VirtualBox

source: vbox/trunk/src/libs/openssl-1.1.0g/crypto/genasm-nasm/sha1-mb-x86_64.S@ 69881

Last change on this file since 69881 was 69881, checked in by vboxsync, 7 years ago

Update OpenSSL to 1.1.0g.
bugref:8070: src/libs maintenance

  • Property svn:eol-style set to native
File size: 147.1 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 mov rcx,QWORD[((OPENSSL_ia32cap_P+4))]
24 bt rcx,61
25 jc NEAR _shaext_shortcut
26 test ecx,268435456
27 jnz NEAR _avx_shortcut
28 mov rax,rsp
29 push rbx
30 push rbp
31 lea rsp,[((-168))+rsp]
32 movaps XMMWORD[rsp],xmm6
33 movaps XMMWORD[16+rsp],xmm7
34 movaps XMMWORD[32+rsp],xmm8
35 movaps XMMWORD[48+rsp],xmm9
36 movaps XMMWORD[(-120)+rax],xmm10
37 movaps XMMWORD[(-104)+rax],xmm11
38 movaps XMMWORD[(-88)+rax],xmm12
39 movaps XMMWORD[(-72)+rax],xmm13
40 movaps XMMWORD[(-56)+rax],xmm14
41 movaps XMMWORD[(-40)+rax],xmm15
42 sub rsp,288
43 and rsp,-256
44 mov QWORD[272+rsp],rax
45$L$body:
46 lea rbp,[K_XX_XX]
47 lea rbx,[256+rsp]
48
49$L$oop_grande:
50 mov DWORD[280+rsp],edx
51 xor edx,edx
52 mov r8,QWORD[rsi]
53 mov ecx,DWORD[8+rsi]
54 cmp ecx,edx
55 cmovg edx,ecx
56 test ecx,ecx
57 mov DWORD[rbx],ecx
58 cmovle r8,rbp
59 mov r9,QWORD[16+rsi]
60 mov ecx,DWORD[24+rsi]
61 cmp ecx,edx
62 cmovg edx,ecx
63 test ecx,ecx
64 mov DWORD[4+rbx],ecx
65 cmovle r9,rbp
66 mov r10,QWORD[32+rsi]
67 mov ecx,DWORD[40+rsi]
68 cmp ecx,edx
69 cmovg edx,ecx
70 test ecx,ecx
71 mov DWORD[8+rbx],ecx
72 cmovle r10,rbp
73 mov r11,QWORD[48+rsi]
74 mov ecx,DWORD[56+rsi]
75 cmp ecx,edx
76 cmovg edx,ecx
77 test ecx,ecx
78 mov DWORD[12+rbx],ecx
79 cmovle r11,rbp
80 test edx,edx
81 jz NEAR $L$done
82
83 movdqu xmm10,XMMWORD[rdi]
84 lea rax,[128+rsp]
85 movdqu xmm11,XMMWORD[32+rdi]
86 movdqu xmm12,XMMWORD[64+rdi]
87 movdqu xmm13,XMMWORD[96+rdi]
88 movdqu xmm14,XMMWORD[128+rdi]
89 movdqa xmm5,XMMWORD[96+rbp]
90 movdqa xmm15,XMMWORD[((-32))+rbp]
91 jmp NEAR $L$oop
92
93ALIGN 32
94$L$oop:
95 movd xmm0,DWORD[r8]
96 lea r8,[64+r8]
97 movd xmm2,DWORD[r9]
98 lea r9,[64+r9]
99 movd xmm3,DWORD[r10]
100 lea r10,[64+r10]
101 movd xmm4,DWORD[r11]
102 lea r11,[64+r11]
103 punpckldq xmm0,xmm3
104 movd xmm1,DWORD[((-60))+r8]
105 punpckldq xmm2,xmm4
106 movd xmm9,DWORD[((-60))+r9]
107 punpckldq xmm0,xmm2
108 movd xmm8,DWORD[((-60))+r10]
109DB 102,15,56,0,197
110 movd xmm7,DWORD[((-60))+r11]
111 punpckldq xmm1,xmm8
112 movdqa xmm8,xmm10
113 paddd xmm14,xmm15
114 punpckldq xmm9,xmm7
115 movdqa xmm7,xmm11
116 movdqa xmm6,xmm11
117 pslld xmm8,5
118 pandn xmm7,xmm13
119 pand xmm6,xmm12
120 punpckldq xmm1,xmm9
121 movdqa xmm9,xmm10
122
123 movdqa XMMWORD[(0-128)+rax],xmm0
124 paddd xmm14,xmm0
125 movd xmm2,DWORD[((-56))+r8]
126 psrld xmm9,27
127 pxor xmm6,xmm7
128 movdqa xmm7,xmm11
129
130 por xmm8,xmm9
131 movd xmm9,DWORD[((-56))+r9]
132 pslld xmm7,30
133 paddd xmm14,xmm6
134
135 psrld xmm11,2
136 paddd xmm14,xmm8
137DB 102,15,56,0,205
138 movd xmm8,DWORD[((-56))+r10]
139 por xmm11,xmm7
140 movd xmm7,DWORD[((-56))+r11]
141 punpckldq xmm2,xmm8
142 movdqa xmm8,xmm14
143 paddd xmm13,xmm15
144 punpckldq xmm9,xmm7
145 movdqa xmm7,xmm10
146 movdqa xmm6,xmm10
147 pslld xmm8,5
148 pandn xmm7,xmm12
149 pand xmm6,xmm11
150 punpckldq xmm2,xmm9
151 movdqa xmm9,xmm14
152
153 movdqa XMMWORD[(16-128)+rax],xmm1
154 paddd xmm13,xmm1
155 movd xmm3,DWORD[((-52))+r8]
156 psrld xmm9,27
157 pxor xmm6,xmm7
158 movdqa xmm7,xmm10
159
160 por xmm8,xmm9
161 movd xmm9,DWORD[((-52))+r9]
162 pslld xmm7,30
163 paddd xmm13,xmm6
164
165 psrld xmm10,2
166 paddd xmm13,xmm8
167DB 102,15,56,0,213
168 movd xmm8,DWORD[((-52))+r10]
169 por xmm10,xmm7
170 movd xmm7,DWORD[((-52))+r11]
171 punpckldq xmm3,xmm8
172 movdqa xmm8,xmm13
173 paddd xmm12,xmm15
174 punpckldq xmm9,xmm7
175 movdqa xmm7,xmm14
176 movdqa xmm6,xmm14
177 pslld xmm8,5
178 pandn xmm7,xmm11
179 pand xmm6,xmm10
180 punpckldq xmm3,xmm9
181 movdqa xmm9,xmm13
182
183 movdqa XMMWORD[(32-128)+rax],xmm2
184 paddd xmm12,xmm2
185 movd xmm4,DWORD[((-48))+r8]
186 psrld xmm9,27
187 pxor xmm6,xmm7
188 movdqa xmm7,xmm14
189
190 por xmm8,xmm9
191 movd xmm9,DWORD[((-48))+r9]
192 pslld xmm7,30
193 paddd xmm12,xmm6
194
195 psrld xmm14,2
196 paddd xmm12,xmm8
197DB 102,15,56,0,221
198 movd xmm8,DWORD[((-48))+r10]
199 por xmm14,xmm7
200 movd xmm7,DWORD[((-48))+r11]
201 punpckldq xmm4,xmm8
202 movdqa xmm8,xmm12
203 paddd xmm11,xmm15
204 punpckldq xmm9,xmm7
205 movdqa xmm7,xmm13
206 movdqa xmm6,xmm13
207 pslld xmm8,5
208 pandn xmm7,xmm10
209 pand xmm6,xmm14
210 punpckldq xmm4,xmm9
211 movdqa xmm9,xmm12
212
213 movdqa XMMWORD[(48-128)+rax],xmm3
214 paddd xmm11,xmm3
215 movd xmm0,DWORD[((-44))+r8]
216 psrld xmm9,27
217 pxor xmm6,xmm7
218 movdqa xmm7,xmm13
219
220 por xmm8,xmm9
221 movd xmm9,DWORD[((-44))+r9]
222 pslld xmm7,30
223 paddd xmm11,xmm6
224
225 psrld xmm13,2
226 paddd xmm11,xmm8
227DB 102,15,56,0,229
228 movd xmm8,DWORD[((-44))+r10]
229 por xmm13,xmm7
230 movd xmm7,DWORD[((-44))+r11]
231 punpckldq xmm0,xmm8
232 movdqa xmm8,xmm11
233 paddd xmm10,xmm15
234 punpckldq xmm9,xmm7
235 movdqa xmm7,xmm12
236 movdqa xmm6,xmm12
237 pslld xmm8,5
238 pandn xmm7,xmm14
239 pand xmm6,xmm13
240 punpckldq xmm0,xmm9
241 movdqa xmm9,xmm11
242
243 movdqa XMMWORD[(64-128)+rax],xmm4
244 paddd xmm10,xmm4
245 movd xmm1,DWORD[((-40))+r8]
246 psrld xmm9,27
247 pxor xmm6,xmm7
248 movdqa xmm7,xmm12
249
250 por xmm8,xmm9
251 movd xmm9,DWORD[((-40))+r9]
252 pslld xmm7,30
253 paddd xmm10,xmm6
254
255 psrld xmm12,2
256 paddd xmm10,xmm8
257DB 102,15,56,0,197
258 movd xmm8,DWORD[((-40))+r10]
259 por xmm12,xmm7
260 movd xmm7,DWORD[((-40))+r11]
261 punpckldq xmm1,xmm8
262 movdqa xmm8,xmm10
263 paddd xmm14,xmm15
264 punpckldq xmm9,xmm7
265 movdqa xmm7,xmm11
266 movdqa xmm6,xmm11
267 pslld xmm8,5
268 pandn xmm7,xmm13
269 pand xmm6,xmm12
270 punpckldq xmm1,xmm9
271 movdqa xmm9,xmm10
272
273 movdqa XMMWORD[(80-128)+rax],xmm0
274 paddd xmm14,xmm0
275 movd xmm2,DWORD[((-36))+r8]
276 psrld xmm9,27
277 pxor xmm6,xmm7
278 movdqa xmm7,xmm11
279
280 por xmm8,xmm9
281 movd xmm9,DWORD[((-36))+r9]
282 pslld xmm7,30
283 paddd xmm14,xmm6
284
285 psrld xmm11,2
286 paddd xmm14,xmm8
287DB 102,15,56,0,205
288 movd xmm8,DWORD[((-36))+r10]
289 por xmm11,xmm7
290 movd xmm7,DWORD[((-36))+r11]
291 punpckldq xmm2,xmm8
292 movdqa xmm8,xmm14
293 paddd xmm13,xmm15
294 punpckldq xmm9,xmm7
295 movdqa xmm7,xmm10
296 movdqa xmm6,xmm10
297 pslld xmm8,5
298 pandn xmm7,xmm12
299 pand xmm6,xmm11
300 punpckldq xmm2,xmm9
301 movdqa xmm9,xmm14
302
303 movdqa XMMWORD[(96-128)+rax],xmm1
304 paddd xmm13,xmm1
305 movd xmm3,DWORD[((-32))+r8]
306 psrld xmm9,27
307 pxor xmm6,xmm7
308 movdqa xmm7,xmm10
309
310 por xmm8,xmm9
311 movd xmm9,DWORD[((-32))+r9]
312 pslld xmm7,30
313 paddd xmm13,xmm6
314
315 psrld xmm10,2
316 paddd xmm13,xmm8
317DB 102,15,56,0,213
318 movd xmm8,DWORD[((-32))+r10]
319 por xmm10,xmm7
320 movd xmm7,DWORD[((-32))+r11]
321 punpckldq xmm3,xmm8
322 movdqa xmm8,xmm13
323 paddd xmm12,xmm15
324 punpckldq xmm9,xmm7
325 movdqa xmm7,xmm14
326 movdqa xmm6,xmm14
327 pslld xmm8,5
328 pandn xmm7,xmm11
329 pand xmm6,xmm10
330 punpckldq xmm3,xmm9
331 movdqa xmm9,xmm13
332
333 movdqa XMMWORD[(112-128)+rax],xmm2
334 paddd xmm12,xmm2
335 movd xmm4,DWORD[((-28))+r8]
336 psrld xmm9,27
337 pxor xmm6,xmm7
338 movdqa xmm7,xmm14
339
340 por xmm8,xmm9
341 movd xmm9,DWORD[((-28))+r9]
342 pslld xmm7,30
343 paddd xmm12,xmm6
344
345 psrld xmm14,2
346 paddd xmm12,xmm8
347DB 102,15,56,0,221
348 movd xmm8,DWORD[((-28))+r10]
349 por xmm14,xmm7
350 movd xmm7,DWORD[((-28))+r11]
351 punpckldq xmm4,xmm8
352 movdqa xmm8,xmm12
353 paddd xmm11,xmm15
354 punpckldq xmm9,xmm7
355 movdqa xmm7,xmm13
356 movdqa xmm6,xmm13
357 pslld xmm8,5
358 pandn xmm7,xmm10
359 pand xmm6,xmm14
360 punpckldq xmm4,xmm9
361 movdqa xmm9,xmm12
362
363 movdqa XMMWORD[(128-128)+rax],xmm3
364 paddd xmm11,xmm3
365 movd xmm0,DWORD[((-24))+r8]
366 psrld xmm9,27
367 pxor xmm6,xmm7
368 movdqa xmm7,xmm13
369
370 por xmm8,xmm9
371 movd xmm9,DWORD[((-24))+r9]
372 pslld xmm7,30
373 paddd xmm11,xmm6
374
375 psrld xmm13,2
376 paddd xmm11,xmm8
377DB 102,15,56,0,229
378 movd xmm8,DWORD[((-24))+r10]
379 por xmm13,xmm7
380 movd xmm7,DWORD[((-24))+r11]
381 punpckldq xmm0,xmm8
382 movdqa xmm8,xmm11
383 paddd xmm10,xmm15
384 punpckldq xmm9,xmm7
385 movdqa xmm7,xmm12
386 movdqa xmm6,xmm12
387 pslld xmm8,5
388 pandn xmm7,xmm14
389 pand xmm6,xmm13
390 punpckldq xmm0,xmm9
391 movdqa xmm9,xmm11
392
393 movdqa XMMWORD[(144-128)+rax],xmm4
394 paddd xmm10,xmm4
395 movd xmm1,DWORD[((-20))+r8]
396 psrld xmm9,27
397 pxor xmm6,xmm7
398 movdqa xmm7,xmm12
399
400 por xmm8,xmm9
401 movd xmm9,DWORD[((-20))+r9]
402 pslld xmm7,30
403 paddd xmm10,xmm6
404
405 psrld xmm12,2
406 paddd xmm10,xmm8
407DB 102,15,56,0,197
408 movd xmm8,DWORD[((-20))+r10]
409 por xmm12,xmm7
410 movd xmm7,DWORD[((-20))+r11]
411 punpckldq xmm1,xmm8
412 movdqa xmm8,xmm10
413 paddd xmm14,xmm15
414 punpckldq xmm9,xmm7
415 movdqa xmm7,xmm11
416 movdqa xmm6,xmm11
417 pslld xmm8,5
418 pandn xmm7,xmm13
419 pand xmm6,xmm12
420 punpckldq xmm1,xmm9
421 movdqa xmm9,xmm10
422
423 movdqa XMMWORD[(160-128)+rax],xmm0
424 paddd xmm14,xmm0
425 movd xmm2,DWORD[((-16))+r8]
426 psrld xmm9,27
427 pxor xmm6,xmm7
428 movdqa xmm7,xmm11
429
430 por xmm8,xmm9
431 movd xmm9,DWORD[((-16))+r9]
432 pslld xmm7,30
433 paddd xmm14,xmm6
434
435 psrld xmm11,2
436 paddd xmm14,xmm8
437DB 102,15,56,0,205
438 movd xmm8,DWORD[((-16))+r10]
439 por xmm11,xmm7
440 movd xmm7,DWORD[((-16))+r11]
441 punpckldq xmm2,xmm8
442 movdqa xmm8,xmm14
443 paddd xmm13,xmm15
444 punpckldq xmm9,xmm7
445 movdqa xmm7,xmm10
446 movdqa xmm6,xmm10
447 pslld xmm8,5
448 pandn xmm7,xmm12
449 pand xmm6,xmm11
450 punpckldq xmm2,xmm9
451 movdqa xmm9,xmm14
452
453 movdqa XMMWORD[(176-128)+rax],xmm1
454 paddd xmm13,xmm1
455 movd xmm3,DWORD[((-12))+r8]
456 psrld xmm9,27
457 pxor xmm6,xmm7
458 movdqa xmm7,xmm10
459
460 por xmm8,xmm9
461 movd xmm9,DWORD[((-12))+r9]
462 pslld xmm7,30
463 paddd xmm13,xmm6
464
465 psrld xmm10,2
466 paddd xmm13,xmm8
467DB 102,15,56,0,213
468 movd xmm8,DWORD[((-12))+r10]
469 por xmm10,xmm7
470 movd xmm7,DWORD[((-12))+r11]
471 punpckldq xmm3,xmm8
472 movdqa xmm8,xmm13
473 paddd xmm12,xmm15
474 punpckldq xmm9,xmm7
475 movdqa xmm7,xmm14
476 movdqa xmm6,xmm14
477 pslld xmm8,5
478 pandn xmm7,xmm11
479 pand xmm6,xmm10
480 punpckldq xmm3,xmm9
481 movdqa xmm9,xmm13
482
483 movdqa XMMWORD[(192-128)+rax],xmm2
484 paddd xmm12,xmm2
485 movd xmm4,DWORD[((-8))+r8]
486 psrld xmm9,27
487 pxor xmm6,xmm7
488 movdqa xmm7,xmm14
489
490 por xmm8,xmm9
491 movd xmm9,DWORD[((-8))+r9]
492 pslld xmm7,30
493 paddd xmm12,xmm6
494
495 psrld xmm14,2
496 paddd xmm12,xmm8
497DB 102,15,56,0,221
498 movd xmm8,DWORD[((-8))+r10]
499 por xmm14,xmm7
500 movd xmm7,DWORD[((-8))+r11]
501 punpckldq xmm4,xmm8
502 movdqa xmm8,xmm12
503 paddd xmm11,xmm15
504 punpckldq xmm9,xmm7
505 movdqa xmm7,xmm13
506 movdqa xmm6,xmm13
507 pslld xmm8,5
508 pandn xmm7,xmm10
509 pand xmm6,xmm14
510 punpckldq xmm4,xmm9
511 movdqa xmm9,xmm12
512
513 movdqa XMMWORD[(208-128)+rax],xmm3
514 paddd xmm11,xmm3
515 movd xmm0,DWORD[((-4))+r8]
516 psrld xmm9,27
517 pxor xmm6,xmm7
518 movdqa xmm7,xmm13
519
520 por xmm8,xmm9
521 movd xmm9,DWORD[((-4))+r9]
522 pslld xmm7,30
523 paddd xmm11,xmm6
524
525 psrld xmm13,2
526 paddd xmm11,xmm8
527DB 102,15,56,0,229
528 movd xmm8,DWORD[((-4))+r10]
529 por xmm13,xmm7
530 movdqa xmm1,XMMWORD[((0-128))+rax]
531 movd xmm7,DWORD[((-4))+r11]
532 punpckldq xmm0,xmm8
533 movdqa xmm8,xmm11
534 paddd xmm10,xmm15
535 punpckldq xmm9,xmm7
536 movdqa xmm7,xmm12
537 movdqa xmm6,xmm12
538 pslld xmm8,5
539 prefetcht0 [63+r8]
540 pandn xmm7,xmm14
541 pand xmm6,xmm13
542 punpckldq xmm0,xmm9
543 movdqa xmm9,xmm11
544
545 movdqa XMMWORD[(224-128)+rax],xmm4
546 paddd xmm10,xmm4
547 psrld xmm9,27
548 pxor xmm6,xmm7
549 movdqa xmm7,xmm12
550 prefetcht0 [63+r9]
551
552 por xmm8,xmm9
553 pslld xmm7,30
554 paddd xmm10,xmm6
555 prefetcht0 [63+r10]
556
557 psrld xmm12,2
558 paddd xmm10,xmm8
559DB 102,15,56,0,197
560 prefetcht0 [63+r11]
561 por xmm12,xmm7
562 movdqa xmm2,XMMWORD[((16-128))+rax]
563 pxor xmm1,xmm3
564 movdqa xmm3,XMMWORD[((32-128))+rax]
565
566 movdqa xmm8,xmm10
567 pxor xmm1,XMMWORD[((128-128))+rax]
568 paddd xmm14,xmm15
569 movdqa xmm7,xmm11
570 pslld xmm8,5
571 pxor xmm1,xmm3
572 movdqa xmm6,xmm11
573 pandn xmm7,xmm13
574 movdqa xmm5,xmm1
575 pand xmm6,xmm12
576 movdqa xmm9,xmm10
577 psrld xmm5,31
578 paddd xmm1,xmm1
579
580 movdqa XMMWORD[(240-128)+rax],xmm0
581 paddd xmm14,xmm0
582 psrld xmm9,27
583 pxor xmm6,xmm7
584
585 movdqa xmm7,xmm11
586 por xmm8,xmm9
587 pslld xmm7,30
588 paddd xmm14,xmm6
589
590 psrld xmm11,2
591 paddd xmm14,xmm8
592 por xmm1,xmm5
593 por xmm11,xmm7
594 pxor xmm2,xmm4
595 movdqa xmm4,XMMWORD[((48-128))+rax]
596
597 movdqa xmm8,xmm14
598 pxor xmm2,XMMWORD[((144-128))+rax]
599 paddd xmm13,xmm15
600 movdqa xmm7,xmm10
601 pslld xmm8,5
602 pxor xmm2,xmm4
603 movdqa xmm6,xmm10
604 pandn xmm7,xmm12
605 movdqa xmm5,xmm2
606 pand xmm6,xmm11
607 movdqa xmm9,xmm14
608 psrld xmm5,31
609 paddd xmm2,xmm2
610
611 movdqa XMMWORD[(0-128)+rax],xmm1
612 paddd xmm13,xmm1
613 psrld xmm9,27
614 pxor xmm6,xmm7
615
616 movdqa xmm7,xmm10
617 por xmm8,xmm9
618 pslld xmm7,30
619 paddd xmm13,xmm6
620
621 psrld xmm10,2
622 paddd xmm13,xmm8
623 por xmm2,xmm5
624 por xmm10,xmm7
625 pxor xmm3,xmm0
626 movdqa xmm0,XMMWORD[((64-128))+rax]
627
628 movdqa xmm8,xmm13
629 pxor xmm3,XMMWORD[((160-128))+rax]
630 paddd xmm12,xmm15
631 movdqa xmm7,xmm14
632 pslld xmm8,5
633 pxor xmm3,xmm0
634 movdqa xmm6,xmm14
635 pandn xmm7,xmm11
636 movdqa xmm5,xmm3
637 pand xmm6,xmm10
638 movdqa xmm9,xmm13
639 psrld xmm5,31
640 paddd xmm3,xmm3
641
642 movdqa XMMWORD[(16-128)+rax],xmm2
643 paddd xmm12,xmm2
644 psrld xmm9,27
645 pxor xmm6,xmm7
646
647 movdqa xmm7,xmm14
648 por xmm8,xmm9
649 pslld xmm7,30
650 paddd xmm12,xmm6
651
652 psrld xmm14,2
653 paddd xmm12,xmm8
654 por xmm3,xmm5
655 por xmm14,xmm7
656 pxor xmm4,xmm1
657 movdqa xmm1,XMMWORD[((80-128))+rax]
658
659 movdqa xmm8,xmm12
660 pxor xmm4,XMMWORD[((176-128))+rax]
661 paddd xmm11,xmm15
662 movdqa xmm7,xmm13
663 pslld xmm8,5
664 pxor xmm4,xmm1
665 movdqa xmm6,xmm13
666 pandn xmm7,xmm10
667 movdqa xmm5,xmm4
668 pand xmm6,xmm14
669 movdqa xmm9,xmm12
670 psrld xmm5,31
671 paddd xmm4,xmm4
672
673 movdqa XMMWORD[(32-128)+rax],xmm3
674 paddd xmm11,xmm3
675 psrld xmm9,27
676 pxor xmm6,xmm7
677
678 movdqa xmm7,xmm13
679 por xmm8,xmm9
680 pslld xmm7,30
681 paddd xmm11,xmm6
682
683 psrld xmm13,2
684 paddd xmm11,xmm8
685 por xmm4,xmm5
686 por xmm13,xmm7
687 pxor xmm0,xmm2
688 movdqa xmm2,XMMWORD[((96-128))+rax]
689
690 movdqa xmm8,xmm11
691 pxor xmm0,XMMWORD[((192-128))+rax]
692 paddd xmm10,xmm15
693 movdqa xmm7,xmm12
694 pslld xmm8,5
695 pxor xmm0,xmm2
696 movdqa xmm6,xmm12
697 pandn xmm7,xmm14
698 movdqa xmm5,xmm0
699 pand xmm6,xmm13
700 movdqa xmm9,xmm11
701 psrld xmm5,31
702 paddd xmm0,xmm0
703
704 movdqa XMMWORD[(48-128)+rax],xmm4
705 paddd xmm10,xmm4
706 psrld xmm9,27
707 pxor xmm6,xmm7
708
709 movdqa xmm7,xmm12
710 por xmm8,xmm9
711 pslld xmm7,30
712 paddd xmm10,xmm6
713
714 psrld xmm12,2
715 paddd xmm10,xmm8
716 por xmm0,xmm5
717 por xmm12,xmm7
718 movdqa xmm15,XMMWORD[rbp]
719 pxor xmm1,xmm3
720 movdqa xmm3,XMMWORD[((112-128))+rax]
721
722 movdqa xmm8,xmm10
723 movdqa xmm6,xmm13
724 pxor xmm1,XMMWORD[((208-128))+rax]
725 paddd xmm14,xmm15
726 pslld xmm8,5
727 pxor xmm6,xmm11
728
729 movdqa xmm9,xmm10
730 movdqa XMMWORD[(64-128)+rax],xmm0
731 paddd xmm14,xmm0
732 pxor xmm1,xmm3
733 psrld xmm9,27
734 pxor xmm6,xmm12
735 movdqa xmm7,xmm11
736
737 pslld xmm7,30
738 movdqa xmm5,xmm1
739 por xmm8,xmm9
740 psrld xmm5,31
741 paddd xmm14,xmm6
742 paddd xmm1,xmm1
743
744 psrld xmm11,2
745 paddd xmm14,xmm8
746 por xmm1,xmm5
747 por xmm11,xmm7
748 pxor xmm2,xmm4
749 movdqa xmm4,XMMWORD[((128-128))+rax]
750
751 movdqa xmm8,xmm14
752 movdqa xmm6,xmm12
753 pxor xmm2,XMMWORD[((224-128))+rax]
754 paddd xmm13,xmm15
755 pslld xmm8,5
756 pxor xmm6,xmm10
757
758 movdqa xmm9,xmm14
759 movdqa XMMWORD[(80-128)+rax],xmm1
760 paddd xmm13,xmm1
761 pxor xmm2,xmm4
762 psrld xmm9,27
763 pxor xmm6,xmm11
764 movdqa xmm7,xmm10
765
766 pslld xmm7,30
767 movdqa xmm5,xmm2
768 por xmm8,xmm9
769 psrld xmm5,31
770 paddd xmm13,xmm6
771 paddd xmm2,xmm2
772
773 psrld xmm10,2
774 paddd xmm13,xmm8
775 por xmm2,xmm5
776 por xmm10,xmm7
777 pxor xmm3,xmm0
778 movdqa xmm0,XMMWORD[((144-128))+rax]
779
780 movdqa xmm8,xmm13
781 movdqa xmm6,xmm11
782 pxor xmm3,XMMWORD[((240-128))+rax]
783 paddd xmm12,xmm15
784 pslld xmm8,5
785 pxor xmm6,xmm14
786
787 movdqa xmm9,xmm13
788 movdqa XMMWORD[(96-128)+rax],xmm2
789 paddd xmm12,xmm2
790 pxor xmm3,xmm0
791 psrld xmm9,27
792 pxor xmm6,xmm10
793 movdqa xmm7,xmm14
794
795 pslld xmm7,30
796 movdqa xmm5,xmm3
797 por xmm8,xmm9
798 psrld xmm5,31
799 paddd xmm12,xmm6
800 paddd xmm3,xmm3
801
802 psrld xmm14,2
803 paddd xmm12,xmm8
804 por xmm3,xmm5
805 por xmm14,xmm7
806 pxor xmm4,xmm1
807 movdqa xmm1,XMMWORD[((160-128))+rax]
808
809 movdqa xmm8,xmm12
810 movdqa xmm6,xmm10
811 pxor xmm4,XMMWORD[((0-128))+rax]
812 paddd xmm11,xmm15
813 pslld xmm8,5
814 pxor xmm6,xmm13
815
816 movdqa xmm9,xmm12
817 movdqa XMMWORD[(112-128)+rax],xmm3
818 paddd xmm11,xmm3
819 pxor xmm4,xmm1
820 psrld xmm9,27
821 pxor xmm6,xmm14
822 movdqa xmm7,xmm13
823
824 pslld xmm7,30
825 movdqa xmm5,xmm4
826 por xmm8,xmm9
827 psrld xmm5,31
828 paddd xmm11,xmm6
829 paddd xmm4,xmm4
830
831 psrld xmm13,2
832 paddd xmm11,xmm8
833 por xmm4,xmm5
834 por xmm13,xmm7
835 pxor xmm0,xmm2
836 movdqa xmm2,XMMWORD[((176-128))+rax]
837
838 movdqa xmm8,xmm11
839 movdqa xmm6,xmm14
840 pxor xmm0,XMMWORD[((16-128))+rax]
841 paddd xmm10,xmm15
842 pslld xmm8,5
843 pxor xmm6,xmm12
844
845 movdqa xmm9,xmm11
846 movdqa XMMWORD[(128-128)+rax],xmm4
847 paddd xmm10,xmm4
848 pxor xmm0,xmm2
849 psrld xmm9,27
850 pxor xmm6,xmm13
851 movdqa xmm7,xmm12
852
853 pslld xmm7,30
854 movdqa xmm5,xmm0
855 por xmm8,xmm9
856 psrld xmm5,31
857 paddd xmm10,xmm6
858 paddd xmm0,xmm0
859
860 psrld xmm12,2
861 paddd xmm10,xmm8
862 por xmm0,xmm5
863 por xmm12,xmm7
864 pxor xmm1,xmm3
865 movdqa xmm3,XMMWORD[((192-128))+rax]
866
867 movdqa xmm8,xmm10
868 movdqa xmm6,xmm13
869 pxor xmm1,XMMWORD[((32-128))+rax]
870 paddd xmm14,xmm15
871 pslld xmm8,5
872 pxor xmm6,xmm11
873
874 movdqa xmm9,xmm10
875 movdqa XMMWORD[(144-128)+rax],xmm0
876 paddd xmm14,xmm0
877 pxor xmm1,xmm3
878 psrld xmm9,27
879 pxor xmm6,xmm12
880 movdqa xmm7,xmm11
881
882 pslld xmm7,30
883 movdqa xmm5,xmm1
884 por xmm8,xmm9
885 psrld xmm5,31
886 paddd xmm14,xmm6
887 paddd xmm1,xmm1
888
889 psrld xmm11,2
890 paddd xmm14,xmm8
891 por xmm1,xmm5
892 por xmm11,xmm7
893 pxor xmm2,xmm4
894 movdqa xmm4,XMMWORD[((208-128))+rax]
895
896 movdqa xmm8,xmm14
897 movdqa xmm6,xmm12
898 pxor xmm2,XMMWORD[((48-128))+rax]
899 paddd xmm13,xmm15
900 pslld xmm8,5
901 pxor xmm6,xmm10
902
903 movdqa xmm9,xmm14
904 movdqa XMMWORD[(160-128)+rax],xmm1
905 paddd xmm13,xmm1
906 pxor xmm2,xmm4
907 psrld xmm9,27
908 pxor xmm6,xmm11
909 movdqa xmm7,xmm10
910
911 pslld xmm7,30
912 movdqa xmm5,xmm2
913 por xmm8,xmm9
914 psrld xmm5,31
915 paddd xmm13,xmm6
916 paddd xmm2,xmm2
917
918 psrld xmm10,2
919 paddd xmm13,xmm8
920 por xmm2,xmm5
921 por xmm10,xmm7
922 pxor xmm3,xmm0
923 movdqa xmm0,XMMWORD[((224-128))+rax]
924
925 movdqa xmm8,xmm13
926 movdqa xmm6,xmm11
927 pxor xmm3,XMMWORD[((64-128))+rax]
928 paddd xmm12,xmm15
929 pslld xmm8,5
930 pxor xmm6,xmm14
931
932 movdqa xmm9,xmm13
933 movdqa XMMWORD[(176-128)+rax],xmm2
934 paddd xmm12,xmm2
935 pxor xmm3,xmm0
936 psrld xmm9,27
937 pxor xmm6,xmm10
938 movdqa xmm7,xmm14
939
940 pslld xmm7,30
941 movdqa xmm5,xmm3
942 por xmm8,xmm9
943 psrld xmm5,31
944 paddd xmm12,xmm6
945 paddd xmm3,xmm3
946
947 psrld xmm14,2
948 paddd xmm12,xmm8
949 por xmm3,xmm5
950 por xmm14,xmm7
951 pxor xmm4,xmm1
952 movdqa xmm1,XMMWORD[((240-128))+rax]
953
954 movdqa xmm8,xmm12
955 movdqa xmm6,xmm10
956 pxor xmm4,XMMWORD[((80-128))+rax]
957 paddd xmm11,xmm15
958 pslld xmm8,5
959 pxor xmm6,xmm13
960
961 movdqa xmm9,xmm12
962 movdqa XMMWORD[(192-128)+rax],xmm3
963 paddd xmm11,xmm3
964 pxor xmm4,xmm1
965 psrld xmm9,27
966 pxor xmm6,xmm14
967 movdqa xmm7,xmm13
968
969 pslld xmm7,30
970 movdqa xmm5,xmm4
971 por xmm8,xmm9
972 psrld xmm5,31
973 paddd xmm11,xmm6
974 paddd xmm4,xmm4
975
976 psrld xmm13,2
977 paddd xmm11,xmm8
978 por xmm4,xmm5
979 por xmm13,xmm7
980 pxor xmm0,xmm2
981 movdqa xmm2,XMMWORD[((0-128))+rax]
982
983 movdqa xmm8,xmm11
984 movdqa xmm6,xmm14
985 pxor xmm0,XMMWORD[((96-128))+rax]
986 paddd xmm10,xmm15
987 pslld xmm8,5
988 pxor xmm6,xmm12
989
990 movdqa xmm9,xmm11
991 movdqa XMMWORD[(208-128)+rax],xmm4
992 paddd xmm10,xmm4
993 pxor xmm0,xmm2
994 psrld xmm9,27
995 pxor xmm6,xmm13
996 movdqa xmm7,xmm12
997
998 pslld xmm7,30
999 movdqa xmm5,xmm0
1000 por xmm8,xmm9
1001 psrld xmm5,31
1002 paddd xmm10,xmm6
1003 paddd xmm0,xmm0
1004
1005 psrld xmm12,2
1006 paddd xmm10,xmm8
1007 por xmm0,xmm5
1008 por xmm12,xmm7
1009 pxor xmm1,xmm3
1010 movdqa xmm3,XMMWORD[((16-128))+rax]
1011
1012 movdqa xmm8,xmm10
1013 movdqa xmm6,xmm13
1014 pxor xmm1,XMMWORD[((112-128))+rax]
1015 paddd xmm14,xmm15
1016 pslld xmm8,5
1017 pxor xmm6,xmm11
1018
1019 movdqa xmm9,xmm10
1020 movdqa XMMWORD[(224-128)+rax],xmm0
1021 paddd xmm14,xmm0
1022 pxor xmm1,xmm3
1023 psrld xmm9,27
1024 pxor xmm6,xmm12
1025 movdqa xmm7,xmm11
1026
1027 pslld xmm7,30
1028 movdqa xmm5,xmm1
1029 por xmm8,xmm9
1030 psrld xmm5,31
1031 paddd xmm14,xmm6
1032 paddd xmm1,xmm1
1033
1034 psrld xmm11,2
1035 paddd xmm14,xmm8
1036 por xmm1,xmm5
1037 por xmm11,xmm7
1038 pxor xmm2,xmm4
1039 movdqa xmm4,XMMWORD[((32-128))+rax]
1040
1041 movdqa xmm8,xmm14
1042 movdqa xmm6,xmm12
1043 pxor xmm2,XMMWORD[((128-128))+rax]
1044 paddd xmm13,xmm15
1045 pslld xmm8,5
1046 pxor xmm6,xmm10
1047
1048 movdqa xmm9,xmm14
1049 movdqa XMMWORD[(240-128)+rax],xmm1
1050 paddd xmm13,xmm1
1051 pxor xmm2,xmm4
1052 psrld xmm9,27
1053 pxor xmm6,xmm11
1054 movdqa xmm7,xmm10
1055
1056 pslld xmm7,30
1057 movdqa xmm5,xmm2
1058 por xmm8,xmm9
1059 psrld xmm5,31
1060 paddd xmm13,xmm6
1061 paddd xmm2,xmm2
1062
1063 psrld xmm10,2
1064 paddd xmm13,xmm8
1065 por xmm2,xmm5
1066 por xmm10,xmm7
1067 pxor xmm3,xmm0
1068 movdqa xmm0,XMMWORD[((48-128))+rax]
1069
1070 movdqa xmm8,xmm13
1071 movdqa xmm6,xmm11
1072 pxor xmm3,XMMWORD[((144-128))+rax]
1073 paddd xmm12,xmm15
1074 pslld xmm8,5
1075 pxor xmm6,xmm14
1076
1077 movdqa xmm9,xmm13
1078 movdqa XMMWORD[(0-128)+rax],xmm2
1079 paddd xmm12,xmm2
1080 pxor xmm3,xmm0
1081 psrld xmm9,27
1082 pxor xmm6,xmm10
1083 movdqa xmm7,xmm14
1084
1085 pslld xmm7,30
1086 movdqa xmm5,xmm3
1087 por xmm8,xmm9
1088 psrld xmm5,31
1089 paddd xmm12,xmm6
1090 paddd xmm3,xmm3
1091
1092 psrld xmm14,2
1093 paddd xmm12,xmm8
1094 por xmm3,xmm5
