bn-586.S@ 83531

Last change on this file since 83531 was 83531, checked in by vboxsync, 5 years ago
setting svn:sync-process=export for openssl-1.1.1f, all files except tests
File size: 21.3 KB

Line
1	.text
2	.globl _bn_mul_add_words
3	.align 4
4	_bn_mul_add_words:
5	L_bn_mul_add_words_begin:
6	pushl %ebp
7	pushl %ebx
8	pushl %esi
9	pushl %edi
10
11	xorl %esi,%esi
12	movl 20(%esp),%edi
13	movl 28(%esp),%ecx
14	movl 24(%esp),%ebx
15	andl $4294967288,%ecx
16	movl 32(%esp),%ebp
17	pushl %ecx
18	jz L000maw_finish
19	.align 4,0x90
20	L001maw_loop:
21	# Round 0
22	movl (%ebx),%eax
23	mull %ebp
24	addl %esi,%eax
25	adcl $0,%edx
26	addl (%edi),%eax
27	adcl $0,%edx
28	movl %eax,(%edi)
29	movl %edx,%esi
30	# Round 4
31	movl 4(%ebx),%eax
32	mull %ebp
33	addl %esi,%eax
34	adcl $0,%edx
35	addl 4(%edi),%eax
36	adcl $0,%edx
37	movl %eax,4(%edi)
38	movl %edx,%esi
39	# Round 8
40	movl 8(%ebx),%eax
41	mull %ebp
42	addl %esi,%eax
43	adcl $0,%edx
44	addl 8(%edi),%eax
45	adcl $0,%edx
46	movl %eax,8(%edi)
47	movl %edx,%esi
48	# Round 12
49	movl 12(%ebx),%eax
50	mull %ebp
51	addl %esi,%eax
52	adcl $0,%edx
53	addl 12(%edi),%eax
54	adcl $0,%edx
55	movl %eax,12(%edi)
56	movl %edx,%esi
57	# Round 16
58	movl 16(%ebx),%eax
59	mull %ebp
60	addl %esi,%eax
61	adcl $0,%edx
62	addl 16(%edi),%eax
63	adcl $0,%edx
64	movl %eax,16(%edi)
65	movl %edx,%esi
66	# Round 20
67	movl 20(%ebx),%eax
68	mull %ebp
69	addl %esi,%eax
70	adcl $0,%edx
71	addl 20(%edi),%eax
72	adcl $0,%edx
73	movl %eax,20(%edi)
74	movl %edx,%esi
75	# Round 24
76	movl 24(%ebx),%eax
77	mull %ebp
78	addl %esi,%eax
79	adcl $0,%edx
80	addl 24(%edi),%eax
81	adcl $0,%edx
82	movl %eax,24(%edi)
83	movl %edx,%esi
84	# Round 28
85	movl 28(%ebx),%eax
86	mull %ebp
87	addl %esi,%eax
88	adcl $0,%edx
89	addl 28(%edi),%eax
90	adcl $0,%edx
91	movl %eax,28(%edi)
92	movl %edx,%esi
93
94	subl $8,%ecx
95	leal 32(%ebx),%ebx
96	leal 32(%edi),%edi
97	jnz L001maw_loop
98	L000maw_finish:
99	movl 32(%esp),%ecx
100	andl $7,%ecx
101	jnz L002maw_finish2
102	jmp L003maw_end
103	L002maw_finish2:
104	# Tail Round 0
105	movl (%ebx),%eax
106	mull %ebp
107	addl %esi,%eax
108	adcl $0,%edx
109	addl (%edi),%eax
110	adcl $0,%edx
111	decl %ecx
112	movl %eax,(%edi)
113	movl %edx,%esi
114	jz L003maw_end
115	# Tail Round 1
116	movl 4(%ebx),%eax
117	mull %ebp
118	addl %esi,%eax
119	adcl $0,%edx
120	addl 4(%edi),%eax
121	adcl $0,%edx
122	decl %ecx
123	movl %eax,4(%edi)
124	movl %edx,%esi
125	jz L003maw_end
126	# Tail Round 2
127	movl 8(%ebx),%eax
128	mull %ebp
129	addl %esi,%eax
130	adcl $0,%edx
131	addl 8(%edi),%eax
132	adcl $0,%edx
133	decl %ecx
134	movl %eax,8(%edi)
135	movl %edx,%esi
136	jz L003maw_end
137	# Tail Round 3
138	movl 12(%ebx),%eax
139	mull %ebp
140	addl %esi,%eax
141	adcl $0,%edx
142	addl 12(%edi),%eax
143	adcl $0,%edx
144	decl %ecx
145	movl %eax,12(%edi)
146	movl %edx,%esi
147	jz L003maw_end
148	# Tail Round 4
149	movl 16(%ebx),%eax
150	mull %ebp
151	addl %esi,%eax
152	adcl $0,%edx
153	addl 16(%edi),%eax
154	adcl $0,%edx
155	decl %ecx
156	movl %eax,16(%edi)
157	movl %edx,%esi
158	jz L003maw_end
159	# Tail Round 5
160	movl 20(%ebx),%eax
161	mull %ebp
162	addl %esi,%eax
163	adcl $0,%edx
164	addl 20(%edi),%eax
165	adcl $0,%edx
166	decl %ecx
167	movl %eax,20(%edi)
168	movl %edx,%esi
169	jz L003maw_end
170	# Tail Round 6
171	movl 24(%ebx),%eax
