x86_64-mont5.S@ 107278

最後變更在這個檔案從107278是 99371,由 vboxsync 提交於 23 月前
openssl-3.1.0: After generating headers and asm (kmk recreate-headers recreate-headers)
檔案大小: 59.7 KB

行
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_gather5
11
12	ALIGN 64
13	bn_mul_mont_gather5:
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_gather5:
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,7
31	jnz NEAR $L$mul_enter
32	mov r11d,DWORD[((OPENSSL_ia32cap_P+8))]
33	jmp NEAR $L$mul4x_enter
34
35	ALIGN 16
36	$L$mul_enter:
37	movd xmm5,DWORD[56+rsp]
38	push rbx
39
40	push rbp
41
42	push r12
43
44	push r13
45
46	push r14
47
48	push r15
49
50
51	neg r9
52	mov r11,rsp
53	lea r10,[((-280))+r9*8+rsp]
54	neg r9
55	and r10,-1024
56
57
58
59
60
61
62
63
64
65	sub r11,r10
66	and r11,-4096
67	lea rsp,[r11*1+r10]
68	mov r11,QWORD[rsp]
69	cmp rsp,r10
70	ja NEAR $L$mul_page_walk
71	jmp NEAR $L$mul_page_walk_done
72
73	$L$mul_page_walk:
74	lea rsp,[((-4096))+rsp]
75	mov r11,QWORD[rsp]
76	cmp rsp,r10
77	ja NEAR $L$mul_page_walk
78	$L$mul_page_walk_done:
79
80	lea r10,[$L$inc]
81	mov QWORD[8+r9*8+rsp],rax
82
83	$L$mul_body:
84
85	lea r12,[128+rdx]
86	movdqa xmm0,XMMWORD[r10]
87	movdqa xmm1,XMMWORD[16+r10]
88	lea r10,[((24-112))+r9*8+rsp]
89	and r10,-16
90
91	pshufd xmm5,xmm5,0
92	movdqa xmm4,xmm1
93	movdqa xmm2,xmm1
94	paddd xmm1,xmm0
95	pcmpeqd xmm0,xmm5
96	DB 0x67
97	movdqa xmm3,xmm4
98	paddd xmm2,xmm1
99	pcmpeqd xmm1,xmm5
100	movdqa XMMWORD[112+r10],xmm0
101	movdqa xmm0,xmm4
102
103	paddd xmm3,xmm2
104	pcmpeqd xmm2,xmm5
105	movdqa XMMWORD[128+r10],xmm1
106	movdqa xmm1,xmm4
107
108	paddd xmm0,xmm3
109	pcmpeqd xmm3,xmm5
110	movdqa XMMWORD[144+r10],xmm2
111	movdqa xmm2,xmm4
112
113	paddd xmm1,xmm0
114	pcmpeqd xmm0,xmm5
115	movdqa XMMWORD[160+r10],xmm3
116	movdqa xmm3,xmm4
117	paddd xmm2,xmm1
118	pcmpeqd xmm1,xmm5
119	movdqa XMMWORD[176+r10],xmm0
120	movdqa xmm0,xmm4
121
122	paddd xmm3,xmm2
123	pcmpeqd xmm2,xmm5
124	movdqa XMMWORD[192+r10],xmm1
125	movdqa xmm1,xmm4
126
127	paddd xmm0,xmm3
128	pcmpeqd xmm3,xmm5
129	movdqa XMMWORD[208+r10],xmm2
130	movdqa xmm2,xmm4
131
132	paddd xmm1,xmm0
133	pcmpeqd xmm0,xmm5
134	movdqa XMMWORD[224+r10],xmm3
135	movdqa xmm3,xmm4
136	paddd xmm2,xmm1
137	pcmpeqd xmm1,xmm5
138	movdqa XMMWORD[240+r10],xmm0
139	movdqa xmm0,xmm4
140
141	paddd xmm3,xmm2
142	pcmpeqd xmm2,xmm5
143	movdqa XMMWORD[256+r10],xmm1
144	movdqa xmm1,xmm4
145
146	paddd xmm0,xmm3
147	pcmpeqd xmm3,xmm5
148	movdqa XMMWORD[272+r10],xmm2
149	movdqa xmm2,xmm4
150
151	paddd xmm1,xmm0
152	pcmpeqd xmm0,xmm5
153	movdqa XMMWORD[288+r10],xmm3
154	movdqa xmm3,xmm4
155	paddd xmm2,xmm1
156	pcmpeqd xmm1,xmm5
157	movdqa XMMWORD[304+r10],xmm0
158
159	paddd xmm3,xmm2
160	DB 0x67
161	pcmpeqd xmm2,xmm5
162	movdqa XMMWORD[320+r10],xmm1
163
164	pcmpeqd xmm3,xmm5
165	movdqa XMMWORD[336+r10],xmm2
166	pand xmm0,XMMWORD[64+r12]
167
168	pand xmm1,XMMWORD[80+r12]
169	pand xmm2,XMMWORD[96+r12]
170	movdqa XMMWORD[352+r10],xmm3
171	pand xmm3,XMMWORD[112+r12]
172	por xmm0,xmm2
173	por xmm1,xmm3
174	movdqa xmm4,XMMWORD[((-128))+r12]
175	movdqa xmm5,XMMWORD[((-112))+r12]
176	movdqa xmm2,XMMWORD[((-96))+r12]
177	pand xmm4,XMMWORD[112+r10]
178	movdqa xmm3,XMMWORD[((-80))+r12]
179	pand xmm5,XMMWORD[128+r10]
180	por xmm0,xmm4
181	pand xmm2,XMMWORD[144+r10]
182	por xmm1,xmm5
183	pand xmm3,XMMWORD[160+r10]
184	por xmm0,xmm2
185	por xmm1,xmm3
186	movdqa xmm4,XMMWORD[((-64))+r12]
187	movdqa xmm5,XMMWORD[((-48))+r12]
188	movdqa xmm2,XMMWORD[((-32))+r12]
189	pand xmm4,XMMWORD[176+r10]
190	movdqa xmm3,XMMWORD[((-16))+r12]
191	pand xmm5,XMMWORD[192+r10]
192	por xmm0,xmm4
193	pand xmm2,XMMWORD[208+r10]
194	por xmm1,xmm5
195	pand xmm3,XMMWORD[224+r10]
196	por xmm0,xmm2
197	por xmm1,xmm3
198	movdqa xmm4,XMMWORD[r12]
199	movdqa xmm5,XMMWORD[16+r12]
200	movdqa xmm2,XMMWORD[32+r12]
201	pand xmm4,XMMWORD[240+r10]
202	movdqa xmm3,XMMWORD[48+r12]
203	pand xmm5,XMMWORD[256+r10]
204	por xmm0,xmm4
205	pand xmm2,XMMWORD[272+r10]
206	por xmm1,xmm5
207	pand xmm3,XMMWORD[288+r10]
208	por xmm0,xmm2
209	por xmm1,xmm3
210	por xmm0,xmm1
211	pshufd xmm1,xmm0,0x4e
212	por xmm0,xmm1
213	lea r12,[256+r12]
214	DB 102,72,15,126,195
215
216	mov r8,QWORD[r8]
217	mov rax,QWORD[rsi]
218
219	xor r14,r14
220	xor r15,r15
221
222	mov rbp,r8
223	mul rbx
224	mov r10,rax
225	mov rax,QWORD[rcx]
226
227	imul rbp,r10
228	mov r11,rdx
229
230	mul rbp
231	add r10,rax
232	mov rax,QWORD[8+rsi]
233	adc rdx,0
234	mov r13,rdx
235
236	lea r15,[1+r15]
237	jmp NEAR $L$1st_enter
238
239	ALIGN 16
240	$L$1st:
241	add r13,rax
242	mov rax,QWORD[r15*8+rsi]
243	adc rdx,0
244	add r13,r11
245	mov r11,r10
246	adc rdx,0
247	mov QWORD[((-16))+r15*8+rsp],r13
248	mov r13,rdx
249
250	$L$1st_enter:
251	mul rbx
252	add r11,rax
253	mov rax,QWORD[r15*8+rcx]
254	adc rdx,0
255	lea r15,[1+r15]
256	mov r10,rdx
257
258	mul rbp
259	cmp r15,r9
260	jne NEAR $L$1st
261
262
263	add r13,rax
264	adc rdx,0
265	add r13,r11
266	adc rdx,0
267	mov QWORD[((-16))+r9*8+rsp],r13
268	mov r13,rdx
269	mov r11,r10
270
271	xor rdx,rdx
272	add r13,r11
273	adc rdx,0
274	mov QWORD[((-8))+r9*8+rsp],r13
275	mov QWORD[r9*8+rsp],rdx
276
277	lea r14,[1+r14]
278	jmp NEAR $L$outer
279	ALIGN 16
280	$L$outer:
281	lea rdx,[((24+128))+r9*8+rsp]
282	and rdx,-16
283	pxor xmm4,xmm4
284	pxor xmm5,xmm5
285	movdqa xmm0,XMMWORD[((-128))+r12]
286	movdqa xmm1,XMMWORD[((-112))+r12]
287	movdqa xmm2,XMMWORD[((-96))+r12]
288	movdqa xmm3,XMMWORD[((-80))+r12]
289	pand xmm0,XMMWORD[((-128))+rdx]
290	pand xmm1,XMMWORD[((-112))+rdx]
291	por xmm4,xmm0
292	pand xmm2,XMMWORD[((-96))+rdx]
293	por xmm5,xmm1
294	pand xmm3,XMMWORD[((-80))+rdx]
295	por xmm4,xmm2
296	por xmm5,xmm3
297	movdqa xmm0,XMMWORD[((-64))+r12]
298	movdqa xmm1,XMMWORD[((-48))+r12]
299	movdqa xmm2,XMMWORD[((-32))+r12]
300	movdqa xmm3,XMMWORD[((-16))+r12]
301	pand xmm0,XMMWORD[((-64))+rdx]
302	pand xmm1,XMMWORD[((-48))+rdx]
303	por xmm4,xmm0
304	pand xmm2,XMMWORD[((-32))+rdx]
305	por xmm5,xmm1
306	pand xmm3,XMMWORD[((-16))+rdx]
307	por xmm4,xmm2
308	por xmm5,xmm3
309	movdqa xmm0,XMMWORD[r12]
310	movdqa xmm1,XMMWORD[16+r12]
311	movdqa xmm2,XMMWORD[32+r12]
312	movdqa xmm3,XMMWORD[48+r12]
313	pand xmm0,XMMWORD[rdx]
314	pand xmm1,XMMWORD[16+rdx]
315	por xmm4,xmm0
316	pand xmm2,XMMWORD[32+rdx]
317	por xmm5,xmm1
318	pand xmm3,XMMWORD[48+rdx]
319	por xmm4,xmm2
320	por xmm5,xmm3
321	movdqa xmm0,XMMWORD[64+r12]
322	movdqa xmm1,XMMWORD[80+r12]
323	movdqa xmm2,XMMWORD[96+r12]
324	movdqa xmm3,XMMWORD[112+r12]
325	pand xmm0,XMMWORD[64+rdx]
326	pand xmm1,XMMWORD[80+rdx]
327	por xmm4,xmm0
328	pand xmm2,XMMWORD[96+rdx]
329	por xmm5,xmm1
330	pand xmm3,XMMWORD[112+rdx]
331	por xmm4,xmm2
332	por xmm5,xmm3
333	por xmm4,xmm5
334	pshufd xmm0,xmm4,0x4e
335	por xmm0,xmm4
336	lea r12,[256+r12]
337
338	mov rax,QWORD[rsi]
339	DB 102,72,15,126,195
340
341	xor r15,r15
342	mov rbp,r8
343	mov r10,QWORD[rsp]
344
345	mul rbx
346	add r10,rax
347	mov rax,QWORD[rcx]
348	adc rdx,0
349
350	imul rbp,r10
351	mov r11,rdx
352
353	mul rbp
354	add r10,rax
355	mov rax,QWORD[8+rsi]
356	adc rdx,0
357	mov r10,QWORD[8+rsp]
358	mov r13,rdx
359
360	lea r15,[1+r15]
361	jmp NEAR $L$inner_enter
362
363	ALIGN 16
364	$L$inner:
365	add r13,rax
366	mov rax,QWORD[r15*8+rsi]
367	adc rdx,0
368	add r13,r10
369	mov r10,QWORD[r15*8+rsp]
370	adc rdx,0
371	mov QWORD[((-16))+r15*8+rsp],r13
372	mov r13,rdx
373
374	$L$inner_enter:
375	mul rbx
376	add r11,rax
377	mov rax,QWORD[r15*8+rcx]
378	adc rdx,0
379	add r10,r11
380	mov r11,rdx
381	adc r11,0
382	lea r15,[1+r15]
383
384	mul rbp
385	cmp r15,r9
386	jne NEAR $L$inner
387
388	add r13,rax
389	adc rdx,0
390	add r13,r10
391	mov r10,QWORD[r9*8+rsp]
392	adc rdx,0
393	mov QWORD[((-16))+r9*8+rsp],r13
394	mov r13,rdx
395
396	xor rdx,rdx
397	add r13,r11
398	adc rdx,0
399	add r13,r10
400	adc rdx,0
401	mov QWORD[((-8))+r9*8+rsp],r13
402	mov QWORD[r9*8+rsp],rdx
403
404	lea r14,[1+r14]
405	cmp r14,r9
406	jb NEAR $L$outer
407
408	xor r14,r14
409	mov rax,QWORD[rsp]
410	lea rsi,[rsp]
411	mov r15,r9
412	jmp NEAR $L$sub
413	ALIGN 16
414	$L$sub: sbb rax,QWORD[r14*8+rcx]
415	mov QWORD[r14*8+rdi],rax
416	mov rax,QWORD[8+r14*8+rsi]
417	lea r14,[1+r14]
418	dec r15
419	jnz NEAR $L$sub
420
421	sbb rax,0
422	mov rbx,-1
423	xor rbx,rax
424	xor r14,r14
425	mov r15,r9
426
427	$L$copy:
428	mov rcx,QWORD[r14*8+rdi]
429	mov rdx,QWORD[r14*8+rsp]
430	and rcx,rbx
431	and rdx,rax
432	mov QWORD[r14*8+rsp],r14
433	or rdx,rcx
434	mov QWORD[r14*8+rdi],rdx
435	lea r14,[1+r14]
436	sub r15,1
437	jnz NEAR $L$copy
438
439	mov rsi,QWORD[8+r9*8+rsp]
440
441	mov rax,1
442
443	mov r15,QWORD[((-48))+rsi]
444
445	mov r14,QWORD[((-40))+rsi]
446
447	mov r13,QWORD[((-32))+rsi]
448
449	mov r12,QWORD[((-24))+rsi]
450
451	mov rbp,QWORD[((-16))+rsi]
452
453	mov rbx,QWORD[((-8))+rsi]
454
455	lea rsp,[rsi]
456
457	$L$mul_epilogue:
458	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
459	mov rsi,QWORD[16+rsp]
460	DB 0F3h,0C3h ;repret
461
462	$L$SEH_end_bn_mul_mont_gather5:
463
464	ALIGN 32
465	bn_mul4x_mont_gather5:
466	mov QWORD[8+rsp],rdi ;WIN64 prologue
467	mov QWORD[16+rsp],rsi
468	mov rax,rsp
469	$L$SEH_begin_bn_mul4x_mont_gather5:
470	mov rdi,rcx
471	mov rsi,rdx
472	mov rdx,r8
473	mov rcx,r9
474	mov r8,QWORD[40+rsp]
475	mov r9,QWORD[48+rsp]
476
477
478
479	DB 0x67
480	mov rax,rsp
481
482	$L$mul4x_enter:
483	and r11d,0x80108
484	cmp r11d,0x80108
485	je NEAR $L$mulx4x_enter
486	push rbx
487
488	push rbp
489
490	push r12
491
492	push r13
493
494	push r14
495
496	push r15
497
498	$L$mul4x_prologue:
499
500	DB 0x67
501	shl r9d,3
502	lea r10,[r9*2+r9]
