x86_64-mont.S@ 88461

Last change on this file since 88461 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: 15.9 KB

Line
1	default rel
2	%define XMMWORD
3	%define YMMWORD
4	%define ZMMWORD
5	section .text code align=64
6
7
8	EXTERN OPENSSL_ia32cap_P
9
10	global bn_mul_mont
11
12	ALIGN 16
13	bn_mul_mont:
14	mov QWORD[8+rsp],rdi ;WIN64 prologue
15	mov QWORD[16+rsp],rsi
16	mov rax,rsp
17	$L$SEH_begin_bn_mul_mont:
18	mov rdi,rcx
19	mov rsi,rdx
20	mov rdx,r8
21	mov rcx,r9
22	mov r8,QWORD[40+rsp]
23	mov r9,QWORD[48+rsp]
24
25
26
27	mov r9d,r9d
28	mov rax,rsp
29
30	test r9d,3
31	jnz NEAR $L$mul_enter
32	cmp r9d,8
33	jb NEAR $L$mul_enter
34	cmp rdx,rsi
35	jne NEAR $L$mul4x_enter
36	test r9d,7
37	jz NEAR $L$sqr8x_enter
38	jmp NEAR $L$mul4x_enter
39
40	ALIGN 16
41	$L$mul_enter:
42	push rbx
43
44	push rbp
45
46	push r12
47
48	push r13
49
50	push r14
51
52	push r15
53
54
55	neg r9
56	mov r11,rsp
57	lea r10,[((-16))+r9*8+rsp]
58	neg r9
59	and r10,-1024
60
61
62
63
64
65
66
67
68
69	sub r11,r10
70	and r11,-4096
71	lea rsp,[r11*1+r10]
72	mov r11,QWORD[rsp]
73	cmp rsp,r10
74	ja NEAR $L$mul_page_walk
75	jmp NEAR $L$mul_page_walk_done
76
77	ALIGN 16
78	$L$mul_page_walk:
79	lea rsp,[((-4096))+rsp]
80	mov r11,QWORD[rsp]
81	cmp rsp,r10
82	ja NEAR $L$mul_page_walk
83	$L$mul_page_walk_done:
84
85	mov QWORD[8+r9*8+rsp],rax
86
87	$L$mul_body:
88	mov r12,rdx
89	mov r8,QWORD[r8]
90	mov rbx,QWORD[r12]
91	mov rax,QWORD[rsi]
92
93	xor r14,r14
94	xor r15,r15
95
96	mov rbp,r8
97	mul rbx
98	mov r10,rax
99	mov rax,QWORD[rcx]
100
101	imul rbp,r10
102	mov r11,rdx
103
104	mul rbp
105	add r10,rax
106	mov rax,QWORD[8+rsi]
107	adc rdx,0
108	mov r13,rdx
109
110	lea r15,[1+r15]
111	jmp NEAR $L$1st_enter
112
113	ALIGN 16
114	$L$1st:
115	add r13,rax
116	mov rax,QWORD[r15*8+rsi]
117	adc rdx,0
118	add r13,r11
119	mov r11,r10
120	adc rdx,0
121	mov QWORD[((-16))+r15*8+rsp],r13
122	mov r13,rdx
123
124	$L$1st_enter:
125	mul rbx
126	add r11,rax
127	mov rax,QWORD[r15*8+rcx]
128	adc rdx,0
129	lea r15,[1+r15]
130	mov r10,rdx
131
132	mul rbp
133	cmp r15,r9
134	jne NEAR $L$1st
135
136	add r13,rax
137	mov rax,QWORD[rsi]
138	adc rdx,0
139	add r13,r11
140	adc rdx,0
141	mov QWORD[((-16))+r15*8+rsp],r13
142	mov r13,rdx
143	mov r11,r10
144
145	xor rdx,rdx
146	add r13,r11
147	adc rdx,0
148	mov QWORD[((-8))+r9*8+rsp],r13
149	mov QWORD[r9*8+rsp],rdx
150
151	lea r14,[1+r14]
152	jmp NEAR $L$outer
153	ALIGN 16
154	$L$outer:
155	mov rbx,QWORD[r14*8+r12]
156	xor r15,r15
157	mov rbp,r8
158	mov r10,QWORD[rsp]