172	mull %ebp
173	addl %esi,%eax
174	adcl $0,%edx
175	addl 24(%edi),%eax
176	adcl $0,%edx
177	movl %eax,24(%edi)
178	movl %edx,%esi
179	L003maw_end:
180	movl %esi,%eax
181	popl %ecx
182	popl %edi
183	popl %esi
184	popl %ebx
185	popl %ebp
186	ret
187	.globl _bn_mul_words
188	.align 4
189	_bn_mul_words:
190	L_bn_mul_words_begin:
191	pushl %ebp
192	pushl %ebx
193	pushl %esi
194	pushl %edi
195
196	xorl %esi,%esi
197	movl 20(%esp),%edi
198	movl 24(%esp),%ebx
199	movl 28(%esp),%ebp
200	movl 32(%esp),%ecx
201	andl $4294967288,%ebp
202	jz L004mw_finish
203	L005mw_loop:
204	# Round 0
205	movl (%ebx),%eax
206	mull %ecx
207	addl %esi,%eax
208	adcl $0,%edx
209	movl %eax,(%edi)
210	movl %edx,%esi
211	# Round 4
212	movl 4(%ebx),%eax
213	mull %ecx
214	addl %esi,%eax
215	adcl $0,%edx
216	movl %eax,4(%edi)
217	movl %edx,%esi
218	# Round 8
219	movl 8(%ebx),%eax
220	mull %ecx
221	addl %esi,%eax
222	adcl $0,%edx
223	movl %eax,8(%edi)
224	movl %edx,%esi
225	# Round 12
226	movl 12(%ebx),%eax
227	mull %ecx
228	addl %esi,%eax
229	adcl $0,%edx
230	movl %eax,12(%edi)
231	movl %edx,%esi
232	# Round 16
233	movl 16(%ebx),%eax
234	mull %ecx
235	addl %esi,%eax
236	adcl $0,%edx
237	movl %eax,16(%edi)
238	movl %edx,%esi
239	# Round 20
240	movl 20(%ebx),%eax
241	mull %ecx
242	addl %esi,%eax
243	adcl $0,%edx
244	movl %eax,20(%edi)
245	movl %edx,%esi
246	# Round 24
247	movl 24(%ebx),%eax
248	mull %ecx
249	addl %esi,%eax
250	adcl $0,%edx
251	movl %eax,24(%edi)
252	movl %edx,%esi
253	# Round 28
254	movl 28(%ebx),%eax
255	mull %ecx
256	addl %esi,%eax
257	adcl $0,%edx
258	movl %eax,28(%edi)
259	movl %edx,%esi
260
261	addl $32,%ebx
262	addl $32,%edi
263	subl $8,%ebp
264	jz L004mw_finish
265	jmp L005mw_loop
266	L004mw_finish:
267	movl 28(%esp),%ebp
268	andl $7,%ebp
269	jnz L006mw_finish2
270	jmp L007mw_end
271	L006mw_finish2:
272	# Tail Round 0
273	movl (%ebx),%eax
274	mull %ecx
275	addl %esi,%eax
276	adcl $0,%edx
277	movl %eax,(%edi)
278	movl %edx,%esi
279	decl %ebp
280	jz L007mw_end
281	# Tail Round 1
282	movl 4(%ebx),%eax
283	mull %ecx
284	addl %esi,%eax
285	adcl $0,%edx
286	movl %eax,4(%edi)
287	movl %edx,%esi
288	decl %ebp
289	jz L007mw_end
290	# Tail Round 2
291	movl 8(%ebx),%eax
292	mull %ecx
293	addl %esi,%eax
294	adcl $0,%edx
295	movl %eax,8(%edi)
296	movl %edx,%esi
297	decl %ebp
298	jz L007mw_end
299	# Tail Round 3
300	movl 12(%ebx),%eax
301	mull %ecx
302	addl %esi,%eax
303	adcl $0,%edx
304	movl %eax,12(%edi)
305	movl %edx,%esi
306	decl %ebp
307	jz L007mw_end
308	# Tail Round 4
309	movl 16(%ebx),%eax
310	mull %ecx
311	addl %esi,%eax
312	adcl $0,%edx
313	movl %eax,16(%edi)
314	movl %edx,%esi
315	decl %ebp
316	jz L007mw_end
317	# Tail Round 5
318	movl 20(%ebx),%eax
319	mull %ecx
320	addl %esi,%eax
321	adcl $0,%edx
322	movl %eax,20(%edi)
323	movl %edx,%esi
324	decl %ebp
325	jz L007mw_end
326	# Tail Round 6
327	movl 24(%ebx),%eax
328	mull %ecx
329	addl %esi,%eax
330	adcl $0,%edx
331	movl %eax,24(%edi)
332	movl %edx,%esi
333	L007mw_end:
334	movl %esi,%eax
335	popl %edi
336	popl %esi
337	popl %ebx
338	popl %ebp
339	ret
340	.globl _bn_sqr_words
341	.align 4
342	_bn_sqr_words:
343	L_bn_sqr_words_begin:
344	pushl %ebp
345	pushl %ebx
346	pushl %esi
347	pushl %edi
348
349	movl 20(%esp),%esi
350	movl 24(%esp),%edi
351	movl 28(%esp),%ebx
352	andl $4294967288,%ebx
353	jz L008sw_finish
354	L009sw_loop:
355	# Round 0
356	movl (%edi),%eax
357	mull %eax
358	movl %eax,(%esi)
359	movl %edx,4(%esi)
360	# Round 4
361	movl 4(%edi),%eax
362	mull %eax
363	movl %eax,8(%esi)
364	movl %edx,12(%esi)
365	# Round 8
366	movl 8(%edi),%eax
367	mull %eax
368	movl %eax,16(%esi)
369	movl %edx,20(%esi)
370	# Round 12
371	movl 12(%edi),%eax
372	mull %eax
373	movl %eax,24(%esi)
374	movl %edx,28(%esi)
375	# Round 16
376	movl 16(%edi),%eax
377	mull %eax
378	movl %eax,32(%esi)
379	movl %edx,36(%esi)
380	# Round 20
381	movl 20(%edi),%eax
382	mull %eax
383	movl %eax,40(%esi)
384	movl %edx,44(%esi)
385	# Round 24
386	movl 24(%edi),%eax
387	mull %eax
388	movl %eax,48(%esi)
389	movl %edx,52(%esi)
390	# Round 28
391	movl 28(%edi),%eax
392	mull %eax
393	movl %eax,56(%esi)
394	movl %edx,60(%esi)
395
396	addl $32,%edi
397	addl $64,%esi
398	subl $8,%ebx
399	jnz L009sw_loop
400	L008sw_finish:
401	movl 28(%esp),%ebx
402	andl $7,%ebx
403	jz L010sw_end
404	# Tail Round 0
405	movl (%edi),%eax
406	mull %eax
407	movl %eax,(%esi)
408	decl %ebx
409	movl %edx,4(%esi)
410	jz L010sw_end
411	# Tail Round 1
412	movl 4(%edi),%eax
413	mull %eax
414	movl %eax,8(%esi)
415	decl %ebx
416	movl %edx,12(%esi)
417	jz L010sw_end
418	# Tail Round 2
419	movl 8(%edi),%eax
420	mull %eax
421	movl %eax,16(%esi)
422	decl %ebx
423	movl %edx,20(%esi)
424	jz L010sw_end
425	# Tail Round 3
426	movl 12(%edi),%eax
427	mull %eax
428	movl %eax,24(%esi)
429	decl %ebx
430	movl %edx,28(%esi)
431	jz L010sw_end
432	# Tail Round 4
433	movl 16(%edi),%eax
434	mull %eax
435	movl %eax,32(%esi)
436	decl %ebx
437	movl %edx,36(%esi)
438	jz L010sw_end
439	# Tail Round 5
440	movl 20(%edi),%eax
441	mull %eax
442	movl %eax,40(%esi)
443	decl %ebx
444	movl %edx,44(%esi)
445	jz L010sw_end
446	# Tail Round 6
447	movl 24(%edi),%eax
448	mull %eax
449	movl %eax,48(%esi)
450	movl %edx,52(%esi)
451	L010sw_end:
452	popl %edi
453	popl %esi
454	popl %ebx
455	popl %ebp
456	ret
457	.globl _bn_div_words
458	.align 4
459	_bn_div_words:
460	L_bn_div_words_begin:
461	movl 4(%esp),%edx
462	movl 8(%esp),%eax
463	movl 12(%esp),%ecx
464	divl %ecx
465	ret
466	.globl _bn_add_words
467	.align 4
468	_bn_add_words:
469	L_bn_add_words_begin:
470	pushl %ebp
471	pushl %ebx
472	pushl %esi
473	pushl %edi
474
475	movl 20(%esp),%ebx
476	movl 24(%esp),%esi
477	movl 28(%esp),%edi
478	movl 32(%esp),%ebp
479	xorl %eax,%eax
480	andl $4294967288,%ebp
481	jz L011aw_finish
482	L012aw_loop:
483	# Round 0
484	movl (%esi),%ecx
485	movl (%edi),%edx
486	addl %eax,%ecx
487	movl $0,%eax
488	adcl %eax,%eax
489	addl %edx,%ecx
490	adcl $0,%eax
491	movl %ecx,(%ebx)
492	# Round 1
493	movl 4(%esi),%ecx
494	movl 4(%edi),%edx
495	addl %eax,%ecx
496	movl $0,%eax
497	adcl %eax,%eax
498	addl %edx,%ecx
499	adcl $0,%eax
500	movl %ecx,4(%ebx)
501	# Round 2
502	movl 8(%esi),%ecx
503	movl 8(%edi),%edx
504	addl %eax,%ecx
505	movl $0,%eax
506	adcl %eax,%eax
507	addl %edx,%ecx
508	adcl $0,%eax
509	movl %ecx,8(%ebx)
510	# Round 3
511	movl 12(%esi),%ecx
512	movl 12(%edi),%edx
513	addl %eax,%ecx
514	movl $0,%eax
515	adcl %eax,%eax
516	addl %edx,%ecx
517	adcl $0,%eax
518	movl %ecx,12(%ebx)
519	# Round 4
520	movl 16(%esi),%ecx
521	movl 16(%edi),%edx
522	addl %eax,%ecx
523	movl $0,%eax
524	adcl %eax,%eax
525	addl %edx,%ecx
526	adcl $0,%eax
527	movl %ecx,16(%ebx)
528	# Round 5
529	movl 20(%esi),%ecx
530	movl 20(%edi),%edx
531	addl %eax,%ecx
532	movl $0,%eax
533	adcl %eax,%eax
534	addl %edx,%ecx
535	adcl $0,%eax
536	movl %ecx,20(%ebx)
537	# Round 6
538	movl 24(%esi),%ecx
539	movl 24(%edi),%edx
540	addl %eax,%ecx
541	movl $0,%eax
542	adcl %eax,%eax
543	addl %edx,%ecx
544	adcl $0,%eax
545	movl %ecx,24(%ebx)
546	# Round 7
547	movl 28(%esi),%ecx
548	movl 28(%edi),%edx
549	addl %eax,%ecx
550	movl $0,%eax
551	adcl %eax,%eax
552	addl %edx,%ecx
553	adcl $0,%eax
554	movl %ecx,28(%ebx)
555
556	addl $32,%esi
557	addl $32,%edi
558	addl $32,%ebx
559	subl $8,%ebp
560	jnz L012aw_loop
561	L011aw_finish:
562	movl 32(%esp),%ebp
563	andl $7,%ebp
564	jz L013aw_end
565	# Tail Round 0
566	movl (%esi),%ecx
567	movl (%edi),%edx
568	addl %eax,%ecx
569	movl $0,%eax
570	adcl %eax,%eax
571	addl %edx,%ecx
572	adcl $0,%eax
573	decl %ebp
574	movl %ecx,(%ebx)
575	jz L013aw_end
576	# Tail Round 1
577	movl 4(%esi),%ecx
578	movl 4(%edi),%edx
579	addl %eax,%ecx
580	movl $0,%eax
581	adcl %eax,%eax
582	addl %edx,%ecx
583	adcl $0,%eax
584	decl %ebp
585	movl %ecx,4(%ebx)
586	jz L013aw_end
587	# Tail Round 2
588	movl 8(%esi),%ecx
589	movl 8(%edi),%edx
590	addl %eax,%ecx
591	movl $0,%eax
592	adcl %eax,%eax
593	addl %edx,%ecx
594	adcl $0,%eax
595	decl %ebp
596	movl %ecx,8(%ebx)
597	jz L013aw_end
598	# Tail Round 3
599	movl 12(%esi),%ecx
600	movl 12(%edi),%edx
601	addl %eax,%ecx
602	movl $0,%eax
603	adcl %eax,%eax
604	addl %edx,%ecx
605	adcl $0,%eax
606	decl %ebp
607	movl %ecx,12(%ebx)
608	jz L013aw_end
609	# Tail Round 4
610	movl 16(%esi),%ecx
611	movl 16(%edi),%edx
612	addl %eax,%ecx
613	movl $0,%eax
614	adcl %eax,%eax
615	addl %edx,%ecx
616	adcl $0,%eax
617	decl %ebp
618	movl %ecx,16(%ebx)
619	jz L013aw_end
620	# Tail Round 5
621	movl 20(%esi),%ecx
622	movl 20(%edi),%edx
623	addl %eax,%ecx
624	movl $0,%eax
625	adcl %eax,%eax
626	addl %edx,%ecx
627	adcl $0,%eax
628	decl %ebp
629	movl %ecx,20(%ebx)
630	jz L013aw_end
631	# Tail Round 6
632	movl 24(%esi),%ecx
633	movl 24(%edi),%edx
634	addl %eax,%ecx
635	movl $0,%eax
636	adcl %eax,%eax
637	addl %edx,%ecx
638	adcl $0,%eax
639	movl %ecx,24(%ebx)
640	L013aw_end:
641	popl %edi
642	popl %esi
643	popl %ebx
644	popl %ebp
645	ret
646	.globl _bn_sub_words
647	.align 4
648	_bn_sub_words:
649	L_bn_sub_words_begin:
650	pushl %ebp
651	pushl %ebx
652	pushl %esi
653	pushl %edi
654
655	movl 20(%esp),%ebx
656	movl 24(%esp),%esi
657	movl 28(%esp),%edi
658	movl 32(%esp),%ebp
659	xorl %eax,%eax
660	andl $4294967288,%ebp
661	jz L014aw_finish
662	L015aw_loop:
663	# Round 0
664	movl (%esi),%ecx
665	movl (%edi),%edx
666	subl %eax,%ecx
667	movl $0,%eax
668	adcl %eax,%eax
669	subl %edx,%ecx
670	adcl $0,%eax
671	movl %ecx,(%ebx)
672	# Round 1
673	movl 4(%esi),%ecx
674	movl 4(%edi),%edx
675	subl %eax,%ecx
676	movl $0,%eax
677	adcl %eax,%eax
678	subl %edx,%ecx
679	adcl $0,%eax
680	movl %ecx,4(%ebx)
681	# Round 2
682	movl 8(%esi),%ecx
683	movl 8(%edi),%edx
684	subl %eax,%ecx
685	movl $0,%eax
686	adcl %eax,%eax
687	subl %edx,%ecx
688	adcl $0,%eax
689	movl %ecx,8(%ebx)
690	# Round 3
691	movl 12(%esi),%ecx
692	movl 12(%edi),%edx
693	subl %eax,%ecx
694	movl $0,%eax
695	adcl %eax,%eax
696	subl %edx,%ecx
697	adcl $0,%eax
698	movl %ecx,12(%ebx)
699	# Round 4
700	movl 16(%esi),%ecx
701	movl 16(%edi),%edx
702	subl %eax,%ecx
703	movl $0,%eax
704	adcl %eax,%eax
705	subl %edx,%ecx
706	adcl $0,%eax
707	movl %ecx,16(%ebx)
708	# Round 5
709	movl 20(%esi),%ecx
710	movl 20(%edi),%edx
711	subl %eax,%ecx
712	movl $0,%eax
713	adcl %eax,%eax
714	subl %edx,%ecx
715	adcl $0,%eax
716	movl %ecx,20(%ebx)
717	# Round 6
718	movl 24(%esi),%ecx
719	movl 24(%edi),%edx
720	subl %eax,%ecx
721	movl $0,%eax
722	adcl %eax,%eax
723	subl %edx,%ecx
724	adcl $0,%eax
725	movl %ecx,24(%ebx)
726	# Round 7
727	movl 28(%esi),%ecx
728	movl 28(%edi),%edx
729	subl %eax,%ecx
730	movl $0,%eax
731	adcl %eax,%eax
732	subl %edx,%ecx
733	adcl $0,%eax
734	movl %ecx,28(%ebx)
735
736	addl $32,%esi
737	addl $32,%edi
738	addl $32,%ebx
739	subl $8,%ebp
740	jnz L015aw_loop
741	L014aw_finish:
742	movl 32(%esp),%ebp
743	andl $7,%ebp
744	jz L016aw_end
745	# Tail Round 0
746	movl (%esi),%ecx
747	movl (%edi),%edx
748	subl %eax,%ecx
749	movl $0,%eax
750	adcl %eax,%eax
751	subl %edx,%ecx
752	adcl $0,%eax
753	decl %ebp
754	movl %ecx,(%ebx)
755	jz L016aw_end
756	# Tail Round 1
757	movl 4(%esi),%ecx
758	movl 4(%edi),%edx
759	subl %eax,%ecx
760	movl $0,%eax
761	adcl %eax,%eax
762	subl %edx,%ecx
763	adcl $0,%eax
764	decl %ebp
765	movl %ecx,4(%ebx)
766	jz L016aw_end
767	# Tail Round 2
768	movl 8(%esi),%ecx
769	movl 8(%edi),%edx
770	subl %eax,%ecx
771	movl $0,%eax
772	adcl %eax,%eax
773	subl %edx,%ecx
774	adcl $0,%eax
775	decl %ebp
776	movl %ecx,8(%ebx)
777	jz L016aw_end
778	# Tail Round 3
779	movl 12(%esi),%ecx
780	movl 12(%edi),%edx
781	subl %eax,%ecx
782	movl $0,%eax
783	adcl %eax,%eax
784	subl %edx,%ecx
785	adcl $0,%eax
786	decl %ebp
787	movl %ecx,12(%ebx)
788	jz L016aw_end
789	# Tail Round 4
790	movl 16(%esi),%ecx
791	movl 16(%edi),%edx
792	subl %eax,%ecx
793	movl $0,%eax
794	adcl %eax,%eax
795	subl %edx,%ecx
796	adcl $0,%eax
797	decl %ebp
798	movl %ecx,16(%ebx)
799	jz L016aw_end
800	# Tail Round 5
801	movl 20(%esi),%ecx
802	movl 20(%edi),%edx
803	subl %eax,%ecx
804	movl $0,%eax
805	adcl %eax,%eax
806	subl %edx,%ecx
807	adcl $0,%eax
808	decl %ebp
809	movl %ecx,20(%ebx)
810	jz L016aw_end
811	# Tail Round 6
812	movl 24(%esi),%ecx
813	movl 24(%edi),%edx
814	subl %eax,%ecx
815	movl $0,%eax
816	adcl %eax,%eax
817	subl %edx,%ecx
818	adcl $0,%eax
819	movl %ecx,24(%ebx)
820	L016aw_end:
821	popl %edi
822	popl %esi
823	popl %ebx
824	popl %ebp
825	ret
826	.globl _bn_sub_part_words
827	.align 4
828	_bn_sub_part_words:
829	L_bn_sub_part_words_begin:
830	pushl %ebp
831	pushl %ebx
832	pushl %esi
833	pushl %edi
834
835	movl 20(%esp),%ebx
836	movl 24(%esp),%esi
837	movl 28(%esp),%edi
838	movl 32(%esp),%ebp
839	xorl %eax,%eax
840	andl $4294967288,%ebp
841	jz L017aw_finish
842	L018aw_loop:
843	# Round 0
844	movl (%esi),%ecx
845	movl (%edi),%edx
846	subl %eax,%ecx
847	movl $0,%eax
848	adcl %eax,%eax
849	subl %edx,%ecx
850	adcl $0,%eax
851	movl %ecx,(%ebx)
852	# Round 1
853	movl 4(%esi),%ecx
854	movl 4(%edi),%edx
855	subl %eax,%ecx
856	movl $0,%eax
857	adcl %eax,%eax
858	subl %edx,%ecx
859	adcl $0,%eax
860	movl %ecx,4(%ebx)
861	# Round 2
862	movl 8(%esi),%ecx
863	movl 8(%edi),%edx
864	subl %eax,%ecx
865	movl $0,%eax
866	adcl %eax,%eax
867	subl %edx,%ecx
868	adcl $0,%eax
869	movl %ecx,8(%ebx)
870	# Round 3
871	movl 12(%esi),%ecx
872	movl 12(%edi),%edx
873	subl %eax,%ecx
874	movl $0,%eax
875	adcl %eax,%eax
876	subl %edx,%ecx
877	adcl $0,%eax
878	movl %ecx,12(%ebx)
879	# Round 4
880	movl 16(%esi),%ecx
881	movl 16(%edi),%edx
882	subl %eax,%ecx
883	movl $0,%eax
884	adcl %eax,%eax
885	subl %edx,%ecx
886	adcl $0,%eax
887	movl %ecx,16(%ebx)
888	# Round 5
889	movl 20(%esi),%ecx
890	movl 20(%edi),%edx
891	subl %eax,%ecx
892	movl $0,%eax
893	adcl %eax,%eax
894	subl %edx,%ecx
895	adcl $0,%eax
896	movl %ecx,20(%ebx)
897	# Round 6
898	movl 24(%esi),%ecx
899	movl 24(%edi),%edx
900	subl %eax,%ecx
901	movl $0,%eax
902	adcl %eax,%eax
903	subl %edx,%ecx
904	adcl $0,%eax
905	movl %ecx,24(%ebx)
906	# Round 7
907	movl 28(%esi),%ecx
908	movl 28(%edi),%edx
909	subl %eax,%ecx
910	movl $0,%eax
911	adcl %eax,%eax
912	subl %edx,%ecx
913	adcl $0,%eax
914	movl %ecx,28(%ebx)
915
916	addl $32,%esi
917	addl $32,%edi
918	addl $32,%ebx
919	subl $8,%ebp
920	jnz L018aw_loop
921	L017aw_finish:
922	movl 32(%esp),%ebp
923	andl $7,%ebp
924	jz L019aw_end
925	# Tail Round 0
926	movl (%esi),%ecx
927	movl (%edi),%edx
928	subl %eax,%ecx
929	movl $0,%eax
930	adcl %eax,%eax
931	subl %edx,%ecx
932	adcl $0,%eax
933	movl %ecx,(%ebx)
934	addl $4,%esi
935	addl $4,%edi
936	addl $4,%ebx
937	decl %ebp
938	jz L019aw_end
939	# Tail Round 1
940	movl (%esi),%ecx
941	movl (%edi),%edx
942	subl %eax,%ecx
943	movl $0,%eax
944	adcl %eax,%eax
945	subl %edx,%ecx
946	adcl $0,%eax
947	movl %ecx,(%ebx)
948	addl $4,%esi
949	addl $4,%edi
950	addl $4,%ebx
951	decl %ebp
952	jz L019aw_end
953	# Tail Round 2
954	movl (%esi),%ecx
955	movl (%edi),%edx
956	subl %eax,%ecx
957	movl $0,%eax
958	adcl %eax,%eax
959	subl %edx,%ecx
960	adcl $0,%eax
961	movl %ecx,(%ebx)
962	addl $4,%esi
963	addl $4,%edi
964	addl $4,%ebx
965	decl %ebp
966	jz L019aw_end
967	# Tail Round 3
968	movl (%esi),%ecx
969	movl (%edi),%edx
970	subl %eax,%ecx
971	movl $0,%eax
972	adcl %eax,%eax
973	subl %edx,%ecx
974	adcl $0,%eax
975	movl %ecx,(%ebx)
976	addl $4,%esi
977	addl $4,%edi
978	addl $4,%ebx
979	decl %ebp
980	jz L019aw_end
981	# Tail Round 4
982	movl (%esi),%ecx
983	movl (%edi),%edx
984	subl %eax,%ecx
985	movl $0,%eax
986	adcl %eax,%eax
987	subl %edx,%ecx
988	adcl $0,%eax
989	movl %ecx,(%ebx)
990	addl $4,%esi
991	addl $4,%edi
992	addl $4,%ebx
993	decl %ebp
994	jz L019aw_end
995	# Tail Round 5
996	movl (%esi),%ecx
997	movl (%edi),%edx
998	subl %eax,%ecx
999	movl $0,%eax
1000	adcl %eax,%eax
1001	subl %edx,%ecx
1002	adcl $0,%eax
1003	movl %ecx,(%ebx)
1004	addl $4,%esi
1005	addl $4,%edi
1006	addl $4,%ebx
1007	decl %ebp
1008	jz L019aw_end
1009	# Tail Round 6
1010	movl (%esi),%ecx
1011	movl (%edi),%edx
1012	subl %eax,%ecx
1013	movl $0,%eax
1014	adcl %eax,%eax
1015	subl %edx,%ecx
1016	adcl $0,%eax
1017	movl %ecx,(%ebx)
1018	addl $4,%esi
1019	addl $4,%edi
1020	addl $4,%ebx
1021	L019aw_end:
1022	cmpl $0,36(%esp)
1023	je L020pw_end
1024	movl 36(%esp),%ebp
1025	cmpl $0,%ebp
1026	je L020pw_end
1027	jge L021pw_pos
1028	# pw_neg
1029	movl $0,%edx
1030	subl %ebp,%edx
1031	movl %edx,%ebp
1032	andl $4294967288,%ebp
1033	jz L022pw_neg_finish
1034	L023pw_neg_loop:
1035	# dl<0 Round 0
1036	movl $0,%ecx
1037	movl (%edi),%edx
1038	subl %eax,%ecx
1039	movl $0,%eax
1040	adcl %eax,%eax
1041	subl %edx,%ecx
1042	adcl $0,%eax
1043	movl %ecx,(%ebx)
1044	# dl<0 Round 1
1045	movl $0,%ecx
1046	movl 4(%edi),%edx
1047	subl %eax,%ecx
1048	movl $0,%eax
1049	adcl %eax,%eax
1050	subl %edx,%ecx
1051	adcl $0,%eax
1052	movl %ecx,4(%ebx)
1053	# dl<0 Round 2
1054	movl $0,%ecx
1055	movl 8(%edi),%edx
1056	subl %eax,%ecx
1057	movl $0,%eax
1058	adcl %eax,%eax
1059	subl %edx,%ecx
1060	adcl $0,%eax
1061	movl %ecx,8(%ebx)
1062	# dl<0 Round 3
1063	movl $0,%ecx
1064	movl 12(%edi),%edx
1065	subl %eax,%ecx
1066	movl $0,%eax
1067	adcl %eax,%eax
1068	subl %edx,%ecx
1069	adcl $0,%eax
1070	movl %ecx,12(%ebx)
1071	# dl<0 Round 4
1072	movl $0,%ecx
1073	movl 16(%edi),%edx
1074	subl %eax,%ecx
1075	movl $0,%eax
1076	adcl %eax,%eax
1077	subl %edx,%ecx
1078	adcl $0,%eax
1079	movl %ecx,16(%ebx)
1080	# dl<0 Round 5
1081	movl $0,%ecx
1082	movl 20(%edi),%edx
1083	subl %eax,%ecx
1084	movl $0,%eax
1085	adcl %eax,%eax
1086	subl %edx,%ecx
1087	adcl $0,%eax
1088	movl %ecx,20(%ebx)
1089	# dl<0 Round 6
1090	movl $0,%ecx
1091	movl 24(%edi),%edx
1092	subl %eax,%ecx
1093	movl $0,%eax
1094	adcl %eax,%eax
1095	subl %edx,%ecx
1096	adcl $0,%eax
1097	movl %ecx,24(%ebx)
1098	# dl<0 Round 7
1099	movl $0,%ecx
1100	movl 28(%edi),%edx
1101	subl %eax,%ecx
1102	movl $0,%eax
1103	adcl %eax,%eax
1104	subl %edx,%ecx
1105	adcl $0,%eax
1106	movl %ecx,28(%ebx)
1107
1108	addl $32,%edi
1109	addl $32,%ebx
1110	subl $8,%ebp
1111	jnz L023pw_neg_loop
1112	L022pw_neg_finish:
1113	movl 36(%esp),%edx
1114	movl $0,%ebp
1115	subl %edx,%ebp
1116	andl $7,%ebp
1117	jz L020pw_end
1118	# dl<0 Tail Round 0
1119	movl $0,%ecx
1120	movl (%edi),%edx
1121	subl %eax,%ecx
1122	movl $0,%eax
1123	adcl %eax,%eax
1124	subl %edx,%ecx
1125	adcl $0,%eax
1126	decl %ebp
1127	movl %ecx,(%ebx)
1128	jz L020pw_end
1129	# dl<0 Tail Round 1
1130	movl $0,%ecx
1131	movl 4(%edi),%edx
1132	subl %eax,%ecx
1133	movl $0,%eax
1134	adcl %eax,%eax
1135	subl %edx,%ecx
1136	adcl $0,%eax
1137	decl %ebp
1138	movl %ecx,4(%ebx)
1139	jz L020pw_end
1140	# dl<0 Tail Round 2
1141	movl $0,%ecx
1142	movl 8(%edi),%edx
1143	subl %eax,%ecx
1144	movl $0,%eax
1145	adcl %eax,%eax
1146	subl %edx,%ecx
1147	adcl $0,%eax
1148	decl %ebp
1149	movl %ecx,8(%ebx)
1150	jz L020pw_end
1151	# dl<0 Tail Round 3
1152	movl $0,%ecx
1153	movl 12(%edi),%edx
1154	subl %eax,%ecx
1155	movl $0,%eax
1156	adcl %eax,%eax
1157	subl %edx,%ecx
1158	adcl $0,%eax
1159	decl %ebp
1160	movl %ecx,12(%ebx)
1161	jz L020pw_end
1162	# dl<0 Tail Round 4
1163	movl $0,%ecx
1164	movl 16(%edi),%edx
1165	subl %eax,%ecx
1166	movl $0,%eax
1167	adcl %eax,%eax
1168	subl %edx,%ecx
1169	adcl $0,%eax
1170	decl %ebp
1171	movl %ecx,16(%ebx)
1172	jz L020pw_end
1173	# dl<0 Tail Round 5
1174	movl $0,%ecx
1175	movl 20(%edi),%edx
1176	subl %eax,%ecx
1177	movl $0,%eax
1178	adcl %eax,%eax
1179	subl %edx,%ecx
1180	adcl $0,%eax
1181	decl %ebp
1182	movl %ecx,20(%ebx)
1183	jz L020pw_end
1184	# dl<0 Tail Round 6
1185	movl $0,%ecx
1186	movl 24(%edi),%edx
1187	subl %eax,%ecx
1188	movl $0,%eax
1189	adcl %eax,%eax
1190	subl %edx,%ecx
1191	adcl $0,%eax
1192	movl %ecx,24(%ebx)
1193	jmp L020pw_end
1194	L021pw_pos:
1195	andl $4294967288,%ebp
1196	jz L024pw_pos_finish
1197	L025pw_pos_loop:
1198	# dl>0 Round 0
1199	movl (%esi),%ecx
1200	subl %eax,%ecx
1201	movl %ecx,(%ebx)
1202	jnc L026pw_nc0
1203	# dl>0 Round 1
1204	movl 4(%esi),%ecx
1205	subl %eax,%ecx
1206	movl %ecx,4(%ebx)
1207	jnc L027pw_nc1
1208	# dl>0 Round 2
1209	movl 8(%esi),%ecx
1210	subl %eax,%ecx
1211	movl %ecx,8(%ebx)
1212	jnc L028pw_nc2
1213	# dl>0 Round 3
1214	movl 12(%esi),%ecx
1215	subl %eax,%ecx
1216	movl %ecx,12(%ebx)
1217	jnc L029pw_nc3
1218	# dl>0 Round 4
1219	movl 16(%esi),%ecx
1220	subl %eax,%ecx
1221	movl %ecx,16(%ebx)
1222	jnc L030pw_nc4
1223	# dl>0 Round 5
1224	movl 20(%esi),%ecx
1225	subl %eax,%ecx
1226	movl %ecx,20(%ebx)
1227	jnc L031pw_nc5
1228	# dl>0 Round 6
1229	movl 24(%esi),%ecx
1230	subl %eax,%ecx
1231	movl %ecx,24(%ebx)
1232	jnc L032pw_nc6
1233	# dl>0 Round 7
1234	movl 28(%esi),%ecx
1235	subl %eax,%ecx
1236	movl %ecx,28(%ebx)
1237	jnc L033pw_nc7
1238
1239	addl $32,%esi
1240	addl $32,%ebx
1241	subl $8,%ebp
1242	jnz L025pw_pos_loop
1243	L024pw_pos_finish:
1244	movl 36(%esp),%ebp
1245	andl $7,%ebp
1246	jz L020pw_end
1247	# dl>0 Tail Round 0
1248	movl (%esi),%ecx
1249	subl %eax,%ecx
1250	movl %ecx,(%ebx)
1251	jnc L034pw_tail_nc0
1252	decl %ebp
1253	jz L020pw_end
1254	# dl>0 Tail Round 1
1255	movl 4(%esi),%ecx
1256	subl %eax,%ecx
1257	movl %ecx,4(%ebx)
1258	jnc L035pw_tail_nc1
1259	decl %ebp
1260	jz L020pw_end
1261	# dl>0 Tail Round 2
1262	movl 8(%esi),%ecx
1263	subl %eax,%ecx
1264	movl %ecx,8(%ebx)
1265	jnc L036pw_tail_nc2
1266	decl %ebp
1267	jz L020pw_end
1268	# dl>0 Tail Round 3
1269	movl 12(%esi),%ecx
1270	subl %eax,%ecx
1271	movl %ecx,12(%ebx)
1272	jnc L037pw_tail_nc3
1273	decl %ebp
1274	jz L020pw_end
1275	# dl>0 Tail Round 4
1276	movl 16(%esi),%ecx
1277	subl %eax,%ecx
1278	movl %ecx,16(%ebx)
1279	jnc L038pw_tail_nc4
1280	decl %ebp
1281	jz L020pw_end
1282	# dl>0 Tail Round 5
1283	movl 20(%esi),%ecx
1284	subl %eax,%ecx
1285	movl %ecx,20(%ebx)
1286	jnc L039pw_tail_nc5
1287	decl %ebp
1288	jz L020pw_end
1289	# dl>0 Tail Round 6
1290	movl 24(%esi),%ecx
1291	subl %eax,%ecx
1292	movl %ecx,24(%ebx)
1293	jnc L040pw_tail_nc6
1294	movl $1,%eax
1295	jmp L020pw_end
1296	L041pw_nc_loop:
1297	movl (%esi),%ecx
1298	movl %ecx,(%ebx)
1299	L026pw_nc0:
1300	movl 4(%esi),%ecx
1301	movl %ecx,4(%ebx)
1302	L027pw_nc1:
1303	movl 8(%esi),%ecx
1304	movl %ecx,8(%ebx)
1305	L028pw_nc2:
1306	movl 12(%esi),%ecx
1307	movl %ecx,12(%ebx)
1308	L029pw_nc3:
1309	movl 16(%esi),%ecx
1310	movl %ecx,16(%ebx)
1311	L030pw_nc4:
1312	movl 20(%esi),%ecx
1313	movl %ecx,20(%ebx)
1314	L031pw_nc5:
1315	movl 24(%esi),%ecx
1316	movl %ecx,24(%ebx)
1317	L032pw_nc6:
1318	movl 28(%esi),%ecx
1319	movl %ecx,28(%ebx)
1320	L033pw_nc7:
1321
1322	addl $32,%esi
1323	addl $32,%ebx
1324	subl $8,%ebp
1325	jnz L041pw_nc_loop
1326	movl 36(%esp),%ebp
1327	andl $7,%ebp
1328	jz L042pw_nc_end
1329	movl (%esi),%ecx
1330	movl %ecx,(%ebx)
1331	L034pw_tail_nc0:
1332	decl %ebp
1333	jz L042pw_nc_end
1334	movl 4(%esi),%ecx
1335	movl %ecx,4(%ebx)
1336	L035pw_tail_nc1:
1337	decl %ebp
1338	jz L042pw_nc_end
1339	movl 8(%esi),%ecx
1340	movl %ecx,8(%ebx)
1341	L036pw_tail_nc2:
1342	decl %ebp
1343	jz L042pw_nc_end
1344	movl 12(%esi),%ecx
1345	movl %ecx,12(%ebx)
1346	L037pw_tail_nc3:
1347	decl %ebp
1348	jz L042pw_nc_end
1349	movl 16(%esi),%ecx
1350	movl %ecx,16(%ebx)
1351	L038pw_tail_nc4:
1352	decl %ebp
1353	jz L042pw_nc_end
1354	movl 20(%esi),%ecx
1355	movl %ecx,20(%ebx)
1356	L039pw_tail_nc5:
1357	decl %ebp
1358	jz L042pw_nc_end
1359	movl 24(%esi),%ecx
1360	movl %ecx,24(%ebx)
1361	L040pw_tail_nc6:
1362	L042pw_nc_end:
1363	movl $0,%eax
1364	L020pw_end:
1365	popl %edi
1366	popl %esi
1367	popl %ebx
1368	popl %ebp
1369	ret

Note: See TracBrowser for help on using the repository browser.

Download in other formats:

Original Format