503	neg r9
504
505
506
507
508
509
510
511
512
513
514	lea r11,[((-320))+r9*2+rsp]
515	mov rbp,rsp
516	sub r11,rdi
517	and r11,4095
518	cmp r10,r11
519	jb NEAR $L$mul4xsp_alt
520	sub rbp,r11
521	lea rbp,[((-320))+r9*2+rbp]
522	jmp NEAR $L$mul4xsp_done
523
524	ALIGN 32
525	$L$mul4xsp_alt:
526	lea r10,[((4096-320))+r9*2]
527	lea rbp,[((-320))+r9*2+rbp]
528	sub r11,r10
529	mov r10,0
530	cmovc r11,r10
531	sub rbp,r11
532	$L$mul4xsp_done:
533	and rbp,-64
534	mov r11,rsp
535	sub r11,rbp
536	and r11,-4096
537	lea rsp,[rbp*1+r11]
538	mov r10,QWORD[rsp]
539	cmp rsp,rbp
540	ja NEAR $L$mul4x_page_walk
541	jmp NEAR $L$mul4x_page_walk_done
542
543	$L$mul4x_page_walk:
544	lea rsp,[((-4096))+rsp]
545	mov r10,QWORD[rsp]
546	cmp rsp,rbp
547	ja NEAR $L$mul4x_page_walk
548	$L$mul4x_page_walk_done:
549
550	neg r9
551
552	mov QWORD[40+rsp],rax
553
554	$L$mul4x_body:
555
556	call mul4x_internal
557
558	mov rsi,QWORD[40+rsp]
559
560	mov rax,1
561
562	mov r15,QWORD[((-48))+rsi]
563
564	mov r14,QWORD[((-40))+rsi]
565
566	mov r13,QWORD[((-32))+rsi]
567
568	mov r12,QWORD[((-24))+rsi]
569
570	mov rbp,QWORD[((-16))+rsi]
571
572	mov rbx,QWORD[((-8))+rsi]
573
574	lea rsp,[rsi]
575
576	$L$mul4x_epilogue:
577	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
578	mov rsi,QWORD[16+rsp]
579	DB 0F3h,0C3h ;repret
580
581	$L$SEH_end_bn_mul4x_mont_gather5:
582
583
584	ALIGN 32
585	mul4x_internal:
586
587	shl r9,5
588	movd xmm5,DWORD[56+rax]
589	lea rax,[$L$inc]
590	lea r13,[128+r9*1+rdx]
591	shr r9,5
592	movdqa xmm0,XMMWORD[rax]
593	movdqa xmm1,XMMWORD[16+rax]
594	lea r10,[((88-112))+r9*1+rsp]
595	lea r12,[128+rdx]
596
597	pshufd xmm5,xmm5,0
598	movdqa xmm4,xmm1
599	DB 0x67,0x67
600	movdqa xmm2,xmm1
601	paddd xmm1,xmm0
602	pcmpeqd xmm0,xmm5
603	DB 0x67
604	movdqa xmm3,xmm4
605	paddd xmm2,xmm1
606	pcmpeqd xmm1,xmm5
607	movdqa XMMWORD[112+r10],xmm0
608	movdqa xmm0,xmm4
609
610	paddd xmm3,xmm2
611	pcmpeqd xmm2,xmm5
612	movdqa XMMWORD[128+r10],xmm1
613	movdqa xmm1,xmm4
614
615	paddd xmm0,xmm3
616	pcmpeqd xmm3,xmm5
617	movdqa XMMWORD[144+r10],xmm2
618	movdqa xmm2,xmm4
619
620	paddd xmm1,xmm0
621	pcmpeqd xmm0,xmm5
622	movdqa XMMWORD[160+r10],xmm3
623	movdqa xmm3,xmm4
624	paddd xmm2,xmm1
625	pcmpeqd xmm1,xmm5
626	movdqa XMMWORD[176+r10],xmm0
627	movdqa xmm0,xmm4
628
629	paddd xmm3,xmm2
630	pcmpeqd xmm2,xmm5
631	movdqa XMMWORD[192+r10],xmm1
632	movdqa xmm1,xmm4
633
634	paddd xmm0,xmm3
635	pcmpeqd xmm3,xmm5
636	movdqa XMMWORD[208+r10],xmm2
637	movdqa xmm2,xmm4
638
639	paddd xmm1,xmm0
640	pcmpeqd xmm0,xmm5
641	movdqa XMMWORD[224+r10],xmm3
642	movdqa xmm3,xmm4
643	paddd xmm2,xmm1
644	pcmpeqd xmm1,xmm5
645	movdqa XMMWORD[240+r10],xmm0
646	movdqa xmm0,xmm4
647
648	paddd xmm3,xmm2
649	pcmpeqd xmm2,xmm5
650	movdqa XMMWORD[256+r10],xmm1
651	movdqa xmm1,xmm4
652
653	paddd xmm0,xmm3
654	pcmpeqd xmm3,xmm5
655	movdqa XMMWORD[272+r10],xmm2
656	movdqa xmm2,xmm4
657
658	paddd xmm1,xmm0
659	pcmpeqd xmm0,xmm5
660	movdqa XMMWORD[288+r10],xmm3
661	movdqa xmm3,xmm4
662	paddd xmm2,xmm1
663	pcmpeqd xmm1,xmm5
664	movdqa XMMWORD[304+r10],xmm0
665
666	paddd xmm3,xmm2
667	DB 0x67
668	pcmpeqd xmm2,xmm5
669	movdqa XMMWORD[320+r10],xmm1
670
671	pcmpeqd xmm3,xmm5
672	movdqa XMMWORD[336+r10],xmm2
673	pand xmm0,XMMWORD[64+r12]
674
675	pand xmm1,XMMWORD[80+r12]
676	pand xmm2,XMMWORD[96+r12]
677	movdqa XMMWORD[352+r10],xmm3
678	pand xmm3,XMMWORD[112+r12]
679	por xmm0,xmm2
680	por xmm1,xmm3
681	movdqa xmm4,XMMWORD[((-128))+r12]
682	movdqa xmm5,XMMWORD[((-112))+r12]
683	movdqa xmm2,XMMWORD[((-96))+r12]
684	pand xmm4,XMMWORD[112+r10]
685	movdqa xmm3,XMMWORD[((-80))+r12]
686	pand xmm5,XMMWORD[128+r10]
687	por xmm0,xmm4
688	pand xmm2,XMMWORD[144+r10]
689	por xmm1,xmm5
690	pand xmm3,XMMWORD[160+r10]
691	por xmm0,xmm2
692	por xmm1,xmm3
693	movdqa xmm4,XMMWORD[((-64))+r12]
694	movdqa xmm5,XMMWORD[((-48))+r12]
695	movdqa xmm2,XMMWORD[((-32))+r12]
696	pand xmm4,XMMWORD[176+r10]
697	movdqa xmm3,XMMWORD[((-16))+r12]
698	pand xmm5,XMMWORD[192+r10]
699	por xmm0,xmm4
700	pand xmm2,XMMWORD[208+r10]
701	por xmm1,xmm5
702	pand xmm3,XMMWORD[224+r10]
703	por xmm0,xmm2
704	por xmm1,xmm3
705	movdqa xmm4,XMMWORD[r12]
706	movdqa xmm5,XMMWORD[16+r12]
707	movdqa xmm2,XMMWORD[32+r12]
708	pand xmm4,XMMWORD[240+r10]
709	movdqa xmm3,XMMWORD[48+r12]
710	pand xmm5,XMMWORD[256+r10]
711	por xmm0,xmm4
712	pand xmm2,XMMWORD[272+r10]
713	por xmm1,xmm5
714	pand xmm3,XMMWORD[288+r10]
715	por xmm0,xmm2
716	por xmm1,xmm3
717	por xmm0,xmm1
718	pshufd xmm1,xmm0,0x4e
719	por xmm0,xmm1
720	lea r12,[256+r12]
721	DB 102,72,15,126,195
722
723	mov QWORD[((16+8))+rsp],r13
724	mov QWORD[((56+8))+rsp],rdi
725
726	mov r8,QWORD[r8]
727	mov rax,QWORD[rsi]
728	lea rsi,[r9*1+rsi]
729	neg r9
730
731	mov rbp,r8
732	mul rbx
733	mov r10,rax
734	mov rax,QWORD[rcx]
735
736	imul rbp,r10
737	lea r14,[((64+8))+rsp]
738	mov r11,rdx
739
740	mul rbp
741	add r10,rax
742	mov rax,QWORD[8+r9*1+rsi]
743	adc rdx,0
744	mov rdi,rdx
745
746	mul rbx
747	add r11,rax
748	mov rax,QWORD[8+rcx]
749	adc rdx,0
750	mov r10,rdx
751
752	mul rbp
753	add rdi,rax
754	mov rax,QWORD[16+r9*1+rsi]
755	adc rdx,0
756	add rdi,r11
757	lea r15,[32+r9]
758	lea rcx,[32+rcx]
759	adc rdx,0
760	mov QWORD[r14],rdi
761	mov r13,rdx
762	jmp NEAR $L$1st4x
763
764	ALIGN 32
765	$L$1st4x:
766	mul rbx
767	add r10,rax
768	mov rax,QWORD[((-16))+rcx]
769	lea r14,[32+r14]
770	adc rdx,0
771	mov r11,rdx
772
773	mul rbp
774	add r13,rax
775	mov rax,QWORD[((-8))+r15*1+rsi]
776	adc rdx,0
777	add r13,r10
778	adc rdx,0
779	mov QWORD[((-24))+r14],r13
780	mov rdi,rdx
781
782	mul rbx
783	add r11,rax
784	mov rax,QWORD[((-8))+rcx]
785	adc rdx,0
786	mov r10,rdx
787
788	mul rbp
789	add rdi,rax
790	mov rax,QWORD[r15*1+rsi]
791	adc rdx,0
792	add rdi,r11
793	adc rdx,0
794	mov QWORD[((-16))+r14],rdi
795	mov r13,rdx
796
797	mul rbx
798	add r10,rax
799	mov rax,QWORD[rcx]
800	adc rdx,0
801	mov r11,rdx
802
803	mul rbp
804	add r13,rax
805	mov rax,QWORD[8+r15*1+rsi]
806	adc rdx,0
807	add r13,r10
808	adc rdx,0
809	mov QWORD[((-8))+r14],r13
810	mov rdi,rdx
811
812	mul rbx
813	add r11,rax
814	mov rax,QWORD[8+rcx]
815	adc rdx,0
816	mov r10,rdx
817
818	mul rbp
819	add rdi,rax
820	mov rax,QWORD[16+r15*1+rsi]
821	adc rdx,0
822	add rdi,r11
823	lea rcx,[32+rcx]
824	adc rdx,0
825	mov QWORD[r14],rdi
826	mov r13,rdx
827
828	add r15,32
829	jnz NEAR $L$1st4x
830
831	mul rbx
832	add r10,rax
833	mov rax,QWORD[((-16))+rcx]
834	lea r14,[32+r14]
835	adc rdx,0
836	mov r11,rdx
837
838	mul rbp
839	add r13,rax
840	mov rax,QWORD[((-8))+rsi]
841	adc rdx,0
842	add r13,r10
843	adc rdx,0
844	mov QWORD[((-24))+r14],r13
845	mov rdi,rdx
846
847	mul rbx
848	add r11,rax
849	mov rax,QWORD[((-8))+rcx]
850	adc rdx,0
851	mov r10,rdx
852
853	mul rbp
854	add rdi,rax
855	mov rax,QWORD[r9*1+rsi]
856	adc rdx,0
857	add rdi,r11
858	adc rdx,0
859	mov QWORD[((-16))+r14],rdi
860	mov r13,rdx
861
862	lea rcx,[r9*1+rcx]
863
864	xor rdi,rdi
865	add r13,r10
866	adc rdi,0
867	mov QWORD[((-8))+r14],r13
868
869	jmp NEAR $L$outer4x
870
871	ALIGN 32
872	$L$outer4x:
873	lea rdx,[((16+128))+r14]
874	pxor xmm4,xmm4
875	pxor xmm5,xmm5
876	movdqa xmm0,XMMWORD[((-128))+r12]
877	movdqa xmm1,XMMWORD[((-112))+r12]
878	movdqa xmm2,XMMWORD[((-96))+r12]
879	movdqa xmm3,XMMWORD[((-80))+r12]
880	pand xmm0,XMMWORD[((-128))+rdx]
881	pand xmm1,XMMWORD[((-112))+rdx]
882	por xmm4,xmm0
883	pand xmm2,XMMWORD[((-96))+rdx]
884	por xmm5,xmm1
885	pand xmm3,XMMWORD[((-80))+rdx]
886	por xmm4,xmm2
887	por xmm5,xmm3
888	movdqa xmm0,XMMWORD[((-64))+r12]
889	movdqa xmm1,XMMWORD[((-48))+r12]
890	movdqa xmm2,XMMWORD[((-32))+r12]
891	movdqa xmm3,XMMWORD[((-16))+r12]
892	pand xmm0,XMMWORD[((-64))+rdx]
893	pand xmm1,XMMWORD[((-48))+rdx]
894	por xmm4,xmm0
895	pand xmm2,XMMWORD[((-32))+rdx]
896	por xmm5,xmm1
897	pand xmm3,XMMWORD[((-16))+rdx]
898	por xmm4,xmm2
899	por xmm5,xmm3
900	movdqa xmm0,XMMWORD[r12]
901	movdqa xmm1,XMMWORD[16+r12]
902	movdqa xmm2,XMMWORD[32+r12]
903	movdqa xmm3,XMMWORD[48+r12]
904	pand xmm0,XMMWORD[rdx]
905	pand xmm1,XMMWORD[16+rdx]
906	por xmm4,xmm0
907	pand xmm2,XMMWORD[32+rdx]
908	por xmm5,xmm1
909	pand xmm3,XMMWORD[48+rdx]
910	por xmm4,xmm2
911	por xmm5,xmm3
912	movdqa xmm0,XMMWORD[64+r12]
913	movdqa xmm1,XMMWORD[80+r12]
914	movdqa xmm2,XMMWORD[96+r12]
915	movdqa xmm3,XMMWORD[112+r12]
916	pand xmm0,XMMWORD[64+rdx]
917	pand xmm1,XMMWORD[80+rdx]
918	por xmm4,xmm0
919	pand xmm2,XMMWORD[96+rdx]
920	por xmm5,xmm1
921	pand xmm3,XMMWORD[112+rdx]
922	por xmm4,xmm2
923	por xmm5,xmm3
924	por xmm4,xmm5
925	pshufd xmm0,xmm4,0x4e
926	por xmm0,xmm4
927	lea r12,[256+r12]
928	DB 102,72,15,126,195
929
930	mov r10,QWORD[r9*1+r14]
931	mov rbp,r8
932	mul rbx
933	add r10,rax
934	mov rax,QWORD[rcx]
935	adc rdx,0
936
937	imul rbp,r10
938	mov r11,rdx
939	mov QWORD[r14],rdi
940
941	lea r14,[r9*1+r14]
942
943	mul rbp
944	add r10,rax
945	mov rax,QWORD[8+r9*1+rsi]
946	adc rdx,0
947	mov rdi,rdx
948
949	mul rbx
950	add r11,rax
951	mov rax,QWORD[8+rcx]
952	adc rdx,0
953	add r11,QWORD[8+r14]
954	adc rdx,0
955	mov r10,rdx
956
957	mul rbp
958	add rdi,rax
959	mov rax,QWORD[16+r9*1+rsi]
960	adc rdx,0
961	add rdi,r11
962	lea r15,[32+r9]
963	lea rcx,[32+rcx]
964	adc rdx,0
965	mov r13,rdx
966	jmp NEAR $L$inner4x
967
968	ALIGN 32
969	$L$inner4x:
970	mul rbx
971	add r10,rax
972	mov rax,QWORD[((-16))+rcx]
973	adc rdx,0
974	add r10,QWORD[16+r14]
975	lea r14,[32+r14]
976	adc rdx,0
977	mov r11,rdx
978
979	mul rbp
980	add r13,rax
981	mov rax,QWORD[((-8))+r15*1+rsi]
982	adc rdx,0
983	add r13,r10
984	adc rdx,0
985	mov QWORD[((-32))+r14],rdi
986	mov rdi,rdx
987
988	mul rbx