159	mul rbx
160	add r10,rax
161	mov rax,QWORD[rcx]
162	adc rdx,0
163
164	imul rbp,r10
165	mov r11,rdx
166
167	mul rbp
168	add r10,rax
169	mov rax,QWORD[8+rsi]
170	adc rdx,0
171	mov r10,QWORD[8+rsp]
172	mov r13,rdx
173
174	lea r15,[1+r15]
175	jmp NEAR $L$inner_enter
176
177	ALIGN 16
178	$L$inner:
179	add r13,rax
180	mov rax,QWORD[r15*8+rsi]
181	adc rdx,0
182	add r13,r10
183	mov r10,QWORD[r15*8+rsp]
184	adc rdx,0
185	mov QWORD[((-16))+r15*8+rsp],r13
186	mov r13,rdx
187
188	$L$inner_enter:
189	mul rbx
190	add r11,rax
191	mov rax,QWORD[r15*8+rcx]
192	adc rdx,0
193	add r10,r11
194	mov r11,rdx
195	adc r11,0
196	lea r15,[1+r15]
197
198	mul rbp
199	cmp r15,r9
200	jne NEAR $L$inner
201
202	add r13,rax
203	mov rax,QWORD[rsi]
204	adc rdx,0
205	add r13,r10
206	mov r10,QWORD[r15*8+rsp]
207	adc rdx,0
208	mov QWORD[((-16))+r15*8+rsp],r13
209	mov r13,rdx
210
211	xor rdx,rdx
212	add r13,r11
213	adc rdx,0
214	add r13,r10
215	adc rdx,0
216	mov QWORD[((-8))+r9*8+rsp],r13
217	mov QWORD[r9*8+rsp],rdx
218
219	lea r14,[1+r14]
220	cmp r14,r9
221	jb NEAR $L$outer
222
223	xor r14,r14
224	mov rax,QWORD[rsp]
225	mov r15,r9
226
227	ALIGN 16
228	$L$sub: sbb rax,QWORD[r14*8+rcx]
229	mov QWORD[r14*8+rdi],rax
230	mov rax,QWORD[8+r14*8+rsp]
231	lea r14,[1+r14]
232	dec r15
233	jnz NEAR $L$sub
234
235	sbb rax,0
236	mov rbx,-1
237	xor rbx,rax
238	xor r14,r14
239	mov r15,r9
240
241	$L$copy:
242	mov rcx,QWORD[r14*8+rdi]
243	mov rdx,QWORD[r14*8+rsp]
244	and rcx,rbx
245	and rdx,rax
246	mov QWORD[r14*8+rsp],r9
247	or rdx,rcx
248	mov QWORD[r14*8+rdi],rdx
249	lea r14,[1+r14]
250	sub r15,1
251	jnz NEAR $L$copy
252
253	mov rsi,QWORD[8+r9*8+rsp]
254
255	mov rax,1
256	mov r15,QWORD[((-48))+rsi]
257
258	mov r14,QWORD[((-40))+rsi]
259
260	mov r13,QWORD[((-32))+rsi]
261
262	mov r12,QWORD[((-24))+rsi]
263
264	mov rbp,QWORD[((-16))+rsi]
265
266	mov rbx,QWORD[((-8))+rsi]
267
268	lea rsp,[rsi]
269
270	$L$mul_epilogue:
271	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
272	mov rsi,QWORD[16+rsp]
273	DB 0F3h,0C3h ;repret
274
275	$L$SEH_end_bn_mul_mont:
276
277	ALIGN 16
278	bn_mul4x_mont:
279	mov QWORD[8+rsp],rdi ;WIN64 prologue
280	mov QWORD[16+rsp],rsi
281	mov rax,rsp
282	$L$SEH_begin_bn_mul4x_mont:
283	mov rdi,rcx
284	mov rsi,rdx
285	mov rdx,r8
286	mov rcx,r9
287	mov r8,QWORD[40+rsp]
288	mov r9,QWORD[48+rsp]
289
290
291
292	mov r9d,r9d
293	mov rax,rsp
294
295	$L$mul4x_enter:
296	push rbx
297
298	push rbp
299
300	push r12
301
302	push r13
303
304	push r14
305
306	push r15
307
308