1095 por xmm14,xmm7
1096 pxor xmm4,xmm1
1097 movdqa xmm1,XMMWORD[((64-128))+rax]
1098
1099 movdqa xmm8,xmm12
1100 movdqa xmm6,xmm10
1101 pxor xmm4,XMMWORD[((160-128))+rax]
1102 paddd xmm11,xmm15
1103 pslld xmm8,5
1104 pxor xmm6,xmm13
1105
1106 movdqa xmm9,xmm12
1107 movdqa XMMWORD[(16-128)+rax],xmm3
1108 paddd xmm11,xmm3
1109 pxor xmm4,xmm1
1110 psrld xmm9,27
1111 pxor xmm6,xmm14
1112 movdqa xmm7,xmm13
1113
1114 pslld xmm7,30
1115 movdqa xmm5,xmm4
1116 por xmm8,xmm9
1117 psrld xmm5,31
1118 paddd xmm11,xmm6
1119 paddd xmm4,xmm4
1120
1121 psrld xmm13,2
1122 paddd xmm11,xmm8
1123 por xmm4,xmm5
1124 por xmm13,xmm7
1125 pxor xmm0,xmm2
1126 movdqa xmm2,XMMWORD[((80-128))+rax]
1127
1128 movdqa xmm8,xmm11
1129 movdqa xmm6,xmm14
1130 pxor xmm0,XMMWORD[((176-128))+rax]
1131 paddd xmm10,xmm15
1132 pslld xmm8,5
1133 pxor xmm6,xmm12
1134
1135 movdqa xmm9,xmm11
1136 movdqa XMMWORD[(32-128)+rax],xmm4
1137 paddd xmm10,xmm4
1138 pxor xmm0,xmm2
1139 psrld xmm9,27
1140 pxor xmm6,xmm13
1141 movdqa xmm7,xmm12
1142
1143 pslld xmm7,30
1144 movdqa xmm5,xmm0
1145 por xmm8,xmm9
1146 psrld xmm5,31
1147 paddd xmm10,xmm6
1148 paddd xmm0,xmm0
1149
1150 psrld xmm12,2
1151 paddd xmm10,xmm8
1152 por xmm0,xmm5
1153 por xmm12,xmm7
1154 pxor xmm1,xmm3
1155 movdqa xmm3,XMMWORD[((96-128))+rax]
1156
1157 movdqa xmm8,xmm10
1158 movdqa xmm6,xmm13
1159 pxor xmm1,XMMWORD[((192-128))+rax]
1160 paddd xmm14,xmm15
1161 pslld xmm8,5
1162 pxor xmm6,xmm11
1163
1164 movdqa xmm9,xmm10
1165 movdqa XMMWORD[(48-128)+rax],xmm0
1166 paddd xmm14,xmm0
1167 pxor xmm1,xmm3
1168 psrld xmm9,27
1169 pxor xmm6,xmm12
1170 movdqa xmm7,xmm11
1171
1172 pslld xmm7,30
1173 movdqa xmm5,xmm1
1174 por xmm8,xmm9
1175 psrld xmm5,31
1176 paddd xmm14,xmm6
1177 paddd xmm1,xmm1
1178
1179 psrld xmm11,2
1180 paddd xmm14,xmm8
1181 por xmm1,xmm5
1182 por xmm11,xmm7
1183 pxor xmm2,xmm4
1184 movdqa xmm4,XMMWORD[((112-128))+rax]
1185
1186 movdqa xmm8,xmm14
1187 movdqa xmm6,xmm12
1188 pxor xmm2,XMMWORD[((208-128))+rax]
1189 paddd xmm13,xmm15
1190 pslld xmm8,5
1191 pxor xmm6,xmm10
1192
1193 movdqa xmm9,xmm14
1194 movdqa XMMWORD[(64-128)+rax],xmm1
1195 paddd xmm13,xmm1
1196 pxor xmm2,xmm4
1197 psrld xmm9,27
1198 pxor xmm6,xmm11
1199 movdqa xmm7,xmm10
1200
1201 pslld xmm7,30
1202 movdqa xmm5,xmm2
1203 por xmm8,xmm9
1204 psrld xmm5,31
1205 paddd xmm13,xmm6
1206 paddd xmm2,xmm2
1207
1208 psrld xmm10,2
1209 paddd xmm13,xmm8
1210 por xmm2,xmm5
1211 por xmm10,xmm7
1212 pxor xmm3,xmm0
1213 movdqa xmm0,XMMWORD[((128-128))+rax]
1214
1215 movdqa xmm8,xmm13
1216 movdqa xmm6,xmm11
1217 pxor xmm3,XMMWORD[((224-128))+rax]
1218 paddd xmm12,xmm15
1219 pslld xmm8,5
1220 pxor xmm6,xmm14
1221
1222 movdqa xmm9,xmm13
1223 movdqa XMMWORD[(80-128)+rax],xmm2
1224 paddd xmm12,xmm2
1225 pxor xmm3,xmm0
1226 psrld xmm9,27
1227 pxor xmm6,xmm10
1228 movdqa xmm7,xmm14
1229
1230 pslld xmm7,30
1231 movdqa xmm5,xmm3
1232 por xmm8,xmm9
1233 psrld xmm5,31
1234 paddd xmm12,xmm6
1235 paddd xmm3,xmm3
1236
1237 psrld xmm14,2
1238 paddd xmm12,xmm8
1239 por xmm3,xmm5
1240 por xmm14,xmm7
1241 pxor xmm4,xmm1
1242 movdqa xmm1,XMMWORD[((144-128))+rax]
1243
1244 movdqa xmm8,xmm12
1245 movdqa xmm6,xmm10
1246 pxor xmm4,XMMWORD[((240-128))+rax]
1247 paddd xmm11,xmm15
1248 pslld xmm8,5
1249 pxor xmm6,xmm13
1250
1251 movdqa xmm9,xmm12
1252 movdqa XMMWORD[(96-128)+rax],xmm3
1253 paddd xmm11,xmm3
1254 pxor xmm4,xmm1
1255 psrld xmm9,27
1256 pxor xmm6,xmm14
1257 movdqa xmm7,xmm13
1258
1259 pslld xmm7,30
1260 movdqa xmm5,xmm4
1261 por xmm8,xmm9
1262 psrld xmm5,31
1263 paddd xmm11,xmm6
1264 paddd xmm4,xmm4
1265
1266 psrld xmm13,2
1267 paddd xmm11,xmm8
1268 por xmm4,xmm5
1269 por xmm13,xmm7
1270 pxor xmm0,xmm2
1271 movdqa xmm2,XMMWORD[((160-128))+rax]
1272
1273 movdqa xmm8,xmm11
1274 movdqa xmm6,xmm14
1275 pxor xmm0,XMMWORD[((0-128))+rax]
1276 paddd xmm10,xmm15
1277 pslld xmm8,5
1278 pxor xmm6,xmm12
1279
1280 movdqa xmm9,xmm11
1281 movdqa XMMWORD[(112-128)+rax],xmm4
1282 paddd xmm10,xmm4
1283 pxor xmm0,xmm2
1284 psrld xmm9,27
1285 pxor xmm6,xmm13
1286 movdqa xmm7,xmm12
1287
1288 pslld xmm7,30
1289 movdqa xmm5,xmm0
1290 por xmm8,xmm9
1291 psrld xmm5,31
1292 paddd xmm10,xmm6
1293 paddd xmm0,xmm0
1294
1295 psrld xmm12,2
1296 paddd xmm10,xmm8
1297 por xmm0,xmm5
1298 por xmm12,xmm7
1299 movdqa xmm15,XMMWORD[32+rbp]
1300 pxor xmm1,xmm3
1301 movdqa xmm3,XMMWORD[((176-128))+rax]
1302
1303 movdqa xmm8,xmm10
1304 movdqa xmm7,xmm13
1305 pxor xmm1,XMMWORD[((16-128))+rax]
1306 pxor xmm1,xmm3
1307 paddd xmm14,xmm15
1308 pslld xmm8,5
1309 movdqa xmm9,xmm10
1310 pand xmm7,xmm12
1311
1312 movdqa xmm6,xmm13
1313 movdqa xmm5,xmm1
1314 psrld xmm9,27
1315 paddd xmm14,xmm7
1316 pxor xmm6,xmm12
1317
1318 movdqa XMMWORD[(128-128)+rax],xmm0
1319 paddd xmm14,xmm0
1320 por xmm8,xmm9
1321 psrld xmm5,31
1322 pand xmm6,xmm11
1323 movdqa xmm7,xmm11
1324
1325 pslld xmm7,30
1326 paddd xmm1,xmm1
1327 paddd xmm14,xmm6
1328
1329 psrld xmm11,2
1330 paddd xmm14,xmm8
1331 por xmm1,xmm5
1332 por xmm11,xmm7
1333 pxor xmm2,xmm4
1334 movdqa xmm4,XMMWORD[((192-128))+rax]
1335
1336 movdqa xmm8,xmm14
1337 movdqa xmm7,xmm12
1338 pxor xmm2,XMMWORD[((32-128))+rax]
1339 pxor xmm2,xmm4
1340 paddd xmm13,xmm15
1341 pslld xmm8,5
1342 movdqa xmm9,xmm14
1343 pand xmm7,xmm11
1344
1345 movdqa xmm6,xmm12
1346 movdqa xmm5,xmm2
1347 psrld xmm9,27
1348 paddd xmm13,xmm7
1349 pxor xmm6,xmm11
1350
1351 movdqa XMMWORD[(144-128)+rax],xmm1
1352 paddd xmm13,xmm1
1353 por xmm8,xmm9
1354 psrld xmm5,31
1355 pand xmm6,xmm10
1356 movdqa xmm7,xmm10
1357
1358 pslld xmm7,30
1359 paddd xmm2,xmm2
1360 paddd xmm13,xmm6
1361
1362 psrld xmm10,2
1363 paddd xmm13,xmm8
1364 por xmm2,xmm5
1365 por xmm10,xmm7
1366 pxor xmm3,xmm0
1367 movdqa xmm0,XMMWORD[((208-128))+rax]
1368
1369 movdqa xmm8,xmm13
1370 movdqa xmm7,xmm11
1371 pxor xmm3,XMMWORD[((48-128))+rax]
1372 pxor xmm3,xmm0
1373 paddd xmm12,xmm15
1374 pslld xmm8,5
1375 movdqa xmm9,xmm13
1376 pand xmm7,xmm10
1377
1378 movdqa xmm6,xmm11
1379 movdqa xmm5,xmm3
1380 psrld xmm9,27
1381 paddd xmm12,xmm7
1382 pxor xmm6,xmm10
1383
1384 movdqa XMMWORD[(160-128)+rax],xmm2
1385 paddd xmm12,xmm2
1386 por xmm8,xmm9
1387 psrld xmm5,31
1388 pand xmm6,xmm14
1389 movdqa xmm7,xmm14
1390
1391 pslld xmm7,30
1392 paddd xmm3,xmm3
1393 paddd xmm12,xmm6
1394
1395 psrld xmm14,2
1396 paddd xmm12,xmm8
1397 por xmm3,xmm5
1398 por xmm14,xmm7
1399 pxor xmm4,xmm1
1400 movdqa xmm1,XMMWORD[((224-128))+rax]
1401
1402 movdqa xmm8,xmm12
1403 movdqa xmm7,xmm10
1404 pxor xmm4,XMMWORD[((64-128))+rax]
1405 pxor xmm4,xmm1
1406 paddd xmm11,xmm15
1407 pslld xmm8,5
1408 movdqa xmm9,xmm12
1409 pand xmm7,xmm14
1410
1411 movdqa xmm6,xmm10
1412 movdqa xmm5,xmm4
1413 psrld xmm9,27
1414 paddd xmm11,xmm7
1415 pxor xmm6,xmm14
1416
1417 movdqa XMMWORD[(176-128)+rax],xmm3
1418 paddd xmm11,xmm3
1419 por xmm8,xmm9
1420 psrld xmm5,31
1421 pand xmm6,xmm13
1422 movdqa xmm7,xmm13
1423
1424 pslld xmm7,30
1425 paddd xmm4,xmm4
1426 paddd xmm11,xmm6
1427
1428 psrld xmm13,2
1429 paddd xmm11,xmm8
1430 por xmm4,xmm5
1431 por xmm13,xmm7
1432 pxor xmm0,xmm2
1433 movdqa xmm2,XMMWORD[((240-128))+rax]
1434
1435 movdqa xmm8,xmm11
1436 movdqa xmm7,xmm14
1437 pxor xmm0,XMMWORD[((80-128))+rax]
1438 pxor xmm0,xmm2
1439 paddd xmm10,xmm15
1440 pslld xmm8,5
1441 movdqa xmm9,xmm11
1442 pand xmm7,xmm13
1443
1444 movdqa xmm6,xmm14
1445 movdqa xmm5,xmm0
1446 psrld xmm9,27
1447 paddd xmm10,xmm7
1448 pxor xmm6,xmm13
1449
1450 movdqa XMMWORD[(192-128)+rax],xmm4
1451 paddd xmm10,xmm4
1452 por xmm8,xmm9
1453 psrld xmm5,31
1454 pand xmm6,xmm12
1455 movdqa xmm7,xmm12
1456
1457 pslld xmm7,30
1458 paddd xmm0,xmm0
1459 paddd xmm10,xmm6
1460
1461 psrld xmm12,2
1462 paddd xmm10,xmm8
1463 por xmm0,xmm5
1464 por xmm12,xmm7
1465 pxor xmm1,xmm3
1466 movdqa xmm3,XMMWORD[((0-128))+rax]
1467
1468 movdqa xmm8,xmm10
1469 movdqa xmm7,xmm13
1470 pxor xmm1,XMMWORD[((96-128))+rax]
1471 pxor xmm1,xmm3
1472 paddd xmm14,xmm15
1473 pslld xmm8,5
1474 movdqa xmm9,xmm10
1475 pand xmm7,xmm12
1476
1477 movdqa xmm6,xmm13
1478 movdqa xmm5,xmm1
1479 psrld xmm9,27
1480 paddd xmm14,xmm7
1481 pxor xmm6,xmm12
1482
1483 movdqa XMMWORD[(208-128)+rax],xmm0
1484 paddd xmm14,xmm0
1485 por xmm8,xmm9
1486 psrld xmm5,31
1487 pand xmm6,xmm11
1488 movdqa xmm7,xmm11
1489
1490 pslld xmm7,30
1491 paddd xmm1,xmm1
1492 paddd xmm14,xmm6
1493
1494 psrld xmm11,2
1495 paddd xmm14,xmm8
1496 por xmm1,xmm5
1497 por xmm11,xmm7
1498 pxor xmm2,xmm4
1499 movdqa xmm4,XMMWORD[((16-128))+rax]
1500
1501 movdqa xmm8,xmm14
1502 movdqa xmm7,xmm12
1503 pxor xmm2,XMMWORD[((112-128))+rax]
1504 pxor xmm2,xmm4
1505 paddd xmm13,xmm15
1506 pslld xmm8,5
1507 movdqa xmm9,xmm14
1508 pand xmm7,xmm11
1509
1510 movdqa xmm6,xmm12
1511 movdqa xmm5,xmm2
1512 psrld xmm9,27
1513 paddd xmm13,xmm7
1514 pxor xmm6,xmm11
1515
1516 movdqa XMMWORD[(224-128)+rax],xmm1
1517 paddd xmm13,xmm1
1518 por xmm8,xmm9
1519 psrld xmm5,31
1520 pand xmm6,xmm10
1521 movdqa xmm7,xmm10
1522
1523 pslld xmm7,30
1524 paddd xmm2,xmm2
1525 paddd xmm13,xmm6
1526
1527 psrld xmm10,2
1528 paddd xmm13,xmm8
1529 por xmm2,xmm5
1530 por xmm10,xmm7
1531 pxor xmm3,xmm0
1532 movdqa xmm0,XMMWORD[((32-128))+rax]
1533
1534 movdqa xmm8,xmm13
1535 movdqa xmm7,xmm11
1536 pxor xmm3,XMMWORD[((128-128))+rax]
1537 pxor xmm3,xmm0
1538 paddd xmm12,xmm15
1539 pslld xmm8,5
1540 movdqa xmm9,xmm13
1541 pand xmm7,xmm10
1542
1543 movdqa xmm6,xmm11
1544 movdqa xmm5,xmm3
1545 psrld xmm9,27
1546 paddd xmm12,xmm7
1547 pxor xmm6,xmm10
1548
1549 movdqa XMMWORD[(240-128)+rax],xmm2
1550 paddd xmm12,xmm2
1551 por xmm8,xmm9
1552 psrld xmm5,31
1553 pand xmm6,xmm14
1554 movdqa xmm7,xmm14
1555
1556 pslld xmm7,30
1557 paddd xmm3,xmm3
1558 paddd xmm12,xmm6
1559
1560 psrld xmm14,2
1561 paddd xmm12,xmm8
1562 por xmm3,xmm5
1563 por xmm14,xmm7
1564 pxor xmm4,xmm1
1565 movdqa xmm1,XMMWORD[((48-128))+rax]
1566
1567 movdqa xmm8,xmm12
1568 movdqa xmm7,xmm10
1569 pxor xmm4,XMMWORD[((144-128))+rax]
1570 pxor xmm4,xmm1
1571 paddd xmm11,xmm15
1572 pslld xmm8,5
1573 movdqa xmm9,xmm12
1574 pand xmm7,xmm14
1575
1576 movdqa xmm6,xmm10
1577 movdqa xmm5,xmm4
1578 psrld xmm9,27
1579 paddd xmm11,xmm7
1580 pxor xmm6,xmm14
1581
1582 movdqa XMMWORD[(0-128)+rax],xmm3
1583 paddd xmm11,xmm3
1584 por xmm8,xmm9
1585 psrld xmm5,31
1586 pand xmm6,xmm13
1587 movdqa xmm7,xmm13
1588
1589 pslld xmm7,30
1590 paddd xmm4,xmm4
1591 paddd xmm11,xmm6
1592
1593 psrld xmm13,2
1594 paddd xmm11,xmm8
1595 por xmm4,xmm5
1596 por xmm13,xmm7
1597 pxor xmm0,xmm2
1598 movdqa xmm2,XMMWORD[((64-128))+rax]
1599
1600 movdqa xmm8,xmm11
1601 movdqa xmm7,xmm14
1602 pxor xmm0,XMMWORD[((160-128))+rax]
1603 pxor xmm0,xmm2
1604 paddd xmm10,xmm15
1605 pslld xmm8,5
1606 movdqa xmm9,xmm11
1607 pand xmm7,xmm13
1608
1609 movdqa xmm6,xmm14
1610 movdqa xmm5,xmm0
1611 psrld xmm9,27
1612 paddd xmm10,xmm7
1613 pxor xmm6,xmm13
1614
1615 movdqa XMMWORD[(16-128)+rax],xmm4
1616 paddd xmm10,xmm4
1617 por xmm8,xmm9
1618 psrld xmm5,31
1619 pand xmm6,xmm12
1620 movdqa xmm7,xmm12
1621
1622 pslld xmm7,30
1623 paddd xmm0,xmm0
1624 paddd xmm10,xmm6
1625
1626 psrld xmm12,2
1627 paddd xmm10,xmm8
1628 por xmm0,xmm5
1629 por xmm12,xmm7
1630 pxor xmm1,xmm3
1631 movdqa xmm3,XMMWORD[((80-128))+rax]
1632
1633 movdqa xmm8,xmm10
1634 movdqa xmm7,xmm13
1635 pxor xmm1,XMMWORD[((176-128))+rax]
1636 pxor xmm1,xmm3
1637 paddd xmm14,xmm15
1638 pslld xmm8,5
1639 movdqa xmm9,xmm10
1640 pand xmm7,xmm12
1641
1642 movdqa xmm6,xmm13
1643 movdqa xmm5,xmm1
1644 psrld xmm9,27
1645 paddd xmm14,xmm7
1646 pxor xmm6,xmm12
1647
1648 movdqa XMMWORD[(32-128)+rax],xmm0
1649 paddd xmm14,xmm0
1650 por xmm8,xmm9
1651 psrld xmm5,31
1652 pand xmm6,xmm11
1653 movdqa xmm7,xmm11
1654
1655 pslld xmm7,30
1656 paddd xmm1,xmm1
1657 paddd xmm14,xmm6
1658
1659 psrld xmm11,2
1660 paddd xmm14,xmm8
1661 por xmm1,xmm5
1662 por xmm11,xmm7
1663 pxor xmm2,xmm4
1664 movdqa xmm4,XMMWORD[((96-128))+rax]
1665
1666 movdqa xmm8,xmm14
1667 movdqa xmm7,xmm12
1668 pxor xmm2,XMMWORD[((192-128))+rax]
1669 pxor xmm2,xmm4
1670 paddd xmm13,xmm15
1671 pslld xmm8,5
1672 movdqa xmm9,xmm14
1673 pand xmm7,xmm11
1674
1675 movdqa xmm6,xmm12
1676 movdqa xmm5,xmm2
1677 psrld xmm9,27
1678 paddd xmm13,xmm7
1679 pxor xmm6,xmm11
1680
1681 movdqa XMMWORD[(48-128)+rax],xmm1
1682 paddd xmm13,xmm1
1683 por xmm8,xmm9
1684 psrld xmm5,31
1685 pand xmm6,xmm10
1686 movdqa xmm7,xmm10
1687
1688 pslld xmm7,30
1689 paddd xmm2,xmm2
1690 paddd xmm13,xmm6
1691
1692 psrld xmm10,2
1693 paddd xmm13,xmm8
1694 por xmm2,xmm5
1695 por xmm10,xmm7
1696 pxor xmm3,xmm0
1697 movdqa xmm0,XMMWORD[((112-128))+rax]
1698
1699 movdqa xmm8,xmm13
1700 movdqa xmm7,xmm11
1701 pxor xmm3,XMMWORD[((208-128))+rax]
1702 pxor xmm3,xmm0
1703 paddd xmm12,xmm15
1704 pslld xmm8,5
1705 movdqa xmm9,xmm13
1706 pand xmm7,xmm10
1707
1708 movdqa xmm6,xmm11
1709 movdqa xmm5,xmm3
1710 psrld xmm9,27
1711 paddd xmm12,xmm7
1712 pxor xmm6,xmm10
1713
1714 movdqa XMMWORD[(64-128)+rax],xmm2
1715 paddd xmm12,xmm2
1716 por xmm8,xmm9
1717 psrld xmm5,31
1718 pand xmm6,xmm14
1719 movdqa xmm7,xmm14
1720
1721 pslld xmm7,30
1722 paddd xmm3,xmm3
1723 paddd xmm12,xmm6
1724
1725 psrld xmm14,2
1726 paddd xmm12,xmm8
1727 por xmm3,xmm5
1728 por xmm14,xmm7
1729 pxor xmm4,xmm1
1730 movdqa xmm1,XMMWORD[((128-128))+rax]
1731
1732 movdqa xmm8,xmm12
1733 movdqa xmm7,xmm10
1734 pxor xmm4,XMMWORD[((224-128))+rax]
1735 pxor xmm4,xmm1
1736 paddd xmm11,xmm15
1737 pslld xmm8,5
1738 movdqa xmm9,xmm12
1739 pand xmm7,xmm14
1740
1741 movdqa xmm6,xmm10
1742 movdqa xmm5,xmm4
1743 psrld xmm9,27
1744 paddd xmm11,xmm7
1745 pxor xmm6,xmm14
1746
1747 movdqa XMMWORD[(80-128)+rax],xmm3
1748 paddd xmm11,xmm3
1749 por xmm8,xmm9
1750 psrld xmm5,31
1751 pand xmm6,xmm13
1752 movdqa xmm7,xmm13
1753
1754 pslld xmm7,30
1755 paddd xmm4,xmm4
1756 paddd xmm11,xmm6
1757
1758 psrld xmm13,2
1759 paddd xmm11,xmm8
1760 por xmm4,xmm5
1761 por xmm13,xmm7
1762 pxor xmm0,xmm2
1763 movdqa xmm2,XMMWORD[((144-128))+rax]
1764
1765 movdqa xmm8,xmm11
1766 movdqa xmm7,xmm14
1767 pxor xmm0,XMMWORD[((240-128))+rax]
1768 pxor xmm0,xmm2
1769 paddd xmm10,xmm15
1770 pslld xmm8,5
1771 movdqa xmm9,xmm11
1772 pand xmm7,xmm13
1773
1774 movdqa xmm6,xmm14
1775 movdqa xmm5,xmm0
1776 psrld xmm9,27
1777 paddd xmm10,xmm7
1778 pxor xmm6,xmm13
1779
1780 movdqa XMMWORD[(96-128)+rax],xmm4
1781 paddd xmm10,xmm4
1782 por xmm8,xmm9
1783 psrld xmm5,31
1784 pand xmm6,xmm12
1785 movdqa xmm7,xmm12
1786
1787 pslld xmm7,30
1788 paddd xmm0,xmm0
1789 paddd xmm10,xmm6
1790
1791 psrld xmm12,2
1792 paddd xmm10,xmm8
1793 por xmm0,xmm5
1794 por xmm12,xmm7
1795 pxor xmm1,xmm3
1796 movdqa xmm3,XMMWORD[((160-128))+rax]
1797
1798 movdqa xmm8,xmm10
1799 movdqa xmm7,xmm13
1800 pxor xmm1,XMMWORD[((0-128))+rax]
1801 pxor xmm1,xmm3
1802 paddd xmm14,xmm15
1803 pslld xmm8,5
1804 movdqa xmm9,xmm10
1805 pand xmm7,xmm12
1806
1807 movdqa xmm6,xmm13
1808 movdqa xmm5,xmm1
1809 psrld xmm9,27
1810 paddd xmm14,xmm7
1811 pxor xmm6,xmm12
1812
1813 movdqa XMMWORD[(112-128)+rax],xmm0
1814 paddd xmm14,xmm0
1815 por xmm8,xmm9
1816 psrld xmm5,31
1817 pand xmm6,xmm11
1818 movdqa xmm7,xmm11
1819
1820 pslld xmm7,30
1821 paddd xmm1,xmm1
1822 paddd xmm14,xmm6
1823
1824 psrld xmm11,2
1825 paddd xmm14,xmm8
1826 por xmm1,xmm5
1827 por xmm11,xmm7
1828 pxor xmm2,xmm4
1829 movdqa xmm4,XMMWORD[((176-128))+rax]
1830
1831 movdqa xmm8,xmm14
1832 movdqa xmm7,xmm12
1833 pxor xmm2,XMMWORD[((16-128))+rax]
1834 pxor xmm2,xmm4
1835 paddd xmm13,xmm15
1836 pslld xmm8,5
1837 movdqa xmm9,xmm14
1838 pand xmm7,xmm11
1839
1840 movdqa xmm6,xmm12
1841 movdqa xmm5,xmm2
1842 psrld xmm9,27
1843 paddd xmm13,xmm7
1844 pxor xmm6,xmm11
1845
1846 movdqa XMMWORD[(128-128)+rax],xmm1
1847 paddd xmm13,xmm1
1848 por xmm8,xmm9
1849 psrld xmm5,31
1850 pand xmm6,xmm10
1851 movdqa xmm7,xmm10
1852
1853 pslld xmm7,30
1854 paddd xmm2,xmm2
1855 paddd xmm13,xmm6
1856
1857 psrld xmm10,2
1858 paddd xmm13,xmm8
1859 por xmm2,xmm5
1860 por xmm10,xmm7
1861 pxor xmm3,xmm0
1862 movdqa xmm0,XMMWORD[((192-128))+rax]
1863
1864 movdqa xmm8,xmm13
1865 movdqa xmm7,xmm11
1866 pxor xmm3,XMMWORD[((32-128))+rax]
1867 pxor xmm3,xmm0
1868 paddd xmm12,xmm15
1869 pslld xmm8,5
1870 movdqa xmm9,xmm13
1871 pand xmm7,xmm10
1872
1873 movdqa xmm6,xmm11
1874 movdqa xmm5,xmm3
1875 psrld xmm9,27
1876 paddd xmm12,xmm7
1877 pxor xmm6,xmm10
1878
1879 movdqa XMMWORD[(144-128)+rax],xmm2
1880 paddd xmm12,xmm2
1881 por xmm8,xmm9
1882 psrld xmm5,31
1883 pand xmm6,xmm14
1884 movdqa xmm7,xmm14
1885
1886 pslld xmm7,30
1887 paddd xmm3,xmm3
1888 paddd xmm12,xmm6
1889
1890 psrld xmm14,2
1891 paddd xmm12,xmm8
1892 por xmm3,xmm5
1893 por xmm14,xmm7
1894 pxor xmm4,xmm1
1895 movdqa xmm1,XMMWORD[((208-128))+rax]
1896
1897 movdqa xmm8,xmm12
1898 movdqa xmm7,xmm10
1899 pxor xmm4,XMMWORD[((48-128))+rax]
1900 pxor xmm4,xmm1
1901 paddd xmm11,xmm15
1902 pslld xmm8,5
1903 movdqa xmm9,xmm12
1904 pand xmm7,xmm14
1905
1906 movdqa xmm6,xmm10
1907 movdqa xmm5,xmm4
1908 psrld xmm9,27
1909 paddd xmm11,xmm7
1910 pxor xmm6,xmm14
1911
1912 movdqa XMMWORD[(160-128)+rax],xmm3
1913 paddd xmm11,xmm3
1914 por xmm8,xmm9
1915 psrld xmm5,31
1916 pand xmm6,xmm13
1917 movdqa xmm7,xmm13
1918
1919 pslld xmm7,30
1920 paddd xmm4,xmm4
1921 paddd xmm11,xmm6
1922
1923 psrld xmm13,2
1924 paddd xmm11,xmm8
1925 por xmm4,xmm5
1926 por xmm13,xmm7
1927 pxor xmm0,xmm2
1928 movdqa xmm2,XMMWORD[((224-128))+rax]
1929
1930 movdqa xmm8,xmm11
1931 movdqa xmm7,xmm14
1932 pxor xmm0,XMMWORD[((64-128))+rax]
1933 pxor xmm0,xmm2
1934 paddd xmm10,xmm15
1935 pslld xmm8,5
1936 movdqa xmm9,xmm11
1937 pand xmm7,xmm13
1938
1939 movdqa xmm6,xmm14
1940 movdqa xmm5,xmm0
1941 psrld xmm9,27
1942 paddd xmm10,xmm7
1943 pxor xmm6,xmm13
1944
1945 movdqa XMMWORD[(176-128)+rax],xmm4
1946 paddd xmm10,xmm4
1947 por xmm8,xmm9
1948 psrld xmm5,31
1949 pand xmm6,xmm12
1950 movdqa xmm7,xmm12
1951
1952 pslld xmm7,30
1953 paddd xmm0,xmm0
1954 paddd xmm10,xmm6
1955
1956 psrld xmm12,2
1957 paddd xmm10,xmm8
1958 por xmm0,xmm5
1959 por xmm12,xmm7
1960 movdqa xmm15,XMMWORD[64+rbp]
1961 pxor xmm1,xmm3
1962 movdqa xmm3,XMMWORD[((240-128))+rax]
1963
1964 movdqa xmm8,xmm10
1965 movdqa xmm6,xmm13
1966 pxor xmm1,XMMWORD[((80-128))+rax]
1967 paddd xmm14,xmm15
1968 pslld xmm8,5
1969 pxor xmm6,xmm11
1970
1971 movdqa xmm9,xmm10
1972 movdqa XMMWORD[(192-128)+rax],xmm0
1973 paddd xmm14,xmm0
1974 pxor xmm1,xmm3
1975 psrld xmm9,27
1976 pxor xmm6,xmm12
1977 movdqa xmm7,xmm11
1978
1979 pslld xmm7,30
1980 movdqa xmm5,xmm1
1981 por xmm8,xmm9
1982 psrld xmm5,31
1983 paddd xmm14,xmm6
1984 paddd xmm1,xmm1
1985
1986 psrld xmm11,2
1987 paddd xmm14,xmm8
1988 por xmm1,xmm5
1989 por xmm11,xmm7
1990 pxor xmm2,xmm4
1991 movdqa xmm4,XMMWORD[((0-128))+rax]
1992
1993 movdqa xmm8,xmm14
1994 movdqa xmm6,xmm12
1995 pxor xmm2,XMMWORD[((96-128))+rax]
1996 paddd xmm13,xmm15
1997 pslld xmm8,5
1998 pxor xmm6,xmm10
1999
2000 movdqa xmm9,xmm14
2001 movdqa XMMWORD[(208-128)+rax],xmm1
2002 paddd xmm13,xmm1
2003 pxor xmm2,xmm4
2004 psrld xmm9,27
2005 pxor xmm6,xmm11
2006 movdqa xmm7,xmm10
2007
2008 pslld xmm7,30
2009 movdqa xmm5,xmm2
2010 por xmm8,xmm9
2011 psrld xmm5,31
2012 paddd xmm13,xmm6
2013 paddd xmm2,xmm2
2014
2015 psrld xmm10,2
2016 paddd xmm13,xmm8
2017 por xmm2,xmm5
2018 por xmm10,xmm7
2019 pxor xmm3,xmm0
2020 movdqa xmm0,XMMWORD[((16-128))+rax]
2021
2022 movdqa xmm8,xmm13
2023 movdqa xmm6,xmm11
2024 pxor xmm3,XMMWORD[((112-128))+rax]
2025 paddd xmm12,xmm15
2026 pslld xmm8,5
2027 pxor xmm6,xmm14
2028
2029 movdqa xmm9,xmm13
2030 movdqa XMMWORD[(224-128)+rax],xmm2
2031 paddd xmm12,xmm2
2032 pxor xmm3,xmm0
2033 psrld xmm9,27
2034 pxor xmm6,xmm10
2035 movdqa xmm7,xmm14
2036
2037 pslld xmm7,30
2038 movdqa xmm5,xmm3
2039 por xmm8,xmm9
2040 psrld xmm5,31
2041 paddd xmm12,xmm6
2042 paddd xmm3,xmm3
2043
2044 psrld xmm14,2
2045 paddd xmm12,xmm8
2046 por xmm3,xmm5
2047 por xmm14,xmm7
2048 pxor xmm4,xmm1
2049 movdqa xmm1,XMMWORD[((32-128))+rax]
2050
2051 movdqa xmm8,xmm12
2052 movdqa xmm6,xmm10
2053 pxor xmm4,XMMWORD[((128-128))+rax]
2054 paddd xmm11,xmm15
2055 pslld xmm8,5
2056 pxor xmm6,xmm13
2057
2058 movdqa xmm9,xmm12
2059 movdqa XMMWORD[(240-128)+rax],xmm3
2060 paddd xmm11,xmm3
2061 pxor xmm4,xmm1
2062 psrld xmm9,27
2063 pxor xmm6,xmm14
2064 movdqa xmm7,xmm13
2065
2066 pslld xmm7,30
2067 movdqa xmm5,xmm4
2068 por xmm8,xmm9
2069 psrld xmm5,31
2070 paddd xmm11,xmm6
2071 paddd xmm4,xmm4
2072
2073 psrld xmm13,2
2074 paddd xmm11,xmm8
2075 por xmm4,xmm5
2076 por xmm13,xmm7
2077 pxor xmm0,xmm2
2078 movdqa xmm2,XMMWORD[((48-128))+rax]
2079
2080 movdqa xmm8,xmm11
2081 movdqa xmm6,xmm14
2082 pxor xmm0,XMMWORD[((144-128))+rax]
2083 paddd xmm10,xmm15
2084 pslld xmm8,5
2085 pxor xmm6,xmm12
2086
2087 movdqa xmm9,xmm11
2088 movdqa XMMWORD[(0-128)+rax],xmm4
2089 paddd xmm10,xmm4
2090 pxor xmm0,xmm2
2091 psrld xmm9,27
2092 pxor xmm6,xmm13
2093 movdqa xmm7,xmm12
2094
2095 pslld xmm7,30
2096 movdqa xmm5,xmm0
2097 por xmm8,xmm9
2098 psrld xmm5,31
2099 paddd xmm10,xmm6
2100 paddd xmm0,xmm0
2101
2102 psrld xmm12,2
2103 paddd xmm10,xmm8
2104 por xmm0,xmm5
2105 por xmm12,xmm7
2106 pxor xmm1,xmm3
2107 movdqa xmm3,XMMWORD[((64-128))+rax]
2108
2109 movdqa xmm8,xmm10
2110 movdqa xmm6,xmm13
2111 pxor xmm1,XMMWORD[((160-128))+rax]
2112 paddd xmm14,xmm15
2113 pslld xmm8,5
2114 pxor xmm6,xmm11
2115
2116 movdqa xmm9,xmm10
2117 movdqa XMMWORD[(16-128)+rax],xmm0
2118 paddd xmm14,xmm0
2119 pxor xmm1,xmm3
2120 psrld xmm9,27
2121 pxor xmm6,xmm12
2122 movdqa xmm7,xmm11
2123
2124 pslld xmm7,30
2125 movdqa xmm5,xmm1
2126 por xmm8,xmm9
2127 psrld xmm5,31
2128 paddd xmm14,xmm6
2129 paddd xmm1,xmm1
2130
2131 psrld xmm11,2
2132 paddd xmm14,xmm8
2133 por xmm1,xmm5
2134 por xmm11,xmm7
2135 pxor xmm2,xmm4
2136 movdqa xmm4,XMMWORD[((80-128))+rax]
2137
2138 movdqa xmm8,xmm14
2139 movdqa xmm6,xmm12
2140 pxor xmm2,XMMWORD[((176-128))+rax]
2141 paddd xmm13,xmm15
2142 pslld xmm8,5
2143 pxor xmm6,xmm10
2144
2145 movdqa xmm9,xmm14
2146 movdqa XMMWORD[(32-128)+rax],xmm1
2147 paddd xmm13,xmm1
2148 pxor xmm2,xmm4
2149 psrld xmm9,27
2150 pxor xmm6,xmm11
2151 movdqa xmm7,xmm10
2152
2153 pslld xmm7,30
2154 movdqa xmm5,xmm2
2155 por xmm8,xmm9
2156 psrld xmm5,31
2157 paddd xmm13,xmm6
2158 paddd xmm2,xmm2
2159
2160 psrld xmm10,2
2161 paddd xmm13,xmm8
2162 por xmm2,xmm5
2163 por xmm10,xmm7
2164 pxor xmm3,xmm0
2165 movdqa xmm0,XMMWORD[((96-128))+rax]
2166
2167 movdqa xmm8,xmm13
2168 movdqa xmm6,xmm11
2169 pxor xmm3,XMMWORD[((192-128))+rax]
2170 paddd xmm12,xmm15
2171 pslld xmm8,5
2172 pxor xmm6,xmm14
2173
2174 movdqa xmm9,xmm13
2175 movdqa XMMWORD[(48-128)+rax],xmm2
2176 paddd xmm12,xmm2
2177 pxor xmm3,xmm0
2178 psrld xmm9,27
2179 pxor xmm6,xmm10
2180 movdqa xmm7,xmm14
2181
2182 pslld xmm7,30
2183 movdqa xmm5,xmm3
2184 por xmm8,xmm9
2185 psrld xmm5,31
2186 paddd xmm12,xmm6
2187 paddd xmm3,xmm3
2188
2189 psrld xmm14,2
2190 paddd xmm12,xmm8
2191 por xmm3,xmm5
2192 por xmm14,xmm7
2193 pxor xmm4,xmm1