989	add r11,rax
990	mov rax,QWORD[((-8))+rcx]
991	adc rdx,0
992	add r11,QWORD[((-8))+r14]
993	adc rdx,0
994	mov r10,rdx
995
996	mul rbp
997	add rdi,rax
998	mov rax,QWORD[r15*1+rsi]
999	adc rdx,0
1000	add rdi,r11
1001	adc rdx,0
1002	mov QWORD[((-24))+r14],r13
1003	mov r13,rdx
1004
1005	mul rbx
1006	add r10,rax
1007	mov rax,QWORD[rcx]
1008	adc rdx,0
1009	add r10,QWORD[r14]
1010	adc rdx,0
1011	mov r11,rdx
1012
1013	mul rbp
1014	add r13,rax
1015	mov rax,QWORD[8+r15*1+rsi]
1016	adc rdx,0
1017	add r13,r10
1018	adc rdx,0
1019	mov QWORD[((-16))+r14],rdi
1020	mov rdi,rdx
1021
1022	mul rbx
1023	add r11,rax
1024	mov rax,QWORD[8+rcx]
1025	adc rdx,0
1026	add r11,QWORD[8+r14]
1027	adc rdx,0
1028	mov r10,rdx
1029
1030	mul rbp
1031	add rdi,rax
1032	mov rax,QWORD[16+r15*1+rsi]
1033	adc rdx,0
1034	add rdi,r11
1035	lea rcx,[32+rcx]
1036	adc rdx,0
1037	mov QWORD[((-8))+r14],r13
1038	mov r13,rdx
1039
1040	add r15,32
1041	jnz NEAR $L$inner4x
1042
1043	mul rbx
1044	add r10,rax
1045	mov rax,QWORD[((-16))+rcx]
1046	adc rdx,0
1047	add r10,QWORD[16+r14]
1048	lea r14,[32+r14]
1049	adc rdx,0
1050	mov r11,rdx
1051
1052	mul rbp
1053	add r13,rax
1054	mov rax,QWORD[((-8))+rsi]
1055	adc rdx,0
1056	add r13,r10
1057	adc rdx,0
1058	mov QWORD[((-32))+r14],rdi
1059	mov rdi,rdx
1060
1061	mul rbx
1062	add r11,rax
1063	mov rax,rbp
1064	mov rbp,QWORD[((-8))+rcx]
1065	adc rdx,0
1066	add r11,QWORD[((-8))+r14]
1067	adc rdx,0
1068	mov r10,rdx
1069
1070	mul rbp
1071	add rdi,rax
1072	mov rax,QWORD[r9*1+rsi]
1073	adc rdx,0
1074	add rdi,r11
1075	adc rdx,0
1076	mov QWORD[((-24))+r14],r13
1077	mov r13,rdx
1078
1079	mov QWORD[((-16))+r14],rdi
1080	lea rcx,[r9*1+rcx]
1081
1082	xor rdi,rdi
1083	add r13,r10
1084	adc rdi,0
1085	add r13,QWORD[r14]
1086	adc rdi,0
1087	mov QWORD[((-8))+r14],r13
1088
1089	cmp r12,QWORD[((16+8))+rsp]
1090	jb NEAR $L$outer4x
1091	xor rax,rax
1092	sub rbp,r13
1093	adc r15,r15
1094	or rdi,r15
1095	sub rax,rdi
1096	lea rbx,[r9*1+r14]
1097	mov r12,QWORD[rcx]
1098	lea rbp,[rcx]
1099	mov rcx,r9
1100	sar rcx,3+2
1101	mov rdi,QWORD[((56+8))+rsp]
1102	dec r12
1103	xor r10,r10
1104	mov r13,QWORD[8+rbp]
1105	mov r14,QWORD[16+rbp]
1106	mov r15,QWORD[24+rbp]
1107	jmp NEAR $L$sqr4x_sub_entry
1108
1109
1110	global bn_power5
1111
1112	ALIGN 32
1113	bn_power5:
1114	mov QWORD[8+rsp],rdi ;WIN64 prologue
1115	mov QWORD[16+rsp],rsi
1116	mov rax,rsp
1117	$L$SEH_begin_bn_power5:
1118	mov rdi,rcx
1119	mov rsi,rdx
1120	mov rdx,r8
1121	mov rcx,r9
1122	mov r8,QWORD[40+rsp]
1123	mov r9,QWORD[48+rsp]
1124
1125
1126
1127	mov rax,rsp
1128
1129	mov r11d,DWORD[((OPENSSL_ia32cap_P+8))]
1130	and r11d,0x80108
1131	cmp r11d,0x80108
1132	je NEAR $L$powerx5_enter
1133	push rbx
1134
1135	push rbp
1136
1137	push r12
1138
1139	push r13
1140
1141	push r14
1142
1143	push r15
1144
1145	$L$power5_prologue:
1146
1147	shl r9d,3
1148	lea r10d,[r9*2+r9]
1149	neg r9
1150	mov r8,QWORD[r8]
1151
1152
1153
1154
1155
1156
1157
1158
1159	lea r11,[((-320))+r9*2+rsp]
1160	mov rbp,rsp
1161	sub r11,rdi
1162	and r11,4095
1163	cmp r10,r11
1164	jb NEAR $L$pwr_sp_alt
1165	sub rbp,r11
1166	lea rbp,[((-320))+r9*2+rbp]
1167	jmp NEAR $L$pwr_sp_done
1168
1169	ALIGN 32
1170	$L$pwr_sp_alt:
1171	lea r10,[((4096-320))+r9*2]
1172	lea rbp,[((-320))+r9*2+rbp]
1173	sub r11,r10
1174	mov r10,0
1175	cmovc r11,r10
1176	sub rbp,r11
1177	$L$pwr_sp_done:
1178	and rbp,-64
1179	mov r11,rsp
1180	sub r11,rbp
1181	and r11,-4096
1182	lea rsp,[rbp*1+r11]
1183	mov r10,QWORD[rsp]
1184	cmp rsp,rbp
1185	ja NEAR $L$pwr_page_walk
1186	jmp NEAR $L$pwr_page_walk_done
1187
1188	$L$pwr_page_walk:
1189	lea rsp,[((-4096))+rsp]
1190	mov r10,QWORD[rsp]
1191	cmp rsp,rbp
1192	ja NEAR $L$pwr_page_walk
1193	$L$pwr_page_walk_done:
1194
1195	mov r10,r9
1196	neg r9
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207	mov QWORD[32+rsp],r8
1208	mov QWORD[40+rsp],rax
1209
1210	$L$power5_body:
1211	DB 102,72,15,110,207
1212	DB 102,72,15,110,209
1213	DB 102,73,15,110,218
1214	DB 102,72,15,110,226
1215
1216	call __bn_sqr8x_internal
1217	call __bn_post4x_internal
1218	call __bn_sqr8x_internal
1219	call __bn_post4x_internal
1220	call __bn_sqr8x_internal
1221	call __bn_post4x_internal
1222	call __bn_sqr8x_internal
1223	call __bn_post4x_internal
1224	call __bn_sqr8x_internal
1225	call __bn_post4x_internal
1226
1227	DB 102,72,15,126,209
1228	DB 102,72,15,126,226
1229	mov rdi,rsi
1230	mov rax,QWORD[40+rsp]
1231	lea r8,[32+rsp]
1232
1233	call mul4x_internal
1234
1235	mov rsi,QWORD[40+rsp]
1236
1237	mov rax,1
1238	mov r15,QWORD[((-48))+rsi]
1239
1240	mov r14,QWORD[((-40))+rsi]
1241
1242	mov r13,QWORD[((-32))+rsi]
1243
1244	mov r12,QWORD[((-24))+rsi]
1245
1246	mov rbp,QWORD[((-16))+rsi]
1247
1248	mov rbx,QWORD[((-8))+rsi]
1249
1250	lea rsp,[rsi]
1251
1252	$L$power5_epilogue:
1253	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
1254	mov rsi,QWORD[16+rsp]
1255	DB 0F3h,0C3h ;repret
1256
1257	$L$SEH_end_bn_power5:
1258
1259	global bn_sqr8x_internal
1260
1261
1262	ALIGN 32
1263	bn_sqr8x_internal:
1264	__bn_sqr8x_internal:
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339	lea rbp,[32+r10]
1340	lea rsi,[r9*1+rsi]
1341
1342	mov rcx,r9
1343
1344
1345	mov r14,QWORD[((-32))+rbp*1+rsi]
1346	lea rdi,[((48+8))+r9*2+rsp]
1347	mov rax,QWORD[((-24))+rbp*1+rsi]
1348	lea rdi,[((-32))+rbp*1+rdi]
1349	mov rbx,QWORD[((-16))+rbp*1+rsi]
1350	mov r15,rax
1351
1352	mul r14
1353	mov r10,rax
1354	mov rax,rbx
1355	mov r11,rdx
1356	mov QWORD[((-24))+rbp*1+rdi],r10
1357
1358	mul r14
1359	add r11,rax
1360	mov rax,rbx
1361	adc rdx,0
1362	mov QWORD[((-16))+rbp*1+rdi],r11
1363	mov r10,rdx
1364
1365
1366	mov rbx,QWORD[((-8))+rbp*1+rsi]
1367	mul r15
1368	mov r12,rax
1369	mov rax,rbx
1370	mov r13,rdx
1371
1372	lea rcx,[rbp]
1373	mul r14
1374	add r10,rax
1375	mov rax,rbx
1376	mov r11,rdx
1377	adc r11,0
1378	add r10,r12
1379	adc r11,0
1380	mov QWORD[((-8))+rcx*1+rdi],r10
1381	jmp NEAR $L$sqr4x_1st
1382
1383	ALIGN 32
1384	$L$sqr4x_1st:
1385	mov rbx,QWORD[rcx*1+rsi]
1386	mul r15
1387	add r13,rax
1388	mov rax,rbx
1389	mov r12,rdx
1390	adc r12,0
1391
1392	mul r14
1393	add r11,rax
1394	mov rax,rbx
1395	mov rbx,QWORD[8+rcx*1+rsi]
1396	mov r10,rdx
1397	adc r10,0
1398	add r11,r13
1399	adc r10,0
1400
1401
1402	mul r15
1403	add r12,rax
1404	mov rax,rbx
1405	mov QWORD[rcx*1+rdi],r11
1406	mov r13,rdx
1407	adc r13,0
1408
1409	mul r14
1410	add r10,rax
1411	mov rax,rbx
1412	mov rbx,QWORD[16+rcx*1+rsi]
1413	mov r11,rdx
1414	adc r11,0
1415	add r10,r12
1416	adc r11,0
1417
1418	mul r15
1419	add r13,rax
1420	mov rax,rbx
1421	mov QWORD[8+rcx*1+rdi],r10
1422	mov r12,rdx
1423	adc r12,0
1424
1425	mul r14
1426	add r11,rax
1427	mov rax,rbx
1428	mov rbx,QWORD[24+rcx*1+rsi]
1429	mov r10,rdx
1430	adc r10,0
1431	add r11,r13
1432	adc r10,0
1433
1434
1435	mul r15
1436	add r12,rax
1437	mov rax,rbx
1438	mov QWORD[16+rcx*1+rdi],r11
1439	mov r13,rdx
1440	adc r13,0
1441	lea rcx,[32+rcx]
1442
1443	mul r14
1444	add r10,rax
1445	mov rax,rbx
1446	mov r11,rdx
1447	adc r11,0
1448	add r10,r12
1449	adc r11,0
1450	mov QWORD[((-8))+rcx*1+rdi],r10
1451
1452	cmp rcx,0
1453	jne NEAR $L$sqr4x_1st
1454
1455	mul r15
1456	add r13,rax
1457	lea rbp,[16+rbp]
1458	adc rdx,0
1459	add r13,r11
1460	adc rdx,0
1461
1462	mov QWORD[rdi],r13
1463	mov r12,rdx
1464	mov QWORD[8+rdi],rdx
1465	jmp NEAR $L$sqr4x_outer
1466
1467	ALIGN 32
1468	$L$sqr4x_outer:
1469	mov r14,QWORD[((-32))+rbp*1+rsi]
1470	lea rdi,[((48+8))+r9*2+rsp]
1471	mov rax,QWORD[((-24))+rbp*1+rsi]
1472	lea rdi,[((-32))+rbp*1+rdi]
1473	mov rbx,QWORD[((-16))+rbp*1+rsi]
1474	mov r15,rax
1475
1476	mul r14
1477	mov r10,QWORD[((-24))+rbp*1+rdi]
1478	add r10,rax
1479	mov rax,rbx
1480	adc rdx,0
1481	mov QWORD[((-24))+rbp*1+rdi],r10
1482	mov r11,rdx
1483
1484	mul r14
1485	add r11,rax
1486	mov rax,rbx
1487	adc rdx,0
1488	add r11,QWORD[((-16))+rbp*1+rdi]
1489	mov r10,rdx
1490	adc r10,0
1491	mov QWORD[((-16))+rbp*1+rdi],r11
1492
1493	xor r12,r12
1494
1495	mov rbx,QWORD[((-8))+rbp*1+rsi]
1496	mul r15
1497	add r12,rax
1498	mov rax,rbx
1499	adc rdx,0
1500	add r12,QWORD[((-8))+rbp*1+rdi]
1501	mov r13,rdx
1502	adc r13,0
1503
1504	mul r14
1505	add r10,rax
1506	mov rax,rbx
1507	adc rdx,0
1508	add r10,r12
1509	mov r11,rdx
1510	adc r11,0
1511	mov QWORD[((-8))+rbp*1+rdi],r10
1512
1513	lea rcx,[rbp]
1514	jmp NEAR $L$sqr4x_inner
1515
1516	ALIGN 32
1517	$L$sqr4x_inner:
1518	mov rbx,QWORD[rcx*1+rsi]
1519	mul r15
1520	add r13,rax
1521	mov rax,rbx
1522	mov r12,rdx
1523	adc r12,0
1524	add r13,QWORD[rcx*1+rdi]
1525	adc r12,0
1526
1527	DB 0x67
1528	mul r14
1529	add r11,rax
1530	mov rax,rbx
1531	mov rbx,QWORD[8+rcx*1+rsi]
1532	mov r10,rdx
1533	adc r10,0
1534	add r11,r13
1535	adc r10,0
1536
1537	mul r15
1538	add r12,rax
1539	mov QWORD[rcx*1+rdi],r11
1540	mov rax,rbx
1541	mov r13,rdx
1542	adc r13,0
1543	add r12,QWORD[8+rcx*1+rdi]
1544	lea rcx,[16+rcx]
1545	adc r13,0
1546
1547	mul r14
1548	add r10,rax
1549	mov rax,rbx
1550	adc rdx,0
1551	add r10,r12
1552	mov r11,rdx
1553	adc r11,0
1554	mov QWORD[((-8))+rcx*1+rdi],r10
1555
1556	cmp rcx,0
1557	jne NEAR $L$sqr4x_inner
1558
1559	DB 0x67
1560	mul r15
1561	add r13,rax
1562	adc rdx,0
1563	add r13,r11
1564	adc rdx,0
1565
1566	mov QWORD[rdi],r13
1567	mov r12,rdx
1568	mov QWORD[8+rdi],rdx
1569
1570	add rbp,16
1571	jnz NEAR $L$sqr4x_outer
1572
1573
1574	mov r14,QWORD[((-32))+rsi]
1575	lea rdi,[((48+8))+r9*2+rsp]
1576	mov rax,QWORD[((-24))+rsi]
1577	lea rdi,[((-32))+rbp*1+rdi]
1578	mov rbx,QWORD[((-16))+rsi]
1579	mov r15,rax
1580
1581	mul r14
1582	add r10,rax
1583	mov rax,rbx
1584	mov r11,rdx
1585	adc r11,0
1586
1587	mul r14
1588	add r11,rax
1589	mov rax,rbx
1590	mov QWORD[((-24))+rdi],r10
1591	mov r10,rdx
1592	adc r10,0
1593	add r11,r13
1594	mov rbx,QWORD[((-8))+rsi]
1595	adc r10,0
1596
1597	mul r15
1598	add r12,rax
1599	mov rax,rbx
1600	mov QWORD[((-16))+rdi],r11
1601	mov r13,rdx
1602	adc r13,0
1603
1604	mul r14
1605	add r10,rax
1606	mov rax,rbx
1607	mov r11,rdx
1608	adc r11,0
1609	add r10,r12
1610	adc r11,0
1611	mov QWORD[((-8))+rdi],r10
1612
1613	mul r15
1614	add r13,rax
1615	mov rax,QWORD[((-16))+rsi]
1616	adc rdx,0
1617	add r13,r11
1618	adc rdx,0
1619
1620	mov QWORD[rdi],r13
1621	mov r12,rdx
1622	mov QWORD[8+rdi],rdx
1623
1624	mul rbx
1625	add rbp,16
1626	xor r14,r14
1627	sub rbp,r9
1628	xor r15,r15
1629
1630	add rax,r12
1631	adc rdx,0
1632	mov QWORD[8+rdi],rax
1633	mov QWORD[16+rdi],rdx
1634	mov QWORD[24+rdi],r15
1635
1636	mov rax,QWORD[((-16))+rbp*1+rsi]
1637	lea rdi,[((48+8))+rsp]
1638	xor r10,r10
1639	mov r11,QWORD[8+rdi]
1640
1641	lea r12,[r10*2+r14]
1642	shr r10,63
1643	lea r13,[r11*2+rcx]
1644	shr r11,63
1645	or r13,r10
1646	mov r10,QWORD[16+rdi]
1647	mov r14,r11
1648	mul rax
1649	neg r15
1650	mov r11,QWORD[24+rdi]
1651	adc r12,rax
1652	mov rax,QWORD[((-8))+rbp*1+rsi]
1653	mov QWORD[rdi],r12
1654	adc r13,rdx
1655
1656	lea rbx,[r10*2+r14]
1657	mov QWORD[8+rdi],r13
1658	sbb r15,r15
1659	shr r10,63
1660	lea r8,[r11*2+rcx]
1661	shr r11,63
1662	or r8,r10
1663	mov r10,QWORD[32+rdi]
1664	mov r14,r11
1665	mul rax
1666	neg r15
1667	mov r11,QWORD[40+rdi]
1668	adc rbx,rax
1669	mov rax,QWORD[rbp*1+rsi]
1670	mov QWORD[16+rdi],rbx
1671	adc r8,rdx
1672	lea rbp,[16+rbp]
1673	mov QWORD[24+rdi],r8
1674	sbb r15,r15
1675	lea rdi,[64+rdi]
1676	jmp NEAR $L$sqr4x_shift_n_add
1677
1678	ALIGN 32
1679	$L$sqr4x_shift_n_add:
1680	lea r12,[r10*2+r14]
1681	shr r10,63
1682	lea r13,[r11*2+rcx]
1683	shr r11,63
1684	or r13,r10
1685	mov r10,QWORD[((-16))+rdi]
1686	mov r14,r11
1687	mul rax
1688	neg r15
1689	mov r11,QWORD[((-8))+rdi]
1690	adc r12,rax
1691	mov rax,QWORD[((-8))+rbp*1+rsi]
1692	mov QWORD[((-32))+rdi],r12
1693	adc r13,rdx
1694
1695	lea rbx,[r10*2+r14]
1696	mov QWORD[((-24))+rdi],r13
1697	sbb r15,r15
1698	shr r10,63
1699	lea r8,[r11*2+rcx]
1700	shr r11,63
1701	or r8,r10
1702	mov r10,QWORD[rdi]
1703	mov r14,r11
1704	mul rax
1705	neg r15
1706	mov r11,QWORD[8+rdi]
1707	adc rbx,rax
1708	mov rax,QWORD[rbp*1+rsi]
1709	mov QWORD[((-16))+rdi],rbx
1710	adc r8,rdx
1711
1712	lea r12,[r10*2+r14]
1713	mov QWORD[((-8))+rdi],r8
1714	sbb r15,r15
1715	shr r10,63
1716	lea r13,[r11*2+rcx]
1717	shr r11,63
1718	or r13,r10
1719	mov r10,QWORD[16+rdi]
1720	mov r14,r11
1721	mul rax
1722	neg r15
1723	mov r11,QWORD[24+rdi]
1724	adc r12,rax
1725	mov rax,QWORD[8+rbp*1+rsi]
1726	mov QWORD[rdi],r12
1727	adc r13,rdx
1728
1729	lea rbx,[r10*2+r14]
1730	mov QWORD[8+rdi],r13
1731	sbb r15,r15
1732	shr r10,63
1733	lea r8,[r11*2+rcx]
1734	shr r11,63
1735	or r8,r10
1736	mov r10,QWORD[32+rdi]
1737	mov r14,r11
1738	mul rax
1739	neg r15
1740	mov r11,QWORD[40+rdi]
1741	adc rbx,rax
1742	mov rax,QWORD[16+rbp*1+rsi]
1743	mov QWORD[16+rdi],rbx
1744	adc r8,rdx
1745	mov QWORD[24+rdi],r8
1746	sbb r15,r15
1747	lea rdi,[64+rdi]
1748	add rbp,32
1749	jnz NEAR $L$sqr4x_shift_n_add
1750
1751	lea r12,[r10*2+r14]
1752	DB 0x67
1753	shr r10,63
1754	lea r13,[r11*2+rcx]
1755	shr r11,63
1756	or r13,r10
1757	mov r10,QWORD[((-16))+rdi]
1758	mov r14,r11
1759	mul rax
1760	neg r15
1761	mov r11,QWORD[((-8))+rdi]
1762	adc r12,rax
1763	mov rax,QWORD[((-8))+rsi]
1764	mov QWORD[((-32))+rdi],r12
1765	adc r13,rdx
1766
1767	lea rbx,[r10*2+r14]
1768	mov QWORD[((-24))+rdi],r13
1769	sbb r15,r15
1770	shr r10,63
1771	lea r8,[r11*2+rcx]
1772	shr r11,63
1773	or r8,r10
1774	mul rax
1775	neg r15
1776	adc rbx,rax
1777	adc r8,rdx
1778	mov QWORD[((-16))+rdi],rbx
1779	mov QWORD[((-8))+rdi],r8
1780	DB 102,72,15,126,213
1781	__bn_sqr8x_reduction:
1782	xor rax,rax
1783	lea rcx,[rbp*1+r9]
1784	lea rdx,[((48+8))+r9*2+rsp]
1785	mov QWORD[((0+8))+rsp],rcx
1786	lea rdi,[((48+8))+r9*1+rsp]
1787	mov QWORD[((8+8))+rsp],rdx
1788	neg r9
1789	jmp NEAR $L$8x_reduction_loop
1790
1791	ALIGN 32
1792	$L$8x_reduction_loop:
1793	lea rdi,[r9*1+rdi]
1794	DB 0x66
1795	mov rbx,QWORD[rdi]
1796	mov r9,QWORD[8+rdi]
1797	mov r10,QWORD[16+rdi]
1798	mov r11,QWORD[24+rdi]
1799	mov r12,QWORD[32+rdi]
1800	mov r13,QWORD[40+rdi]
1801	mov r14,QWORD[48+rdi]
1802	mov r15,QWORD[56+rdi]
1803	mov QWORD[rdx],rax
1804	lea rdi,[64+rdi]
1805
1806	DB 0x67
1807	mov r8,rbx
1808	imul rbx,QWORD[((32+8))+rsp]
1809	mov rax,QWORD[rbp]
1810	mov ecx,8
1811	jmp NEAR $L$8x_reduce
1812
1813	ALIGN 32
1814	$L$8x_reduce:
1815	mul rbx
1816	mov rax,QWORD[8+rbp]
1817	neg r8
1818	mov r8,rdx
1819	adc r8,0
1820
1821	mul rbx
1822	add r9,rax
1823	mov rax,QWORD[16+rbp]
1824	adc rdx,0
1825	add r8,r9
1826	mov QWORD[((48-8+8))+rcx*8+rsp],rbx
1827	mov r9,rdx
1828	adc r9,0
1829
1830	mul rbx
1831	add r10,rax
1832	mov rax,QWORD[24+rbp]
1833	adc rdx,0
1834	add r9,r10
1835	mov rsi,QWORD[((32+8))+rsp]
1836	mov r10,rdx
1837	adc r10,0
1838
1839	mul rbx
1840	add r11,rax
1841	mov rax,QWORD[32+rbp]
1842	adc rdx,0
1843	imul rsi,r8
1844	add r10,r11
1845	mov r11,rdx
1846	adc r11,0
1847
1848	mul rbx
1849	add r12,rax
1850	mov rax,QWORD[40+rbp]
1851	adc rdx,0
1852	add r11,r12
1853	mov r12,rdx
1854	adc r12,0
1855
1856	mul rbx
1857	add r13,rax
1858	mov rax,QWORD[48+rbp]
1859	adc rdx,0
1860	add r12,r13
1861	mov r13,rdx
1862	adc r13,0
1863
1864	mul rbx
1865	add r14,rax
1866	mov rax,QWORD[56+rbp]
1867	adc rdx,0
1868	add r13,r14
1869	mov r14,rdx
1870	adc r14,0
1871
1872	mul rbx
1873	mov rbx,rsi
1874	add r15,rax
1875	mov rax,QWORD[rbp]
1876	adc rdx,0
1877	add r14,r15
1878	mov r15,rdx
1879	adc r15,0
1880
1881	dec ecx
1882	jnz NEAR $L$8x_reduce
1883
1884	lea rbp,[64+rbp]
1885	xor rax,rax
1886	mov rdx,QWORD[((8+8))+rsp]
1887	cmp rbp,QWORD[((0+8))+rsp]
1888	jae NEAR $L$8x_no_tail
1889
1890	DB 0x66
1891	add r8,QWORD[rdi]
1892	adc r9,QWORD[8+rdi]
1893	adc r10,QWORD[16+rdi]
1894	adc r11,QWORD[24+rdi]
1895	adc r12,QWORD[32+rdi]
1896	adc r13,QWORD[40+rdi]
1897	adc r14,QWORD[48+rdi]
1898	adc r15,QWORD[56+rdi]
1899	sbb rsi,rsi
1900
1901	mov rbx,QWORD[((48+56+8))+rsp]
1902	mov ecx,8
1903	mov rax,QWORD[rbp]
1904	jmp NEAR $L$8x_tail
1905
1906	ALIGN 32
1907	$L$8x_tail:
1908	mul rbx
1909	add r8,rax
1910	mov rax,QWORD[8+rbp]
1911	mov QWORD[rdi],r8
1912	mov r8,rdx
1913	adc r8,0
1914
1915	mul rbx
1916	add r9,rax
1917	mov rax,QWORD[16+rbp]
1918	adc rdx,0
1919	add r8,r9
1920	lea rdi,[8+rdi]
1921	mov r9,rdx
1922	adc r9,0
1923
1924	mul rbx
1925	add r10,rax
1926	mov rax,QWORD[24+rbp]
1927	adc rdx,0
1928	add r9,r10
1929	mov r10,rdx
1930	adc r10,0
1931
1932	mul rbx
1933	add r11,rax
1934	mov rax,QWORD[32+rbp]
1935	adc rdx,0
1936	add r10,r11
1937	mov r11,rdx
1938	adc r11,0
1939
1940	mul rbx
1941	add r12,rax
1942	mov rax,QWORD[40+rbp]
1943	adc rdx,0
1944	add r11,r12
1945	mov r12,rdx
1946	adc r12,0
1947
1948	mul rbx
1949	add r13,rax
1950	mov rax,QWORD[48+rbp]
1951	adc rdx,0
1952	add r12,r13
1953	mov r13,rdx
1954	adc r13,0
1955
1956	mul rbx
1957	add r14,rax
1958	mov rax,QWORD[56+rbp]
1959	adc rdx,0
1960	add r13,r14
1961	mov r14,rdx
1962	adc r14,0
1963
1964	mul rbx
1965	mov rbx,QWORD[((48-16+8))+rcx*8+rsp]
1966	add r15,rax
1967	adc rdx,0
1968	add r14,r15
1969	mov rax,QWORD[rbp]
1970	mov r15,rdx
1971	adc r15,0
1972
1973	dec ecx
1974	jnz NEAR $L$8x_tail
1975
1976	lea rbp,[64+rbp]
1977	mov rdx,QWORD[((8+8))+rsp]
1978	cmp rbp,QWORD[((0+8))+rsp]
1979	jae NEAR $L$8x_tail_done
1980
1981	mov rbx,QWORD[((48+56+8))+rsp]
1982	neg rsi
1983	mov rax,QWORD[rbp]
1984	adc r8,QWORD[rdi]
1985	adc r9,QWORD[8+rdi]
1986	adc r10,QWORD[16+rdi]
1987	adc r11,QWORD[24+rdi]
1988	adc r12,QWORD[32+rdi]
1989	adc r13,QWORD[40+rdi]
1990	adc r14,QWORD[48+rdi]
1991	adc r15,QWORD[56+rdi]
1992	sbb rsi,rsi
1993
1994	mov ecx,8
1995	jmp NEAR $L$8x_tail
1996
1997	ALIGN 32
1998	$L$8x_tail_done:
1999	xor rax,rax
2000	add r8,QWORD[rdx]
2001	adc r9,0
2002	adc r10,0
2003	adc r11,0
2004	adc r12,0
2005	adc r13,0
2006	adc r14,0
2007	adc r15,0
2008	adc rax,0
2009
2010	neg rsi
2011	$L$8x_no_tail:
2012	adc r8,QWORD[rdi]
2013	adc r9,QWORD[8+rdi]
2014	adc r10,QWORD[16+rdi]
2015	adc r11,QWORD[24+rdi]
2016	adc r12,QWORD[32+rdi]
2017	adc r13,QWORD[40+rdi]
2018	adc r14,QWORD[48+rdi]
2019	adc r15,QWORD[56+rdi]
2020	adc rax,0
2021	mov rcx,QWORD[((-8))+rbp]
2022	xor rsi,rsi
2023
2024	DB 102,72,15,126,213
2025
2026	mov QWORD[rdi],r8
2027	mov QWORD[8+rdi],r9
2028	DB 102,73,15,126,217
2029	mov QWORD[16+rdi],r10
2030	mov QWORD[24+rdi],r11
2031	mov QWORD[32+rdi],r12
2032	mov QWORD[40+rdi],r13
2033	mov QWORD[48+rdi],r14
2034	mov QWORD[56+rdi],r15
2035	lea rdi,[64+rdi]
2036
2037	cmp rdi,rdx
2038	jb NEAR $L$8x_reduction_loop
2039	DB 0F3h,0C3h ;repret
2040
2041
2042
2043	ALIGN 32
2044	__bn_post4x_internal:
2045
2046	mov r12,QWORD[rbp]
2047	lea rbx,[r9*1+rdi]
2048	mov rcx,r9
2049	DB 102,72,15,126,207
2050	neg rax
2051	DB 102,72,15,126,206
2052	sar rcx,3+2
2053	dec r12
2054	xor r10,r10
2055	mov r13,QWORD[8+rbp]
2056	mov r14,QWORD[16+rbp]
2057	mov r15,QWORD[24+rbp]
2058	jmp NEAR $L$sqr4x_sub_entry
2059
2060	ALIGN 16
2061	$L$sqr4x_sub:
2062	mov r12,QWORD[rbp]
2063	mov r13,QWORD[8+rbp]
2064	mov r14,QWORD[16+rbp]
2065	mov r15,QWORD[24+rbp]
2066	$L$sqr4x_sub_entry:
2067	lea rbp,[32+rbp]
2068	not r12
2069	not r13
2070	not r14
2071	not r15
2072	and r12,rax
2073	and r13,rax
2074	and r14,rax
2075	and r15,rax
2076
2077	neg r10
2078	adc r12,QWORD[rbx]
2079	adc r13,QWORD[8+rbx]
2080	adc r14,QWORD[16+rbx]
2081	adc r15,QWORD[24+rbx]
2082	mov QWORD[rdi],r12
2083	lea rbx,[32+rbx]
2084	mov QWORD[8+rdi],r13
2085	sbb r10,r10
2086	mov QWORD[16+rdi],r14
2087	mov QWORD[24+rdi],r15
2088	lea rdi,[32+rdi]
2089
2090	inc rcx
2091	jnz NEAR $L$sqr4x_sub
2092
2093	mov r10,r9
2094	neg r9
2095	DB 0F3h,0C3h ;repret
2096
2097
2098
2099	ALIGN 32
2100	bn_mulx4x_mont_gather5:
2101	mov QWORD[8+rsp],rdi ;WIN64 prologue
2102	mov QWORD[16+rsp],rsi
2103	mov rax,rsp
2104	$L$SEH_begin_bn_mulx4x_mont_gather5:
2105	mov rdi,rcx
2106	mov rsi,rdx
2107	mov rdx,r8
2108	mov rcx,r9
2109	mov r8,QWORD[40+rsp]
2110	mov r9,QWORD[48+rsp]
2111
2112
2113
2114	mov rax,rsp
2115
2116	$L$mulx4x_enter:
2117	push rbx
2118
2119	push rbp
2120
2121	push r12
2122
2123	push r13
2124
2125	push r14
2126
2127	push r15
2128
2129	$L$mulx4x_prologue:
2130
2131	shl r9d,3
2132	lea r10,[r9*2+r9]
2133	neg r9
2134	mov r8,QWORD[r8]
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145	lea r11,[((-320))+r9*2+rsp]
2146	mov rbp,rsp
2147	sub r11,rdi
2148	and r11,4095
2149	cmp r10,r11
2150	jb NEAR $L$mulx4xsp_alt
2151	sub rbp,r11
2152	lea rbp,[((-320))+r9*2+rbp]
2153	jmp NEAR $L$mulx4xsp_done
2154
2155	$L$mulx4xsp_alt:
2156	lea r10,[((4096-320))+r9*2]
2157	lea rbp,[((-320))+r9*2+rbp]
2158	sub r11,r10
2159	mov r10,0
2160	cmovc r11,r10
2161	sub rbp,r11
2162	$L$mulx4xsp_done:
2163	and rbp,-64
2164	mov r11,rsp
2165	sub r11,rbp
2166	and r11,-4096
2167	lea rsp,[rbp*1+r11]
2168	mov r10,QWORD[rsp]
2169	cmp rsp,rbp
2170	ja NEAR $L$mulx4x_page_walk
2171	jmp NEAR $L$mulx4x_page_walk_done
2172
2173	$L$mulx4x_page_walk:
2174	lea rsp,[((-4096))+rsp]
2175	mov r10,QWORD[rsp]
2176	cmp rsp,rbp
2177	ja NEAR $L$mulx4x_page_walk
2178	$L$mulx4x_page_walk_done:
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192	mov QWORD[32+rsp],r8
2193	mov QWORD[40+rsp],rax
2194
2195	$L$mulx4x_body:
2196	call mulx4x_internal
2197
2198	mov rsi,QWORD[40+rsp]
2199
2200	mov rax,1
2201
2202	mov r15,QWORD[((-48))+rsi]
2203
2204	mov r14,QWORD[((-40))+rsi]
2205
2206	mov r13,QWORD[((-32))+rsi]
2207
2208	mov r12,QWORD[((-24))+rsi]
2209
2210	mov rbp,QWORD[((-16))+rsi]
2211
2212	mov rbx,QWORD[((-8))+rsi]
2213
2214	lea rsp,[rsi]
2215
2216	$L$mulx4x_epilogue:
2217	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2218	mov rsi,QWORD[16+rsp]
2219	DB 0F3h,0C3h ;repret
2220
2221	$L$SEH_end_bn_mulx4x_mont_gather5:
2222
2223
2224	ALIGN 32
2225	mulx4x_internal:
2226
2227	mov QWORD[8+rsp],r9
2228	mov r10,r9
2229	neg r9
2230	shl r9,5
2231	neg r10
2232	lea r13,[128+r9*1+rdx]
2233	shr r9,5+5
2234	movd xmm5,DWORD[56+rax]
2235	sub r9,1
2236	lea rax,[$L$inc]
2237	mov QWORD[((16+8))+rsp],r13
2238	mov QWORD[((24+8))+rsp],r9
2239	mov QWORD[((56+8))+rsp],rdi
2240	movdqa xmm0,XMMWORD[rax]
2241	movdqa xmm1,XMMWORD[16+rax]
2242	lea r10,[((88-112))+r10*1+rsp]
2243	lea rdi,[128+rdx]
2244
2245	pshufd xmm5,xmm5,0
2246	movdqa xmm4,xmm1
2247	DB 0x67
2248	movdqa xmm2,xmm1
2249	DB 0x67
2250	paddd xmm1,xmm0
2251	pcmpeqd xmm0,xmm5
2252	movdqa xmm3,xmm4
2253	paddd xmm2,xmm1
2254	pcmpeqd xmm1,xmm5
2255	movdqa XMMWORD[112+r10],xmm0
2256	movdqa xmm0,xmm4
2257
2258	paddd xmm3,xmm2
2259	pcmpeqd xmm2,xmm5
2260	movdqa XMMWORD[128+r10],xmm1
2261	movdqa xmm1,xmm4
2262
2263	paddd xmm0,xmm3
2264	pcmpeqd xmm3,xmm5
2265	movdqa XMMWORD[144+r10],xmm2
2266	movdqa xmm2,xmm4
2267
2268	paddd xmm1,xmm0
2269	pcmpeqd xmm0,xmm5
2270	movdqa XMMWORD[160+r10],xmm3
2271	movdqa xmm3,xmm4
2272	paddd xmm2,xmm1
2273	pcmpeqd xmm1,xmm5
2274	movdqa XMMWORD[176+r10],xmm0
2275	movdqa xmm0,xmm4
2276
2277	paddd xmm3,xmm2
2278	pcmpeqd xmm2,xmm5
2279	movdqa XMMWORD[192+r10],xmm1
2280	movdqa xmm1,xmm4
2281
2282	paddd xmm0,xmm3
2283	pcmpeqd xmm3,xmm5
2284	movdqa XMMWORD[208+r10],xmm2
2285	movdqa xmm2,xmm4
2286
2287	paddd xmm1,xmm0
2288	pcmpeqd xmm0,xmm5
2289	movdqa XMMWORD[224+r10],xmm3
2290	movdqa xmm3,xmm4
2291	paddd xmm2,xmm1
2292	pcmpeqd xmm1,xmm5
2293	movdqa XMMWORD[240+r10],xmm0
2294	movdqa xmm0,xmm4
2295
2296	paddd xmm3,xmm2
2297	pcmpeqd xmm2,xmm5
2298	movdqa XMMWORD[256+r10],xmm1
2299	movdqa xmm1,xmm4
2300
2301	paddd xmm0,xmm3
2302	pcmpeqd xmm3,xmm5
2303	movdqa XMMWORD[272+r10],xmm2
2304	movdqa xmm2,xmm4
2305
2306	paddd xmm1,xmm0
2307	pcmpeqd xmm0,xmm5
2308	movdqa XMMWORD[288+r10],xmm3
2309	movdqa xmm3,xmm4
2310	DB 0x67
2311	paddd xmm2,xmm1
2312	pcmpeqd xmm1,xmm5
2313	movdqa XMMWORD[304+r10],xmm0
2314
2315	paddd xmm3,xmm2
2316	pcmpeqd xmm2,xmm5
2317	movdqa XMMWORD[320+r10],xmm1
2318
2319	pcmpeqd xmm3,xmm5
2320	movdqa XMMWORD[336+r10],xmm2
2321
2322	pand xmm0,XMMWORD[64+rdi]
2323	pand xmm1,XMMWORD[80+rdi]
2324	pand xmm2,XMMWORD[96+rdi]
2325	movdqa XMMWORD[352+r10],xmm3
2326	pand xmm3,XMMWORD[112+rdi]
2327	por xmm0,xmm2
2328	por xmm1,xmm3
2329	movdqa xmm4,XMMWORD[((-128))+rdi]
2330	movdqa xmm5,XMMWORD[((-112))+rdi]
2331	movdqa xmm2,XMMWORD[((-96))+rdi]
2332	pand xmm4,XMMWORD[112+r10]
2333	movdqa xmm3,XMMWORD[((-80))+rdi]
2334	pand xmm5,XMMWORD[128+r10]
2335	por xmm0,xmm4
2336	pand xmm2,XMMWORD[144+r10]
2337	por xmm1,xmm5
2338	pand xmm3,XMMWORD[160+r10]
2339	por xmm0,xmm2
2340	por xmm1,xmm3
2341	movdqa xmm4,XMMWORD[((-64))+rdi]
2342	movdqa xmm5,XMMWORD[((-48))+rdi]
2343	movdqa xmm2,XMMWORD[((-32))+rdi]
2344	pand xmm4,XMMWORD[176+r10]
2345	movdqa xmm3,XMMWORD[((-16))+rdi]
2346	pand xmm5,XMMWORD[192+r10]
2347	por xmm0,xmm4
2348	pand xmm2,XMMWORD[208+r10]
2349	por xmm1,xmm5
2350	pand xmm3,XMMWORD[224+r10]
2351	por xmm0,xmm2
2352	por xmm1,xmm3
2353	movdqa xmm4,XMMWORD[rdi]
2354	movdqa xmm5,XMMWORD[16+rdi]
2355	movdqa xmm2,XMMWORD[32+rdi]
2356	pand xmm4,XMMWORD[240+r10]
2357	movdqa xmm3,XMMWORD[48+rdi]
2358	pand xmm5,XMMWORD[256+r10]
2359	por xmm0,xmm4
2360	pand xmm2,XMMWORD[272+r10]
2361	por xmm1,xmm5
2362	pand xmm3,XMMWORD[288+r10]
2363	por xmm0,xmm2
2364	por xmm1,xmm3
2365	pxor xmm0,xmm1
2366	pshufd xmm1,xmm0,0x4e
2367	por xmm0,xmm1
2368	lea rdi,[256+rdi]
2369	DB 102,72,15,126,194
2370	lea rbx,[((64+32+8))+rsp]
2371
2372	mov r9,rdx
2373	mulx rax,r8,QWORD[rsi]
2374	mulx r12,r11,QWORD[8+rsi]
2375	add r11,rax
2376	mulx r13,rax,QWORD[16+rsi]
2377	adc r12,rax
2378	adc r13,0
2379	mulx r14,rax,QWORD[24+rsi]
2380
2381	mov r15,r8
2382	imul r8,QWORD[((32+8))+rsp]
2383	xor rbp,rbp
2384	mov rdx,r8
2385
2386	mov QWORD[((8+8))+rsp],rdi
2387
2388	lea rsi,[32+rsi]
2389	adcx r13,rax
2390	adcx r14,rbp
2391
2392	mulx r10,rax,QWORD[rcx]
2393	adcx r15,rax
2394	adox r10,r11
2395	mulx r11,rax,QWORD[8+rcx]
2396	adcx r10,rax
2397	adox r11,r12
2398	mulx r12,rax,QWORD[16+rcx]
2399	mov rdi,QWORD[((24+8))+rsp]
2400	mov QWORD[((-32))+rbx],r10
2401	adcx r11,rax
2402	adox r12,r13
2403	mulx r15,rax,QWORD[24+rcx]
2404	mov rdx,r9
2405	mov QWORD[((-24))+rbx],r11
2406	adcx r12,rax
2407	adox r15,rbp
2408	lea rcx,[32+rcx]
2409	mov QWORD[((-16))+rbx],r12
2410	jmp NEAR $L$mulx4x_1st
2411
2412	ALIGN 32
2413	$L$mulx4x_1st:
2414	adcx r15,rbp
2415	mulx rax,r10,QWORD[rsi]
2416	adcx r10,r14
2417	mulx r14,r11,QWORD[8+rsi]
2418	adcx r11,rax
2419	mulx rax,r12,QWORD[16+rsi]
2420	adcx r12,r14
2421	mulx r14,r13,QWORD[24+rsi]
2422	DB 0x67,0x67
2423	mov rdx,r8
2424	adcx r13,rax
2425	adcx r14,rbp
2426	lea rsi,[32+rsi]
2427	lea rbx,[32+rbx]
2428
2429	adox r10,r15
2430	mulx r15,rax,QWORD[rcx]
2431	adcx r10,rax
2432	adox r11,r15
2433	mulx r15,rax,QWORD[8+rcx]
2434	adcx r11,rax
2435	adox r12,r15
2436	mulx r15,rax,QWORD[16+rcx]
2437	mov QWORD[((-40))+rbx],r10
2438	adcx r12,rax
2439	mov QWORD[((-32))+rbx],r11
2440	adox r13,r15
2441	mulx r15,rax,QWORD[24+rcx]
2442	mov rdx,r9
2443	mov QWORD[((-24))+rbx],r12
2444	adcx r13,rax
2445	adox r15,rbp
2446	lea rcx,[32+rcx]
2447	mov QWORD[((-16))+rbx],r13
2448
2449	dec rdi
2450	jnz NEAR $L$mulx4x_1st
2451
2452	mov rax,QWORD[8+rsp]
2453	adc r15,rbp
2454	lea rsi,[rax*1+rsi]
2455	add r14,r15
2456	mov rdi,QWORD[((8+8))+rsp]
2457	adc rbp,rbp
2458	mov QWORD[((-8))+rbx],r14
2459	jmp NEAR $L$mulx4x_outer
2460
2461	ALIGN 32
2462	$L$mulx4x_outer:
2463	lea r10,[((16-256))+rbx]
2464	pxor xmm4,xmm4
2465	DB 0x67,0x67
2466	pxor xmm5,xmm5
2467	movdqa xmm0,XMMWORD[((-128))+rdi]
2468	movdqa xmm1,XMMWORD[((-112))+rdi]
2469	movdqa xmm2,XMMWORD[((-96))+rdi]
2470	pand xmm0,XMMWORD[256+r10]
2471	movdqa xmm3,XMMWORD[((-80))+rdi]
2472	pand xmm1,XMMWORD[272+r10]
2473	por xmm4,xmm0
2474	pand xmm2,XMMWORD[288+r10]
2475	por xmm5,xmm1
2476	pand xmm3,XMMWORD[304+r10]
2477	por xmm4,xmm2
2478	por xmm5,xmm3
2479	movdqa xmm0,XMMWORD[((-64))+rdi]
2480	movdqa xmm1,XMMWORD[((-48))+rdi]
2481	movdqa xmm2,XMMWORD[((-32))+rdi]
2482	pand xmm0,XMMWORD[320+r10]
2483	movdqa xmm3,XMMWORD[((-16))+rdi]
2484	pand xmm1,XMMWORD[336+r10]
2485	por xmm4,xmm0
2486	pand xmm2,XMMWORD[352+r10]
2487	por xmm5,xmm1
2488	pand xmm3,XMMWORD[368+r10]
2489	por xmm4,xmm2
2490	por xmm5,xmm3
2491	movdqa xmm0,XMMWORD[rdi]
2492	movdqa xmm1,XMMWORD[16+rdi]
2493	movdqa xmm2,XMMWORD[32+rdi]
2494	pand xmm0,XMMWORD[384+r10]
2495	movdqa xmm3,XMMWORD[48+rdi]
2496	pand xmm1,XMMWORD[400+r10]
2497	por xmm4,xmm0
2498	pand xmm2,XMMWORD[416+r10]
2499	por xmm5,xmm1
2500	pand xmm3,XMMWORD[432+r10]
2501	por xmm4,xmm2
2502	por xmm5,xmm3
2503	movdqa xmm0,XMMWORD[64+rdi]
2504	movdqa xmm1,XMMWORD[80+rdi]
2505	movdqa xmm2,XMMWORD[96+rdi]
2506	pand xmm0,XMMWORD[448+r10]
2507	movdqa xmm3,XMMWORD[112+rdi]
2508	pand xmm1,XMMWORD[464+r10]
2509	por xmm4,xmm0
2510	pand xmm2,XMMWORD[480+r10]
2511	por xmm5,xmm1
2512	pand xmm3,XMMWORD[496+r10]
2513	por xmm4,xmm2
2514	por xmm5,xmm3
2515	por xmm4,xmm5
2516	pshufd xmm0,xmm4,0x4e
2517	por xmm0,xmm4
2518	lea rdi,[256+rdi]
2519	DB 102,72,15,126,194
2520
2521	mov QWORD[rbx],rbp
2522	lea rbx,[32+rax*1+rbx]
2523	mulx r11,r8,QWORD[rsi]
2524	xor rbp,rbp
2525	mov r9,rdx
2526	mulx r12,r14,QWORD[8+rsi]
2527	adox r8,QWORD[((-32))+rbx]
2528	adcx r11,r14
2529	mulx r13,r15,QWORD[16+rsi]
2530	adox r11,QWORD[((-24))+rbx]
2531	adcx r12,r15
2532	mulx r14,rdx,QWORD[24+rsi]
2533	adox r12,QWORD[((-16))+rbx]
2534	adcx r13,rdx
2535	lea rcx,[rax*1+rcx]
2536	lea rsi,[32+rsi]
2537	adox r13,QWORD[((-8))+rbx]
2538	adcx r14,rbp
2539	adox r14,rbp
2540
2541	mov r15,r8
2542	imul r8,QWORD[((32+8))+rsp]
2543
2544	mov rdx,r8
2545	xor rbp,rbp
2546	mov QWORD[((8+8))+rsp],rdi
2547
2548	mulx r10,rax,QWORD[rcx]
2549	adcx r15,rax
2550	adox r10,r11
2551	mulx r11,rax,QWORD[8+rcx]
2552	adcx r10,rax
2553	adox r11,r12
2554	mulx r12,rax,QWORD[16+rcx]
2555	adcx r11,rax
2556	adox r12,r13
2557	mulx r15,rax,QWORD[24+rcx]
2558	mov rdx,r9
2559	mov rdi,QWORD[((24+8))+rsp]
2560	mov QWORD[((-32))+rbx],r10
2561	adcx r12,rax
2562	mov QWORD[((-24))+rbx],r11
2563	adox r15,rbp
2564	mov QWORD[((-16))+rbx],r12
2565	lea rcx,[32+rcx]
2566	jmp NEAR $L$mulx4x_inner
2567
2568	ALIGN 32
2569	$L$mulx4x_inner:
2570	mulx rax,r10,QWORD[rsi]
2571	adcx r15,rbp
2572	adox r10,r14
2573	mulx r14,r11,QWORD[8+rsi]
2574	adcx r10,QWORD[rbx]
2575	adox r11,rax
2576	mulx rax,r12,QWORD[16+rsi]
2577	adcx r11,QWORD[8+rbx]
2578	adox r12,r14
2579	mulx r14,r13,QWORD[24+rsi]
2580	mov rdx,r8
2581	adcx r12,QWORD[16+rbx]
2582	adox r13,rax
2583	adcx r13,QWORD[24+rbx]
2584	adox r14,rbp
2585	lea rsi,[32+rsi]
2586	lea rbx,[32+rbx]
2587	adcx r14,rbp
2588
2589	adox r10,r15
2590	mulx r15,rax,QWORD[rcx]
2591	adcx r10,rax
2592	adox r11,r15
2593	mulx r15,rax,QWORD[8+rcx]
2594	adcx r11,rax
2595	adox r12,r15
2596	mulx r15,rax,QWORD[16+rcx]
2597	mov QWORD[((-40))+rbx],r10
2598	adcx r12,rax
2599	adox r13,r15
2600	mov QWORD[((-32))+rbx],r11
2601	mulx r15,rax,QWORD[24+rcx]
2602	mov rdx,r9
2603	lea rcx,[32+rcx]
2604	mov QWORD[((-24))+rbx],r12
2605	adcx r13,rax
2606	adox r15,rbp
2607	mov QWORD[((-16))+rbx],r13
2608
2609	dec rdi
2610	jnz NEAR $L$mulx4x_inner
2611
2612	mov rax,QWORD[((0+8))+rsp]
2613	adc r15,rbp
2614	sub rdi,QWORD[rbx]
2615	mov rdi,QWORD[((8+8))+rsp]
2616	mov r10,QWORD[((16+8))+rsp]
2617	adc r14,r15
2618	lea rsi,[rax*1+rsi]
2619	adc rbp,rbp
2620	mov QWORD[((-8))+rbx],r14
2621
2622	cmp rdi,r10
2623	jb NEAR $L$mulx4x_outer
2624
2625	mov r10,QWORD[((-8))+rcx]
2626	mov r8,rbp
2627	mov r12,QWORD[rax*1+rcx]
2628	lea rbp,[rax*1+rcx]
2629	mov rcx,rax
2630	lea rdi,[rax*1+rbx]
2631	xor eax,eax
2632	xor r15,r15
2633	sub r10,r14
2634	adc r15,r15
2635	or r8,r15
2636	sar rcx,3+2
2637	sub rax,r8
2638	mov rdx,QWORD[((56+8))+rsp]
2639	dec r12
2640	mov r13,QWORD[8+rbp]
2641	xor r8,r8
2642	mov r14,QWORD[16+rbp]
2643	mov r15,QWORD[24+rbp]
2644	jmp NEAR $L$sqrx4x_sub_entry
2645
2646
2647
2648	ALIGN 32
2649	bn_powerx5:
2650	mov QWORD[8+rsp],rdi ;WIN64 prologue
2651	mov QWORD[16+rsp],rsi
2652	mov rax,rsp
2653	$L$SEH_begin_bn_powerx5:
2654	mov rdi,rcx
2655	mov rsi,rdx
2656	mov rdx,r8
2657	mov rcx,r9
2658	mov r8,QWORD[40+rsp]
2659	mov r9,QWORD[48+rsp]
2660
2661
2662
2663	mov rax,rsp
2664
2665	$L$powerx5_enter:
2666	push rbx
2667
2668	push rbp
2669
2670	push r12
2671
2672	push r13
2673
2674	push r14
2675
2676	push r15
2677
2678	$L$powerx5_prologue:
2679
2680	shl r9d,3
2681	lea r10,[r9*2+r9]
2682	neg r9
2683	mov r8,QWORD[r8]
2684
2685
2686
2687
2688
2689
2690
2691
2692	lea r11,[((-320))+r9*2+rsp]
2693	mov rbp,rsp
2694	sub r11,rdi
2695	and r11,4095
2696	cmp r10,r11
2697	jb NEAR $L$pwrx_sp_alt
2698	sub rbp,r11
2699	lea rbp,[((-320))+r9*2+rbp]
2700	jmp NEAR $L$pwrx_sp_done
2701
2702	ALIGN 32
2703	$L$pwrx_sp_alt:
2704	lea r10,[((4096-320))+r9*2]
2705	lea rbp,[((-320))+r9*2+rbp]
2706	sub r11,r10
2707	mov r10,0
2708	cmovc r11,r10
2709	sub rbp,r11
2710	$L$pwrx_sp_done:
2711	and rbp,-64
2712	mov r11,rsp
2713	sub r11,rbp
2714	and r11,-4096
2715	lea rsp,[rbp*1+r11]
2716	mov r10,QWORD[rsp]
2717	cmp rsp,rbp
2718	ja NEAR $L$pwrx_page_walk
2719	jmp NEAR $L$pwrx_page_walk_done
2720
2721	$L$pwrx_page_walk:
2722	lea rsp,[((-4096))+rsp]
2723	mov r10,QWORD[rsp]
2724	cmp rsp,rbp
2725	ja NEAR $L$pwrx_page_walk
2726	$L$pwrx_page_walk_done:
2727
2728	mov r10,r9
2729	neg r9
2730
2731
2732
2733
2734
2735
2736
2737
2738
2739
2740
2741
2742	pxor xmm0,xmm0
2743	DB 102,72,15,110,207
2744	DB 102,72,15,110,209
2745	DB 102,73,15,110,218
2746	DB 102,72,15,110,226
2747	mov QWORD[32+rsp],r8
2748	mov QWORD[40+rsp],rax
2749
2750	$L$powerx5_body:
2751
2752	call __bn_sqrx8x_internal
2753	call __bn_postx4x_internal
2754	call __bn_sqrx8x_internal
2755	call __bn_postx4x_internal
2756	call __bn_sqrx8x_internal
2757	call __bn_postx4x_internal
2758	call __bn_sqrx8x_internal
2759	call __bn_postx4x_internal
2760	call __bn_sqrx8x_internal
2761	call __bn_postx4x_internal
2762
2763	mov r9,r10
2764	mov rdi,rsi
2765	DB 102,72,15,126,209
2766	DB 102,72,15,126,226
2767	mov rax,QWORD[40+rsp]
2768
2769	call mulx4x_internal
2770
2771	mov rsi,QWORD[40+rsp]
2772
2773	mov rax,1
2774
2775	mov r15,QWORD[((-48))+rsi]
2776
2777	mov r14,QWORD[((-40))+rsi]
2778
2779	mov r13,QWORD[((-32))+rsi]
2780
2781	mov r12,QWORD[((-24))+rsi]
2782
2783	mov rbp,QWORD[((-16))+rsi]
2784
2785	mov rbx,QWORD[((-8))+rsi]
2786
2787	lea rsp,[rsi]
2788
2789	$L$powerx5_epilogue:
2790	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2791	mov rsi,QWORD[16+rsp]
2792	DB 0F3h,0C3h ;repret
2793
2794	$L$SEH_end_bn_powerx5:
2795
2796	global bn_sqrx8x_internal
2797
2798
2799	ALIGN 32
2800	bn_sqrx8x_internal:
2801	__bn_sqrx8x_internal:
2802
2803
2804
2805
2806
2807
2808
2809
2810
2811
2812
2813
2814
2815
2816
2817
2818
2819
2820
2821
2822
2823
2824
2825
2826
2827
2828
2829
2830
2831
2832
2833
2834
2835
2836
2837
2838
2839
2840
2841
2842
2843	lea rdi,[((48+8))+rsp]
2844	lea rbp,[r9*1+rsi]
2845	mov QWORD[((0+8))+rsp],r9
2846	mov QWORD[((8+8))+rsp],rbp
2847	jmp NEAR $L$sqr8x_zero_start
2848
2849	ALIGN 32
2850	DB 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
2851	$L$sqrx8x_zero:
2852	DB 0x3e
2853	movdqa XMMWORD[rdi],xmm0
2854	movdqa XMMWORD[16+rdi],xmm0
2855	movdqa XMMWORD[32+rdi],xmm0
2856	movdqa XMMWORD[48+rdi],xmm0
2857	$L$sqr8x_zero_start:
2858	movdqa XMMWORD[64+rdi],xmm0
2859	movdqa XMMWORD[80+rdi],xmm0
2860	movdqa XMMWORD[96+rdi],xmm0
2861	movdqa XMMWORD[112+rdi],xmm0
2862	lea rdi,[128+rdi]
2863	sub r9,64
2864	jnz NEAR $L$sqrx8x_zero
2865
2866	mov rdx,QWORD[rsi]
2867
2868	xor r10,r10
2869	xor r11,r11
2870	xor r12,r12
2871	xor r13,r13
2872	xor r14,r14
2873	xor r15,r15
2874	lea rdi,[((48+8))+rsp]
2875	xor rbp,rbp
2876	jmp NEAR $L$sqrx8x_outer_loop
2877
2878	ALIGN 32
2879	$L$sqrx8x_outer_loop:
2880	mulx rax,r8,QWORD[8+rsi]
2881	adcx r8,r9
2882	adox r10,rax
2883	mulx rax,r9,QWORD[16+rsi]
2884	adcx r9,r10
2885	adox r11,rax
2886	DB 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00
2887	adcx r10,r11
2888	adox r12,rax
2889	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00
2890	adcx r11,r12
2891	adox r13,rax
2892	mulx rax,r12,QWORD[40+rsi]
2893	adcx r12,r13
2894	adox r14,rax
2895	mulx rax,r13,QWORD[48+rsi]
2896	adcx r13,r14
2897	adox rax,r15
2898	mulx r15,r14,QWORD[56+rsi]
2899	mov rdx,QWORD[8+rsi]
2900	adcx r14,rax
2901	adox r15,rbp
2902	adc r15,QWORD[64+rdi]
2903	mov QWORD[8+rdi],r8
2904	mov QWORD[16+rdi],r9
2905	sbb rcx,rcx
2906	xor rbp,rbp
2907
2908
2909	mulx rbx,r8,QWORD[16+rsi]
2910	mulx rax,r9,QWORD[24+rsi]
2911	adcx r8,r10
2912	adox r9,rbx
2913	mulx rbx,r10,QWORD[32+rsi]
2914	adcx r9,r11
2915	adox r10,rax
2916	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00
2917	adcx r10,r12
2918	adox r11,rbx
2919	DB 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00
2920	adcx r11,r13
2921	adox r12,r14
2922	DB 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00
2923	mov rdx,QWORD[16+rsi]
2924	adcx r12,rax
2925	adox r13,rbx
2926	adcx r13,r15
2927	adox r14,rbp
2928	adcx r14,rbp
2929
2930	mov QWORD[24+rdi],r8
2931	mov QWORD[32+rdi],r9
2932
2933	mulx rbx,r8,QWORD[24+rsi]
2934	mulx rax,r9,QWORD[32+rsi]
2935	adcx r8,r10
2936	adox r9,rbx
2937	mulx rbx,r10,QWORD[40+rsi]
2938	adcx r9,r11
2939	adox r10,rax
2940	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00
2941	adcx r10,r12
2942	adox r11,r13
2943	DB 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00
2944	DB 0x3e
2945	mov rdx,QWORD[24+rsi]
2946	adcx r11,rbx
2947	adox r12,rax
2948	adcx r12,r14
2949	mov QWORD[40+rdi],r8
2950	mov QWORD[48+rdi],r9
2951	mulx rax,r8,QWORD[32+rsi]
2952	adox r13,rbp
2953	adcx r13,rbp
2954
2955	mulx rbx,r9,QWORD[40+rsi]
2956	adcx r8,r10
2957	adox r9,rax
2958	mulx rax,r10,QWORD[48+rsi]
2959	adcx r9,r11
2960	adox r10,r12
2961	mulx r12,r11,QWORD[56+rsi]
2962	mov rdx,QWORD[32+rsi]
2963	mov r14,QWORD[40+rsi]
2964	adcx r10,rbx
2965	adox r11,rax
2966	mov r15,QWORD[48+rsi]
2967	adcx r11,r13
2968	adox r12,rbp
2969	adcx r12,rbp
2970
2971	mov QWORD[56+rdi],r8
2972	mov QWORD[64+rdi],r9
2973
2974	mulx rax,r9,r14
2975	mov r8,QWORD[56+rsi]
2976	adcx r9,r10
2977	mulx rbx,r10,r15
2978	adox r10,rax
2979	adcx r10,r11
2980	mulx rax,r11,r8
2981	mov rdx,r14
2982	adox r11,rbx
2983	adcx r11,r12
2984
2985	adcx rax,rbp
2986
2987	mulx rbx,r14,r15
2988	mulx r13,r12,r8
2989	mov rdx,r15
2990	lea rsi,[64+rsi]
2991	adcx r11,r14
2992	adox r12,rbx
2993	adcx r12,rax
2994	adox r13,rbp
2995
2996	DB 0x67,0x67
2997	mulx r14,r8,r8
2998	adcx r13,r8
2999	adcx r14,rbp
3000
3001	cmp rsi,QWORD[((8+8))+rsp]
3002	je NEAR $L$sqrx8x_outer_break
3003
3004	neg rcx
3005	mov rcx,-8
3006	mov r15,rbp
3007	mov r8,QWORD[64+rdi]
3008	adcx r9,QWORD[72+rdi]
3009	adcx r10,QWORD[80+rdi]
3010	adcx r11,QWORD[88+rdi]
3011	adc r12,QWORD[96+rdi]
3012	adc r13,QWORD[104+rdi]
3013	adc r14,QWORD[112+rdi]
3014	adc r15,QWORD[120+rdi]
3015	lea rbp,[rsi]
3016	lea rdi,[128+rdi]
3017	sbb rax,rax
3018
3019	mov rdx,QWORD[((-64))+rsi]
3020	mov QWORD[((16+8))+rsp],rax
3021	mov QWORD[((24+8))+rsp],rdi
3022
3023
3024	xor eax,eax
3025	jmp NEAR $L$sqrx8x_loop
3026
3027	ALIGN 32
3028	$L$sqrx8x_loop:
3029	mov rbx,r8
3030	mulx r8,rax,QWORD[rbp]
3031	adcx rbx,rax
3032	adox r8,r9
3033
3034	mulx r9,rax,QWORD[8+rbp]
3035	adcx r8,rax
3036	adox r9,r10
3037
3038	mulx r10,rax,QWORD[16+rbp]
3039	adcx r9,rax
3040	adox r10,r11
3041
3042	mulx r11,rax,QWORD[24+rbp]
3043	adcx r10,rax
3044	adox r11,r12
3045
3046	DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
3047	adcx r11,rax
3048	adox r12,r13
3049
3050	mulx r13,rax,QWORD[40+rbp]
3051	adcx r12,rax
3052	adox r13,r14
3053
3054	mulx r14,rax,QWORD[48+rbp]
3055	mov QWORD[rcx*8+rdi],rbx
3056	mov ebx,0
3057	adcx r13,rax
3058	adox r14,r15
3059
3060	DB 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00
3061	mov rdx,QWORD[8+rcx*8+rsi]
3062	adcx r14,rax
3063	adox r15,rbx
3064	adcx r15,rbx
3065
3066	DB 0x67
3067	inc rcx
3068	jnz NEAR $L$sqrx8x_loop
3069
3070	lea rbp,[64+rbp]
3071	mov rcx,-8
3072	cmp rbp,QWORD[((8+8))+rsp]
3073	je NEAR $L$sqrx8x_break
3074
3075	sub rbx,QWORD[((16+8))+rsp]
3076	DB 0x66
3077	mov rdx,QWORD[((-64))+rsi]
3078	adcx r8,QWORD[rdi]
3079	adcx r9,QWORD[8+rdi]
3080	adc r10,QWORD[16+rdi]
3081	adc r11,QWORD[24+rdi]
3082	adc r12,QWORD[32+rdi]
3083	adc r13,QWORD[40+rdi]
3084	adc r14,QWORD[48+rdi]
3085	adc r15,QWORD[56+rdi]
3086	lea rdi,[64+rdi]
3087	DB 0x67
3088	sbb rax,rax
3089	xor ebx,ebx
3090	mov QWORD[((16+8))+rsp],rax
3091	jmp NEAR $L$sqrx8x_loop
3092
3093	ALIGN 32
3094	$L$sqrx8x_break:
3095	xor rbp,rbp
3096	sub rbx,QWORD[((16+8))+rsp]
3097	adcx r8,rbp
3098	mov rcx,QWORD[((24+8))+rsp]
3099	adcx r9,rbp
3100	mov rdx,QWORD[rsi]
3101	adc r10,0
3102	mov QWORD[rdi],r8
3103	adc r11,0
3104	adc r12,0
3105	adc r13,0
3106	adc r14,0
3107	adc r15,0
3108	cmp rdi,rcx
3109	je NEAR $L$sqrx8x_outer_loop
3110
3111	mov QWORD[8+rdi],r9
3112	mov r9,QWORD[8+rcx]
3113	mov QWORD[16+rdi],r10
3114	mov r10,QWORD[16+rcx]
3115	mov QWORD[24+rdi],r11
3116	mov r11,QWORD[24+rcx]
3117	mov QWORD[32+rdi],r12
3118	mov r12,QWORD[32+rcx]
3119	mov QWORD[40+rdi],r13
3120	mov r13,QWORD[40+rcx]
3121	mov QWORD[48+rdi],r14
3122	mov r14,QWORD[48+rcx]
3123	mov QWORD[56+rdi],r15
3124	mov r15,QWORD[56+rcx]
3125	mov rdi,rcx
3126	jmp NEAR $L$sqrx8x_outer_loop
3127
3128	ALIGN 32
3129	$L$sqrx8x_outer_break:
3130	mov QWORD[72+rdi],r9
3131	DB 102,72,15,126,217
3132	mov QWORD[80+rdi],r10
3133	mov QWORD[88+rdi],r11
3134	mov QWORD[96+rdi],r12
3135	mov QWORD[104+rdi],r13
3136	mov QWORD[112+rdi],r14
3137	lea rdi,[((48+8))+rsp]
3138	mov rdx,QWORD[rcx*1+rsi]
3139
3140	mov r11,QWORD[8+rdi]
3141	xor r10,r10
3142	mov r9,QWORD[((0+8))+rsp]
3143	adox r11,r11
3144	mov r12,QWORD[16+rdi]
3145	mov r13,QWORD[24+rdi]
3146
3147
3148	ALIGN 32
3149	$L$sqrx4x_shift_n_add:
3150	mulx rbx,rax,rdx
3151	adox r12,r12
3152	adcx rax,r10
3153	DB 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00
3154	DB 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00
3155	adox r13,r13
3156	adcx rbx,r11
3157	mov r11,QWORD[40+rdi]
3158	mov QWORD[rdi],rax
3159	mov QWORD[8+rdi],rbx
3160
3161	mulx rbx,rax,rdx
3162	adox r10,r10
3163	adcx rax,r12
3164	mov rdx,QWORD[16+rcx*1+rsi]
3165	mov r12,QWORD[48+rdi]
3166	adox r11,r11
3167	adcx rbx,r13
3168	mov r13,QWORD[56+rdi]
3169	mov QWORD[16+rdi],rax
3170	mov QWORD[24+rdi],rbx
3171
3172	mulx rbx,rax,rdx
3173	adox r12,r12
3174	adcx rax,r10
3175	mov rdx,QWORD[24+rcx*1+rsi]
3176	lea rcx,[32+rcx]
3177	mov r10,QWORD[64+rdi]
3178	adox r13,r13
3179	adcx rbx,r11
3180	mov r11,QWORD[72+rdi]
3181	mov QWORD[32+rdi],rax
3182	mov QWORD[40+rdi],rbx
3183
3184	mulx rbx,rax,rdx
3185	adox r10,r10
3186	adcx rax,r12
3187	jrcxz $L$sqrx4x_shift_n_add_break
3188	DB 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00
3189	adox r11,r11
3190	adcx rbx,r13
3191	mov r12,QWORD[80+rdi]
3192	mov r13,QWORD[88+rdi]
3193	mov QWORD[48+rdi],rax
3194	mov QWORD[56+rdi],rbx
3195	lea rdi,[64+rdi]
3196	nop
3197	jmp NEAR $L$sqrx4x_shift_n_add
3198
3199	ALIGN 32
3200	$L$sqrx4x_shift_n_add_break:
3201	adcx rbx,r13
3202	mov QWORD[48+rdi],rax
3203	mov QWORD[56+rdi],rbx
3204	lea rdi,[64+rdi]
3205	DB 102,72,15,126,213
3206	__bn_sqrx8x_reduction:
3207	xor eax,eax
3208	mov rbx,QWORD[((32+8))+rsp]
3209	mov rdx,QWORD[((48+8))+rsp]
3210	lea rcx,[((-64))+r9*1+rbp]
3211
3212	mov QWORD[((0+8))+rsp],rcx
3213	mov QWORD[((8+8))+rsp],rdi
3214
3215	lea rdi,[((48+8))+rsp]
3216	jmp NEAR $L$sqrx8x_reduction_loop
3217
3218	ALIGN 32
3219	$L$sqrx8x_reduction_loop:
3220	mov r9,QWORD[8+rdi]
3221	mov r10,QWORD[16+rdi]
3222	mov r11,QWORD[24+rdi]
3223	mov r12,QWORD[32+rdi]
3224	mov r8,rdx
3225	imul rdx,rbx
3226	mov r13,QWORD[40+rdi]
3227	mov r14,QWORD[48+rdi]
3228	mov r15,QWORD[56+rdi]
3229	mov QWORD[((24+8))+rsp],rax
3230
3231	lea rdi,[64+rdi]
3232	xor rsi,rsi
3233	mov rcx,-8
3234	jmp NEAR $L$sqrx8x_reduce
3235
3236	ALIGN 32
3237	$L$sqrx8x_reduce:
3238	mov rbx,r8
3239	mulx r8,rax,QWORD[rbp]
3240	adcx rax,rbx
3241	adox r8,r9
3242
3243	mulx r9,rbx,QWORD[8+rbp]
3244	adcx r8,rbx
3245	adox r9,r10
3246
3247	mulx r10,rbx,QWORD[16+rbp]
3248	adcx r9,rbx
3249	adox r10,r11
3250
3251	mulx r11,rbx,QWORD[24+rbp]
3252	adcx r10,rbx
3253	adox r11,r12
3254
3255	DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00
3256	mov rax,rdx
3257	mov rdx,r8
3258	adcx r11,rbx
3259	adox r12,r13
3260
3261	mulx rdx,rbx,QWORD[((32+8))+rsp]
3262	mov rdx,rax
3263	mov QWORD[((64+48+8))+rcx*8+rsp],rax
3264
3265	mulx r13,rax,QWORD[40+rbp]
3266	adcx r12,rax
3267	adox r13,r14
3268
3269	mulx r14,rax,QWORD[48+rbp]
3270	adcx r13,rax
3271	adox r14,r15
3272
3273	mulx r15,rax,QWORD[56+rbp]
3274	mov rdx,rbx
3275	adcx r14,rax
3276	adox r15,rsi
3277	adcx r15,rsi
3278
3279	DB 0x67,0x67,0x67
3280	inc rcx
3281	jnz NEAR $L$sqrx8x_reduce
3282
3283	mov rax,rsi
3284	cmp rbp,QWORD[((0+8))+rsp]
3285	jae NEAR $L$sqrx8x_no_tail
3286
3287	mov rdx,QWORD[((48+8))+rsp]
3288	add r8,QWORD[rdi]
3289	lea rbp,[64+rbp]
3290	mov rcx,-8
3291	adcx r9,QWORD[8+rdi]
3292	adcx r10,QWORD[16+rdi]
3293	adc r11,QWORD[24+rdi]
3294	adc r12,QWORD[32+rdi]
3295	adc r13,QWORD[40+rdi]
3296	adc r14,QWORD[48+rdi]
3297	adc r15,QWORD[56+rdi]
3298	lea rdi,[64+rdi]
3299	sbb rax,rax
3300
3301	xor rsi,rsi
3302	mov QWORD[((16+8))+rsp],rax
3303	jmp NEAR $L$sqrx8x_tail
3304
3305	ALIGN 32
3306	$L$sqrx8x_tail:
3307	mov rbx,r8
3308	mulx r8,rax,QWORD[rbp]
3309	adcx rbx,rax
3310	adox r8,r9
3311
3312	mulx r9,rax,QWORD[8+rbp]
3313	adcx r8,rax
3314	adox r9,r10
3315
3316	mulx r10,rax,QWORD[16+rbp]
3317	adcx r9,rax
3318	adox r10,r11
3319
3320	mulx r11,rax,QWORD[24+rbp]
3321	adcx r10,rax
3322	adox r11,r12
3323
3324	DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
3325	adcx r11,rax
3326	adox r12,r13
3327
3328	mulx r13,rax,QWORD[40+rbp]
3329	adcx r12,rax
3330	adox r13,r14
3331
3332	mulx r14,rax,QWORD[48+rbp]
3333	adcx r13,rax
3334	adox r14,r15
3335
3336	mulx r15,rax,QWORD[56+rbp]
3337	mov rdx,QWORD[((72+48+8))+rcx*8+rsp]
3338	adcx r14,rax
3339	adox r15,rsi
3340	mov QWORD[rcx*8+rdi],rbx
3341	mov rbx,r8
3342	adcx r15,rsi
3343
3344	inc rcx
3345	jnz NEAR $L$sqrx8x_tail
3346
3347	cmp rbp,QWORD[((0+8))+rsp]
3348	jae NEAR $L$sqrx8x_tail_done
3349
3350	sub rsi,QWORD[((16+8))+rsp]
3351	mov rdx,QWORD[((48+8))+rsp]
3352	lea rbp,[64+rbp]
3353	adc r8,QWORD[rdi]
3354	adc r9,QWORD[8+rdi]
3355	adc r10,QWORD[16+rdi]
3356	adc r11,QWORD[24+rdi]
3357	adc r12,QWORD[32+rdi]
3358	adc r13,QWORD[40+rdi]
3359	adc r14,QWORD[48+rdi]
3360	adc r15,QWORD[56+rdi]
3361	lea rdi,[64+rdi]
3362	sbb rax,rax
3363	sub rcx,8
3364
3365	xor rsi,rsi
3366	mov QWORD[((16+8))+rsp],rax
3367	jmp NEAR $L$sqrx8x_tail
3368
3369	ALIGN 32
3370	$L$sqrx8x_tail_done:
3371	xor rax,rax
3372	add r8,QWORD[((24+8))+rsp]
3373	adc r9,0
3374	adc r10,0
3375	adc r11,0
3376	adc r12,0
3377	adc r13,0
3378	adc r14,0
3379	adc r15,0
3380	adc rax,0
3381
3382	sub rsi,QWORD[((16+8))+rsp]
3383	$L$sqrx8x_no_tail:
3384	adc r8,QWORD[rdi]
3385	DB 102,72,15,126,217
3386	adc r9,QWORD[8+rdi]
3387	mov rsi,QWORD[56+rbp]
3388	DB 102,72,15,126,213
3389	adc r10,QWORD[16+rdi]
3390	adc r11,QWORD[24+rdi]
3391	adc r12,QWORD[32+rdi]
3392	adc r13,QWORD[40+rdi]
3393	adc r14,QWORD[48+rdi]
3394	adc r15,QWORD[56+rdi]
3395	adc rax,0
3396
3397	mov rbx,QWORD[((32+8))+rsp]
3398	mov rdx,QWORD[64+rcx*1+rdi]
3399
3400	mov QWORD[rdi],r8
3401	lea r8,[64+rdi]
3402	mov QWORD[8+rdi],r9
3403	mov QWORD[16+rdi],r10
3404	mov QWORD[24+rdi],r11
3405	mov QWORD[32+rdi],r12
3406	mov QWORD[40+rdi],r13
3407	mov QWORD[48+rdi],r14
3408	mov QWORD[56+rdi],r15
3409
3410	lea rdi,[64+rcx*1+rdi]
3411	cmp r8,QWORD[((8+8))+rsp]
3412	jb NEAR $L$sqrx8x_reduction_loop
3413	DB 0F3h,0C3h ;repret
3414
3415
3416	ALIGN 32
3417	__bn_postx4x_internal:
3418
3419	mov r12,QWORD[rbp]
3420	mov r10,rcx
3421	mov r9,rcx
3422	neg rax
3423	sar rcx,3+2
3424
3425	DB 102,72,15,126,202
3426	DB 102,72,15,126,206
3427	dec r12
3428	mov r13,QWORD[8+rbp]
3429	xor r8,r8
3430	mov r14,QWORD[16+rbp]
3431	mov r15,QWORD[24+rbp]
3432	jmp NEAR $L$sqrx4x_sub_entry
3433
3434	ALIGN 16
3435	$L$sqrx4x_sub:
3436	mov r12,QWORD[rbp]
3437	mov r13,QWORD[8+rbp]
3438	mov r14,QWORD[16+rbp]
3439	mov r15,QWORD[24+rbp]
3440	$L$sqrx4x_sub_entry:
3441	andn r12,r12,rax
3442	lea rbp,[32+rbp]
3443	andn r13,r13,rax
3444	andn r14,r14,rax
3445	andn r15,r15,rax
3446
3447	neg r8
3448	adc r12,QWORD[rdi]
3449	adc r13,QWORD[8+rdi]
3450	adc r14,QWORD[16+rdi]
3451	adc r15,QWORD[24+rdi]
3452	mov QWORD[rdx],r12
3453	lea rdi,[32+rdi]
3454	mov QWORD[8+rdx],r13
3455	sbb r8,r8
3456	mov QWORD[16+rdx],r14
3457	mov QWORD[24+rdx],r15
3458	lea rdx,[32+rdx]
3459
3460	inc rcx
3461	jnz NEAR $L$sqrx4x_sub
3462
3463	neg r9
3464
3465	DB 0F3h,0C3h ;repret
3466
3467
3468	global bn_get_bits5
3469
3470	ALIGN 16
3471	bn_get_bits5:
3472
3473	lea r10,[rcx]
3474	lea r11,[1+rcx]
3475	mov ecx,edx
3476	shr edx,4
3477	and ecx,15
3478	lea eax,[((-8))+rcx]
3479	cmp ecx,11
3480	cmova r10,r11
3481	cmova ecx,eax
3482	movzx eax,WORD[rdx*2+r10]
3483	shr eax,cl
3484	and eax,31
3485	DB 0F3h,0C3h ;repret
3486
3487
3488
3489	global bn_scatter5
3490
3491	ALIGN 16
3492	bn_scatter5:
3493
3494	cmp edx,0
3495	jz NEAR $L$scatter_epilogue
3496	lea r8,[r9*8+r8]
3497	$L$scatter:
3498	mov rax,QWORD[rcx]
3499	lea rcx,[8+rcx]
3500	mov QWORD[r8],rax
3501	lea r8,[256+r8]
3502	sub edx,1
3503	jnz NEAR $L$scatter
3504	$L$scatter_epilogue:
3505	DB 0F3h,0C3h ;repret
3506
3507
3508
3509	global bn_gather5
3510
3511	ALIGN 32
3512	bn_gather5:
3513	$L$SEH_begin_bn_gather5:
3514
3515
3516	DB 0x4c,0x8d,0x14,0x24
3517	DB 0x48,0x81,0xec,0x08,0x01,0x00,0x00
3518	lea rax,[$L$inc]
3519	and rsp,-16
3520
3521	movd xmm5,r9d
3522	movdqa xmm0,XMMWORD[rax]
3523	movdqa xmm1,XMMWORD[16+rax]
3524	lea r11,[128+r8]
3525	lea rax,[128+rsp]
3526
3527	pshufd xmm5,xmm5,0
3528	movdqa xmm4,xmm1
3529	movdqa xmm2,xmm1
3530	paddd xmm1,xmm0
3531	pcmpeqd xmm0,xmm5
3532	movdqa xmm3,xmm4
3533
3534	paddd xmm2,xmm1
3535	pcmpeqd xmm1,xmm5
3536	movdqa XMMWORD[(-128)+rax],xmm0
3537	movdqa xmm0,xmm4
3538
3539	paddd xmm3,xmm2
3540	pcmpeqd xmm2,xmm5
3541	movdqa XMMWORD[(-112)+rax],xmm1
3542	movdqa xmm1,xmm4
3543
3544	paddd xmm0,xmm3
3545	pcmpeqd xmm3,xmm5
3546	movdqa XMMWORD[(-96)+rax],xmm2
3547	movdqa xmm2,xmm4
3548	paddd xmm1,xmm0
3549	pcmpeqd xmm0,xmm5
3550	movdqa XMMWORD[(-80)+rax],xmm3
3551	movdqa xmm3,xmm4
3552
3553	paddd xmm2,xmm1
3554	pcmpeqd xmm1,xmm5
3555	movdqa XMMWORD[(-64)+rax],xmm0
3556	movdqa xmm0,xmm4
3557
3558	paddd xmm3,xmm2
3559	pcmpeqd xmm2,xmm5
3560	movdqa XMMWORD[(-48)+rax],xmm1
3561	movdqa xmm1,xmm4
3562
3563	paddd xmm0,xmm3
3564	pcmpeqd xmm3,xmm5
3565	movdqa XMMWORD[(-32)+rax],xmm2
3566	movdqa xmm2,xmm4
3567	paddd xmm1,xmm0
3568	pcmpeqd xmm0,xmm5
3569	movdqa XMMWORD[(-16)+rax],xmm3
3570	movdqa xmm3,xmm4
3571
3572	paddd xmm2,xmm1
3573	pcmpeqd xmm1,xmm5
3574	movdqa XMMWORD[rax],xmm0
3575	movdqa xmm0,xmm4
3576
3577	paddd xmm3,xmm2
3578	pcmpeqd xmm2,xmm5
3579	movdqa XMMWORD[16+rax],xmm1
3580	movdqa xmm1,xmm4
3581
3582	paddd xmm0,xmm3
3583	pcmpeqd xmm3,xmm5
3584	movdqa XMMWORD[32+rax],xmm2
3585	movdqa xmm2,xmm4
3586	paddd xmm1,xmm0
3587	pcmpeqd xmm0,xmm5
3588	movdqa XMMWORD[48+rax],xmm3
3589	movdqa xmm3,xmm4
3590
3591	paddd xmm2,xmm1
3592	pcmpeqd xmm1,xmm5
3593	movdqa XMMWORD[64+rax],xmm0
3594	movdqa xmm0,xmm4
3595
3596	paddd xmm3,xmm2
3597	pcmpeqd xmm2,xmm5
3598	movdqa XMMWORD[80+rax],xmm1
3599	movdqa xmm1,xmm4
3600
3601	paddd xmm0,xmm3
3602	pcmpeqd xmm3,xmm5
3603	movdqa XMMWORD[96+rax],xmm2
3604	movdqa xmm2,xmm4
3605	movdqa XMMWORD[112+rax],xmm3
3606	jmp NEAR $L$gather
3607
3608	ALIGN 32
3609	$L$gather:
3610	pxor xmm4,xmm4
3611	pxor xmm5,xmm5
3612	movdqa xmm0,XMMWORD[((-128))+r11]
3613	movdqa xmm1,XMMWORD[((-112))+r11]
3614	movdqa xmm2,XMMWORD[((-96))+r11]
3615	pand xmm0,XMMWORD[((-128))+rax]
3616	movdqa xmm3,XMMWORD[((-80))+r11]
3617	pand xmm1,XMMWORD[((-112))+rax]
3618	por xmm4,xmm0
3619	pand xmm2,XMMWORD[((-96))+rax]
3620	por xmm5,xmm1
3621	pand xmm3,XMMWORD[((-80))+rax]
3622	por xmm4,xmm2
3623	por xmm5,xmm3
3624	movdqa xmm0,XMMWORD[((-64))+r11]
3625	movdqa xmm1,XMMWORD[((-48))+r11]
3626	movdqa xmm2,XMMWORD[((-32))+r11]
3627	pand xmm0,XMMWORD[((-64))+rax]
3628	movdqa xmm3,XMMWORD[((-16))+r11]
3629	pand xmm1,XMMWORD[((-48))+rax]
3630	por xmm4,xmm0
3631	pand xmm2,XMMWORD[((-32))+rax]
3632	por xmm5,xmm1
3633	pand xmm3,XMMWORD[((-16))+rax]
3634	por xmm4,xmm2
3635	por xmm5,xmm3
3636	movdqa xmm0,XMMWORD[r11]
3637	movdqa xmm1,XMMWORD[16+r11]
3638	movdqa xmm2,XMMWORD[32+r11]
3639	pand xmm0,XMMWORD[rax]
3640	movdqa xmm3,XMMWORD[48+r11]
3641	pand xmm1,XMMWORD[16+rax]
3642	por xmm4,xmm0
3643	pand xmm2,XMMWORD[32+rax]
3644	por xmm5,xmm1
3645	pand xmm3,XMMWORD[48+rax]
3646	por xmm4,xmm2
3647	por xmm5,xmm3
3648	movdqa xmm0,XMMWORD[64+r11]
3649	movdqa xmm1,XMMWORD[80+r11]
3650	movdqa xmm2,XMMWORD[96+r11]
3651	pand xmm0,XMMWORD[64+rax]
3652	movdqa xmm3,XMMWORD[112+r11]
3653	pand xmm1,XMMWORD[80+rax]
3654	por xmm4,xmm0
3655	pand xmm2,XMMWORD[96+rax]
3656	por xmm5,xmm1
3657	pand xmm3,XMMWORD[112+rax]
3658	por xmm4,xmm2
3659	por xmm5,xmm3
3660	por xmm4,xmm5
3661	lea r11,[256+r11]
3662	pshufd xmm0,xmm4,0x4e
3663	por xmm0,xmm4
3664	movq QWORD[rcx],xmm0
3665	lea rcx,[8+rcx]
3666	sub edx,1
3667	jnz NEAR $L$gather
3668
3669	lea rsp,[r10]
3670	DB 0F3h,0C3h ;repret
3671	$L$SEH_end_bn_gather5:
3672
3673
3674	ALIGN 64
3675	$L$inc:
3676	DD 0,0,1,1
3677	DD 2,2,2,2
3678	DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
3679	DB 112,108,105,99,97,116,105,111,110,32,119,105,116,104,32,115
3680	DB 99,97,116,116,101,114,47,103,97,116,104,101,114,32,102,111
3681	DB 114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79
3682	DB 71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111
3683	DB 112,101,110,115,115,108,46,111,114,103,62,0
3684	EXTERN __imp_RtlVirtualUnwind
3685
3686	ALIGN 16
3687	mul_handler:
3688	push rsi
3689	push rdi
3690	push rbx
3691	push rbp
3692	push r12
3693	push r13
3694	push r14
3695	push r15
3696	pushfq
3697	sub rsp,64
3698
3699	mov rax,QWORD[120+r8]
3700	mov rbx,QWORD[248+r8]
3701
3702	mov rsi,QWORD[8+r9]
3703	mov r11,QWORD[56+r9]
3704
3705	mov r10d,DWORD[r11]
3706	lea r10,[r10*1+rsi]
3707	cmp rbx,r10
3708	jb NEAR $L$common_seh_tail
3709
3710	mov r10d,DWORD[4+r11]
3711	lea r10,[r10*1+rsi]
3712	cmp rbx,r10
3713	jb NEAR $L$common_pop_regs
3714
3715	mov rax,QWORD[152+r8]
3716
3717	mov r10d,DWORD[8+r11]
3718	lea r10,[r10*1+rsi]
3719	cmp rbx,r10
3720	jae NEAR $L$common_seh_tail
3721
3722	lea r10,[$L$mul_epilogue]
3723	cmp rbx,r10
3724	ja NEAR $L$body_40
3725
3726	mov r10,QWORD[192+r8]
3727	mov rax,QWORD[8+r10*8+rax]
3728
3729	jmp NEAR $L$common_pop_regs
3730
3731	$L$body_40:
3732	mov rax,QWORD[40+rax]
3733	$L$common_pop_regs:
3734	mov rbx,QWORD[((-8))+rax]
3735	mov rbp,QWORD[((-16))+rax]
3736	mov r12,QWORD[((-24))+rax]
3737	mov r13,QWORD[((-32))+rax]
3738	mov r14,QWORD[((-40))+rax]
3739	mov r15,QWORD[((-48))+rax]
3740	mov QWORD[144+r8],rbx
3741	mov QWORD[160+r8],rbp
3742	mov QWORD[216+r8],r12
3743	mov QWORD[224+r8],r13
3744	mov QWORD[232+r8],r14
3745	mov QWORD[240+r8],r15
3746
3747	$L$common_seh_tail:
3748	mov rdi,QWORD[8+rax]
3749	mov rsi,QWORD[16+rax]
3750	mov QWORD[152+r8],rax
3751	mov QWORD[168+r8],rsi
3752	mov QWORD[176+r8],rdi
3753
3754	mov rdi,QWORD[40+r9]
3755	mov rsi,r8
3756	mov ecx,154
3757	DD 0xa548f3fc
3758
3759	mov rsi,r9
3760	xor rcx,rcx
3761	mov rdx,QWORD[8+rsi]
3762	mov r8,QWORD[rsi]
3763	mov r9,QWORD[16+rsi]
3764	mov r10,QWORD[40+rsi]
3765	lea r11,[56+rsi]
3766	lea r12,[24+rsi]
3767	mov QWORD[32+rsp],r10
3768	mov QWORD[40+rsp],r11
3769	mov QWORD[48+rsp],r12
3770	mov QWORD[56+rsp],rcx
3771	call QWORD[__imp_RtlVirtualUnwind]
3772
3773	mov eax,1
3774	add rsp,64
3775	popfq
3776	pop r15
3777	pop r14
3778	pop r13
3779	pop r12
3780	pop rbp
3781	pop rbx
3782	pop rdi
3783	pop rsi
3784	DB 0F3h,0C3h ;repret
3785
3786
3787	section .pdata rdata align=4
3788	ALIGN 4
3789	DD $L$SEH_begin_bn_mul_mont_gather5 wrt ..imagebase
3790	DD $L$SEH_end_bn_mul_mont_gather5 wrt ..imagebase
3791	DD $L$SEH_info_bn_mul_mont_gather5 wrt ..imagebase
3792
3793	DD $L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase
3794	DD $L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase
3795	DD $L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase
3796
3797	DD $L$SEH_begin_bn_power5 wrt ..imagebase
3798	DD $L$SEH_end_bn_power5 wrt ..imagebase
3799	DD $L$SEH_info_bn_power5 wrt ..imagebase
3800	DD $L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase
3801	DD $L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase
3802	DD $L$SEH_info_bn_mulx4x_mont_gather5 wrt ..imagebase
3803
3804	DD $L$SEH_begin_bn_powerx5 wrt ..imagebase
3805	DD $L$SEH_end_bn_powerx5 wrt ..imagebase
3806	DD $L$SEH_info_bn_powerx5 wrt ..imagebase
3807	DD $L$SEH_begin_bn_gather5 wrt ..imagebase
3808	DD $L$SEH_end_bn_gather5 wrt ..imagebase
3809	DD $L$SEH_info_bn_gather5 wrt ..imagebase
3810
3811	section .xdata rdata align=8
3812	ALIGN 8
3813	$L$SEH_info_bn_mul_mont_gather5:
3814	DB 9,0,0,0
3815	DD mul_handler wrt ..imagebase
3816	DD $L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
3817	ALIGN 8
3818	$L$SEH_info_bn_mul4x_mont_gather5:
3819	DB 9,0,0,0
3820	DD mul_handler wrt ..imagebase
3821	DD $L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
3822	ALIGN 8
3823	$L$SEH_info_bn_power5:
3824	DB 9,0,0,0
3825	DD mul_handler wrt ..imagebase
3826	DD $L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase
3827	ALIGN 8
3828	$L$SEH_info_bn_mulx4x_mont_gather5:
3829	DB 9,0,0,0
3830	DD mul_handler wrt ..imagebase
3831	DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase
3832	ALIGN 8
3833	$L$SEH_info_bn_powerx5:
3834	DB 9,0,0,0
3835	DD mul_handler wrt ..imagebase
3836	DD $L$powerx5_prologue wrt ..imagebase,$L$powerx5_body wrt ..imagebase,$L$powerx5_epilogue wrt ..imagebase
3837	ALIGN 8
3838	$L$SEH_info_bn_gather5:
3839	DB 0x01,0x0b,0x03,0x0a
3840	DB 0x0b,0x01,0x21,0x00
3841	DB 0x04,0xa3,0x00,0x00
3842	ALIGN 8

注意: 瀏覽 TracBrowser 來幫助您使用儲存庫瀏覽器

以其他格式下載:

原來格式