309	neg r9
310	mov r11,rsp
311	lea r10,[((-32))+r9*8+rsp]
312	neg r9
313	and r10,-1024
314
315	sub r11,r10
316	and r11,-4096
317	lea rsp,[r11*1+r10]
318	mov r11,QWORD[rsp]
319	cmp rsp,r10
320	ja NEAR $L$mul4x_page_walk
321	jmp NEAR $L$mul4x_page_walk_done
322
323	$L$mul4x_page_walk:
324	lea rsp,[((-4096))+rsp]
325	mov r11,QWORD[rsp]
326	cmp rsp,r10
327	ja NEAR $L$mul4x_page_walk
328	$L$mul4x_page_walk_done:
329
330	mov QWORD[8+r9*8+rsp],rax
331
332	$L$mul4x_body:
333	mov QWORD[16+r9*8+rsp],rdi
334	mov r12,rdx
335	mov r8,QWORD[r8]
336	mov rbx,QWORD[r12]
337	mov rax,QWORD[rsi]
338
339	xor r14,r14
340	xor r15,r15
341
342	mov rbp,r8
343	mul rbx
344	mov r10,rax
345	mov rax,QWORD[rcx]
346
347	imul rbp,r10
348	mov r11,rdx
349
350	mul rbp
351	add r10,rax
352	mov rax,QWORD[8+rsi]
353	adc rdx,0
354	mov rdi,rdx
355
356	mul rbx
357	add r11,rax
358	mov rax,QWORD[8+rcx]
359	adc rdx,0
360	mov r10,rdx
361
362	mul rbp
363	add rdi,rax
364	mov rax,QWORD[16+rsi]
365	adc rdx,0
366	add rdi,r11
367	lea r15,[4+r15]
368	adc rdx,0
369	mov QWORD[rsp],rdi
370	mov r13,rdx
371	jmp NEAR $L$1st4x
372	ALIGN 16
373	$L$1st4x:
374	mul rbx
375	add r10,rax
376	mov rax,QWORD[((-16))+r15*8+rcx]
377	adc rdx,0
378	mov r11,rdx
379
380	mul rbp
381	add r13,rax
382	mov rax,QWORD[((-8))+r15*8+rsi]
383	adc rdx,0
384	add r13,r10
385	adc rdx,0
386	mov QWORD[((-24))+r15*8+rsp],r13
387	mov rdi,rdx
388
389	mul rbx
390	add r11,rax
391	mov rax,QWORD[((-8))+r15*8+rcx]
392	adc rdx,0
393	mov r10,rdx
394
395	mul rbp
396	add rdi,rax
397	mov rax,QWORD[r15*8+rsi]
398	adc rdx,0
399	add rdi,r11
400	adc rdx,0
401	mov QWORD[((-16))+r15*8+rsp],rdi
402	mov r13,rdx
403
404	mul rbx
405	add r10,rax
406	mov rax,QWORD[r15*8+rcx]
407	adc rdx,0
408	mov r11,rdx
409
410	mul rbp
411	add r13,rax
412	mov rax,QWORD[8+r15*8+rsi]
413	adc rdx,0
414	add r13,r10
415	adc rdx,0
416	mov QWORD[((-8))+r15*8+rsp],r13
417	mov rdi,rdx
418
419	mul rbx
420	add r11,rax
421	mov rax,QWORD[8+r15*8+rcx]
422	adc rdx,0
423	lea r15,[4+r15]
424	mov r10,rdx
425
426	mul rbp
427	add rdi,rax
428	mov rax,QWORD[((-16))+r15*8+rsi]
429	adc rdx,0
430	add rdi,r11
431	adc rdx,0
432	mov QWORD[((-32))+r15*8+rsp],rdi
433	mov r13,rdx
434	cmp r15,r9
435	jb NEAR $L$1st4x
436
437	mul rbx
438	add r10,rax
439	mov rax,QWORD[((-16))+r15*8+rcx]
440	adc rdx,0
441	mov r11,rdx
442
443	mul rbp
444	add r13,rax
445	mov rax,QWORD[((-8))+r15*8+rsi]
446	adc rdx,0
447	add r13,r10
448	adc rdx,0
449	mov QWORD[((-24))+r15*8+rsp],r13
450	mov rdi,rdx
451
452	mul rbx
453	add r11,rax
454	mov rax,QWORD[((-8))+r15*8+rcx]
455	adc rdx,0
456	mov r10,rdx
457
458	mul rbp
459	add rdi,rax
460	mov rax,QWORD[rsi]
461	adc rdx,0
462	add rdi,r11
463	adc rdx,0
464	mov QWORD[((-16))+r15*8+rsp],rdi
465	mov r13,rdx
466
467	xor rdi,rdi
468	add r13,r10
469	adc rdi,0
470	mov QWORD[((-8))+r15*8+rsp],r13
471	mov QWORD[r15*8+rsp],rdi
472
473	lea r14,[1+r14]
474	ALIGN 4
475	$L$outer4x:
476	mov rbx,QWORD[r14*8+r12]
477	xor r15,r15
478	mov r10,QWORD[rsp]
479	mov rbp,r8
480	mul rbx
481	add r10,rax
482	mov rax,QWORD[rcx]
483	adc rdx,0
484
485	imul rbp,r10
486	mov r11,rdx
487
488	mul rbp
489	add r10,rax
490	mov rax,QWORD[8+rsi]
491	adc rdx,0
492	mov rdi,rdx
493
494	mul rbx
495	add r11,rax
496	mov rax,QWORD[8+rcx]
497	adc rdx,0
498	add r11,QWORD[8+rsp]
499	adc rdx,0
500	mov r10,rdx
501
502	mul rbp
503	add rdi,rax
504	mov rax,QWORD[16+rsi]
505	adc rdx,0
506	add rdi,r11
507	lea r15,[4+r15]
508	adc rdx,0
509	mov QWORD[rsp],rdi
510	mov r13,rdx
511	jmp NEAR $L$inner4x
512	ALIGN 16
513	$L$inner4x:
514	mul rbx
515	add r10,rax
516	mov rax,QWORD[((-16))+r15*8+rcx]
517	adc rdx,0
518	add r10,QWORD[((-16))+r15*8+rsp]
519	adc rdx,0
520	mov r11,rdx
521
522	mul rbp
523	add r13,rax
524	mov rax,QWORD[((-8))+r15*8+rsi]
525	adc rdx,0
526	add r13,r10
527	adc rdx,0
528	mov QWORD[((-24))+r15*8+rsp],r13
529	mov rdi,rdx
530
531	mul rbx
532	add r11,rax
533	mov rax,QWORD[((-8))+r15*8+rcx]
534	adc rdx,0
535	add r11,QWORD[((-8))+r15*8+rsp]
536	adc rdx,0
537	mov r10,rdx
538
539	mul rbp
540	add rdi,rax
541	mov rax,QWORD[r15*8+rsi]
542	adc rdx,0
543	add rdi,r11
544	adc rdx,0
545	mov QWORD[((-16))+r15*8+rsp],rdi
546	mov r13,rdx
547
548	mul rbx
549	add r10,rax
550	mov rax,QWORD[r15*8+rcx]
551	adc rdx,0
552	add r10,QWORD[r15*8+rsp]
553	adc rdx,0
554	mov r11,rdx
555
556	mul rbp
557	add r13,rax
558	mov rax,QWORD[8+r15*8+rsi]
559	adc rdx,0
560	add r13,r10
561	adc rdx,0
562	mov QWORD[((-8))+r15*8+rsp],r13
563	mov rdi,rdx
564
565	mul rbx
566	add r11,rax
567	mov rax,QWORD[8+r15*8+rcx]
568	adc rdx,0
569	add r11,QWORD[8+r15*8+rsp]
570	adc rdx,0
571	lea r15,[4+r15]
572	mov r10,rdx
573
574	mul rbp
575	add rdi,rax
576	mov rax,QWORD[((-16))+r15*8+rsi]
577	adc rdx,0
578	add rdi,r11
579	adc rdx,0
580	mov QWORD[((-32))+r15*8+rsp],rdi
581	mov r13,rdx
582	cmp r15,r9
583	jb NEAR $L$inner4x
584
585	mul rbx
586	add r10,rax
587	mov rax,QWORD[((-16))+r15*8+rcx]
588	adc rdx,0
589	add r10,QWORD[((-16))+r15*8+rsp]
590	adc rdx,0
591	mov r11,rdx
592
593	mul rbp
594	add r13,rax
595	mov rax,QWORD[((-8))+r15*8+rsi]
596	adc rdx,0
597	add r13,r10
598	adc rdx,0
599	mov QWORD[((-24))+r15*8+rsp],r13
600	mov rdi,rdx
601
602	mul rbx
603	add r11,rax
604	mov rax,QWORD[((-8))+r15*8+rcx]
605	adc rdx,0
606	add r11,QWORD[((-8))+r15*8+rsp]
607	adc rdx,0
608	lea r14,[1+r14]
609	mov r10,rdx
610
611	mul rbp
612	add rdi,rax
613	mov rax,QWORD[rsi]
614	adc rdx,0
615	add rdi,r11
616	adc rdx,0
617	mov QWORD[((-16))+r15*8+rsp],rdi
618	mov r13,rdx
619
620	xor rdi,rdi
621	add r13,r10
622	adc rdi,0
623	add r13,QWORD[r9*8+rsp]
624	adc rdi,0
625	mov QWORD[((-8))+r15*8+rsp],r13
626	mov QWORD[r15*8+rsp],rdi
627
628	cmp r14,r9
629	jb NEAR $L$outer4x
630	mov rdi,QWORD[16+r9*8+rsp]
631	lea r15,[((-4))+r9]
632	mov rax,QWORD[rsp]
633	mov rdx,QWORD[8+rsp]
634	shr r15,2
635	lea rsi,[rsp]
636	xor r14,r14
637
638	sub rax,QWORD[rcx]
639	mov rbx,QWORD[16+rsi]
640	mov rbp,QWORD[24+rsi]
641	sbb rdx,QWORD[8+rcx]
642
643	$L$sub4x:
644	mov QWORD[r14*8+rdi],rax
645	mov QWORD[8+r14*8+rdi],rdx
646	sbb rbx,QWORD[16+r14*8+rcx]
647	mov rax,QWORD[32+r14*8+rsi]
648	mov rdx,QWORD[40+r14*8+rsi]
649	sbb rbp,QWORD[24+r14*8+rcx]
650	mov QWORD[16+r14*8+rdi],rbx
651	mov QWORD[24+r14*8+rdi],rbp
652	sbb rax,QWORD[32+r14*8+rcx]
653	mov rbx,QWORD[48+r14*8+rsi]
654	mov rbp,QWORD[56+r14*8+rsi]
655	sbb rdx,QWORD[40+r14*8+rcx]
656	lea r14,[4+r14]
657	dec r15
658	jnz NEAR $L$sub4x
659
660	mov QWORD[r14*8+rdi],rax
661	mov rax,QWORD[32+r14*8+rsi]
662	sbb rbx,QWORD[16+r14*8+rcx]
663	mov QWORD[8+r14*8+rdi],rdx
664	sbb rbp,QWORD[24+r14*8+rcx]
665	mov QWORD[16+r14*8+rdi],rbx
666
667	sbb rax,0
668	mov QWORD[24+r14*8+rdi],rbp
669	pxor xmm0,xmm0
670	DB 102,72,15,110,224
671	pcmpeqd xmm5,xmm5
672	pshufd xmm4,xmm4,0
673	mov r15,r9
674	pxor xmm5,xmm4
675	shr r15,2
676	xor eax,eax
677
678	jmp NEAR $L$copy4x
679	ALIGN 16
680	$L$copy4x:
681	movdqa xmm1,XMMWORD[rax*1+rsp]
682	movdqu xmm2,XMMWORD[rax*1+rdi]
683	pand xmm1,xmm4
684	pand xmm2,xmm5
685	movdqa xmm3,XMMWORD[16+rax*1+rsp]
686	movdqa XMMWORD[rax*1+rsp],xmm0
687	por xmm1,xmm2
688	movdqu xmm2,XMMWORD[16+rax*1+rdi]
689	movdqu XMMWORD[rax*1+rdi],xmm1
690	pand xmm3,xmm4
691	pand xmm2,xmm5
692	movdqa XMMWORD[16+rax*1+rsp],xmm0
693	por xmm3,xmm2
694	movdqu XMMWORD[16+rax*1+rdi],xmm3
695	lea rax,[32+rax]
696	dec r15
697	jnz NEAR $L$copy4x
698	mov rsi,QWORD[8+r9*8+rsp]
699
700	mov rax,1
701	mov r15,QWORD[((-48))+rsi]
702
703	mov r14,QWORD[((-40))+rsi]
704
705	mov r13,QWORD[((-32))+rsi]
706
707	mov r12,QWORD[((-24))+rsi]
708
709	mov rbp,QWORD[((-16))+rsi]
710
711	mov rbx,QWORD[((-8))+rsi]
712
713	lea rsp,[rsi]
714
715	$L$mul4x_epilogue:
716	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
717	mov rsi,QWORD[16+rsp]
718	DB 0F3h,0C3h ;repret
719
720	$L$SEH_end_bn_mul4x_mont:
721	EXTERN bn_sqr8x_internal
722
723
724	ALIGN 32
725	bn_sqr8x_mont:
726	mov QWORD[8+rsp],rdi ;WIN64 prologue
727	mov QWORD[16+rsp],rsi
728	mov rax,rsp
729	$L$SEH_begin_bn_sqr8x_mont:
730	mov rdi,rcx
731	mov rsi,rdx
732	mov rdx,r8
733	mov rcx,r9
734	mov r8,QWORD[40+rsp]
735	mov r9,QWORD[48+rsp]
736
737
738
739	mov rax,rsp
740
741	$L$sqr8x_enter:
742	push rbx
743
744	push rbp
745
746	push r12
747
748	push r13
749
750	push r14
751
752	push r15
753
754	$L$sqr8x_prologue:
755
756	mov r10d,r9d
757	shl r9d,3
758	shl r10,3+2
759	neg r9
760
761
762
763
764
765
766	lea r11,[((-64))+r9*2+rsp]
767	mov rbp,rsp
768	mov r8,QWORD[r8]
769	sub r11,rsi
770	and r11,4095
771	cmp r10,r11
772	jb NEAR $L$sqr8x_sp_alt
773	sub rbp,r11
774	lea rbp,[((-64))+r9*2+rbp]
775	jmp NEAR $L$sqr8x_sp_done
776
777	ALIGN 32
778	$L$sqr8x_sp_alt:
779	lea r10,[((4096-64))+r9*2]
780	lea rbp,[((-64))+r9*2+rbp]
781	sub r11,r10
782	mov r10,0
783	cmovc r11,r10
784	sub rbp,r11
785	$L$sqr8x_sp_done:
786	and rbp,-64
787	mov r11,rsp
788	sub r11,rbp
789	and r11,-4096
790	lea rsp,[rbp*1+r11]
791	mov r10,QWORD[rsp]
792	cmp rsp,rbp
793	ja NEAR $L$sqr8x_page_walk
794	jmp NEAR $L$sqr8x_page_walk_done
795
796	ALIGN 16
797	$L$sqr8x_page_walk:
798	lea rsp,[((-4096))+rsp]
799	mov r10,QWORD[rsp]
800	cmp rsp,rbp
801	ja NEAR $L$sqr8x_page_walk
802	$L$sqr8x_page_walk_done:
803
804	mov r10,r9
805	neg r9
806
807	mov QWORD[32+rsp],r8
808	mov QWORD[40+rsp],rax
809
810	$L$sqr8x_body:
811
812	DB 102,72,15,110,209
813	pxor xmm0,xmm0
814	DB 102,72,15,110,207
815	DB 102,73,15,110,218
816	call bn_sqr8x_internal
817
818
819
820
821	lea rbx,[r9*1+rdi]
822	mov rcx,r9
823	mov rdx,r9
824	DB 102,72,15,126,207
825	sar rcx,3+2
826	jmp NEAR $L$sqr8x_sub
827
828	ALIGN 32
829	$L$sqr8x_sub:
830	mov r12,QWORD[rbx]
831	mov r13,QWORD[8+rbx]
832	mov r14,QWORD[16+rbx]
833	mov r15,QWORD[24+rbx]
834	lea rbx,[32+rbx]
835	sbb r12,QWORD[rbp]
836	sbb r13,QWORD[8+rbp]
837	sbb r14,QWORD[16+rbp]
838	sbb r15,QWORD[24+rbp]
839	lea rbp,[32+rbp]
840	mov QWORD[rdi],r12
841	mov QWORD[8+rdi],r13
842	mov QWORD[16+rdi],r14
843	mov QWORD[24+rdi],r15
844	lea rdi,[32+rdi]
845	inc rcx
846	jnz NEAR $L$sqr8x_sub
847
848	sbb rax,0
849	lea rbx,[r9*1+rbx]
850	lea rdi,[r9*1+rdi]
851
852	DB 102,72,15,110,200
853	pxor xmm0,xmm0
854	pshufd xmm1,xmm1,0
855	mov rsi,QWORD[40+rsp]
856
857	jmp NEAR $L$sqr8x_cond_copy
858
859	ALIGN 32
860	$L$sqr8x_cond_copy:
861	movdqa xmm2,XMMWORD[rbx]
862	movdqa xmm3,XMMWORD[16+rbx]
863	lea rbx,[32+rbx]
864	movdqu xmm4,XMMWORD[rdi]
865	movdqu xmm5,XMMWORD[16+rdi]
866	lea rdi,[32+rdi]
867	movdqa XMMWORD[(-32)+rbx],xmm0
868	movdqa XMMWORD[(-16)+rbx],xmm0
869	movdqa XMMWORD[(-32)+rdx*1+rbx],xmm0
870	movdqa XMMWORD[(-16)+rdx*1+rbx],xmm0
871	pcmpeqd xmm0,xmm1
872	pand xmm2,xmm1
873	pand xmm3,xmm1
874	pand xmm4,xmm0
875	pand xmm5,xmm0
876	pxor xmm0,xmm0
877	por xmm4,xmm2
878	por xmm5,xmm3
879	movdqu XMMWORD[(-32)+rdi],xmm4
880	movdqu XMMWORD[(-16)+rdi],xmm5
881	add r9,32
882	jnz NEAR $L$sqr8x_cond_copy
883
884	mov rax,1
885	mov r15,QWORD[((-48))+rsi]
886
887	mov r14,QWORD[((-40))+rsi]
888
889	mov r13,QWORD[((-32))+rsi]
890
891	mov r12,QWORD[((-24))+rsi]
892
893	mov rbp,QWORD[((-16))+rsi]
894
895	mov rbx,QWORD[((-8))+rsi]
896
897	lea rsp,[rsi]
898
899	$L$sqr8x_epilogue:
900	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
901	mov rsi,QWORD[16+rsp]
902	DB 0F3h,0C3h ;repret
903
904	$L$SEH_end_bn_sqr8x_mont:
905	DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
906	DB 112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56
907	DB 54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83
908	DB 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115
909	DB 115,108,46,111,114,103,62,0
910	ALIGN 16
911	EXTERN __imp_RtlVirtualUnwind
912
913	ALIGN 16
914	mul_handler:
915	push rsi
916	push rdi
917	push rbx
918	push rbp
919	push r12
920	push r13
921	push r14
922	push r15
923	pushfq
924	sub rsp,64
925
926	mov rax,QWORD[120+r8]
927	mov rbx,QWORD[248+r8]
928
929	mov rsi,QWORD[8+r9]
930	mov r11,QWORD[56+r9]
931
932	mov r10d,DWORD[r11]
933	lea r10,[r10*1+rsi]
934	cmp rbx,r10
935	jb NEAR $L$common_seh_tail
936
937	mov rax,QWORD[152+r8]
938
939	mov r10d,DWORD[4+r11]
940	lea r10,[r10*1+rsi]
941	cmp rbx,r10
942	jae NEAR $L$common_seh_tail
943
944	mov r10,QWORD[192+r8]
945	mov rax,QWORD[8+r10*8+rax]
946
947	jmp NEAR $L$common_pop_regs
948
949
950
951	ALIGN 16
952	sqr_handler:
953	push rsi
954	push rdi
955	push rbx
956	push rbp
957	push r12
958	push r13
959	push r14
960	push r15
961	pushfq
962	sub rsp,64
963
964	mov rax,QWORD[120+r8]
965	mov rbx,QWORD[248+r8]
966
967	mov rsi,QWORD[8+r9]
968	mov r11,QWORD[56+r9]
969
970	mov r10d,DWORD[r11]
971	lea r10,[r10*1+rsi]
972	cmp rbx,r10
973	jb NEAR $L$common_seh_tail
974
975	mov r10d,DWORD[4+r11]
976	lea r10,[r10*1+rsi]
977	cmp rbx,r10
978	jb NEAR $L$common_pop_regs
979
980	mov rax,QWORD[152+r8]
981
982	mov r10d,DWORD[8+r11]
983	lea r10,[r10*1+rsi]
984	cmp rbx,r10
985	jae NEAR $L$common_seh_tail
986
987	mov rax,QWORD[40+rax]
988
989	$L$common_pop_regs:
990	mov rbx,QWORD[((-8))+rax]
991	mov rbp,QWORD[((-16))+rax]
992	mov r12,QWORD[((-24))+rax]
993	mov r13,QWORD[((-32))+rax]
994	mov r14,QWORD[((-40))+rax]
995	mov r15,QWORD[((-48))+rax]
996	mov QWORD[144+r8],rbx
997	mov QWORD[160+r8],rbp
998	mov QWORD[216+r8],r12
999	mov QWORD[224+r8],r13
1000	mov QWORD[232+r8],r14
1001	mov QWORD[240+r8],r15
1002
1003	$L$common_seh_tail:
1004	mov rdi,QWORD[8+rax]
1005	mov rsi,QWORD[16+rax]
1006	mov QWORD[152+r8],rax
1007	mov QWORD[168+r8],rsi
1008	mov QWORD[176+r8],rdi
1009
1010	mov rdi,QWORD[40+r9]
1011	mov rsi,r8
1012	mov ecx,154
1013	DD 0xa548f3fc
1014
1015	mov rsi,r9
1016	xor rcx,rcx
1017	mov rdx,QWORD[8+rsi]
1018	mov r8,QWORD[rsi]
1019	mov r9,QWORD[16+rsi]
1020	mov r10,QWORD[40+rsi]
1021	lea r11,[56+rsi]
1022	lea r12,[24+rsi]
1023	mov QWORD[32+rsp],r10
1024	mov QWORD[40+rsp],r11
1025	mov QWORD[48+rsp],r12
1026	mov QWORD[56+rsp],rcx
1027	call QWORD[__imp_RtlVirtualUnwind]
1028
1029	mov eax,1
1030	add rsp,64
1031	popfq
1032	pop r15
1033	pop r14
1034	pop r13
1035	pop r12
1036	pop rbp
1037	pop rbx
1038	pop rdi
1039	pop rsi
1040	DB 0F3h,0C3h ;repret
1041
1042
1043	section .pdata rdata align=4
1044	ALIGN 4
1045	DD $L$SEH_begin_bn_mul_mont wrt ..imagebase
1046	DD $L$SEH_end_bn_mul_mont wrt ..imagebase
1047	DD $L$SEH_info_bn_mul_mont wrt ..imagebase
1048
1049	DD $L$SEH_begin_bn_mul4x_mont wrt ..imagebase
1050	DD $L$SEH_end_bn_mul4x_mont wrt ..imagebase
1051	DD $L$SEH_info_bn_mul4x_mont wrt ..imagebase
1052
1053	DD $L$SEH_begin_bn_sqr8x_mont wrt ..imagebase
1054	DD $L$SEH_end_bn_sqr8x_mont wrt ..imagebase
1055	DD $L$SEH_info_bn_sqr8x_mont wrt ..imagebase
1056	section .xdata rdata align=8
1057	ALIGN 8
1058	$L$SEH_info_bn_mul_mont:
1059	DB 9,0,0,0
1060	DD mul_handler wrt ..imagebase
1061	DD $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
1062	$L$SEH_info_bn_mul4x_mont:
1063	DB 9,0,0,0
1064	DD mul_handler wrt ..imagebase
1065	DD $L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
1066	$L$SEH_info_bn_sqr8x_mont:
1067	DB 9,0,0,0
1068	DD sqr_handler wrt ..imagebase
1069	DD $L$sqr8x_prologue wrt ..imagebase,$L$sqr8x_body wrt ..imagebase,$L$sqr8x_epilogue wrt ..imagebase
1070	ALIGN 8

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

Download in other formats:

Original Format