2194 movdqa xmm1,XMMWORD[((112-128))+rax]
2195
2196 movdqa xmm8,xmm12
2197 movdqa xmm6,xmm10
2198 pxor xmm4,XMMWORD[((208-128))+rax]
2199 paddd xmm11,xmm15
2200 pslld xmm8,5
2201 pxor xmm6,xmm13
2202
2203 movdqa xmm9,xmm12
2204 movdqa XMMWORD[(64-128)+rax],xmm3
2205 paddd xmm11,xmm3
2206 pxor xmm4,xmm1
2207 psrld xmm9,27
2208 pxor xmm6,xmm14
2209 movdqa xmm7,xmm13
2210
2211 pslld xmm7,30
2212 movdqa xmm5,xmm4
2213 por xmm8,xmm9
2214 psrld xmm5,31
2215 paddd xmm11,xmm6
2216 paddd xmm4,xmm4
2217
2218 psrld xmm13,2
2219 paddd xmm11,xmm8
2220 por xmm4,xmm5
2221 por xmm13,xmm7
2222 pxor xmm0,xmm2
2223 movdqa xmm2,XMMWORD[((128-128))+rax]
2224
2225 movdqa xmm8,xmm11
2226 movdqa xmm6,xmm14
2227 pxor xmm0,XMMWORD[((224-128))+rax]
2228 paddd xmm10,xmm15
2229 pslld xmm8,5
2230 pxor xmm6,xmm12
2231
2232 movdqa xmm9,xmm11
2233 movdqa XMMWORD[(80-128)+rax],xmm4
2234 paddd xmm10,xmm4
2235 pxor xmm0,xmm2
2236 psrld xmm9,27
2237 pxor xmm6,xmm13
2238 movdqa xmm7,xmm12
2239
2240 pslld xmm7,30
2241 movdqa xmm5,xmm0
2242 por xmm8,xmm9
2243 psrld xmm5,31
2244 paddd xmm10,xmm6
2245 paddd xmm0,xmm0
2246
2247 psrld xmm12,2
2248 paddd xmm10,xmm8
2249 por xmm0,xmm5
2250 por xmm12,xmm7
2251 pxor xmm1,xmm3
2252 movdqa xmm3,XMMWORD[((144-128))+rax]
2253
2254 movdqa xmm8,xmm10
2255 movdqa xmm6,xmm13
2256 pxor xmm1,XMMWORD[((240-128))+rax]
2257 paddd xmm14,xmm15
2258 pslld xmm8,5
2259 pxor xmm6,xmm11
2260
2261 movdqa xmm9,xmm10
2262 movdqa XMMWORD[(96-128)+rax],xmm0
2263 paddd xmm14,xmm0
2264 pxor xmm1,xmm3
2265 psrld xmm9,27
2266 pxor xmm6,xmm12
2267 movdqa xmm7,xmm11
2268
2269 pslld xmm7,30
2270 movdqa xmm5,xmm1
2271 por xmm8,xmm9
2272 psrld xmm5,31
2273 paddd xmm14,xmm6
2274 paddd xmm1,xmm1
2275
2276 psrld xmm11,2
2277 paddd xmm14,xmm8
2278 por xmm1,xmm5
2279 por xmm11,xmm7
2280 pxor xmm2,xmm4
2281 movdqa xmm4,XMMWORD[((160-128))+rax]
2282
2283 movdqa xmm8,xmm14
2284 movdqa xmm6,xmm12
2285 pxor xmm2,XMMWORD[((0-128))+rax]
2286 paddd xmm13,xmm15
2287 pslld xmm8,5
2288 pxor xmm6,xmm10
2289
2290 movdqa xmm9,xmm14
2291 movdqa XMMWORD[(112-128)+rax],xmm1
2292 paddd xmm13,xmm1
2293 pxor xmm2,xmm4
2294 psrld xmm9,27
2295 pxor xmm6,xmm11
2296 movdqa xmm7,xmm10
2297
2298 pslld xmm7,30
2299 movdqa xmm5,xmm2
2300 por xmm8,xmm9
2301 psrld xmm5,31
2302 paddd xmm13,xmm6
2303 paddd xmm2,xmm2
2304
2305 psrld xmm10,2
2306 paddd xmm13,xmm8
2307 por xmm2,xmm5
2308 por xmm10,xmm7
2309 pxor xmm3,xmm0
2310 movdqa xmm0,XMMWORD[((176-128))+rax]
2311
2312 movdqa xmm8,xmm13
2313 movdqa xmm6,xmm11
2314 pxor xmm3,XMMWORD[((16-128))+rax]
2315 paddd xmm12,xmm15
2316 pslld xmm8,5
2317 pxor xmm6,xmm14
2318
2319 movdqa xmm9,xmm13
2320 paddd xmm12,xmm2
2321 pxor xmm3,xmm0
2322 psrld xmm9,27
2323 pxor xmm6,xmm10
2324 movdqa xmm7,xmm14
2325
2326 pslld xmm7,30
2327 movdqa xmm5,xmm3
2328 por xmm8,xmm9
2329 psrld xmm5,31
2330 paddd xmm12,xmm6
2331 paddd xmm3,xmm3
2332
2333 psrld xmm14,2
2334 paddd xmm12,xmm8
2335 por xmm3,xmm5
2336 por xmm14,xmm7
2337 pxor xmm4,xmm1
2338 movdqa xmm1,XMMWORD[((192-128))+rax]
2339
2340 movdqa xmm8,xmm12
2341 movdqa xmm6,xmm10
2342 pxor xmm4,XMMWORD[((32-128))+rax]
2343 paddd xmm11,xmm15
2344 pslld xmm8,5
2345 pxor xmm6,xmm13
2346
2347 movdqa xmm9,xmm12
2348 paddd xmm11,xmm3
2349 pxor xmm4,xmm1
2350 psrld xmm9,27
2351 pxor xmm6,xmm14
2352 movdqa xmm7,xmm13
2353
2354 pslld xmm7,30
2355 movdqa xmm5,xmm4
2356 por xmm8,xmm9
2357 psrld xmm5,31
2358 paddd xmm11,xmm6
2359 paddd xmm4,xmm4
2360
2361 psrld xmm13,2
2362 paddd xmm11,xmm8
2363 por xmm4,xmm5
2364 por xmm13,xmm7
2365 pxor xmm0,xmm2
2366 movdqa xmm2,XMMWORD[((208-128))+rax]
2367
2368 movdqa xmm8,xmm11
2369 movdqa xmm6,xmm14
2370 pxor xmm0,XMMWORD[((48-128))+rax]
2371 paddd xmm10,xmm15
2372 pslld xmm8,5
2373 pxor xmm6,xmm12
2374
2375 movdqa xmm9,xmm11
2376 paddd xmm10,xmm4
2377 pxor xmm0,xmm2
2378 psrld xmm9,27
2379 pxor xmm6,xmm13
2380 movdqa xmm7,xmm12
2381
2382 pslld xmm7,30
2383 movdqa xmm5,xmm0
2384 por xmm8,xmm9
2385 psrld xmm5,31
2386 paddd xmm10,xmm6
2387 paddd xmm0,xmm0
2388
2389 psrld xmm12,2
2390 paddd xmm10,xmm8
2391 por xmm0,xmm5
2392 por xmm12,xmm7
2393 pxor xmm1,xmm3
2394 movdqa xmm3,XMMWORD[((224-128))+rax]
2395
2396 movdqa xmm8,xmm10
2397 movdqa xmm6,xmm13
2398 pxor xmm1,XMMWORD[((64-128))+rax]
2399 paddd xmm14,xmm15
2400 pslld xmm8,5
2401 pxor xmm6,xmm11
2402
2403 movdqa xmm9,xmm10
2404 paddd xmm14,xmm0
2405 pxor xmm1,xmm3
2406 psrld xmm9,27
2407 pxor xmm6,xmm12
2408 movdqa xmm7,xmm11
2409
2410 pslld xmm7,30
2411 movdqa xmm5,xmm1
2412 por xmm8,xmm9
2413 psrld xmm5,31
2414 paddd xmm14,xmm6
2415 paddd xmm1,xmm1
2416
2417 psrld xmm11,2
2418 paddd xmm14,xmm8
2419 por xmm1,xmm5
2420 por xmm11,xmm7
2421 pxor xmm2,xmm4
2422 movdqa xmm4,XMMWORD[((240-128))+rax]
2423
2424 movdqa xmm8,xmm14
2425 movdqa xmm6,xmm12
2426 pxor xmm2,XMMWORD[((80-128))+rax]
2427 paddd xmm13,xmm15
2428 pslld xmm8,5
2429 pxor xmm6,xmm10
2430
2431 movdqa xmm9,xmm14
2432 paddd xmm13,xmm1
2433 pxor xmm2,xmm4
2434 psrld xmm9,27
2435 pxor xmm6,xmm11
2436 movdqa xmm7,xmm10
2437
2438 pslld xmm7,30
2439 movdqa xmm5,xmm2
2440 por xmm8,xmm9
2441 psrld xmm5,31
2442 paddd xmm13,xmm6
2443 paddd xmm2,xmm2
2444
2445 psrld xmm10,2
2446 paddd xmm13,xmm8
2447 por xmm2,xmm5
2448 por xmm10,xmm7
2449 pxor xmm3,xmm0
2450 movdqa xmm0,XMMWORD[((0-128))+rax]
2451
2452 movdqa xmm8,xmm13
2453 movdqa xmm6,xmm11
2454 pxor xmm3,XMMWORD[((96-128))+rax]
2455 paddd xmm12,xmm15
2456 pslld xmm8,5
2457 pxor xmm6,xmm14
2458
2459 movdqa xmm9,xmm13
2460 paddd xmm12,xmm2
2461 pxor xmm3,xmm0
2462 psrld xmm9,27
2463 pxor xmm6,xmm10
2464 movdqa xmm7,xmm14
2465
2466 pslld xmm7,30
2467 movdqa xmm5,xmm3
2468 por xmm8,xmm9
2469 psrld xmm5,31
2470 paddd xmm12,xmm6
2471 paddd xmm3,xmm3
2472
2473 psrld xmm14,2
2474 paddd xmm12,xmm8
2475 por xmm3,xmm5
2476 por xmm14,xmm7
2477 pxor xmm4,xmm1
2478 movdqa xmm1,XMMWORD[((16-128))+rax]
2479
2480 movdqa xmm8,xmm12
2481 movdqa xmm6,xmm10
2482 pxor xmm4,XMMWORD[((112-128))+rax]
2483 paddd xmm11,xmm15
2484 pslld xmm8,5
2485 pxor xmm6,xmm13
2486
2487 movdqa xmm9,xmm12
2488 paddd xmm11,xmm3
2489 pxor xmm4,xmm1
2490 psrld xmm9,27
2491 pxor xmm6,xmm14
2492 movdqa xmm7,xmm13
2493
2494 pslld xmm7,30
2495 movdqa xmm5,xmm4
2496 por xmm8,xmm9
2497 psrld xmm5,31
2498 paddd xmm11,xmm6
2499 paddd xmm4,xmm4
2500
2501 psrld xmm13,2
2502 paddd xmm11,xmm8
2503 por xmm4,xmm5
2504 por xmm13,xmm7
2505 movdqa xmm8,xmm11
2506 paddd xmm10,xmm15
2507 movdqa xmm6,xmm14
2508 pslld xmm8,5
2509 pxor xmm6,xmm12
2510
2511 movdqa xmm9,xmm11
2512 paddd xmm10,xmm4
2513 psrld xmm9,27
2514 movdqa xmm7,xmm12
2515 pxor xmm6,xmm13
2516
2517 pslld xmm7,30
2518 por xmm8,xmm9
2519 paddd xmm10,xmm6
2520
2521 psrld xmm12,2
2522 paddd xmm10,xmm8
2523 por xmm12,xmm7
2524 movdqa xmm0,XMMWORD[rbx]
2525 mov ecx,1
2526 cmp ecx,DWORD[rbx]
2527 pxor xmm8,xmm8
2528 cmovge r8,rbp
2529 cmp ecx,DWORD[4+rbx]
2530 movdqa xmm1,xmm0
2531 cmovge r9,rbp
2532 cmp ecx,DWORD[8+rbx]
2533 pcmpgtd xmm1,xmm8
2534 cmovge r10,rbp
2535 cmp ecx,DWORD[12+rbx]
2536 paddd xmm0,xmm1
2537 cmovge r11,rbp
2538
2539 movdqu xmm6,XMMWORD[rdi]
2540 pand xmm10,xmm1
2541 movdqu xmm7,XMMWORD[32+rdi]
2542 pand xmm11,xmm1
2543 paddd xmm10,xmm6
2544 movdqu xmm8,XMMWORD[64+rdi]
2545 pand xmm12,xmm1
2546 paddd xmm11,xmm7
2547 movdqu xmm9,XMMWORD[96+rdi]
2548 pand xmm13,xmm1
2549 paddd xmm12,xmm8
2550 movdqu xmm5,XMMWORD[128+rdi]
2551 pand xmm14,xmm1
2552 movdqu XMMWORD[rdi],xmm10
2553 paddd xmm13,xmm9
2554 movdqu XMMWORD[32+rdi],xmm11
2555 paddd xmm14,xmm5
2556 movdqu XMMWORD[64+rdi],xmm12
2557 movdqu XMMWORD[96+rdi],xmm13
2558 movdqu XMMWORD[128+rdi],xmm14
2559
2560 movdqa XMMWORD[rbx],xmm0
2561 movdqa xmm5,XMMWORD[96+rbp]
2562 movdqa xmm15,XMMWORD[((-32))+rbp]
2563 dec edx
2564 jnz NEAR $L$oop
2565
2566 mov edx,DWORD[280+rsp]
2567 lea rdi,[16+rdi]
2568 lea rsi,[64+rsi]
2569 dec edx
2570 jnz NEAR $L$oop_grande
2571
2572$L$done:
2573 mov rax,QWORD[272+rsp]
2574 movaps xmm6,XMMWORD[((-184))+rax]
2575 movaps xmm7,XMMWORD[((-168))+rax]
2576 movaps xmm8,XMMWORD[((-152))+rax]
2577 movaps xmm9,XMMWORD[((-136))+rax]
2578 movaps xmm10,XMMWORD[((-120))+rax]
2579 movaps xmm11,XMMWORD[((-104))+rax]
2580 movaps xmm12,XMMWORD[((-88))+rax]
2581 movaps xmm13,XMMWORD[((-72))+rax]
2582 movaps xmm14,XMMWORD[((-56))+rax]
2583 movaps xmm15,XMMWORD[((-40))+rax]
2584 mov rbp,QWORD[((-16))+rax]
2585 mov rbx,QWORD[((-8))+rax]
2586 lea rsp,[rax]
2587$L$epilogue:
2588 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2589 mov rsi,QWORD[16+rsp]
2590 DB 0F3h,0C3h ;repret
2591$L$SEH_end_sha1_multi_block:
2592
2593ALIGN 32
2594sha1_multi_block_shaext:
2595 mov QWORD[8+rsp],rdi ;WIN64 prologue
2596 mov QWORD[16+rsp],rsi
2597 mov rax,rsp
2598$L$SEH_begin_sha1_multi_block_shaext:
2599 mov rdi,rcx
2600 mov rsi,rdx
2601 mov rdx,r8
2602
2603
2604_shaext_shortcut:
2605 mov rax,rsp
2606 push rbx
2607 push rbp
2608 lea rsp,[((-168))+rsp]
2609 movaps XMMWORD[rsp],xmm6
2610 movaps XMMWORD[16+rsp],xmm7
2611 movaps XMMWORD[32+rsp],xmm8
2612 movaps XMMWORD[48+rsp],xmm9
2613 movaps XMMWORD[(-120)+rax],xmm10
2614 movaps XMMWORD[(-104)+rax],xmm11
2615 movaps XMMWORD[(-88)+rax],xmm12
2616 movaps XMMWORD[(-72)+rax],xmm13
2617 movaps XMMWORD[(-56)+rax],xmm14
2618 movaps XMMWORD[(-40)+rax],xmm15
2619 sub rsp,288
2620 shl edx,1
2621 and rsp,-256
2622 lea rdi,[64+rdi]
2623 mov QWORD[272+rsp],rax
2624$L$body_shaext:
2625 lea rbx,[256+rsp]
2626 movdqa xmm3,XMMWORD[((K_XX_XX+128))]
2627
2628$L$oop_grande_shaext:
2629 mov DWORD[280+rsp],edx
2630 xor edx,edx
2631 mov r8,QWORD[rsi]
2632 mov ecx,DWORD[8+rsi]
2633 cmp ecx,edx
2634 cmovg edx,ecx
2635 test ecx,ecx
2636 mov DWORD[rbx],ecx
2637 cmovle r8,rsp
2638 mov r9,QWORD[16+rsi]
2639 mov ecx,DWORD[24+rsi]
2640 cmp ecx,edx
2641 cmovg edx,ecx
2642 test ecx,ecx
2643 mov DWORD[4+rbx],ecx
2644 cmovle r9,rsp
2645 test edx,edx
2646 jz NEAR $L$done_shaext
2647
2648 movq xmm0,QWORD[((0-64))+rdi]
2649 movq xmm4,QWORD[((32-64))+rdi]
2650 movq xmm5,QWORD[((64-64))+rdi]
2651 movq xmm6,QWORD[((96-64))+rdi]
2652 movq xmm7,QWORD[((128-64))+rdi]
2653
2654 punpckldq xmm0,xmm4
2655 punpckldq xmm5,xmm6
2656
2657 movdqa xmm8,xmm0
2658 punpcklqdq xmm0,xmm5
2659 punpckhqdq xmm8,xmm5
2660
2661 pshufd xmm1,xmm7,63
2662 pshufd xmm9,xmm7,127
2663 pshufd xmm0,xmm0,27
2664 pshufd xmm8,xmm8,27
2665 jmp NEAR $L$oop_shaext
2666
2667ALIGN 32
2668$L$oop_shaext:
2669 movdqu xmm4,XMMWORD[r8]
2670 movdqu xmm11,XMMWORD[r9]
2671 movdqu xmm5,XMMWORD[16+r8]
2672 movdqu xmm12,XMMWORD[16+r9]
2673 movdqu xmm6,XMMWORD[32+r8]
2674DB 102,15,56,0,227
2675 movdqu xmm13,XMMWORD[32+r9]
2676DB 102,68,15,56,0,219
2677 movdqu xmm7,XMMWORD[48+r8]
2678 lea r8,[64+r8]
2679DB 102,15,56,0,235
2680 movdqu xmm14,XMMWORD[48+r9]
2681 lea r9,[64+r9]
2682DB 102,68,15,56,0,227
2683
2684 movdqa XMMWORD[80+rsp],xmm1
2685 paddd xmm1,xmm4
2686 movdqa XMMWORD[112+rsp],xmm9
2687 paddd xmm9,xmm11
2688 movdqa XMMWORD[64+rsp],xmm0
2689 movdqa xmm2,xmm0
2690 movdqa XMMWORD[96+rsp],xmm8
2691 movdqa xmm10,xmm8
2692DB 15,58,204,193,0
2693DB 15,56,200,213
2694DB 69,15,58,204,193,0
2695DB 69,15,56,200,212
2696DB 102,15,56,0,243
2697 prefetcht0 [127+r8]
2698DB 15,56,201,229
2699DB 102,68,15,56,0,235
2700 prefetcht0 [127+r9]
2701DB 69,15,56,201,220
2702
2703DB 102,15,56,0,251
2704 movdqa xmm1,xmm0
2705DB 102,68,15,56,0,243
2706 movdqa xmm9,xmm8
2707DB 15,58,204,194,0
2708DB 15,56,200,206
2709DB 69,15,58,204,194,0
2710DB 69,15,56,200,205
2711 pxor xmm4,xmm6
2712DB 15,56,201,238
2713 pxor xmm11,xmm13
2714DB 69,15,56,201,229
2715 movdqa xmm2,xmm0
2716 movdqa xmm10,xmm8
2717DB 15,58,204,193,0
2718DB 15,56,200,215
2719DB 69,15,58,204,193,0
2720DB 69,15,56,200,214
2721DB 15,56,202,231
2722DB 69,15,56,202,222
2723 pxor xmm5,xmm7
2724DB 15,56,201,247
2725 pxor xmm12,xmm14
2726DB 69,15,56,201,238
2727 movdqa xmm1,xmm0
2728 movdqa xmm9,xmm8
2729DB 15,58,204,194,0
2730DB 15,56,200,204
2731DB 69,15,58,204,194,0
2732DB 69,15,56,200,203
2733DB 15,56,202,236
2734DB 69,15,56,202,227
2735 pxor xmm6,xmm4
2736DB 15,56,201,252
2737 pxor xmm13,xmm11
2738DB 69,15,56,201,243
2739 movdqa xmm2,xmm0
2740 movdqa xmm10,xmm8
2741DB 15,58,204,193,0
2742DB 15,56,200,213
2743DB 69,15,58,204,193,0
2744DB 69,15,56,200,212
2745DB 15,56,202,245
2746DB 69,15,56,202,236
2747 pxor xmm7,xmm5
2748DB 15,56,201,229
2749 pxor xmm14,xmm12
2750DB 69,15,56,201,220
2751 movdqa xmm1,xmm0
2752 movdqa xmm9,xmm8
2753DB 15,58,204,194,1
2754DB 15,56,200,206
2755DB 69,15,58,204,194,1
2756DB 69,15,56,200,205
2757DB 15,56,202,254
2758DB 69,15,56,202,245
2759 pxor xmm4,xmm6
2760DB 15,56,201,238
2761 pxor xmm11,xmm13
2762DB 69,15,56,201,229
2763 movdqa xmm2,xmm0
2764 movdqa xmm10,xmm8
2765DB 15,58,204,193,1
2766DB 15,56,200,215
2767DB 69,15,58,204,193,1
2768DB 69,15,56,200,214
2769DB 15,56,202,231
2770DB 69,15,56,202,222
2771 pxor xmm5,xmm7
2772DB 15,56,201,247
2773 pxor xmm12,xmm14
2774DB 69,15,56,201,238
2775 movdqa xmm1,xmm0
2776 movdqa xmm9,xmm8
2777DB 15,58,204,194,1
2778DB 15,56,200,204
2779DB 69,15,58,204,194,1
2780DB 69,15,56,200,203
2781DB 15,56,202,236
2782DB 69,15,56,202,227
2783 pxor xmm6,xmm4
2784DB 15,56,201,252
2785 pxor xmm13,xmm11
2786DB 69,15,56,201,243
2787 movdqa xmm2,xmm0
2788 movdqa xmm10,xmm8
2789DB 15,58,204,193,1
2790DB 15,56,200,213
2791DB 69,15,58,204,193,1
2792DB 69,15,56,200,212
2793DB 15,56,202,245
2794DB 69,15,56,202,236
2795 pxor xmm7,xmm5
2796DB 15,56,201,229
2797 pxor xmm14,xmm12
2798DB 69,15,56,201,220
2799 movdqa xmm1,xmm0
2800 movdqa xmm9,xmm8
2801DB 15,58,204,194,1
2802DB 15,56,200,206
2803DB 69,15,58,204,194,1
2804DB 69,15,56,200,205
2805DB 15,56,202,254
2806DB 69,15,56,202,245
2807 pxor xmm4,xmm6
2808DB 15,56,201,238
2809 pxor xmm11,xmm13
2810DB 69,15,56,201,229
2811 movdqa xmm2,xmm0
2812 movdqa xmm10,xmm8
2813DB 15,58,204,193,2
2814DB 15,56,200,215
2815DB 69,15,58,204,193,2
2816DB 69,15,56,200,214
2817DB 15,56,202,231
2818DB 69,15,56,202,222
2819 pxor xmm5,xmm7
2820DB 15,56,201,247
2821 pxor xmm12,xmm14
2822DB 69,15,56,201,238
2823 movdqa xmm1,xmm0
2824 movdqa xmm9,xmm8
2825DB 15,58,204,194,2
2826DB 15,56,200,204
2827DB 69,15,58,204,194,2
2828DB 69,15,56,200,203
2829DB 15,56,202,236
2830DB 69,15,56,202,227
2831 pxor xmm6,xmm4
2832DB 15,56,201,252
2833 pxor xmm13,xmm11
2834DB 69,15,56,201,243
2835 movdqa xmm2,xmm0
2836 movdqa xmm10,xmm8
2837DB 15,58,204,193,2
2838DB 15,56,200,213
2839DB 69,15,58,204,193,2
2840DB 69,15,56,200,212
2841DB 15,56,202,245
2842DB 69,15,56,202,236
2843 pxor xmm7,xmm5
2844DB 15,56,201,229
2845 pxor xmm14,xmm12
2846DB 69,15,56,201,220
2847 movdqa xmm1,xmm0
2848 movdqa xmm9,xmm8
2849DB 15,58,204,194,2
2850DB 15,56,200,206
2851DB 69,15,58,204,194,2
2852DB 69,15,56,200,205
2853DB 15,56,202,254
2854DB 69,15,56,202,245
2855 pxor xmm4,xmm6
2856DB 15,56,201,238
2857 pxor xmm11,xmm13
2858DB 69,15,56,201,229
2859 movdqa xmm2,xmm0
2860 movdqa xmm10,xmm8
2861DB 15,58,204,193,2
2862DB 15,56,200,215
2863DB 69,15,58,204,193,2
2864DB 69,15,56,200,214
2865DB 15,56,202,231
2866DB 69,15,56,202,222
2867 pxor xmm5,xmm7
2868DB 15,56,201,247
2869 pxor xmm12,xmm14
2870DB 69,15,56,201,238
2871 movdqa xmm1,xmm0
2872 movdqa xmm9,xmm8
2873DB 15,58,204,194,3
2874DB 15,56,200,204
2875DB 69,15,58,204,194,3
2876DB 69,15,56,200,203
2877DB 15,56,202,236
2878DB 69,15,56,202,227
2879 pxor xmm6,xmm4
2880DB 15,56,201,252
2881 pxor xmm13,xmm11
2882DB 69,15,56,201,243
2883 movdqa xmm2,xmm0
2884 movdqa xmm10,xmm8
2885DB 15,58,204,193,3
2886DB 15,56,200,213
2887DB 69,15,58,204,193,3
2888DB 69,15,56,200,212
2889DB 15,56,202,245
2890DB 69,15,56,202,236
2891 pxor xmm7,xmm5
2892 pxor xmm14,xmm12
2893
2894 mov ecx,1
2895 pxor xmm4,xmm4
2896 cmp ecx,DWORD[rbx]
2897 cmovge r8,rsp
2898
2899 movdqa xmm1,xmm0
2900 movdqa xmm9,xmm8
2901DB 15,58,204,194,3
2902DB 15,56,200,206
2903DB 69,15,58,204,194,3
2904DB 69,15,56,200,205
2905DB 15,56,202,254
2906DB 69,15,56,202,245
2907
2908 cmp ecx,DWORD[4+rbx]
2909 cmovge r9,rsp
2910 movq xmm6,QWORD[rbx]
2911
2912 movdqa xmm2,xmm0
2913 movdqa xmm10,xmm8
2914DB 15,58,204,193,3
2915DB 15,56,200,215
2916DB 69,15,58,204,193,3
2917DB 69,15,56,200,214
2918
2919 pshufd xmm11,xmm6,0x00
2920 pshufd xmm12,xmm6,0x55
2921 movdqa xmm7,xmm6
2922 pcmpgtd xmm11,xmm4
2923 pcmpgtd xmm12,xmm4
2924
2925 movdqa xmm1,xmm0
2926 movdqa xmm9,xmm8
2927DB 15,58,204,194,3
2928DB 15,56,200,204
2929DB 69,15,58,204,194,3
2930DB 68,15,56,200,204
2931
2932 pcmpgtd xmm7,xmm4
2933 pand xmm0,xmm11
2934 pand xmm1,xmm11
2935 pand xmm8,xmm12
2936 pand xmm9,xmm12
2937 paddd xmm6,xmm7
2938
2939 paddd xmm0,XMMWORD[64+rsp]
2940 paddd xmm1,XMMWORD[80+rsp]
2941 paddd xmm8,XMMWORD[96+rsp]
2942 paddd xmm9,XMMWORD[112+rsp]
2943
2944 movq QWORD[rbx],xmm6
2945 dec edx
2946 jnz NEAR $L$oop_shaext
2947
2948 mov edx,DWORD[280+rsp]
2949
2950 pshufd xmm0,xmm0,27
2951 pshufd xmm8,xmm8,27
2952
2953 movdqa xmm6,xmm0
2954 punpckldq xmm0,xmm8
2955 punpckhdq xmm6,xmm8
2956 punpckhdq xmm1,xmm9
2957 movq QWORD[(0-64)+rdi],xmm0
2958 psrldq xmm0,8
2959 movq QWORD[(64-64)+rdi],xmm6
2960 psrldq xmm6,8
2961 movq QWORD[(32-64)+rdi],xmm0
2962 psrldq xmm1,8
2963 movq QWORD[(96-64)+rdi],xmm6
2964 movq QWORD[(128-64)+rdi],xmm1
2965
2966 lea rdi,[8+rdi]
2967 lea rsi,[32+rsi]
2968 dec edx
2969 jnz NEAR $L$oop_grande_shaext
2970
2971$L$done_shaext:
2972
2973 movaps xmm6,XMMWORD[((-184))+rax]
2974 movaps xmm7,XMMWORD[((-168))+rax]
2975 movaps xmm8,XMMWORD[((-152))+rax]
2976 movaps xmm9,XMMWORD[((-136))+rax]
2977 movaps xmm10,XMMWORD[((-120))+rax]
2978 movaps xmm11,XMMWORD[((-104))+rax]
2979 movaps xmm12,XMMWORD[((-88))+rax]
2980 movaps xmm13,XMMWORD[((-72))+rax]
2981 movaps xmm14,XMMWORD[((-56))+rax]
2982 movaps xmm15,XMMWORD[((-40))+rax]
2983 mov rbp,QWORD[((-16))+rax]
2984 mov rbx,QWORD[((-8))+rax]
2985 lea rsp,[rax]
2986$L$epilogue_shaext:
2987 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2988 mov rsi,QWORD[16+rsp]
2989 DB 0F3h,0C3h ;repret
2990$L$SEH_end_sha1_multi_block_shaext:
2991
2992ALIGN 32
2993sha1_multi_block_avx:
2994 mov QWORD[8+rsp],rdi ;WIN64 prologue
2995 mov QWORD[16+rsp],rsi
2996 mov rax,rsp
2997$L$SEH_begin_sha1_multi_block_avx:
2998 mov rdi,rcx
2999 mov rsi,rdx
3000 mov rdx,r8
3001
3002
3003_avx_shortcut:
3004 shr rcx,32
3005 cmp edx,2
3006 jb NEAR $L$avx
3007 test ecx,32
3008 jnz NEAR _avx2_shortcut
3009 jmp NEAR $L$avx
3010ALIGN 32
3011$L$avx:
3012 mov rax,rsp
3013 push rbx
3014 push rbp
3015 lea rsp,[((-168))+rsp]
3016 movaps XMMWORD[rsp],xmm6
3017 movaps XMMWORD[16+rsp],xmm7
3018 movaps XMMWORD[32+rsp],xmm8
3019 movaps XMMWORD[48+rsp],xmm9
3020 movaps XMMWORD[(-120)+rax],xmm10
3021 movaps XMMWORD[(-104)+rax],xmm11
3022 movaps XMMWORD[(-88)+rax],xmm12
3023 movaps XMMWORD[(-72)+rax],xmm13
3024 movaps XMMWORD[(-56)+rax],xmm14
3025 movaps XMMWORD[(-40)+rax],xmm15
3026 sub rsp,288
3027 and rsp,-256
3028 mov QWORD[272+rsp],rax
3029$L$body_avx:
3030 lea rbp,[K_XX_XX]
3031 lea rbx,[256+rsp]
3032
3033 vzeroupper
3034$L$oop_grande_avx:
3035 mov DWORD[280+rsp],edx
3036 xor edx,edx
3037 mov r8,QWORD[rsi]
3038 mov ecx,DWORD[8+rsi]
3039 cmp ecx,edx
3040 cmovg edx,ecx
3041 test ecx,ecx
3042 mov DWORD[rbx],ecx
3043 cmovle r8,rbp
3044 mov r9,QWORD[16+rsi]
3045 mov ecx,DWORD[24+rsi]
3046 cmp ecx,edx
3047 cmovg edx,ecx
3048 test ecx,ecx
3049 mov DWORD[4+rbx],ecx
3050 cmovle r9,rbp
3051 mov r10,QWORD[32+rsi]
3052 mov ecx,DWORD[40+rsi]
3053 cmp ecx,edx
3054 cmovg edx,ecx
3055 test ecx,ecx
3056 mov DWORD[8+rbx],ecx
3057 cmovle r10,rbp
3058 mov r11,QWORD[48+rsi]
3059 mov ecx,DWORD[56+rsi]
3060 cmp ecx,edx
3061 cmovg edx,ecx
3062 test ecx,ecx
3063 mov DWORD[12+rbx],ecx
3064 cmovle r11,rbp
3065 test edx,edx
3066 jz NEAR $L$done_avx
3067
3068 vmovdqu xmm10,XMMWORD[rdi]
3069 lea rax,[128+rsp]
3070 vmovdqu xmm11,XMMWORD[32+rdi]
3071 vmovdqu xmm12,XMMWORD[64+rdi]
3072 vmovdqu xmm13,XMMWORD[96+rdi]
3073 vmovdqu xmm14,XMMWORD[128+rdi]
3074 vmovdqu xmm5,XMMWORD[96+rbp]
3075 jmp NEAR $L$oop_avx
3076
3077ALIGN 32
3078$L$oop_avx:
3079 vmovdqa xmm15,XMMWORD[((-32))+rbp]
3080 vmovd xmm0,DWORD[r8]
3081 lea r8,[64+r8]
3082 vmovd xmm2,DWORD[r9]
3083 lea r9,[64+r9]
3084 vpinsrd xmm0,xmm0,DWORD[r10],1
3085 lea r10,[64+r10]
3086 vpinsrd xmm2,xmm2,DWORD[r11],1
3087 lea r11,[64+r11]
3088 vmovd xmm1,DWORD[((-60))+r8]
3089 vpunpckldq xmm0,xmm0,xmm2
3090 vmovd xmm9,DWORD[((-60))+r9]
3091 vpshufb xmm0,xmm0,xmm5
3092 vpinsrd xmm1,xmm1,DWORD[((-60))+r10],1
3093 vpinsrd xmm9,xmm9,DWORD[((-60))+r11],1
3094 vpaddd xmm14,xmm14,xmm15
3095 vpslld xmm8,xmm10,5
3096 vpandn xmm7,xmm11,xmm13
3097 vpand xmm6,xmm11,xmm12
3098
3099 vmovdqa XMMWORD[(0-128)+rax],xmm0
3100 vpaddd xmm14,xmm14,xmm0
3101 vpunpckldq xmm1,xmm1,xmm9
3102 vpsrld xmm9,xmm10,27
3103 vpxor xmm6,xmm6,xmm7
3104 vmovd xmm2,DWORD[((-56))+r8]
3105
3106 vpslld xmm7,xmm11,30
3107 vpor xmm8,xmm8,xmm9
3108 vmovd xmm9,DWORD[((-56))+r9]
3109 vpaddd xmm14,xmm14,xmm6
3110
3111 vpsrld xmm11,xmm11,2
3112 vpaddd xmm14,xmm14,xmm8
3113 vpshufb xmm1,xmm1,xmm5
3114 vpor xmm11,xmm11,xmm7
3115 vpinsrd xmm2,xmm2,DWORD[((-56))+r10],1
3116 vpinsrd xmm9,xmm9,DWORD[((-56))+r11],1
3117 vpaddd xmm13,xmm13,xmm15
3118 vpslld xmm8,xmm14,5
3119 vpandn xmm7,xmm10,xmm12
3120 vpand xmm6,xmm10,xmm11
3121
3122 vmovdqa XMMWORD[(16-128)+rax],xmm1
3123 vpaddd xmm13,xmm13,xmm1
3124 vpunpckldq xmm2,xmm2,xmm9
3125 vpsrld xmm9,xmm14,27
3126 vpxor xmm6,xmm6,xmm7
3127 vmovd xmm3,DWORD[((-52))+r8]
3128
3129 vpslld xmm7,xmm10,30
3130 vpor xmm8,xmm8,xmm9
3131 vmovd xmm9,DWORD[((-52))+r9]
3132 vpaddd xmm13,xmm13,xmm6
3133
3134 vpsrld xmm10,xmm10,2
3135 vpaddd xmm13,xmm13,xmm8
3136 vpshufb xmm2,xmm2,xmm5
3137 vpor xmm10,xmm10,xmm7
3138 vpinsrd xmm3,xmm3,DWORD[((-52))+r10],1
3139 vpinsrd xmm9,xmm9,DWORD[((-52))+r11],1
3140 vpaddd xmm12,xmm12,xmm15
3141 vpslld xmm8,xmm13,5
3142 vpandn xmm7,xmm14,xmm11
3143 vpand xmm6,xmm14,xmm10
3144
3145 vmovdqa XMMWORD[(32-128)+rax],xmm2
3146 vpaddd xmm12,xmm12,xmm2
3147 vpunpckldq xmm3,xmm3,xmm9
3148 vpsrld xmm9,xmm13,27
3149 vpxor xmm6,xmm6,xmm7
3150 vmovd xmm4,DWORD[((-48))+r8]
3151
3152 vpslld xmm7,xmm14,30
3153 vpor xmm8,xmm8,xmm9
3154 vmovd xmm9,DWORD[((-48))+r9]
3155 vpaddd xmm12,xmm12,xmm6
3156
3157 vpsrld xmm14,xmm14,2
3158 vpaddd xmm12,xmm12,xmm8
3159 vpshufb xmm3,xmm3,xmm5
3160 vpor xmm14,xmm14,xmm7
3161 vpinsrd xmm4,xmm4,DWORD[((-48))+r10],1
3162 vpinsrd xmm9,xmm9,DWORD[((-48))+r11],1
3163 vpaddd xmm11,xmm11,xmm15
3164 vpslld xmm8,xmm12,5
3165 vpandn xmm7,xmm13,xmm10
3166 vpand xmm6,xmm13,xmm14
3167
3168 vmovdqa XMMWORD[(48-128)+rax],xmm3
3169 vpaddd xmm11,xmm11,xmm3
3170 vpunpckldq xmm4,xmm4,xmm9
3171 vpsrld xmm9,xmm12,27
3172 vpxor xmm6,xmm6,xmm7
3173 vmovd xmm0,DWORD[((-44))+r8]
3174
3175 vpslld xmm7,xmm13,30
3176 vpor xmm8,xmm8,xmm9
3177 vmovd xmm9,DWORD[((-44))+r9]
3178 vpaddd xmm11,xmm11,xmm6
3179
3180 vpsrld xmm13,xmm13,2
3181 vpaddd xmm11,xmm11,xmm8
3182 vpshufb xmm4,xmm4,xmm5
3183 vpor xmm13,xmm13,xmm7
3184 vpinsrd xmm0,xmm0,DWORD[((-44))+r10],1
3185 vpinsrd xmm9,xmm9,DWORD[((-44))+r11],1
3186 vpaddd xmm10,xmm10,xmm15
3187 vpslld xmm8,xmm11,5
3188 vpandn xmm7,xmm12,xmm14
3189 vpand xmm6,xmm12,xmm13
3190
3191 vmovdqa XMMWORD[(64-128)+rax],xmm4
3192 vpaddd xmm10,xmm10,xmm4
3193 vpunpckldq xmm0,xmm0,xmm9
3194 vpsrld xmm9,xmm11,27
3195 vpxor xmm6,xmm6,xmm7
3196 vmovd xmm1,DWORD[((-40))+r8]
3197
3198 vpslld xmm7,xmm12,30
3199 vpor xmm8,xmm8,xmm9
3200 vmovd xmm9,DWORD[((-40))+r9]
3201 vpaddd xmm10,xmm10,xmm6
3202
3203 vpsrld xmm12,xmm12,2
3204 vpaddd xmm10,xmm10,xmm8
3205 vpshufb xmm0,xmm0,xmm5
3206 vpor xmm12,xmm12,xmm7
3207 vpinsrd xmm1,xmm1,DWORD[((-40))+r10],1
3208 vpinsrd xmm9,xmm9,DWORD[((-40))+r11],1
3209 vpaddd xmm14,xmm14,xmm15
3210 vpslld xmm8,xmm10,5
3211 vpandn xmm7,xmm11,xmm13
3212 vpand xmm6,xmm11,xmm12
3213
3214 vmovdqa XMMWORD[(80-128)+rax],xmm0
3215 vpaddd xmm14,xmm14,xmm0
3216 vpunpckldq xmm1,xmm1,xmm9
3217 vpsrld xmm9,xmm10,27
3218 vpxor xmm6,xmm6,xmm7
3219 vmovd xmm2,DWORD[((-36))+r8]
3220
3221 vpslld xmm7,xmm11,30
3222 vpor xmm8,xmm8,xmm9
3223 vmovd xmm9,DWORD[((-36))+r9]
3224 vpaddd xmm14,xmm14,xmm6
3225
3226 vpsrld xmm11,xmm11,2
3227 vpaddd xmm14,xmm14,xmm8
3228 vpshufb xmm1,xmm1,xmm5
3229 vpor xmm11,xmm11,xmm7
3230 vpinsrd xmm2,xmm2,DWORD[((-36))+r10],1
3231 vpinsrd xmm9,xmm9,DWORD[((-36))+r11],1
3232 vpaddd xmm13,xmm13,xmm15
3233 vpslld xmm8,xmm14,5
3234 vpandn xmm7,xmm10,xmm12
3235 vpand xmm6,xmm10,xmm11
3236
3237 vmovdqa XMMWORD[(96-128)+rax],xmm1
3238 vpaddd xmm13,xmm13,xmm1
3239 vpunpckldq xmm2,xmm2,xmm9
3240 vpsrld xmm9,xmm14,27
3241 vpxor xmm6,xmm6,xmm7
3242 vmovd xmm3,DWORD[((-32))+r8]
3243
3244 vpslld xmm7,xmm10,30
3245 vpor xmm8,xmm8,xmm9
3246 vmovd xmm9,DWORD[((-32))+r9]
3247 vpaddd xmm13,xmm13,xmm6
3248
3249 vpsrld xmm10,xmm10,2
3250 vpaddd xmm13,xmm13,xmm8
3251 vpshufb xmm2,xmm2,xmm5
3252 vpor xmm10,xmm10,xmm7
3253 vpinsrd xmm3,xmm3,DWORD[((-32))+r10],1
3254 vpinsrd xmm9,xmm9,DWORD[((-32))+r11],1
3255 vpaddd xmm12,xmm12,xmm15
3256 vpslld xmm8,xmm13,5
3257 vpandn xmm7,xmm14,xmm11
3258 vpand xmm6,xmm14,xmm10
3259
3260 vmovdqa XMMWORD[(112-128)+rax],xmm2
3261 vpaddd xmm12,xmm12,xmm2
3262 vpunpckldq xmm3,xmm3,xmm9
3263 vpsrld xmm9,xmm13,27
3264 vpxor xmm6,xmm6,xmm7
3265 vmovd xmm4,DWORD[((-28))+r8]
3266
3267 vpslld xmm7,xmm14,30
3268 vpor xmm8,xmm8,xmm9
3269 vmovd xmm9,DWORD[((-28))+r9]
3270 vpaddd xmm12,xmm12,xmm6
3271
3272 vpsrld xmm14,xmm14,2
3273 vpaddd xmm12,xmm12,xmm8
3274 vpshufb xmm3,xmm3,xmm5
3275 vpor xmm14,xmm14,xmm7
3276 vpinsrd xmm4,xmm4,DWORD[((-28))+r10],1
3277 vpinsrd xmm9,xmm9,DWORD[((-28))+r11],1
3278 vpaddd xmm11,xmm11,xmm15
3279 vpslld xmm8,xmm12,5
3280 vpandn xmm7,xmm13,xmm10
3281 vpand xmm6,xmm13,xmm14
3282
3283 vmovdqa XMMWORD[(128-128)+rax],xmm3
3284 vpaddd xmm11,xmm11,xmm3
3285 vpunpckldq xmm4,xmm4,xmm9
3286 vpsrld xmm9,xmm12,27
3287 vpxor xmm6,xmm6,xmm7
3288 vmovd xmm0,DWORD[((-24))+r8]
3289
3290 vpslld xmm7,xmm13,30
3291 vpor xmm8,xmm8,xmm9
3292 vmovd xmm9,DWORD[((-24))+r9]
3293 vpaddd xmm11,xmm11,xmm6
3294
3295 vpsrld xmm13,xmm13,2
3296 vpaddd xmm11,xmm11,xmm8
3297 vpshufb xmm4,xmm4,xmm5
3298 vpor xmm13,xmm13,xmm7
3299 vpinsrd xmm0,xmm0,DWORD[((-24))+r10],1
3300 vpinsrd xmm9,xmm9,DWORD[((-24))+r11],1
3301 vpaddd xmm10,xmm10,xmm15
3302 vpslld xmm8,xmm11,5
3303 vpandn xmm7,xmm12,xmm14
3304 vpand xmm6,xmm12,xmm13
3305
3306 vmovdqa XMMWORD[(144-128)+rax],xmm4
3307 vpaddd xmm10,xmm10,xmm4
3308 vpunpckldq xmm0,xmm0,xmm9
3309 vpsrld xmm9,xmm11,27
3310 vpxor xmm6,xmm6,xmm7
3311 vmovd xmm1,DWORD[((-20))+r8]
3312
3313 vpslld xmm7,xmm12,30
3314 vpor xmm8,xmm8,xmm9
3315 vmovd xmm9,DWORD[((-20))+r9]
3316 vpaddd xmm10,xmm10,xmm6
3317
3318 vpsrld xmm12,xmm12,2
3319 vpaddd xmm10,xmm10,xmm8
3320 vpshufb xmm0,xmm0,xmm5
3321 vpor xmm12,xmm12,xmm7
3322 vpinsrd xmm1,xmm1,DWORD[((-20))+r10],1
3323 vpinsrd xmm9,xmm9,DWORD[((-20))+r11],1
3324 vpaddd xmm14,xmm14,xmm15
3325 vpslld xmm8,xmm10,5
3326 vpandn xmm7,xmm11,xmm13
3327 vpand xmm6,xmm11,xmm12
3328
3329 vmovdqa XMMWORD[(160-128)+rax],xmm0
3330 vpaddd xmm14,xmm14,xmm0
3331 vpunpckldq xmm1,xmm1,xmm9
3332 vpsrld xmm9,xmm10,27
3333 vpxor xmm6,xmm6,xmm7
3334 vmovd xmm2,DWORD[((-16))+r8]
3335
3336 vpslld xmm7,xmm11,30
3337 vpor xmm8,xmm8,xmm9
3338 vmovd xmm9,DWORD[((-16))+r9]
3339 vpaddd xmm14,xmm14,xmm6
3340
3341 vpsrld xmm11,xmm11,2
3342 vpaddd xmm14,xmm14,xmm8
3343 vpshufb xmm1,xmm1,xmm5
3344 vpor xmm11,xmm11,xmm7
3345 vpinsrd xmm2,xmm2,DWORD[((-16))+r10],1
3346 vpinsrd xmm9,xmm9,DWORD[((-16))+r11],1
3347 vpaddd xmm13,xmm13,xmm15
3348 vpslld xmm8,xmm14,5
3349 vpandn xmm7,xmm10,xmm12
3350 vpand xmm6,xmm10,xmm11
3351
3352 vmovdqa XMMWORD[(176-128)+rax],xmm1
3353 vpaddd xmm13,xmm13,xmm1
3354 vpunpckldq xmm2,xmm2,xmm9
3355 vpsrld xmm9,xmm14,27
3356 vpxor xmm6,xmm6,xmm7
3357 vmovd xmm3,DWORD[((-12))+r8]
3358
3359 vpslld xmm7,xmm10,30
3360 vpor xmm8,xmm8,xmm9
3361 vmovd xmm9,DWORD[((-12))+r9]
3362 vpaddd xmm13,xmm13,xmm6
3363
3364 vpsrld xmm10,xmm10,2
3365 vpaddd xmm13,xmm13,xmm8
3366 vpshufb xmm2,xmm2,xmm5
3367 vpor xmm10,xmm10,xmm7
3368 vpinsrd xmm3,xmm3,DWORD[((-12))+r10],1
3369 vpinsrd xmm9,xmm9,DWORD[((-12))+r11],1
3370 vpaddd xmm12,xmm12,xmm15
3371 vpslld xmm8,xmm13,5
3372 vpandn xmm7,xmm14,xmm11
3373 vpand xmm6,xmm14,xmm10
3374
3375 vmovdqa XMMWORD[(192-128)+rax],xmm2
3376 vpaddd xmm12,xmm12,xmm2
3377 vpunpckldq xmm3,xmm3,xmm9
3378 vpsrld xmm9,xmm13,27
3379 vpxor xmm6,xmm6,xmm7
3380 vmovd xmm4,DWORD[((-8))+r8]
3381
3382 vpslld xmm7,xmm14,30
3383 vpor xmm8,xmm8,xmm9
3384 vmovd xmm9,DWORD[((-8))+r9]
3385 vpaddd xmm12,xmm12,xmm6
3386
3387 vpsrld xmm14,xmm14,2
3388 vpaddd xmm12,xmm12,xmm8
3389 vpshufb xmm3,xmm3,xmm5
3390 vpor xmm14,xmm14,xmm7
3391 vpinsrd xmm4,xmm4,DWORD[((-8))+r10],1
3392 vpinsrd xmm9,xmm9,DWORD[((-8))+r11],1
3393 vpaddd xmm11,xmm11,xmm15
3394 vpslld xmm8,xmm12,5
3395 vpandn xmm7,xmm13,xmm10
3396 vpand xmm6,xmm13,xmm14
3397
3398 vmovdqa XMMWORD[(208-128)+rax],xmm3
3399 vpaddd xmm11,xmm11,xmm3
3400 vpunpckldq xmm4,xmm4,xmm9
3401 vpsrld xmm9,xmm12,27
3402 vpxor xmm6,xmm6,xmm7
3403 vmovd xmm0,DWORD[((-4))+r8]
3404
3405 vpslld xmm7,xmm13,30
3406 vpor xmm8,xmm8,xmm9
3407 vmovd xmm9,DWORD[((-4))+r9]
3408 vpaddd xmm11,xmm11,xmm6
3409
3410 vpsrld xmm13,xmm13,2
3411 vpaddd xmm11,xmm11,xmm8
3412 vpshufb xmm4,xmm4,xmm5
3413 vpor xmm13,xmm13,xmm7
3414 vmovdqa xmm1,XMMWORD[((0-128))+rax]
3415 vpinsrd xmm0,xmm0,DWORD[((-4))+r10],1
3416 vpinsrd xmm9,xmm9,DWORD[((-4))+r11],1
3417 vpaddd xmm10,xmm10,xmm15
3418 prefetcht0 [63+r8]
3419 vpslld xmm8,xmm11,5
3420 vpandn xmm7,xmm12,xmm14
3421 vpand xmm6,xmm12,xmm13
3422
3423 vmovdqa XMMWORD[(224-128)+rax],xmm4
3424 vpaddd xmm10,xmm10,xmm4
3425 vpunpckldq xmm0,xmm0,xmm9
3426 vpsrld xmm9,xmm11,27
3427 prefetcht0 [63+r9]
3428 vpxor xmm6,xmm6,xmm7
3429
3430 vpslld xmm7,xmm12,30
3431 vpor xmm8,xmm8,xmm9
3432 prefetcht0 [63+r10]
3433 vpaddd xmm10,xmm10,xmm6
3434
3435 vpsrld xmm12,xmm12,2
3436 vpaddd xmm10,xmm10,xmm8
3437 prefetcht0 [63+r11]
3438 vpshufb xmm0,xmm0,xmm5
3439 vpor xmm12,xmm12,xmm7
3440 vmovdqa xmm2,XMMWORD[((16-128))+rax]
3441 vpxor xmm1,xmm1,xmm3
3442 vmovdqa xmm3,XMMWORD[((32-128))+rax]
3443
3444 vpaddd xmm14,xmm14,xmm15
3445 vpslld xmm8,xmm10,5
3446 vpandn xmm7,xmm11,xmm13
3447
3448 vpand xmm6,xmm11,xmm12
3449
3450 vmovdqa XMMWORD[(240-128)+rax],xmm0
3451 vpaddd xmm14,xmm14,xmm0
3452 vpxor xmm1,xmm1,XMMWORD[((128-128))+rax]
3453 vpsrld xmm9,xmm10,27
3454 vpxor xmm6,xmm6,xmm7
3455 vpxor xmm1,xmm1,xmm3
3456
3457
3458 vpslld xmm7,xmm11,30
3459 vpor xmm8,xmm8,xmm9
3460 vpaddd xmm14,xmm14,xmm6
3461
3462 vpsrld xmm5,xmm1,31
3463 vpaddd xmm1,xmm1,xmm1
3464
3465 vpsrld xmm11,xmm11,2
3466
3467 vpaddd xmm14,xmm14,xmm8
3468 vpor xmm1,xmm1,xmm5
3469 vpor xmm11,xmm11,xmm7
3470 vpxor xmm2,xmm2,xmm4
3471 vmovdqa xmm4,XMMWORD[((48-128))+rax]
3472
3473 vpaddd xmm13,xmm13,xmm15
3474 vpslld xmm8,xmm14,5
3475 vpandn xmm7,xmm10,xmm12
3476
3477 vpand xmm6,xmm10,xmm11
3478
3479 vmovdqa XMMWORD[(0-128)+rax],xmm1
3480 vpaddd xmm13,xmm13,xmm1
3481 vpxor xmm2,xmm2,XMMWORD[((144-128))+rax]
3482 vpsrld xmm9,xmm14,27
3483 vpxor xmm6,xmm6,xmm7
3484 vpxor xmm2,xmm2,xmm4
3485
3486
3487 vpslld xmm7,xmm10,30
3488 vpor xmm8,xmm8,xmm9
3489 vpaddd xmm13,xmm13,xmm6
3490
3491 vpsrld xmm5,xmm2,31
3492 vpaddd xmm2,xmm2,xmm2
3493
3494 vpsrld xmm10,xmm10,2
3495
3496 vpaddd xmm13,xmm13,xmm8
3497 vpor xmm2,xmm2,xmm5
3498 vpor xmm10,xmm10,xmm7
3499 vpxor xmm3,xmm3,xmm0
3500 vmovdqa xmm0,XMMWORD[((64-128))+rax]
3501
3502 vpaddd xmm12,xmm12,xmm15
3503 vpslld xmm8,xmm13,5
3504 vpandn xmm7,xmm14,xmm11
3505
3506 vpand xmm6,xmm14,xmm10
3507
3508 vmovdqa XMMWORD[(16-128)+rax],xmm2
3509 vpaddd xmm12,xmm12,xmm2
3510 vpxor xmm3,xmm3,XMMWORD[((160-128))+rax]
3511 vpsrld xmm9,xmm13,27
3512 vpxor xmm6,xmm6,xmm7
3513 vpxor xmm3,xmm3,xmm0
3514
3515
3516 vpslld xmm7,xmm14,30
3517 vpor xmm8,xmm8,xmm9
3518 vpaddd xmm12,xmm12,xmm6
3519
3520 vpsrld xmm5,xmm3,31
3521 vpaddd xmm3,xmm3,xmm3
3522
3523 vpsrld xmm14,xmm14,2
3524
3525 vpaddd xmm12,xmm12,xmm8
3526 vpor xmm3,xmm3,xmm5
3527 vpor xmm14,xmm14,xmm7
3528 vpxor xmm4,xmm4,xmm1
3529 vmovdqa xmm1,XMMWORD[((80-128))+rax]
3530
3531 vpaddd xmm11,xmm11,xmm15
3532 vpslld xmm8,xmm12,5
3533 vpandn xmm7,xmm13,xmm10
3534
3535 vpand xmm6,xmm13,xmm14
3536
3537 vmovdqa XMMWORD[(32-128)+rax],xmm3
3538 vpaddd xmm11,xmm11,xmm3
3539 vpxor xmm4,xmm4,XMMWORD[((176-128))+rax]
3540 vpsrld xmm9,xmm12,27
3541 vpxor xmm6,xmm6,xmm7
3542 vpxor xmm4,xmm4,xmm1
3543
3544
3545 vpslld xmm7,xmm13,30
3546 vpor xmm8,xmm8,xmm9
3547 vpaddd xmm11,xmm11,xmm6
3548
3549 vpsrld xmm5,xmm4,31
3550 vpaddd xmm4,xmm4,xmm4
3551
3552 vpsrld xmm13,xmm13,2
3553
3554 vpaddd xmm11,xmm11,xmm8
3555 vpor xmm4,xmm4,xmm5
3556 vpor xmm13,xmm13,xmm7
3557 vpxor xmm0,xmm0,xmm2
3558 vmovdqa xmm2,XMMWORD[((96-128))+rax]
3559
3560 vpaddd xmm10,xmm10,xmm15
3561 vpslld xmm8,xmm11,5
3562 vpandn xmm7,xmm12,xmm14
3563
3564 vpand xmm6,xmm12,xmm13
3565
3566 vmovdqa XMMWORD[(48-128)+rax],xmm4
3567 vpaddd xmm10,xmm10,xmm4
3568 vpxor xmm0,xmm0,XMMWORD[((192-128))+rax]
3569 vpsrld xmm9,xmm11,27
3570 vpxor xmm6,xmm6,xmm7
3571 vpxor xmm0,xmm0,xmm2
3572
3573
3574 vpslld xmm7,xmm12,30
3575 vpor xmm8,xmm8,xmm9
3576 vpaddd xmm10,xmm10,xmm6
3577
3578 vpsrld xmm5,xmm0,31
3579 vpaddd xmm0,xmm0,xmm0
3580
3581 vpsrld xmm12,xmm12,2
3582
3583 vpaddd xmm10,xmm10,xmm8
3584 vpor xmm0,xmm0,xmm5
3585 vpor xmm12,xmm12,xmm7
3586 vmovdqa xmm15,XMMWORD[rbp]
3587 vpxor xmm1,xmm1,xmm3
3588 vmovdqa xmm3,XMMWORD[((112-128))+rax]
3589
3590 vpslld xmm8,xmm10,5
3591 vpaddd xmm14,xmm14,xmm15
3592 vpxor xmm6,xmm13,xmm11
3593 vmovdqa XMMWORD[(64-128)+rax],xmm0
3594 vpaddd xmm14,xmm14,xmm0
3595 vpxor xmm1,xmm1,XMMWORD[((208-128))+rax]
3596 vpsrld xmm9,xmm10,27
3597 vpxor xmm6,xmm6,xmm12
3598 vpxor xmm1,xmm1,xmm3
3599
3600 vpslld xmm7,xmm11,30
3601 vpor xmm8,xmm8,xmm9
3602 vpaddd xmm14,xmm14,xmm6
3603 vpsrld xmm5,xmm1,31
3604 vpaddd xmm1,xmm1,xmm1
3605
3606 vpsrld xmm11,xmm11,2
3607 vpaddd xmm14,xmm14,xmm8
3608 vpor xmm1,xmm1,xmm5
3609 vpor xmm11,xmm11,xmm7
3610 vpxor xmm2,xmm2,xmm4
3611 vmovdqa xmm4,XMMWORD[((128-128))+rax]
3612
3613 vpslld xmm8,xmm14,5
3614 vpaddd xmm13,xmm13,xmm15
3615 vpxor xmm6,xmm12,xmm10
3616 vmovdqa XMMWORD[(80-128)+rax],xmm1
3617 vpaddd xmm13,xmm13,xmm1
3618 vpxor xmm2,xmm2,XMMWORD[((224-128))+rax]
3619 vpsrld xmm9,xmm14,27
3620 vpxor xmm6,xmm6,xmm11
3621 vpxor xmm2,xmm2,xmm4
3622
3623 vpslld xmm7,xmm10,30
3624 vpor xmm8,xmm8,xmm9
3625 vpaddd xmm13,xmm13,xmm6
3626 vpsrld xmm5,xmm2,31
3627 vpaddd xmm2,xmm2,xmm2
3628
3629 vpsrld xmm10,xmm10,2
3630 vpaddd xmm13,xmm13,xmm8
3631 vpor xmm2,xmm2,xmm5
3632 vpor xmm10,xmm10,xmm7
3633 vpxor xmm3,xmm3,xmm0
3634 vmovdqa xmm0,XMMWORD[((144-128))+rax]
3635
3636 vpslld xmm8,xmm13,5
3637 vpaddd xmm12,xmm12,xmm15
3638 vpxor xmm6,xmm11,xmm14
3639 vmovdqa XMMWORD[(96-128)+rax],xmm2
3640 vpaddd xmm12,xmm12,xmm2
3641 vpxor xmm3,xmm3,XMMWORD[((240-128))+rax]
3642 vpsrld xmm9,xmm13,27
3643 vpxor xmm6,xmm6,xmm10
3644 vpxor xmm3,xmm3,xmm0
3645
3646 vpslld xmm7,xmm14,30
3647 vpor xmm8,xmm8,xmm9
3648 vpaddd xmm12,xmm12,xmm6
3649 vpsrld xmm5,xmm3,31
3650 vpaddd xmm3,xmm3,xmm3
3651
3652 vpsrld xmm14,xmm14,2
3653 vpaddd xmm12,xmm12,xmm8
3654 vpor xmm3,xmm3,xmm5
3655 vpor xmm14,xmm14,xmm7
3656 vpxor xmm4,xmm4,xmm1
3657 vmovdqa xmm1,XMMWORD[((160-128))+rax]
3658
3659 vpslld xmm8,xmm12,5
3660 vpaddd xmm11,xmm11,xmm15
3661 vpxor xmm6,xmm10,xmm13
3662 vmovdqa XMMWORD[(112-128)+rax],xmm3
3663 vpaddd xmm11,xmm11,xmm3
3664 vpxor xmm4,xmm4,XMMWORD[((0-128))+rax]
3665 vpsrld xmm9,xmm12,27
3666 vpxor xmm6,xmm6,xmm14
3667 vpxor xmm4,xmm4,xmm1
3668
3669 vpslld xmm7,xmm13,30
3670 vpor xmm8,xmm8,xmm9
3671 vpaddd xmm11,xmm11,xmm6
3672 vpsrld xmm5,xmm4,31
3673 vpaddd xmm4,xmm4,xmm4
3674
3675 vpsrld xmm13,xmm13,2
3676 vpaddd xmm11,xmm11,xmm8
3677 vpor xmm4,xmm4,xmm5
3678 vpor xmm13,xmm13,xmm7
3679 vpxor xmm0,xmm0,xmm2
3680 vmovdqa xmm2,XMMWORD[((176-128))+rax]
3681
3682 vpslld xmm8,xmm11,5
3683 vpaddd xmm10,xmm10,xmm15
3684 vpxor xmm6,xmm14,xmm12
3685 vmovdqa XMMWORD[(128-128)+rax],xmm4
3686 vpaddd xmm10,xmm10,xmm4
3687 vpxor xmm0,xmm0,XMMWORD[((16-128))+rax]
3688 vpsrld xmm9,xmm11,27
3689 vpxor xmm6,xmm6,xmm13
3690 vpxor xmm0,xmm0,xmm2
3691
3692 vpslld xmm7,xmm12,30
3693 vpor xmm8,xmm8,xmm9
3694 vpaddd xmm10,xmm10,xmm6
3695 vpsrld xmm5,xmm0,31
3696 vpaddd xmm0,xmm0,xmm0
3697
3698 vpsrld xmm12,xmm12,2
3699 vpaddd xmm10,xmm10,xmm8
3700 vpor xmm0,xmm0,xmm5
3701 vpor xmm12,xmm12,xmm7
3702 vpxor xmm1,xmm1,xmm3
3703 vmovdqa xmm3,XMMWORD[((192-128))+rax]
3704
3705 vpslld xmm8,xmm10,5
3706 vpaddd xmm14,xmm14,xmm15
3707 vpxor xmm6,xmm13,xmm11
3708 vmovdqa XMMWORD[(144-128)+rax],xmm0
3709 vpaddd xmm14,xmm14,xmm0
3710 vpxor xmm1,xmm1,XMMWORD[((32-128))+rax]
3711 vpsrld xmm9,xmm10,27
3712 vpxor xmm6,xmm6,xmm12
3713 vpxor xmm1,xmm1,xmm3
3714
3715 vpslld xmm7,xmm11,30
3716 vpor xmm8,xmm8,xmm9
3717 vpaddd xmm14,xmm14,xmm6
3718 vpsrld xmm5,xmm1,31
3719 vpaddd xmm1,xmm1,xmm1
3720
3721 vpsrld xmm11,xmm11,2
3722 vpaddd xmm14,xmm14,xmm8
3723 vpor xmm1,xmm1,xmm5
3724 vpor xmm11,xmm11,xmm7
3725 vpxor xmm2,xmm2,xmm4
3726 vmovdqa xmm4,XMMWORD[((208-128))+rax]
3727
3728 vpslld xmm8,xmm14,5
3729 vpaddd xmm13,xmm13,xmm15
3730 vpxor xmm6,xmm12,xmm10
3731 vmovdqa XMMWORD[(160-128)+rax],xmm1
3732 vpaddd xmm13,xmm13,xmm1
3733 vpxor xmm2,xmm2,XMMWORD[((48-128))+rax]
3734 vpsrld xmm9,xmm14,27
3735 vpxor xmm6,xmm6,xmm11
3736 vpxor xmm2,xmm2,xmm4
3737
3738 vpslld xmm7,xmm10,30
3739 vpor xmm8,xmm8,xmm9
3740 vpaddd xmm13,xmm13,xmm6
3741 vpsrld xmm5,xmm2,31
3742 vpaddd xmm2,xmm2,xmm2
3743
3744 vpsrld xmm10,xmm10,2
3745 vpaddd xmm13,xmm13,xmm8
3746 vpor xmm2,xmm2,xmm5
3747 vpor xmm10,xmm10,xmm7
3748 vpxor xmm3,xmm3,xmm0
3749 vmovdqa xmm0,XMMWORD[((224-128))+rax]
3750
3751 vpslld xmm8,xmm13,5
3752 vpaddd xmm12,xmm12,xmm15
3753 vpxor xmm6,xmm11,xmm14
3754 vmovdqa XMMWORD[(176-128)+rax],xmm2
3755 vpaddd xmm12,xmm12,xmm2
3756 vpxor xmm3,xmm3,XMMWORD[((64-128))+rax]
3757 vpsrld xmm9,xmm13,27
3758 vpxor xmm6,xmm6,xmm10
3759 vpxor xmm3,xmm3,xmm0
3760
3761 vpslld xmm7,xmm14,30
3762 vpor xmm8,xmm8,xmm9
3763 vpaddd xmm12,xmm12,xmm6
3764 vpsrld xmm5,xmm3,31
3765 vpaddd xmm3,xmm3,xmm3
3766
3767 vpsrld xmm14,xmm14,2
3768 vpaddd xmm12,xmm12,xmm8
3769 vpor xmm3,xmm3,xmm5
3770 vpor xmm14,xmm14,xmm7
3771 vpxor xmm4,xmm4,xmm1
3772 vmovdqa xmm1,XMMWORD[((240-128))+rax]
3773
3774 vpslld xmm8,xmm12,5
3775 vpaddd xmm11,xmm11,xmm15
3776 vpxor xmm6,xmm10,xmm13
3777 vmovdqa XMMWORD[(192-128)+rax],xmm3
3778 vpaddd xmm11,xmm11,xmm3
3779 vpxor xmm4,xmm4,XMMWORD[((80-128))+rax]
3780 vpsrld xmm9,xmm12,27
3781 vpxor xmm6,xmm6,xmm14
3782 vpxor xmm4,xmm4,xmm1
3783
3784 vpslld xmm7,xmm13,30
3785 vpor xmm8,xmm8,xmm9
3786 vpaddd xmm11,xmm11,xmm6
3787 vpsrld xmm5,xmm4,31
3788 vpaddd xmm4,xmm4,xmm4
3789
3790 vpsrld xmm13,xmm13,2
3791 vpaddd xmm11,xmm11,xmm8
3792 vpor xmm4,xmm4,xmm5
3793 vpor xmm13,xmm13,xmm7
3794 vpxor xmm0,xmm0,xmm2
3795 vmovdqa xmm2,XMMWORD[((0-128))+rax]
3796
3797 vpslld xmm8,xmm11,5
3798 vpaddd xmm10,xmm10,xmm15
3799 vpxor xmm6,xmm14,xmm12
3800 vmovdqa XMMWORD[(208-128)+rax],xmm4
3801 vpaddd xmm10,xmm10,xmm4
3802 vpxor xmm0,xmm0,XMMWORD[((96-128))+rax]
3803 vpsrld xmm9,xmm11,27
3804 vpxor xmm6,xmm6,xmm13
3805 vpxor xmm0,xmm0,xmm2
3806
3807 vpslld xmm7,xmm12,30
3808 vpor xmm8,xmm8,xmm9
3809 vpaddd xmm10,xmm10,xmm6
3810 vpsrld xmm5,xmm0,31
3811 vpaddd xmm0,xmm0,xmm0
3812
3813 vpsrld xmm12,xmm12,2
3814 vpaddd xmm10,xmm10,xmm8
3815 vpor xmm0,xmm0,xmm5
3816 vpor xmm12,xmm12,xmm7
3817 vpxor xmm1,xmm1,xmm3
3818 vmovdqa xmm3,XMMWORD[((16-128))+rax]
3819
3820 vpslld xmm8,xmm10,5
3821 vpaddd xmm14,xmm14,xmm15
3822 vpxor xmm6,xmm13,xmm11
3823 vmovdqa XMMWORD[(224-128)+rax],xmm0
3824 vpaddd xmm14,xmm14,xmm0
3825 vpxor xmm1,xmm1,XMMWORD[((112-128))+rax]
3826 vpsrld xmm9,xmm10,27
3827 vpxor xmm6,xmm6,xmm12
3828 vpxor xmm1,xmm1,xmm3
3829
3830 vpslld xmm7,xmm11,30
3831 vpor xmm8,xmm8,xmm9
3832 vpaddd xmm14,xmm14,xmm6
3833 vpsrld xmm5,xmm1,31
3834 vpaddd xmm1,xmm1,xmm1
3835
3836 vpsrld xmm11,xmm11,2
3837 vpaddd xmm14,xmm14,xmm8
3838 vpor xmm1,xmm1,xmm5
3839 vpor xmm11,xmm11,xmm7
3840 vpxor xmm2,xmm2,xmm4
3841 vmovdqa xmm4,XMMWORD[((32-128))+rax]
3842
3843 vpslld xmm8,xmm14,5
3844 vpaddd xmm13,xmm13,xmm15
3845 vpxor xmm6,xmm12,xmm10
3846 vmovdqa XMMWORD[(240-128)+rax],xmm1
3847 vpaddd xmm13,xmm13,xmm1
3848 vpxor xmm2,xmm2,XMMWORD[((128-128))+rax]
3849 vpsrld xmm9,xmm14,27
3850 vpxor xmm6,xmm6,xmm11
3851 vpxor xmm2,xmm2,xmm4
3852
3853 vpslld xmm7,xmm10,30
3854 vpor xmm8,xmm8,xmm9
3855 vpaddd xmm13,xmm13,xmm6
3856 vpsrld xmm5,xmm2,31
3857 vpaddd xmm2,xmm2,xmm2
3858
3859 vpsrld xmm10,xmm10,2
3860 vpaddd xmm13,xmm13,xmm8
3861 vpor xmm2,xmm2,xmm5
3862 vpor xmm10,xmm10,xmm7
3863 vpxor xmm3,xmm3,xmm0
3864 vmovdqa xmm0,XMMWORD[((48-128))+rax]
3865
3866 vpslld xmm8,xmm13,5
3867 vpaddd xmm12,xmm12,xmm15
3868 vpxor xmm6,xmm11,xmm14
3869 vmovdqa XMMWORD[(0-128)+rax],xmm2
3870 vpaddd xmm12,xmm12,xmm2
3871 vpxor xmm3,xmm3,XMMWORD[((144-128))+rax]
3872 vpsrld xmm9,xmm13,27
3873 vpxor xmm6,xmm6,xmm10
3874 vpxor xmm3,xmm3,xmm0
3875
3876 vpslld xmm7,xmm14,30
3877 vpor xmm8,xmm8,xmm9
3878 vpaddd xmm12,xmm12,xmm6
3879 vpsrld xmm5,xmm3,31
3880 vpaddd xmm3,xmm3,xmm3
3881
3882 vpsrld xmm14,xmm14,2
3883 vpaddd xmm12,xmm12,xmm8
3884 vpor xmm3,xmm3,xmm5
3885 vpor xmm14,xmm14,xmm7
3886 vpxor xmm4,xmm4,xmm1
3887 vmovdqa xmm1,XMMWORD[((64-128))+rax]
3888
3889 vpslld xmm8,xmm12,5
3890 vpaddd xmm11,xmm11,xmm15
3891 vpxor xmm6,xmm10,xmm13
3892 vmovdqa XMMWORD[(16-128)+rax],xmm3
3893 vpaddd xmm11,xmm11,xmm3
3894 vpxor xmm4,xmm4,XMMWORD[((160-128))+rax]
3895 vpsrld xmm9,xmm12,27
3896 vpxor xmm6,xmm6,xmm14
3897 vpxor xmm4,xmm4,xmm1
3898
3899 vpslld xmm7,xmm13,30
3900 vpor xmm8,xmm8,xmm9
3901 vpaddd xmm11,xmm11,xmm6
3902 vpsrld xmm5,xmm4,31
3903 vpaddd xmm4,xmm4,xmm4
3904
3905 vpsrld xmm13,xmm13,2
3906 vpaddd xmm11,xmm11,xmm8
3907 vpor xmm4,xmm4,xmm5
3908 vpor xmm13,xmm13,xmm7
3909 vpxor xmm0,xmm0,xmm2
3910 vmovdqa xmm2,XMMWORD[((80-128))+rax]
3911
3912 vpslld xmm8,xmm11,5
3913 vpaddd xmm10,xmm10,xmm15
3914 vpxor xmm6,xmm14,xmm12
3915 vmovdqa XMMWORD[(32-128)+rax],xmm4
3916 vpaddd xmm10,xmm10,xmm4
3917 vpxor xmm0,xmm0,XMMWORD[((176-128))+rax]
3918 vpsrld xmm9,xmm11,27
3919 vpxor xmm6,xmm6,xmm13
3920 vpxor xmm0,xmm0,xmm2
3921
3922 vpslld xmm7,xmm12,30
3923 vpor xmm8,xmm8,xmm9
3924 vpaddd xmm10,xmm10,xmm6
3925 vpsrld xmm5,xmm0,31
3926 vpaddd xmm0,xmm0,xmm0
3927
3928 vpsrld xmm12,xmm12,2
3929 vpaddd xmm10,xmm10,xmm8
3930 vpor xmm0,xmm0,xmm5
3931 vpor xmm12,xmm12,xmm7
3932 vpxor xmm1,xmm1,xmm3
3933 vmovdqa xmm3,XMMWORD[((96-128))+rax]
3934
3935 vpslld xmm8,xmm10,5
3936 vpaddd xmm14,xmm14,xmm15
3937 vpxor xmm6,xmm13,xmm11
3938 vmovdqa XMMWORD[(48-128)+rax],xmm0
3939 vpaddd xmm14,xmm14,xmm0
3940 vpxor xmm1,xmm1,XMMWORD[((192-128))+rax]
3941 vpsrld xmm9,xmm10,27
3942 vpxor xmm6,xmm6,xmm12
3943 vpxor xmm1,xmm1,xmm3
3944
3945 vpslld xmm7,xmm11,30
3946 vpor xmm8,xmm8,xmm9
3947 vpaddd xmm14,xmm14,xmm6
3948 vpsrld xmm5,xmm1,31
3949 vpaddd xmm1,xmm1,xmm1
3950
3951 vpsrld xmm11,xmm11,2
3952 vpaddd xmm14,xmm14,xmm8
3953 vpor xmm1,xmm1,xmm5
3954 vpor xmm11,xmm11,xmm7
3955 vpxor xmm2,xmm2,xmm4
3956 vmovdqa xmm4,XMMWORD[((112-128))+rax]
3957
3958 vpslld xmm8,xmm14,5
3959 vpaddd xmm13,xmm13,xmm15
3960 vpxor xmm6,xmm12,xmm10
3961 vmovdqa XMMWORD[(64-128)+rax],xmm1
3962 vpaddd xmm13,xmm13,xmm1
3963 vpxor xmm2,xmm2,XMMWORD[((208-128))+rax]
3964 vpsrld xmm9,xmm14,27
3965 vpxor xmm6,xmm6,xmm11
3966 vpxor xmm2,xmm2,xmm4
3967
3968 vpslld xmm7,xmm10,30
3969 vpor xmm8,xmm8,xmm9
3970 vpaddd xmm13,xmm13,xmm6
3971 vpsrld xmm5,xmm2,31
3972 vpaddd xmm2,xmm2,xmm2
3973
3974 vpsrld xmm10,xmm10,2
3975 vpaddd xmm13,xmm13,xmm8
3976 vpor xmm2,xmm2,xmm5
3977 vpor xmm10,xmm10,xmm7
3978 vpxor xmm3,xmm3,xmm0
3979 vmovdqa xmm0,XMMWORD[((128-128))+rax]
3980
3981 vpslld xmm8,xmm13,5
3982 vpaddd xmm12,xmm12,xmm15
3983 vpxor xmm6,xmm11,xmm14
3984 vmovdqa XMMWORD[(80-128)+rax],xmm2
3985 vpaddd xmm12,xmm12,xmm2
3986 vpxor xmm3,xmm3,XMMWORD[((224-128))+rax]
3987 vpsrld xmm9,xmm13,27
3988 vpxor xmm6,xmm6,xmm10
3989 vpxor xmm3,xmm3,xmm0
3990
3991 vpslld xmm7,xmm14,30
3992 vpor xmm8,xmm8,xmm9
3993 vpaddd xmm12,xmm12,xmm6
3994 vpsrld xmm5,xmm3,31
3995 vpaddd xmm3,xmm3,xmm3
3996
3997 vpsrld xmm14,xmm14,2
3998 vpaddd xmm12,xmm12,xmm8
3999 vpor xmm3,xmm3,xmm5
4000 vpor xmm14,xmm14,xmm7
4001 vpxor xmm4,xmm4,xmm1
4002 vmovdqa xmm1,XMMWORD[((144-128))+rax]
4003
4004 vpslld xmm8,xmm12,5
4005 vpaddd xmm11,xmm11,xmm15
4006 vpxor xmm6,xmm10,xmm13
4007 vmovdqa XMMWORD[(96-128)+rax],xmm3
4008 vpaddd xmm11,xmm11,xmm3
4009 vpxor xmm4,xmm4,XMMWORD[((240-128))+rax]
4010 vpsrld xmm9,xmm12,27
4011 vpxor xmm6,xmm6,xmm14
4012 vpxor xmm4,xmm4,xmm1
4013
4014 vpslld xmm7,xmm13,30
4015 vpor xmm8,xmm8,xmm9
4016 vpaddd xmm11,xmm11,xmm6
4017 vpsrld xmm5,xmm4,31
4018 vpaddd xmm4,xmm4,xmm4
4019
4020 vpsrld xmm13,xmm13,2
4021 vpaddd xmm11,xmm11,xmm8
4022 vpor xmm4,xmm4,xmm5
4023 vpor xmm13,xmm13,xmm7
4024 vpxor xmm0,xmm0,xmm2
4025 vmovdqa xmm2,XMMWORD[((160-128))+rax]
4026
4027 vpslld xmm8,xmm11,5
4028 vpaddd xmm10,xmm10,xmm15
4029 vpxor xmm6,xmm14,xmm12
4030 vmovdqa XMMWORD[(112-128)+rax],xmm4
4031 vpaddd xmm10,xmm10,xmm4
4032 vpxor xmm0,xmm0,XMMWORD[((0-128))+rax]
4033 vpsrld xmm9,xmm11,27
4034 vpxor xmm6,xmm6,xmm13
4035 vpxor xmm0,xmm0,xmm2
4036
4037 vpslld xmm7,xmm12,30
4038 vpor xmm8,xmm8,xmm9
4039 vpaddd xmm10,xmm10,xmm6
4040 vpsrld xmm5,xmm0,31
4041 vpaddd xmm0,xmm0,xmm0
4042
4043 vpsrld xmm12,xmm12,2
4044 vpaddd xmm10,xmm10,xmm8
4045 vpor xmm0,xmm0,xmm5
4046 vpor xmm12,xmm12,xmm7
4047 vmovdqa xmm15,XMMWORD[32+rbp]
4048 vpxor xmm1,xmm1,xmm3
4049 vmovdqa xmm3,XMMWORD[((176-128))+rax]
4050
4051 vpaddd xmm14,xmm14,xmm15
4052 vpslld xmm8,xmm10,5
4053 vpand xmm7,xmm13,xmm12
4054 vpxor xmm1,xmm1,XMMWORD[((16-128))+rax]
4055
4056 vpaddd xmm14,xmm14,xmm7
4057 vpsrld xmm9,xmm10,27
4058 vpxor xmm6,xmm13,xmm12
4059 vpxor xmm1,xmm1,xmm3
4060
4061 vmovdqu XMMWORD[(128-128)+rax],xmm0
4062 vpaddd xmm14,xmm14,xmm0
4063 vpor xmm8,xmm8,xmm9
4064 vpsrld xmm5,xmm1,31
4065 vpand xmm6,xmm6,xmm11
4066 vpaddd xmm1,xmm1,xmm1
4067
4068 vpslld xmm7,xmm11,30
4069 vpaddd xmm14,xmm14,xmm6
4070
4071 vpsrld xmm11,xmm11,2
4072 vpaddd xmm14,xmm14,xmm8
4073 vpor xmm1,xmm1,xmm5
4074 vpor xmm11,xmm11,xmm7
4075 vpxor xmm2,xmm2,xmm4
4076 vmovdqa xmm4,XMMWORD[((192-128))+rax]
4077
4078 vpaddd xmm13,xmm13,xmm15
4079 vpslld xmm8,xmm14,5
4080 vpand xmm7,xmm12,xmm11
4081 vpxor xmm2,xmm2,XMMWORD[((32-128))+rax]
4082
4083 vpaddd xmm13,xmm13,xmm7
4084 vpsrld xmm9,xmm14,27
4085 vpxor xmm6,xmm12,xmm11
4086 vpxor xmm2,xmm2,xmm4
4087
4088 vmovdqu XMMWORD[(144-128)+rax],xmm1
4089 vpaddd xmm13,xmm13,xmm1
4090 vpor xmm8,xmm8,xmm9
4091 vpsrld xmm5,xmm2,31
4092 vpand xmm6,xmm6,xmm10
4093 vpaddd xmm2,xmm2,xmm2
4094
4095 vpslld xmm7,xmm10,30
4096 vpaddd xmm13,xmm13,xmm6
4097
4098 vpsrld xmm10,xmm10,2
4099 vpaddd xmm13,xmm13,xmm8
4100 vpor xmm2,xmm2,xmm5
4101 vpor xmm10,xmm10,xmm7
4102 vpxor xmm3,xmm3,xmm0
4103 vmovdqa xmm0,XMMWORD[((208-128))+rax]
4104
4105 vpaddd xmm12,xmm12,xmm15
4106 vpslld xmm8,xmm13,5
4107 vpand xmm7,xmm11,xmm10
4108 vpxor xmm3,xmm3,XMMWORD[((48-128))+rax]
4109
4110 vpaddd xmm12,xmm12,xmm7
4111 vpsrld xmm9,xmm13,27
4112 vpxor xmm6,xmm11,xmm10
4113 vpxor xmm3,xmm3,xmm0
4114
4115 vmovdqu XMMWORD[(160-128)+rax],xmm2
4116 vpaddd xmm12,xmm12,xmm2
4117 vpor xmm8,xmm8,xmm9
4118 vpsrld xmm5,xmm3,31
4119 vpand xmm6,xmm6,xmm14
4120 vpaddd xmm3,xmm3,xmm3
4121
4122 vpslld xmm7,xmm14,30
4123 vpaddd xmm12,xmm12,xmm6
4124
4125 vpsrld xmm14,xmm14,2
4126 vpaddd xmm12,xmm12,xmm8
4127 vpor xmm3,xmm3,xmm5
4128 vpor xmm14,xmm14,xmm7
4129 vpxor xmm4,xmm4,xmm1
4130 vmovdqa xmm1,XMMWORD[((224-128))+rax]
4131
4132 vpaddd xmm11,xmm11,xmm15
4133 vpslld xmm8,xmm12,5
4134 vpand xmm7,xmm10,xmm14
4135 vpxor xmm4,xmm4,XMMWORD[((64-128))+rax]
4136
4137 vpaddd xmm11,xmm11,xmm7
4138 vpsrld xmm9,xmm12,27
4139 vpxor xmm6,xmm10,xmm14
4140 vpxor xmm4,xmm4,xmm1
4141
4142 vmovdqu XMMWORD[(176-128)+rax],xmm3
4143 vpaddd xmm11,xmm11,xmm3
4144 vpor xmm8,xmm8,xmm9
4145 vpsrld xmm5,xmm4,31
4146 vpand xmm6,xmm6,xmm13
4147 vpaddd xmm4,xmm4,xmm4
4148
4149 vpslld xmm7,xmm13,30
4150 vpaddd xmm11,xmm11,xmm6
4151
4152 vpsrld xmm13,xmm13,2
4153 vpaddd xmm11,xmm11,xmm8
4154 vpor xmm4,xmm4,xmm5
4155 vpor xmm13,xmm13,xmm7
4156 vpxor xmm0,xmm0,xmm2
4157 vmovdqa xmm2,XMMWORD[((240-128))+rax]
4158
4159 vpaddd xmm10,xmm10,xmm15
4160 vpslld xmm8,xmm11,5
4161 vpand xmm7,xmm14,xmm13
4162 vpxor xmm0,xmm0,XMMWORD[((80-128))+rax]
4163
4164 vpaddd xmm10,xmm10,xmm7
4165 vpsrld xmm9,xmm11,27
4166 vpxor xmm6,xmm14,xmm13
4167 vpxor xmm0,xmm0,xmm2
4168
4169 vmovdqu XMMWORD[(192-128)+rax],xmm4
4170 vpaddd xmm10,xmm10,xmm4
4171 vpor xmm8,xmm8,xmm9
4172 vpsrld xmm5,xmm0,31
4173 vpand xmm6,xmm6,xmm12
4174 vpaddd xmm0,xmm0,xmm0
4175
4176 vpslld xmm7,xmm12,30
4177 vpaddd xmm10,xmm10,xmm6
4178
4179 vpsrld xmm12,xmm12,2
4180 vpaddd xmm10,xmm10,xmm8
4181 vpor xmm0,xmm0,xmm5
4182 vpor xmm12,xmm12,xmm7
4183 vpxor xmm1,xmm1,xmm3
4184 vmovdqa xmm3,XMMWORD[((0-128))+rax]
4185
4186 vpaddd xmm14,xmm14,xmm15
4187 vpslld xmm8,xmm10,5
4188 vpand xmm7,xmm13,xmm12
4189 vpxor xmm1,xmm1,XMMWORD[((96-128))+rax]
4190
4191 vpaddd xmm14,xmm14,xmm7
4192 vpsrld xmm9,xmm10,27
4193 vpxor xmm6,xmm13,xmm12
4194 vpxor xmm1,xmm1,xmm3
4195
4196 vmovdqu XMMWORD[(208-128)+rax],xmm0
4197 vpaddd xmm14,xmm14,xmm0
4198 vpor xmm8,xmm8,xmm9
4199 vpsrld xmm5,xmm1,31
4200 vpand xmm6,xmm6,xmm11
4201 vpaddd xmm1,xmm1,xmm1
4202
4203 vpslld xmm7,xmm11,30
4204 vpaddd xmm14,xmm14,xmm6
4205
4206 vpsrld xmm11,xmm11,2
4207 vpaddd xmm14,xmm14,xmm8
4208 vpor xmm1,xmm1,xmm5
4209 vpor xmm11,xmm11,xmm7
4210 vpxor xmm2,xmm2,xmm4
4211 vmovdqa xmm4,XMMWORD[((16-128))+rax]
4212
4213 vpaddd xmm13,xmm13,xmm15
4214 vpslld xmm8,xmm14,5
4215 vpand xmm7,xmm12,xmm11
4216 vpxor xmm2,xmm2,XMMWORD[((112-128))+rax]
4217
4218 vpaddd xmm13,xmm13,xmm7
4219 vpsrld xmm9,xmm14,27
4220 vpxor xmm6,xmm12,xmm11
4221 vpxor xmm2,xmm2,xmm4
4222
4223 vmovdqu XMMWORD[(224-128)+rax],xmm1
4224 vpaddd xmm13,xmm13,xmm1
4225 vpor xmm8,xmm8,xmm9
4226 vpsrld xmm5,xmm2,31
4227 vpand xmm6,xmm6,xmm10
4228 vpaddd xmm2,xmm2,xmm2
4229
4230 vpslld xmm7,xmm10,30
4231 vpaddd xmm13,xmm13,xmm6
4232
4233 vpsrld xmm10,xmm10,2
4234 vpaddd xmm13,xmm13,xmm8
4235 vpor xmm2,xmm2,xmm5
4236 vpor xmm10,xmm10,xmm7
4237 vpxor xmm3,xmm3,xmm0
4238 vmovdqa xmm0,XMMWORD[((32-128))+rax]
4239
4240 vpaddd xmm12,xmm12,xmm15
4241 vpslld xmm8,xmm13,5
4242 vpand xmm7,xmm11,xmm10
4243 vpxor xmm3,xmm3,XMMWORD[((128-128))+rax]
4244
4245 vpaddd xmm12,xmm12,xmm7
4246 vpsrld xmm9,xmm13,27
4247 vpxor xmm6,xmm11,xmm10
4248 vpxor xmm3,xmm3,xmm0
4249
4250 vmovdqu XMMWORD[(240-128)+rax],xmm2
4251 vpaddd xmm12,xmm12,xmm2
4252 vpor xmm8,xmm8,xmm9
4253 vpsrld xmm5,xmm3,31
4254 vpand xmm6,xmm6,xmm14
4255 vpaddd xmm3,xmm3,xmm3
4256
4257 vpslld xmm7,xmm14,30
4258 vpaddd xmm12,xmm12,xmm6
4259
4260 vpsrld xmm14,xmm14,2
4261 vpaddd xmm12,xmm12,xmm8
4262 vpor xmm3,xmm3,xmm5
4263 vpor xmm14,xmm14,xmm7
4264 vpxor xmm4,xmm4,xmm1
4265 vmovdqa xmm1,XMMWORD[((48-128))+rax]
4266
4267 vpaddd xmm11,xmm11,xmm15
4268 vpslld xmm8,xmm12,5
4269 vpand xmm7,xmm10,xmm14
4270 vpxor xmm4,xmm4,XMMWORD[((144-128))+rax]
4271
4272 vpaddd xmm11,xmm11,xmm7
4273 vpsrld xmm9,xmm12,27
4274 vpxor xmm6,xmm10,xmm14
4275 vpxor xmm4,xmm4,xmm1
4276
4277 vmovdqu XMMWORD[(0-128)+rax],xmm3
4278 vpaddd xmm11,xmm11,xmm3
4279 vpor xmm8,xmm8,xmm9
4280 vpsrld xmm5,xmm4,31
4281 vpand xmm6,xmm6,xmm13
4282 vpaddd xmm4,xmm4,xmm4
4283
4284 vpslld xmm7,xmm13,30
4285 vpaddd xmm11,xmm11,xmm6
4286
4287 vpsrld xmm13,xmm13,2
4288 vpaddd xmm11,xmm11,xmm8
4289 vpor xmm4,xmm4,xmm5
4290 vpor xmm13,xmm13,xmm7
4291 vpxor xmm0,xmm0,xmm2
4292 vmovdqa xmm2,XMMWORD[((64-128))+rax]
4293
4294 vpaddd xmm10,xmm10,xmm15
4295 vpslld xmm8,xmm11,5
4296 vpand xmm7,xmm14,xmm13
4297 vpxor xmm0,xmm0,XMMWORD[((160-128))+rax]
4298
4299 vpaddd xmm10,xmm10,xmm7
4300 vpsrld xmm9,xmm11,27
4301 vpxor xmm6,xmm14,xmm13
4302 vpxor xmm0,xmm0,xmm2
4303
4304 vmovdqu XMMWORD[(16-128)+rax],xmm4
4305 vpaddd xmm10,xmm10,xmm4
4306 vpor xmm8,xmm8,xmm9
4307 vpsrld xmm5,xmm0,31
4308 vpand xmm6,xmm6,xmm12
4309 vpaddd xmm0,xmm0,xmm0
4310
4311 vpslld xmm7,xmm12,30
4312 vpaddd xmm10,xmm10,xmm6
4313
4314 vpsrld xmm12,xmm12,2
4315 vpaddd xmm10,xmm10,xmm8
4316 vpor xmm0,xmm0,xmm5
4317 vpor xmm12,xmm12,xmm7
4318 vpxor xmm1,xmm1,xmm3
4319 vmovdqa xmm3,XMMWORD[((80-128))+rax]
4320
4321 vpaddd xmm14,xmm14,xmm15
4322 vpslld xmm8,xmm10,5
4323 vpand xmm7,xmm13,xmm12
4324 vpxor xmm1,xmm1,XMMWORD[((176-128))+rax]
4325
4326 vpaddd xmm14,xmm14,xmm7
4327 vpsrld xmm9,xmm10,27
4328 vpxor xmm6,xmm13,xmm12
4329 vpxor xmm1,xmm1,xmm3
4330
4331 vmovdqu XMMWORD[(32-128)+rax],xmm0
4332 vpaddd xmm14,xmm14,xmm0
4333 vpor xmm8,xmm8,xmm9
4334 vpsrld xmm5,xmm1,31
4335 vpand xmm6,xmm6,xmm11
4336 vpaddd xmm1,xmm1,xmm1
4337
4338 vpslld xmm7,xmm11,30
4339 vpaddd xmm14,xmm14,xmm6
4340
4341 vpsrld xmm11,xmm11,2
4342 vpaddd xmm14,xmm14,xmm8
4343 vpor xmm1,xmm1,xmm5
4344 vpor xmm11,xmm11,xmm7
4345 vpxor xmm2,xmm2,xmm4
4346 vmovdqa xmm4,XMMWORD[((96-128))+rax]
4347
4348 vpaddd xmm13,xmm13,xmm15
4349 vpslld xmm8,xmm14,5
4350 vpand xmm7,xmm12,xmm11
4351 vpxor xmm2,xmm2,XMMWORD[((192-128))+rax]
4352
4353 vpaddd xmm13,xmm13,xmm7
4354 vpsrld xmm9,xmm14,27
4355 vpxor xmm6,xmm12,xmm11
4356 vpxor xmm2,xmm2,xmm4
4357
4358 vmovdqu XMMWORD[(48-128)+rax],xmm1
4359 vpaddd xmm13,xmm13,xmm1
4360 vpor xmm8,xmm8,xmm9
4361 vpsrld xmm5,xmm2,31
4362 vpand xmm6,xmm6,xmm10
4363 vpaddd xmm2,xmm2,xmm2
4364
4365 vpslld xmm7,xmm10,30
4366 vpaddd xmm13,xmm13,xmm6
4367
4368 vpsrld xmm10,xmm10,2
4369 vpaddd xmm13,xmm13,xmm8
4370 vpor xmm2,xmm2,xmm5
4371 vpor xmm10,xmm10,xmm7
4372 vpxor xmm3,xmm3,xmm0
4373 vmovdqa xmm0,XMMWORD[((112-128))+rax]
4374
4375 vpaddd xmm12,xmm12,xmm15
4376 vpslld xmm8,xmm13,5
4377 vpand xmm7,xmm11,xmm10
4378 vpxor xmm3,xmm3,XMMWORD[((208-128))+rax]
4379
4380 vpaddd xmm12,xmm12,xmm7
4381 vpsrld xmm9,xmm13,27
4382 vpxor xmm6,xmm11,xmm10
4383 vpxor xmm3,xmm3,xmm0
4384
4385 vmovdqu XMMWORD[(64-128)+rax],xmm2
4386 vpaddd xmm12,xmm12,xmm2
4387 vpor xmm8,xmm8,xmm9
4388 vpsrld xmm5,xmm3,31
4389 vpand xmm6,xmm6,xmm14
4390 vpaddd xmm3,xmm3,xmm3
4391
4392 vpslld xmm7,xmm14,30
4393 vpaddd xmm12,xmm12,xmm6
4394
4395 vpsrld xmm14,xmm14,2
4396 vpaddd xmm12,xmm12,xmm8
4397 vpor xmm3,xmm3,xmm5
4398 vpor xmm14,xmm14,xmm7
4399 vpxor xmm4,xmm4,xmm1
4400 vmovdqa xmm1,XMMWORD[((128-128))+rax]
4401
4402 vpaddd xmm11,xmm11,xmm15
4403 vpslld xmm8,xmm12,5
4404 vpand xmm7,xmm10,xmm14
4405 vpxor xmm4,xmm4,XMMWORD[((224-128))+rax]
4406
4407 vpaddd xmm11,xmm11,xmm7
4408 vpsrld xmm9,xmm12,27
4409 vpxor xmm6,xmm10,xmm14
4410 vpxor xmm4,xmm4,xmm1
4411
4412 vmovdqu XMMWORD[(80-128)+rax],xmm3
4413 vpaddd xmm11,xmm11,xmm3
4414 vpor xmm8,xmm8,xmm9
4415 vpsrld xmm5,xmm4,31
4416 vpand xmm6,xmm6,xmm13
4417 vpaddd xmm4,xmm4,xmm4
4418
4419 vpslld xmm7,xmm13,30
4420 vpaddd xmm11,xmm11,xmm6
4421
4422 vpsrld xmm13,xmm13,2
4423 vpaddd xmm11,xmm11,xmm8
4424 vpor xmm4,xmm4,xmm5
4425 vpor xmm13,xmm13,xmm7
4426 vpxor xmm0,xmm0,xmm2
4427 vmovdqa xmm2,XMMWORD[((144-128))+rax]
4428
4429 vpaddd xmm10,xmm10,xmm15
4430 vpslld xmm8,xmm11,5
4431 vpand xmm7,xmm14,xmm13
4432 vpxor xmm0,xmm0,XMMWORD[((240-128))+rax]
4433
4434 vpaddd xmm10,xmm10,xmm7
4435 vpsrld xmm9,xmm11,27
4436 vpxor xmm6,xmm14,xmm13
4437 vpxor xmm0,xmm0,xmm2
4438
4439 vmovdqu XMMWORD[(96-128)+rax],xmm4
4440 vpaddd xmm10,xmm10,xmm4
4441 vpor xmm8,xmm8,xmm9
4442 vpsrld xmm5,xmm0,31
4443 vpand xmm6,xmm6,xmm12
4444 vpaddd xmm0,xmm0,xmm0
4445
4446 vpslld xmm7,xmm12,30
4447 vpaddd xmm10,xmm10,xmm6
4448
4449 vpsrld xmm12,xmm12,2
4450 vpaddd xmm10,xmm10,xmm8
4451 vpor xmm0,xmm0,xmm5
4452 vpor xmm12,xmm12,xmm7
4453 vpxor xmm1,xmm1,xmm3
4454 vmovdqa xmm3,XMMWORD[((160-128))+rax]
4455
4456 vpaddd xmm14,xmm14,xmm15
4457 vpslld xmm8,xmm10,5
4458 vpand xmm7,xmm13,xmm12
4459 vpxor xmm1,xmm1,XMMWORD[((0-128))+rax]
4460
4461 vpaddd xmm14,xmm14,xmm7
4462 vpsrld xmm9,xmm10,27
4463 vpxor xmm6,xmm13,xmm12
4464 vpxor xmm1,xmm1,xmm3
4465
4466 vmovdqu XMMWORD[(112-128)+rax],xmm0
4467 vpaddd xmm14,xmm14,xmm0
4468 vpor xmm8,xmm8,xmm9
4469 vpsrld xmm5,xmm1,31
4470 vpand xmm6,xmm6,xmm11
4471 vpaddd xmm1,xmm1,xmm1
4472
4473 vpslld xmm7,xmm11,30
4474 vpaddd xmm14,xmm14,xmm6
4475
4476 vpsrld xmm11,xmm11,2
4477 vpaddd xmm14,xmm14,xmm8
4478 vpor xmm1,xmm1,xmm5
4479 vpor xmm11,xmm11,xmm7
4480 vpxor xmm2,xmm2,xmm4
4481 vmovdqa xmm4,XMMWORD[((176-128))+rax]
4482
4483 vpaddd xmm13,xmm13,xmm15
4484 vpslld xmm8,xmm14,5
4485 vpand xmm7,xmm12,xmm11
4486 vpxor xmm2,xmm2,XMMWORD[((16-128))+rax]
4487
4488 vpaddd xmm13,xmm13,xmm7
4489 vpsrld xmm9,xmm14,27
4490 vpxor xmm6,xmm12,xmm11
4491 vpxor xmm2,xmm2,xmm4
4492
4493 vmovdqu XMMWORD[(128-128)+rax],xmm1
4494 vpaddd xmm13,xmm13,xmm1
4495 vpor xmm8,xmm8,xmm9
4496 vpsrld xmm5,xmm2,31
4497 vpand xmm6,xmm6,xmm10
4498 vpaddd xmm2,xmm2,xmm2
4499
4500 vpslld xmm7,xmm10,30
4501 vpaddd xmm13,xmm13,xmm6
4502
4503 vpsrld xmm10,xmm10,2
4504 vpaddd xmm13,xmm13,xmm8
4505 vpor xmm2,xmm2,xmm5
4506 vpor xmm10,xmm10,xmm7
4507 vpxor xmm3,xmm3,xmm0
4508 vmovdqa xmm0,XMMWORD[((192-128))+rax]
4509
4510 vpaddd xmm12,xmm12,xmm15
4511 vpslld xmm8,xmm13,5
4512 vpand xmm7,xmm11,xmm10
4513 vpxor xmm3,xmm3,XMMWORD[((32-128))+rax]
4514
4515 vpaddd xmm12,xmm12,xmm7
4516 vpsrld xmm9,xmm13,27
4517 vpxor xmm6,xmm11,xmm10
4518 vpxor xmm3,xmm3,xmm0
4519
4520 vmovdqu XMMWORD[(144-128)+rax],xmm2
4521 vpaddd xmm12,xmm12,xmm2
4522 vpor xmm8,xmm8,xmm9
4523 vpsrld xmm5,xmm3,31
4524 vpand xmm6,xmm6,xmm14
4525 vpaddd xmm3,xmm3,xmm3
4526
4527 vpslld xmm7,xmm14,30
4528 vpaddd xmm12,xmm12,xmm6
4529
4530 vpsrld xmm14,xmm14,2
4531 vpaddd xmm12,xmm12,xmm8
4532 vpor xmm3,xmm3,xmm5
4533 vpor xmm14,xmm14,xmm7
4534 vpxor xmm4,xmm4,xmm1
4535 vmovdqa xmm1,XMMWORD[((208-128))+rax]
4536
4537 vpaddd xmm11,xmm11,xmm15
4538 vpslld xmm8,xmm12,5
4539 vpand xmm7,xmm10,xmm14
4540 vpxor xmm4,xmm4,XMMWORD[((48-128))+rax]
4541
4542 vpaddd xmm11,xmm11,xmm7
4543 vpsrld xmm9,xmm12,27
4544 vpxor xmm6,xmm10,xmm14
4545 vpxor xmm4,xmm4,xmm1
4546
4547 vmovdqu XMMWORD[(160-128)+rax],xmm3
4548 vpaddd xmm11,xmm11,xmm3
4549 vpor xmm8,xmm8,xmm9
4550 vpsrld xmm5,xmm4,31
4551 vpand xmm6,xmm6,xmm13
4552 vpaddd xmm4,xmm4,xmm4
4553
4554 vpslld xmm7,xmm13,30
4555 vpaddd xmm11,xmm11,xmm6
4556
4557 vpsrld xmm13,xmm13,2
4558 vpaddd xmm11,xmm11,xmm8
4559 vpor xmm4,xmm4,xmm5
4560 vpor xmm13,xmm13,xmm7
4561 vpxor xmm0,xmm0,xmm2
4562 vmovdqa xmm2,XMMWORD[((224-128))+rax]
4563
4564 vpaddd xmm10,xmm10,xmm15
4565 vpslld xmm8,xmm11,5
4566 vpand xmm7,xmm14,xmm13
4567 vpxor xmm0,xmm0,XMMWORD[((64-128))+rax]
4568
4569 vpaddd xmm10,xmm10,xmm7
4570 vpsrld xmm9,xmm11,27
4571 vpxor xmm6,xmm14,xmm13
4572 vpxor xmm0,xmm0,xmm2
4573
4574 vmovdqu XMMWORD[(176-128)+rax],xmm4
4575 vpaddd xmm10,xmm10,xmm4
4576 vpor xmm8,xmm8,xmm9
4577 vpsrld xmm5,xmm0,31
4578 vpand xmm6,xmm6,xmm12
4579 vpaddd xmm0,xmm0,xmm0
4580
4581 vpslld xmm7,xmm12,30
4582 vpaddd xmm10,xmm10,xmm6
4583
4584 vpsrld xmm12,xmm12,2
4585 vpaddd xmm10,xmm10,xmm8
4586 vpor xmm0,xmm0,xmm5
4587 vpor xmm12,xmm12,xmm7
4588 vmovdqa xmm15,XMMWORD[64+rbp]
4589 vpxor xmm1,xmm1,xmm3
4590 vmovdqa xmm3,XMMWORD[((240-128))+rax]
4591
4592 vpslld xmm8,xmm10,5
4593 vpaddd xmm14,xmm14,xmm15
4594 vpxor xmm6,xmm13,xmm11
4595 vmovdqa XMMWORD[(192-128)+rax],xmm0
4596 vpaddd xmm14,xmm14,xmm0
4597 vpxor xmm1,xmm1,XMMWORD[((80-128))+rax]
4598 vpsrld xmm9,xmm10,27
4599 vpxor xmm6,xmm6,xmm12
4600 vpxor xmm1,xmm1,xmm3
4601
4602 vpslld xmm7,xmm11,30
4603 vpor xmm8,xmm8,xmm9
4604 vpaddd xmm14,xmm14,xmm6
4605 vpsrld xmm5,xmm1,31
4606 vpaddd xmm1,xmm1,xmm1
4607
4608 vpsrld xmm11,xmm11,2
4609 vpaddd xmm14,xmm14,xmm8
4610 vpor xmm1,xmm1,xmm5
4611 vpor xmm11,xmm11,xmm7
4612 vpxor xmm2,xmm2,xmm4
4613 vmovdqa xmm4,XMMWORD[((0-128))+rax]
4614
4615 vpslld xmm8,xmm14,5
4616 vpaddd xmm13,xmm13,xmm15
4617 vpxor xmm6,xmm12,xmm10
4618 vmovdqa XMMWORD[(208-128)+rax],xmm1
4619 vpaddd xmm13,xmm13,xmm1
4620 vpxor xmm2,xmm2,XMMWORD[((96-128))+rax]
4621 vpsrld xmm9,xmm14,27
4622 vpxor xmm6,xmm6,xmm11
4623 vpxor xmm2,xmm2,xmm4
4624
4625 vpslld xmm7,xmm10,30
4626 vpor xmm8,xmm8,xmm9
4627 vpaddd xmm13,xmm13,xmm6
4628 vpsrld xmm5,xmm2,31
4629 vpaddd xmm2,xmm2,xmm2
4630
4631 vpsrld xmm10,xmm10,2
4632 vpaddd xmm13,xmm13,xmm8
4633 vpor xmm2,xmm2,xmm5
4634 vpor xmm10,xmm10,xmm7
4635 vpxor xmm3,xmm3,xmm0
4636 vmovdqa xmm0,XMMWORD[((16-128))+rax]
4637
4638 vpslld xmm8,xmm13,5
4639 vpaddd xmm12,xmm12,xmm15
4640 vpxor xmm6,xmm11,xmm14
4641 vmovdqa XMMWORD[(224-128)+rax],xmm2
4642 vpaddd xmm12,xmm12,xmm2
4643 vpxor xmm3,xmm3,XMMWORD[((112-128))+rax]
4644 vpsrld xmm9,xmm13,27
4645 vpxor xmm6,xmm6,xmm10
4646 vpxor xmm3,xmm3,xmm0
4647
4648 vpslld xmm7,xmm14,30
4649 vpor xmm8,xmm8,xmm9
4650 vpaddd xmm12,xmm12,xmm6
4651 vpsrld xmm5,xmm3,31
4652 vpaddd xmm3,xmm3,xmm3
4653
4654 vpsrld xmm14,xmm14,2
4655 vpaddd xmm12,xmm12,xmm8
4656 vpor xmm3,xmm3,xmm5
4657 vpor xmm14,xmm14,xmm7
4658 vpxor xmm4,xmm4,xmm1
4659 vmovdqa xmm1,XMMWORD[((32-128))+rax]
4660
4661 vpslld xmm8,xmm12,5
4662 vpaddd xmm11,xmm11,xmm15
4663 vpxor xmm6,xmm10,xmm13
4664 vmovdqa XMMWORD[(240-128)+rax],xmm3
4665 vpaddd xmm11,xmm11,xmm3
4666 vpxor xmm4,xmm4,XMMWORD[((128-128))+rax]
4667 vpsrld xmm9,xmm12,27
4668 vpxor xmm6,xmm6,xmm14
4669 vpxor xmm4,xmm4,xmm1
4670
4671 vpslld xmm7,xmm13,30
4672 vpor xmm8,xmm8,xmm9
4673 vpaddd xmm11,xmm11,xmm6
4674 vpsrld xmm5,xmm4,31
4675 vpaddd xmm4,xmm4,xmm4
4676
4677 vpsrld xmm13,xmm13,2
4678 vpaddd xmm11,xmm11,xmm8
4679 vpor xmm4,xmm4,xmm5
4680 vpor xmm13,xmm13,xmm7
4681 vpxor xmm0,xmm0,xmm2
4682 vmovdqa xmm2,XMMWORD[((48-128))+rax]
4683
4684 vpslld xmm8,xmm11,5
4685 vpaddd xmm10,xmm10,xmm15
4686 vpxor xmm6,xmm14,xmm12
4687 vmovdqa XMMWORD[(0-128)+rax],xmm4
4688 vpaddd xmm10,xmm10,xmm4
4689 vpxor xmm0,xmm0,XMMWORD[((144-128))+rax]
4690 vpsrld xmm9,xmm11,27
4691 vpxor xmm6,xmm6,xmm13
4692 vpxor xmm0,xmm0,xmm2
4693
4694 vpslld xmm7,xmm12,30
4695 vpor xmm8,xmm8,xmm9
4696 vpaddd xmm10,xmm10,xmm6
4697 vpsrld xmm5,xmm0,31
4698 vpaddd xmm0,xmm0,xmm0
4699
4700 vpsrld xmm12,xmm12,2
4701 vpaddd xmm10,xmm10,xmm8
4702 vpor xmm0,xmm0,xmm5
4703 vpor xmm12,xmm12,xmm7
4704 vpxor xmm1,xmm1,xmm3
4705 vmovdqa xmm3,XMMWORD[((64-128))+rax]
4706
4707 vpslld xmm8,xmm10,5
4708 vpaddd xmm14,xmm14,xmm15
4709 vpxor xmm6,xmm13,xmm11
4710 vmovdqa XMMWORD[(16-128)+rax],xmm0
4711 vpaddd xmm14,xmm14,xmm0
4712 vpxor xmm1,xmm1,XMMWORD[((160-128))+rax]
4713 vpsrld xmm9,xmm10,27
4714 vpxor xmm6,xmm6,xmm12
4715 vpxor xmm1,xmm1,xmm3
4716
4717 vpslld xmm7,xmm11,30
4718 vpor xmm8,xmm8,xmm9
4719 vpaddd xmm14,xmm14,xmm6
4720 vpsrld xmm5,xmm1,31
4721 vpaddd xmm1,xmm1,xmm1
4722
4723 vpsrld xmm11,xmm11,2
4724 vpaddd xmm14,xmm14,xmm8
4725 vpor xmm1,xmm1,xmm5
4726 vpor xmm11,xmm11,xmm7
4727 vpxor xmm2,xmm2,xmm4
4728 vmovdqa xmm4,XMMWORD[((80-128))+rax]
4729
4730 vpslld xmm8,xmm14,5
4731 vpaddd xmm13,xmm13,xmm15
4732 vpxor xmm6,xmm12,xmm10
4733 vmovdqa XMMWORD[(32-128)+rax],xmm1
4734 vpaddd xmm13,xmm13,xmm1
4735 vpxor xmm2,xmm2,XMMWORD[((176-128))+rax]
4736 vpsrld xmm9,xmm14,27
4737 vpxor xmm6,xmm6,xmm11
4738 vpxor xmm2,xmm2,xmm4
4739
4740 vpslld xmm7,xmm10,30
4741 vpor xmm8,xmm8,xmm9
4742 vpaddd xmm13,xmm13,xmm6
4743 vpsrld xmm5,xmm2,31
4744 vpaddd xmm2,xmm2,xmm2
4745
4746 vpsrld xmm10,xmm10,2
4747 vpaddd xmm13,xmm13,xmm8
4748 vpor xmm2,xmm2,xmm5
4749 vpor xmm10,xmm10,xmm7
4750 vpxor xmm3,xmm3,xmm0
4751 vmovdqa xmm0,XMMWORD[((96-128))+rax]
4752
4753 vpslld xmm8,xmm13,5
4754 vpaddd xmm12,xmm12,xmm15
4755 vpxor xmm6,xmm11,xmm14
4756 vmovdqa XMMWORD[(48-128)+rax],xmm2
4757 vpaddd xmm12,xmm12,xmm2
4758 vpxor xmm3,xmm3,XMMWORD[((192-128))+rax]
4759 vpsrld xmm9,xmm13,27
4760 vpxor xmm6,xmm6,xmm10
4761 vpxor xmm3,xmm3,xmm0
4762
4763 vpslld xmm7,xmm14,30
4764 vpor xmm8,xmm8,xmm9
4765 vpaddd xmm12,xmm12,xmm6
4766 vpsrld xmm5,xmm3,31
4767 vpaddd xmm3,xmm3,xmm3
4768
4769 vpsrld xmm14,xmm14,2
4770 vpaddd xmm12,xmm12,xmm8
4771 vpor xmm3,xmm3,xmm5
4772 vpor xmm14,xmm14,xmm7
4773 vpxor xmm4,xmm4,xmm1
4774 vmovdqa xmm1,XMMWORD[((112-128))+rax]
4775
4776 vpslld xmm8,xmm12,5
4777 vpaddd xmm11,xmm11,xmm15
4778 vpxor xmm6,xmm10,xmm13
4779 vmovdqa XMMWORD[(64-128)+rax],xmm3
4780 vpaddd xmm11,xmm11,xmm3
4781 vpxor xmm4,xmm4,XMMWORD[((208-128))+rax]
4782 vpsrld xmm9,xmm12,27
4783 vpxor xmm6,xmm6,xmm14
4784 vpxor xmm4,xmm4,xmm1
4785
4786 vpslld xmm7,xmm13,30
4787 vpor xmm8,xmm8,xmm9
4788 vpaddd xmm11,xmm11,xmm6
4789 vpsrld xmm5,xmm4,31
4790 vpaddd xmm4,xmm4,xmm4
4791
4792 vpsrld xmm13,xmm13,2
4793 vpaddd xmm11,xmm11,xmm8
4794 vpor xmm4,xmm4,xmm5
4795 vpor xmm13,xmm13,xmm7
4796 vpxor xmm0,xmm0,xmm2
4797 vmovdqa xmm2,XMMWORD[((128-128))+rax]
4798
4799 vpslld xmm8,xmm11,5
4800 vpaddd xmm10,xmm10,xmm15
4801 vpxor xmm6,xmm14,xmm12
4802 vmovdqa XMMWORD[(80-128)+rax],xmm4
4803 vpaddd xmm10,xmm10,xmm4
4804 vpxor xmm0,xmm0,XMMWORD[((224-128))+rax]
4805 vpsrld xmm9,xmm11,27
4806 vpxor xmm6,xmm6,xmm13
4807 vpxor xmm0,xmm0,xmm2
4808
4809 vpslld xmm7,xmm12,30
4810 vpor xmm8,xmm8,xmm9
4811 vpaddd xmm10,xmm10,xmm6
4812 vpsrld xmm5,xmm0,31
4813 vpaddd xmm0,xmm0,xmm0
4814
4815 vpsrld xmm12,xmm12,2
4816 vpaddd xmm10,xmm10,xmm8
4817 vpor xmm0,xmm0,xmm5
4818 vpor xmm12,xmm12,xmm7
4819 vpxor xmm1,xmm1,xmm3
4820 vmovdqa xmm3,XMMWORD[((144-128))+rax]
4821
4822 vpslld xmm8,xmm10,5
4823 vpaddd xmm14,xmm14,xmm15
4824 vpxor xmm6,xmm13,xmm11
4825 vmovdqa XMMWORD[(96-128)+rax],xmm0
4826 vpaddd xmm14,xmm14,xmm0
4827 vpxor xmm1,xmm1,XMMWORD[((240-128))+rax]
4828 vpsrld xmm9,xmm10,27
4829 vpxor xmm6,xmm6,xmm12
4830 vpxor xmm1,xmm1,xmm3
4831
4832 vpslld xmm7,xmm11,30
4833 vpor xmm8,xmm8,xmm9
4834 vpaddd xmm14,xmm14,xmm6
4835 vpsrld xmm5,xmm1,31
4836 vpaddd xmm1,xmm1,xmm1
4837
4838 vpsrld xmm11,xmm11,2
4839 vpaddd xmm14,xmm14,xmm8
4840 vpor xmm1,xmm1,xmm5
4841 vpor xmm11,xmm11,xmm7
4842 vpxor xmm2,xmm2,xmm4
4843 vmovdqa xmm4,XMMWORD[((160-128))+rax]
4844
4845 vpslld xmm8,xmm14,5
4846 vpaddd xmm13,xmm13,xmm15
4847 vpxor xmm6,xmm12,xmm10
4848 vmovdqa XMMWORD[(112-128)+rax],xmm1
4849 vpaddd xmm13,xmm13,xmm1
4850 vpxor xmm2,xmm2,XMMWORD[((0-128))+rax]
4851 vpsrld xmm9,xmm14,27
4852 vpxor xmm6,xmm6,xmm11
4853 vpxor xmm2,xmm2,xmm4
4854
4855 vpslld xmm7,xmm10,30
4856 vpor xmm8,xmm8,xmm9
4857 vpaddd xmm13,xmm13,xmm6
4858 vpsrld xmm5,xmm2,31
4859 vpaddd xmm2,xmm2,xmm2
4860
4861 vpsrld xmm10,xmm10,2
4862 vpaddd xmm13,xmm13,xmm8
4863 vpor xmm2,xmm2,xmm5
4864 vpor xmm10,xmm10,xmm7
4865 vpxor xmm3,xmm3,xmm0
4866 vmovdqa xmm0,XMMWORD[((176-128))+rax]
4867
4868 vpslld xmm8,xmm13,5
4869 vpaddd xmm12,xmm12,xmm15
4870 vpxor xmm6,xmm11,xmm14
4871 vpaddd xmm12,xmm12,xmm2
4872 vpxor xmm3,xmm3,XMMWORD[((16-128))+rax]
4873 vpsrld xmm9,xmm13,27
4874 vpxor xmm6,xmm6,xmm10
4875 vpxor xmm3,xmm3,xmm0
4876
4877 vpslld xmm7,xmm14,30
4878 vpor xmm8,xmm8,xmm9
4879 vpaddd xmm12,xmm12,xmm6
4880 vpsrld xmm5,xmm3,31
4881 vpaddd xmm3,xmm3,xmm3
4882
4883 vpsrld xmm14,xmm14,2
4884 vpaddd xmm12,xmm12,xmm8
4885 vpor xmm3,xmm3,xmm5
4886 vpor xmm14,xmm14,xmm7
4887 vpxor xmm4,xmm4,xmm1
4888 vmovdqa xmm1,XMMWORD[((192-128))+rax]
4889
4890 vpslld xmm8,xmm12,5
4891 vpaddd xmm11,xmm11,xmm15
4892 vpxor xmm6,xmm10,xmm13
4893 vpaddd xmm11,xmm11,xmm3
4894 vpxor xmm4,xmm4,XMMWORD[((32-128))+rax]
4895 vpsrld xmm9,xmm12,27
4896 vpxor xmm6,xmm6,xmm14
4897 vpxor xmm4,xmm4,xmm1
4898
4899 vpslld xmm7,xmm13,30
4900 vpor xmm8,xmm8,xmm9
4901 vpaddd xmm11,xmm11,xmm6
4902 vpsrld xmm5,xmm4,31
4903 vpaddd xmm4,xmm4,xmm4
4904
4905 vpsrld xmm13,xmm13,2
4906 vpaddd xmm11,xmm11,xmm8
4907 vpor xmm4,xmm4,xmm5
4908 vpor xmm13,xmm13,xmm7
4909 vpxor xmm0,xmm0,xmm2
4910 vmovdqa xmm2,XMMWORD[((208-128))+rax]
4911
4912 vpslld xmm8,xmm11,5
4913 vpaddd xmm10,xmm10,xmm15
4914 vpxor xmm6,xmm14,xmm12
4915 vpaddd xmm10,xmm10,xmm4
4916 vpxor xmm0,xmm0,XMMWORD[((48-128))+rax]
4917 vpsrld xmm9,xmm11,27
4918 vpxor xmm6,xmm6,xmm13
4919 vpxor xmm0,xmm0,xmm2
4920
4921 vpslld xmm7,xmm12,30
4922 vpor xmm8,xmm8,xmm9
4923 vpaddd xmm10,xmm10,xmm6
4924 vpsrld xmm5,xmm0,31
4925 vpaddd xmm0,xmm0,xmm0
4926
4927 vpsrld xmm12,xmm12,2
4928 vpaddd xmm10,xmm10,xmm8
4929 vpor xmm0,xmm0,xmm5
4930 vpor xmm12,xmm12,xmm7
4931 vpxor xmm1,xmm1,xmm3
4932 vmovdqa xmm3,XMMWORD[((224-128))+rax]
4933
4934 vpslld xmm8,xmm10,5
4935 vpaddd xmm14,xmm14,xmm15
4936 vpxor xmm6,xmm13,xmm11
4937 vpaddd xmm14,xmm14,xmm0
4938 vpxor xmm1,xmm1,XMMWORD[((64-128))+rax]
4939 vpsrld xmm9,xmm10,27
4940 vpxor xmm6,xmm6,xmm12
4941 vpxor xmm1,xmm1,xmm3
4942
4943 vpslld xmm7,xmm11,30
4944 vpor xmm8,xmm8,xmm9
4945 vpaddd xmm14,xmm14,xmm6
4946 vpsrld xmm5,xmm1,31
4947 vpaddd xmm1,xmm1,xmm1
4948
4949 vpsrld xmm11,xmm11,2
4950 vpaddd xmm14,xmm14,xmm8
4951 vpor xmm1,xmm1,xmm5
4952 vpor xmm11,xmm11,xmm7
4953 vpxor xmm2,xmm2,xmm4
4954 vmovdqa xmm4,XMMWORD[((240-128))+rax]
4955
4956 vpslld xmm8,xmm14,5
4957 vpaddd xmm13,xmm13,xmm15
4958 vpxor xmm6,xmm12,xmm10
4959 vpaddd xmm13,xmm13,xmm1
4960 vpxor xmm2,xmm2,XMMWORD[((80-128))+rax]
4961 vpsrld xmm9,xmm14,27
4962 vpxor xmm6,xmm6,xmm11
4963 vpxor xmm2,xmm2,xmm4
4964
4965 vpslld xmm7,xmm10,30
4966 vpor xmm8,xmm8,xmm9
4967 vpaddd xmm13,xmm13,xmm6
4968 vpsrld xmm5,xmm2,31
4969 vpaddd xmm2,xmm2,xmm2
4970
4971 vpsrld xmm10,xmm10,2
4972 vpaddd xmm13,xmm13,xmm8
4973 vpor xmm2,xmm2,xmm5
4974 vpor xmm10,xmm10,xmm7
4975 vpxor xmm3,xmm3,xmm0
4976 vmovdqa xmm0,XMMWORD[((0-128))+rax]
4977
4978 vpslld xmm8,xmm13,5
4979 vpaddd xmm12,xmm12,xmm15
4980 vpxor xmm6,xmm11,xmm14
4981 vpaddd xmm12,xmm12,xmm2
4982 vpxor xmm3,xmm3,XMMWORD[((96-128))+rax]
4983 vpsrld xmm9,xmm13,27
4984 vpxor xmm6,xmm6,xmm10
4985 vpxor xmm3,xmm3,xmm0
4986
4987 vpslld xmm7,xmm14,30
4988 vpor xmm8,xmm8,xmm9
4989 vpaddd xmm12,xmm12,xmm6
4990 vpsrld xmm5,xmm3,31
4991 vpaddd xmm3,xmm3,xmm3
4992
4993 vpsrld xmm14,xmm14,2
4994 vpaddd xmm12,xmm12,xmm8
4995 vpor xmm3,xmm3,xmm5
4996 vpor xmm14,xmm14,xmm7
4997 vpxor xmm4,xmm4,xmm1
4998 vmovdqa xmm1,XMMWORD[((16-128))+rax]
4999
5000 vpslld xmm8,xmm12,5
5001 vpaddd xmm11,xmm11,xmm15
5002 vpxor xmm6,xmm10,xmm13
5003 vpaddd xmm11,xmm11,xmm3
5004 vpxor xmm4,xmm4,XMMWORD[((112-128))+rax]
5005 vpsrld xmm9,xmm12,27
5006 vpxor xmm6,xmm6,xmm14
5007 vpxor xmm4,xmm4,xmm1
5008
5009 vpslld xmm7,xmm13,30
5010 vpor xmm8,xmm8,xmm9
5011 vpaddd xmm11,xmm11,xmm6
5012 vpsrld xmm5,xmm4,31
5013 vpaddd xmm4,xmm4,xmm4
5014
5015 vpsrld xmm13,xmm13,2
5016 vpaddd xmm11,xmm11,xmm8
5017 vpor xmm4,xmm4,xmm5
5018 vpor xmm13,xmm13,xmm7
5019 vpslld xmm8,xmm11,5
5020 vpaddd xmm10,xmm10,xmm15
5021 vpxor xmm6,xmm14,xmm12
5022
5023 vpsrld xmm9,xmm11,27
5024 vpaddd xmm10,xmm10,xmm4
5025 vpxor xmm6,xmm6,xmm13
5026
5027 vpslld xmm7,xmm12,30
5028 vpor xmm8,xmm8,xmm9
5029 vpaddd xmm10,xmm10,xmm6
5030
5031 vpsrld xmm12,xmm12,2
5032 vpaddd xmm10,xmm10,xmm8
5033 vpor xmm12,xmm12,xmm7
5034 mov ecx,1
5035 cmp ecx,DWORD[rbx]
5036 cmovge r8,rbp
5037 cmp ecx,DWORD[4+rbx]
5038 cmovge r9,rbp
5039 cmp ecx,DWORD[8+rbx]
5040 cmovge r10,rbp
5041 cmp ecx,DWORD[12+rbx]
5042 cmovge r11,rbp
5043 vmovdqu xmm6,XMMWORD[rbx]
5044 vpxor xmm8,xmm8,xmm8
5045 vmovdqa xmm7,xmm6
5046 vpcmpgtd xmm7,xmm7,xmm8
5047 vpaddd xmm6,xmm6,xmm7
5048
5049 vpand xmm10,xmm10,xmm7
5050 vpand xmm11,xmm11,xmm7
5051 vpaddd xmm10,xmm10,XMMWORD[rdi]
5052 vpand xmm12,xmm12,xmm7
5053 vpaddd xmm11,xmm11,XMMWORD[32+rdi]
5054 vpand xmm13,xmm13,xmm7
5055 vpaddd xmm12,xmm12,XMMWORD[64+rdi]
5056 vpand xmm14,xmm14,xmm7
5057 vpaddd xmm13,xmm13,XMMWORD[96+rdi]
5058 vpaddd xmm14,xmm14,XMMWORD[128+rdi]
5059 vmovdqu XMMWORD[rdi],xmm10
5060 vmovdqu XMMWORD[32+rdi],xmm11
5061 vmovdqu XMMWORD[64+rdi],xmm12
5062 vmovdqu XMMWORD[96+rdi],xmm13
5063 vmovdqu XMMWORD[128+rdi],xmm14
5064
5065 vmovdqu XMMWORD[rbx],xmm6
5066 vmovdqu xmm5,XMMWORD[96+rbp]
5067 dec edx
5068 jnz NEAR $L$oop_avx
5069
5070 mov edx,DWORD[280+rsp]
5071 lea rdi,[16+rdi]
5072 lea rsi,[64+rsi]
5073 dec edx
5074 jnz NEAR $L$oop_grande_avx
5075
5076$L$done_avx:
5077 mov rax,QWORD[272+rsp]
5078 vzeroupper
5079 movaps xmm6,XMMWORD[((-184))+rax]
5080 movaps xmm7,XMMWORD[((-168))+rax]
5081 movaps xmm8,XMMWORD[((-152))+rax]
5082 movaps xmm9,XMMWORD[((-136))+rax]
5083 movaps xmm10,XMMWORD[((-120))+rax]
5084 movaps xmm11,XMMWORD[((-104))+rax]
5085 movaps xmm12,XMMWORD[((-88))+rax]
5086 movaps xmm13,XMMWORD[((-72))+rax]
5087 movaps xmm14,XMMWORD[((-56))+rax]
5088 movaps xmm15,XMMWORD[((-40))+rax]
5089 mov rbp,QWORD[((-16))+rax]
5090 mov rbx,QWORD[((-8))+rax]
5091 lea rsp,[rax]
5092$L$epilogue_avx:
5093 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
5094 mov rsi,QWORD[16+rsp]
5095 DB 0F3h,0C3h ;repret
5096$L$SEH_end_sha1_multi_block_avx:
5097
5098ALIGN 32
5099sha1_multi_block_avx2:
5100 mov QWORD[8+rsp],rdi ;WIN64 prologue
5101 mov QWORD[16+rsp],rsi
5102 mov rax,rsp
5103$L$SEH_begin_sha1_multi_block_avx2:
5104 mov rdi,rcx
5105 mov rsi,rdx
5106 mov rdx,r8
5107
5108
5109_avx2_shortcut:
5110 mov rax,rsp
5111 push rbx
5112 push rbp
5113 push r12
5114 push r13
5115 push r14
5116 push r15
5117 lea rsp,[((-168))+rsp]
5118 movaps XMMWORD[rsp],xmm6
5119 movaps XMMWORD[16+rsp],xmm7
5120 movaps XMMWORD[32+rsp],xmm8
5121 movaps XMMWORD[48+rsp],xmm9
5122 movaps XMMWORD[64+rsp],xmm10
5123 movaps XMMWORD[80+rsp],xmm11
5124 movaps XMMWORD[(-120)+rax],xmm12
5125 movaps XMMWORD[(-104)+rax],xmm13
5126 movaps XMMWORD[(-88)+rax],xmm14
5127 movaps XMMWORD[(-72)+rax],xmm15
5128 sub rsp,576
5129 and rsp,-256
5130 mov QWORD[544+rsp],rax
5131$L$body_avx2:
5132 lea rbp,[K_XX_XX]
5133 shr edx,1
5134
5135 vzeroupper
5136$L$oop_grande_avx2:
5137 mov DWORD[552+rsp],edx
5138 xor edx,edx
5139 lea rbx,[512+rsp]
5140 mov r12,QWORD[rsi]
5141 mov ecx,DWORD[8+rsi]
5142 cmp ecx,edx
5143 cmovg edx,ecx
5144 test ecx,ecx
5145 mov DWORD[rbx],ecx
5146 cmovle r12,rbp
5147 mov r13,QWORD[16+rsi]
5148 mov ecx,DWORD[24+rsi]
5149 cmp ecx,edx
5150 cmovg edx,ecx
5151 test ecx,ecx
5152 mov DWORD[4+rbx],ecx
5153 cmovle r13,rbp
5154 mov r14,QWORD[32+rsi]
5155 mov ecx,DWORD[40+rsi]
5156 cmp ecx,edx
5157 cmovg edx,ecx
5158 test ecx,ecx
5159 mov DWORD[8+rbx],ecx
5160 cmovle r14,rbp
5161 mov r15,QWORD[48+rsi]
5162 mov ecx,DWORD[56+rsi]
5163 cmp ecx,edx
5164 cmovg edx,ecx
5165 test ecx,ecx
5166 mov DWORD[12+rbx],ecx
5167 cmovle r15,rbp
5168 mov r8,QWORD[64+rsi]
5169 mov ecx,DWORD[72+rsi]
5170 cmp ecx,edx
5171 cmovg edx,ecx
5172 test ecx,ecx
5173 mov DWORD[16+rbx],ecx
5174 cmovle r8,rbp
5175 mov r9,QWORD[80+rsi]
5176 mov ecx,DWORD[88+rsi]
5177 cmp ecx,edx
5178 cmovg edx,ecx
5179 test ecx,ecx
5180 mov DWORD[20+rbx],ecx
5181 cmovle r9,rbp
5182 mov r10,QWORD[96+rsi]
5183 mov ecx,DWORD[104+rsi]
5184 cmp ecx,edx
5185 cmovg edx,ecx
5186 test ecx,ecx
5187 mov DWORD[24+rbx],ecx
5188 cmovle r10,rbp
5189 mov r11,QWORD[112+rsi]
5190 mov ecx,DWORD[120+rsi]
5191 cmp ecx,edx
5192 cmovg edx,ecx
5193 test ecx,ecx
5194 mov DWORD[28+rbx],ecx
5195 cmovle r11,rbp
5196 vmovdqu ymm0,YMMWORD[rdi]
5197 lea rax,[128+rsp]
5198 vmovdqu ymm1,YMMWORD[32+rdi]
5199 lea rbx,[((256+128))+rsp]
5200 vmovdqu ymm2,YMMWORD[64+rdi]
5201 vmovdqu ymm3,YMMWORD[96+rdi]
5202 vmovdqu ymm4,YMMWORD[128+rdi]
5203 vmovdqu ymm9,YMMWORD[96+rbp]
5204 jmp NEAR $L$oop_avx2
5205
5206ALIGN 32
5207$L$oop_avx2:
5208 vmovdqa ymm15,YMMWORD[((-32))+rbp]
5209 vmovd xmm10,DWORD[r12]
5210 lea r12,[64+r12]
5211 vmovd xmm12,DWORD[r8]
5212 lea r8,[64+r8]
5213 vmovd xmm7,DWORD[r13]
5214 lea r13,[64+r13]
5215 vmovd xmm6,DWORD[r9]
5216 lea r9,[64+r9]
5217 vpinsrd xmm10,xmm10,DWORD[r14],1
5218 lea r14,[64+r14]
5219 vpinsrd xmm12,xmm12,DWORD[r10],1
5220 lea r10,[64+r10]
5221 vpinsrd xmm7,xmm7,DWORD[r15],1
5222 lea r15,[64+r15]
5223 vpunpckldq ymm10,ymm10,ymm7
5224 vpinsrd xmm6,xmm6,DWORD[r11],1
5225 lea r11,[64+r11]
5226 vpunpckldq ymm12,ymm12,ymm6
5227 vmovd xmm11,DWORD[((-60))+r12]
5228 vinserti128 ymm10,ymm10,xmm12,1
5229 vmovd xmm8,DWORD[((-60))+r8]
5230 vpshufb ymm10,ymm10,ymm9
5231 vmovd xmm7,DWORD[((-60))+r13]
5232 vmovd xmm6,DWORD[((-60))+r9]
5233 vpinsrd xmm11,xmm11,DWORD[((-60))+r14],1
5234 vpinsrd xmm8,xmm8,DWORD[((-60))+r10],1
5235 vpinsrd xmm7,xmm7,DWORD[((-60))+r15],1
5236 vpunpckldq ymm11,ymm11,ymm7
5237 vpinsrd xmm6,xmm6,DWORD[((-60))+r11],1
5238 vpunpckldq ymm8,ymm8,ymm6
5239 vpaddd ymm4,ymm4,ymm15
5240 vpslld ymm7,ymm0,5
5241 vpandn ymm6,ymm1,ymm3
5242 vpand ymm5,ymm1,ymm2
5243
5244 vmovdqa YMMWORD[(0-128)+rax],ymm10
5245 vpaddd ymm4,ymm4,ymm10
5246 vinserti128 ymm11,ymm11,xmm8,1
5247 vpsrld ymm8,ymm0,27
5248 vpxor ymm5,ymm5,ymm6
5249 vmovd xmm12,DWORD[((-56))+r12]
5250
5251 vpslld ymm6,ymm1,30
5252 vpor ymm7,ymm7,ymm8
5253 vmovd xmm8,DWORD[((-56))+r8]
5254 vpaddd ymm4,ymm4,ymm5
5255
5256 vpsrld ymm1,ymm1,2
5257 vpaddd ymm4,ymm4,ymm7
5258 vpshufb ymm11,ymm11,ymm9
5259 vpor ymm1,ymm1,ymm6
5260 vmovd xmm7,DWORD[((-56))+r13]
5261 vmovd xmm6,DWORD[((-56))+r9]
5262 vpinsrd xmm12,xmm12,DWORD[((-56))+r14],1
5263 vpinsrd xmm8,xmm8,DWORD[((-56))+r10],1
5264 vpinsrd xmm7,xmm7,DWORD[((-56))+r15],1
5265 vpunpckldq ymm12,ymm12,ymm7
5266 vpinsrd xmm6,xmm6,DWORD[((-56))+r11],1
5267 vpunpckldq ymm8,ymm8,ymm6
5268 vpaddd ymm3,ymm3,ymm15
5269 vpslld ymm7,ymm4,5
5270 vpandn ymm6,ymm0,ymm2
5271 vpand ymm5,ymm0,ymm1
5272
5273 vmovdqa YMMWORD[(32-128)+rax],ymm11
5274 vpaddd ymm3,ymm3,ymm11
5275 vinserti128 ymm12,ymm12,xmm8,1
5276 vpsrld ymm8,ymm4,27
5277 vpxor ymm5,ymm5,ymm6
5278 vmovd xmm13,DWORD[((-52))+r12]
5279
5280 vpslld ymm6,ymm0,30
5281 vpor ymm7,ymm7,ymm8
5282 vmovd xmm8,DWORD[((-52))+r8]
5283 vpaddd ymm3,ymm3,ymm5
5284
5285 vpsrld ymm0,ymm0,2
5286 vpaddd ymm3,ymm3,ymm7
5287 vpshufb ymm12,ymm12,ymm9
5288 vpor ymm0,ymm0,ymm6
5289 vmovd xmm7,DWORD[((-52))+r13]
5290 vmovd xmm6,DWORD[((-52))+r9]
5291 vpinsrd xmm13,xmm13,DWORD[((-52))+r14],1
5292 vpinsrd xmm8,xmm8,DWORD[((-52))+r10],1
5293 vpinsrd xmm7,xmm7,DWORD[((-52))+r15],1
5294 vpunpckldq ymm13,ymm13,ymm7
5295 vpinsrd xmm6,xmm6,DWORD[((-52))+r11],1
5296 vpunpckldq ymm8,ymm8,ymm6
5297 vpaddd ymm2,ymm2,ymm15
5298 vpslld ymm7,ymm3,5
5299 vpandn ymm6,ymm4,ymm1
5300 vpand ymm5,ymm4,ymm0
5301
5302 vmovdqa YMMWORD[(64-128)+rax],ymm12
5303 vpaddd ymm2,ymm2,ymm12
5304 vinserti128 ymm13,ymm13,xmm8,1
5305 vpsrld ymm8,ymm3,27
5306 vpxor ymm5,ymm5,ymm6
5307 vmovd xmm14,DWORD[((-48))+r12]
5308
5309 vpslld ymm6,ymm4,30
5310 vpor ymm7,ymm7,ymm8
5311 vmovd xmm8,DWORD[((-48))+r8]
5312 vpaddd ymm2,ymm2,ymm5
5313
5314 vpsrld ymm4,ymm4,2
5315 vpaddd ymm2,ymm2,ymm7
5316 vpshufb ymm13,ymm13,ymm9
5317 vpor ymm4,ymm4,ymm6
5318 vmovd xmm7,DWORD[((-48))+r13]
5319 vmovd xmm6,DWORD[((-48))+r9]
5320 vpinsrd xmm14,xmm14,DWORD[((-48))+r14],1
5321 vpinsrd xmm8,xmm8,DWORD[((-48))+r10],1
5322 vpinsrd xmm7,xmm7,DWORD[((-48))+r15],1
5323 vpunpckldq ymm14,ymm14,ymm7
5324 vpinsrd xmm6,xmm6,DWORD[((-48))+r11],1
5325 vpunpckldq ymm8,ymm8,ymm6
5326 vpaddd ymm1,ymm1,ymm15
5327 vpslld ymm7,ymm2,5
5328 vpandn ymm6,ymm3,ymm0
5329 vpand ymm5,ymm3,ymm4
5330
5331 vmovdqa YMMWORD[(96-128)+rax],ymm13
5332 vpaddd ymm1,ymm1,ymm13
5333 vinserti128 ymm14,ymm14,xmm8,1
5334 vpsrld ymm8,ymm2,27
5335 vpxor ymm5,ymm5,ymm6
5336 vmovd xmm10,DWORD[((-44))+r12]
5337
5338 vpslld ymm6,ymm3,30
5339 vpor ymm7,ymm7,ymm8
5340 vmovd xmm8,DWORD[((-44))+r8]
5341 vpaddd ymm1,ymm1,ymm5
5342
5343 vpsrld ymm3,ymm3,2
5344 vpaddd ymm1,ymm1,ymm7
5345 vpshufb ymm14,ymm14,ymm9
5346 vpor ymm3,ymm3,ymm6
5347 vmovd xmm7,DWORD[((-44))+r13]
5348 vmovd xmm6,DWORD[((-44))+r9]
5349 vpinsrd xmm10,xmm10,DWORD[((-44))+r14],1
5350 vpinsrd xmm8,xmm8,DWORD[((-44))+r10],1
5351 vpinsrd xmm7,xmm7,DWORD[((-44))+r15],1
5352 vpunpckldq ymm10,ymm10,ymm7
5353 vpinsrd xmm6,xmm6,DWORD[((-44))+r11],1
5354 vpunpckldq ymm8,ymm8,ymm6
5355 vpaddd ymm0,ymm0,ymm15
5356 vpslld ymm7,ymm1,5
5357 vpandn ymm6,ymm2,ymm4
5358 vpand ymm5,ymm2,ymm3
5359
5360 vmovdqa YMMWORD[(128-128)+rax],ymm14
5361 vpaddd ymm0,ymm0,ymm14
5362 vinserti128 ymm10,ymm10,xmm8,1
5363 vpsrld ymm8,ymm1,27
5364 vpxor ymm5,ymm5,ymm6
5365 vmovd xmm11,DWORD[((-40))+r12]
5366
5367 vpslld ymm6,ymm2,30
5368 vpor ymm7,ymm7,ymm8
5369 vmovd xmm8,DWORD[((-40))+r8]
5370 vpaddd ymm0,ymm0,ymm5
5371
5372 vpsrld ymm2,ymm2,2
5373 vpaddd ymm0,ymm0,ymm7
5374 vpshufb ymm10,ymm10,ymm9
5375 vpor ymm2,ymm2,ymm6
5376 vmovd xmm7,DWORD[((-40))+r13]
5377 vmovd xmm6,DWORD[((-40))+r9]
5378 vpinsrd xmm11,xmm11,DWORD[((-40))+r14],1
5379 vpinsrd xmm8,xmm8,DWORD[((-40))+r10],1
5380 vpinsrd xmm7,xmm7,DWORD[((-40))+r15],1
5381 vpunpckldq ymm11,ymm11,ymm7
5382 vpinsrd xmm6,xmm6,DWORD[((-40))+r11],1
5383 vpunpckldq ymm8,ymm8,ymm6
5384 vpaddd ymm4,ymm4,ymm15
5385 vpslld ymm7,ymm0,5
5386 vpandn ymm6,ymm1,ymm3
5387 vpand ymm5,ymm1,ymm2
5388
5389 vmovdqa YMMWORD[(160-128)+rax],ymm10
5390 vpaddd ymm4,ymm4,ymm10
5391 vinserti128 ymm11,ymm11,xmm8,1
5392 vpsrld ymm8,ymm0,27
5393 vpxor ymm5,ymm5,ymm6
5394 vmovd xmm12,DWORD[((-36))+r12]
5395
5396 vpslld ymm6,ymm1,30
5397 vpor ymm7,ymm7,ymm8
5398 vmovd xmm8,DWORD[((-36))+r8]
5399 vpaddd ymm4,ymm4,ymm5
5400
5401 vpsrld ymm1,ymm1,2
5402 vpaddd ymm4,ymm4,ymm7
5403 vpshufb ymm11,ymm11,ymm9
5404 vpor ymm1,ymm1,ymm6
5405 vmovd xmm7,DWORD[((-36))+r13]
5406 vmovd xmm6,DWORD[((-36))+r9]
5407 vpinsrd xmm12,xmm12,DWORD[((-36))+r14],1
5408 vpinsrd xmm8,xmm8,DWORD[((-36))+r10],1
5409 vpinsrd xmm7,xmm7,DWORD[((-36))+r15],1
5410 vpunpckldq ymm12,ymm12,ymm7
5411 vpinsrd xmm6,xmm6,DWORD[((-36))+r11],1
5412 vpunpckldq ymm8,ymm8,ymm6
5413 vpaddd ymm3,ymm3,ymm15
5414 vpslld ymm7,ymm4,5
5415 vpandn ymm6,ymm0,ymm2
5416 vpand ymm5,ymm0,ymm1
5417
5418 vmovdqa YMMWORD[(192-128)+rax],ymm11
5419 vpaddd ymm3,ymm3,ymm11
5420 vinserti128 ymm12,ymm12,xmm8,1
5421 vpsrld ymm8,ymm4,27
5422 vpxor ymm5,ymm5,ymm6
5423 vmovd xmm13,DWORD[((-32))+r12]
5424
5425 vpslld ymm6,ymm0,30
5426 vpor ymm7,ymm7,ymm8
5427 vmovd xmm8,DWORD[((-32))+r8]
5428 vpaddd ymm3,ymm3,ymm5
5429
5430 vpsrld ymm0,ymm0,2
5431 vpaddd ymm3,ymm3,ymm7
5432 vpshufb ymm12,ymm12,ymm9
5433 vpor ymm0,ymm0,ymm6
5434 vmovd xmm7,DWORD[((-32))+r13]
5435 vmovd xmm6,DWORD[((-32))+r9]
5436 vpinsrd xmm13,xmm13,DWORD[((-32))+r14],1
5437 vpinsrd xmm8,xmm8,DWORD[((-32))+r10],1
5438 vpinsrd xmm7,xmm7,DWORD[((-32))+r15],1
5439 vpunpckldq ymm13,ymm13,ymm7
5440 vpinsrd xmm6,xmm6,DWORD[((-32))+r11],1
5441 vpunpckldq ymm8,ymm8,ymm6
5442 vpaddd ymm2,ymm2,ymm15
5443 vpslld ymm7,ymm3,5
5444 vpandn ymm6,ymm4,ymm1
5445 vpand ymm5,ymm4,ymm0
5446
5447 vmovdqa YMMWORD[(224-128)+rax],ymm12
5448 vpaddd ymm2,ymm2,ymm12
5449 vinserti128 ymm13,ymm13,xmm8,1
5450 vpsrld ymm8,ymm3,27
5451 vpxor ymm5,ymm5,ymm6
5452 vmovd xmm14,DWORD[((-28))+r12]
5453
5454 vpslld ymm6,ymm4,30
5455 vpor ymm7,ymm7,ymm8
5456 vmovd xmm8,DWORD[((-28))+r8]
5457 vpaddd ymm2,ymm2,ymm5
5458
5459 vpsrld ymm4,ymm4,2
5460 vpaddd ymm2,ymm2,ymm7
5461 vpshufb ymm13,ymm13,ymm9
5462 vpor ymm4,ymm4,ymm6
5463 vmovd xmm7,DWORD[((-28))+r13]
5464 vmovd xmm6,DWORD[((-28))+r9]
5465 vpinsrd xmm14,xmm14,DWORD[((-28))+r14],1
5466 vpinsrd xmm8,xmm8,DWORD[((-28))+r10],1
5467 vpinsrd xmm7,xmm7,DWORD[((-28))+r15],1
5468 vpunpckldq ymm14,ymm14,ymm7
5469 vpinsrd xmm6,xmm6,DWORD[((-28))+r11],1
5470 vpunpckldq ymm8,ymm8,ymm6
5471 vpaddd ymm1,ymm1,ymm15
5472 vpslld ymm7,ymm2,5
5473 vpandn ymm6,ymm3,ymm0
5474 vpand ymm5,ymm3,ymm4
5475
5476 vmovdqa YMMWORD[(256-256-128)+rbx],ymm13
5477 vpaddd ymm1,ymm1,ymm13
5478 vinserti128 ymm14,ymm14,xmm8,1
5479 vpsrld ymm8,ymm2,27
5480 vpxor ymm5,ymm5,ymm6
5481 vmovd xmm10,DWORD[((-24))+r12]
5482
5483 vpslld ymm6,ymm3,30
5484 vpor ymm7,ymm7,ymm8
5485 vmovd xmm8,DWORD[((-24))+r8]
5486 vpaddd ymm1,ymm1,ymm5
5487
5488 vpsrld ymm3,ymm3,2
5489 vpaddd ymm1,ymm1,ymm7
5490 vpshufb ymm14,ymm14,ymm9
5491 vpor ymm3,ymm3,ymm6
5492 vmovd xmm7,DWORD[((-24))+r13]
5493 vmovd xmm6,DWORD[((-24))+r9]
5494 vpinsrd xmm10,xmm10,DWORD[((-24))+r14],1
5495 vpinsrd xmm8,xmm8,DWORD[((-24))+r10],1
5496 vpinsrd xmm7,xmm7,DWORD[((-24))+r15],1
5497 vpunpckldq ymm10,ymm10,ymm7
5498 vpinsrd xmm6,xmm6,DWORD[((-24))+r11],1
5499 vpunpckldq ymm8,ymm8,ymm6
5500 vpaddd ymm0,ymm0,ymm15
5501 vpslld ymm7,ymm1,5
5502 vpandn ymm6,ymm2,ymm4
5503 vpand ymm5,ymm2,ymm3
5504
5505 vmovdqa YMMWORD[(288-256-128)+rbx],ymm14
5506 vpaddd ymm0,ymm0,ymm14
5507 vinserti128 ymm10,ymm10,xmm8,1
5508 vpsrld ymm8,ymm1,27
5509 vpxor ymm5,ymm5,ymm6
5510 vmovd xmm11,DWORD[((-20))+r12]
5511
5512 vpslld ymm6,ymm2,30
5513 vpor ymm7,ymm7,ymm8
5514 vmovd xmm8,DWORD[((-20))+r8]
5515 vpaddd ymm0,ymm0,ymm5
5516
5517 vpsrld ymm2,ymm2,2
5518 vpaddd ymm0,ymm0,ymm7
5519 vpshufb ymm10,ymm10,ymm9
5520 vpor ymm2,ymm2,ymm6
5521 vmovd xmm7,DWORD[((-20))+r13]
5522 vmovd xmm6,DWORD[((-20))+r9]
5523 vpinsrd xmm11,xmm11,DWORD[((-20))+r14],1
5524 vpinsrd xmm8,xmm8,DWORD[((-20))+r10],1
5525 vpinsrd xmm7,xmm7,DWORD[((-20))+r15],1
5526 vpunpckldq ymm11,ymm11,ymm7
5527 vpinsrd xmm6,xmm6,DWORD[((-20))+r11],1
5528 vpunpckldq ymm8,ymm8,ymm6
5529 vpaddd ymm4,ymm4,ymm15
5530 vpslld ymm7,ymm0,5
5531 vpandn ymm6,ymm1,ymm3
5532 vpand ymm5,ymm1,ymm2
5533
5534 vmovdqa YMMWORD[(320-256-128)+rbx],ymm10
5535 vpaddd ymm4,ymm4,ymm10
5536 vinserti128 ymm11,ymm11,xmm8,1
5537 vpsrld ymm8,ymm0,27
5538 vpxor ymm5,ymm5,ymm6
5539 vmovd xmm12,DWORD[((-16))+r12]
5540
5541 vpslld ymm6,ymm1,30
5542 vpor ymm7,ymm7,ymm8
5543 vmovd xmm8,DWORD[((-16))+r8]
5544 vpaddd ymm4,ymm4,ymm5
5545
5546 vpsrld ymm1,ymm1,2
5547 vpaddd ymm4,ymm4,ymm7
5548 vpshufb ymm11,ymm11,ymm9
5549 vpor ymm1,ymm1,ymm6
5550 vmovd xmm7,DWORD[((-16))+r13]
5551 vmovd xmm6,DWORD[((-16))+r9]
5552 vpinsrd xmm12,xmm12,DWORD[((-16))+r14],1
5553 vpinsrd xmm8,xmm8,DWORD[((-16))+r10],1
5554 vpinsrd xmm7,xmm7,DWORD[((-16))+r15],1
5555 vpunpckldq ymm12,ymm12,ymm7
5556 vpinsrd xmm6,xmm6,DWORD[((-16))+r11],1
5557 vpunpckldq ymm8,ymm8,ymm6
5558 vpaddd ymm3,ymm3,ymm15
5559 vpslld ymm7,ymm4,5
5560 vpandn ymm6,ymm0,ymm2
5561 vpand ymm5,ymm0,ymm1
5562
5563 vmovdqa YMMWORD[(352-256-128)+rbx],ymm11
5564 vpaddd ymm3,ymm3,ymm11
5565 vinserti128 ymm12,ymm12,xmm8,1
5566 vpsrld ymm8,ymm4,27
5567 vpxor ymm5,ymm5,ymm6
5568 vmovd xmm13,DWORD[((-12))+r12]
5569
5570 vpslld ymm6,ymm0,30
5571 vpor ymm7,ymm7,ymm8
5572 vmovd xmm8,DWORD[((-12))+r8]
5573 vpaddd ymm3,ymm3,ymm5
5574
5575 vpsrld ymm0,ymm0,2
5576 vpaddd ymm3,ymm3,ymm7
5577 vpshufb ymm12,ymm12,ymm9
5578 vpor ymm0,ymm0,ymm6
5579 vmovd xmm7,DWORD[((-12))+r13]
5580 vmovd xmm6,DWORD[((-12))+r9]
5581 vpinsrd xmm13,xmm13,DWORD[((-12))+r14],1
5582 vpinsrd xmm8,xmm8,DWORD[((-12))+r10],1
5583 vpinsrd xmm7,xmm7,DWORD[((-12))+r15],1
5584 vpunpckldq ymm13,ymm13,ymm7
5585 vpinsrd xmm6,xmm6,DWORD[((-12))+r11],1
5586 vpunpckldq ymm8,ymm8,ymm6
5587 vpaddd ymm2,ymm2,ymm15
5588 vpslld ymm7,ymm3,5
5589 vpandn ymm6,ymm4,ymm1
5590 vpand ymm5,ymm4,ymm0
5591
5592 vmovdqa YMMWORD[(384-256-128)+rbx],ymm12
5593 vpaddd ymm2,ymm2,ymm12
5594 vinserti128 ymm13,ymm13,xmm8,1
5595 vpsrld ymm8,ymm3,27
5596 vpxor ymm5,ymm5,ymm6
5597 vmovd xmm14,DWORD[((-8))+r12]
5598
5599 vpslld ymm6,ymm4,30
5600 vpor ymm7,ymm7,ymm8
5601 vmovd xmm8,DWORD[((-8))+r8]
5602 vpaddd ymm2,ymm2,ymm5
5603
5604 vpsrld ymm4,ymm4,2
5605 vpaddd ymm2,ymm2,ymm7
5606 vpshufb ymm13,ymm13,ymm9
5607 vpor ymm4,ymm4,ymm6
5608 vmovd xmm7,DWORD[((-8))+r13]
5609 vmovd xmm6,DWORD[((-8))+r9]
5610 vpinsrd xmm14,xmm14,DWORD[((-8))+r14],1
5611 vpinsrd xmm8,xmm8,DWORD[((-8))+r10],1
5612 vpinsrd xmm7,xmm7,DWORD[((-8))+r15],1
5613 vpunpckldq ymm14,ymm14,ymm7
5614 vpinsrd xmm6,xmm6,DWORD[((-8))+r11],1
5615 vpunpckldq ymm8,ymm8,ymm6
5616 vpaddd ymm1,ymm1,ymm15
5617 vpslld ymm7,ymm2,5
5618 vpandn ymm6,ymm3,ymm0
5619 vpand ymm5,ymm3,ymm4
5620
5621 vmovdqa YMMWORD[(416-256-128)+rbx],ymm13
5622 vpaddd ymm1,ymm1,ymm13
5623 vinserti128 ymm14,ymm14,xmm8,1
5624 vpsrld ymm8,ymm2,27
5625 vpxor ymm5,ymm5,ymm6
5626 vmovd xmm10,DWORD[((-4))+r12]
5627
5628 vpslld ymm6,ymm3,30
5629 vpor ymm7,ymm7,ymm8
5630 vmovd xmm8,DWORD[((-4))+r8]
5631 vpaddd ymm1,ymm1,ymm5
5632
5633 vpsrld ymm3,ymm3,2
5634 vpaddd ymm1,ymm1,ymm7
5635 vpshufb ymm14,ymm14,ymm9
5636 vpor ymm3,ymm3,ymm6
5637 vmovdqa ymm11,YMMWORD[((0-128))+rax]
5638 vmovd xmm7,DWORD[((-4))+r13]
5639 vmovd xmm6,DWORD[((-4))+r9]
5640 vpinsrd xmm10,xmm10,DWORD[((-4))+r14],1
5641 vpinsrd xmm8,xmm8,DWORD[((-4))+r10],1
5642 vpinsrd xmm7,xmm7,DWORD[((-4))+r15],1
5643 vpunpckldq ymm10,ymm10,ymm7
5644 vpinsrd xmm6,xmm6,DWORD[((-4))+r11],1
5645 vpunpckldq ymm8,ymm8,ymm6
5646 vpaddd ymm0,ymm0,ymm15
5647 prefetcht0 [63+r12]
5648 vpslld ymm7,ymm1,5
5649 vpandn ymm6,ymm2,ymm4
5650 vpand ymm5,ymm2,ymm3
5651
5652 vmovdqa YMMWORD[(448-256-128)+rbx],ymm14
5653 vpaddd ymm0,ymm0,ymm14
5654 vinserti128 ymm10,ymm10,xmm8,1
5655 vpsrld ymm8,ymm1,27
5656 prefetcht0 [63+r13]
5657 vpxor ymm5,ymm5,ymm6
5658
5659 vpslld ymm6,ymm2,30
5660 vpor ymm7,ymm7,ymm8
5661 prefetcht0 [63+r14]
5662 vpaddd ymm0,ymm0,ymm5
5663
5664 vpsrld ymm2,ymm2,2
5665 vpaddd ymm0,ymm0,ymm7
5666 prefetcht0 [63+r15]
5667 vpshufb ymm10,ymm10,ymm9
5668 vpor ymm2,ymm2,ymm6
5669 vmovdqa ymm12,YMMWORD[((32-128))+rax]
5670 vpxor ymm11,ymm11,ymm13
5671 vmovdqa ymm13,YMMWORD[((64-128))+rax]
5672
5673 vpaddd ymm4,ymm4,ymm15
5674 vpslld ymm7,ymm0,5
5675 vpandn ymm6,ymm1,ymm3
5676 prefetcht0 [63+r8]
5677 vpand ymm5,ymm1,ymm2
5678
5679 vmovdqa YMMWORD[(480-256-128)+rbx],ymm10
5680 vpaddd ymm4,ymm4,ymm10
5681 vpxor ymm11,ymm11,YMMWORD[((256-256-128))+rbx]
5682 vpsrld ymm8,ymm0,27
5683 vpxor ymm5,ymm5,ymm6
5684 vpxor ymm11,ymm11,ymm13
5685 prefetcht0 [63+r9]
5686
5687 vpslld ymm6,ymm1,30
5688 vpor ymm7,ymm7,ymm8
5689 vpaddd ymm4,ymm4,ymm5
5690 prefetcht0 [63+r10]
5691 vpsrld ymm9,ymm11,31
5692 vpaddd ymm11,ymm11,ymm11
5693
5694 vpsrld ymm1,ymm1,2
5695 prefetcht0 [63+r11]
5696 vpaddd ymm4,ymm4,ymm7
5697 vpor ymm11,ymm11,ymm9
5698 vpor ymm1,ymm1,ymm6
5699 vpxor ymm12,ymm12,ymm14
5700 vmovdqa ymm14,YMMWORD[((96-128))+rax]
5701
5702 vpaddd ymm3,ymm3,ymm15
5703 vpslld ymm7,ymm4,5
5704 vpandn ymm6,ymm0,ymm2
5705
5706 vpand ymm5,ymm0,ymm1
5707
5708 vmovdqa YMMWORD[(0-128)+rax],ymm11
5709 vpaddd ymm3,ymm3,ymm11
5710 vpxor ymm12,ymm12,YMMWORD[((288-256-128))+rbx]
5711 vpsrld ymm8,ymm4,27
5712 vpxor ymm5,ymm5,ymm6
5713 vpxor ymm12,ymm12,ymm14
5714
5715
5716 vpslld ymm6,ymm0,30
5717 vpor ymm7,ymm7,ymm8
5718 vpaddd ymm3,ymm3,ymm5
5719
5720 vpsrld ymm9,ymm12,31
5721 vpaddd ymm12,ymm12,ymm12
5722
5723 vpsrld ymm0,ymm0,2
5724
5725 vpaddd ymm3,ymm3,ymm7
5726 vpor ymm12,ymm12,ymm9
5727 vpor ymm0,ymm0,ymm6
5728 vpxor ymm13,ymm13,ymm10
5729 vmovdqa ymm10,YMMWORD[((128-128))+rax]
5730
5731 vpaddd ymm2,ymm2,ymm15
5732 vpslld ymm7,ymm3,5
5733 vpandn ymm6,ymm4,ymm1
5734
5735 vpand ymm5,ymm4,ymm0
5736
5737 vmovdqa YMMWORD[(32-128)+rax],ymm12
5738 vpaddd ymm2,ymm2,ymm12
5739 vpxor ymm13,ymm13,YMMWORD[((320-256-128))+rbx]
5740 vpsrld ymm8,ymm3,27
5741 vpxor ymm5,ymm5,ymm6
5742 vpxor ymm13,ymm13,ymm10
5743
5744
5745 vpslld ymm6,ymm4,30
5746 vpor ymm7,ymm7,ymm8
5747 vpaddd ymm2,ymm2,ymm5
5748
5749 vpsrld ymm9,ymm13,31
5750 vpaddd ymm13,ymm13,ymm13
5751
5752 vpsrld ymm4,ymm4,2
5753
5754 vpaddd ymm2,ymm2,ymm7
5755 vpor ymm13,ymm13,ymm9
5756 vpor ymm4,ymm4,ymm6
5757 vpxor ymm14,ymm14,ymm11
5758 vmovdqa ymm11,YMMWORD[((160-128))+rax]
5759
5760 vpaddd ymm1,ymm1,ymm15
5761 vpslld ymm7,ymm2,5
5762 vpandn ymm6,ymm3,ymm0
5763
5764 vpand ymm5,ymm3,ymm4
5765
5766 vmovdqa YMMWORD[(64-128)+rax],ymm13
5767 vpaddd ymm1,ymm1,ymm13
5768 vpxor ymm14,ymm14,YMMWORD[((352-256-128))+rbx]
5769 vpsrld ymm8,ymm2,27
5770 vpxor ymm5,ymm5,ymm6
5771 vpxor ymm14,ymm14,ymm11
5772
5773
5774 vpslld ymm6,ymm3,30
5775 vpor ymm7,ymm7,ymm8
5776 vpaddd ymm1,ymm1,ymm5
5777
5778 vpsrld ymm9,ymm14,31
5779 vpaddd ymm14,ymm14,ymm14
5780
5781 vpsrld ymm3,ymm3,2
5782
5783 vpaddd ymm1,ymm1,ymm7
5784 vpor ymm14,ymm14,ymm9
5785 vpor ymm3,ymm3,ymm6
5786 vpxor ymm10,ymm10,ymm12
5787 vmovdqa ymm12,YMMWORD[((192-128))+rax]
5788
5789 vpaddd ymm0,ymm0,ymm15
5790 vpslld ymm7,ymm1,5
5791 vpandn ymm6,ymm2,ymm4
5792
5793 vpand ymm5,ymm2,ymm3
5794
5795 vmovdqa YMMWORD[(96-128)+rax],ymm14
5796 vpaddd ymm0,ymm0,ymm14
5797 vpxor ymm10,ymm10,YMMWORD[((384-256-128))+rbx]
5798 vpsrld ymm8,ymm1,27
5799 vpxor ymm5,ymm5,ymm6
5800 vpxor ymm10,ymm10,ymm12
5801
5802
5803 vpslld ymm6,ymm2,30
5804 vpor ymm7,ymm7,ymm8
5805 vpaddd ymm0,ymm0,ymm5
5806
5807 vpsrld ymm9,ymm10,31
5808 vpaddd ymm10,ymm10,ymm10
5809
5810 vpsrld ymm2,ymm2,2
5811
5812 vpaddd ymm0,ymm0,ymm7
5813 vpor ymm10,ymm10,ymm9
5814 vpor ymm2,ymm2,ymm6
5815 vmovdqa ymm15,YMMWORD[rbp]
5816 vpxor ymm11,ymm11,ymm13
5817 vmovdqa ymm13,YMMWORD[((224-128))+rax]
5818
5819 vpslld ymm7,ymm0,5
5820 vpaddd ymm4,ymm4,ymm15
5821 vpxor ymm5,ymm3,ymm1
5822 vmovdqa YMMWORD[(128-128)+rax],ymm10
5823 vpaddd ymm4,ymm4,ymm10
5824 vpxor ymm11,ymm11,YMMWORD[((416-256-128))+rbx]
5825 vpsrld ymm8,ymm0,27
5826 vpxor ymm5,ymm5,ymm2
5827 vpxor ymm11,ymm11,ymm13
5828
5829 vpslld ymm6,ymm1,30
5830 vpor ymm7,ymm7,ymm8
5831 vpaddd ymm4,ymm4,ymm5
5832 vpsrld ymm9,ymm11,31
5833 vpaddd ymm11,ymm11,ymm11
5834
5835 vpsrld ymm1,ymm1,2
5836 vpaddd ymm4,ymm4,ymm7
5837 vpor ymm11,ymm11,ymm9
5838 vpor ymm1,ymm1,ymm6
5839 vpxor ymm12,ymm12,ymm14
5840 vmovdqa ymm14,YMMWORD[((256-256-128))+rbx]
5841
5842 vpslld ymm7,ymm4,5
5843 vpaddd ymm3,ymm3,ymm15
5844 vpxor ymm5,ymm2,ymm0
5845 vmovdqa YMMWORD[(160-128)+rax],ymm11
5846 vpaddd ymm3,ymm3,ymm11
5847 vpxor ymm12,ymm12,YMMWORD[((448-256-128))+rbx]
5848 vpsrld ymm8,ymm4,27
5849 vpxor ymm5,ymm5,ymm1
5850 vpxor ymm12,ymm12,ymm14
5851
5852 vpslld ymm6,ymm0,30
5853 vpor ymm7,ymm7,ymm8
5854 vpaddd ymm3,ymm3,ymm5
5855 vpsrld ymm9,ymm12,31
5856 vpaddd ymm12,ymm12,ymm12
5857
5858 vpsrld ymm0,ymm0,2
5859 vpaddd ymm3,ymm3,ymm7
5860 vpor ymm12,ymm12,ymm9
5861 vpor ymm0,ymm0,ymm6
5862 vpxor ymm13,ymm13,ymm10
5863 vmovdqa ymm10,YMMWORD[((288-256-128))+rbx]
5864
5865 vpslld ymm7,ymm3,5
5866 vpaddd ymm2,ymm2,ymm15
5867 vpxor ymm5,ymm1,ymm4
5868 vmovdqa YMMWORD[(192-128)+rax],ymm12
5869 vpaddd ymm2,ymm2,ymm12
5870 vpxor ymm13,ymm13,YMMWORD[((480-256-128))+rbx]
5871 vpsrld ymm8,ymm3,27
5872 vpxor ymm5,ymm5,ymm0
5873 vpxor ymm13,ymm13,ymm10
5874
5875 vpslld ymm6,ymm4,30
5876 vpor ymm7,ymm7,ymm8
5877 vpaddd ymm2,ymm2,ymm5
5878 vpsrld ymm9,ymm13,31
5879 vpaddd ymm13,ymm13,ymm13
5880
5881 vpsrld ymm4,ymm4,2
5882 vpaddd ymm2,ymm2,ymm7
5883 vpor ymm13,ymm13,ymm9
5884 vpor ymm4,ymm4,ymm6
5885 vpxor ymm14,ymm14,ymm11
5886 vmovdqa ymm11,YMMWORD[((320-256-128))+rbx]
5887
5888 vpslld ymm7,ymm2,5
5889 vpaddd ymm1,ymm1,ymm15
5890 vpxor ymm5,ymm0,ymm3
5891 vmovdqa YMMWORD[(224-128)+rax],ymm13
5892 vpaddd ymm1,ymm1,ymm13
5893 vpxor ymm14,ymm14,YMMWORD[((0-128))+rax]
5894 vpsrld ymm8,ymm2,27
5895 vpxor ymm5,ymm5,ymm4
5896 vpxor ymm14,ymm14,ymm11
5897
5898 vpslld ymm6,ymm3,30
5899 vpor ymm7,ymm7,ymm8
5900 vpaddd ymm1,ymm1,ymm5
5901 vpsrld ymm9,ymm14,31
5902 vpaddd ymm14,ymm14,ymm14
5903
5904 vpsrld ymm3,ymm3,2
5905 vpaddd ymm1,ymm1,ymm7
5906 vpor ymm14,ymm14,ymm9
5907 vpor ymm3,ymm3,ymm6
5908 vpxor ymm10,ymm10,ymm12
5909 vmovdqa ymm12,YMMWORD[((352-256-128))+rbx]
5910
5911 vpslld ymm7,ymm1,5
5912 vpaddd ymm0,ymm0,ymm15
5913 vpxor ymm5,ymm4,ymm2
5914 vmovdqa YMMWORD[(256-256-128)+rbx],ymm14
5915 vpaddd ymm0,ymm0,ymm14
5916 vpxor ymm10,ymm10,YMMWORD[((32-128))+rax]
5917 vpsrld ymm8,ymm1,27
5918 vpxor ymm5,ymm5,ymm3
5919 vpxor ymm10,ymm10,ymm12
5920
5921 vpslld ymm6,ymm2,30
5922 vpor ymm7,ymm7,ymm8
5923 vpaddd ymm0,ymm0,ymm5
5924 vpsrld ymm9,ymm10,31
5925 vpaddd ymm10,ymm10,ymm10
5926
5927 vpsrld ymm2,ymm2,2
5928 vpaddd ymm0,ymm0,ymm7
5929 vpor ymm10,ymm10,ymm9
5930 vpor ymm2,ymm2,ymm6
5931 vpxor ymm11,ymm11,ymm13
5932 vmovdqa ymm13,YMMWORD[((384-256-128))+rbx]
5933
5934 vpslld ymm7,ymm0,5
5935 vpaddd ymm4,ymm4,ymm15
5936 vpxor ymm5,ymm3,ymm1
5937 vmovdqa YMMWORD[(288-256-128)+rbx],ymm10
5938 vpaddd ymm4,ymm4,ymm10
5939 vpxor ymm11,ymm11,YMMWORD[((64-128))+rax]
5940 vpsrld ymm8,ymm0,27
5941 vpxor ymm5,ymm5,ymm2
5942 vpxor ymm11,ymm11,ymm13
5943
5944 vpslld ymm6,ymm1,30
5945 vpor ymm7,ymm7,ymm8
5946 vpaddd ymm4,ymm4,ymm5
5947 vpsrld ymm9,ymm11,31
5948 vpaddd ymm11,ymm11,ymm11
5949
5950 vpsrld ymm1,ymm1,2
5951 vpaddd ymm4,ymm4,ymm7
5952 vpor ymm11,ymm11,ymm9
5953 vpor ymm1,ymm1,ymm6
5954 vpxor ymm12,ymm12,ymm14
5955 vmovdqa ymm14,YMMWORD[((416-256-128))+rbx]
5956
5957 vpslld ymm7,ymm4,5
5958 vpaddd ymm3,ymm3,ymm15
5959 vpxor ymm5,ymm2,ymm0
5960 vmovdqa YMMWORD[(320-256-128)+rbx],ymm11
5961 vpaddd ymm3,ymm3,ymm11
5962 vpxor ymm12,ymm12,YMMWORD[((96-128))+rax]
5963 vpsrld ymm8,ymm4,27
5964 vpxor ymm5,ymm5,ymm1
5965 vpxor ymm12,ymm12,ymm14
5966
5967 vpslld ymm6,ymm0,30
5968 vpor ymm7,ymm7,ymm8
5969 vpaddd ymm3,ymm3,ymm5
5970 vpsrld ymm9,ymm12,31
5971 vpaddd ymm12,ymm12,ymm12
5972
5973 vpsrld ymm0,ymm0,2
5974 vpaddd ymm3,ymm3,ymm7
5975 vpor ymm12,ymm12,ymm9
5976 vpor ymm0,ymm0,ymm6
5977 vpxor ymm13,ymm13,ymm10
5978 vmovdqa ymm10,YMMWORD[((448-256-128))+rbx]
5979
5980 vpslld ymm7,ymm3,5
5981 vpaddd ymm2,ymm2,ymm15
5982 vpxor ymm5,ymm1,ymm4
5983 vmovdqa YMMWORD[(352-256-128)+rbx],ymm12
5984 vpaddd ymm2,ymm2,ymm12
5985 vpxor ymm13,ymm13,YMMWORD[((128-128))+rax]
5986 vpsrld ymm8,ymm3,27
5987 vpxor ymm5,ymm5,ymm0
5988 vpxor ymm13,ymm13,ymm10
5989
5990 vpslld ymm6,ymm4,30
5991 vpor ymm7,ymm7,ymm8
5992 vpaddd ymm2,ymm2,ymm5
5993 vpsrld ymm9,ymm13,31
5994 vpaddd ymm13,ymm13,ymm13
5995
5996 vpsrld ymm4,ymm4,2
5997 vpaddd ymm2,ymm2,ymm7
5998 vpor ymm13,ymm13,ymm9
5999 vpor ymm4,ymm4,ymm6
6000 vpxor ymm14,ymm14,ymm11
6001 vmovdqa ymm11,YMMWORD[((480-256-128))+rbx]
6002
6003 vpslld ymm7,ymm2,5
6004 vpaddd ymm1,ymm1,ymm15
6005 vpxor ymm5,ymm0,ymm3
6006 vmovdqa YMMWORD[(384-256-128)+rbx],ymm13
6007 vpaddd ymm1,ymm1,ymm13
6008 vpxor ymm14,ymm14,YMMWORD[((160-128))+rax]
6009 vpsrld ymm8,ymm2,27
6010 vpxor ymm5,ymm5,ymm4
6011 vpxor ymm14,ymm14,ymm11
6012
6013 vpslld ymm6,ymm3,30
6014 vpor ymm7,ymm7,ymm8
6015 vpaddd ymm1,ymm1,ymm5
6016 vpsrld ymm9,ymm14,31
6017 vpaddd ymm14,ymm14,ymm14
6018
6019 vpsrld ymm3,ymm3,2
6020 vpaddd ymm1,ymm1,ymm7
6021 vpor ymm14,ymm14,ymm9
6022 vpor ymm3,ymm3,ymm6
6023 vpxor ymm10,ymm10,ymm12
6024 vmovdqa ymm12,YMMWORD[((0-128))+rax]
6025
6026 vpslld ymm7,ymm1,5
6027 vpaddd ymm0,ymm0,ymm15
6028 vpxor ymm5,ymm4,ymm2
6029 vmovdqa YMMWORD[(416-256-128)+rbx],ymm14
6030 vpaddd ymm0,ymm0,ymm14
6031 vpxor ymm10,ymm10,YMMWORD[((192-128))+rax]
6032 vpsrld ymm8,ymm1,27
6033 vpxor ymm5,ymm5,ymm3
6034 vpxor ymm10,ymm10,ymm12
6035
6036 vpslld ymm6,ymm2,30
6037 vpor ymm7,ymm7,ymm8
6038 vpaddd ymm0,ymm0,ymm5
6039 vpsrld ymm9,ymm10,31
6040 vpaddd ymm10,ymm10,ymm10
6041
6042 vpsrld ymm2,ymm2,2
6043 vpaddd ymm0,ymm0,ymm7
6044 vpor ymm10,ymm10,ymm9
6045 vpor ymm2,ymm2,ymm6
6046 vpxor ymm11,ymm11,ymm13
6047 vmovdqa ymm13,YMMWORD[((32-128))+rax]
6048
6049 vpslld ymm7,ymm0,5
6050 vpaddd ymm4,ymm4,ymm15
6051 vpxor ymm5,ymm3,ymm1
6052 vmovdqa YMMWORD[(448-256-128)+rbx],ymm10
6053 vpaddd ymm4,ymm4,ymm10
6054 vpxor ymm11,ymm11,YMMWORD[((224-128))+rax]
6055 vpsrld ymm8,ymm0,27
6056 vpxor ymm5,ymm5,ymm2
6057 vpxor ymm11,ymm11,ymm13
6058
6059 vpslld ymm6,ymm1,30
6060 vpor ymm7,ymm7,ymm8
6061 vpaddd ymm4,ymm4,ymm5
6062 vpsrld ymm9,ymm11,31
6063 vpaddd ymm11,ymm11,ymm11
6064
6065 vpsrld ymm1,ymm1,2
6066 vpaddd ymm4,ymm4,ymm7
6067 vpor ymm11,ymm11,ymm9
6068 vpor ymm1,ymm1,ymm6
6069 vpxor ymm12,ymm12,ymm14
6070 vmovdqa ymm14,YMMWORD[((64-128))+rax]
6071
6072 vpslld ymm7,ymm4,5
6073 vpaddd ymm3,ymm3,ymm15
6074 vpxor ymm5,ymm2,ymm0
6075 vmovdqa YMMWORD[(480-256-128)+rbx],ymm11
6076 vpaddd ymm3,ymm3,ymm11
6077 vpxor ymm12,ymm12,YMMWORD[((256-256-128))+rbx]
6078 vpsrld ymm8,ymm4,27
6079 vpxor ymm5,ymm5,ymm1
6080 vpxor ymm12,ymm12,ymm14
6081
6082 vpslld ymm6,ymm0,30
6083 vpor ymm7,ymm7,ymm8
6084 vpaddd ymm3,ymm3,ymm5
6085 vpsrld ymm9,ymm12,31
6086 vpaddd ymm12,ymm12,ymm12
6087
6088 vpsrld ymm0,ymm0,2
6089 vpaddd ymm3,ymm3,ymm7
6090 vpor ymm12,ymm12,ymm9
6091 vpor ymm0,ymm0,ymm6
6092 vpxor ymm13,ymm13,ymm10
6093 vmovdqa ymm10,YMMWORD[((96-128))+rax]
6094
6095 vpslld ymm7,ymm3,5
6096 vpaddd ymm2,ymm2,ymm15
6097 vpxor ymm5,ymm1,ymm4
6098 vmovdqa YMMWORD[(0-128)+rax],ymm12
6099 vpaddd ymm2,ymm2,ymm12
6100 vpxor ymm13,ymm13,YMMWORD[((288-256-128))+rbx]
6101 vpsrld ymm8,ymm3,27
6102 vpxor ymm5,ymm5,ymm0
6103 vpxor ymm13,ymm13,ymm10
6104
6105 vpslld ymm6,ymm4,30
6106 vpor ymm7,ymm7,ymm8
6107 vpaddd ymm2,ymm2,ymm5
6108 vpsrld ymm9,ymm13,31
6109 vpaddd ymm13,ymm13,ymm13
6110
6111 vpsrld ymm4,ymm4,2
6112 vpaddd ymm2,ymm2,ymm7
6113 vpor ymm13,ymm13,ymm9
6114 vpor ymm4,ymm4,ymm6
6115 vpxor ymm14,ymm14,ymm11
6116 vmovdqa ymm11,YMMWORD[((128-128))+rax]
6117
6118 vpslld ymm7,ymm2,5
6119 vpaddd ymm1,ymm1,ymm15
6120 vpxor ymm5,ymm0,ymm3
6121 vmovdqa YMMWORD[(32-128)+rax],ymm13
6122 vpaddd ymm1,ymm1,ymm13
6123 vpxor ymm14,ymm14,YMMWORD[((320-256-128))+rbx]
6124 vpsrld ymm8,ymm2,27
6125 vpxor ymm5,ymm5,ymm4
6126 vpxor ymm14,ymm14,ymm11
6127
6128 vpslld ymm6,ymm3,30
6129 vpor ymm7,ymm7,ymm8
6130 vpaddd ymm1,ymm1,ymm5
6131 vpsrld ymm9,ymm14,31
6132 vpaddd ymm14,ymm14,ymm14
6133
6134 vpsrld ymm3,ymm3,2
6135 vpaddd ymm1,ymm1,ymm7
6136 vpor ymm14,ymm14,ymm9
6137 vpor ymm3,ymm3,ymm6
6138 vpxor ymm10,ymm10,ymm12
6139 vmovdqa ymm12,YMMWORD[((160-128))+rax]
6140
6141 vpslld ymm7,ymm1,5
6142 vpaddd ymm0,ymm0,ymm15
6143 vpxor ymm5,ymm4,ymm2
6144 vmovdqa YMMWORD[(64-128)+rax],ymm14
6145 vpaddd ymm0,ymm0,ymm14
6146 vpxor ymm10,ymm10,YMMWORD[((352-256-128))+rbx]
6147 vpsrld ymm8,ymm1,27
6148 vpxor ymm5,ymm5,ymm3
6149 vpxor ymm10,ymm10,ymm12
6150
6151 vpslld ymm6,ymm2,30
6152 vpor ymm7,ymm7,ymm8
6153 vpaddd ymm0,ymm0,ymm5
6154 vpsrld ymm9,ymm10,31
6155 vpaddd ymm10,ymm10,ymm10
6156
6157 vpsrld ymm2,ymm2,2
6158 vpaddd ymm0,ymm0,ymm7
6159 vpor ymm10,ymm10,ymm9
6160 vpor ymm2,ymm2,ymm6
6161 vpxor ymm11,ymm11,ymm13
6162 vmovdqa ymm13,YMMWORD[((192-128))+rax]
6163
6164 vpslld ymm7,ymm0,5
6165 vpaddd ymm4,ymm4,ymm15
6166 vpxor ymm5,ymm3,ymm1
6167 vmovdqa YMMWORD[(96-128)+rax],ymm10
6168 vpaddd ymm4,ymm4,ymm10
6169 vpxor ymm11,ymm11,YMMWORD[((384-256-128))+rbx]
6170 vpsrld ymm8,ymm0,27
6171 vpxor ymm5,ymm5,ymm2
6172 vpxor ymm11,ymm11,ymm13
6173
6174 vpslld ymm6,ymm1,30
6175 vpor ymm7,ymm7,ymm8
6176 vpaddd ymm4,ymm4,ymm5
6177 vpsrld ymm9,ymm11,31
6178 vpaddd ymm11,ymm11,ymm11
6179
6180 vpsrld ymm1,ymm1,2
6181 vpaddd ymm4,ymm4,ymm7
6182 vpor ymm11,ymm11,ymm9
6183 vpor ymm1,ymm1,ymm6
6184 vpxor ymm12,ymm12,ymm14
6185 vmovdqa ymm14,YMMWORD[((224-128))+rax]
6186
6187 vpslld ymm7,ymm4,5
6188 vpaddd ymm3,ymm3,ymm15
6189 vpxor ymm5,ymm2,ymm0
6190 vmovdqa YMMWORD[(128-128)+rax],ymm11
6191 vpaddd ymm3,ymm3,ymm11
6192 vpxor ymm12,ymm12,YMMWORD[((416-256-128))+rbx]
6193 vpsrld ymm8,ymm4,27
6194 vpxor ymm5,ymm5,ymm1
6195 vpxor ymm12,ymm12,ymm14
6196
6197 vpslld ymm6,ymm0,30
6198 vpor ymm7,ymm7,ymm8
6199 vpaddd ymm3,ymm3,ymm5
6200 vpsrld ymm9,ymm12,31
6201 vpaddd ymm12,ymm12,ymm12
6202
6203 vpsrld ymm0,ymm0,2
6204 vpaddd ymm3,ymm3,ymm7
6205 vpor ymm12,ymm12,ymm9
6206 vpor ymm0,ymm0,ymm6
6207 vpxor ymm13,ymm13,ymm10
6208 vmovdqa ymm10,YMMWORD[((256-256-128))+rbx]
6209
6210 vpslld ymm7,ymm3,5
6211 vpaddd ymm2,ymm2,ymm15
6212 vpxor ymm5,ymm1,ymm4
6213 vmovdqa YMMWORD[(160-128)+rax],ymm12
6214 vpaddd ymm2,ymm2,ymm12
6215 vpxor ymm13,ymm13,YMMWORD[((448-256-128))+rbx]
6216 vpsrld ymm8,ymm3,27
6217 vpxor ymm5,ymm5,ymm0
6218 vpxor ymm13,ymm13,ymm10
6219
6220 vpslld ymm6,ymm4,30
6221 vpor ymm7,ymm7,ymm8
6222 vpaddd ymm2,ymm2,ymm5
6223 vpsrld ymm9,ymm13,31
6224 vpaddd ymm13,ymm13,ymm13
6225
6226 vpsrld ymm4,ymm4,2
6227 vpaddd ymm2,ymm2,ymm7
6228 vpor ymm13,ymm13,ymm9
6229 vpor ymm4,ymm4,ymm6
6230 vpxor ymm14,ymm14,ymm11
6231 vmovdqa ymm11,YMMWORD[((288-256-128))+rbx]
6232
6233 vpslld ymm7,ymm2,5
6234 vpaddd ymm1,ymm1,ymm15
6235 vpxor ymm5,ymm0,ymm3
6236 vmovdqa YMMWORD[(192-128)+rax],ymm13
6237 vpaddd ymm1,ymm1,ymm13
6238 vpxor ymm14,ymm14,YMMWORD[((480-256-128))+rbx]
6239 vpsrld ymm8,ymm2,27
6240 vpxor ymm5,ymm5,ymm4
6241 vpxor ymm14,ymm14,ymm11
6242
6243 vpslld ymm6,ymm3,30
6244 vpor ymm7,ymm7,ymm8
6245 vpaddd ymm1,ymm1,ymm5
6246 vpsrld ymm9,ymm14,31
6247 vpaddd ymm14,ymm14,ymm14
6248
6249 vpsrld ymm3,ymm3,2
6250 vpaddd ymm1,ymm1,ymm7
6251 vpor ymm14,ymm14,ymm9
6252 vpor ymm3,ymm3,ymm6
6253 vpxor ymm10,ymm10,ymm12
6254 vmovdqa ymm12,YMMWORD[((320-256-128))+rbx]
6255
6256 vpslld ymm7,ymm1,5
6257 vpaddd ymm0,ymm0,ymm15
6258 vpxor ymm5,ymm4,ymm2
6259 vmovdqa YMMWORD[(224-128)+rax],ymm14
6260 vpaddd ymm0,ymm0,ymm14
6261 vpxor ymm10,ymm10,YMMWORD[((0-128))+rax]
6262 vpsrld ymm8,ymm1,27
6263 vpxor ymm5,ymm5,ymm3
6264 vpxor ymm10,ymm10,ymm12
6265
6266 vpslld ymm6,ymm2,30
6267 vpor ymm7,ymm7,ymm8
6268 vpaddd ymm0,ymm0,ymm5
6269 vpsrld ymm9,ymm10,31
6270 vpaddd ymm10,ymm10,ymm10
6271
6272 vpsrld ymm2,ymm2,2
6273 vpaddd ymm0,ymm0,ymm7
6274 vpor ymm10,ymm10,ymm9
6275 vpor ymm2,ymm2,ymm6
6276 vmovdqa ymm15,YMMWORD[32+rbp]
6277 vpxor ymm11,ymm11,ymm13
6278 vmovdqa ymm13,YMMWORD[((352-256-128))+rbx]
6279
6280 vpaddd ymm4,ymm4,ymm15
6281 vpslld ymm7,ymm0,5
6282 vpand ymm6,ymm3,ymm2
6283 vpxor ymm11,ymm11,YMMWORD[((32-128))+rax]
6284
6285 vpaddd ymm4,ymm4,ymm6
6286 vpsrld ymm8,ymm0,27
6287 vpxor ymm5,ymm3,ymm2
6288 vpxor ymm11,ymm11,ymm13
6289
6290 vmovdqu YMMWORD[(256-256-128)+rbx],ymm10
6291 vpaddd ymm4,ymm4,ymm10
6292 vpor ymm7,ymm7,ymm8
6293 vpsrld ymm9,ymm11,31
6294 vpand ymm5,ymm5,ymm1
6295 vpaddd ymm11,ymm11,ymm11
6296
6297 vpslld ymm6,ymm1,30
6298 vpaddd ymm4,ymm4,ymm5
6299
6300 vpsrld ymm1,ymm1,2
6301 vpaddd ymm4,ymm4,ymm7
6302 vpor ymm11,ymm11,ymm9
6303 vpor ymm1,ymm1,ymm6
6304 vpxor ymm12,ymm12,ymm14
6305 vmovdqa ymm14,YMMWORD[((384-256-128))+rbx]
6306
6307 vpaddd ymm3,ymm3,ymm15
6308 vpslld ymm7,ymm4,5
6309 vpand ymm6,ymm2,ymm1
6310 vpxor ymm12,ymm12,YMMWORD[((64-128))+rax]
6311
6312 vpaddd ymm3,ymm3,ymm6
6313 vpsrld ymm8,ymm4,27
6314 vpxor ymm5,ymm2,ymm1
6315 vpxor ymm12,ymm12,ymm14
6316
6317 vmovdqu YMMWORD[(288-256-128)+rbx],ymm11
6318 vpaddd ymm3,ymm3,ymm11
6319 vpor ymm7,ymm7,ymm8
6320 vpsrld ymm9,ymm12,31
6321 vpand ymm5,ymm5,ymm0
6322 vpaddd ymm12,ymm12,ymm12
6323
6324 vpslld ymm6,ymm0,30
6325 vpaddd ymm3,ymm3,ymm5
6326
6327 vpsrld ymm0,ymm0,2
6328 vpaddd ymm3,ymm3,ymm7
6329 vpor ymm12,ymm12,ymm9
6330 vpor ymm0,ymm0,ymm6
6331 vpxor ymm13,ymm13,ymm10
6332 vmovdqa ymm10,YMMWORD[((416-256-128))+rbx]
6333
6334 vpaddd ymm2,ymm2,ymm15
6335 vpslld ymm7,ymm3,5
6336 vpand ymm6,ymm1,ymm0
6337 vpxor ymm13,ymm13,YMMWORD[((96-128))+rax]
6338
6339 vpaddd ymm2,ymm2,ymm6
6340 vpsrld ymm8,ymm3,27
6341 vpxor ymm5,ymm1,ymm0
6342 vpxor ymm13,ymm13,ymm10
6343
6344 vmovdqu YMMWORD[(320-256-128)+rbx],ymm12
6345 vpaddd ymm2,ymm2,ymm12
6346 vpor ymm7,ymm7,ymm8
6347 vpsrld ymm9,ymm13,31
6348 vpand ymm5,ymm5,ymm4
6349 vpaddd ymm13,ymm13,ymm13
6350
6351 vpslld ymm6,ymm4,30
6352 vpaddd ymm2,ymm2,ymm5
6353
6354 vpsrld ymm4,ymm4,2
6355 vpaddd ymm2,ymm2,ymm7
6356 vpor ymm13,ymm13,ymm9
6357 vpor ymm4,ymm4,ymm6
6358 vpxor ymm14,ymm14,ymm11
6359 vmovdqa ymm11,YMMWORD[((448-256-128))+rbx]
6360
6361 vpaddd ymm1,ymm1,ymm15
6362 vpslld ymm7,ymm2,5
6363 vpand ymm6,ymm0,ymm4
6364 vpxor ymm14,ymm14,YMMWORD[((128-128))+rax]
6365
6366 vpaddd ymm1,ymm1,ymm6
6367 vpsrld ymm8,ymm2,27
6368 vpxor ymm5,ymm0,ymm4
6369 vpxor ymm14,ymm14,ymm11
6370
6371 vmovdqu YMMWORD[(352-256-128)+rbx],ymm13
6372 vpaddd ymm1,ymm1,ymm13
6373 vpor ymm7,ymm7,ymm8
6374 vpsrld ymm9,ymm14,31
6375 vpand ymm5,ymm5,ymm3
6376 vpaddd ymm14,ymm14,ymm14
6377
6378 vpslld ymm6,ymm3,30
6379 vpaddd ymm1,ymm1,ymm5
6380
6381 vpsrld ymm3,ymm3,2
6382 vpaddd ymm1,ymm1,ymm7
6383 vpor ymm14,ymm14,ymm9
6384 vpor ymm3,ymm3,ymm6
6385 vpxor ymm10,ymm10,ymm12
6386 vmovdqa ymm12,YMMWORD[((480-256-128))+rbx]
6387
6388 vpaddd ymm0,ymm0,ymm15
6389 vpslld ymm7,ymm1,5
6390 vpand ymm6,ymm4,ymm3
6391 vpxor ymm10,ymm10,YMMWORD[((160-128))+rax]
6392
6393 vpaddd ymm0,ymm0,ymm6
6394 vpsrld ymm8,ymm1,27
6395 vpxor ymm5,ymm4,ymm3
6396 vpxor ymm10,ymm10,ymm12
6397
6398 vmovdqu YMMWORD[(384-256-128)+rbx],ymm14
6399 vpaddd ymm0,ymm0,ymm14
6400 vpor ymm7,ymm7,ymm8
6401 vpsrld ymm9,ymm10,31
6402 vpand ymm5,ymm5,ymm2
6403 vpaddd ymm10,ymm10,ymm10
6404
6405 vpslld ymm6,ymm2,30
6406 vpaddd ymm0,ymm0,ymm5
6407
6408 vpsrld ymm2,ymm2,2
6409 vpaddd ymm0,ymm0,ymm7
6410 vpor ymm10,ymm10,ymm9
6411 vpor ymm2,ymm2,ymm6
6412 vpxor ymm11,ymm11,ymm13
6413 vmovdqa ymm13,YMMWORD[((0-128))+rax]
6414
6415 vpaddd ymm4,ymm4,ymm15
6416 vpslld ymm7,ymm0,5
6417 vpand ymm6,ymm3,ymm2
6418 vpxor ymm11,ymm11,YMMWORD[((192-128))+rax]
6419
6420 vpaddd ymm4,ymm4,ymm6
6421 vpsrld ymm8,ymm0,27
6422 vpxor ymm5,ymm3,ymm2
6423 vpxor ymm11,ymm11,ymm13
6424
6425 vmovdqu YMMWORD[(416-256-128)+rbx],ymm10
6426 vpaddd ymm4,ymm4,ymm10
6427 vpor ymm7,ymm7,ymm8
6428 vpsrld ymm9,ymm11,31
6429 vpand ymm5,ymm5,ymm1
6430 vpaddd ymm11,ymm11,ymm11
6431
6432 vpslld ymm6,ymm1,30
6433 vpaddd ymm4,ymm4,ymm5
6434
6435 vpsrld ymm1,ymm1,2
6436 vpaddd ymm4,ymm4,ymm7
6437 vpor ymm11,ymm11,ymm9
6438 vpor ymm1,ymm1,ymm6
6439 vpxor ymm12,ymm12,ymm14
6440 vmovdqa ymm14,YMMWORD[((32-128))+rax]
6441
6442 vpaddd ymm3,ymm3,ymm15
6443 vpslld ymm7,ymm4,5
6444 vpand ymm6,ymm2,ymm1
6445 vpxor ymm12,ymm12,YMMWORD[((224-128))+rax]
6446
6447 vpaddd ymm3,ymm3,ymm6
6448 vpsrld ymm8,ymm4,27
6449 vpxor ymm5,ymm2,ymm1
6450 vpxor ymm12,ymm12,ymm14
6451
6452 vmovdqu YMMWORD[(448-256-128)+rbx],ymm11
6453 vpaddd ymm3,ymm3,ymm11
6454 vpor ymm7,ymm7,ymm8
6455 vpsrld ymm9,ymm12,31
6456 vpand ymm5,ymm5,ymm0
6457 vpaddd ymm12,ymm12,ymm12
6458
6459 vpslld ymm6,ymm0,30
6460 vpaddd ymm3,ymm3,ymm5
6461
6462 vpsrld ymm0,ymm0,2
6463 vpaddd ymm3,ymm3,ymm7
6464 vpor ymm12,ymm12,ymm9
6465 vpor ymm0,ymm0,ymm6
6466 vpxor ymm13,ymm13,ymm10
6467 vmovdqa ymm10,YMMWORD[((64-128))+rax]
6468
6469 vpaddd ymm2,ymm2,ymm15
6470 vpslld ymm7,ymm3,5
6471 vpand ymm6,ymm1,ymm0
6472 vpxor ymm13,ymm13,YMMWORD[((256-256-128))+rbx]
6473
6474 vpaddd ymm2,ymm2,ymm6
6475 vpsrld ymm8,ymm3,27
6476 vpxor ymm5,ymm1,ymm0
6477 vpxor ymm13,ymm13,ymm10
6478
6479 vmovdqu YMMWORD[(480-256-128)+rbx],ymm12
6480 vpaddd ymm2,ymm2,ymm12
6481 vpor ymm7,ymm7,ymm8
6482 vpsrld ymm9,ymm13,31
6483 vpand ymm5,ymm5,ymm4
6484 vpaddd ymm13,ymm13,ymm13
6485
6486 vpslld ymm6,ymm4,30
6487 vpaddd ymm2,ymm2,ymm5
6488
6489 vpsrld ymm4,ymm4,2
6490 vpaddd ymm2,ymm2,ymm7
6491 vpor ymm13,ymm13,ymm9
6492 vpor ymm4,ymm4,ymm6
6493 vpxor ymm14,ymm14,ymm11
6494 vmovdqa ymm11,YMMWORD[((96-128))+rax]
6495
6496 vpaddd ymm1,ymm1,ymm15
6497 vpslld ymm7,ymm2,5
6498 vpand ymm6,ymm0,ymm4
6499 vpxor ymm14,ymm14,YMMWORD[((288-256-128))+rbx]
6500
6501 vpaddd ymm1,ymm1,ymm6
6502 vpsrld ymm8,ymm2,27
6503 vpxor ymm5,ymm0,ymm4
6504 vpxor ymm14,ymm14,ymm11
6505
6506 vmovdqu YMMWORD[(0-128)+rax],ymm13
6507 vpaddd ymm1,ymm1,ymm13
6508 vpor ymm7,ymm7,ymm8
6509 vpsrld ymm9,ymm14,31
6510 vpand ymm5,ymm5,ymm3
6511 vpaddd ymm14,ymm14,ymm14
6512
6513 vpslld ymm6,ymm3,30
6514 vpaddd ymm1,ymm1,ymm5
6515
6516 vpsrld ymm3,ymm3,2
6517 vpaddd ymm1,ymm1,ymm7
6518 vpor ymm14,ymm14,ymm9
6519 vpor ymm3,ymm3,ymm6
6520 vpxor ymm10,ymm10,ymm12
6521 vmovdqa ymm12,YMMWORD[((128-128))+rax]
6522
6523 vpaddd ymm0,ymm0,ymm15
6524 vpslld ymm7,ymm1,5
6525 vpand ymm6,ymm4,ymm3
6526 vpxor ymm10,ymm10,YMMWORD[((320-256-128))+rbx]
6527
6528 vpaddd ymm0,ymm0,ymm6
6529 vpsrld ymm8,ymm1,27
6530 vpxor ymm5,ymm4,ymm3
6531 vpxor ymm10,ymm10,ymm12
6532
6533 vmovdqu YMMWORD[(32-128)+rax],ymm14
6534 vpaddd ymm0,ymm0,ymm14
6535 vpor ymm7,ymm7,ymm8
6536 vpsrld ymm9,ymm10,31
6537 vpand ymm5,ymm5,ymm2
6538 vpaddd ymm10,ymm10,ymm10
6539
6540 vpslld ymm6,ymm2,30
6541 vpaddd ymm0,ymm0,ymm5
6542
6543 vpsrld ymm2,ymm2,2
6544 vpaddd ymm0,ymm0,ymm7
6545 vpor ymm10,ymm10,ymm9
6546 vpor ymm2,ymm2,ymm6
6547 vpxor ymm11,ymm11,ymm13
6548 vmovdqa ymm13,YMMWORD[((160-128))+rax]
6549
6550 vpaddd ymm4,ymm4,ymm15
6551 vpslld ymm7,ymm0,5
6552 vpand ymm6,ymm3,ymm2
6553 vpxor ymm11,ymm11,YMMWORD[((352-256-128))+rbx]
6554
6555 vpaddd ymm4,ymm4,ymm6
6556 vpsrld ymm8,ymm0,27
6557 vpxor ymm5,ymm3,ymm2
6558 vpxor ymm11,ymm11,ymm13
6559
6560 vmovdqu YMMWORD[(64-128)+rax],ymm10
6561 vpaddd ymm4,ymm4,ymm10
6562 vpor ymm7,ymm7,ymm8
6563 vpsrld ymm9,ymm11,31
6564 vpand ymm5,ymm5,ymm1
6565 vpaddd ymm11,ymm11,ymm11
6566
6567 vpslld ymm6,ymm1,30
6568 vpaddd ymm4,ymm4,ymm5
6569
6570 vpsrld ymm1,ymm1,2
6571 vpaddd ymm4,ymm4,ymm7
6572 vpor ymm11,ymm11,ymm9
6573 vpor ymm1,ymm1,ymm6
6574 vpxor ymm12,ymm12,ymm14
6575 vmovdqa ymm14,YMMWORD[((192-128))+rax]
6576
6577 vpaddd ymm3,ymm3,ymm15
6578 vpslld ymm7,ymm4,5
6579 vpand ymm6,ymm2,ymm1
6580 vpxor ymm12,ymm12,YMMWORD[((384-256-128))+rbx]
6581
6582 vpaddd ymm3,ymm3,ymm6
6583 vpsrld ymm8,ymm4,27
6584 vpxor ymm5,ymm2,ymm1
6585 vpxor ymm12,ymm12,ymm14
6586
6587 vmovdqu YMMWORD[(96-128)+rax],ymm11
6588 vpaddd ymm3,ymm3,ymm11
6589 vpor ymm7,ymm7,ymm8
6590 vpsrld ymm9,ymm12,31
6591 vpand ymm5,ymm5,ymm0
6592 vpaddd ymm12,ymm12,ymm12
6593
6594 vpslld ymm6,ymm0,30
6595 vpaddd ymm3,ymm3,ymm5
6596
6597 vpsrld ymm0,ymm0,2
6598 vpaddd ymm3,ymm3,ymm7
6599 vpor ymm12,ymm12,ymm9
6600 vpor ymm0,ymm0,ymm6
6601 vpxor ymm13,ymm13,ymm10
6602 vmovdqa ymm10,YMMWORD[((224-128))+rax]
6603
6604 vpaddd ymm2,ymm2,ymm15
6605 vpslld ymm7,ymm3,5
6606 vpand ymm6,ymm1,ymm0
6607 vpxor ymm13,ymm13,YMMWORD[((416-256-128))+rbx]
6608
6609 vpaddd ymm2,ymm2,ymm6
6610 vpsrld ymm8,ymm3,27
6611 vpxor ymm5,ymm1,ymm0
6612 vpxor ymm13,ymm13,ymm10
6613
6614 vmovdqu YMMWORD[(128-128)+rax],ymm12
6615 vpaddd ymm2,ymm2,ymm12
6616 vpor ymm7,ymm7,ymm8
6617 vpsrld ymm9,ymm13,31
6618 vpand ymm5,ymm5,ymm4
6619 vpaddd ymm13,ymm13,ymm13
6620
6621 vpslld ymm6,ymm4,30
6622 vpaddd ymm2,ymm2,ymm5
6623
6624 vpsrld ymm4,ymm4,2
6625 vpaddd ymm2,ymm2,ymm7
6626 vpor ymm13,ymm13,ymm9
6627 vpor ymm4,ymm4,ymm6
6628 vpxor ymm14,ymm14,ymm11
6629 vmovdqa ymm11,YMMWORD[((256-256-128))+rbx]
6630
6631 vpaddd ymm1,ymm1,ymm15
6632 vpslld ymm7,ymm2,5
6633 vpand ymm6,ymm0,ymm4
6634 vpxor ymm14,ymm14,YMMWORD[((448-256-128))+rbx]
6635
6636 vpaddd ymm1,ymm1,ymm6
6637 vpsrld ymm8,ymm2,27
6638 vpxor ymm5,ymm0,ymm4
6639 vpxor ymm14,ymm14,ymm11
6640
6641 vmovdqu YMMWORD[(160-128)+rax],ymm13
6642 vpaddd ymm1,ymm1,ymm13
6643 vpor ymm7,ymm7,ymm8
6644 vpsrld ymm9,ymm14,31
6645 vpand ymm5,ymm5,ymm3
6646 vpaddd ymm14,ymm14,ymm14
6647
6648 vpslld ymm6,ymm3,30
6649 vpaddd ymm1,ymm1,ymm5
6650
6651 vpsrld ymm3,ymm3,2
6652 vpaddd ymm1,ymm1,ymm7
6653 vpor ymm14,ymm14,ymm9
6654 vpor ymm3,ymm3,ymm6
6655 vpxor ymm10,ymm10,ymm12
6656 vmovdqa ymm12,YMMWORD[((288-256-128))+rbx]
6657
6658 vpaddd ymm0,ymm0,ymm15
6659 vpslld ymm7,ymm1,5
6660 vpand ymm6,ymm4,ymm3
6661 vpxor ymm10,ymm10,YMMWORD[((480-256-128))+rbx]
6662
6663 vpaddd ymm0,ymm0,ymm6
6664 vpsrld ymm8,ymm1,27
6665 vpxor ymm5,ymm4,ymm3
6666 vpxor ymm10,ymm10,ymm12
6667
6668 vmovdqu YMMWORD[(192-128)+rax],ymm14
6669 vpaddd ymm0,ymm0,ymm14
6670 vpor ymm7,ymm7,ymm8
6671 vpsrld ymm9,ymm10,31
6672 vpand ymm5,ymm5,ymm2
6673 vpaddd ymm10,ymm10,ymm10
6674
6675 vpslld ymm6,ymm2,30
6676 vpaddd ymm0,ymm0,ymm5
6677
6678 vpsrld ymm2,ymm2,2
6679 vpaddd ymm0,ymm0,ymm7
6680 vpor ymm10,ymm10,ymm9
6681 vpor ymm2,ymm2,ymm6
6682 vpxor ymm11,ymm11,ymm13
6683 vmovdqa ymm13,YMMWORD[((320-256-128))+rbx]
6684
6685 vpaddd ymm4,ymm4,ymm15
6686 vpslld ymm7,ymm0,5
6687 vpand ymm6,ymm3,ymm2
6688 vpxor ymm11,ymm11,YMMWORD[((0-128))+rax]
6689
6690 vpaddd ymm4,ymm4,ymm6
6691 vpsrld ymm8,ymm0,27
6692 vpxor ymm5,ymm3,ymm2
6693 vpxor ymm11,ymm11,ymm13
6694
6695 vmovdqu YMMWORD[(224-128)+rax],ymm10
6696 vpaddd ymm4,ymm4,ymm10
6697 vpor ymm7,ymm7,ymm8
6698 vpsrld ymm9,ymm11,31
6699 vpand ymm5,ymm5,ymm1
6700 vpaddd ymm11,ymm11,ymm11
6701
6702 vpslld ymm6,ymm1,30
6703 vpaddd ymm4,ymm4,ymm5
6704
6705 vpsrld ymm1,ymm1,2
6706 vpaddd ymm4,ymm4,ymm7
6707 vpor ymm11,ymm11,ymm9
6708 vpor ymm1,ymm1,ymm6
6709 vpxor ymm12,ymm12,ymm14
6710 vmovdqa ymm14,YMMWORD[((352-256-128))+rbx]
6711
6712 vpaddd ymm3,ymm3,ymm15
6713 vpslld ymm7,ymm4,5
6714 vpand ymm6,ymm2,ymm1
6715 vpxor ymm12,ymm12,YMMWORD[((32-128))+rax]
6716
6717 vpaddd ymm3,ymm3,ymm6
6718 vpsrld ymm8,ymm4,27
6719 vpxor ymm5,ymm2,ymm1
6720 vpxor ymm12,ymm12,ymm14
6721
6722 vmovdqu YMMWORD[(256-256-128)+rbx],ymm11
6723 vpaddd ymm3,ymm3,ymm11
6724 vpor ymm7,ymm7,ymm8
6725 vpsrld ymm9,ymm12,31
6726 vpand ymm5,ymm5,ymm0
6727 vpaddd ymm12,ymm12,ymm12
6728
6729 vpslld ymm6,ymm0,30
6730 vpaddd ymm3,ymm3,ymm5
6731
6732 vpsrld ymm0,ymm0,2
6733 vpaddd ymm3,ymm3,ymm7
6734 vpor ymm12,ymm12,ymm9
6735 vpor ymm0,ymm0,ymm6
6736 vpxor ymm13,ymm13,ymm10
6737 vmovdqa ymm10,YMMWORD[((384-256-128))+rbx]
6738
6739 vpaddd ymm2,ymm2,ymm15
6740 vpslld ymm7,ymm3,5
6741 vpand ymm6,ymm1,ymm0
6742 vpxor ymm13,ymm13,YMMWORD[((64-128))+rax]
6743
6744 vpaddd ymm2,ymm2,ymm6
6745 vpsrld ymm8,ymm3,27
6746 vpxor ymm5,ymm1,ymm0
6747 vpxor ymm13,ymm13,ymm10
6748
6749 vmovdqu YMMWORD[(288-256-128)+rbx],ymm12
6750 vpaddd ymm2,ymm2,ymm12
6751 vpor ymm7,ymm7,ymm8
6752 vpsrld ymm9,ymm13,31
6753 vpand ymm5,ymm5,ymm4
6754 vpaddd ymm13,ymm13,ymm13
6755
6756 vpslld ymm6,ymm4,30
6757 vpaddd ymm2,ymm2,ymm5
6758
6759 vpsrld ymm4,ymm4,2
6760 vpaddd ymm2,ymm2,ymm7
6761 vpor ymm13,ymm13,ymm9
6762 vpor ymm4,ymm4,ymm6
6763 vpxor ymm14,ymm14,ymm11
6764 vmovdqa ymm11,YMMWORD[((416-256-128))+rbx]
6765
6766 vpaddd ymm1,ymm1,ymm15
6767 vpslld ymm7,ymm2,5
6768 vpand ymm6,ymm0,ymm4
6769 vpxor ymm14,ymm14,YMMWORD[((96-128))+rax]
6770
6771 vpaddd ymm1,ymm1,ymm6
6772 vpsrld ymm8,ymm2,27
6773 vpxor ymm5,ymm0,ymm4
6774 vpxor ymm14,ymm14,ymm11
6775
6776 vmovdqu YMMWORD[(320-256-128)+rbx],ymm13
6777 vpaddd ymm1,ymm1,ymm13
6778 vpor ymm7,ymm7,ymm8
6779 vpsrld ymm9,ymm14,31
6780 vpand ymm5,ymm5,ymm3
6781 vpaddd ymm14,ymm14,ymm14
6782
6783 vpslld ymm6,ymm3,30
6784 vpaddd ymm1,ymm1,ymm5
6785
6786 vpsrld ymm3,ymm3,2
6787 vpaddd ymm1,ymm1,ymm7
6788 vpor ymm14,ymm14,ymm9
6789 vpor ymm3,ymm3,ymm6
6790 vpxor ymm10,ymm10,ymm12
6791 vmovdqa ymm12,YMMWORD[((448-256-128))+rbx]
6792
6793 vpaddd ymm0,ymm0,ymm15
6794 vpslld ymm7,ymm1,5
6795 vpand ymm6,ymm4,ymm3
6796 vpxor ymm10,ymm10,YMMWORD[((128-128))+rax]
6797
6798 vpaddd ymm0,ymm0,ymm6
6799 vpsrld ymm8,ymm1,27
6800 vpxor ymm5,ymm4,ymm3
6801 vpxor ymm10,ymm10,ymm12
6802
6803 vmovdqu YMMWORD[(352-256-128)+rbx],ymm14
6804 vpaddd ymm0,ymm0,ymm14
6805 vpor ymm7,ymm7,ymm8
6806 vpsrld ymm9,ymm10,31
6807 vpand ymm5,ymm5,ymm2
6808 vpaddd ymm10,ymm10,ymm10
6809
6810 vpslld ymm6,ymm2,30
6811 vpaddd ymm0,ymm0,ymm5
6812
6813 vpsrld ymm2,ymm2,2
6814 vpaddd ymm0,ymm0,ymm7
6815 vpor ymm10,ymm10,ymm9
6816 vpor ymm2,ymm2,ymm6
6817 vmovdqa ymm15,YMMWORD[64+rbp]
6818 vpxor ymm11,ymm11,ymm13
6819 vmovdqa ymm13,YMMWORD[((480-256-128))+rbx]
6820
6821 vpslld ymm7,ymm0,5
6822 vpaddd ymm4,ymm4,ymm15
6823 vpxor ymm5,ymm3,ymm1
6824 vmovdqa YMMWORD[(384-256-128)+rbx],ymm10
6825 vpaddd ymm4,ymm4,ymm10
6826 vpxor ymm11,ymm11,YMMWORD[((160-128))+rax]
6827 vpsrld ymm8,ymm0,27
6828 vpxor ymm5,ymm5,ymm2
6829 vpxor ymm11,ymm11,ymm13
6830
6831 vpslld ymm6,ymm1,30
6832 vpor ymm7,ymm7,ymm8
6833 vpaddd ymm4,ymm4,ymm5
6834 vpsrld ymm9,ymm11,31
6835 vpaddd ymm11,ymm11,ymm11
6836
6837 vpsrld ymm1,ymm1,2
6838 vpaddd ymm4,ymm4,ymm7
6839 vpor ymm11,ymm11,ymm9
6840 vpor ymm1,ymm1,ymm6
6841 vpxor ymm12,ymm12,ymm14
6842 vmovdqa ymm14,YMMWORD[((0-128))+rax]
6843
6844 vpslld ymm7,ymm4,5
6845 vpaddd ymm3,ymm3,ymm15
6846 vpxor ymm5,ymm2,ymm0
6847 vmovdqa YMMWORD[(416-256-128)+rbx],ymm11
6848 vpaddd ymm3,ymm3,ymm11
6849 vpxor ymm12,ymm12,YMMWORD[((192-128))+rax]
6850 vpsrld ymm8,ymm4,27
6851 vpxor ymm5,ymm5,ymm1
6852 vpxor ymm12,ymm12,ymm14
6853
6854 vpslld ymm6,ymm0,30
6855 vpor ymm7,ymm7,ymm8
6856 vpaddd ymm3,ymm3,ymm5
6857 vpsrld ymm9,ymm12,31
6858 vpaddd ymm12,ymm12,ymm12
6859
6860 vpsrld ymm0,ymm0,2
6861 vpaddd ymm3,ymm3,ymm7
6862 vpor ymm12,ymm12,ymm9
6863 vpor ymm0,ymm0,ymm6
6864 vpxor ymm13,ymm13,ymm10
6865 vmovdqa ymm10,YMMWORD[((32-128))+rax]
6866
6867 vpslld ymm7,ymm3,5
6868 vpaddd ymm2,ymm2,ymm15
6869 vpxor ymm5,ymm1,ymm4
6870 vmovdqa YMMWORD[(448-256-128)+rbx],ymm12
6871 vpaddd ymm2,ymm2,ymm12
6872 vpxor ymm13,ymm13,YMMWORD[((224-128))+rax]
6873 vpsrld ymm8,ymm3,27
6874 vpxor ymm5,ymm5,ymm0
6875 vpxor ymm13,ymm13,ymm10
6876
6877 vpslld ymm6,ymm4,30
6878 vpor ymm7,ymm7,ymm8
6879 vpaddd ymm2,ymm2,ymm5
6880 vpsrld ymm9,ymm13,31
6881 vpaddd ymm13,ymm13,ymm13
6882
6883 vpsrld ymm4,ymm4,2
6884 vpaddd ymm2,ymm2,ymm7
6885 vpor ymm13,ymm13,ymm9
6886 vpor ymm4,ymm4,ymm6
6887 vpxor ymm14,ymm14,ymm11
6888 vmovdqa ymm11,YMMWORD[((64-128))+rax]
6889
6890 vpslld ymm7,ymm2,5
6891 vpaddd ymm1,ymm1,ymm15
6892 vpxor ymm5,ymm0,ymm3
6893 vmovdqa YMMWORD[(480-256-128)+rbx],ymm13
6894 vpaddd ymm1,ymm1,ymm13
6895 vpxor ymm14,ymm14,YMMWORD[((256-256-128))+rbx]
6896 vpsrld ymm8,ymm2,27
6897 vpxor ymm5,ymm5,ymm4
6898 vpxor ymm14,ymm14,ymm11
6899
6900 vpslld ymm6,ymm3,30
6901 vpor ymm7,ymm7,ymm8
6902 vpaddd ymm1,ymm1,ymm5
6903 vpsrld ymm9,ymm14,31
6904 vpaddd ymm14,ymm14,ymm14
6905
6906 vpsrld ymm3,ymm3,2
6907 vpaddd ymm1,ymm1,ymm7
6908 vpor ymm14,ymm14,ymm9
6909 vpor ymm3,ymm3,ymm6
6910 vpxor ymm10,ymm10,ymm12
6911 vmovdqa ymm12,YMMWORD[((96-128))+rax]
6912
6913 vpslld ymm7,ymm1,5
6914 vpaddd ymm0,ymm0,ymm15
6915 vpxor ymm5,ymm4,ymm2
6916 vmovdqa YMMWORD[(0-128)+rax],ymm14
6917 vpaddd ymm0,ymm0,ymm14
6918 vpxor ymm10,ymm10,YMMWORD[((288-256-128))+rbx]
6919 vpsrld ymm8,ymm1,27
6920 vpxor ymm5,ymm5,ymm3
6921 vpxor ymm10,ymm10,ymm12
6922
6923 vpslld ymm6,ymm2,30
6924 vpor ymm7,ymm7,ymm8
6925 vpaddd ymm0,ymm0,ymm5
6926 vpsrld ymm9,ymm10,31
6927 vpaddd ymm10,ymm10,ymm10
6928
6929 vpsrld ymm2,ymm2,2
6930 vpaddd ymm0,ymm0,ymm7
6931 vpor ymm10,ymm10,ymm9
6932 vpor ymm2,ymm2,ymm6
6933 vpxor ymm11,ymm11,ymm13
6934 vmovdqa ymm13,YMMWORD[((128-128))+rax]
6935
6936 vpslld ymm7,ymm0,5
6937 vpaddd ymm4,ymm4,ymm15
6938 vpxor ymm5,ymm3,ymm1
6939 vmovdqa YMMWORD[(32-128)+rax],ymm10
6940 vpaddd ymm4,ymm4,ymm10
6941 vpxor ymm11,ymm11,YMMWORD[((320-256-128))+rbx]
6942 vpsrld ymm8,ymm0,27
6943 vpxor ymm5,ymm5,ymm2
6944 vpxor ymm11,ymm11,ymm13
6945
6946 vpslld ymm6,ymm1,30
6947 vpor ymm7,ymm7,ymm8
6948 vpaddd ymm4,ymm4,ymm5
6949 vpsrld ymm9,ymm11,31
6950 vpaddd ymm11,ymm11,ymm11
6951
6952 vpsrld ymm1,ymm1,2
6953 vpaddd ymm4,ymm4,ymm7
6954 vpor ymm11,ymm11,ymm9
6955 vpor ymm1,ymm1,ymm6
6956 vpxor ymm12,ymm12,ymm14
6957 vmovdqa ymm14,YMMWORD[((160-128))+rax]
6958
6959 vpslld ymm7,ymm4,5
6960 vpaddd ymm3,ymm3,ymm15
6961 vpxor ymm5,ymm2,ymm0
6962 vmovdqa YMMWORD[(64-128)+rax],ymm11
6963 vpaddd ymm3,ymm3,ymm11
6964 vpxor ymm12,ymm12,YMMWORD[((352-256-128))+rbx]
6965 vpsrld ymm8,ymm4,27
6966 vpxor ymm5,ymm5,ymm1
6967 vpxor ymm12,ymm12,ymm14
6968
6969 vpslld ymm6,ymm0,30
6970 vpor ymm7,ymm7,ymm8
6971 vpaddd ymm3,ymm3,ymm5
6972 vpsrld ymm9,ymm12,31
6973 vpaddd ymm12,ymm12,ymm12
6974
6975 vpsrld ymm0,ymm0,2
6976 vpaddd ymm3,ymm3,ymm7
6977 vpor ymm12,ymm12,ymm9
6978 vpor ymm0,ymm0,ymm6
6979 vpxor ymm13,ymm13,ymm10
6980 vmovdqa ymm10,YMMWORD[((192-128))+rax]
6981
6982 vpslld ymm7,ymm3,5
6983 vpaddd ymm2,ymm2,ymm15
6984 vpxor ymm5,ymm1,ymm4
6985 vmovdqa YMMWORD[(96-128)+rax],ymm12
6986 vpaddd ymm2,ymm2,ymm12
6987 vpxor ymm13,ymm13,YMMWORD[((384-256-128))+rbx]
6988 vpsrld ymm8,ymm3,27
6989 vpxor ymm5,ymm5,ymm0
6990 vpxor ymm13,ymm13,ymm10
6991
6992 vpslld ymm6,ymm4,30
6993 vpor ymm7,ymm7,ymm8
6994 vpaddd ymm2,ymm2,ymm5
6995 vpsrld ymm9,ymm13,31
6996 vpaddd ymm13,ymm13,ymm13
6997
6998 vpsrld ymm4,ymm4,2
6999 vpaddd ymm2,ymm2,ymm7
7000 vpor ymm13,ymm13,ymm9
7001 vpor ymm4,ymm4,ymm6
7002 vpxor ymm14,ymm14,ymm11
7003 vmovdqa ymm11,YMMWORD[((224-128))+rax]
7004
7005 vpslld ymm7,ymm2,5
7006 vpaddd ymm1,ymm1,ymm15
7007 vpxor ymm5,ymm0,ymm3
7008 vmovdqa YMMWORD[(128-128)+rax],ymm13
7009 vpaddd ymm1,ymm1,ymm13
7010 vpxor ymm14,ymm14,YMMWORD[((416-256-128))+rbx]
7011 vpsrld ymm8,ymm2,27
7012 vpxor ymm5,ymm5,ymm4
7013 vpxor ymm14,ymm14,ymm11
7014
7015 vpslld ymm6,ymm3,30
7016 vpor ymm7,ymm7,ymm8
7017 vpaddd ymm1,ymm1,ymm5
7018 vpsrld ymm9,ymm14,31
7019 vpaddd ymm14,ymm14,ymm14
7020
7021 vpsrld ymm3,ymm3,2
7022 vpaddd ymm1,ymm1,ymm7
7023 vpor ymm14,ymm14,ymm9
7024 vpor ymm3,ymm3,ymm6
7025 vpxor ymm10,ymm10,ymm12
7026 vmovdqa ymm12,YMMWORD[((256-256-128))+rbx]
7027
7028 vpslld ymm7,ymm1,5
7029 vpaddd ymm0,ymm0,ymm15
7030 vpxor ymm5,ymm4,ymm2
7031 vmovdqa YMMWORD[(160-128)+rax],ymm14
7032 vpaddd ymm0,ymm0,ymm14
7033 vpxor ymm10,ymm10,YMMWORD[((448-256-128))+rbx]
7034 vpsrld ymm8,ymm1,27
7035 vpxor ymm5,ymm5,ymm3
7036 vpxor ymm10,ymm10,ymm12
7037
7038 vpslld ymm6,ymm2,30
7039 vpor ymm7,ymm7,ymm8
7040 vpaddd ymm0,ymm0,ymm5
7041 vpsrld ymm9,ymm10,31
7042 vpaddd ymm10,ymm10,ymm10
7043
7044 vpsrld ymm2,ymm2,2
7045 vpaddd ymm0,ymm0,ymm7
7046 vpor ymm10,ymm10,ymm9
7047 vpor ymm2,ymm2,ymm6
7048 vpxor ymm11,ymm11,ymm13
7049 vmovdqa ymm13,YMMWORD[((288-256-128))+rbx]
7050
7051 vpslld ymm7,ymm0,5
7052 vpaddd ymm4,ymm4,ymm15
7053 vpxor ymm5,ymm3,ymm1
7054 vmovdqa YMMWORD[(192-128)+rax],ymm10
7055 vpaddd ymm4,ymm4,ymm10
7056 vpxor ymm11,ymm11,YMMWORD[((480-256-128))+rbx]
7057 vpsrld ymm8,ymm0,27
7058 vpxor ymm5,ymm5,ymm2
7059 vpxor ymm11,ymm11,ymm13
7060
7061 vpslld ymm6,ymm1,30
7062 vpor ymm7,ymm7,ymm8
7063 vpaddd ymm4,ymm4,ymm5
7064 vpsrld ymm9,ymm11,31
7065 vpaddd ymm11,ymm11,ymm11
7066
7067 vpsrld ymm1,ymm1,2
7068 vpaddd ymm4,ymm4,ymm7
7069 vpor ymm11,ymm11,ymm9
7070 vpor ymm1,ymm1,ymm6
7071 vpxor ymm12,ymm12,ymm14
7072 vmovdqa ymm14,YMMWORD[((320-256-128))+rbx]
7073
7074 vpslld ymm7,ymm4,5
7075 vpaddd ymm3,ymm3,ymm15
7076 vpxor ymm5,ymm2,ymm0
7077 vmovdqa YMMWORD[(224-128)+rax],ymm11
7078 vpaddd ymm3,ymm3,ymm11
7079 vpxor ymm12,ymm12,YMMWORD[((0-128))+rax]
7080 vpsrld ymm8,ymm4,27
7081 vpxor ymm5,ymm5,ymm1
7082 vpxor ymm12,ymm12,ymm14
7083
7084 vpslld ymm6,ymm0,30
7085 vpor ymm7,ymm7,ymm8
7086 vpaddd ymm3,ymm3,ymm5
7087 vpsrld ymm9,ymm12,31
7088 vpaddd ymm12,ymm12,ymm12
7089
7090 vpsrld ymm0,ymm0,2
7091 vpaddd ymm3,ymm3,ymm7
7092 vpor ymm12,ymm12,ymm9
7093 vpor ymm0,ymm0,ymm6
7094 vpxor ymm13,ymm13,ymm10
7095 vmovdqa ymm10,YMMWORD[((352-256-128))+rbx]
7096
7097 vpslld ymm7,ymm3,5
7098 vpaddd ymm2,ymm2,ymm15
7099 vpxor ymm5,ymm1,ymm4
7100 vpaddd ymm2,ymm2,ymm12
7101 vpxor ymm13,ymm13,YMMWORD[((32-128))+rax]
7102 vpsrld ymm8,ymm3,27
7103 vpxor ymm5,ymm5,ymm0
7104 vpxor ymm13,ymm13,ymm10
7105
7106 vpslld ymm6,ymm4,30
7107 vpor ymm7,ymm7,ymm8
7108 vpaddd ymm2,ymm2,ymm5
7109 vpsrld ymm9,ymm13,31
7110 vpaddd ymm13,ymm13,ymm13
7111
7112 vpsrld ymm4,ymm4,2
7113 vpaddd ymm2,ymm2,ymm7
7114 vpor ymm13,ymm13,ymm9
7115 vpor ymm4,ymm4,ymm6
7116 vpxor ymm14,ymm14,ymm11
7117 vmovdqa ymm11,YMMWORD[((384-256-128))+rbx]
7118
7119 vpslld ymm7,ymm2,5
7120 vpaddd ymm1,ymm1,ymm15
7121 vpxor ymm5,ymm0,ymm3
7122 vpaddd ymm1,ymm1,ymm13
7123 vpxor ymm14,ymm14,YMMWORD[((64-128))+rax]
7124 vpsrld ymm8,ymm2,27
7125 vpxor ymm5,ymm5,ymm4
7126 vpxor ymm14,ymm14,ymm11
7127
7128 vpslld ymm6,ymm3,30
7129 vpor ymm7,ymm7,ymm8
7130 vpaddd ymm1,ymm1,ymm5
7131 vpsrld ymm9,ymm14,31
7132 vpaddd ymm14,ymm14,ymm14
7133
7134 vpsrld ymm3,ymm3,2
7135 vpaddd ymm1,ymm1,ymm7
7136 vpor ymm14,ymm14,ymm9
7137 vpor ymm3,ymm3,ymm6
7138 vpxor ymm10,ymm10,ymm12
7139 vmovdqa ymm12,YMMWORD[((416-256-128))+rbx]
7140
7141 vpslld ymm7,ymm1,5
7142 vpaddd ymm0,ymm0,ymm15
7143 vpxor ymm5,ymm4,ymm2
7144 vpaddd ymm0,ymm0,ymm14
7145 vpxor ymm10,ymm10,YMMWORD[((96-128))+rax]
7146 vpsrld ymm8,ymm1,27
7147 vpxor ymm5,ymm5,ymm3
7148 vpxor ymm10,ymm10,ymm12
7149
7150 vpslld ymm6,ymm2,30
7151 vpor ymm7,ymm7,ymm8
7152 vpaddd ymm0,ymm0,ymm5
7153 vpsrld ymm9,ymm10,31
7154 vpaddd ymm10,ymm10,ymm10
7155
7156 vpsrld ymm2,ymm2,2
7157 vpaddd ymm0,ymm0,ymm7
7158 vpor ymm10,ymm10,ymm9
7159 vpor ymm2,ymm2,ymm6
7160 vpxor ymm11,ymm11,ymm13
7161 vmovdqa ymm13,YMMWORD[((448-256-128))+rbx]
7162
7163 vpslld ymm7,ymm0,5
7164 vpaddd ymm4,ymm4,ymm15
7165 vpxor ymm5,ymm3,ymm1
7166 vpaddd ymm4,ymm4,ymm10
7167 vpxor ymm11,ymm11,YMMWORD[((128-128))+rax]
7168 vpsrld ymm8,ymm0,27
7169 vpxor ymm5,ymm5,ymm2
7170 vpxor ymm11,ymm11,ymm13
7171
7172 vpslld ymm6,ymm1,30
7173 vpor ymm7,ymm7,ymm8
7174 vpaddd ymm4,ymm4,ymm5
7175 vpsrld ymm9,ymm11,31
7176 vpaddd ymm11,ymm11,ymm11
7177
7178 vpsrld ymm1,ymm1,2
7179 vpaddd ymm4,ymm4,ymm7
7180 vpor ymm11,ymm11,ymm9
7181 vpor ymm1,ymm1,ymm6
7182 vpxor ymm12,ymm12,ymm14
7183 vmovdqa ymm14,YMMWORD[((480-256-128))+rbx]
7184
7185 vpslld ymm7,ymm4,5
7186 vpaddd ymm3,ymm3,ymm15
7187 vpxor ymm5,ymm2,ymm0
7188 vpaddd ymm3,ymm3,ymm11
7189 vpxor ymm12,ymm12,YMMWORD[((160-128))+rax]
7190 vpsrld ymm8,ymm4,27
7191 vpxor ymm5,ymm5,ymm1
7192 vpxor ymm12,ymm12,ymm14
7193
7194 vpslld ymm6,ymm0,30
7195 vpor ymm7,ymm7,ymm8
7196 vpaddd ymm3,ymm3,ymm5
7197 vpsrld ymm9,ymm12,31
7198 vpaddd ymm12,ymm12,ymm12
7199
7200 vpsrld ymm0,ymm0,2
7201 vpaddd ymm3,ymm3,ymm7
7202 vpor ymm12,ymm12,ymm9
7203 vpor ymm0,ymm0,ymm6
7204 vpxor ymm13,ymm13,ymm10
7205 vmovdqa ymm10,YMMWORD[((0-128))+rax]
7206
7207 vpslld ymm7,ymm3,5
7208 vpaddd ymm2,ymm2,ymm15
7209 vpxor ymm5,ymm1,ymm4
7210 vpaddd ymm2,ymm2,ymm12
7211 vpxor ymm13,ymm13,YMMWORD[((192-128))+rax]
7212 vpsrld ymm8,ymm3,27
7213 vpxor ymm5,ymm5,ymm0
7214 vpxor ymm13,ymm13,ymm10
7215
7216 vpslld ymm6,ymm4,30
7217 vpor ymm7,ymm7,ymm8
7218 vpaddd ymm2,ymm2,ymm5
7219 vpsrld ymm9,ymm13,31
7220 vpaddd ymm13,ymm13,ymm13
7221
7222 vpsrld ymm4,ymm4,2
7223 vpaddd ymm2,ymm2,ymm7
7224 vpor ymm13,ymm13,ymm9
7225 vpor ymm4,ymm4,ymm6
7226 vpxor ymm14,ymm14,ymm11
7227 vmovdqa ymm11,YMMWORD[((32-128))+rax]
7228
7229 vpslld ymm7,ymm2,5
7230 vpaddd ymm1,ymm1,ymm15
7231 vpxor ymm5,ymm0,ymm3
7232 vpaddd ymm1,ymm1,ymm13
7233 vpxor ymm14,ymm14,YMMWORD[((224-128))+rax]
7234 vpsrld ymm8,ymm2,27
7235 vpxor ymm5,ymm5,ymm4
7236 vpxor ymm14,ymm14,ymm11
7237
7238 vpslld ymm6,ymm3,30
7239 vpor ymm7,ymm7,ymm8
7240 vpaddd ymm1,ymm1,ymm5
7241 vpsrld ymm9,ymm14,31
7242 vpaddd ymm14,ymm14,ymm14
7243
7244 vpsrld ymm3,ymm3,2
7245 vpaddd ymm1,ymm1,ymm7
7246 vpor ymm14,ymm14,ymm9
7247 vpor ymm3,ymm3,ymm6
7248 vpslld ymm7,ymm1,5
7249 vpaddd ymm0,ymm0,ymm15
7250 vpxor ymm5,ymm4,ymm2
7251
7252 vpsrld ymm8,ymm1,27
7253 vpaddd ymm0,ymm0,ymm14
7254 vpxor ymm5,ymm5,ymm3
7255
7256 vpslld ymm6,ymm2,30
7257 vpor ymm7,ymm7,ymm8
7258 vpaddd ymm0,ymm0,ymm5
7259
7260 vpsrld ymm2,ymm2,2
7261 vpaddd ymm0,ymm0,ymm7
7262 vpor ymm2,ymm2,ymm6
7263 mov ecx,1
7264 lea rbx,[512+rsp]
7265 cmp ecx,DWORD[rbx]
7266 cmovge r12,rbp
7267 cmp ecx,DWORD[4+rbx]
7268 cmovge r13,rbp
7269 cmp ecx,DWORD[8+rbx]
7270 cmovge r14,rbp
7271 cmp ecx,DWORD[12+rbx]
7272 cmovge r15,rbp
7273 cmp ecx,DWORD[16+rbx]
7274 cmovge r8,rbp
7275 cmp ecx,DWORD[20+rbx]
7276 cmovge r9,rbp
7277 cmp ecx,DWORD[24+rbx]
7278 cmovge r10,rbp
7279 cmp ecx,DWORD[28+rbx]
7280 cmovge r11,rbp
7281 vmovdqu ymm5,YMMWORD[rbx]
7282 vpxor ymm7,ymm7,ymm7
7283 vmovdqa ymm6,ymm5
7284 vpcmpgtd ymm6,ymm6,ymm7
7285 vpaddd ymm5,ymm5,ymm6
7286
7287 vpand ymm0,ymm0,ymm6
7288 vpand ymm1,ymm1,ymm6
7289 vpaddd ymm0,ymm0,YMMWORD[rdi]
7290 vpand ymm2,ymm2,ymm6
7291 vpaddd ymm1,ymm1,YMMWORD[32+rdi]
7292 vpand ymm3,ymm3,ymm6
7293 vpaddd ymm2,ymm2,YMMWORD[64+rdi]
7294 vpand ymm4,ymm4,ymm6
7295 vpaddd ymm3,ymm3,YMMWORD[96+rdi]
7296 vpaddd ymm4,ymm4,YMMWORD[128+rdi]
7297 vmovdqu YMMWORD[rdi],ymm0
7298 vmovdqu YMMWORD[32+rdi],ymm1
7299 vmovdqu YMMWORD[64+rdi],ymm2
7300 vmovdqu YMMWORD[96+rdi],ymm3
7301 vmovdqu YMMWORD[128+rdi],ymm4
7302
7303 vmovdqu YMMWORD[rbx],ymm5
7304 lea rbx,[((256+128))+rsp]
7305 vmovdqu ymm9,YMMWORD[96+rbp]
7306 dec edx
7307 jnz NEAR $L$oop_avx2
7308
7309
7310
7311
7312
7313
7314
7315$L$done_avx2:
7316 mov rax,QWORD[544+rsp]
7317 vzeroupper
7318 movaps xmm6,XMMWORD[((-216))+rax]
7319 movaps xmm7,XMMWORD[((-200))+rax]
7320 movaps xmm8,XMMWORD[((-184))+rax]
7321 movaps xmm9,XMMWORD[((-168))+rax]
7322 movaps xmm10,XMMWORD[((-152))+rax]
7323 movaps xmm11,XMMWORD[((-136))+rax]
7324 movaps xmm12,XMMWORD[((-120))+rax]
7325 movaps xmm13,XMMWORD[((-104))+rax]
7326 movaps xmm14,XMMWORD[((-88))+rax]
7327 movaps xmm15,XMMWORD[((-72))+rax]
7328 mov r15,QWORD[((-48))+rax]
7329 mov r14,QWORD[((-40))+rax]
7330 mov r13,QWORD[((-32))+rax]
7331 mov r12,QWORD[((-24))+rax]
7332 mov rbp,QWORD[((-16))+rax]
7333 mov rbx,QWORD[((-8))+rax]
7334 lea rsp,[rax]
7335$L$epilogue_avx2:
7336 mov rdi,QWORD[8+rsp] ;WIN64 epilogue
7337 mov rsi,QWORD[16+rsp]
7338 DB 0F3h,0C3h ;repret
7339$L$SEH_end_sha1_multi_block_avx2:
7340
7341ALIGN 256
7342 DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
7343 DD 0x5a827999,0x5a827999,0x5a827999,0x5a827999
7344K_XX_XX:
7345 DD 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1
7346 DD 0x6ed9eba1,0x6ed9eba1,0x6ed9eba1,0x6ed9eba1
7347 DD 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc
7348 DD 0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc,0x8f1bbcdc
7349 DD 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6
7350 DD 0xca62c1d6,0xca62c1d6,0xca62c1d6,0xca62c1d6
7351 DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
7352 DD 0x00010203,0x04050607,0x08090a0b,0x0c0d0e0f
7353DB 0xf,0xe,0xd,0xc,0xb,0xa,0x9,0x8,0x7,0x6,0x5,0x4,0x3,0x2,0x1,0x0
7354DB 83,72,65,49,32,109,117,108,116,105,45,98,108,111,99,107
7355DB 32,116,114,97,110,115,102,111,114,109,32,102,111,114,32,120
7356DB 56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77
7357DB 83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110
7358DB 115,115,108,46,111,114,103,62,0
7359EXTERN __imp_RtlVirtualUnwind
7360
7361ALIGN 16
7362se_handler:
7363 push rsi
7364 push rdi
7365 push rbx
7366 push rbp
7367 push r12
7368 push r13
7369 push r14
7370 push r15
7371 pushfq
7372 sub rsp,64
7373
7374 mov rax,QWORD[120+r8]
7375 mov rbx,QWORD[248+r8]
7376
7377 mov rsi,QWORD[8+r9]
7378 mov r11,QWORD[56+r9]
7379
7380 mov r10d,DWORD[r11]
7381 lea r10,[r10*1+rsi]
7382 cmp rbx,r10
7383 jb NEAR $L$in_prologue
7384
7385 mov rax,QWORD[152+r8]
7386
7387 mov r10d,DWORD[4+r11]
7388 lea r10,[r10*1+rsi]
7389 cmp rbx,r10
7390 jae NEAR $L$in_prologue
7391
7392 mov rax,QWORD[272+rax]
7393
7394 mov rbx,QWORD[((-8))+rax]
7395 mov rbp,QWORD[((-16))+rax]
7396 mov QWORD[144+r8],rbx
7397 mov QWORD[160+r8],rbp
7398
7399 lea rsi,[((-24-160))+rax]
7400 lea rdi,[512+r8]
7401 mov ecx,20
7402 DD 0xa548f3fc
7403
7404$L$in_prologue:
7405 mov rdi,QWORD[8+rax]
7406 mov rsi,QWORD[16+rax]
7407 mov QWORD[152+r8],rax
7408 mov QWORD[168+r8],rsi
7409 mov QWORD[176+r8],rdi
7410
7411 mov rdi,QWORD[40+r9]
7412 mov rsi,r8
7413 mov ecx,154
7414 DD 0xa548f3fc
7415
7416 mov rsi,r9
7417 xor rcx,rcx
7418 mov rdx,QWORD[8+rsi]
7419 mov r8,QWORD[rsi]
7420 mov r9,QWORD[16+rsi]
7421 mov r10,QWORD[40+rsi]
7422 lea r11,[56+rsi]
7423 lea r12,[24+rsi]
7424 mov QWORD[32+rsp],r10
7425 mov QWORD[40+rsp],r11
7426 mov QWORD[48+rsp],r12
7427 mov QWORD[56+rsp],rcx
7428 call QWORD[__imp_RtlVirtualUnwind]
7429
7430 mov eax,1
7431 add rsp,64
7432 popfq
7433 pop r15
7434 pop r14
7435 pop r13
7436 pop r12
7437 pop rbp
7438 pop rbx
7439 pop rdi
7440 pop rsi
7441 DB 0F3h,0C3h ;repret
7442
7443
7444ALIGN 16
7445avx2_handler:
7446 push rsi
7447 push rdi
7448 push rbx
7449 push rbp
7450 push r12
7451 push r13
7452 push r14
7453 push r15
7454 pushfq
7455 sub rsp,64
7456
7457 mov rax,QWORD[120+r8]
7458 mov rbx,QWORD[248+r8]
7459
7460 mov rsi,QWORD[8+r9]
7461 mov r11,QWORD[56+r9]
7462
7463 mov r10d,DWORD[r11]
7464 lea r10,[r10*1+rsi]
7465 cmp rbx,r10
7466 jb NEAR $L$in_prologue
7467
7468 mov rax,QWORD[152+r8]
7469
7470 mov r10d,DWORD[4+r11]
7471 lea r10,[r10*1+rsi]
7472 cmp rbx,r10
7473 jae NEAR $L$in_prologue
7474
7475 mov rax,QWORD[544+r8]
7476
7477 mov rbx,QWORD[((-8))+rax]
7478 mov rbp,QWORD[((-16))+rax]
7479 mov r12,QWORD[((-24))+rax]
7480 mov r13,QWORD[((-32))+rax]
7481 mov r14,QWORD[((-40))+rax]
7482 mov r15,QWORD[((-48))+rax]
7483 mov QWORD[144+r8],rbx
7484 mov QWORD[160+r8],rbp
7485 mov QWORD[216+r8],r12
7486 mov QWORD[224+r8],r13
7487 mov QWORD[232+r8],r14
7488 mov QWORD[240+r8],r15
7489
7490 lea rsi,[((-56-160))+rax]
7491 lea rdi,[512+r8]
7492 mov ecx,20
7493 DD 0xa548f3fc
7494
7495 jmp NEAR $L$in_prologue
7496
7497section .pdata rdata align=4
7498ALIGN 4
7499 DD $L$SEH_begin_sha1_multi_block wrt ..imagebase
7500 DD $L$SEH_end_sha1_multi_block wrt ..imagebase
7501 DD $L$SEH_info_sha1_multi_block wrt ..imagebase
7502 DD $L$SEH_begin_sha1_multi_block_shaext wrt ..imagebase
7503 DD $L$SEH_end_sha1_multi_block_shaext wrt ..imagebase
7504 DD $L$SEH_info_sha1_multi_block_shaext wrt ..imagebase
7505 DD $L$SEH_begin_sha1_multi_block_avx wrt ..imagebase
7506 DD $L$SEH_end_sha1_multi_block_avx wrt ..imagebase
7507 DD $L$SEH_info_sha1_multi_block_avx wrt ..imagebase
7508 DD $L$SEH_begin_sha1_multi_block_avx2 wrt ..imagebase
7509 DD $L$SEH_end_sha1_multi_block_avx2 wrt ..imagebase
7510 DD $L$SEH_info_sha1_multi_block_avx2 wrt ..imagebase
7511section .xdata rdata align=8
7512ALIGN 8
7513$L$SEH_info_sha1_multi_block:
7514DB 9,0,0,0
7515 DD se_handler wrt ..imagebase
7516 DD $L$body wrt ..imagebase,$L$epilogue wrt ..imagebase
7517$L$SEH_info_sha1_multi_block_shaext:
7518DB 9,0,0,0
7519 DD se_handler wrt ..imagebase
7520 DD $L$body_shaext wrt ..imagebase,$L$epilogue_shaext wrt ..imagebase
7521$L$SEH_info_sha1_multi_block_avx:
7522DB 9,0,0,0
7523 DD se_handler wrt ..imagebase
7524 DD $L$body_avx wrt ..imagebase,$L$epilogue_avx wrt ..imagebase
7525$L$SEH_info_sha1_multi_block_avx2:
7526DB 9,0,0,0
7527 DD avx2_handler wrt ..imagebase
7528 DD $L$body_avx2 wrt ..imagebase,$L$epilogue_avx2 wrt ..imagebase
Note: See TracBrowser for help on using the repository browser.

© 2025 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette