
4946 vpaddd xmm4,xmm4,xmm4
4947
4948 vpsrld xmm13,xmm13,2
4949 vpaddd xmm11,xmm11,xmm8
4950 vpor xmm4,xmm4,xmm5
4951 vpor xmm13,xmm13,xmm7
4952 vpxor xmm0,xmm0,xmm2
4953 vmovdqa xmm2,XMMWORD[((208-128))+rax]
4954
4955 vpslld xmm8,xmm11,5
4956 vpaddd xmm10,xmm10,xmm15
4957 vpxor xmm6,xmm14,xmm12
4958 vpaddd xmm10,xmm10,xmm4
4959 vpxor xmm0,xmm0,XMMWORD[((48-128))+rax]
4960 vpsrld xmm9,xmm11,27
4961 vpxor xmm6,xmm6,xmm13
4962 vpxor xmm0,xmm0,xmm2
4963
4964 vpslld xmm7,xmm12,30
4965 vpor xmm8,xmm8,xmm9
4966 vpaddd xmm10,xmm10,xmm6
4967 vpsrld xmm5,xmm0,31
4968 vpaddd xmm0,xmm0,xmm0
4969
4970 vpsrld xmm12,xmm12,2
4971 vpaddd xmm10,xmm10,xmm8
4972 vpor xmm0,xmm0,xmm5
4973 vpor xmm12,xmm12,xmm7
4974 vpxor xmm1,xmm1,xmm3
4975 vmovdqa xmm3,XMMWORD[((224-128))+rax]
4976
4977 vpslld xmm8,xmm10,5
4978 vpaddd xmm14,xmm14,xmm15
4979 vpxor xmm6,xmm13,xmm11
4980 vpaddd xmm14,xmm14,xmm0
4981 vpxor xmm1,xmm1,XMMWORD[((64-128))+rax]
4982 vpsrld xmm9,xmm10,27
4983 vpxor xmm6,xmm6,xmm12
4984 vpxor xmm1,xmm1,xmm3
4985
4986 vpslld xmm7,xmm11,30
4987 vpor xmm8,xmm8,xmm9
4988 vpaddd xmm14,xmm14,xmm6
4989 vpsrld xmm5,xmm1,31
4990 vpaddd xmm1,xmm1,xmm1
4991
4992 vpsrld xmm11,xmm11,2
4993 vpaddd xmm14,xmm14,xmm8
4994 vpor xmm1,xmm1,xmm5
4995 vpor xmm11,xmm11,xmm7
4996 vpxor xmm2,xmm2,xmm4
4997 vmovdqa xmm4,XMMWORD[((240-128))+rax]
4998
4999 vpslld xmm8,xmm14,5
5000 vpaddd xmm13,xmm13,xmm15
5001 vpxor xmm6,xmm12,xmm10
5002 vpaddd xmm13,xmm13,xmm1
5003 vpxor xmm2,xmm2,XMMWORD[((80-128))+rax]
5004 vpsrld xmm9,xmm14,27
5005 vpxor xmm6,xmm6,xmm11
5006 vpxor xmm2,xmm2,xmm4
5007
5008 vpslld xmm7,xmm10,30
5009 vpor xmm8,xmm8,xmm9
5010 vpaddd xmm13,xmm13,xmm6
5011 vpsrld xmm5,xmm2,31
5012 vpaddd xmm2,xmm2,xmm2
5013
5014 vpsrld xmm10,xmm10,2
5015 vpaddd xmm13,xmm13,xmm8
5016 vpor xmm2,xmm2,xmm5
5017 vpor xmm10,xmm10,xmm7
5018 vpxor xmm3,xmm3,xmm0
5019 vmovdqa xmm0,XMMWORD[((0-128))+rax]
5020
5021 vpslld xmm8,xmm13,5
5022 vpaddd xmm12,xmm12,xmm15
5023 vpxor xmm6,xmm11,xmm14
5024 vpaddd xmm12,xmm12,xmm2
5025 vpxor xmm3,xmm3,XMMWORD[((96-128))+rax]
5026 vpsrld xmm9,xmm13,27
5027 vpxor xmm6,xmm6,xmm10
5028 vpxor xmm3,xmm3,xmm0
5029
5030 vpslld xmm7,xmm14,30
5031 vpor xmm8,xmm8,xmm9
5032 vpaddd xmm12,xmm12,xmm6
5033 vpsrld xmm5,xmm3,31
5034 vpaddd xmm3,xmm3,xmm3
5035
5036 vpsrld xmm14,xmm14,2
5037 vpaddd xmm12,xmm12,xmm8
5038 vpor xmm3,xmm3,xmm5
5039 vpor xmm14,xmm14,xmm7
5040 vpxor xmm4,xmm4,xmm1
5041 vmovdqa xmm1,XMMWORD[((16-128))+rax]
5042
5043 vpslld xmm8,xmm12,5
5044 vpaddd xmm11,xmm11,xmm15
5045 vpxor xmm6,xmm10,xmm13
5046 vpaddd xmm11,xmm11,xmm3
5047 vpxor xmm4,xmm4,XMMWORD[((112-128))+rax]
5048 vpsrld xmm9,xmm12,27
5049 vpxor xmm6,xmm6,xmm14
5050 vpxor xmm4,xmm4,xmm1
5051
5052 vpslld xmm7,xmm13,30
5053 vpor xmm8,xmm8,xmm9
5054 vpaddd xmm11,xmm11,xmm6
5055 vpsrld xmm5,xmm4,31
5056 vpaddd xmm4,xmm4,xmm4
5057
5058 vpsrld xmm13,xmm13,2
5059 vpaddd xmm11,xmm11,xmm8
5060 vpor xmm4,xmm4,xmm5
5061 vpor xmm13,xmm13,xmm7
5062 vpslld xmm8,xmm11,5
5063 vpaddd xmm10,xmm10,xmm15
5064 vpxor xmm6,xmm14,xmm12
5065
5066 vpsrld xmm9,xmm11,27
5067 vpaddd xmm10,xmm10,xmm4
5068 vpxor xmm6,xmm6,xmm13
5069
5070 vpslld xmm7,xmm12,30
5071 vpor xmm8,xmm8,xmm9
5072 vpaddd xmm10,xmm10,xmm6
5073
5074 vpsrld xmm12,xmm12,2
5075 vpaddd xmm10,xmm10,xmm8
5076 vpor xmm12,xmm12,xmm7
5077 mov ecx,1
5078 cmp ecx,DWORD[rbx]
5079 cmovge r8,rbp
5080 cmp ecx,DWORD[4+rbx]
5081 cmovge r9,rbp
5082 cmp ecx,DWORD[8+rbx]
5083 cmovge r10,rbp
5084 cmp ecx,DWORD[12+rbx]
5085 cmovge r11,rbp
5086 vmovdqu xmm6,XMMWORD[rbx]
5087 vpxor xmm8,xmm8,xmm8
5088 vmovdqa xmm7,xmm6
5089 vpcmpgtd xmm7,xmm7,xmm8
5090 vpaddd xmm6,xmm6,xmm7
5091
5092 vpand xmm10,xmm10,xmm7
5093 vpand xmm11,xmm11,xmm7
5094 vpaddd xmm10,xmm10,XMMWORD[rdi]
5095 vpand xmm12,xmm12,xmm7
5096 vpaddd xmm11,xmm11,XMMWORD[32+rdi]
5097 vpand xmm13,xmm13,xmm7
5098 vpaddd xmm12,xmm12,XMMWORD[64+rdi]
5099 vpand xmm14,xmm14,xmm7
5100 vpaddd xmm13,xmm13,XMMWORD[96+rdi]
5101 vpaddd xmm14,xmm14,XMMWORD[128+rdi]
5102 vmovdqu XMMWORD[rdi],xmm10
5103 vmovdqu XMMWORD[32+rdi],xmm11
5104 vmovdqu XMMWORD[64+rdi],xmm12
5105 vmovdqu XMMWORD[96+rdi],xmm13
5106 vmovdqu XMMWORD[128+rdi],xmm14
5107
5108 vmovdqu XMMWORD[rbx],xmm6
5109 vmovdqu xmm5,XMMWORD[96+rbp]
5110 dec edx
5111 jnz NEAR $L$oop_avx
5112
5113 mov edx,DWORD[280+rsp]
5114 lea rdi,[16+rdi]
5115 lea rsi,[64+rsi]
5116 dec edx
5117 jnz NEAR $L$oop_grande_avx
5118
5119$L$done_avx:
5120 mov rax,QWORD[272+rsp]
5121
5122 vzeroupper
5123 movaps xmm6,XMMWORD[((-184))+rax]
5124 movaps xmm7,XMMWORD[((-168))+rax]
5125 movaps xmm8,XMMWORD[((-152))+rax]
5126 movaps xmm9,XMMWORD[((-136))+rax]
5127 movaps xmm10,XMMWORD[((-120))+rax]
5128 movaps xmm11,XMMWORD[((-104))+rax]
5129 movaps xmm12,XMMWORD[((-88))+rax]
5130 movaps xmm13,XMMWORD[((-72))+rax]
5131 movaps xmm14,XMMWORD[((-56))+rax]
5132 movaps xmm15,XMMWORD[((-40))+rax]
5133 mov rbp,QWORD[((-16))+rax]
5134
5135 mov rbx,QWORD[((-8))+rax]
5136
5137 lea rsp,[rax]
5138
5139$L$epilogue_avx:
5140 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
5141 mov rsi,QWORD[16+rsp]
5142 DB 0F3h,0C3h ;repret
5143
5144$L$SEH_end_sha1_multi_block_avx:
5145
5146ALIGN 32
5147sha1_multi_block_avx2:
5148 mov QWORD[8+rsp],rdi ;WIN64 prologue
5149 mov QWORD[16+rsp],rsi
5150 mov rax,rsp
5151$L$SEH_begin_sha1_multi_block_avx2:
5152 mov rdi,rcx
5153 mov rsi,rdx
5154 mov rdx,r8
5155
5156
5157
5158_avx2_shortcut:
5159 mov rax,rsp
5160
5161 push rbx
5162
5163 push rbp
5164
5165 push r12
5166
5167 push r13
5168
5169 push r14
5170
5171 push r15
5172
5173 lea rsp,[((-168))+rsp]
5174 movaps XMMWORD[rsp],xmm6
5175 movaps XMMWORD[16+rsp],xmm7
5176 movaps XMMWORD[32+rsp],xmm8
5177 movaps XMMWORD[48+rsp],xmm9
5178 movaps XMMWORD[64+rsp],xmm10
5179 movaps XMMWORD[80+rsp],xmm11
5180 movaps XMMWORD[(-120)+rax],xmm12
5181 movaps XMMWORD[(-104)+rax],xmm13
5182 movaps XMMWORD[(-88)+rax],xmm14
5183 movaps XMMWORD[(-72)+rax],xmm15
5184 sub rsp,576
5185 and rsp,-256
5186 mov QWORD[544+rsp],rax
5187
5188$L$body_avx2:
5189 lea rbp,[K_XX_XX]
5190 shr edx,1
5191
5192 vzeroupper
5193$L$oop_grande_avx2:
5194 mov DWORD[552+rsp],edx
5195 xor edx,edx
5196 lea rbx,[512+rsp]
5197
5198 mov r12,QWORD[rsi]
5199
5200 mov ecx,DWORD[8+rsi]
5201 cmp ecx,edx
5202 cmovg edx,ecx
5203 test ecx,ecx
5204 mov DWORD[rbx],ecx
5205 cmovle r12,rbp
5206
5207 mov r13,QWORD[16+rsi]
5208
5209 mov ecx,DWORD[24+rsi]
5210 cmp ecx,edx
5211 cmovg edx,ecx
5212 test ecx,ecx
5213 mov DWORD[4+rbx],ecx
5214 cmovle r13,rbp
5215
5216 mov r14,QWORD[32+rsi]
5217
5218 mov ecx,DWORD[40+rsi]
5219 cmp ecx,edx
5220 cmovg edx,ecx
5221 test ecx,ecx
5222 mov DWORD[8+rbx],ecx
5223 cmovle r14,rbp
5224
5225 mov r15,QWORD[48+rsi]
5226
5227 mov ecx,DWORD[56+rsi]
5228 cmp ecx,edx
5229 cmovg edx,ecx
5230 test ecx,ecx
5231 mov DWORD[12+rbx],ecx
5232 cmovle r15,rbp
5233
5234 mov r8,QWORD[64+rsi]
5235
5236 mov ecx,DWORD[72+rsi]
5237 cmp ecx,edx
5238 cmovg edx,ecx
5239 test ecx,ecx
5240 mov DWORD[16+rbx],ecx
5241 cmovle r8,rbp
5242
5243 mov r9,QWORD[80+rsi]
5244
5245 mov ecx,DWORD[88+rsi]
5246 cmp ecx,edx
5247 cmovg edx,ecx
5248 test ecx,ecx
5249 mov DWORD[20+rbx],ecx
5250 cmovle r9,rbp
5251
5252 mov r10,QWORD[96+rsi]
5253
5254 mov ecx,DWORD[104+rsi]
5255 cmp ecx,edx
5256 cmovg edx,ecx
5257 test ecx,ecx
5258 mov DWORD[24+rbx],ecx
5259 cmovle r10,rbp
5260
5261 mov r11,QWORD[112+rsi]
5262
5263 mov ecx,DWORD[120+rsi]
5264 cmp ecx,edx
5265 cmovg edx,ecx
5266 test ecx,ecx
5267 mov DWORD[28+rbx],ecx
5268 cmovle r11,rbp
5269 vmovdqu ymm0,YMMWORD[rdi]
5270 lea rax,[128+rsp]
5271 vmovdqu ymm1,YMMWORD[32+rdi]
5272 lea rbx,[((256+128))+rsp]
5273 vmovdqu ymm2,YMMWORD[64+rdi]
5274 vmovdqu ymm3,YMMWORD[96+rdi]
5275 vmovdqu ymm4,YMMWORD[128+rdi]
5276 vmovdqu ymm9,YMMWORD[96+rbp]
5277 jmp NEAR $L$oop_avx2
5278
5279ALIGN 32
5280$L$oop_avx2:
5281 vmovdqa ymm15,YMMWORD[((-32))+rbp]
5282 vmovd xmm10,DWORD[r12]
5283 lea r12,[64+r12]
5284 vmovd xmm12,DWORD[r8]
5285 lea r8,[64+r8]
5286 vmovd xmm7,DWORD[r13]
5287 lea r13,[64+r13]
5288 vmovd xmm6,DWORD[r9]
5289 lea r9,[64+r9]
5290 vpinsrd xmm10,xmm10,DWORD[r14],1
5291 lea r14,[64+r14]
5292 vpinsrd xmm12,xmm12,DWORD[r10],1
5293 lea r10,[64+r10]
5294 vpinsrd xmm7,xmm7,DWORD[r15],1
5295 lea r15,[64+r15]
5296 vpunpckldq ymm10,ymm10,ymm7
5297 vpinsrd xmm6,xmm6,DWORD[r11],1
5298 lea r11,[64+r11]
5299 vpunpckldq ymm12,ymm12,ymm6
5300 vmovd xmm11,DWORD[((-60))+r12]
5301 vinserti128 ymm10,ymm10,xmm12,1
5302 vmovd xmm8,DWORD[((-60))+r8]
5303 vpshufb ymm10,ymm10,ymm9
5304 vmovd xmm7,DWORD[((-60))+r13]
5305 vmovd xmm6,DWORD[((-60))+r9]
5306 vpinsrd xmm11,xmm11,DWORD[((-60))+r14],1
5307 vpinsrd xmm8,xmm8,DWORD[((-60))+r10],1
5308 vpinsrd xmm7,xmm7,DWORD[((-60))+r15],1
5309 vpunpckldq ymm11,ymm11,ymm7
5310 vpinsrd xmm6,xmm6,DWORD[((-60))+r11],1
5311 vpunpckldq ymm8,ymm8,ymm6
5312 vpaddd ymm4,ymm4,ymm15
5313 vpslld ymm7,ymm0,5
5314 vpandn ymm6,ymm1,ymm3
5315 vpand ymm5,ymm1,ymm2
5316
5317 vmovdqa YMMWORD[(0-128)+rax],ymm10
5318 vpaddd ymm4,ymm4,ymm10
5319 vinserti128 ymm11,ymm11,xmm8,1
5320 vpsrld ymm8,ymm0,27
5321 vpxor ymm5,ymm5,ymm6
5322 vmovd xmm12,DWORD[((-56))+r12]
5323
5324 vpslld ymm6,ymm1,30
5325 vpor ymm7,ymm7,ymm8
5326 vmovd xmm8,DWORD[((-56))+r8]
5327 vpaddd ymm4,ymm4,ymm5
5328
5329 vpsrld ymm1,ymm1,2
5330 vpaddd ymm4,ymm4,ymm7
5331 vpshufb ymm11,ymm11,ymm9
5332 vpor ymm1,ymm1,ymm6
5333 vmovd xmm7,DWORD[((-56))+r13]
5334 vmovd xmm6,DWORD[((-56))+r9]
5335 vpinsrd xmm12,xmm12,DWORD[((-56))+r14],1
5336 vpinsrd xmm8,xmm8,DWORD[((-56))+r10],1
5337 vpinsrd xmm7,xmm7,DWORD[((-56))+r15],1
5338 vpunpckldq ymm12,ymm12,ymm7
5339 vpinsrd xmm6,xmm6,DWORD[((-56))+r11],1
5340 vpunpckldq ymm8,ymm8,ymm6
5341 vpaddd ymm3,ymm3,ymm15
5342 vpslld ymm7,ymm4,5
5343 vpandn ymm6,ymm0,ymm2
5344 vpand ymm5,ymm0,ymm1
5345
5346 vmovdqa YMMWORD[(32-128)+rax],ymm11
5347 vpaddd ymm3,ymm3,ymm11
5348 vinserti128 ymm12,ymm12,xmm8,1
5349 vpsrld ymm8,ymm4,27
5350 vpxor ymm5,ymm5,ymm6
5351 vmovd xmm13,DWORD[((-52))+r12]
5352
5353 vpslld ymm6,ymm0,30
5354 vpor ymm7,ymm7,ymm8
5355 vmovd xmm8,DWORD[((-52))+r8]
5356 vpaddd ymm3,ymm3,ymm5
5357
5358 vpsrld ymm0,ymm0,2
5359 vpaddd ymm3,ymm3,ymm7
5360 vpshufb ymm12,ymm12,ymm9
5361 vpor ymm0,ymm0,ymm6
5362 vmovd xmm7,DWORD[((-52))+r13]
5363 vmovd xmm6,DWORD[((-52))+r9]
5364 vpinsrd xmm13,xmm13,DWORD[((-52))+r14],1
5365 vpinsrd xmm8,xmm8,DWORD[((-52))+r10],1
5366 vpinsrd xmm7,xmm7,DWORD[((-52))+r15],1
5367 vpunpckldq ymm13,ymm13,ymm7
5368 vpinsrd xmm6,xmm6,DWORD[((-52))+r11],1
5369 vpunpckldq ymm8,ymm8,ymm6
5370 vpaddd ymm2,ymm2,ymm15
5371 vpslld ymm7,ymm3,5
5372 vpandn ymm6,ymm4,ymm1
5373 vpand ymm5,ymm4,ymm0
5374
5375 vmovdqa YMMWORD[(64-128)+rax],ymm12
5376 vpaddd ymm2,ymm2,ymm12
5377 vinserti128 ymm13,ymm13,xmm8,1
5378 vpsrld ymm8,ymm3,27
5379 vpxor ymm5,ymm5,ymm6
5380 vmovd xmm14,DWORD[((-48))+r12]
5381
5382 vpslld ymm6,ymm4,30
5383 vpor ymm7,ymm7,ymm8
5384 vmovd xmm8,DWORD[((-48))+r8]
5385 vpaddd ymm2,ymm2,ymm5
5386
5387 vpsrld ymm4,ymm4,2
5388 vpaddd ymm2,ymm2,ymm7
5389 vpshufb ymm13,ymm13,ymm9
5390 vpor ymm4,ymm4,ymm6
5391 vmovd xmm7,DWORD[((-48))+r13]
5392 vmovd xmm6,DWORD[((-48))+r9]
5393 vpinsrd xmm14,xmm14,DWORD[((-48))+r14],1
5394 vpinsrd xmm8,xmm8,DWORD[((-48))+r10],1
5395 vpinsrd xmm7,xmm7,DWORD[((-48))+r15],1
5396 vpunpckldq ymm14,ymm14,ymm7
5397 vpinsrd xmm6,xmm6,DWORD[((-48))+r11],1
5398 vpunpckldq ymm8,ymm8,ymm6
5399 vpaddd ymm1,ymm1,ymm15
5400 vpslld ymm7,ymm2,5
5401 vpandn ymm6,ymm3,ymm0
5402 vpand ymm5,ymm3,ymm4
5403
5404 vmovdqa YMMWORD[(96-128)+rax],ymm13
5405 vpaddd ymm1,ymm1,ymm13
5406 vinserti128 ymm14,ymm14,xmm8,1
5407 vpsrld ymm8,ymm2,27
5408 vpxor ymm5,ymm5,ymm6
5409 vmovd xmm10,DWORD[((-44))+r12]
5410
5411 vpslld ymm6,ymm3,30
5412 vpor ymm7,ymm7,ymm8
5413 vmovd xmm8,DWORD[((-44))+r8]
5414 vpaddd ymm1,ymm1,ymm5
5415
5416 vpsrld ymm3,ymm3,2
5417 vpaddd ymm1,ymm1,ymm7
5418 vpshufb ymm14,ymm14,ymm9
5419 vpor ymm3,ymm3,ymm6
5420 vmovd xmm7,DWORD[((-44))+r13]
5421 vmovd xmm6,DWORD[((-44))+r9]
5422 vpinsrd xmm10,xmm10,DWORD[((-44))+r14],1
5423 vpinsrd xmm8,xmm8,DWORD[((-44))+r10],1
5424 vpinsrd xmm7,xmm7,DWORD[((-44))+r15],1
5425 vpunpckldq ymm10,ymm10,ymm7
5426 vpinsrd xmm6,xmm6,DWORD[((-44))+r11],1
5427 vpunpckldq ymm8,ymm8,ymm6
5428 vpaddd ymm0,ymm0,ymm15
5429 vpslld ymm7,ymm1,5
5430 vpandn ymm6,ymm2,ymm4
5431 vpand ymm5,ymm2,ymm3
5432
5433 vmovdqa YMMWORD[(128-128)+rax],ymm14
5434 vpaddd ymm0,ymm0,ymm14
5435 vinserti128 ymm10,ymm10,xmm8,1
5436 vpsrld ymm8,ymm1,27
5437 vpxor ymm5,ymm5,ymm6
5438 vmovd xmm11,DWORD[((-40))+r12]
5439
5440 vpslld ymm6,ymm2,30
5441 vpor ymm7,ymm7,ymm8
5442 vmovd xmm8,DWORD[((-40))+r8]
5443 vpaddd ymm0,ymm0,ymm5
5444
5445 vpsrld ymm2,ymm2,2
5446 vpaddd ymm0,ymm0,ymm7
5447 vpshufb ymm10,ymm10,ymm9
5448 vpor ymm2,ymm2,ymm6
5449 vmovd xmm7,DWORD[((-40))+r13]
5450 vmovd xmm6,DWORD[((-40))+r9]
5451 vpinsrd xmm11,xmm11,DWORD[((-40))+r14],1
5452 vpinsrd xmm8,xmm8,DWORD[((-40))+r10],1
5453 vpinsrd xmm7,xmm7,DWORD[((-40))+r15],1
5454 vpunpckldq ymm11,ymm11,ymm7
5455 vpinsrd xmm6,xmm6,DWORD[((-40))+r11],1
5456 vpunpckldq ymm8,ymm8,ymm6
5457 vpaddd ymm4,ymm4,ymm15
5458 vpslld ymm7,ymm0,5
5459 vpandn ymm6,ymm1,ymm3
5460 vpand ymm5,ymm1,ymm2
5461
5462 vmovdqa YMMWORD[(160-128)+rax],ymm10
5463 vpaddd ymm4,ymm4,ymm10
5464 vinserti128 ymm11,ymm11,xmm8,1
5465 vpsrld ymm8,ymm0,27
5466 vpxor ymm5,ymm5,ymm6
5467 vmovd xmm12,DWORD[((-36))+r12]
5468
5469 vpslld ymm6,ymm1,30
5470 vpor ymm7,ymm7,ymm8
5471 vmovd xmm8,DWORD[((-36))+r8]
5472 vpaddd ymm4,ymm4,ymm5
5473
5474 vpsrld ymm1,ymm1,2
5475 vpaddd ymm4,ymm4,ymm7
5476 vpshufb ymm11,ymm11,ymm9
5477 vpor ymm1,ymm1,ymm6
5478 vmovd xmm7,DWORD[((-36))+r13]
5479 vmovd xmm6,DWORD[((-36))+r9]
5480 vpinsrd xmm12,xmm12,DWORD[((-36))+r14],1
5481 vpinsrd xmm8,xmm8,DWORD[((-36))+r10],1
5482 vpinsrd xmm7,xmm7,DWORD[((-36))+r15],1
5483 vpunpckldq ymm12,ymm12,ymm7
5484 vpinsrd xmm6,xmm6,DWORD[((-36))+r11],1
5485 vpunpckldq ymm8,ymm8,ymm6
5486 vpaddd ymm3,ymm3,ymm15
5487 vpslld ymm7,ymm4,5
5488 vpandn ymm6,ymm0,ymm2
5489 vpand ymm5,ymm0,ymm1
5490
5491 vmovdqa YMMWORD[(192-128)+rax],ymm11
5492 vpaddd ymm3,ymm3,ymm11
5493 vinserti128 ymm12,ymm12,xmm8,1
5494 vpsrld ymm8,ymm4,27
5495 vpxor ymm5,ymm5,ymm6
5496 vmovd xmm13,DWORD[((-32))+r12]
5497
5498 vpslld ymm6,ymm0,30
5499 vpor ymm7,ymm7,ymm8
5500 vmovd xmm8,DWORD[((-32))+r8]
5501 vpaddd ymm3,ymm3,ymm5
5502
5503 vpsrld ymm0,ymm0,2
5504 vpaddd ymm3,ymm3,ymm7
5505 vpshufb ymm12,ymm12,ymm9
5506 vpor ymm0,ymm0,ymm6
5507 vmovd xmm7,DWORD[((-32))+r13]
5508 vmovd xmm6,DWORD[((-32))+r9]
5509 vpinsrd xmm13,xmm13,DWORD[((-32))+r14],1
5510 vpinsrd xmm8,xmm8,DWORD[((-32))+r10],1
5511 vpinsrd xmm7,xmm7,DWORD[((-32))+r15],1
5512 vpunpckldq ymm13,ymm13,ymm7
5513 vpinsrd xmm6,xmm6,DWORD[((-32))+r11],1
5514 vpunpckldq ymm8,ymm8,ymm6
5515 vpaddd ymm2,ymm2,ymm15
5516 vpslld ymm7,ymm3,5
5517 vpandn ymm6,ymm4,ymm1
5518 vpand ymm5,ymm4,ymm0
5519
5520 vmovdqa YMMWORD[(224-128)+rax],ymm12
5521 vpaddd ymm2,ymm2,ymm12
5522 vinserti128 ymm13,ymm13,xmm8,1
5523 vpsrld ymm8,ymm3,27
5524 vpxor ymm5,ymm5,ymm6
5525 vmovd xmm14,DWORD[((-28))+r12]
5526
5527 vpslld ymm6,ymm4,30
5528 vpor ymm7,ymm7,ymm8
5529 vmovd xmm8,DWORD[((-28))+r8]
5530 vpaddd ymm2,ymm2,ymm5
5531
5532 vpsrld ymm4,ymm4,2
5533 vpaddd ymm2,ymm2,ymm7
5534 vpshufb ymm13,ymm13,ymm9
5535 vpor ymm4,ymm4,ymm6
5536 vmovd xmm7,DWORD[((-28))+r13]
5537 vmovd xmm6,DWORD[((-28))+r9]
5538 vpinsrd xmm14,xmm14,DWORD[((-28))+r14],1
5539 vpinsrd xmm8,xmm8,DWORD[((-28))+r10],1
5540 vpinsrd xmm7,xmm7,DWORD[((-28))+r15],1
5541 vpunpckldq ymm14,ymm14,ymm7
5542 vpinsrd xmm6,xmm6,DWORD[((-28))+r11],1
5543 vpunpckldq ymm8,ymm8,ymm6
5544 vpaddd ymm1,ymm1,ymm15
5545 vpslld ymm7,ymm2,5
5546 vpandn ymm6,ymm3,ymm0
5547 vpand ymm5,ymm3,ymm4
5548
5549 vmovdqa YMMWORD[(256-256-128)+rbx],ymm13
5550 vpaddd ymm1,ymm1,ymm13
5551 vinserti128 ymm14,ymm14,xmm8,1
5552 vpsrld ymm8,ymm2,27
5553 vpxor ymm5,ymm5,ymm6
5554 vmovd xmm10,DWORD[((-24))+r12]
5555
5556 vpslld ymm6,ymm3,30
5557 vpor ymm7,ymm7,ymm8
5558 vmovd xmm8,DWORD[((-24))+r8]
5559 vpaddd ymm1,ymm1,ymm5
5560
5561 vpsrld ymm3,ymm3,2
5562 vpaddd ymm1,ymm1,ymm7
5563 vpshufb ymm14,ymm14,ymm9
5564 vpor ymm3,ymm3,ymm6
5565 vmovd xmm7,DWORD[((-24))+r13]
5566 vmovd xmm6,DWORD[((-24))+r9]
5567 vpinsrd xmm10,xmm10,DWORD[((-24))+r14],1
5568 vpinsrd xmm8,xmm8,DWORD[((-24))+r10],1
5569 vpinsrd xmm7,xmm7,DWORD[((-24))+r15],1
5570 vpunpckldq ymm10,ymm10,ymm7
5571 vpinsrd xmm6,xmm6,DWORD[((-24))+r11],1
5572 vpunpckldq ymm8,ymm8,ymm6
5573 vpaddd ymm0,ymm0,ymm15
5574 vpslld ymm7,ymm1,5
5575 vpandn ymm6,ymm2,ymm4
5576 vpand ymm5,ymm2,ymm3
5577
5578 vmovdqa YMMWORD[(288-256-128)+rbx],ymm14
5579 vpaddd ymm0,ymm0,ymm14
5580 vinserti128 ymm10,ymm10,xmm8,1
5581 vpsrld ymm8,ymm1,27
5582 vpxor ymm5,ymm5,ymm6
5583 vmovd xmm11,DWORD[((-20))+r12]
5584
5585 vpslld ymm6,ymm2,30
5586 vpor ymm7,ymm7,ymm8
5587 vmovd xmm8,DWORD[((-20))+r8]
5588 vpaddd ymm0,ymm0,ymm5
5589
5590 vpsrld ymm2,ymm2,2
5591 vpaddd ymm0,ymm0,ymm7
5592 vpshufb ymm10,ymm10,ymm9
5593 vpor ymm2,ymm2,ymm6
5594 vmovd xmm7,DWORD[((-20))+r13]
5595 vmovd xmm6,DWORD[((-20))+r9]
5596 vpinsrd xmm11,xmm11,DWORD[((-20))+r14],1
5597 vpinsrd xmm8,xmm8,DWORD[((-20))+r10],1
5598 vpinsrd xmm7,xmm7,DWORD[((-20))+r15],1
5599 vpunpckldq ymm11,ymm11,ymm7
5600 vpinsrd xmm6,xmm6,DWORD[((-20))+r11],1
5601 vpunpckldq ymm8,ymm8,ymm6
5602 vpaddd ymm4,ymm4,ymm15
5603 vpslld ymm7,ymm0,5
5604 vpandn ymm6,ymm1,ymm3
5605 vpand ymm5,ymm1,ymm2
5606
5607 vmovdqa YMMWORD[(320-256-128)+rbx],ymm10
5608 vpaddd ymm4,ymm4,ymm10
5609 vinserti128 ymm11,ymm11,xmm8,1
5610 vpsrld ymm8,ymm0,27
5611 vpxor ymm5,ymm5,ymm6
5612 vmovd xmm12,DWORD[((-16))+r12]
5613
5614 vpslld ymm6,ymm1,30
5615 vpor ymm7,ymm7,ymm8
5616 vmovd xmm8,DWORD[((-16))+r8]
5617 vpaddd ymm4,ymm4,ymm5
5618
5619 vpsrld ymm1,ymm1,2
5620 vpaddd ymm4,ymm4,ymm7
5621 vpshufb ymm11,ymm11,ymm9
5622 vpor ymm1,ymm1,ymm6
5623 vmovd xmm7,DWORD[((-16))+r13]
5624 vmovd xmm6,DWORD[((-16))+r9]
5625 vpinsrd xmm12,xmm12,DWORD[((-16))+r14],1
5626 vpinsrd xmm8,xmm8,DWORD[((-16))+r10],1
5627 vpinsrd xmm7,xmm7,DWORD[((-16))+r15],1
5628 vpunpckldq ymm12,ymm12,ymm7
5629 vpinsrd xmm6,xmm6,DWORD[((-16))+r11],1
5630 vpunpckldq ymm8,ymm8,ymm6
5631 vpaddd ymm3,ymm3,ymm15
5632 vpslld ymm7,ymm4,5
5633 vpandn ymm6,ymm0,ymm2
5634 vpand ymm5,ymm0,ymm1
5635
5636 vmovdqa YMMWORD[(352-256-128)+rbx],ymm11
5637 vpaddd ymm3,ymm3,ymm11
5638 vinserti128 ymm12,ymm12,xmm8,1
5639 vpsrld ymm8,ymm4,27
5640 vpxor ymm5,ymm5,ymm6
5641 vmovd xmm13,DWORD[((-12))+r12]
5642
5643 vpslld ymm6,ymm0,30
5644 vpor ymm7,ymm7,ymm8
5645 vmovd xmm8,DWORD[((-12))+r8]
5646 vpaddd ymm3,ymm3,ymm5
5647
5648 vpsrld ymm0,ymm0,2
5649 vpaddd ymm3,ymm3,ymm7
5650 vpshufb ymm12,ymm12,ymm9
5651 vpor ymm0,ymm0,ymm6
5652 vmovd xmm7,DWORD[((-12))+r13]
5653 vmovd xmm6,DWORD[((-12))+r9]
5654 vpinsrd xmm13,xmm13,DWORD[((-12))+r14],1
5655 vpinsrd xmm8,xmm8,DWORD[((-12))+r10],1
5656 vpinsrd xmm7,xmm7,DWORD[((-12))+r15],1
5657 vpunpckldq ymm13,ymm13,ymm7
5658 vpinsrd xmm6,xmm6,DWORD[((-12))+r11],1
5659 vpunpckldq ymm8,ymm8,ymm6
5660 vpaddd ymm2,ymm2,ymm15
5661 vpslld ymm7,ymm3,5
5662 vpandn ymm6,ymm4,ymm1
5663 vpand ymm5,ymm4,ymm0
5664
5665 vmovdqa YMMWORD[(384-256-128)+rbx],ymm12
5666 vpaddd ymm2,ymm2,ymm12
5667 vinserti128 ymm13,ymm13,xmm8,1
5668 vpsrld ymm8,ymm3,27
5669 vpxor ymm5,ymm5,ymm6
5670 vmovd xmm14,DWORD[((-8))+r12]
5671
5672 vpslld ymm6,ymm4,30
5673 vpor ymm7,ymm7,ymm8
5674 vmovd xmm8,DWORD[((-8))+r8]
5675 vpaddd ymm2,ymm2,ymm5
5676
5677 vpsrld ymm4,ymm4,2
5678 vpaddd ymm2,ymm2,ymm7
5679 vpshufb ymm13,ymm13,ymm9
5680 vpor ymm4,ymm4,ymm6
5681 vmovd xmm7,DWORD[((-8))+r13]
5682 vmovd xmm6,DWORD[((-8))+r9]
5683 vpinsrd xmm14,xmm14,DWORD[((-8))+r14],1
5684 vpinsrd xmm8,xmm8,DWORD[((-8))+r10],1
5685 vpinsrd xmm7,xmm7,DWORD[((-8))+r15],1
5686 vpunpckldq ymm14,ymm14,ymm7
5687 vpinsrd xmm6,xmm6,DWORD[((-8))+r11],1
5688 vpunpckldq ymm8,ymm8,ymm6
5689 vpaddd ymm1,ymm1,ymm15
5690 vpslld ymm7,ymm2,5
5691 vpandn ymm6,ymm3,ymm0
5692 vpand ymm5,ymm3,ymm4
5693
5694 vmovdqa YMMWORD[(416-256-128)+rbx],ymm13
5695 vpaddd ymm1,ymm1,ymm13
5696 vinserti128 ymm14,ymm14,xmm8,1
5697 vpsrld ymm8,ymm2,27
5698 vpxor ymm5,ymm5,ymm6
5699 vmovd xmm10,DWORD[((-4))+r12]
5700
5701 vpslld ymm6,ymm3,30
5702 vpor ymm7,ymm7,ymm8
5703 vmovd xmm8,DWORD[((-4))+r8]
5704 vpaddd ymm1,ymm1,ymm5
5705
5706 vpsrld ymm3,ymm3,2
5707 vpaddd ymm1,ymm1,ymm7
5708 vpshufb ymm14,ymm14,ymm9
5709 vpor ymm3,ymm3,ymm6
5710 vmovdqa ymm11,YMMWORD[((0-128))+rax]
5711 vmovd xmm7,DWORD[((-4))+r13]
5712 vmovd xmm6,DWORD[((-4))+r9]
5713 vpinsrd xmm10,xmm10,DWORD[((-4))+r14],1
5714 vpinsrd xmm8,xmm8,DWORD[((-4))+r10],1
5715 vpinsrd xmm7,xmm7,DWORD[((-4))+r15],1
5716 vpunpckldq ymm10,ymm10,ymm7
5717 vpinsrd xmm6,xmm6,DWORD[((-4))+r11],1
5718 vpunpckldq ymm8,ymm8,ymm6
5719 vpaddd ymm0,ymm0,ymm15
5720 prefetcht0 [63+r12]
5721 vpslld ymm7,ymm1,5
5722 vpandn ymm6,ymm2,ymm4
5723 vpand ymm5,ymm2,ymm3
5724
5725 vmovdqa YMMWORD[(448-256-128)+rbx],ymm14
5726 vpaddd ymm0,ymm0,ymm14
5727 vinserti128 ymm10,ymm10,xmm8,1
5728 vpsrld ymm8,ymm1,27
5729 prefetcht0 [63+r13]
5730 vpxor ymm5,ymm5,ymm6
5731
5732 vpslld ymm6,ymm2,30
5733 vpor ymm7,ymm7,ymm8
5734 prefetcht0 [63+r14]
5735 vpaddd ymm0,ymm0,ymm5
5736
5737 vpsrld ymm2,ymm2,2
5738 vpaddd ymm0,ymm0,ymm7
5739 prefetcht0 [63+r15]
5740 vpshufb ymm10,ymm10,ymm9
5741 vpor ymm2,ymm2,ymm6
5742 vmovdqa ymm12,YMMWORD[((32-128))+rax]
5743 vpxor ymm11,ymm11,ymm13
5744 vmovdqa ymm13,YMMWORD[((64-128))+rax]
5745
5746 vpaddd ymm4,ymm4,ymm15
5747 vpslld ymm7,ymm0,5
5748 vpandn ymm6,ymm1,ymm3
5749 prefetcht0 [63+r8]
5750 vpand ymm5,ymm1,ymm2
5751
5752 vmovdqa YMMWORD[(480-256-128)+rbx],ymm10
5753 vpaddd ymm4,ymm4,ymm10
5754 vpxor ymm11,ymm11,YMMWORD[((256-256-128))+rbx]
5755 vpsrld ymm8,ymm0,27
5756 vpxor ymm5,ymm5,ymm6
5757 vpxor ymm11,ymm11,ymm13
5758 prefetcht0 [63+r9]
5759
5760 vpslld ymm6,ymm1,30
5761 vpor ymm7,ymm7,ymm8
5762 vpaddd ymm4,ymm4,ymm5
5763 prefetcht0 [63+r10]
5764 vpsrld ymm9,ymm11,31
5765 vpaddd ymm11,ymm11,ymm11
5766
5767 vpsrld ymm1,ymm1,2
5768 prefetcht0 [63+r11]
5769 vpaddd ymm4,ymm4,ymm7
5770 vpor ymm11,ymm11,ymm9
5771 vpor ymm1,ymm1,ymm6
5772 vpxor ymm12,ymm12,ymm14
5773 vmovdqa ymm14,YMMWORD[((96-128))+rax]
5774
5775 vpaddd ymm3,ymm3,ymm15
5776 vpslld ymm7,ymm4,5
5777 vpandn ymm6,ymm0,ymm2
5778
5779 vpand ymm5,ymm0,ymm1
5780
5781 vmovdqa YMMWORD[(0-128)+rax],ymm11
5782 vpaddd ymm3,ymm3,ymm11
5783 vpxor ymm12,ymm12,YMMWORD[((288-256-128))+rbx]
5784 vpsrld ymm8,ymm4,27
5785 vpxor ymm5,ymm5,ymm6
5786 vpxor ymm12,ymm12,ymm14
5787
5788
5789 vpslld ymm6,ymm0,30
5790 vpor ymm7,ymm7,ymm8
5791 vpaddd ymm3,ymm3,ymm5
5792
5793 vpsrld ymm9,ymm12,31
5794 vpaddd ymm12,ymm12,ymm12
5795
5796 vpsrld ymm0,ymm0,2
5797
5798 vpaddd ymm3,ymm3,ymm7
5799 vpor ymm12,ymm12,ymm9
5800 vpor ymm0,ymm0,ymm6
5801 vpxor ymm13,ymm13,ymm10
5802 vmovdqa ymm10,YMMWORD[((128-128))+rax]
5803
5804 vpaddd ymm2,ymm2,ymm15
5805 vpslld ymm7,ymm3,5
5806 vpandn ymm6,ymm4,ymm1
5807
5808 vpand ymm5,ymm4,ymm0
5809
5810 vmovdqa YMMWORD[(32-128)+rax],ymm12
5811 vpaddd ymm2,ymm2,ymm12
5812 vpxor ymm13,ymm13,YMMWORD[((320-256-128))+rbx]
5813 vpsrld ymm8,ymm3,27
5814 vpxor ymm5,ymm5,ymm6
5815 vpxor ymm13,ymm13,ymm10
5816
5817
5818 vpslld ymm6,ymm4,30
5819 vpor ymm7,ymm7,ymm8
5820 vpaddd ymm2,ymm2,ymm5
5821
5822 vpsrld ymm9,ymm13,31
5823 vpaddd ymm13,ymm13,ymm13
5824
5825 vpsrld ymm4,ymm4,2
5826
5827 vpaddd ymm2,ymm2,ymm7
5828 vpor ymm13,ymm13,ymm9
5829 vpor ymm4,ymm4,ymm6
5830 vpxor ymm14,ymm14,ymm11
5831 vmovdqa ymm11,YMMWORD[((160-128))+rax]
5832
5833 vpaddd ymm1,ymm1,ymm15
5834 vpslld ymm7,ymm2,5
5835 vpandn ymm6,ymm3,ymm0
5836
5837 vpand ymm5,ymm3,ymm4
5838
5839 vmovdqa YMMWORD[(64-128)+rax],ymm13
5840 vpaddd ymm1,ymm1,ymm13
5841 vpxor ymm14,ymm14,YMMWORD[((352-256-128))+rbx]
5842 vpsrld ymm8,ymm2,27
5843 vpxor ymm5,ymm5,ymm6
5844 vpxor ymm14,ymm14,ymm11
5845
5846
5847 vpslld ymm6,ymm3,30
5848 vpor ymm7,ymm7,ymm8
5849 vpaddd ymm1,ymm1,ymm5
5850
5851 vpsrld ymm9,ymm14,31
5852 vpaddd ymm14,ymm14,ymm14
5853
5854 vpsrld ymm3,ymm3,2
5855
5856 vpaddd ymm1,ymm1,ymm7
5857 vpor ymm14,ymm14,ymm9
5858 vpor ymm3,ymm3,ymm6
5859 vpxor ymm10,ymm10,ymm12
5860 vmovdqa ymm12,YMMWORD[((192-128))+rax]
5861
5862 vpaddd ymm0,ymm0,ymm15
5863 vpslld ymm7,ymm1,5
5864 vpandn ymm6,ymm2,ymm4
5865
5866 vpand ymm5,ymm2,ymm3
5867
5868 vmovdqa YMMWORD[(96-128)+rax],ymm14
5869 vpaddd ymm0,ymm0,ymm14
5870 vpxor ymm10,ymm10,YMMWORD[((384-256-128))+rbx]
5871 vpsrld ymm8,ymm1,27
5872 vpxor ymm5,ymm5,ymm6
5873 vpxor ymm10,ymm10,ymm12
5874
5875
5876 vpslld ymm6,ymm2,30
5877 vpor ymm7,ymm7,ymm8
5878 vpaddd ymm0,ymm0,ymm5
5879
5880 vpsrld ymm9,ymm10,31
5881 vpaddd ymm10,ymm10,ymm10
5882
5883 vpsrld ymm2,ymm2,2
5884
5885 vpaddd ymm0,ymm0,ymm7
5886 vpor ymm10,ymm10,ymm9
5887 vpor ymm2,ymm2,ymm6
5888 vmovdqa ymm15,YMMWORD[rbp]
5889 vpxor ymm11,ymm11,ymm13
5890 vmovdqa ymm13,YMMWORD[((224-128))+rax]
5891
5892 vpslld ymm7,ymm0,5
5893 vpaddd ymm4,ymm4,ymm15
5894 vpxor ymm5,ymm3,ymm1
5895 vmovdqa YMMWORD[(128-128)+rax],ymm10
5896 vpaddd ymm4,ymm4,ymm10
5897 vpxor ymm11,ymm11,YMMWORD[((416-256-128))+rbx]
5898 vpsrld ymm8,ymm0,27
5899 vpxor ymm5,ymm5,ymm2
5900 vpxor ymm11,ymm11,ymm13
5901
5902 vpslld ymm6,ymm1,30
5903 vpor ymm7,ymm7,ymm8
5904 vpaddd ymm4,ymm4,ymm5
5905 vpsrld ymm9,ymm11,31
5906 vpaddd ymm11,ymm11,ymm11
5907
5908 vpsrld ymm1,ymm1,2
5909 vpaddd ymm4,ymm4,ymm7
5910 vpor ymm11,ymm11,ymm9
5911 vpor ymm1,ymm1,ymm6
5912 vpxor ymm12,ymm12,ymm14
5913 vmovdqa ymm14,YMMWORD[((256-256-128))+rbx]
5914
5915 vpslld ymm7,ymm4,5
5916 vpaddd ymm3,ymm3,ymm15
5917 vpxor ymm5,ymm2,ymm0
5918 vmovdqa YMMWORD[(160-128)+rax],ymm11
5919 vpaddd ymm3,ymm3,ymm11
5920 vpxor ymm12,ymm12,YMMWORD[((448-256-128))+rbx]
5921 vpsrld ymm8,ymm4,27
5922 vpxor ymm5,ymm5,ymm1
5923 vpxor ymm12,ymm12,ymm14
5924
5925 vpslld ymm6,ymm0,30
5926 vpor ymm7,ymm7,ymm8
5927 vpaddd ymm3,ymm3,ymm5
5928 vpsrld ymm9,ymm12,31
5929 vpaddd ymm12,ymm12,ymm12
5930
5931 vpsrld ymm0,ymm0,2
5932 vpaddd ymm3,ymm3,ymm7
5933 vpor ymm12,ymm12,ymm9
5934 vpor ymm0,ymm0,ymm6
5935 vpxor ymm13,ymm13,ymm10
5936 vmovdqa ymm10,YMMWORD[((288-256-128))+rbx]
5937
5938 vpslld ymm7,ymm3,5
5939 vpaddd ymm2,ymm2,ymm15
5940 vpxor ymm5,ymm1,ymm4
5941 vmovdqa YMMWORD[(192-128)+rax],ymm12
5942 vpaddd ymm2,ymm2,ymm12
5943 vpxor ymm13,ymm13,YMMWORD[((480-256-128))+rbx]
5944 vpsrld ymm8,ymm3,27
5945 vpxor ymm5,ymm5,ymm0
5946 vpxor ymm13,ymm13,ymm10
5947
5948 vpslld ymm6,ymm4,30
5949 vpor ymm7,ymm7,ymm8
5950 vpaddd ymm2,ymm2,ymm5
5951 vpsrld ymm9,ymm13,31
5952 vpaddd ymm13,ymm13,ymm13
5953
5954 vpsrld ymm4,ymm4,2
5955 vpaddd ymm2,ymm2,ymm7
5956 vpor ymm13,ymm13,ymm9
5957 vpor ymm4,ymm4,ymm6
5958 vpxor ymm14,ymm14,ymm11
5959 vmovdqa ymm11,YMMWORD[((320-256-128))+rbx]
5960
5961 vpslld ymm7,ymm2,5
5962 vpaddd ymm1,ymm1,ymm15
5963 vpxor ymm5,ymm0,ymm3
5964 vmovdqa YMMWORD[(224-128)+rax],ymm13
5965 vpaddd ymm1,ymm1,ymm13
5966 vpxor ymm14,ymm14,YMMWORD[((0-128))+rax]
5967 vpsrld ymm8,ymm2,27
5968 vpxor ymm5,ymm5,ymm4
5969 vpxor ymm14,ymm14,ymm11
5970
5971 vpslld ymm6,ymm3,30
5972 vpor ymm7,ymm7,ymm8
5973 vpaddd ymm1,ymm1,ymm5
5974 vpsrld ymm9,ymm14,31
5975 vpaddd ymm14,ymm14,ymm14
5976
5977 vpsrld ymm3,ymm3,2
5978 vpaddd ymm1,ymm1,ymm7
5979 vpor ymm14,ymm14,ymm9
5980 vpor ymm3,ymm3,ymm6
5981 vpxor ymm10,ymm10,ymm12
5982 vmovdqa ymm12,YMMWORD[((352-256-128))+rbx]
5983
5984 vpslld ymm7,ymm1,5
5985 vpaddd ymm0,ymm0,ymm15
5986 vpxor ymm5,ymm4,ymm2
5987 vmovdqa YMMWORD[(256-256-128)+rbx],ymm14
5988 vpaddd ymm0,ymm0,ymm14
5989 vpxor ymm10,ymm10,YMMWORD[((32-128))+rax]
5990 vpsrld ymm8,ymm1,27
5991 vpxor ymm5,ymm5,ymm3
5992 vpxor ymm10,ymm10,ymm12
5993
5994 vpslld ymm6,ymm2,30
5995 vpor ymm7,ymm7,ymm8
5996 vpaddd ymm0,ymm0,ymm5
5997 vpsrld ymm9,ymm10,31
5998 vpaddd ymm10,ymm10,ymm10
5999
6000 vpsrld ymm2,ymm2,2
6001 vpaddd ymm0,ymm0,ymm7
6002 vpor ymm10,ymm10,ymm9
6003 vpor ymm2,ymm2,ymm6
6004 vpxor ymm11,ymm11,ymm13
6005 vmovdqa ymm13,YMMWORD[((384-256-128))+rbx]
6006
6007 vpslld ymm7,ymm0,5
6008 vpaddd ymm4,ymm4,ymm15
6009 vpxor ymm5,ymm3,ymm1
6010 vmovdqa YMMWORD[(288-256-128)+rbx],ymm10
6011 vpaddd ymm4,ymm4,ymm10
6012 vpxor ymm11,ymm11,YMMWORD[((64-128))+rax]
6013 vpsrld ymm8,ymm0,27
6014 vpxor ymm5,ymm5,ymm2
6015 vpxor ymm11,ymm11,ymm13
6016
6017 vpslld ymm6,ymm1,30
6018 vpor ymm7,ymm7,ymm8
6019 vpaddd ymm4,ymm4,ymm5
6020 vpsrld ymm9,ymm11,31
6021 vpaddd ymm11,ymm11,ymm11
6022
6023 vpsrld ymm1,ymm1,2
6024 vpaddd ymm4,ymm4,ymm7
6025 vpor ymm11,ymm11,ymm9
6026 vpor ymm1,ymm1,ymm6
6027 vpxor ymm12,ymm12,ymm14
6028 vmovdqa ymm14,YMMWORD[((416-256-128))+rbx]
6029
6030 vpslld ymm7,ymm4,5
6031 vpaddd ymm3,ymm3,ymm15
6032 vpxor ymm5,ymm2,ymm0
6033 vmovdqa YMMWORD[(320-256-128)+rbx],ymm11
6034 vpaddd ymm3,ymm3,ymm11
6035 vpxor ymm12,ymm12,YMMWORD[((96-128))+rax]
6036 vpsrld ymm8,ymm4,27
6037 vpxor ymm5,ymm5,ymm1
6038 vpxor ymm12,ymm12,ymm14
6039
6040 vpslld ymm6,ymm0,30
6041 vpor ymm7,ymm7,ymm8
6042 vpaddd ymm3,ymm3,ymm5
6043 vpsrld ymm9,ymm12,31
6044 vpaddd ymm12,ymm12,ymm12
6045
6046 vpsrld ymm0,ymm0,2
6047 vpaddd ymm3,ymm3,ymm7
6048 vpor ymm12,ymm12,ymm9
6049 vpor ymm0,ymm0,ymm6
6050 vpxor ymm13,ymm13,ymm10
6051 vmovdqa ymm10,YMMWORD[((448-256-128))+rbx]
6052
6053 vpslld ymm7,ymm3,5
6054 vpaddd ymm2,ymm2,ymm15
6055 vpxor ymm5,ymm1,ymm4
6056 vmovdqa YMMWORD[(352-256-128)+rbx],ymm12
6057 vpaddd ymm2,ymm2,ymm12
6058 vpxor ymm13,ymm13,YMMWORD[((128-128))+rax]
6059 vpsrld ymm8,ymm3,27
6060 vpxor ymm5,ymm5,ymm0
6061 vpxor ymm13,ymm13,ymm10
6062
6063 vpslld ymm6,ymm4,30
6064 vpor ymm7,ymm7,ymm8
6065 vpaddd ymm2,ymm2,ymm5
6066 vpsrld ymm9,ymm13,31
6067 vpaddd ymm13,ymm13,ymm13
6068
6069 vpsrld ymm4,ymm4,2
6070 vpaddd ymm2,ymm2,ymm7
6071 vpor ymm13,ymm13,ymm9
6072 vpor ymm4,ymm4,ymm6
6073 vpxor ymm14,ymm14,ymm11
6074 vmovdqa ymm11,YMMWORD[((480-256-128))+rbx]
6075
6076 vpslld ymm7,ymm2,5
6077 vpaddd ymm1,ymm1,ymm15
6078 vpxor ymm5,ymm0,ymm3
6079 vmovdqa YMMWORD[(384-256-128)+rbx],ymm13
6080 vpaddd ymm1,ymm1,ymm13
6081 vpxor ymm14,ymm14,YMMWORD[((160-128))+rax]
6082 vpsrld ymm8,ymm2,27
6083 vpxor ymm5,ymm5,ymm4
6084 vpxor ymm14,ymm14,ymm11
6085
6086 vpslld ymm6,ymm3,30
6087 vpor ymm7,ymm7,ymm8
6088 vpaddd ymm1,ymm1,ymm5
6089 vpsrld ymm9,ymm14,31
6090 vpaddd ymm14,ymm14,ymm14
6091
6092 vpsrld ymm3,ymm3,2
6093 vpaddd ymm1,ymm1,ymm7
6094 vpor ymm14,ymm14,ymm9
6095 vpor ymm3,ymm3,ymm6
6096 vpxor ymm10,ymm10,ymm12
6097 vmovdqa ymm12,YMMWORD[((0-128))+rax]
6098
6099 vpslld ymm7,ymm1,5
6100 vpaddd ymm0,ymm0,ymm15
6101 vpxor ymm5,ymm4,ymm2
6102 vmovdqa YMMWORD[(416-256-128)+rbx],ymm14
6103 vpaddd ymm0,ymm0,ymm14
6104 vpxor ymm10,ymm10,YMMWORD[((192-128))+rax]
6105 vpsrld ymm8,ymm1,27
6106 vpxor ymm5,ymm5,ymm3
6107 vpxor ymm10,ymm10,ymm12
6108
6109 vpslld ymm6,ymm2,30
6110 vpor ymm7,ymm7,ymm8
6111 vpaddd ymm0,ymm0,ymm5
6112 vpsrld ymm9,ymm10,31
6113 vpaddd ymm10,ymm10,ymm10
6114
6115 vpsrld ymm2,ymm2,2
6116 vpaddd ymm0,ymm0,ymm7
6117 vpor ymm10,ymm10,ymm9
6118 vpor ymm2,ymm2,ymm6
6119 vpxor ymm11,ymm11,ymm13
6120 vmovdqa ymm13,YMMWORD[((32-128))+rax]
6121
6122 vpslld ymm7,ymm0,5
6123 vpaddd ymm4,ymm4,ymm15
6124 vpxor ymm5,ymm3,ymm1
6125 vmovdqa YMMWORD[(448-256-128)+rbx],ymm10
6126 vpaddd ymm4,ymm4,ymm10
6127 vpxor ymm11,ymm11,YMMWORD[((224-128))+rax]
6128 vpsrld ymm8,ymm0,27
6129 vpxor ymm5,ymm5,ymm2
6130 vpxor ymm11,ymm11,ymm13
6131
6132 vpslld ymm6,ymm1,30
6133 vpor ymm7,ymm7,ymm8
6134 vpaddd ymm4,ymm4,ymm5
6135 vpsrld ymm9,ymm11,31
6136 vpaddd ymm11,ymm11,ymm11
6137
6138 vpsrld ymm1,ymm1,2
6139 vpaddd ymm4,ymm4,ymm7
6140 vpor ymm11,ymm11,ymm9
6141 vpor ymm1,ymm1,ymm6
6142 vpxor ymm12,ymm12,ymm14
6143 vmovdqa ymm14,YMMWORD[((64-128))+rax]
6144
6145 vpslld ymm7,ymm4,5
6146 vpaddd ymm3,ymm3,ymm15
6147 vpxor ymm5,ymm2,ymm0
6148 vmovdqa YMMWORD[(480-256-128)+rbx],ymm11
6149 vpaddd ymm3,ymm3,ymm11
6150 vpxor ymm12,ymm12,YMMWORD[((256-256-128))+rbx]
6151 vpsrld ymm8,ymm4,27
6152 vpxor ymm5,ymm5,ymm1
6153 vpxor ymm12,ymm12,ymm14
6154
6155 vpslld ymm6,ymm0,30
6156 vpor ymm7,ymm7,ymm8
6157 vpaddd ymm3,ymm3,ymm5
6158 vpsrld ymm9,ymm12,31
6159 vpaddd ymm12,ymm12,ymm12
6160
6161 vpsrld ymm0,ymm0,2
6162 vpaddd ymm3,ymm3,ymm7
6163 vpor ymm12,ymm12,ymm9
6164 vpor ymm0,ymm0,ymm6
6165 vpxor ymm13,ymm13,ymm10
6166 vmovdqa ymm10,YMMWORD[((96-128))+rax]
6167
6168 vpslld ymm7,ymm3,5
6169 vpaddd ymm2,ymm2,ymm15
6170 vpxor ymm5,ymm1,ymm4
6171 vmovdqa YMMWORD[(0-128)+rax],ymm12
6172 vpaddd ymm2,ymm2,ymm12
6173 vpxor ymm13,ymm13,YMMWORD[((288-256-128))+rbx]
6174 vpsrld ymm8,ymm3,27
6175 vpxor ymm5,ymm5,ymm0
6176 vpxor ymm13,ymm13,ymm10
6177
6178 vpslld ymm6,ymm4,30
6179 vpor ymm7,ymm7,ymm8
6180 vpaddd ymm2,ymm2,ymm5
6181 vpsrld ymm9,ymm13,31
6182 vpaddd ymm13,ymm13,ymm13
6183
6184 vpsrld ymm4,ymm4,2
6185 vpaddd ymm2,ymm2,ymm7
6186 vpor ymm13,ymm13,ymm9
6187 vpor ymm4,ymm4,ymm6
6188 vpxor ymm14,ymm14,ymm11
6189 vmovdqa ymm11,YMMWORD[((128-128))+rax]
6190
6191 vpslld ymm7,ymm2,5
6192 vpaddd ymm1,ymm1,ymm15
6193 vpxor ymm5,ymm0,ymm3
6194 vmovdqa YMMWORD[(32-128)+rax],ymm13
6195 vpaddd ymm1,ymm1,ymm13
6196 vpxor ymm14,ymm14,YMMWORD[((320-256-128))+rbx]
6197 vpsrld ymm8,ymm2,27
6198 vpxor ymm5,ymm5,ymm4
6199 vpxor ymm14,ymm14,ymm11
6200
6201 vpslld ymm6,ymm3,30
6202 vpor ymm7,ymm7,ymm8
6203 vpaddd ymm1,ymm1,ymm5
6204 vpsrld ymm9,ymm14,31
6205 vpaddd ymm14,ymm14,ymm14
6206
6207 vpsrld ymm3,ymm3,2
6208 vpaddd ymm1,ymm1,ymm7
6209 vpor ymm14,ymm14,ymm9
6210 vpor ymm3,ymm3,ymm6
6211 vpxor ymm10,ymm10,ymm12
6212 vmovdqa ymm12,YMMWORD[((160-128))+rax]
6213
6214 vpslld ymm7,ymm1,5
6215 vpaddd ymm0,ymm0,ymm15
6216 vpxor ymm5,ymm4,ymm2
6217 vmovdqa YMMWORD[(64-128)+rax],ymm14
6218 vpaddd ymm0,ymm0,ymm14
6219 vpxor ymm10,ymm10,YMMWORD[((352-256-128))+rbx]
6220 vpsrld ymm8,ymm1,27
6221 vpxor ymm5,ymm5,ymm3
6222 vpxor ymm10,ymm10,ymm12
6223
6224 vpslld ymm6,ymm2,30
6225 vpor ymm7,ymm7,ymm8
6226 vpaddd ymm0,ymm0,ymm5
6227 vpsrld ymm9,ymm10,31
6228 vpaddd ymm10,ymm10,ymm10
6229
6230 vpsrld ymm2,ymm2,2
6231 vpaddd ymm0,ymm0,ymm7
6232 vpor ymm10,ymm10,ymm9
6233 vpor ymm2,ymm2,ymm6
6234 vpxor ymm11,ymm11,ymm13
6235 vmovdqa ymm13,YMMWORD[((192-128))+rax]
6236
6237 vpslld ymm7,ymm0,5
6238 vpaddd ymm4,ymm4,ymm15
6239 vpxor ymm5,ymm3,ymm1
6240 vmovdqa YMMWORD[(96-128)+rax],ymm10
6241 vpaddd ymm4,ymm4,ymm10
6242 vpxor ymm11,ymm11,YMMWORD[((384-256-128))+rbx]
6243 vpsrld ymm8,ymm0,27
6244 vpxor ymm5,ymm5,ymm2
6245 vpxor ymm11,ymm11,ymm13
6246
6247 vpslld ymm6,ymm1,30
6248 vpor ymm7,ymm7,ymm8
6249 vpaddd ymm4,ymm4,ymm5
6250 vpsrld ymm9,ymm11,31
6251 vpaddd ymm11,ymm11,ymm11
6252
6253 vpsrld ymm1,ymm1,2
6254 vpaddd ymm4,ymm4,ymm7
6255 vpor ymm11,ymm11,ymm9
6256 vpor ymm1,ymm1,ymm6
6257 vpxor ymm12,ymm12,ymm14
6258 vmovdqa ymm14,YMMWORD[((224-128))+rax]
6259
6260 vpslld ymm7,ymm4,5
6261 vpaddd ymm3,ymm3,ymm15
6262 vpxor ymm5,ymm2,ymm0
6263 vmovdqa YMMWORD[(128-128)+rax],ymm11
6264 vpaddd ymm3,ymm3,ymm11
6265 vpxor ymm12,ymm12,YMMWORD[((416-256-128))+rbx]
6266 vpsrld ymm8,ymm4,27
6267 vpxor ymm5,ymm5,ymm1
6268 vpxor ymm12,ymm12,ymm14
6269
6270 vpslld ymm6,ymm0,30
6271 vpor ymm7,ymm7,ymm8
6272 vpaddd ymm3,ymm3,ymm5
6273 vpsrld ymm9,ymm12,31
6274 vpaddd ymm12,ymm12,ymm12
6275
6276 vpsrld ymm0,ymm0,2
6277 vpaddd ymm3,ymm3,ymm7
6278 vpor ymm12,ymm12,ymm9
6279 vpor ymm0,ymm0,ymm6
6280 vpxor ymm13,ymm13,ymm10
6281 vmovdqa ymm10,YMMWORD[((256-256-128))+rbx]
6282
6283 vpslld ymm7,ymm3,5
6284 vpaddd ymm2,ymm2,ymm15
6285 vpxor ymm5,ymm1,ymm4
6286 vmovdqa YMMWORD[(160-128)+rax],ymm12
6287 vpaddd ymm2,ymm2,ymm12
6288 vpxor ymm13,ymm13,YMMWORD[((448-256-128))+rbx]
6289 vpsrld ymm8,ymm3,27
6290 vpxor ymm5,ymm5,ymm0
6291 vpxor ymm13,ymm13,ymm10
6292
6293 vpslld ymm6,ymm4,30
6294 vpor ymm7,ymm7,ymm8
6295 vpaddd ymm2,ymm2,ymm5
6296 vpsrld ymm9,ymm13,31
6297 vpaddd ymm13,ymm13,ymm13
6298
6299 vpsrld ymm4,ymm4,2
6300 vpaddd ymm2,ymm2,ymm7
6301 vpor ymm13,ymm13,ymm9
6302 vpor ymm4,ymm4,ymm6
6303 vpxor ymm14,ymm14,ymm11
6304 vmovdqa ymm11,YMMWORD[((288-256-128))+rbx]
6305
6306 vpslld ymm7,ymm2,5
6307 vpaddd ymm1,ymm1,ymm15
6308 vpxor ymm5,ymm0,ymm3
6309 vmovdqa YMMWORD[(192-128)+rax],ymm13
6310 vpaddd ymm1,ymm1,ymm13
6311 vpxor ymm14,ymm14,YMMWORD[((480-256-128))+rbx]
6312 vpsrld ymm8,ymm2,27
6313 vpxor ymm5,ymm5,ymm4
6314 vpxor ymm14,ymm14,ymm11
6315
6316 vpslld ymm6,ymm3,30
6317 vpor ymm7,ymm7,ymm8
6318 vpaddd ymm1,ymm1,ymm5
6319 vpsrld ymm9,ymm14,31
6320 vpaddd ymm14,ymm14,ymm14
6321
6322 vpsrld ymm3,ymm3,2
6323 vpaddd ymm1,ymm1,ymm7
6324 vpor ymm14,ymm14,ymm9
6325 vpor ymm3,ymm3,ymm6
6326 vpxor ymm10,ymm10,ymm12
6327 vmovdqa ymm12,YMMWORD[((320-256-128))+rbx]
6328
6329 vpslld ymm7,ymm1,5
6330 vpaddd ymm0,ymm0,ymm15
6331 vpxor ymm5,ymm4,ymm2
6332 vmovdqa YMMWORD[(224-128)+rax],ymm14
6333 vpaddd ymm0,ymm0,ymm14
6334 vpxor ymm10,ymm10,YMMWORD[((0-128))+rax]
6335 vpsrld ymm8,ymm1,27
6336 vpxor ymm5,ymm5,ymm3
6337 vpxor ymm10,ymm10,ymm12
6338
6339 vpslld ymm6,ymm2,30
6340 vpor ymm7,ymm7,ymm8
6341 vpaddd ymm0,ymm0,ymm5
6342 vpsrld ymm9,ymm10,31
6343 vpaddd ymm10,ymm10,ymm10
6344
6345 vpsrld ymm2,ymm2,2
6346 vpaddd ymm0,ymm0,ymm7
6347 vpor ymm10,ymm10,ymm9
6348 vpor ymm2,ymm2,ymm6
6349 vmovdqa ymm15,YMMWORD[32+rbp]
6350 vpxor ymm11,ymm11,ymm13
6351 vmovdqa ymm13,YMMWORD[((352-256-128))+rbx]
6352
6353 vpaddd ymm4,ymm4,ymm15
6354 vpslld ymm7,ymm0,5
6355 vpand ymm6,ymm3,ymm2
6356 vpxor ymm11,ymm11,YMMWORD[((32-128))+rax]
6357
6358 vpaddd ymm4,ymm4,ymm6
6359 vpsrld ymm8,ymm0,27
6360 vpxor ymm5,ymm3,ymm2
6361 vpxor ymm11,ymm11,ymm13
6362
6363 vmovdqu YMMWORD[(256-256-128)+rbx],ymm10
6364 vpaddd ymm4,ymm4,ymm10
6365 vpor ymm7,ymm7,ymm8
6366 vpsrld ymm9,ymm11,31
6367 vpand ymm5,ymm5,ymm1
6368 vpaddd ymm11,ymm11,ymm11
6369
6370 vpslld ymm6,ymm1,30
6371 vpaddd ymm4,ymm4,ymm5
6372
6373 vpsrld ymm1,ymm1,2
6374 vpaddd ymm4,ymm4,ymm7
6375 vpor ymm11,ymm11,ymm9
6376 vpor ymm1,ymm1,ymm6
6377 vpxor ymm12,ymm12,ymm14
6378 vmovdqa ymm14,YMMWORD[((384-256-128))+rbx]
6379
6380 vpaddd ymm3,ymm3,ymm15
6381 vpslld ymm7,ymm4,5
6382 vpand ymm6,ymm2,ymm1
6383 vpxor ymm12,ymm12,YMMWORD[((64-128))+rax]
6384
6385 vpaddd ymm3,ymm3,ymm6
6386 vpsrld ymm8,ymm4,27
6387 vpxor ymm5,ymm2,ymm1
6388 vpxor ymm12,ymm12,ymm14
6389
6390 vmovdqu YMMWORD[(288-256-128)+rbx],ymm11
6391 vpaddd ymm3,ymm3,ymm11
6392 vpor ymm7,ymm7,ymm8
6393 vpsrld ymm9,ymm12,31
6394 vpand ymm5,ymm5,ymm0
6395 vpaddd ymm12,ymm12,ymm12
6396
6397 vpslld ymm6,ymm0,30
6398 vpaddd ymm3,ymm3,ymm5
6399
6400 vpsrld ymm0,ymm0,2
6401 vpaddd ymm3,ymm3,ymm7
6402 vpor ymm12,ymm12,ymm9
6403 vpor ymm0,ymm0,ymm6
6404 vpxor ymm13,ymm13,ymm10
6405 vmovdqa ymm10,YMMWORD[((416-256-128))+rbx]
6406
6407 vpaddd ymm2,ymm2,ymm15
6408 vpslld ymm7,ymm3,5
6409 vpand ymm6,ymm1,ymm0
6410 vpxor ymm13,ymm13,YMMWORD[((96-128))+rax]
6411
6412 vpaddd ymm2,ymm2,ymm6
6413 vpsrld ymm8,ymm3,27
6414 vpxor ymm5,ymm1,ymm0
6415 vpxor ymm13,ymm13,ymm10
6416
6417 vmovdqu YMMWORD[(320-256-128)+rbx],ymm12
6418 vpaddd ymm2,ymm2,ymm12
6419 vpor ymm7,ymm7,ymm8
6420 vpsrld ymm9,ymm13,31
6421 vpand ymm5,ymm5,ymm4
6422 vpaddd ymm13,ymm13,ymm13
6423
6424 vpslld ymm6,ymm4,30
6425 vpaddd ymm2,ymm2,ymm5
6426
6427 vpsrld ymm4,ymm4,2
6428 vpaddd ymm2,ymm2,ymm7
6429 vpor ymm13,ymm13,ymm9
6430 vpor ymm4,ymm4,ymm6
6431 vpxor ymm14,ymm14,ymm11
6432 vmovdqa ymm11,YMMWORD[((448-256-128))+rbx]
6433
6434 vpaddd ymm1,ymm1,ymm15
6435 vpslld ymm7,ymm2,5
6436 vpand ymm6,ymm0,ymm4
6437 vpxor ymm14,ymm14,YMMWORD[((128-128))+rax]
6438
6439 vpaddd ymm1,ymm1,ymm6
6440 vpsrld ymm8,ymm2,27
6441 vpxor ymm5,ymm0,ymm4
6442 vpxor ymm14,ymm14,ymm11
6443
6444 vmovdqu YMMWORD[(352-256-128)+rbx],ymm13
6445 vpaddd ymm1,ymm1,ymm13
6446 vpor ymm7,ymm7,ymm8
6447 vpsrld ymm9,ymm14,31
6448 vpand ymm5,ymm5,ymm3
6449 vpaddd ymm14,ymm14,ymm14
6450
6451 vpslld ymm6,ymm3,30
6452 vpaddd ymm1,ymm1,ymm5
6453
6454 vpsrld ymm3,ymm3,2
6455 vpaddd ymm1,ymm1,ymm7
6456 vpor ymm14,ymm14,ymm9
6457 vpor ymm3,ymm3,ymm6
6458 vpxor ymm10,ymm10,ymm12
6459 vmovdqa ymm12,YMMWORD[((480-256-128))+rbx]
6460
6461 vpaddd ymm0,ymm0,ymm15
6462 vpslld ymm7,ymm1,5
6463 vpand ymm6,ymm4,ymm3
6464 vpxor ymm10,ymm10,YMMWORD[((160-128))+rax]
6465
6466 vpaddd ymm0,ymm0,ymm6
6467 vpsrld ymm8,ymm1,27
6468 vpxor ymm5,ymm4,ymm3
6469 vpxor ymm10,ymm10,ymm12
6470
6471 vmovdqu YMMWORD[(384-256-128)+rbx],ymm14
6472 vpaddd ymm0,ymm0,ymm14
6473 vpor ymm7,ymm7,ymm8
6474 vpsrld ymm9,ymm10,31
6475 vpand ymm5,ymm5,ymm2
6476 vpaddd ymm10,ymm10,ymm10
6477
6478 vpslld ymm6,ymm2,30
6479 vpaddd ymm0,ymm0,ymm5
6480
6481 vpsrld ymm2,ymm2,2
6482 vpaddd ymm0,ymm0,ymm7
6483 vpor ymm10,ymm10,ymm9
6484 vpor ymm2,ymm2,ymm6
6485 vpxor ymm11,ymm11,ymm13
6486 vmovdqa ymm13,YMMWORD[((0-128))+rax]
6487
6488 vpaddd ymm4,ymm4,ymm15
6489 vpslld ymm7,ymm0,5
6490 vpand ymm6,ymm3,ymm2
6491 vpxor ymm11,ymm11,YMMWORD[((192-128))+rax]
6492
6493 vpaddd ymm4,ymm4,ymm6
6494 vpsrld ymm8,ymm0,27
6495 vpxor ymm5,ymm3,ymm2
6496 vpxor ymm11,ymm11,ymm13
6497
6498 vmovdqu YMMWORD[(416-256-128)+rbx],ymm10
6499 vpaddd ymm4,ymm4,ymm10
6500 vpor ymm7,ymm7,ymm8
6501 vpsrld ymm9,ymm11,31
6502 vpand ymm5,ymm5,ymm1
6503 vpaddd ymm11,ymm11,ymm11
6504
6505 vpslld ymm6,ymm1,30
6506 vpaddd ymm4,ymm4,ymm5
6507
6508 vpsrld ymm1,ymm1,2
6509 vpaddd ymm4,ymm4,ymm7
6510 vpor ymm11,ymm11,ymm9
6511 vpor ymm1,ymm1,ymm6
6512 vpxor ymm12,ymm12,ymm14
6513 vmovdqa ymm14,YMMWORD[((32-128))+rax]
6514
6515 vpaddd ymm3,ymm3,ymm15
6516 vpslld ymm7,ymm4,5
6517 vpand ymm6,ymm2,ymm1
6518 vpxor ymm12,ymm12,YMMWORD[((224-128))+rax]
6519
6520 vpaddd ymm3,ymm3,ymm6
6521 vpsrld ymm8,ymm4,27
6522 vpxor ymm5,ymm2,ymm1
6523 vpxor ymm12,ymm12,ymm14
6524
6525 vmovdqu YMMWORD[(448-256-128)+rbx],ymm11
6526 vpaddd ymm3,ymm3,ymm11
6527 vpor ymm7,ymm7,ymm8
6528 vpsrld ymm9,ymm12,31
6529 vpand ymm5,ymm5,ymm0
6530 vpaddd ymm12,ymm12,ymm12
6531
6532 vpslld ymm6,ymm0,30
6533 vpaddd ymm3,ymm3,ymm5
6534
6535 vpsrld ymm0,ymm0,2
6536 vpaddd ymm3,ymm3,ymm7
6537 vpor ymm12,ymm12,ymm9
6538 vpor ymm0,ymm0,ymm6
6539 vpxor ymm13,ymm13,ymm10
6540 vmovdqa ymm10,YMMWORD[((64-128))+rax]
6541
6542 vpaddd ymm2,ymm2,ymm15
6543 vpslld ymm7,ymm3,5
6544 vpand ymm6,ymm1,ymm0
6545 vpxor ymm13,ymm13,YMMWORD[((256-256-128))+rbx]
6546
6547 vpaddd ymm2,ymm2,ymm6
6548 vpsrld ymm8,ymm3,27
6549 vpxor ymm5,ymm1,ymm0
6550 vpxor ymm13,ymm13,ymm10
6551
6552 vmovdqu YMMWORD[(480-256-128)+rbx],ymm12
6553 vpaddd ymm2,ymm2,ymm12
6554 vpor ymm7,ymm7,ymm8
6555 vpsrld ymm9,ymm13,31
6556 vpand ymm5,ymm5,ymm4
6557 vpaddd ymm13,ymm13,ymm13
6558
6559 vpslld ymm6,ymm4,30
6560 vpaddd ymm2,ymm2,ymm5
6561
6562 vpsrld ymm4,ymm4,2
6563 vpaddd ymm2,ymm2,ymm7
6564 vpor ymm13,ymm13,ymm9
6565 vpor ymm4,ymm4,ymm6
6566 vpxor ymm14,ymm14,ymm11
6567 vmovdqa ymm11,YMMWORD[((96-128))+rax]
6568
6569 vpaddd ymm1,ymm1,ymm15
6570 vpslld ymm7,ymm2,5
6571 vpand ymm6,ymm0,ymm4
6572 vpxor ymm14,ymm14,YMMWORD[((288-256-128))+rbx]
6573
6574 vpaddd ymm1,ymm1,ymm6
6575 vpsrld ymm8,ymm2,27
6576 vpxor ymm5,ymm0,ymm4
6577 vpxor ymm14,ymm14,ymm11
6578
6579 vmovdqu YMMWORD[(0-128)+rax],ymm13
6580 vpaddd ymm1,ymm1,ymm13
6581 vpor ymm7,ymm7,ymm8
6582 vpsrld ymm9,ymm14,31
6583 vpand ymm5,ymm5,ymm3
6584 vpaddd ymm14,ymm14,ymm14
6585
6586 vpslld ymm6,ymm3,30
6587 vpaddd ymm1,ymm1,ymm5
6588
6589 vpsrld ymm3,ymm3,2
6590 vpaddd ymm1,ymm1,ymm7
6591 vpor ymm14,ymm14,ymm9
6592 vpor ymm3,ymm3,ymm6
6593 vpxor ymm10,ymm10,ymm12
6594 vmovdqa ymm12,YMMWORD[((128-128))+rax]
6595
6596 vpaddd ymm0,ymm0,ymm15
6597 vpslld ymm7,ymm1,5
6598 vpand ymm6,ymm4,ymm3
6599 vpxor ymm10,ymm10,YMMWORD[((320-256-128))+rbx]
6600
6601 vpaddd ymm0,ymm0,ymm6
6602 vpsrld ymm8,ymm1,27
6603 vpxor ymm5,ymm4,ymm3
6604 vpxor ymm10,ymm10,ymm12
6605
6606 vmovdqu YMMWORD[(32-128)+rax],ymm14
6607 vpaddd ymm0,ymm0,ymm14
6608 vpor ymm7,ymm7,ymm8
6609 vpsrld ymm9,ymm10,31
6610 vpand ymm5,ymm5,ymm2
6611 vpaddd ymm10,ymm10,ymm10
6612
6613 vpslld ymm6,ymm2,30
6614 vpaddd ymm0,ymm0,ymm5
6615
6616 vpsrld ymm2,ymm2,2
6617 vpaddd ymm0,ymm0,ymm7
6618 vpor ymm10,ymm10,ymm9
6619 vpor ymm2,ymm2,ymm6
6620 vpxor ymm11,ymm11,ymm13
6621 vmovdqa ymm13,YMMWORD[((160-128))+rax]
6622
6623 vpaddd ymm4,ymm4,ymm15
6624 vpslld ymm7,ymm0,5
6625 vpand ymm6,ymm3,ymm2
6626 vpxor ymm11,ymm11,YMMWORD[((352-256-128))+rbx]
6627
6628 vpaddd ymm4,ymm4,ymm6
6629 vpsrld ymm8,ymm0,27
6630 vpxor ymm5,ymm3,ymm2
6631 vpxor ymm11,ymm11,ymm13
6632
6633 vmovdqu YMMWORD[(64-128)+rax],ymm10
6634 vpaddd ymm4,ymm4,ymm10
6635 vpor ymm7,ymm7,ymm8
6636 vpsrld ymm9,ymm11,31
6637 vpand ymm5,ymm5,ymm1
6638 vpaddd ymm11,ymm11,ymm11
6639
6640 vpslld ymm6,ymm1,30
6641 vpaddd ymm4,ymm4,ymm5
6642
6643 vpsrld ymm1,ymm1,2
6644 vpaddd ymm4,ymm4,ymm7
6645 vpor ymm11,ymm11,ymm9
6646 vpor ymm1,ymm1,ymm6
6647 vpxor ymm12,ymm12,ymm14
6648 vmovdqa ymm14,YMMWORD[((192-128))+rax]
6649
6650 vpaddd ymm3,ymm3,ymm15
6651 vpslld ymm7,ymm4,5
6652 vpand ymm6,ymm2,ymm1
6653 vpxor ymm12,ymm12,YMMWORD[((384-256-128))+rbx]
6654
6655 vpaddd ymm3,ymm3,ymm6
6656 vpsrld ymm8,ymm4,27
6657 vpxor ymm5,ymm2,ymm1
6658 vpxor ymm12,ymm12,ymm14
6659
6660 vmovdqu YMMWORD[(96-128)+rax],ymm11
6661 vpaddd ymm3,ymm3,ymm11
6662 vpor ymm7,ymm7,ymm8
6663 vpsrld ymm9,ymm12,31
6664 vpand ymm5,ymm5,ymm0
6665 vpaddd ymm12,ymm12,ymm12
6666
6667 vpslld ymm6,ymm0,30
6668 vpaddd ymm3,ymm3,ymm5
6669
6670 vpsrld ymm0,ymm0,2
6671 vpaddd ymm3,ymm3,ymm7
6672 vpor ymm12,ymm12,ymm9
6673 vpor ymm0,ymm0,ymm6
6674 vpxor ymm13,ymm13,ymm10
6675 vmovdqa ymm10,YMMWORD[((224-128))+rax]
6676
6677 vpaddd ymm2,ymm2,ymm15
6678 vpslld ymm7,ymm3,5
6679 vpand ymm6,ymm1,ymm0
6680 vpxor ymm13,ymm13,YMMWORD[((416-256-128))+rbx]
6681
6682 vpaddd ymm2,ymm2,ymm6
6683 vpsrld ymm8,ymm3,27
6684 vpxor ymm5,ymm1,ymm0
6685 vpxor ymm13,ymm13,ymm10
6686
6687 vmovdqu YMMWORD[(128-128)+rax],ymm12
6688 vpaddd ymm2,ymm2,ymm12
6689 vpor ymm7,ymm7,ymm8
6690 vpsrld ymm9,ymm13,31
6691 vpand ymm5,ymm5,ymm4
6692 vpaddd ymm13,ymm13,ymm13
6693
6694 vpslld ymm6,ymm4,30
6695 vpaddd ymm2,ymm2,ymm5
6696
6697 vpsrld ymm4,ymm4,2
6698 vpaddd ymm2,ymm2,ymm7
6699 vpor ymm13,ymm13,ymm9
6700 vpor ymm4,ymm4,ymm6
6701 vpxor ymm14,ymm14,ymm11
6702 vmovdqa ymm11,YMMWORD[((256-256-128))+rbx]
6703
6704 vpaddd ymm1,ymm1,ymm15
6705 vpslld ymm7,ymm2,5
6706 vpand ymm6,ymm0,ymm4
6707 vpxor ymm14,ymm14,YMMWORD[((448-256-128))+rbx]
6708
6709 vpaddd ymm1,ymm1,ymm6
6710 vpsrld ymm8,ymm2,27
6711 vpxor ymm5,ymm0,ymm4
6712 vpxor ymm14,ymm14,ymm11
6713
6714 vmovdqu YMMWORD[(160-128)+rax],ymm13
6715 vpaddd ymm1,ymm1,ymm13
6716 vpor ymm7,ymm7,ymm8
6717 vpsrld ymm9,ymm14,31
6718 vpand ymm5,ymm5,ymm3
6719 vpaddd ymm14,ymm14,ymm14
6720
6721 vpslld ymm6,ymm3,30
6722 vpaddd ymm1,ymm1,ymm5
6723
6724 vpsrld ymm3,ymm3,2
6725 vpaddd ymm1,ymm1,ymm7
6726 vpor ymm14,ymm14,ymm9
6727 vpor ymm3,ymm3,ymm6
6728 vpxor ymm10,ymm10,ymm12
6729 vmovdqa ymm12,YMMWORD[((288-256-128))+rbx]
6730
6731 vpaddd ymm0,ymm0,ymm15
6732 vpslld ymm7,ymm1,5
6733 vpand ymm6,ymm4,ymm3
6734 vpxor ymm10,ymm10,YMMWORD[((480-256-128))+rbx]
6735
6736 vpaddd ymm0,ymm0,ymm6
6737 vpsrld ymm8,ymm1,27
6738 vpxor ymm5,ymm4,ymm3
6739 vpxor ymm10,ymm10,ymm12
6740
6741 vmovdqu YMMWORD[(192-128)+rax],ymm14
6742 vpaddd ymm0,ymm0,ymm14
6743 vpor ymm7,ymm7,ymm8
6744 vpsrld ymm9,ymm10,31
6745 vpand ymm5,ymm5,ymm2
6746 vpaddd ymm10,ymm10,ymm10
6747
6748 vpslld ymm6,ymm2,30
6749 vpaddd ymm0,ymm0,ymm5
6750
6751 vpsrld ymm2,ymm2,2
6752 vpaddd ymm0,ymm0,ymm7
6753 vpor ymm10,ymm10,ymm9
6754 vpor ymm2,ymm2,ymm6
6755 vpxor ymm11,ymm11,ymm13
6756 vmovdqa ymm13,YMMWORD[((320-256-128))+rbx]
6757
6758 vpaddd ymm4,ymm4,ymm15
6759 vpslld ymm7,ymm0,5
6760 vpand ymm6,ymm3,ymm2
6761 vpxor ymm11,ymm11,YMMWORD[((0-128))+rax]
6762
6763 vpaddd ymm4,ymm4,ymm6
6764 vpsrld ymm8,ymm0,27
6765 vpxor ymm5,ymm3,ymm2
6766 vpxor ymm11,ymm11,ymm13
6767
6768 vmovdqu YMMWORD[(224-128)+rax],ymm10
6769 vpaddd ymm4,ymm4,ymm10
6770 vpor ymm7,ymm7,ymm8
6771 vpsrld ymm9,ymm11,31
6772 vpand ymm5,ymm5,ymm1
6773 vpaddd ymm11,ymm11,ymm11
6774
6775 vpslld ymm6,ymm1,30
6776 vpaddd ymm4,ymm4,ymm5
6777
6778 vpsrld ymm1,ymm1,2
6779 vpaddd ymm4,ymm4,ymm7
6780 vpor ymm11,ymm11,ymm9
6781 vpor ymm1,ymm1,ymm6
6782 vpxor ymm12,ymm12,ymm14
6783 vmovdqa ymm14,YMMWORD[((352-256-128))+rbx]
6784
6785 vpaddd ymm3,ymm3,ymm15
6786 vpslld ymm7,ymm4,5
6787 vpand ymm6,ymm2,ymm1
6788 vpxor ymm12,ymm12,YMMWORD[((32-128))+rax]
6789
6790 vpaddd ymm3,ymm3,ymm6
6791 vpsrld ymm8,ymm4,27
6792 vpxor ymm5,ymm2,ymm1
6793 vpxor ymm12,ymm12,ymm14
6794
6795 vmovdqu YMMWORD[(256-256-128)+rbx],ymm11
6796 vpaddd ymm3,ymm3,ymm11
6797 vpor ymm7,ymm7,ymm8
6798 vpsrld ymm9,ymm12,31
6799 vpand ymm5,ymm5,ymm0
6800 vpaddd ymm12,ymm12,ymm12
6801
6802 vpslld ymm6,ymm0,30
6803 vpaddd ymm3,ymm3,ymm5
6804
6805 vpsrld ymm0,ymm0,2
6806 vpaddd ymm3,ymm3,ymm7
6807 vpor ymm12,ymm12,ymm9
6808 vpor ymm0,ymm0,ymm6
6809 vpxor ymm13,ymm13,ymm10
6810 vmovdqa ymm10,YMMWORD[((384-256-128))+rbx]
6811
6812 vpaddd ymm2,ymm2,ymm15
6813 vpslld ymm7,ymm3,5
6814 vpand ymm6,ymm1,ymm0
6815 vpxor ymm13,ymm13,YMMWORD[((64-128))+rax]
6816
6817 vpaddd ymm2,ymm2,ymm6
6818 vpsrld ymm8,ymm3,27
6819 vpxor ymm5,ymm1,ymm0
6820 vpxor ymm13,ymm13,ymm10
6821
6822 vmovdqu YMMWORD[(288-256-128)+rbx],ymm12
6823 vpaddd ymm2,ymm2,ymm12
6824 vpor ymm7,ymm7,ymm8
6825 vpsrld ymm9,ymm13,31
6826 vpand ymm5,ymm5,ymm4
6827 vpaddd ymm13,ymm13,ymm13
6828
6829 vpslld ymm6,ymm4,30
6830 vpaddd ymm2,ymm2,ymm5
6831
6832 vpsrld ymm4,ymm4,2
6833 vpaddd ymm2,ymm2,ymm7
6834 vpor ymm13,ymm13,ymm9
6835 vpor ymm4,ymm4,ymm6
6836 vpxor ymm14,ymm14,ymm11
6837 vmovdqa ymm11,YMMWORD[((416-256-128))+rbx]
6838
6839 vpaddd ymm1,ymm1,ymm15
6840 vpslld ymm7,ymm2,5
6841 vpand ymm6,ymm0,ymm4
6842 vpxor ymm14,ymm14,YMMWORD[((96-128))+rax]
6843
6844 vpaddd ymm1,ymm1,ymm6
6845 vpsrld ymm8,ymm2,27
6846 vpxor ymm5,ymm0,ymm4
6847 vpxor ymm14,ymm14,ymm11
6848
6849 vmovdqu YMMWORD[(320-256-128)+rbx],ymm13
6850 vpaddd ymm1,ymm1,ymm13
6851 vpor ymm7,ymm7,ymm8
6852 vpsrld ymm9,ymm14,31
6853 vpand ymm5,ymm5,ymm3
6854 vpaddd ymm14,ymm14,ymm14
6855
6856 vpslld ymm6,ymm3,30
6857 vpaddd ymm1,ymm1,ymm5
6858
6859 vpsrld ymm3,ymm3,2
6860 vpaddd ymm1,ymm1,ymm7
6861 vpor ymm14,ymm14,ymm9
6862 vpor ymm3,ymm3,ymm6
6863 vpxor ymm10,ymm10,ymm12
6864 vmovdqa ymm12,YMMWORD[((448-256-128))+rbx]
6865
6866 vpaddd ymm0,ymm0,ymm15
6867 vpslld ymm7,ymm1,5
6868 vpand ymm6,ymm4,ymm3
6869 vpxor ymm10,ymm10,YMMWORD[((128-128))+rax]
6870
6871 vpaddd ymm0,ymm0,ymm6
6872 vpsrld ymm8,ymm1,27
6873 vpxor ymm5,ymm4,ymm3
6874 vpxor ymm10,ymm10,ymm12
6875
6876 vmovdqu YMMWORD[(352-256-128)+rbx],ymm14
6877 vpaddd ymm0,ymm0,ymm14
6878 vpor ymm7,ymm7,ymm8
6879 vpsrld ymm9,ymm10,31
6880 vpand ymm5,ymm5,ymm2
6881 vpaddd ymm10,ymm10,ymm10
6882
6883 vpslld ymm6,ymm2,30
6884 vpaddd ymm0,ymm0,ymm5
6885
6886 vpsrld ymm2,ymm2,2
6887 vpaddd ymm0,ymm0,ymm7
6888 vpor ymm10,ymm10,ymm9
6889 vpor ymm2,ymm2,ymm6
6890 vmovdqa ymm15,YMMWORD[64+rbp]
6891 vpxor ymm11,ymm11,ymm13
6892 vmovdqa ymm13,YMMWORD[((480-256-128))+rbx]
6893
6894 vpslld ymm7,ymm0,5
6895 vpaddd ymm4,ymm4,ymm15
6896 vpxor ymm5,ymm3,ymm1
6897 vmovdqa YMMWORD[(384-256-128)+rbx],ymm10
6898 vpaddd ymm4,ymm4,ymm10
6899 vpxor ymm11,ymm11,YMMWORD[((160-128))+rax]
6900 vpsrld ymm8,ymm0,27
6901 vpxor ymm5,ymm5,ymm2
6902 vpxor ymm11,ymm11,ymm13
6903
6904 vpslld ymm6,ymm1,30
6905 vpor ymm7,ymm7,ymm8
6906 vpaddd ymm4,ymm4,ymm5
6907 vpsrld ymm9,ymm11,31
6908 vpaddd ymm11,ymm11,ymm11
6909
6910 vpsrld ymm1,ymm1,2
6911 vpaddd ymm4,ymm4,ymm7
6912 vpor ymm11,ymm11,ymm9
6913 vpor ymm1,ymm1,ymm6
6914 vpxor ymm12,ymm12,ymm14
6915 vmovdqa ymm14,YMMWORD[((0-128))+rax]
6916
6917 vpslld ymm7,ymm4,5
6918 vpaddd ymm3,ymm3,ymm15
6919 vpxor ymm5,ymm2,ymm0
6920 vmovdqa YMMWORD[(416-256-128)+rbx],ymm11
6921 vpaddd ymm3,ymm3,ymm11
6922 vpxor ymm12,ymm12,YMMWORD[((192-128))+rax]
6923 vpsrld ymm8,ymm4,27
6924 vpxor ymm5,ymm5,ymm1
6925 vpxor ymm12,ymm12,ymm14
6926
6927 vpslld ymm6,ymm0,30
6928 vpor ymm7,ymm7,ymm8
6929 vpaddd ymm3,ymm3,ymm5
6930 vpsrld ymm9,ymm12,31
6931 vpaddd ymm12,ymm12,ymm12
6932
6933 vpsrld ymm0,ymm0,2
6934 vpaddd ymm3,ymm3,ymm7
6935 vpor ymm12,ymm12,ymm9
6936 vpor ymm0,ymm0,ymm6
6937 vpxor ymm13,ymm13,ymm10
6938 vmovdqa ymm10,YMMWORD[((32-128))+rax]
6939
6940 vpslld ymm7,ymm3,5
6941 vpaddd ymm2,ymm2,ymm15
6942 vpxor ymm5,ymm1,ymm4
6943 vmovdqa YMMWORD[(448-256-128)+rbx],ymm12
6944 vpaddd ymm2,ymm2,ymm12
6945 vpxor ymm13,ymm13,YMMWORD[((224-128))+rax]
6946 vpsrld ymm8,ymm3,27
6947 vpxor ymm5,ymm5,ymm0
6948 vpxor ymm13,ymm13,ymm10
6949
6950 vpslld ymm6,ymm4,30
6951 vpor ymm7,ymm7,ymm8
6952 vpaddd ymm2,ymm2,ymm5
6953 vpsrld ymm9,ymm13,31
6954 vpaddd ymm13,ymm13,ymm13
6955
6956 vpsrld ymm4,ymm4,2
6957 vpaddd ymm2,ymm2,ymm7
6958 vpor ymm13,ymm13,ymm9
6959 vpor ymm4,ymm4,ymm6
6960 vpxor ymm14,ymm14,ymm11
6961 vmovdqa ymm11,YMMWORD[((64-128))+rax]
6962
6963 vpslld ymm7,ymm2,5
6964 vpaddd ymm1,ymm1,ymm15
6965 vpxor ymm5,ymm0,ymm3
6966 vmovdqa YMMWORD[(480-256-128)+rbx],ymm13
6967 vpaddd ymm1,ymm1,ymm13
6968 vpxor ymm14,ymm14,YMMWORD[((256-256-128))+rbx]
6969 vpsrld ymm8,ymm2,27
6970 vpxor ymm5,ymm5,ymm4
6971 vpxor ymm14,ymm14,ymm11
6972
6973 vpslld ymm6,ymm3,30
6974 vpor ymm7,ymm7,ymm8
6975 vpaddd ymm1,ymm1,ymm5
6976 vpsrld ymm9,ymm14,31
6977 vpaddd ymm14,ymm14,ymm14
6978
6979 vpsrld ymm3,ymm3,2
6980 vpaddd ymm1,ymm1,ymm7
6981 vpor ymm14,ymm14,ymm9
6982 vpor ymm3,ymm3,ymm6
6983 vpxor ymm10,ymm10,ymm12
6984 vmovdqa ymm12,YMMWORD[((96-128))+rax]
6985
6986 vpslld ymm7,ymm1,5
6987 vpaddd ymm0,ymm0,ymm15
6988 vpxor ymm5,ymm4,ymm2
6989 vmovdqa YMMWORD[(0-128)+rax],ymm14
6990 vpaddd ymm0,ymm0,ymm14
6991 vpxor ymm10,ymm10,YMMWORD[((288-256-128))+rbx]
6992 vpsrld ymm8,ymm1,27
6993 vpxor ymm5,ymm5,ymm3
6994 vpxor ymm10,ymm10,ymm12
6995
6996 vpslld ymm6,ymm2,30
6997 vpor ymm7,ymm7,ymm8
6998 vpaddd ymm0,ymm0,ymm5
6999 vpsrld ymm9,ymm10,31
7000 vpaddd ymm10,ymm10,ymm10
7001
7002 vpsrld ymm2,ymm2,2
7003 vpaddd ymm0,ymm0,ymm7
7004 vpor ymm10,ymm10,ymm9
7005 vpor ymm2,ymm2,ymm6
7006 vpxor ymm11,ymm11,ymm13
7007 vmovdqa ymm13,YMMWORD[((128-128))+rax]
7008
7009 vpslld ymm7,ymm0,5
7010 vpaddd ymm4,ymm4,ymm15
7011 vpxor ymm5,ymm3,ymm1
7012 vmovdqa YMMWORD[(32-128)+rax],ymm10
7013 vpaddd ymm4,ymm4,ymm10
7014 vpxor ymm11,ymm11,YMMWORD[((320-256-128))+rbx]
7015 vpsrld ymm8,ymm0,27
7016 vpxor ymm5,ymm5,ymm2
7017 vpxor ymm11,ymm11,ymm13
7018
7019 vpslld ymm6,ymm1,30
7020 vpor ymm7,ymm7,ymm8
7021 vpaddd ymm4,ymm4,ymm5
7022 vpsrld ymm9,ymm11,31
7023 vpaddd ymm11,ymm11,ymm11
7024
7025 vpsrld ymm1,ymm1,2
7026 vpaddd ymm4,ymm4,ymm7
7027 vpor ymm11,ymm11,ymm9
7028 vpor ymm1,ymm1,ymm6
7029 vpxor ymm12,ymm12,ymm14
7030 vmovdqa ymm14,YMMWORD[((160-128))+rax]
7031
7032 vpslld ymm7,ymm4,5
7033 vpaddd ymm3,ymm3,ymm15
7034 vpxor ymm5,ymm2,ymm0
7035 vmovdqa YMMWORD[(64-128)+rax],ymm11
7036 vpaddd ymm3,ymm3,ymm11
7037 vpxor ymm12,ymm12,YMMWORD[((352-256-128))+rbx]
7038 vpsrld ymm8,ymm4,27
7039 vpxor ymm5,ymm5,ymm1
7040 vpxor ymm12,ymm12,ymm14
7041
7042 vpslld ymm6,ymm0,30
7043 vpor ymm7,ymm7,ymm8
7044 vpaddd ymm3,ymm3,ymm5
7045 vpsrld ymm9,ymm12,31
7046 vpaddd ymm12,ymm12,ymm12
7047
7048 vpsrld ymm0,ymm0,2
7049 vpaddd ymm3,ymm3,ymm7
7050 vpor ymm12,ymm12,ymm9
7051 vpor ymm0,ymm0,ymm6
7052 vpxor ymm13,ymm13,ymm10
7053 vmovdqa ymm10,YMMWORD[((192-128))+rax]
7054
7055 vpslld ymm7,ymm3,5
7056 vpaddd ymm2,ymm2,ymm15
7057 vpxor ymm5,ymm1,ymm4
7058 vmovdqa YMMWORD[(96-128)+rax],ymm12
7059 vpaddd ymm2,ymm2,ymm12
7060 vpxor ymm13,ymm13,YMMWORD[((384-256-128))+rbx]
7061 vpsrld ymm8,ymm3,27
7062 vpxor ymm5,ymm5,ymm0
7063 vpxor ymm13,ymm13,ymm10
7064
7065 vpslld ymm6,ymm4,30
7066 vpor ymm7,ymm7,ymm8
7067 vpaddd ymm2,ymm2,ymm5
7068 vpsrld ymm9,ymm13,31
7069 vpaddd ymm13,ymm13,ymm13
7070
7071 vpsrld ymm4,ymm4,2
7072 vpaddd ymm2,ymm2,ymm7
7073 vpor ymm13,ymm13,ymm9
7074 vpor ymm4,ymm4,ymm6
7075 vpxor ymm14,ymm14,ymm11
7076 vmovdqa ymm11,YMMWORD[((224-128))+rax]
7077
7078 vpslld ymm7,ymm2,5
7079 vpaddd ymm1,ymm1,ymm15
7080 vpxor ymm5,ymm0,ymm3
7081 vmovdqa YMMWORD[(128-128)+rax],ymm13
7082 vpaddd ymm1,ymm1,ymm13
7083 vpxor ymm14,ymm14,YMMWORD[((416-256-128))+rbx]
7084 vpsrld ymm8,ymm2,27
7085 vpxor ymm5,ymm5,ymm4
7086 vpxor ymm14,ymm14,ymm11
7087
7088 vpslld ymm6,ymm3,30
7089 vpor ymm7,ymm7,ymm8
7090 vpaddd ymm1,ymm1,ymm5
7091 vpsrld ymm9,ymm14,31
7092 vpaddd ymm14,ymm14,ymm14
7093
7094 vpsrld ymm3,ymm3,2
7095 vpaddd ymm1,ymm1,ymm7
7096 vpor ymm14,ymm14,ymm9
7097 vpor ymm3,ymm3,ymm6
7098 vpxor ymm10,ymm10,ymm12
7099 vmovdqa ymm12,YMMWORD[((256-256-128))+rbx]
7100
7101 vpslld ymm7,ymm1,5
7102 vpaddd ymm0,ymm0,ymm15
7103 vpxor ymm5,ymm4,ymm2
7104 vmovdqa YMMWORD[(160-128)+rax],ymm14
7105 vpaddd ymm0,ymm0,ymm14
7106 vpxor ymm10,ymm10,YMMWORD[((448-256-128))+rbx]
7107 vpsrld ymm8,ymm1,27
7108 vpxor ymm5,ymm5,ymm3
7109 vpxor ymm10,ymm10,ymm12
7110
7111 vpslld ymm6,ymm2,30
7112 vpor ymm7,ymm7,ymm8
7113 vpaddd ymm0,ymm0,ymm5
7114 vpsrld ymm9,ymm10,31
7115 vpaddd ymm10,ymm10,ymm10
7116
7117 vpsrld ymm2,ymm2,2
7118 vpaddd ymm0,ymm0,ymm7
7119 vpor ymm10,ymm10,ymm9
7120 vpor ymm2,ymm2,ymm6
7121 vpxor ymm11,ymm11,ymm13
7122 vmovdqa ymm13,YMMWORD[((288-256-128))+rbx]
7123
7124 vpslld ymm7,ymm0,5
7125 vpaddd ymm4,ymm4,ymm15
7126 vpxor ymm5,ymm3,ymm1
7127 vmovdqa YMMWORD[(192-128)+rax],ymm10
7128 vpaddd ymm4,ymm4,ymm10
7129 vpxor ymm11,ymm11,YMMWORD[((480-256-128))+rbx]
7130 vpsrld ymm8,ymm0,27
7131 vpxor ymm5,ymm5,ymm2
7132 vpxor ymm11,ymm11,ymm13
7133
7134 vpslld ymm6,ymm1,30
7135 vpor ymm7,ymm7,ymm8
7136 vpaddd ymm4,ymm4,ymm5
7137 vpsrld ymm9,ymm11,31
7138 vpaddd ymm11,ymm11,ymm11
7139
7140 vpsrld ymm1,ymm1,2
7141 vpaddd ymm4,ymm4,ymm7
7142 vpor ymm11,ymm11,ymm9
7143 vpor ymm1,ymm1,ymm6
7144 vpxor ymm12,ymm12,ymm14
7145 vmovdqa ymm14,YMMWORD[((320-256-128))+rbx]
7146
7147 vpslld ymm7,ymm4,5
7148 vpaddd ymm3,ymm3,ymm15
7149 vpxor ymm5,ymm2,ymm0
7150 vmovdqa YMMWORD[(224-128)+rax],ymm11
7151 vpaddd ymm3,ymm3,ymm11
7152 vpxor ymm12,ymm12,YMMWORD[((0-128))+rax]
7153 vpsrld ymm8,ymm4,27
7154 vpxor ymm5,ymm5,ymm1
7155 vpxor ymm12,ymm12,ymm14
7156
7157 vpslld ymm6,ymm0,30
7158 vpor ymm7,ymm7,ymm8
7159 vpaddd ymm3,ymm3,ymm5
7160 vpsrld ymm9,ymm12,31
7161 vpaddd ymm12,ymm12,ymm12
7162
7163 vpsrld ymm0,ymm0,2
7164 vpaddd ymm3,ymm3,ymm7
7165 vpor ymm12,ymm12,ymm9
7166 vpor ymm0,ymm0,ymm6
7167 vpxor ymm13,ymm13,ymm10
7168 vmovdqa ymm10,YMMWORD[((352-256-128))+rbx]
7169
7170 vpslld ymm7,ymm3,5
7171 vpaddd ymm2,ymm2,ymm15
7172 vpxor ymm5,ymm1,ymm4
7173 vpaddd ymm2,ymm2,ymm12
7174 vpxor ymm13,ymm13,YMMWORD[((32-128))+rax]
7175 vpsrld ymm8,ymm3,27
7176 vpxor ymm5,ymm5,ymm0
7177 vpxor ymm13,ymm13,ymm10
7178
7179 vpslld ymm6,ymm4,30
7180 vpor ymm7,ymm7,ymm8
7181 vpaddd ymm2,ymm2,ymm5
7182 vpsrld ymm9,ymm13,31
7183 vpaddd ymm13,ymm13,ymm13
7184
7185 vpsrld ymm4,ymm4,2
7186 vpaddd ymm2,ymm2,ymm7
7187 vpor ymm13,ymm13,ymm9
7188 vpor ymm4,ymm4,ymm6
7189 vpxor ymm14,ymm14,ymm11
7190 vmovdqa ymm11,YMMWORD[((384-256-128))+rbx]
7191
7192 vpslld ymm7,ymm2,5
7193 vpaddd ymm1,ymm1,ymm15
7194 vpxor ymm5,ymm0,ymm3
7195 vpaddd ymm1,ymm1,ymm13
7196 vpxor ymm14,ymm14,YMMWORD[((64-128))+rax]
7197 vpsrld ymm8,ymm2,27
7198 vpxor ymm5,ymm5,ymm4
7199 vpxor ymm14,ymm14,ymm11
7200
7201 vpslld ymm6,ymm3,30
7202 vpor ymm7,ymm7,ymm8
7203 vpaddd ymm1,ymm1,ymm5
7204 vpsrld ymm9,ymm14,31
7205 vpaddd ymm14,ymm14,ymm14
7206
7207 vpsrld ymm3,ymm3,2
7208 vpaddd ymm1,ymm1,ymm7
7209 vpor ymm14,ymm14,ymm9
7210 vpor ymm3,ymm3,ymm6
7211 vpxor ymm10,ymm10,ymm12
7212 vmovdqa ymm12,YMMWORD[((416-256-128))+rbx]
7213
7214 vpslld ymm7,ymm1,5
7215 vpaddd ymm0,ymm0,ymm15
7216 vpxor ymm5,ymm4,ymm2
7217 vpaddd ymm0,ymm0,ymm14
7218 vpxor ymm10,ymm10,YMMWORD[((96-128))+rax]
7219 vpsrld ymm8,ymm1,27
7220 vpxor ymm5,ymm5,ymm3
7221 vpxor ymm10,ymm10,ymm12
7222
7223 vpslld ymm6,ymm2,30
7224 vpor ymm7,ymm7,ymm8
7225 vpaddd ymm0,ymm0,ymm5
7226 vpsrld ymm9,ymm10,31
7227 vpaddd ymm10,ymm10,ymm10
7228
7229 vpsrld ymm2,ymm2,2
7230 vpaddd ymm0,ymm0,ymm7
7231 vpor ymm10,ymm10,ymm9
7232 vpor ymm2,ymm2,ymm6
7233 vpxor ymm11,ymm11,ymm13
7234 vmovdqa ymm13,YMMWORD[((448-256-128))+rbx]
7235
7236 vpslld ymm7,ymm0,5
7237 vpaddd ymm4,ymm4,ymm15
7238 vpxor ymm5,ymm3,ymm1
7239 vpaddd ymm4,ymm4,ymm10
7240 vpxor ymm11,ymm11,YMMWORD[((128-128))+rax]
7241 vpsrld ymm8,ymm0,27
7242 vpxor ymm5,ymm5,ymm2
7243 vpxor ymm11,ymm11,ymm13
7244
7245 vpslld ymm6,ymm1,30
7246 vpor ymm7,ymm7,ymm8
7247 vpaddd ymm4,ymm4,ymm5
7248 vpsrld ymm9,ymm11,31
7249 vpaddd ymm11,ymm11,ymm11
7250
7251 vpsrld ymm1,ymm1,2
7252 vpaddd ymm4,ymm4,ymm7
7253 vpor ymm11,ymm11,ymm9
7254 vpor ymm1,ymm1,ymm6
7255 vpxor ymm12,ymm12,ymm14
7256 vmovdqa ymm14,YMMWORD[((480-256-128))+rbx]
7257
7258 vpslld ymm7,ymm4,5
7259 vpaddd ymm3,ymm3,ymm15
7260 vpxor ymm5,ymm2,ymm0
7261 vpaddd ymm3,ymm3,ymm11
7262 vpxor ymm12,ymm12,YMMWORD[((160-128))+rax]
7263 vpsrld ymm8,ymm4,27
7264 vpxor ymm5,ymm5,ymm1
7265 vpxor ymm12,ymm12,ymm14
7266
7267 vpslld ymm6,ymm0,30
7268 vpor ymm7,ymm7,ymm8
7269 vpaddd ymm3,ymm3,ymm5
7270 vpsrld ymm9,ymm12,31
7271 vpaddd ymm12,ymm12,ymm12
7272
7273 vpsrld ymm0,ymm0,2
7274 vpaddd ymm3,ymm3,ymm7
7275 vpor ymm12,ymm12,ymm9
7276 vpor ymm0,ymm0,ymm6
7277 vpxor ymm13,ymm13,ymm10
7278 vmovdqa ymm10,YMMWORD[((0-128))+rax]
7279
7280 vpslld ymm7,ymm3,5
7281 vpaddd ymm2,ymm2,ymm15
7282 vpxor ymm5,ymm1,ymm4
7283 vpaddd ymm2,ymm2,ymm12
7284 vpxor ymm13,ymm13,YMMWORD[((192-128))+rax]
7285 vpsrld ymm8,ymm3,27
7286 vpxor ymm5,ymm5,ymm0
7287 vpxor ymm13,ymm13,ymm10
7288
7289 vpslld ymm6,ymm4,30
7290 vpor ymm7,ymm7,ymm8
7291 vpaddd ymm2,ymm2,ymm5
7292 vpsrld ymm9,ymm13,31
7293 vpaddd ymm13,ymm13,ymm13
7294
7295 vpsrld ymm4,ymm4,2
7296 vpaddd ymm2,ymm2,ymm7
7297 vpor ymm13,ymm13,ymm9
7298 vpor ymm4,ymm4,ymm6
7299 vpxor ymm14,ymm14,ymm11
7300 vmovdqa ymm11,YMMWORD[((32-128))+rax]
7301
7302 vpslld ymm7,ymm2,5
7303 vpaddd ymm1,ymm1,ymm15
7304 vpxor ymm5,ymm0,ymm3
7305 vpaddd ymm1,ymm1,ymm13
7306 vpxor ymm14,ymm14,YMMWORD[((224-128))+rax]
7307 vpsrld ymm8,ymm2,27
7308 vpxor ymm5,ymm5,ymm4
7309 vpxor ymm14,ymm14,ymm11
7310
7311 vpslld ymm6,ymm3,30
7312 vpor ymm7,ymm7,ymm8
7313 vpaddd ymm1,ymm1,ymm5
7314 vpsrld ymm9,ymm14,31
7315 vpaddd ymm14,ymm14,ymm14
7316
7317 vpsrld ymm3,ymm3,2
7318 vpaddd ymm1,ymm1,ymm7
7319 vpor ymm14,ymm14,ymm9
7320 vpor ymm3,ymm3,ymm6
7321 vpslld ymm7,ymm1,5
7322 vpaddd ymm0,ymm0,ymm15
7323 vpxor ymm5,ymm4,ymm2
7324
7325 vpsrld ymm8,ymm1,27
7326 vpaddd ymm0,ymm0,ymm14
7327 vpxor ymm5,ymm5,ymm3
7328
7329 vpslld ymm6,ymm2,30
7330 vpor ymm7,ymm7,ymm8
7331 vpaddd ymm0,ymm0,ymm5
7332
7333 vpsrld ymm2,ymm2,2
7334 vpaddd ymm0,ymm0,ymm7
7335 vpor ymm2,ymm2,ymm6
7336 mov ecx,1
7337 lea rbx,[512+rsp]
7338 cmp ecx,DWORD[rbx]
7339 cmovge r12,rbp
7340 cmp ecx,DWORD[4+rbx]
7341 cmovge r13,rbp
7342 cmp ecx,DWORD[8+rbx]
7343 cmovge r14,rbp
7344 cmp ecx,DWORD[12+rbx]
7345 cmovge r15,rbp
7346 cmp ecx,DWORD[16+rbx]
7347 cmovge r8,rbp
7348 cmp ecx,DWORD[20+rbx]
7349 cmovge r9,rbp
7350 cmp ecx,DWORD[24+rbx]
7351 cmovge r10,rbp
7352 cmp ecx,DWORD[28+rbx]
7353 cmovge r11,rbp
7354 vmovdqu ymm5,YMMWORD[rbx]
7355 vpxor ymm7,ymm7,ymm7
7356 vmovdqa ymm6,ymm5
7357 vpcmpgtd ymm6,ymm6,ymm7
7358 vpaddd ymm5,ymm5,ymm6
7359
7360 vpand ymm0,ymm0,ymm6
7361 vpand ymm1,ymm1,ymm6
7362 vpaddd ymm0,ymm0,YMMWORD[rdi]
7363 vpand ymm2,ymm2,ymm6
7364 vpaddd ymm1,ymm1,YMMWORD[32+rdi]
7365 vpand ymm3,ymm3,ymm6
7366 vpaddd ymm2,ymm2,YMMWORD[64+rdi]
7367 vpand ymm4,ymm4,ymm6
7368 vpaddd ymm3,ymm3,YMMWORD[96+rdi]
7369 vpaddd ymm4,ymm4,YMMWORD[128+rdi]
7370 vmovdqu YMMWORD[rdi],ymm0
7371 vmovdqu YMMWORD[32+rdi],ymm1
7372 vmovdqu YMMWORD[64+rdi],ymm2
7373 vmovdqu YMMWORD[96+rdi],ymm3
7374 vmovdqu YMMWORD[128+rdi],ymm4
7375
7376 vmovdqu YMMWORD[rbx],ymm5
7377 lea rbx,[((256+128))+rsp]
7378 vmovdqu ymm9,YMMWORD[96+rbp]
7379 dec edx
7380 jnz NEAR $L$oop_avx2
7381
7382
7383
7384
7385
7386
7387
7388$L$done_avx2:
7389 mov rax,QWORD[544+rsp]
7390
7391 vzeroupper
7392 movaps xmm6,XMMWORD[((-216))+rax]
7393 movaps xmm7,XMMWORD[((-200))+rax]
7394 movaps xmm8,XMMWORD[((-184))+rax]
7395 movaps xmm9,XMMWORD[((-168))+rax]
7396 movaps xmm10,XMMWORD[((-152))+rax]
7397 movaps xmm11,XMMWORD[((-136))+rax]
7398 movaps xmm12,XMMWORD[((-120))+rax]
7399 movaps xmm13,XMMWORD[((-104))+rax]
7400 movaps xmm14,XMMWORD[((-88))+rax]
7401 movaps xmm15,XMMWORD[((-72))+rax]
7402 mov r15,QWORD[((-48))+rax]
7403
7404 mov r14,QWORD[((-40))+rax]
7405
7406 mov r13,QWORD[((-32))+rax]
7407
7408 mov r12,QWORD[((-24))+rax]
7409
7410 mov rbp,QWORD[((-16))+rax]
7411
7412 mov rbx,QWORD[((-8))+rax]
7413
7414 lea rsp,[rax]
7415
7416$L$epilogue_avx2:
7417 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
7418 mov rsi,QWORD[16+rsp]
7419 DB 0F3h,0C3h ;repret
7420
7421$L$SEH_end_sha1_multi_block_avx2:
7422
7423ALIGN 256
7424 DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
7425 DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
7426K_XX_XX:
7427 DD 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1
7428 DD 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1
7429 DD 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc
7430 DD 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc
7431 DD 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6
7432 DD 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6
7433 DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
7434 DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
7435DB 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
7436DB 83,72,65,49,32,109,117,108,116,105,45,98,108,111,99,107
7437DB 32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,120
7438DB 56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77
7439DB 83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110
7440DB 115,115,108,46,111,114,103,62,0
7441EXTERN __imp_RtlVirtualUnwind
7442
7443ALIGN 16
7444se_handler:
7445 push rsi
7446 push rdi
7447 push rbx
7448 push rbp
7449 push r12
7450 push r13
7451 push r14
7452 push r15
7453 pushfq
7454 sub rsp,64
7455
7456 mov rax,QWORD[120+r8]
7457 mov rbx,QWORD[248+r8]
7458
7459 mov rsi,QWORD[8+r9]
7460 mov r11,QWORD[56+r9]
7461
7462 mov r10d,DWORD[r11]
7463 lea r10,[r10*1+rsi]
7464 cmp rbx,r10
7465 jb NEAR $L$in_prologue
7466
7467 mov rax,QWORD[152+r8]
7468
7469 mov r10d,DWORD[4+r11]
7470 lea r10,[r10*1+rsi]
7471 cmp rbx,r10
7472 jae NEAR $L$in_prologue
7473
7474 mov rax,QWORD[272+rax]
7475
7476 mov rbx,QWORD[((-8))+rax]
7477 mov rbp,QWORD[((-16))+rax]
7478 mov QWORD[144+r8],rbx
7479 mov QWORD[160+r8],rbp
7480
7481 lea rsi,[((-24-160))+rax]
7482 lea rdi,[512+r8]
7483 mov ecx,20
7484 DD 0xa548f3fc
7485
7486$L$in_prologue:
7487 mov rdi,QWORD[8+rax]
7488 mov rsi,QWORD[16+rax]
7489 mov QWORD[152+r8],rax
7490 mov QWORD[168+r8],rsi
7491 mov QWORD[176+r8],rdi
7492
7493 mov rdi,QWORD[40+r9]
7494 mov rsi,r8
7495 mov ecx,154
7496 DD 0xa548f3fc
7497
7498 mov rsi,r9
7499 xor rcx,rcx
7500 mov rdx,QWORD[8+rsi]
7501 mov r8,QWORD[rsi]
7502 mov r9,QWORD[16+rsi]
7503 mov r10,QWORD[40+rsi]
7504 lea r11,[56+rsi]
7505 lea r12,[24+rsi]
7506 mov QWORD[32+rsp],r10
7507 mov QWORD[40+rsp],r11
7508 mov QWORD[48+rsp],r12
7509 mov QWORD[56+rsp],rcx
7510 call QWORD[__imp_RtlVirtualUnwind]
7511
7512 mov eax,1
7513 add rsp,64
7514 popfq
7515 pop r15
7516 pop r14
7517 pop r13
7518 pop r12
7519 pop rbp
7520 pop rbx
7521 pop rdi
7522 pop rsi
7523 DB 0F3h,0C3h ;repret
7524
7525
7526ALIGN 16
7527avx2_handler:
7528 push rsi
7529 push rdi
7530 push rbx
7531 push rbp
7532 push r12
7533 push r13
7534 push r14
7535 push r15
7536 pushfq
7537 sub rsp,64
7538
7539 mov rax,QWORD[120+r8]
7540 mov rbx,QWORD[248+r8]
7541
7542 mov rsi,QWORD[8+r9]
7543 mov r11,QWORD[56+r9]
7544
7545 mov r10d,DWORD[r11]
7546 lea r10,[r10*1+rsi]
7547 cmp rbx,r10
7548 jb NEAR $L$in_prologue
7549
7550 mov rax,QWORD[152+r8]
7551
7552 mov r10d,DWORD[4+r11]
7553 lea r10,[r10*1+rsi]
7554 cmp rbx,r10
7555 jae NEAR $L$in_prologue
7556
7557 mov rax,QWORD[544+r8]
7558
7559 mov rbx,QWORD[((-8))+rax]
7560 mov rbp,QWORD[((-16))+rax]
7561 mov r12,QWORD[((-24))+rax]
7562 mov r13,QWORD[((-32))+rax]
7563 mov r14,QWORD[((-40))+rax]
7564 mov r15,QWORD[((-48))+rax]
7565 mov QWORD[144+r8],rbx
7566 mov QWORD[160+r8],rbp
7567 mov QWORD[216+r8],r12
7568 mov QWORD[224+r8],r13
7569 mov QWORD[232+r8],r14
7570 mov QWORD[240+r8],r15
7571
7572 lea rsi,[((-56-160))+rax]
7573 lea rdi,[512+r8]
7574 mov ecx,20
7575 DD 0xa548f3fc
7576
7577 jmp NEAR $L$in_prologue
7578
7579section .pdata rdata align=4
7580ALIGN 4
7581 DD $L$SEH_begin_sha1_multi_block wrt ..imagebase
7582 DD $L$SEH_end_sha1_multi_block wrt ..imagebase
7583 DD $L$SEH_info_sha1_multi_block wrt ..imagebase
7584 DD $L$SEH_begin_sha1_multi_block_shaext wrt ..imagebase
7585 DD $L$SEH_end_sha1_multi_block_shaext wrt ..imagebase
7586 DD $L$SEH_info_sha1_multi_block_shaext wrt ..imagebase
7587 DD $L$SEH_begin_sha1_multi_block_avx wrt ..imagebase
7588 DD $L$SEH_end_sha1_multi_block_avx wrt ..imagebase
7589 DD $L$SEH_info_sha1_multi_block_avx wrt ..imagebase
7590 DD $L$SEH_begin_sha1_multi_block_avx2 wrt ..imagebase
7591 DD $L$SEH_end_sha1_multi_block_avx2 wrt ..imagebase
7592 DD $L$SEH_info_sha1_multi_block_avx2 wrt ..imagebase
7593section .xdata rdata align=8
7594ALIGN 8
7595$L$SEH_info_sha1_multi_block:
7596DB 9,0,0,0
7597 DD se_handler wrt ..imagebase
7598 DD $L$body wrt ..imagebase,$L$epilogue wrt ..imagebase
7599$L$SEH_info_sha1_multi_block_shaext:
7600DB 9,0,0,0
7601 DD se_handler wrt ..imagebase
7602 DD $L$body_shaext wrt ..imagebase,$L$epilogue_shaext wrt ..imagebase
7603$L$SEH_info_sha1_multi_block_avx:
7604DB 9,0,0,0
7605 DD se_handler wrt ..imagebase
7606 DD $L$body_avx wrt ..imagebase,$L$epilogue_avx wrt ..imagebase
7607$L$SEH_info_sha1_multi_block_avx2:
7608DB 9,0,0,0
7609 DD avx2_handler wrt ..imagebase
7610 DD $L$body_avx2 wrt ..imagebase,$L$epilogue_avx2 wrt ..imagebase
Note: See TracBrowser for help on using the repository browser.

© 2024 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette