VirtualBox

source: vbox/trunk/src/libs/openssl-1.1.1g/crypto/genasm-nasm/co-586.S@ 86352

最後變更 在這個檔案從86352是 83531,由 vboxsync 提交於 5 年 前

setting svn:sync-process=export for openssl-1.1.1f, all files except tests
檔案大小: 19.9 KB
行 
1%ifidn __OUTPUT_FORMAT__,obj
2section code use32 class=code align=64
3%elifidn __OUTPUT_FORMAT__,win32
4section .text code align=64
5%else
6section .text code
7%endif
8global _bn_mul_comba8
9align 16
10_bn_mul_comba8:
11L$_bn_mul_comba8_begin:
12 push esi
13 mov esi,DWORD [12+esp]
14 push edi
15 mov edi,DWORD [20+esp]
16 push ebp
17 push ebx
18 xor ebx,ebx
19 mov eax,DWORD [esi]
20 xor ecx,ecx
21 mov edx,DWORD [edi]
22 ; ################## Calculate word 0
23 xor ebp,ebp
24 ; mul a[0]*b[0]
25 mul edx
26 add ebx,eax
27 mov eax,DWORD [20+esp]
28 adc ecx,edx
29 mov edx,DWORD [edi]
30 adc ebp,0
31 mov DWORD [eax],ebx
32 mov eax,DWORD [4+esi]
33 ; saved r[0]
34 ; ################## Calculate word 1
35 xor ebx,ebx
36 ; mul a[1]*b[0]
37 mul edx
38 add ecx,eax
39 mov eax,DWORD [esi]
40 adc ebp,edx
41 mov edx,DWORD [4+edi]
42 adc ebx,0
43 ; mul a[0]*b[1]
44 mul edx
45 add ecx,eax
46 mov eax,DWORD [20+esp]
47 adc ebp,edx
48 mov edx,DWORD [edi]
49 adc ebx,0
50 mov DWORD [4+eax],ecx
51 mov eax,DWORD [8+esi]
52 ; saved r[1]
53 ; ################## Calculate word 2
54 xor ecx,ecx
55 ; mul a[2]*b[0]
56 mul edx
57 add ebp,eax
58 mov eax,DWORD [4+esi]
59 adc ebx,edx
60 mov edx,DWORD [4+edi]
61 adc ecx,0
62 ; mul a[1]*b[1]
63 mul edx
64 add ebp,eax
65 mov eax,DWORD [esi]
66 adc ebx,edx
67 mov edx,DWORD [8+edi]
68 adc ecx,0
69 ; mul a[0]*b[2]
70 mul edx
71 add ebp,eax
72 mov eax,DWORD [20+esp]
73 adc ebx,edx
74 mov edx,DWORD [edi]
75 adc ecx,0
76 mov DWORD [8+eax],ebp
77 mov eax,DWORD [12+esi]
78 ; saved r[2]
79 ; ################## Calculate word 3
80 xor ebp,ebp
81 ; mul a[3]*b[0]
82 mul edx
83 add ebx,eax
84 mov eax,DWORD [8+esi]
85 adc ecx,edx
86 mov edx,DWORD [4+edi]
87 adc ebp,0
88 ; mul a[2]*b[1]
89 mul edx
90 add ebx,eax
91 mov eax,DWORD [4+esi]
92 adc ecx,edx
93 mov edx,DWORD [8+edi]
94 adc ebp,0
95 ; mul a[1]*b[2]
96 mul edx
97 add ebx,eax
98 mov eax,DWORD [esi]
99 adc ecx,edx
100 mov edx,DWORD [12+edi]
101 adc ebp,0
102 ; mul a[0]*b[3]
103 mul edx
104 add ebx,eax
105 mov eax,DWORD [20+esp]
106 adc ecx,edx
107 mov edx,DWORD [edi]
108 adc ebp,0
109 mov DWORD [12+eax],ebx
110 mov eax,DWORD [16+esi]
111 ; saved r[3]
112 ; ################## Calculate word 4
113 xor ebx,ebx
114 ; mul a[4]*b[0]
115 mul edx
116 add ecx,eax
117 mov eax,DWORD [12+esi]
118 adc ebp,edx
119 mov edx,DWORD [4+edi]
120 adc ebx,0
121 ; mul a[3]*b[1]
122 mul edx
123 add ecx,eax
124 mov eax,DWORD [8+esi]
125 adc ebp,edx
126 mov edx,DWORD [8+edi]
127 adc ebx,0
128 ; mul a[2]*b[2]
129 mul edx
130 add ecx,eax
131 mov eax,DWORD [4+esi]
132 adc ebp,edx
133 mov edx,DWORD [12+edi]
134 adc ebx,0
135 ; mul a[1]*b[3]
136 mul edx
137 add ecx,eax
138 mov eax,DWORD [esi]
139 adc ebp,edx
140 mov edx,DWORD [16+edi]
141 adc ebx,0
142 ; mul a[0]*b[4]
143 mul edx
144 add ecx,eax
145 mov eax,DWORD [20+esp]
146 adc ebp,edx
147 mov edx,DWORD [edi]
148 adc ebx,0
149 mov DWORD [16+eax],ecx
150 mov eax,DWORD [20+esi]
151 ; saved r[4]
152 ; ################## Calculate word 5
153 xor ecx,ecx
154 ; mul a[5]*b[0]
155 mul edx
156 add ebp,eax
157 mov eax,DWORD [16+esi]
158 adc ebx,edx
159 mov edx,DWORD [4+edi]
160 adc ecx,0
161 ; mul a[4]*b[1]
162 mul edx
163 add ebp,eax
164 mov eax,DWORD [12+esi]
165 adc ebx,edx
166 mov edx,DWORD [8+edi]
167 adc ecx,0
168 ; mul a[3]*b[2]
169 mul edx
170 add ebp,eax
171 mov eax,DWORD [8+esi]
172 adc ebx,edx
173 mov edx,DWORD [12+edi]
174 adc ecx,0
175 ; mul a[2]*b[3]
176 mul edx
177 add ebp,eax
178 mov eax,DWORD [4+esi]
179 adc ebx,edx
180 mov edx,DWORD [16+edi]
181 adc ecx,0
182 ; mul a[1]*b[4]
183 mul edx
184 add ebp,eax
185 mov eax,DWORD [esi]
186 adc ebx,edx
187 mov edx,DWORD [20+edi]
188 adc ecx,0
189 ; mul a[0]*b[5]
190 mul edx
191 add ebp,eax
192 mov eax,DWORD [20+esp]
193 adc ebx,edx
194 mov edx,DWORD [edi]
195 adc ecx,0
196 mov DWORD [20+eax],ebp
197 mov eax,DWORD [24+esi]
198 ; saved r[5]
199 ; ################## Calculate word 6
200 xor ebp,ebp
201 ; mul a[6]*b[0]
202 mul edx
203 add ebx,eax
204 mov eax,DWORD [20+esi]
205 adc ecx,edx
206 mov edx,DWORD [4+edi]
207 adc ebp,0
208 ; mul a[5]*b[1]
209 mul edx
210 add ebx,eax
211 mov eax,DWORD [16+esi]
212 adc ecx,edx
213 mov edx,DWORD [8+edi]
214 adc ebp,0
215 ; mul a[4]*b[2]
216 mul edx
217 add ebx,eax
218 mov eax,DWORD [12+esi]
219 adc ecx,edx
220 mov edx,DWORD [12+edi]
221 adc ebp,0
222 ; mul a[3]*b[3]
223 mul edx
224 add ebx,eax
225 mov eax,DWORD [8+esi]
226 adc ecx,edx
227 mov edx,DWORD [16+edi]
228 adc ebp,0
229 ; mul a[2]*b[4]
230 mul edx
231 add ebx,eax
232 mov eax,DWORD [4+esi]
233 adc ecx,edx
234 mov edx,DWORD [20+edi]
235 adc ebp,0
236 ; mul a[1]*b[5]
237 mul edx
238 add ebx,eax
239 mov eax,DWORD [esi]
240 adc ecx,edx
241 mov edx,DWORD [24+edi]
242 adc ebp,0
243 ; mul a[0]*b[6]
244 mul edx
245 add ebx,eax
246 mov eax,DWORD [20+esp]
247 adc ecx,edx
248 mov edx,DWORD [edi]
249 adc ebp,0
250 mov DWORD [24+eax],ebx
251 mov eax,DWORD [28+esi]
252 ; saved r[6]
253 ; ################## Calculate word 7
254 xor ebx,ebx
255 ; mul a[7]*b[0]
256 mul edx
257 add ecx,eax
258 mov eax,DWORD [24+esi]
259 adc ebp,edx
260 mov edx,DWORD [4+edi]
261 adc ebx,0
262 ; mul a[6]*b[1]
263 mul edx
264 add ecx,eax
265 mov eax,DWORD [20+esi]
266 adc ebp,edx
267 mov edx,DWORD [8+edi]
268 adc ebx,0
269 ; mul a[5]*b[2]
270 mul edx
271 add ecx,eax
272 mov eax,DWORD [16+esi]
273 adc ebp,edx
274 mov edx,DWORD [12+edi]
275 adc ebx,0
276 ; mul a[4]*b[3]
277 mul edx
278 add ecx,eax
279 mov eax,DWORD [12+esi]
280 adc ebp,edx
281 mov edx,DWORD [16+edi]
282 adc ebx,0
283 ; mul a[3]*b[4]
284 mul edx
285 add ecx,eax
286 mov eax,DWORD [8+esi]
287 adc ebp,edx
288 mov edx,DWORD [20+edi]
289 adc ebx,0
290 ; mul a[2]*b[5]
291 mul edx
292 add ecx,eax
293 mov eax,DWORD [4+esi]
294 adc ebp,edx
295 mov edx,DWORD [24+edi]
296 adc ebx,0
297 ; mul a[1]*b[6]
298 mul edx
299 add ecx,eax
300 mov eax,DWORD [esi]
301 adc ebp,edx
302 mov edx,DWORD [28+edi]
303 adc ebx,0
304 ; mul a[0]*b[7]
305 mul edx
306 add ecx,eax
307 mov eax,DWORD [20+esp]
308 adc ebp,edx
309 mov edx,DWORD [4+edi]
310 adc ebx,0
311 mov DWORD [28+eax],ecx
312 mov eax,DWORD [28+esi]
313 ; saved r[7]
314 ; ################## Calculate word 8
315 xor ecx,ecx
316 ; mul a[7]*b[1]
317 mul edx
318 add ebp,eax
319 mov eax,DWORD [24+esi]
320 adc ebx,edx
321 mov edx,DWORD [8+edi]
322 adc ecx,0
323 ; mul a[6]*b[2]
324 mul edx
325 add ebp,eax
326 mov eax,DWORD [20+esi]
327 adc ebx,edx
328 mov edx,DWORD [12+edi]
329 adc ecx,0
330 ; mul a[5]*b[3]
331 mul edx
332 add ebp,eax
333 mov eax,DWORD [16+esi]
334 adc ebx,edx
335 mov edx,DWORD [16+edi]
336 adc ecx,0
337 ; mul a[4]*b[4]
338 mul edx
339 add ebp,eax
340 mov eax,DWORD [12+esi]
341 adc ebx,edx
342 mov edx,DWORD [20+edi]
343 adc ecx,0
344 ; mul a[3]*b[5]
345 mul edx
346 add ebp,eax
347 mov eax,DWORD [8+esi]
348 adc ebx,edx
349 mov edx,DWORD [24+edi]
350 adc ecx,0
351 ; mul a[2]*b[6]
352 mul edx
353 add ebp,eax
354 mov eax,DWORD [4+esi]
355 adc ebx,edx
356 mov edx,DWORD [28+edi]
357 adc ecx,0
358 ; mul a[1]*b[7]
359 mul edx
360 add ebp,eax
361 mov eax,DWORD [20+esp]
362 adc ebx,edx
363 mov edx,DWORD [8+edi]
364 adc ecx,0
365 mov DWORD [32+eax],ebp
366 mov eax,DWORD [28+esi]
367 ; saved r[8]
368 ; ################## Calculate word 9
369 xor ebp,ebp
370 ; mul a[7]*b[2]
371 mul edx
372 add ebx,eax
373 mov eax,DWORD [24+esi]
374 adc ecx,edx
375 mov edx,DWORD [12+edi]
376 adc ebp,0
377 ; mul a[6]*b[3]
378 mul edx
379 add ebx,eax
380 mov eax,DWORD [20+esi]
381 adc ecx,edx
382 mov edx,DWORD [16+edi]
383 adc ebp,0
384 ; mul a[5]*b[4]
385 mul edx
386 add ebx,eax
387 mov eax,DWORD [16+esi]
388 adc ecx,edx
389 mov edx,DWORD [20+edi]
390 adc ebp,0
391 ; mul a[4]*b[5]
392 mul edx
393 add ebx,eax
394 mov eax,DWORD [12+esi]
395 adc ecx,edx
396 mov edx,DWORD [24+edi]
397 adc ebp,0
398 ; mul a[3]*b[6]
399 mul edx
400 add ebx,eax
401 mov eax,DWORD [8+esi]
402 adc ecx,edx
403 mov edx,DWORD [28+edi]
404 adc ebp,0
405 ; mul a[2]*b[7]
406 mul edx
407 add ebx,eax
408 mov eax,DWORD [20+esp]
409 adc ecx,edx
410 mov edx,DWORD [12+edi]
411 adc ebp,0
412 mov DWORD [36+eax],ebx
413 mov eax,DWORD [28+esi]
414 ; saved r[9]
415 ; ################## Calculate word 10
416 xor ebx,ebx
417 ; mul a[7]*b[3]
418 mul edx
419 add ecx,eax
420 mov eax,DWORD [24+esi]
421 adc ebp,edx
422 mov edx,DWORD [16+edi]
423 adc ebx,0
424 ; mul a[6]*b[4]
425 mul edx
426 add ecx,eax
427 mov eax,DWORD [20+esi]
428 adc ebp,edx
429 mov edx,DWORD [20+edi]
430 adc ebx,0
431 ; mul a[5]*b[5]
432 mul edx
433 add ecx,eax
434 mov eax,DWORD [16+esi]
435 adc ebp,edx
436 mov edx,DWORD [24+edi]
437 adc ebx,0
438 ; mul a[4]*b[6]
439 mul edx
440 add ecx,eax
441 mov eax,DWORD [12+esi]
442 adc ebp,edx
443 mov edx,DWORD [28+edi]
444 adc ebx,0
445 ; mul a[3]*b[7]
446 mul edx
447 add ecx,eax
448 mov eax,DWORD [20+esp]
449 adc ebp,edx
450 mov edx,DWORD [16+edi]
451 adc ebx,0
452 mov DWORD [40+eax],ecx
453 mov eax,DWORD [28+esi]
454 ; saved r[10]
455 ; ################## Calculate word 11
456 xor ecx,ecx
457 ; mul a[7]*b[4]
458 mul edx
459 add ebp,eax
460 mov eax,DWORD [24+esi]
461 adc ebx,edx
462 mov edx,DWORD [20+edi]
463 adc ecx,0
464 ; mul a[6]*b[5]
465 mul edx
466 add ebp,eax
467 mov eax,DWORD [20+esi]
468 adc ebx,edx
469 mov edx,DWORD [24+edi]
470 adc ecx,0
471 ; mul a[5]*b[6]
472 mul edx
473 add ebp,eax
474 mov eax,DWORD [16+esi]
475 adc ebx,edx
476 mov edx,DWORD [28+edi]
477 adc ecx,0
478 ; mul a[4]*b[7]
479 mul edx
480 add ebp,eax
481 mov eax,DWORD [20+esp]
482 adc ebx,edx
483 mov edx,DWORD [20+edi]
484 adc ecx,0
485 mov DWORD [44+eax],ebp
486 mov eax,DWORD [28+esi]
487 ; saved r[11]
488 ; ################## Calculate word 12
489 xor ebp,ebp
490 ; mul a[7]*b[5]
491 mul edx
492 add ebx,eax
493 mov eax,DWORD [24+esi]
494 adc ecx,edx
495 mov edx,DWORD [24+edi]
496 adc ebp,0
497 ; mul a[6]*b[6]
498 mul edx
499 add ebx,eax
500 mov eax,DWORD [20+esi]
501 adc ecx,edx
502 mov edx,DWORD [28+edi]
503 adc ebp,0
504 ; mul a[5]*b[7]
505 mul edx
506 add ebx,eax
507 mov eax,DWORD [20+esp]
508 adc ecx,edx
509 mov edx,DWORD [24+edi]
510 adc ebp,0
511 mov DWORD [48+eax],ebx
512 mov eax,DWORD [28+esi]
513 ; saved r[12]
514 ; ################## Calculate word 13
515 xor ebx,ebx
516 ; mul a[7]*b[6]
517 mul edx
518 add ecx,eax
519 mov eax,DWORD [24+esi]
520 adc ebp,edx
521 mov edx,DWORD [28+edi]
522 adc ebx,0
523 ; mul a[6]*b[7]
524 mul edx
525 add ecx,eax
526 mov eax,DWORD [20+esp]
527 adc ebp,edx
528 mov edx,DWORD [28+edi]
529 adc ebx,0
530 mov DWORD [52+eax],ecx
531 mov eax,DWORD [28+esi]
532 ; saved r[13]
533 ; ################## Calculate word 14
534 xor ecx,ecx
535 ; mul a[7]*b[7]
536 mul edx
537 add ebp,eax
538 mov eax,DWORD [20+esp]
539 adc ebx,edx
540 adc ecx,0
541 mov DWORD [56+eax],ebp
542 ; saved r[14]
543 ; save r[15]
544 mov DWORD [60+eax],ebx
545 pop ebx
546 pop ebp
547 pop edi
548 pop esi
549 ret
550global _bn_mul_comba4
551align 16
552_bn_mul_comba4:
553L$_bn_mul_comba4_begin:
554 push esi
555 mov esi,DWORD [12+esp]
556 push edi
557 mov edi,DWORD [20+esp]
558 push ebp
559 push ebx
560 xor ebx,ebx
561 mov eax,DWORD [esi]
562 xor ecx,ecx
563 mov edx,DWORD [edi]
564 ; ################## Calculate word 0
565 xor ebp,ebp
566 ; mul a[0]*b[0]
567 mul edx
568 add ebx,eax
569 mov eax,DWORD [20+esp]
570 adc ecx,edx
571 mov edx,DWORD [edi]
572 adc ebp,0
573 mov DWORD [eax],ebx
574 mov eax,DWORD [4+esi]
575 ; saved r[0]
576 ; ################## Calculate word 1
577 xor ebx,ebx
578 ; mul a[1]*b[0]
579 mul edx
580 add ecx,eax
581 mov eax,DWORD [esi]
582 adc ebp,edx
583 mov edx,DWORD [4+edi]
584 adc ebx,0
585 ; mul a[0]*b[1]
586 mul edx
587 add ecx,eax
588 mov eax,DWORD [20+esp]
589 adc ebp,edx
590 mov edx,DWORD [edi]
591 adc ebx,0
592 mov DWORD [4+eax],ecx
593 mov eax,DWORD [8+esi]
594 ; saved r[1]
595 ; ################## Calculate word 2
596 xor ecx,ecx
597 ; mul a[2]*b[0]
598 mul edx
599 add ebp,eax
600 mov eax,DWORD [4+esi]
601 adc ebx,edx
602 mov edx,DWORD [4+edi]
603 adc ecx,0
604 ; mul a[1]*b[1]
605 mul edx
606 add ebp,eax
607 mov eax,DWORD [esi]
608 adc ebx,edx
609 mov edx,DWORD [8+edi]
610 adc ecx,0
611 ; mul a[0]*b[2]
612 mul edx
613 add ebp,eax
614 mov eax,DWORD [20+esp]
615 adc ebx,edx
616 mov edx,DWORD [edi]
617 adc ecx,0
618 mov DWORD [8+eax],ebp
619 mov eax,DWORD [12+esi]
620 ; saved r[2]
621 ; ################## Calculate word 3
622 xor ebp,ebp
623 ; mul a[3]*b[0]
624 mul edx
625 add ebx,eax
626 mov eax,DWORD [8+esi]
627 adc ecx,edx
628 mov edx,DWORD [4+edi]
629 adc ebp,0
630 ; mul a[2]*b[1]
631 mul edx
632 add ebx,eax
633 mov eax,DWORD [4+esi]
634 adc ecx,edx
635 mov edx,DWORD [8+edi]
636 adc ebp,0
637 ; mul a[1]*b[2]
638 mul edx
639 add ebx,eax
640 mov eax,DWORD [esi]
641 adc ecx,edx
642 mov edx,DWORD [12+edi]
643 adc ebp,0
644 ; mul a[0]*b[3]
645 mul edx
646 add ebx,eax
647 mov eax,DWORD [20+esp]
648 adc ecx,edx
649 mov edx,DWORD [4+edi]
650 adc ebp,0
651 mov DWORD [12+eax],ebx
652 mov eax,DWORD [12+esi]
653 ; saved r[3]
654 ; ################## Calculate word 4
655 xor ebx,ebx
656 ; mul a[3]*b[1]
657 mul edx
658 add ecx,eax
659 mov eax,DWORD [8+esi]
660 adc ebp,edx
661 mov edx,DWORD [8+edi]
662 adc ebx,0
663 ; mul a[2]*b[2]
664 mul edx
665 add ecx,eax
666 mov eax,DWORD [4+esi]
667 adc ebp,edx
668 mov edx,DWORD [12+edi]
669 adc ebx,0
670 ; mul a[1]*b[3]
671 mul edx
672 add ecx,eax
673 mov eax,DWORD [20+esp]
674 adc ebp,edx
675 mov edx,DWORD [8+edi]
676 adc ebx,0
677 mov DWORD [16+eax],ecx
678 mov eax,DWORD [12+esi]
679 ; saved r[4]
680 ; ################## Calculate word 5
681 xor ecx,ecx
682 ; mul a[3]*b[2]
683 mul edx
684 add ebp,eax
685 mov eax,DWORD [8+esi]
686 adc ebx,edx
687 mov edx,DWORD [12+edi]
688 adc ecx,0
689 ; mul a[2]*b[3]
690 mul edx
691 add ebp,eax
692 mov eax,DWORD [20+esp]
693 adc ebx,edx
694 mov edx,DWORD [12+edi]
695 adc ecx,0
696 mov DWORD [20+eax],ebp
697 mov eax,DWORD [12+esi]
698 ; saved r[5]
699 ; ################## Calculate word 6
700 xor ebp,ebp
701 ; mul a[3]*b[3]
702 mul edx
703 add ebx,eax
704 mov eax,DWORD [20+esp]
705 adc ecx,edx
706 adc ebp,0
707 mov DWORD [24+eax],ebx
708 ; saved r[6]
709 ; save r[7]
710 mov DWORD [28+eax],ecx
711 pop ebx
712 pop ebp
713 pop edi
714 pop esi
715 ret
716global _bn_sqr_comba8
717align 16
718_bn_sqr_comba8:
719L$_bn_sqr_comba8_begin:
720 push esi
721 push edi
722 push ebp
723 push ebx
724 mov edi,DWORD [20+esp]
725 mov esi,DWORD [24+esp]
726 xor ebx,ebx
727 xor ecx,ecx
728 mov eax,DWORD [esi]
729 ; ############### Calculate word 0
730 xor ebp,ebp
731 ; sqr a[0]*a[0]
732 mul eax
733 add ebx,eax
734 adc ecx,edx
735 mov edx,DWORD [esi]
736 adc ebp,0
737 mov DWORD [edi],ebx
738 mov eax,DWORD [4+esi]
739 ; saved r[0]
740 ; ############### Calculate word 1
741 xor ebx,ebx
742 ; sqr a[1]*a[0]
743 mul edx
744 add eax,eax
745 adc edx,edx
746 adc ebx,0
747 add ecx,eax
748 adc ebp,edx
749 mov eax,DWORD [8+esi]
750 adc ebx,0
751 mov DWORD [4+edi],ecx
752 mov edx,DWORD [esi]
753 ; saved r[1]
754 ; ############### Calculate word 2
755 xor ecx,ecx
756 ; sqr a[2]*a[0]
757 mul edx
758 add eax,eax
759 adc edx,edx
760 adc ecx,0
761 add ebp,eax
762 adc ebx,edx
763 mov eax,DWORD [4+esi]
764 adc ecx,0
765 ; sqr a[1]*a[1]
766 mul eax
767 add ebp,eax
768 adc ebx,edx
769 mov edx,DWORD [esi]
770 adc ecx,0
771 mov DWORD [8+edi],ebp
772 mov eax,DWORD [12+esi]
773 ; saved r[2]
774 ; ############### Calculate word 3
775 xor ebp,ebp
776 ; sqr a[3]*a[0]
777 mul edx
778 add eax,eax
779 adc edx,edx
780 adc ebp,0
781 add ebx,eax
782 adc ecx,edx
783 mov eax,DWORD [8+esi]
784 adc ebp,0
785 mov edx,DWORD [4+esi]
786 ; sqr a[2]*a[1]
787 mul edx
788 add eax,eax
789 adc edx,edx
790 adc ebp,0
791 add ebx,eax
792 adc ecx,edx
793 mov eax,DWORD [16+esi]
794 adc ebp,0
795 mov DWORD [12+edi],ebx
796 mov edx,DWORD [esi]
797 ; saved r[3]
798 ; ############### Calculate word 4
799 xor ebx,ebx
800 ; sqr a[4]*a[0]
801 mul edx
802 add eax,eax
803 adc edx,edx
804 adc ebx,0
805 add ecx,eax
806 adc ebp,edx
807 mov eax,DWORD [12+esi]
808 adc ebx,0
809 mov edx,DWORD [4+esi]
810 ; sqr a[3]*a[1]
811 mul edx
812 add eax,eax
813 adc edx,edx
814 adc ebx,0
815 add ecx,eax
816 adc ebp,edx
817 mov eax,DWORD [8+esi]
818 adc ebx,0
819 ; sqr a[2]*a[2]
820 mul eax
821 add ecx,eax
822 adc ebp,edx
823 mov edx,DWORD [esi]
824 adc ebx,0
825 mov DWORD [16+edi],ecx
826 mov eax,DWORD [20+esi]
827 ; saved r[4]
828 ; ############### Calculate word 5
829 xor ecx,ecx
830 ; sqr a[5]*a[0]
831 mul edx
832 add eax,eax
833 adc edx,edx
834 adc ecx,0
835 add ebp,eax
836 adc ebx,edx
837 mov eax,DWORD [16+esi]
838 adc ecx,0
839 mov edx,DWORD [4+esi]
840 ; sqr a[4]*a[1]
841 mul edx
842 add eax,eax
843 adc edx,edx
844 adc ecx,0
845 add ebp,eax
846 adc ebx,edx
847 mov eax,DWORD [12+esi]
848 adc ecx,0
849 mov edx,DWORD [8+esi]
850 ; sqr a[3]*a[2]
851 mul edx
852 add eax,eax
853 adc edx,edx
854 adc ecx,0
855 add ebp,eax
856 adc ebx,edx
857 mov eax,DWORD [24+esi]
858 adc ecx,0
859 mov DWORD [20+edi],ebp
860 mov edx,DWORD [esi]
861 ; saved r[5]
862 ; ############### Calculate word 6
863 xor ebp,ebp
864 ; sqr a[6]*a[0]
865 mul edx
866 add eax,eax
867 adc edx,edx
868 adc ebp,0
869 add ebx,eax
870 adc ecx,edx
871 mov eax,DWORD [20+esi]
872 adc ebp,0
873 mov edx,DWORD [4+esi]
874 ; sqr a[5]*a[1]
875 mul edx
876 add eax,eax
877 adc edx,edx
878 adc ebp,0
879 add ebx,eax
880 adc ecx,edx
881 mov eax,DWORD [16+esi]
882 adc ebp,0
883 mov edx,DWORD [8+esi]
884 ; sqr a[4]*a[2]
885 mul edx
886 add eax,eax
887 adc edx,edx
888 adc ebp,0
889 add ebx,eax
890 adc ecx,edx
891 mov eax,DWORD [12+esi]
892 adc ebp,0
893 ; sqr a[3]*a[3]
894 mul eax
895 add ebx,eax
896 adc ecx,edx
897 mov edx,DWORD [esi]
898 adc ebp,0
899 mov DWORD [24+edi],ebx
900 mov eax,DWORD [28+esi]
901 ; saved r[6]
902 ; ############### Calculate word 7
903 xor ebx,ebx
904 ; sqr a[7]*a[0]
905 mul edx
906 add eax,eax
907 adc edx,edx
908 adc ebx,0
909 add ecx,eax
910 adc ebp,edx
911 mov eax,DWORD [24+esi]
912 adc ebx,0
913 mov edx,DWORD [4+esi]
914 ; sqr a[6]*a[1]
915 mul edx
916 add eax,eax
917 adc edx,edx
918 adc ebx,0
919 add ecx,eax
920 adc ebp,edx
921 mov eax,DWORD [20+esi]
922 adc ebx,0
923 mov edx,DWORD [8+esi]
924 ; sqr a[5]*a[2]
925 mul edx
926 add eax,eax
927 adc edx,edx
928 adc ebx,0
929 add ecx,eax
930 adc ebp,edx
931 mov eax,DWORD [16+esi]
932 adc ebx,0
933 mov edx,DWORD [12+esi]
934 ; sqr a[4]*a[3]
935 mul edx
936 add eax,eax
937 adc edx,edx
938 adc ebx,0
939 add ecx,eax
940 adc ebp,edx
941 mov eax,DWORD [28+esi]
942 adc ebx,0
943 mov DWORD [28+edi],ecx
944 mov edx,DWORD [4+esi]
945 ; saved r[7]
946 ; ############### Calculate word 8
947 xor ecx,ecx
948 ; sqr a[7]*a[1]
949 mul edx
950 add eax,eax
951 adc edx,edx
952 adc ecx,0
953 add ebp,eax
954 adc ebx,edx
955 mov eax,DWORD [24+esi]
956 adc ecx,0
957 mov edx,DWORD [8+esi]
958 ; sqr a[6]*a[2]
959 mul edx
960 add eax,eax
961 adc edx,edx
962 adc ecx,0
963 add ebp,eax
964 adc ebx,edx
965 mov eax,DWORD [20+esi]
966 adc ecx,0
967 mov edx,DWORD [12+esi]
968 ; sqr a[5]*a[3]
969 mul edx
970 add eax,eax
971 adc edx,edx
972 adc ecx,0
973 add ebp,eax
974 adc ebx,edx
975 mov eax,DWORD [16+esi]
976 adc ecx,0
977 ; sqr a[4]*a[4]
978 mul eax
979 add ebp,eax
980 adc ebx,edx
981 mov edx,DWORD [8+esi]
982 adc ecx,0
983 mov DWORD [32+edi],ebp
984 mov eax,DWORD [28+esi]
985 ; saved r[8]
986 ; ############### Calculate word 9
987 xor ebp,ebp
988 ; sqr a[7]*a[2]
989 mul edx
990 add eax,eax
991 adc edx,edx
992 adc ebp,0
993 add ebx,eax
994 adc ecx,edx
995 mov eax,DWORD [24+esi]
996 adc ebp,0
997 mov edx,DWORD [12+esi]
998 ; sqr a[6]*a[3]
999 mul edx
1000 add eax,eax
1001 adc edx,edx
1002 adc ebp,0
1003 add ebx,eax
1004 adc ecx,edx
1005 mov eax,DWORD [20+esi]
1006 adc ebp,0
1007 mov edx,DWORD [16+esi]
1008 ; sqr a[5]*a[4]
1009 mul edx
1010 add eax,eax
1011 adc edx,edx
1012 adc ebp,0
1013 add ebx,eax
1014 adc ecx,edx
1015 mov eax,DWORD [28+esi]
1016 adc ebp,0
1017 mov DWORD [36+edi],ebx
1018 mov edx,DWORD [12+esi]
1019 ; saved r[9]
1020 ; ############### Calculate word 10
1021 xor ebx,ebx
1022 ; sqr a[7]*a[3]
1023 mul edx
1024 add eax,eax
1025 adc edx,edx
1026 adc ebx,0
1027 add ecx,eax
1028 adc ebp,edx
1029 mov eax,DWORD [24+esi]
1030 adc ebx,0
1031 mov edx,DWORD [16+esi]
1032 ; sqr a[6]*a[4]
1033 mul edx
1034 add eax,eax
1035 adc edx,edx
1036 adc ebx,0
1037 add ecx,eax
1038 adc ebp,edx
1039 mov eax,DWORD [20+esi]
1040 adc ebx,0
1041 ; sqr a[5]*a[5]
1042 mul eax
1043 add ecx,eax
1044 adc ebp,edx
1045 mov edx,DWORD [16+esi]
1046 adc ebx,0
1047 mov DWORD [40+edi],ecx
1048 mov eax,DWORD [28+esi]
1049 ; saved r[10]
1050 ; ############### Calculate word 11
1051 xor ecx,ecx
1052 ; sqr a[7]*a[4]
1053 mul edx
1054 add eax,eax
1055 adc edx,edx
1056 adc ecx,0
1057 add ebp,eax
1058 adc ebx,edx
1059 mov eax,DWORD [24+esi]
1060 adc ecx,0
1061 mov edx,DWORD [20+esi]
1062 ; sqr a[6]*a[5]
1063 mul edx
1064 add eax,eax
1065 adc edx,edx
1066 adc ecx,0
1067 add ebp,eax
1068 adc ebx,edx
1069 mov eax,DWORD [28+esi]
1070 adc ecx,0
1071 mov DWORD [44+edi],ebp
1072 mov edx,DWORD [20+esi]
1073 ; saved r[11]
1074 ; ############### Calculate word 12
1075 xor ebp,ebp
1076 ; sqr a[7]*a[5]
1077 mul edx
1078 add eax,eax
1079 adc edx,edx
1080 adc ebp,0
1081 add ebx,eax
1082 adc ecx,edx
1083 mov eax,DWORD [24+esi]
1084 adc ebp,0
1085 ; sqr a[6]*a[6]
1086 mul eax
1087 add ebx,eax
1088 adc ecx,edx
1089 mov edx,DWORD [24+esi]
1090 adc ebp,0
1091 mov DWORD [48+edi],ebx
1092 mov eax,DWORD [28+esi]
1093 ; saved r[12]
1094 ; ############### Calculate word 13
1095 xor ebx,ebx
1096 ; sqr a[7]*a[6]
1097 mul edx
1098 add eax,eax
1099 adc edx,edx
1100 adc ebx,0
1101 add ecx,eax
1102 adc ebp,edx
1103 mov eax,DWORD [28+esi]
1104 adc ebx,0
1105 mov DWORD [52+edi],ecx
1106 ; saved r[13]
1107 ; ############### Calculate word 14
1108 xor ecx,ecx
1109 ; sqr a[7]*a[7]
1110 mul eax
1111 add ebp,eax
1112 adc ebx,edx
1113 adc ecx,0
1114 mov DWORD [56+edi],ebp
1115 ; saved r[14]
1116 mov DWORD [60+edi],ebx
1117 pop ebx
1118 pop ebp
1119 pop edi
1120 pop esi
1121 ret
1122global _bn_sqr_comba4
1123align 16
1124_bn_sqr_comba4:
1125L$_bn_sqr_comba4_begin:
1126 push esi
1127 push edi
1128 push ebp
1129 push ebx
1130 mov edi,DWORD [20+esp]
1131 mov esi,DWORD [24+esp]
1132 xor ebx,ebx
1133 xor ecx,ecx
1134 mov eax,DWORD [esi]
1135 ; ############### Calculate word 0
1136 xor ebp,ebp
1137 ; sqr a[0]*a[0]
1138 mul eax
1139 add ebx,eax
1140 adc ecx,edx
1141 mov edx,DWORD [esi]
1142 adc ebp,0
1143 mov DWORD [edi],ebx
1144 mov eax,DWORD [4+esi]
1145 ; saved r[0]
1146 ; ############### Calculate word 1
1147 xor ebx,ebx
1148 ; sqr a[1]*a[0]
1149 mul edx
1150 add eax,eax
1151 adc edx,edx
1152 adc ebx,0
1153 add ecx,eax
1154 adc ebp,edx
1155 mov eax,DWORD [8+esi]
1156 adc ebx,0
1157 mov DWORD [4+edi],ecx
1158 mov edx,DWORD [esi]
1159 ; saved r[1]
1160 ; ############### Calculate word 2
1161 xor ecx,ecx
1162 ; sqr a[2]*a[0]
1163 mul edx
1164 add eax,eax
1165 adc edx,edx
1166 adc ecx,0
1167 add ebp,eax
1168 adc ebx,edx
1169 mov eax,DWORD [4+esi]
1170 adc ecx,0
1171 ; sqr a[1]*a[1]
1172 mul eax
1173 add ebp,eax
1174 adc ebx,edx
1175 mov edx,DWORD [esi]
1176 adc ecx,0
1177 mov DWORD [8+edi],ebp
1178 mov eax,DWORD [12+esi]
1179 ; saved r[2]
1180 ; ############### Calculate word 3
1181 xor ebp,ebp
1182 ; sqr a[3]*a[0]
1183 mul edx
1184 add eax,eax
1185 adc edx,edx
1186 adc ebp,0
1187 add ebx,eax
1188 adc ecx,edx
1189 mov eax,DWORD [8+esi]
1190 adc ebp,0
1191 mov edx,DWORD [4+esi]
1192 ; sqr a[2]*a[1]
1193 mul edx
1194 add eax,eax
1195 adc edx,edx
1196 adc ebp,0
1197 add ebx,eax
1198 adc ecx,edx
1199 mov eax,DWORD [12+esi]
1200 adc ebp,0
1201 mov DWORD [12+edi],ebx
1202 mov edx,DWORD [4+esi]
1203 ; saved r[3]
1204 ; ############### Calculate word 4
1205 xor ebx,ebx
1206 ; sqr a[3]*a[1]
1207 mul edx
1208 add eax,eax
1209 adc edx,edx
1210 adc ebx,0
1211 add ecx,eax
1212 adc ebp,edx
1213 mov eax,DWORD [8+esi]
1214 adc ebx,0
1215 ; sqr a[2]*a[2]
1216 mul eax
1217 add ecx,eax
1218 adc ebp,edx
1219 mov edx,DWORD [8+esi]
1220 adc ebx,0
1221 mov DWORD [16+edi],ecx
1222 mov eax,DWORD [12+esi]
1223 ; saved r[4]
1224 ; ############### Calculate word 5
1225 xor ecx,ecx
1226 ; sqr a[3]*a[2]
1227 mul edx
1228 add eax,eax
1229 adc edx,edx
1230 adc ecx,0
1231 add ebp,eax
1232 adc ebx,edx
1233 mov eax,DWORD [12+esi]
1234 adc ecx,0
1235 mov DWORD [20+edi],ebp
1236 ; saved r[5]
1237 ; ############### Calculate word 6
1238 xor ebp,ebp
1239 ; sqr a[3]*a[3]
1240 mul eax
1241 add ebx,eax
1242 adc ecx,edx
1243 adc ebp,0
1244 mov DWORD [24+edi],ebx
1245 ; saved r[6]
1246 mov DWORD [28+edi],ecx
1247 pop ebx
1248 pop ebp
1249 pop edi
1250 pop esi
1251 ret
注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

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