x86_64-mont5.S@ 95219

Last change on this file since 95219 was 95219, checked in by vboxsync, 3 years ago
libs/openssl: Switched to v3.0.3, bugref:10128
File size: 62.7 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_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	global bn_from_montgomery
2099
2100	ALIGN 32
2101	bn_from_montgomery:
2102
2103	test DWORD[48+rsp],7
2104	jz NEAR bn_from_mont8x
2105	xor eax,eax
2106	DB 0F3h,0C3h ;repret
2107
2108
2109
2110
2111	ALIGN 32
2112	bn_from_mont8x:
2113	mov QWORD[8+rsp],rdi ;WIN64 prologue
2114	mov QWORD[16+rsp],rsi
2115	mov rax,rsp
2116	$L$SEH_begin_bn_from_mont8x:
2117	mov rdi,rcx
2118	mov rsi,rdx
2119	mov rdx,r8
2120	mov rcx,r9
2121	mov r8,QWORD[40+rsp]
2122	mov r9,QWORD[48+rsp]
2123
2124
2125
2126	DB 0x67
2127	mov rax,rsp
2128
2129	push rbx
2130
2131	push rbp
2132
2133	push r12
2134
2135	push r13
2136
2137	push r14
2138
2139	push r15
2140
2141	$L$from_prologue:
2142
2143	shl r9d,3
2144	lea r10,[r9*2+r9]
2145	neg r9
2146	mov r8,QWORD[r8]
2147
2148
2149
2150
2151
2152
2153
2154
2155	lea r11,[((-320))+r9*2+rsp]
2156	mov rbp,rsp
2157	sub r11,rdi
2158	and r11,4095
2159	cmp r10,r11
2160	jb NEAR $L$from_sp_alt
2161	sub rbp,r11
2162	lea rbp,[((-320))+r9*2+rbp]
2163	jmp NEAR $L$from_sp_done
2164
2165	ALIGN 32
2166	$L$from_sp_alt:
2167	lea r10,[((4096-320))+r9*2]
2168	lea rbp,[((-320))+r9*2+rbp]
2169	sub r11,r10
2170	mov r10,0
2171	cmovc r11,r10
2172	sub rbp,r11
2173	$L$from_sp_done:
2174	and rbp,-64
2175	mov r11,rsp
2176	sub r11,rbp
2177	and r11,-4096
2178	lea rsp,[rbp*1+r11]
2179	mov r10,QWORD[rsp]
2180	cmp rsp,rbp
2181	ja NEAR $L$from_page_walk
2182	jmp NEAR $L$from_page_walk_done
2183
2184	$L$from_page_walk:
2185	lea rsp,[((-4096))+rsp]
2186	mov r10,QWORD[rsp]
2187	cmp rsp,rbp
2188	ja NEAR $L$from_page_walk
2189	$L$from_page_walk_done:
2190
2191	mov r10,r9
2192	neg r9
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203	mov QWORD[32+rsp],r8
2204	mov QWORD[40+rsp],rax
2205
2206	$L$from_body:
2207	mov r11,r9
2208	lea rax,[48+rsp]
2209	pxor xmm0,xmm0
2210	jmp NEAR $L$mul_by_1
2211
2212	ALIGN 32
2213	$L$mul_by_1:
2214	movdqu xmm1,XMMWORD[rsi]
2215	movdqu xmm2,XMMWORD[16+rsi]
2216	movdqu xmm3,XMMWORD[32+rsi]
2217	movdqa XMMWORD[r9*1+rax],xmm0
2218	movdqu xmm4,XMMWORD[48+rsi]
2219	movdqa XMMWORD[16+r9*1+rax],xmm0
2220	DB 0x48,0x8d,0xb6,0x40,0x00,0x00,0x00
2221	movdqa XMMWORD[rax],xmm1
2222	movdqa XMMWORD[32+r9*1+rax],xmm0
2223	movdqa XMMWORD[16+rax],xmm2
2224	movdqa XMMWORD[48+r9*1+rax],xmm0
2225	movdqa XMMWORD[32+rax],xmm3
2226	movdqa XMMWORD[48+rax],xmm4
2227	lea rax,[64+rax]
2228	sub r11,64
2229	jnz NEAR $L$mul_by_1
2230
2231	DB 102,72,15,110,207
2232	DB 102,72,15,110,209
2233	DB 0x67
2234	mov rbp,rcx
2235	DB 102,73,15,110,218
2236	mov r11d,DWORD[((OPENSSL_ia32cap_P+8))]
2237	and r11d,0x80108
2238	cmp r11d,0x80108
2239	jne NEAR $L$from_mont_nox
2240
2241	lea rdi,[r9*1+rax]
2242	call __bn_sqrx8x_reduction
2243	call __bn_postx4x_internal
2244
2245	pxor xmm0,xmm0
2246	lea rax,[48+rsp]
2247	jmp NEAR $L$from_mont_zero
2248
2249	ALIGN 32
2250	$L$from_mont_nox:
2251	call __bn_sqr8x_reduction
2252	call __bn_post4x_internal
2253
2254	pxor xmm0,xmm0
2255	lea rax,[48+rsp]
2256	jmp NEAR $L$from_mont_zero
2257
2258	ALIGN 32
2259	$L$from_mont_zero:
2260	mov rsi,QWORD[40+rsp]
2261
2262	movdqa XMMWORD[rax],xmm0
2263	movdqa XMMWORD[16+rax],xmm0
2264	movdqa XMMWORD[32+rax],xmm0
2265	movdqa XMMWORD[48+rax],xmm0
2266	lea rax,[64+rax]
2267	sub r9,32
2268	jnz NEAR $L$from_mont_zero
2269
2270	mov rax,1
2271	mov r15,QWORD[((-48))+rsi]
2272
2273	mov r14,QWORD[((-40))+rsi]
2274
2275	mov r13,QWORD[((-32))+rsi]
2276
2277	mov r12,QWORD[((-24))+rsi]
2278
2279	mov rbp,QWORD[((-16))+rsi]
2280
2281	mov rbx,QWORD[((-8))+rsi]
2282
2283	lea rsp,[rsi]
2284
2285	$L$from_epilogue:
2286	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2287	mov rsi,QWORD[16+rsp]
2288	DB 0F3h,0C3h ;repret
2289
2290	$L$SEH_end_bn_from_mont8x:
2291
2292	ALIGN 32
2293	bn_mulx4x_mont_gather5:
2294	mov QWORD[8+rsp],rdi ;WIN64 prologue
2295	mov QWORD[16+rsp],rsi
2296	mov rax,rsp
2297	$L$SEH_begin_bn_mulx4x_mont_gather5:
2298	mov rdi,rcx
2299	mov rsi,rdx
2300	mov rdx,r8
2301	mov rcx,r9
2302	mov r8,QWORD[40+rsp]
2303	mov r9,QWORD[48+rsp]
2304
2305
2306
2307	mov rax,rsp
2308
2309	$L$mulx4x_enter:
2310	push rbx
2311
2312	push rbp
2313
2314	push r12
2315
2316	push r13
2317
2318	push r14
2319
2320	push r15
2321
2322	$L$mulx4x_prologue:
2323
2324	shl r9d,3
2325	lea r10,[r9*2+r9]
2326	neg r9
2327	mov r8,QWORD[r8]
2328
2329
2330
2331
2332
2333
2334
2335
2336
2337
2338	lea r11,[((-320))+r9*2+rsp]
2339	mov rbp,rsp
2340	sub r11,rdi
2341	and r11,4095
2342	cmp r10,r11
2343	jb NEAR $L$mulx4xsp_alt
2344	sub rbp,r11
2345	lea rbp,[((-320))+r9*2+rbp]
2346	jmp NEAR $L$mulx4xsp_done
2347
2348	$L$mulx4xsp_alt:
2349	lea r10,[((4096-320))+r9*2]
2350	lea rbp,[((-320))+r9*2+rbp]
2351	sub r11,r10
2352	mov r10,0
2353	cmovc r11,r10
2354	sub rbp,r11
2355	$L$mulx4xsp_done:
2356	and rbp,-64
2357	mov r11,rsp
2358	sub r11,rbp
2359	and r11,-4096
2360	lea rsp,[rbp*1+r11]
2361	mov r10,QWORD[rsp]
2362	cmp rsp,rbp
2363	ja NEAR $L$mulx4x_page_walk
2364	jmp NEAR $L$mulx4x_page_walk_done
2365
2366	$L$mulx4x_page_walk:
2367	lea rsp,[((-4096))+rsp]
2368	mov r10,QWORD[rsp]
2369	cmp rsp,rbp
2370	ja NEAR $L$mulx4x_page_walk
2371	$L$mulx4x_page_walk_done:
2372
2373
2374
2375
2376
2377
2378
2379
2380
2381
2382
2383
2384
2385	mov QWORD[32+rsp],r8
2386	mov QWORD[40+rsp],rax
2387
2388	$L$mulx4x_body:
2389	call mulx4x_internal
2390
2391	mov rsi,QWORD[40+rsp]
2392
2393	mov rax,1
2394
2395	mov r15,QWORD[((-48))+rsi]
2396
2397	mov r14,QWORD[((-40))+rsi]
2398
2399	mov r13,QWORD[((-32))+rsi]
2400
2401	mov r12,QWORD[((-24))+rsi]
2402
2403	mov rbp,QWORD[((-16))+rsi]
2404
2405	mov rbx,QWORD[((-8))+rsi]
2406
2407	lea rsp,[rsi]
2408
2409	$L$mulx4x_epilogue:
2410	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2411	mov rsi,QWORD[16+rsp]
2412	DB 0F3h,0C3h ;repret
2413
2414	$L$SEH_end_bn_mulx4x_mont_gather5:
2415
2416
2417	ALIGN 32
2418	mulx4x_internal:
2419
2420	mov QWORD[8+rsp],r9
2421	mov r10,r9
2422	neg r9
2423	shl r9,5
2424	neg r10
2425	lea r13,[128+r9*1+rdx]
2426	shr r9,5+5
2427	movd xmm5,DWORD[56+rax]
2428	sub r9,1
2429	lea rax,[$L$inc]
2430	mov QWORD[((16+8))+rsp],r13
2431	mov QWORD[((24+8))+rsp],r9
2432	mov QWORD[((56+8))+rsp],rdi
2433	movdqa xmm0,XMMWORD[rax]
2434	movdqa xmm1,XMMWORD[16+rax]
2435	lea r10,[((88-112))+r10*1+rsp]
2436	lea rdi,[128+rdx]
2437
2438	pshufd xmm5,xmm5,0
2439	movdqa xmm4,xmm1
2440	DB 0x67
2441	movdqa xmm2,xmm1
2442	DB 0x67
2443	paddd xmm1,xmm0
2444	pcmpeqd xmm0,xmm5
2445	movdqa xmm3,xmm4
2446	paddd xmm2,xmm1
2447	pcmpeqd xmm1,xmm5
2448	movdqa XMMWORD[112+r10],xmm0
2449	movdqa xmm0,xmm4
2450
2451	paddd xmm3,xmm2
2452	pcmpeqd xmm2,xmm5
2453	movdqa XMMWORD[128+r10],xmm1
2454	movdqa xmm1,xmm4
2455
2456	paddd xmm0,xmm3
2457	pcmpeqd xmm3,xmm5
2458	movdqa XMMWORD[144+r10],xmm2
2459	movdqa xmm2,xmm4
2460
2461	paddd xmm1,xmm0
2462	pcmpeqd xmm0,xmm5
2463	movdqa XMMWORD[160+r10],xmm3
2464	movdqa xmm3,xmm4
2465	paddd xmm2,xmm1
2466	pcmpeqd xmm1,xmm5
2467	movdqa XMMWORD[176+r10],xmm0
2468	movdqa xmm0,xmm4
2469
2470	paddd xmm3,xmm2
2471	pcmpeqd xmm2,xmm5
2472	movdqa XMMWORD[192+r10],xmm1
2473	movdqa xmm1,xmm4
2474
2475	paddd xmm0,xmm3
2476	pcmpeqd xmm3,xmm5
2477	movdqa XMMWORD[208+r10],xmm2
2478	movdqa xmm2,xmm4
2479
2480	paddd xmm1,xmm0
2481	pcmpeqd xmm0,xmm5
2482	movdqa XMMWORD[224+r10],xmm3
2483	movdqa xmm3,xmm4
2484	paddd xmm2,xmm1
2485	pcmpeqd xmm1,xmm5
2486	movdqa XMMWORD[240+r10],xmm0
2487	movdqa xmm0,xmm4
2488
2489	paddd xmm3,xmm2
2490	pcmpeqd xmm2,xmm5
2491	movdqa XMMWORD[256+r10],xmm1
2492	movdqa xmm1,xmm4
2493
2494	paddd xmm0,xmm3
2495	pcmpeqd xmm3,xmm5
2496	movdqa XMMWORD[272+r10],xmm2
2497	movdqa xmm2,xmm4
2498
2499	paddd xmm1,xmm0
2500	pcmpeqd xmm0,xmm5
2501	movdqa XMMWORD[288+r10],xmm3
2502	movdqa xmm3,xmm4
2503	DB 0x67
2504	paddd xmm2,xmm1
2505	pcmpeqd xmm1,xmm5
2506	movdqa XMMWORD[304+r10],xmm0
2507
2508	paddd xmm3,xmm2
2509	pcmpeqd xmm2,xmm5
2510	movdqa XMMWORD[320+r10],xmm1
2511
2512	pcmpeqd xmm3,xmm5
2513	movdqa XMMWORD[336+r10],xmm2
2514
2515	pand xmm0,XMMWORD[64+rdi]
2516	pand xmm1,XMMWORD[80+rdi]
2517	pand xmm2,XMMWORD[96+rdi]
2518	movdqa XMMWORD[352+r10],xmm3
2519	pand xmm3,XMMWORD[112+rdi]
2520	por xmm0,xmm2
2521	por xmm1,xmm3
2522	movdqa xmm4,XMMWORD[((-128))+rdi]
2523	movdqa xmm5,XMMWORD[((-112))+rdi]
2524	movdqa xmm2,XMMWORD[((-96))+rdi]
2525	pand xmm4,XMMWORD[112+r10]
2526	movdqa xmm3,XMMWORD[((-80))+rdi]
2527	pand xmm5,XMMWORD[128+r10]
2528	por xmm0,xmm4
2529	pand xmm2,XMMWORD[144+r10]
2530	por xmm1,xmm5
2531	pand xmm3,XMMWORD[160+r10]
2532	por xmm0,xmm2
2533	por xmm1,xmm3
2534	movdqa xmm4,XMMWORD[((-64))+rdi]
2535	movdqa xmm5,XMMWORD[((-48))+rdi]
2536	movdqa xmm2,XMMWORD[((-32))+rdi]
2537	pand xmm4,XMMWORD[176+r10]
2538	movdqa xmm3,XMMWORD[((-16))+rdi]
2539	pand xmm5,XMMWORD[192+r10]
2540	por xmm0,xmm4
2541	pand xmm2,XMMWORD[208+r10]
2542	por xmm1,xmm5
2543	pand xmm3,XMMWORD[224+r10]
2544	por xmm0,xmm2
2545	por xmm1,xmm3
2546	movdqa xmm4,XMMWORD[rdi]
2547	movdqa xmm5,XMMWORD[16+rdi]
2548	movdqa xmm2,XMMWORD[32+rdi]
2549	pand xmm4,XMMWORD[240+r10]
2550	movdqa xmm3,XMMWORD[48+rdi]
2551	pand xmm5,XMMWORD[256+r10]
2552	por xmm0,xmm4
2553	pand xmm2,XMMWORD[272+r10]
2554	por xmm1,xmm5
2555	pand xmm3,XMMWORD[288+r10]
2556	por xmm0,xmm2
2557	por xmm1,xmm3
2558	pxor xmm0,xmm1
2559	pshufd xmm1,xmm0,0x4e
2560	por xmm0,xmm1
2561	lea rdi,[256+rdi]
2562	DB 102,72,15,126,194
2563	lea rbx,[((64+32+8))+rsp]
2564
2565	mov r9,rdx
2566	mulx rax,r8,QWORD[rsi]
2567	mulx r12,r11,QWORD[8+rsi]
2568	add r11,rax
2569	mulx r13,rax,QWORD[16+rsi]
2570	adc r12,rax
2571	adc r13,0
2572	mulx r14,rax,QWORD[24+rsi]
2573
2574	mov r15,r8
2575	imul r8,QWORD[((32+8))+rsp]
2576	xor rbp,rbp
2577	mov rdx,r8
2578
2579	mov QWORD[((8+8))+rsp],rdi
2580
2581	lea rsi,[32+rsi]
2582	adcx r13,rax
2583	adcx r14,rbp
2584
2585	mulx r10,rax,QWORD[rcx]
2586	adcx r15,rax
2587	adox r10,r11
2588	mulx r11,rax,QWORD[8+rcx]
2589	adcx r10,rax
2590	adox r11,r12
2591	mulx r12,rax,QWORD[16+rcx]
2592	mov rdi,QWORD[((24+8))+rsp]
2593	mov QWORD[((-32))+rbx],r10
2594	adcx r11,rax
2595	adox r12,r13
2596	mulx r15,rax,QWORD[24+rcx]
2597	mov rdx,r9
2598	mov QWORD[((-24))+rbx],r11
2599	adcx r12,rax
2600	adox r15,rbp
2601	lea rcx,[32+rcx]
2602	mov QWORD[((-16))+rbx],r12
2603	jmp NEAR $L$mulx4x_1st
2604
2605	ALIGN 32
2606	$L$mulx4x_1st:
2607	adcx r15,rbp
2608	mulx rax,r10,QWORD[rsi]
2609	adcx r10,r14
2610	mulx r14,r11,QWORD[8+rsi]
2611	adcx r11,rax
2612	mulx rax,r12,QWORD[16+rsi]
2613	adcx r12,r14
2614	mulx r14,r13,QWORD[24+rsi]
2615	DB 0x67,0x67
2616	mov rdx,r8
2617	adcx r13,rax
2618	adcx r14,rbp
2619	lea rsi,[32+rsi]
2620	lea rbx,[32+rbx]
2621
2622	adox r10,r15
2623	mulx r15,rax,QWORD[rcx]
2624	adcx r10,rax
2625	adox r11,r15
2626	mulx r15,rax,QWORD[8+rcx]
2627	adcx r11,rax
2628	adox r12,r15
2629	mulx r15,rax,QWORD[16+rcx]
2630	mov QWORD[((-40))+rbx],r10
2631	adcx r12,rax
2632	mov QWORD[((-32))+rbx],r11
2633	adox r13,r15
2634	mulx r15,rax,QWORD[24+rcx]
2635	mov rdx,r9
2636	mov QWORD[((-24))+rbx],r12
2637	adcx r13,rax
2638	adox r15,rbp
2639	lea rcx,[32+rcx]
2640	mov QWORD[((-16))+rbx],r13
2641
2642	dec rdi
2643	jnz NEAR $L$mulx4x_1st
2644
2645	mov rax,QWORD[8+rsp]
2646	adc r15,rbp
2647	lea rsi,[rax*1+rsi]
2648	add r14,r15
2649	mov rdi,QWORD[((8+8))+rsp]
2650	adc rbp,rbp
2651	mov QWORD[((-8))+rbx],r14
2652	jmp NEAR $L$mulx4x_outer
2653
2654	ALIGN 32
2655	$L$mulx4x_outer:
2656	lea r10,[((16-256))+rbx]
2657	pxor xmm4,xmm4
2658	DB 0x67,0x67
2659	pxor xmm5,xmm5
2660	movdqa xmm0,XMMWORD[((-128))+rdi]
2661	movdqa xmm1,XMMWORD[((-112))+rdi]
2662	movdqa xmm2,XMMWORD[((-96))+rdi]
2663	pand xmm0,XMMWORD[256+r10]
2664	movdqa xmm3,XMMWORD[((-80))+rdi]
2665	pand xmm1,XMMWORD[272+r10]
2666	por xmm4,xmm0
2667	pand xmm2,XMMWORD[288+r10]
2668	por xmm5,xmm1
2669	pand xmm3,XMMWORD[304+r10]
2670	por xmm4,xmm2
2671	por xmm5,xmm3
2672	movdqa xmm0,XMMWORD[((-64))+rdi]
2673	movdqa xmm1,XMMWORD[((-48))+rdi]
2674	movdqa xmm2,XMMWORD[((-32))+rdi]
2675	pand xmm0,XMMWORD[320+r10]
2676	movdqa xmm3,XMMWORD[((-16))+rdi]
2677	pand xmm1,XMMWORD[336+r10]
2678	por xmm4,xmm0
2679	pand xmm2,XMMWORD[352+r10]
2680	por xmm5,xmm1
2681	pand xmm3,XMMWORD[368+r10]
2682	por xmm4,xmm2
2683	por xmm5,xmm3
2684	movdqa xmm0,XMMWORD[rdi]
2685	movdqa xmm1,XMMWORD[16+rdi]
2686	movdqa xmm2,XMMWORD[32+rdi]
2687	pand xmm0,XMMWORD[384+r10]
2688	movdqa xmm3,XMMWORD[48+rdi]
2689	pand xmm1,XMMWORD[400+r10]
2690	por xmm4,xmm0
2691	pand xmm2,XMMWORD[416+r10]
2692	por xmm5,xmm1
2693	pand xmm3,XMMWORD[432+r10]
2694	por xmm4,xmm2
2695	por xmm5,xmm3
2696	movdqa xmm0,XMMWORD[64+rdi]
2697	movdqa xmm1,XMMWORD[80+rdi]
2698	movdqa xmm2,XMMWORD[96+rdi]
2699	pand xmm0,XMMWORD[448+r10]
2700	movdqa xmm3,XMMWORD[112+rdi]
2701	pand xmm1,XMMWORD[464+r10]
2702	por xmm4,xmm0
2703	pand xmm2,XMMWORD[480+r10]
2704	por xmm5,xmm1
2705	pand xmm3,XMMWORD[496+r10]
2706	por xmm4,xmm2
2707	por xmm5,xmm3
2708	por xmm4,xmm5
2709	pshufd xmm0,xmm4,0x4e
2710	por xmm0,xmm4
2711	lea rdi,[256+rdi]
2712	DB 102,72,15,126,194
2713
2714	mov QWORD[rbx],rbp
2715	lea rbx,[32+rax*1+rbx]
2716	mulx r11,r8,QWORD[rsi]
2717	xor rbp,rbp
2718	mov r9,rdx
2719	mulx r12,r14,QWORD[8+rsi]
2720	adox r8,QWORD[((-32))+rbx]
2721	adcx r11,r14
2722	mulx r13,r15,QWORD[16+rsi]
2723	adox r11,QWORD[((-24))+rbx]
2724	adcx r12,r15
2725	mulx r14,rdx,QWORD[24+rsi]
2726	adox r12,QWORD[((-16))+rbx]
2727	adcx r13,rdx
2728	lea rcx,[rax*1+rcx]
2729	lea rsi,[32+rsi]
2730	adox r13,QWORD[((-8))+rbx]
2731	adcx r14,rbp
2732	adox r14,rbp
2733
2734	mov r15,r8
2735	imul r8,QWORD[((32+8))+rsp]
2736
2737	mov rdx,r8
2738	xor rbp,rbp
2739	mov QWORD[((8+8))+rsp],rdi
2740
2741	mulx r10,rax,QWORD[rcx]
2742	adcx r15,rax
2743	adox r10,r11
2744	mulx r11,rax,QWORD[8+rcx]
2745	adcx r10,rax
2746	adox r11,r12
2747	mulx r12,rax,QWORD[16+rcx]
2748	adcx r11,rax
2749	adox r12,r13
2750	mulx r15,rax,QWORD[24+rcx]
2751	mov rdx,r9
2752	mov rdi,QWORD[((24+8))+rsp]
2753	mov QWORD[((-32))+rbx],r10
2754	adcx r12,rax
2755	mov QWORD[((-24))+rbx],r11
2756	adox r15,rbp
2757	mov QWORD[((-16))+rbx],r12
2758	lea rcx,[32+rcx]
2759	jmp NEAR $L$mulx4x_inner
2760
2761	ALIGN 32
2762	$L$mulx4x_inner:
2763	mulx rax,r10,QWORD[rsi]
2764	adcx r15,rbp
2765	adox r10,r14
2766	mulx r14,r11,QWORD[8+rsi]
2767	adcx r10,QWORD[rbx]
2768	adox r11,rax
2769	mulx rax,r12,QWORD[16+rsi]
2770	adcx r11,QWORD[8+rbx]
2771	adox r12,r14
2772	mulx r14,r13,QWORD[24+rsi]
2773	mov rdx,r8
2774	adcx r12,QWORD[16+rbx]
2775	adox r13,rax
2776	adcx r13,QWORD[24+rbx]
2777	adox r14,rbp
2778	lea rsi,[32+rsi]
2779	lea rbx,[32+rbx]
2780	adcx r14,rbp
2781
2782	adox r10,r15
2783	mulx r15,rax,QWORD[rcx]
2784	adcx r10,rax
2785	adox r11,r15
2786	mulx r15,rax,QWORD[8+rcx]
2787	adcx r11,rax
2788	adox r12,r15
2789	mulx r15,rax,QWORD[16+rcx]
2790	mov QWORD[((-40))+rbx],r10
2791	adcx r12,rax
2792	adox r13,r15
2793	mov QWORD[((-32))+rbx],r11
2794	mulx r15,rax,QWORD[24+rcx]
2795	mov rdx,r9
2796	lea rcx,[32+rcx]
2797	mov QWORD[((-24))+rbx],r12
2798	adcx r13,rax
2799	adox r15,rbp
2800	mov QWORD[((-16))+rbx],r13
2801
2802	dec rdi
2803	jnz NEAR $L$mulx4x_inner
2804
2805	mov rax,QWORD[((0+8))+rsp]
2806	adc r15,rbp
2807	sub rdi,QWORD[rbx]
2808	mov rdi,QWORD[((8+8))+rsp]
2809	mov r10,QWORD[((16+8))+rsp]
2810	adc r14,r15
2811	lea rsi,[rax*1+rsi]
2812	adc rbp,rbp
2813	mov QWORD[((-8))+rbx],r14
2814
2815	cmp rdi,r10
2816	jb NEAR $L$mulx4x_outer
2817
2818	mov r10,QWORD[((-8))+rcx]
2819	mov r8,rbp
2820	mov r12,QWORD[rax*1+rcx]
2821	lea rbp,[rax*1+rcx]
2822	mov rcx,rax
2823	lea rdi,[rax*1+rbx]
2824	xor eax,eax
2825	xor r15,r15
2826	sub r10,r14
2827	adc r15,r15
2828	or r8,r15
2829	sar rcx,3+2
2830	sub rax,r8
2831	mov rdx,QWORD[((56+8))+rsp]
2832	dec r12
2833	mov r13,QWORD[8+rbp]
2834	xor r8,r8
2835	mov r14,QWORD[16+rbp]
2836	mov r15,QWORD[24+rbp]
2837	jmp NEAR $L$sqrx4x_sub_entry
2838
2839
2840
2841	ALIGN 32
2842	bn_powerx5:
2843	mov QWORD[8+rsp],rdi ;WIN64 prologue
2844	mov QWORD[16+rsp],rsi
2845	mov rax,rsp
2846	$L$SEH_begin_bn_powerx5:
2847	mov rdi,rcx
2848	mov rsi,rdx
2849	mov rdx,r8
2850	mov rcx,r9
2851	mov r8,QWORD[40+rsp]
2852	mov r9,QWORD[48+rsp]
2853
2854
2855
2856	mov rax,rsp
2857
2858	$L$powerx5_enter:
2859	push rbx
2860
2861	push rbp
2862
2863	push r12
2864
2865	push r13
2866
2867	push r14
2868
2869	push r15
2870
2871	$L$powerx5_prologue:
2872
2873	shl r9d,3
2874	lea r10,[r9*2+r9]
2875	neg r9
2876	mov r8,QWORD[r8]
2877
2878
2879
2880
2881
2882
2883
2884
2885	lea r11,[((-320))+r9*2+rsp]
2886	mov rbp,rsp
2887	sub r11,rdi
2888	and r11,4095
2889	cmp r10,r11
2890	jb NEAR $L$pwrx_sp_alt
2891	sub rbp,r11
2892	lea rbp,[((-320))+r9*2+rbp]
2893	jmp NEAR $L$pwrx_sp_done
2894
2895	ALIGN 32
2896	$L$pwrx_sp_alt:
2897	lea r10,[((4096-320))+r9*2]
2898	lea rbp,[((-320))+r9*2+rbp]
2899	sub r11,r10
2900	mov r10,0
2901	cmovc r11,r10
2902	sub rbp,r11
2903	$L$pwrx_sp_done:
2904	and rbp,-64
2905	mov r11,rsp
2906	sub r11,rbp
2907	and r11,-4096
2908	lea rsp,[rbp*1+r11]
2909	mov r10,QWORD[rsp]
2910	cmp rsp,rbp
2911	ja NEAR $L$pwrx_page_walk
2912	jmp NEAR $L$pwrx_page_walk_done
2913
2914	$L$pwrx_page_walk:
2915	lea rsp,[((-4096))+rsp]
2916	mov r10,QWORD[rsp]
2917	cmp rsp,rbp
2918	ja NEAR $L$pwrx_page_walk
2919	$L$pwrx_page_walk_done:
2920
2921	mov r10,r9
2922	neg r9
2923
2924
2925
2926
2927
2928
2929
2930
2931
2932
2933
2934
2935	pxor xmm0,xmm0
2936	DB 102,72,15,110,207
2937	DB 102,72,15,110,209
2938	DB 102,73,15,110,218
2939	DB 102,72,15,110,226
2940	mov QWORD[32+rsp],r8
2941	mov QWORD[40+rsp],rax
2942
2943	$L$powerx5_body:
2944
2945	call __bn_sqrx8x_internal
2946	call __bn_postx4x_internal
2947	call __bn_sqrx8x_internal
2948	call __bn_postx4x_internal
2949	call __bn_sqrx8x_internal
2950	call __bn_postx4x_internal
2951	call __bn_sqrx8x_internal
2952	call __bn_postx4x_internal
2953	call __bn_sqrx8x_internal
2954	call __bn_postx4x_internal
2955
2956	mov r9,r10
2957	mov rdi,rsi
2958	DB 102,72,15,126,209
2959	DB 102,72,15,126,226
2960	mov rax,QWORD[40+rsp]
2961
2962	call mulx4x_internal
2963
2964	mov rsi,QWORD[40+rsp]
2965
2966	mov rax,1
2967
2968	mov r15,QWORD[((-48))+rsi]
2969
2970	mov r14,QWORD[((-40))+rsi]
2971
2972	mov r13,QWORD[((-32))+rsi]
2973
2974	mov r12,QWORD[((-24))+rsi]
2975
2976	mov rbp,QWORD[((-16))+rsi]
2977
2978	mov rbx,QWORD[((-8))+rsi]
2979
2980	lea rsp,[rsi]
2981
2982	$L$powerx5_epilogue:
2983	mov rdi,QWORD[8+rsp] ;WIN64 epilogue
2984	mov rsi,QWORD[16+rsp]
2985	DB 0F3h,0C3h ;repret
2986
2987	$L$SEH_end_bn_powerx5:
2988
2989	global bn_sqrx8x_internal
2990
2991
2992	ALIGN 32
2993	bn_sqrx8x_internal:
2994	__bn_sqrx8x_internal:
2995
2996
2997
2998
2999
3000
3001
3002
3003
3004
3005
3006
3007
3008
3009
3010
3011
3012
3013
3014
3015
3016
3017
3018
3019
3020
3021
3022
3023
3024
3025
3026
3027
3028
3029
3030
3031
3032
3033
3034
3035
3036	lea rdi,[((48+8))+rsp]
3037	lea rbp,[r9*1+rsi]
3038	mov QWORD[((0+8))+rsp],r9
3039	mov QWORD[((8+8))+rsp],rbp
3040	jmp NEAR $L$sqr8x_zero_start
3041
3042	ALIGN 32
3043	DB 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
3044	$L$sqrx8x_zero:
3045	DB 0x3e
3046	movdqa XMMWORD[rdi],xmm0
3047	movdqa XMMWORD[16+rdi],xmm0
3048	movdqa XMMWORD[32+rdi],xmm0
3049	movdqa XMMWORD[48+rdi],xmm0
3050	$L$sqr8x_zero_start:
3051	movdqa XMMWORD[64+rdi],xmm0
3052	movdqa XMMWORD[80+rdi],xmm0
3053	movdqa XMMWORD[96+rdi],xmm0
3054	movdqa XMMWORD[112+rdi],xmm0
3055	lea rdi,[128+rdi]
3056	sub r9,64
3057	jnz NEAR $L$sqrx8x_zero
3058
3059	mov rdx,QWORD[rsi]
3060
3061	xor r10,r10
3062	xor r11,r11
3063	xor r12,r12
3064	xor r13,r13
3065	xor r14,r14
3066	xor r15,r15
3067	lea rdi,[((48+8))+rsp]
3068	xor rbp,rbp
3069	jmp NEAR $L$sqrx8x_outer_loop
3070
3071	ALIGN 32
3072	$L$sqrx8x_outer_loop:
3073	mulx rax,r8,QWORD[8+rsi]
3074	adcx r8,r9
3075	adox r10,rax
3076	mulx rax,r9,QWORD[16+rsi]
3077	adcx r9,r10
3078	adox r11,rax
3079	DB 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00
3080	adcx r10,r11
3081	adox r12,rax
3082	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00
3083	adcx r11,r12
3084	adox r13,rax
3085	mulx rax,r12,QWORD[40+rsi]
3086	adcx r12,r13
3087	adox r14,rax
3088	mulx rax,r13,QWORD[48+rsi]
3089	adcx r13,r14
3090	adox rax,r15
3091	mulx r15,r14,QWORD[56+rsi]
3092	mov rdx,QWORD[8+rsi]
3093	adcx r14,rax
3094	adox r15,rbp
3095	adc r15,QWORD[64+rdi]
3096	mov QWORD[8+rdi],r8
3097	mov QWORD[16+rdi],r9
3098	sbb rcx,rcx
3099	xor rbp,rbp
3100
3101
3102	mulx rbx,r8,QWORD[16+rsi]
3103	mulx rax,r9,QWORD[24+rsi]
3104	adcx r8,r10
3105	adox r9,rbx
3106	mulx rbx,r10,QWORD[32+rsi]
3107	adcx r9,r11
3108	adox r10,rax
3109	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00
3110	adcx r10,r12
3111	adox r11,rbx
3112	DB 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00
3113	adcx r11,r13
3114	adox r12,r14
3115	DB 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00
3116	mov rdx,QWORD[16+rsi]
3117	adcx r12,rax
3118	adox r13,rbx
3119	adcx r13,r15
3120	adox r14,rbp
3121	adcx r14,rbp
3122
3123	mov QWORD[24+rdi],r8
3124	mov QWORD[32+rdi],r9
3125
3126	mulx rbx,r8,QWORD[24+rsi]
3127	mulx rax,r9,QWORD[32+rsi]
3128	adcx r8,r10
3129	adox r9,rbx
3130	mulx rbx,r10,QWORD[40+rsi]
3131	adcx r9,r11
3132	adox r10,rax
3133	DB 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00
3134	adcx r10,r12
3135	adox r11,r13
3136	DB 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00
3137	DB 0x3e
3138	mov rdx,QWORD[24+rsi]
3139	adcx r11,rbx
3140	adox r12,rax
3141	adcx r12,r14
3142	mov QWORD[40+rdi],r8
3143	mov QWORD[48+rdi],r9
3144	mulx rax,r8,QWORD[32+rsi]
3145	adox r13,rbp
3146	adcx r13,rbp
3147
3148	mulx rbx,r9,QWORD[40+rsi]
3149	adcx r8,r10
3150	adox r9,rax
3151	mulx rax,r10,QWORD[48+rsi]
3152	adcx r9,r11
3153	adox r10,r12
3154	mulx r12,r11,QWORD[56+rsi]
3155	mov rdx,QWORD[32+rsi]
3156	mov r14,QWORD[40+rsi]
3157	adcx r10,rbx
3158	adox r11,rax
3159	mov r15,QWORD[48+rsi]
3160	adcx r11,r13
3161	adox r12,rbp
3162	adcx r12,rbp
3163
3164	mov QWORD[56+rdi],r8
3165	mov QWORD[64+rdi],r9
3166
3167	mulx rax,r9,r14
3168	mov r8,QWORD[56+rsi]
3169	adcx r9,r10
3170	mulx rbx,r10,r15
3171	adox r10,rax
3172	adcx r10,r11
3173	mulx rax,r11,r8
3174	mov rdx,r14
3175	adox r11,rbx
3176	adcx r11,r12
3177
3178	adcx rax,rbp
3179
3180	mulx rbx,r14,r15
3181	mulx r13,r12,r8
3182	mov rdx,r15
3183	lea rsi,[64+rsi]
3184	adcx r11,r14
3185	adox r12,rbx
3186	adcx r12,rax
3187	adox r13,rbp
3188
3189	DB 0x67,0x67
3190	mulx r14,r8,r8
3191	adcx r13,r8
3192	adcx r14,rbp
3193
3194	cmp rsi,QWORD[((8+8))+rsp]
3195	je NEAR $L$sqrx8x_outer_break
3196
3197	neg rcx
3198	mov rcx,-8
3199	mov r15,rbp
3200	mov r8,QWORD[64+rdi]
3201	adcx r9,QWORD[72+rdi]
3202	adcx r10,QWORD[80+rdi]
3203	adcx r11,QWORD[88+rdi]
3204	adc r12,QWORD[96+rdi]
3205	adc r13,QWORD[104+rdi]
3206	adc r14,QWORD[112+rdi]
3207	adc r15,QWORD[120+rdi]
3208	lea rbp,[rsi]
3209	lea rdi,[128+rdi]
3210	sbb rax,rax
3211
3212	mov rdx,QWORD[((-64))+rsi]
3213	mov QWORD[((16+8))+rsp],rax
3214	mov QWORD[((24+8))+rsp],rdi
3215
3216
3217	xor eax,eax
3218	jmp NEAR $L$sqrx8x_loop
3219
3220	ALIGN 32
3221	$L$sqrx8x_loop:
3222	mov rbx,r8
3223	mulx r8,rax,QWORD[rbp]
3224	adcx rbx,rax
3225	adox r8,r9
3226
3227	mulx r9,rax,QWORD[8+rbp]
3228	adcx r8,rax
3229	adox r9,r10
3230
3231	mulx r10,rax,QWORD[16+rbp]
3232	adcx r9,rax
3233	adox r10,r11
3234
3235	mulx r11,rax,QWORD[24+rbp]
3236	adcx r10,rax
3237	adox r11,r12
3238
3239	DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
3240	adcx r11,rax
3241	adox r12,r13
3242
3243	mulx r13,rax,QWORD[40+rbp]
3244	adcx r12,rax
3245	adox r13,r14
3246
3247	mulx r14,rax,QWORD[48+rbp]
3248	mov QWORD[rcx*8+rdi],rbx
3249	mov ebx,0
3250	adcx r13,rax
3251	adox r14,r15
3252
3253	DB 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00
3254	mov rdx,QWORD[8+rcx*8+rsi]
3255	adcx r14,rax
3256	adox r15,rbx
3257	adcx r15,rbx
3258
3259	DB 0x67
3260	inc rcx
3261	jnz NEAR $L$sqrx8x_loop
3262
3263	lea rbp,[64+rbp]
3264	mov rcx,-8
3265	cmp rbp,QWORD[((8+8))+rsp]
3266	je NEAR $L$sqrx8x_break
3267
3268	sub rbx,QWORD[((16+8))+rsp]
3269	DB 0x66
3270	mov rdx,QWORD[((-64))+rsi]
3271	adcx r8,QWORD[rdi]
3272	adcx r9,QWORD[8+rdi]
3273	adc r10,QWORD[16+rdi]
3274	adc r11,QWORD[24+rdi]
3275	adc r12,QWORD[32+rdi]
3276	adc r13,QWORD[40+rdi]
3277	adc r14,QWORD[48+rdi]
3278	adc r15,QWORD[56+rdi]
3279	lea rdi,[64+rdi]
3280	DB 0x67
3281	sbb rax,rax
3282	xor ebx,ebx
3283	mov QWORD[((16+8))+rsp],rax
3284	jmp NEAR $L$sqrx8x_loop
3285
3286	ALIGN 32
3287	$L$sqrx8x_break:
3288	xor rbp,rbp
3289	sub rbx,QWORD[((16+8))+rsp]
3290	adcx r8,rbp
3291	mov rcx,QWORD[((24+8))+rsp]
3292	adcx r9,rbp
3293	mov rdx,QWORD[rsi]
3294	adc r10,0
3295	mov QWORD[rdi],r8
3296	adc r11,0
3297	adc r12,0
3298	adc r13,0
3299	adc r14,0
3300	adc r15,0
3301	cmp rdi,rcx
3302	je NEAR $L$sqrx8x_outer_loop
3303
3304	mov QWORD[8+rdi],r9
3305	mov r9,QWORD[8+rcx]
3306	mov QWORD[16+rdi],r10
3307	mov r10,QWORD[16+rcx]
3308	mov QWORD[24+rdi],r11
3309	mov r11,QWORD[24+rcx]
3310	mov QWORD[32+rdi],r12
3311	mov r12,QWORD[32+rcx]
3312	mov QWORD[40+rdi],r13
3313	mov r13,QWORD[40+rcx]
3314	mov QWORD[48+rdi],r14
3315	mov r14,QWORD[48+rcx]
3316	mov QWORD[56+rdi],r15
3317	mov r15,QWORD[56+rcx]
3318	mov rdi,rcx
3319	jmp NEAR $L$sqrx8x_outer_loop
3320
3321	ALIGN 32
3322	$L$sqrx8x_outer_break:
3323	mov QWORD[72+rdi],r9
3324	DB 102,72,15,126,217
3325	mov QWORD[80+rdi],r10
3326	mov QWORD[88+rdi],r11
3327	mov QWORD[96+rdi],r12
3328	mov QWORD[104+rdi],r13
3329	mov QWORD[112+rdi],r14
3330	lea rdi,[((48+8))+rsp]
3331	mov rdx,QWORD[rcx*1+rsi]
3332
3333	mov r11,QWORD[8+rdi]
3334	xor r10,r10
3335	mov r9,QWORD[((0+8))+rsp]
3336	adox r11,r11
3337	mov r12,QWORD[16+rdi]
3338	mov r13,QWORD[24+rdi]
3339
3340
3341	ALIGN 32
3342	$L$sqrx4x_shift_n_add:
3343	mulx rbx,rax,rdx
3344	adox r12,r12
3345	adcx rax,r10
3346	DB 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00
3347	DB 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00
3348	adox r13,r13
3349	adcx rbx,r11
3350	mov r11,QWORD[40+rdi]
3351	mov QWORD[rdi],rax
3352	mov QWORD[8+rdi],rbx
3353
3354	mulx rbx,rax,rdx
3355	adox r10,r10
3356	adcx rax,r12
3357	mov rdx,QWORD[16+rcx*1+rsi]
3358	mov r12,QWORD[48+rdi]
3359	adox r11,r11
3360	adcx rbx,r13
3361	mov r13,QWORD[56+rdi]
3362	mov QWORD[16+rdi],rax
3363	mov QWORD[24+rdi],rbx
3364
3365	mulx rbx,rax,rdx
3366	adox r12,r12
3367	adcx rax,r10
3368	mov rdx,QWORD[24+rcx*1+rsi]
3369	lea rcx,[32+rcx]
3370	mov r10,QWORD[64+rdi]
3371	adox r13,r13
3372	adcx rbx,r11
3373	mov r11,QWORD[72+rdi]
3374	mov QWORD[32+rdi],rax
3375	mov QWORD[40+rdi],rbx
3376
3377	mulx rbx,rax,rdx
3378	adox r10,r10
3379	adcx rax,r12
3380	jrcxz $L$sqrx4x_shift_n_add_break
3381	DB 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00
3382	adox r11,r11
3383	adcx rbx,r13
3384	mov r12,QWORD[80+rdi]
3385	mov r13,QWORD[88+rdi]
3386	mov QWORD[48+rdi],rax
3387	mov QWORD[56+rdi],rbx
3388	lea rdi,[64+rdi]
3389	nop
3390	jmp NEAR $L$sqrx4x_shift_n_add
3391
3392	ALIGN 32
3393	$L$sqrx4x_shift_n_add_break:
3394	adcx rbx,r13
3395	mov QWORD[48+rdi],rax
3396	mov QWORD[56+rdi],rbx
3397	lea rdi,[64+rdi]
3398	DB 102,72,15,126,213
3399	__bn_sqrx8x_reduction:
3400	xor eax,eax
3401	mov rbx,QWORD[((32+8))+rsp]
3402	mov rdx,QWORD[((48+8))+rsp]
3403	lea rcx,[((-64))+r9*1+rbp]
3404
3405	mov QWORD[((0+8))+rsp],rcx
3406	mov QWORD[((8+8))+rsp],rdi
3407
3408	lea rdi,[((48+8))+rsp]
3409	jmp NEAR $L$sqrx8x_reduction_loop
3410
3411	ALIGN 32
3412	$L$sqrx8x_reduction_loop:
3413	mov r9,QWORD[8+rdi]
3414	mov r10,QWORD[16+rdi]
3415	mov r11,QWORD[24+rdi]
3416	mov r12,QWORD[32+rdi]
3417	mov r8,rdx
3418	imul rdx,rbx
3419	mov r13,QWORD[40+rdi]
3420	mov r14,QWORD[48+rdi]
3421	mov r15,QWORD[56+rdi]
3422	mov QWORD[((24+8))+rsp],rax
3423
3424	lea rdi,[64+rdi]
3425	xor rsi,rsi
3426	mov rcx,-8
3427	jmp NEAR $L$sqrx8x_reduce
3428
3429	ALIGN 32
3430	$L$sqrx8x_reduce:
3431	mov rbx,r8
3432	mulx r8,rax,QWORD[rbp]
3433	adcx rax,rbx
3434	adox r8,r9
3435
3436	mulx r9,rbx,QWORD[8+rbp]
3437	adcx r8,rbx
3438	adox r9,r10
3439
3440	mulx r10,rbx,QWORD[16+rbp]
3441	adcx r9,rbx
3442	adox r10,r11
3443
3444	mulx r11,rbx,QWORD[24+rbp]
3445	adcx r10,rbx
3446	adox r11,r12
3447
3448	DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00
3449	mov rax,rdx
3450	mov rdx,r8
3451	adcx r11,rbx
3452	adox r12,r13
3453
3454	mulx rdx,rbx,QWORD[((32+8))+rsp]
3455	mov rdx,rax
3456	mov QWORD[((64+48+8))+rcx*8+rsp],rax
3457
3458	mulx r13,rax,QWORD[40+rbp]
3459	adcx r12,rax
3460	adox r13,r14
3461
3462	mulx r14,rax,QWORD[48+rbp]
3463	adcx r13,rax
3464	adox r14,r15
3465
3466	mulx r15,rax,QWORD[56+rbp]
3467	mov rdx,rbx
3468	adcx r14,rax
3469	adox r15,rsi
3470	adcx r15,rsi
3471
3472	DB 0x67,0x67,0x67
3473	inc rcx
3474	jnz NEAR $L$sqrx8x_reduce
3475
3476	mov rax,rsi
3477	cmp rbp,QWORD[((0+8))+rsp]
3478	jae NEAR $L$sqrx8x_no_tail
3479
3480	mov rdx,QWORD[((48+8))+rsp]
3481	add r8,QWORD[rdi]
3482	lea rbp,[64+rbp]
3483	mov rcx,-8
3484	adcx r9,QWORD[8+rdi]
3485	adcx r10,QWORD[16+rdi]
3486	adc r11,QWORD[24+rdi]
3487	adc r12,QWORD[32+rdi]
3488	adc r13,QWORD[40+rdi]
3489	adc r14,QWORD[48+rdi]
3490	adc r15,QWORD[56+rdi]
3491	lea rdi,[64+rdi]
3492	sbb rax,rax
3493
3494	xor rsi,rsi
3495	mov QWORD[((16+8))+rsp],rax
3496	jmp NEAR $L$sqrx8x_tail
3497
3498	ALIGN 32
3499	$L$sqrx8x_tail:
3500	mov rbx,r8
3501	mulx r8,rax,QWORD[rbp]
3502	adcx rbx,rax
3503	adox r8,r9
3504
3505	mulx r9,rax,QWORD[8+rbp]
3506	adcx r8,rax
3507	adox r9,r10
3508
3509	mulx r10,rax,QWORD[16+rbp]
3510	adcx r9,rax
3511	adox r10,r11
3512
3513	mulx r11,rax,QWORD[24+rbp]
3514	adcx r10,rax
3515	adox r11,r12
3516
3517	DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00
3518	adcx r11,rax
3519	adox r12,r13
3520
3521	mulx r13,rax,QWORD[40+rbp]
3522	adcx r12,rax
3523	adox r13,r14
3524
3525	mulx r14,rax,QWORD[48+rbp]
3526	adcx r13,rax
3527	adox r14,r15
3528
3529	mulx r15,rax,QWORD[56+rbp]
3530	mov rdx,QWORD[((72+48+8))+rcx*8+rsp]
3531	adcx r14,rax
3532	adox r15,rsi
3533	mov QWORD[rcx*8+rdi],rbx
3534	mov rbx,r8
3535	adcx r15,rsi
3536
3537	inc rcx
3538	jnz NEAR $L$sqrx8x_tail
3539
3540	cmp rbp,QWORD[((0+8))+rsp]
3541	jae NEAR $L$sqrx8x_tail_done
3542
3543	sub rsi,QWORD[((16+8))+rsp]
3544	mov rdx,QWORD[((48+8))+rsp]
3545	lea rbp,[64+rbp]
3546	adc r8,QWORD[rdi]
3547	adc r9,QWORD[8+rdi]
3548	adc r10,QWORD[16+rdi]
3549	adc r11,QWORD[24+rdi]
3550	adc r12,QWORD[32+rdi]
3551	adc r13,QWORD[40+rdi]
3552	adc r14,QWORD[48+rdi]
3553	adc r15,QWORD[56+rdi]
3554	lea rdi,[64+rdi]
3555	sbb rax,rax
3556	sub rcx,8
3557
3558	xor rsi,rsi
3559	mov QWORD[((16+8))+rsp],rax
3560	jmp NEAR $L$sqrx8x_tail
3561
3562	ALIGN 32
3563	$L$sqrx8x_tail_done:
3564	xor rax,rax
3565	add r8,QWORD[((24+8))+rsp]
3566	adc r9,0
3567	adc r10,0
3568	adc r11,0
3569	adc r12,0
3570	adc r13,0
3571	adc r14,0
3572	adc r15,0
3573	adc rax,0
3574
3575	sub rsi,QWORD[((16+8))+rsp]
3576	$L$sqrx8x_no_tail:
3577	adc r8,QWORD[rdi]
3578	DB 102,72,15,126,217
3579	adc r9,QWORD[8+rdi]
3580	mov rsi,QWORD[56+rbp]
3581	DB 102,72,15,126,213
3582	adc r10,QWORD[16+rdi]
3583	adc r11,QWORD[24+rdi]
3584	adc r12,QWORD[32+rdi]
3585	adc r13,QWORD[40+rdi]
3586	adc r14,QWORD[48+rdi]
3587	adc r15,QWORD[56+rdi]
3588	adc rax,0
3589
3590	mov rbx,QWORD[((32+8))+rsp]
3591	mov rdx,QWORD[64+rcx*1+rdi]
3592
3593	mov QWORD[rdi],r8
3594	lea r8,[64+rdi]
3595	mov QWORD[8+rdi],r9
3596	mov QWORD[16+rdi],r10
3597	mov QWORD[24+rdi],r11
3598	mov QWORD[32+rdi],r12
3599	mov QWORD[40+rdi],r13
3600	mov QWORD[48+rdi],r14
3601	mov QWORD[56+rdi],r15
3602
3603	lea rdi,[64+rcx*1+rdi]
3604	cmp r8,QWORD[((8+8))+rsp]
3605	jb NEAR $L$sqrx8x_reduction_loop
3606	DB 0F3h,0C3h ;repret
3607
3608
3609	ALIGN 32
3610	__bn_postx4x_internal:
3611
3612	mov r12,QWORD[rbp]
3613	mov r10,rcx
3614	mov r9,rcx
3615	neg rax
3616	sar rcx,3+2
3617
3618	DB 102,72,15,126,202
3619	DB 102,72,15,126,206
3620	dec r12
3621	mov r13,QWORD[8+rbp]
3622	xor r8,r8
3623	mov r14,QWORD[16+rbp]
3624	mov r15,QWORD[24+rbp]
3625	jmp NEAR $L$sqrx4x_sub_entry
3626
3627	ALIGN 16
3628	$L$sqrx4x_sub:
3629	mov r12,QWORD[rbp]
3630	mov r13,QWORD[8+rbp]
3631	mov r14,QWORD[16+rbp]
3632	mov r15,QWORD[24+rbp]
3633	$L$sqrx4x_sub_entry:
3634	andn r12,r12,rax
3635	lea rbp,[32+rbp]
3636	andn r13,r13,rax
3637	andn r14,r14,rax
3638	andn r15,r15,rax
3639
3640	neg r8
3641	adc r12,QWORD[rdi]
3642	adc r13,QWORD[8+rdi]
3643	adc r14,QWORD[16+rdi]
3644	adc r15,QWORD[24+rdi]
3645	mov QWORD[rdx],r12
3646	lea rdi,[32+rdi]
3647	mov QWORD[8+rdx],r13
3648	sbb r8,r8
3649	mov QWORD[16+rdx],r14
3650	mov QWORD[24+rdx],r15
3651	lea rdx,[32+rdx]
3652
3653	inc rcx
3654	jnz NEAR $L$sqrx4x_sub
3655
3656	neg r9
3657
3658	DB 0F3h,0C3h ;repret
3659
3660
3661	global bn_get_bits5
3662
3663	ALIGN 16
3664	bn_get_bits5:
3665
3666	lea r10,[rcx]
3667	lea r11,[1+rcx]
3668	mov ecx,edx
3669	shr edx,4
3670	and ecx,15
3671	lea eax,[((-8))+rcx]
3672	cmp ecx,11
3673	cmova r10,r11
3674	cmova ecx,eax
3675	movzx eax,WORD[rdx*2+r10]
3676	shr eax,cl
3677	and eax,31
3678	DB 0F3h,0C3h ;repret
3679
3680
3681
3682	global bn_scatter5
3683
3684	ALIGN 16
3685	bn_scatter5:
3686
3687	cmp edx,0
3688	jz NEAR $L$scatter_epilogue
3689	lea r8,[r9*8+r8]
3690	$L$scatter:
3691	mov rax,QWORD[rcx]
3692	lea rcx,[8+rcx]
3693	mov QWORD[r8],rax
3694	lea r8,[256+r8]
3695	sub edx,1
3696	jnz NEAR $L$scatter
3697	$L$scatter_epilogue:
3698	DB 0F3h,0C3h ;repret
3699
3700
3701
3702	global bn_gather5
3703
3704	ALIGN 32
3705	bn_gather5:
3706	$L$SEH_begin_bn_gather5:
3707
3708
3709	DB 0x4c,0x8d,0x14,0x24
3710	DB 0x48,0x81,0xec,0x08,0x01,0x00,0x00
3711	lea rax,[$L$inc]
3712	and rsp,-16
3713
3714	movd xmm5,r9d
3715	movdqa xmm0,XMMWORD[rax]
3716	movdqa xmm1,XMMWORD[16+rax]
3717	lea r11,[128+r8]
3718	lea rax,[128+rsp]
3719
3720	pshufd xmm5,xmm5,0
3721	movdqa xmm4,xmm1
3722	movdqa xmm2,xmm1
3723	paddd xmm1,xmm0
3724	pcmpeqd xmm0,xmm5
3725	movdqa xmm3,xmm4
3726
3727	paddd xmm2,xmm1
3728	pcmpeqd xmm1,xmm5
3729	movdqa XMMWORD[(-128)+rax],xmm0
3730	movdqa xmm0,xmm4
3731
3732	paddd xmm3,xmm2
3733	pcmpeqd xmm2,xmm5
3734	movdqa XMMWORD[(-112)+rax],xmm1
3735	movdqa xmm1,xmm4
3736
3737	paddd xmm0,xmm3
3738	pcmpeqd xmm3,xmm5
3739	movdqa XMMWORD[(-96)+rax],xmm2
3740	movdqa xmm2,xmm4
3741	paddd xmm1,xmm0
3742	pcmpeqd xmm0,xmm5
3743	movdqa XMMWORD[(-80)+rax],xmm3
3744	movdqa xmm3,xmm4
3745
3746	paddd xmm2,xmm1
3747	pcmpeqd xmm1,xmm5
3748	movdqa XMMWORD[(-64)+rax],xmm0
3749	movdqa xmm0,xmm4
3750
3751	paddd xmm3,xmm2
3752	pcmpeqd xmm2,xmm5
3753	movdqa XMMWORD[(-48)+rax],xmm1
3754	movdqa xmm1,xmm4
3755
3756	paddd xmm0,xmm3
3757	pcmpeqd xmm3,xmm5
3758	movdqa XMMWORD[(-32)+rax],xmm2
3759	movdqa xmm2,xmm4
3760	paddd xmm1,xmm0
3761	pcmpeqd xmm0,xmm5
3762	movdqa XMMWORD[(-16)+rax],xmm3
3763	movdqa xmm3,xmm4
3764
3765	paddd xmm2,xmm1
3766	pcmpeqd xmm1,xmm5
3767	movdqa XMMWORD[rax],xmm0
3768	movdqa xmm0,xmm4
3769
3770	paddd xmm3,xmm2
3771	pcmpeqd xmm2,xmm5
3772	movdqa XMMWORD[16+rax],xmm1
3773	movdqa xmm1,xmm4
3774
3775	paddd xmm0,xmm3
3776	pcmpeqd xmm3,xmm5
3777	movdqa XMMWORD[32+rax],xmm2
3778	movdqa xmm2,xmm4
3779	paddd xmm1,xmm0
3780	pcmpeqd xmm0,xmm5
3781	movdqa XMMWORD[48+rax],xmm3
3782	movdqa xmm3,xmm4
3783
3784	paddd xmm2,xmm1
3785	pcmpeqd xmm1,xmm5
3786	movdqa XMMWORD[64+rax],xmm0
3787	movdqa xmm0,xmm4
3788
3789	paddd xmm3,xmm2
3790	pcmpeqd xmm2,xmm5
3791	movdqa XMMWORD[80+rax],xmm1
3792	movdqa xmm1,xmm4
3793
3794	paddd xmm0,xmm3
3795	pcmpeqd xmm3,xmm5
3796	movdqa XMMWORD[96+rax],xmm2
3797	movdqa xmm2,xmm4
3798	movdqa XMMWORD[112+rax],xmm3
3799	jmp NEAR $L$gather
3800
3801	ALIGN 32
3802	$L$gather:
3803	pxor xmm4,xmm4
3804	pxor xmm5,xmm5
3805	movdqa xmm0,XMMWORD[((-128))+r11]
3806	movdqa xmm1,XMMWORD[((-112))+r11]
3807	movdqa xmm2,XMMWORD[((-96))+r11]
3808	pand xmm0,XMMWORD[((-128))+rax]
3809	movdqa xmm3,XMMWORD[((-80))+r11]
3810	pand xmm1,XMMWORD[((-112))+rax]
3811	por xmm4,xmm0
3812	pand xmm2,XMMWORD[((-96))+rax]
3813	por xmm5,xmm1
3814	pand xmm3,XMMWORD[((-80))+rax]
3815	por xmm4,xmm2
3816	por xmm5,xmm3
3817	movdqa xmm0,XMMWORD[((-64))+r11]
3818	movdqa xmm1,XMMWORD[((-48))+r11]
3819	movdqa xmm2,XMMWORD[((-32))+r11]
3820	pand xmm0,XMMWORD[((-64))+rax]
3821	movdqa xmm3,XMMWORD[((-16))+r11]
3822	pand xmm1,XMMWORD[((-48))+rax]
3823	por xmm4,xmm0
3824	pand xmm2,XMMWORD[((-32))+rax]
3825	por xmm5,xmm1
3826	pand xmm3,XMMWORD[((-16))+rax]
3827	por xmm4,xmm2
3828	por xmm5,xmm3
3829	movdqa xmm0,XMMWORD[r11]
3830	movdqa xmm1,XMMWORD[16+r11]
3831	movdqa xmm2,XMMWORD[32+r11]
3832	pand xmm0,XMMWORD[rax]
3833	movdqa xmm3,XMMWORD[48+r11]
3834	pand xmm1,XMMWORD[16+rax]
3835	por xmm4,xmm0
3836	pand xmm2,XMMWORD[32+rax]
3837	por xmm5,xmm1
3838	pand xmm3,XMMWORD[48+rax]
3839	por xmm4,xmm2
3840	por xmm5,xmm3
3841	movdqa xmm0,XMMWORD[64+r11]
3842	movdqa xmm1,XMMWORD[80+r11]
3843	movdqa xmm2,XMMWORD[96+r11]
3844	pand xmm0,XMMWORD[64+rax]
3845	movdqa xmm3,XMMWORD[112+r11]
3846	pand xmm1,XMMWORD[80+rax]
3847	por xmm4,xmm0
3848	pand xmm2,XMMWORD[96+rax]
3849	por xmm5,xmm1
3850	pand xmm3,XMMWORD[112+rax]
3851	por xmm4,xmm2
3852	por xmm5,xmm3
3853	por xmm4,xmm5
3854	lea r11,[256+r11]
3855	pshufd xmm0,xmm4,0x4e
3856	por xmm0,xmm4
3857	movq QWORD[rcx],xmm0
3858	lea rcx,[8+rcx]
3859	sub edx,1
3860	jnz NEAR $L$gather
3861
3862	lea rsp,[r10]
3863	DB 0F3h,0C3h ;repret
3864	$L$SEH_end_bn_gather5:
3865
3866
3867	ALIGN 64
3868	$L$inc:
3869	DD 0,0,1,1
3870	DD 2,2,2,2
3871	DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105
3872	DB 112,108,105,99,97,116,105,111,110,32,119,105,116,104,32,115
3873	DB 99,97,116,116,101,114,47,103,97,116,104,101,114,32,102,111
3874	DB 114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79
3875	DB 71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111
3876	DB 112,101,110,115,115,108,46,111,114,103,62,0
3877	EXTERN __imp_RtlVirtualUnwind
3878
3879	ALIGN 16
3880	mul_handler:
3881	push rsi
3882	push rdi
3883	push rbx
3884	push rbp
3885	push r12
3886	push r13
3887	push r14
3888	push r15
3889	pushfq
3890	sub rsp,64
3891
3892	mov rax,QWORD[120+r8]
3893	mov rbx,QWORD[248+r8]
3894
3895	mov rsi,QWORD[8+r9]
3896	mov r11,QWORD[56+r9]
3897
3898	mov r10d,DWORD[r11]
3899	lea r10,[r10*1+rsi]
3900	cmp rbx,r10
3901	jb NEAR $L$common_seh_tail
3902
3903	mov r10d,DWORD[4+r11]
3904	lea r10,[r10*1+rsi]
3905	cmp rbx,r10
3906	jb NEAR $L$common_pop_regs
3907
3908	mov rax,QWORD[152+r8]
3909
3910	mov r10d,DWORD[8+r11]
3911	lea r10,[r10*1+rsi]
3912	cmp rbx,r10
3913	jae NEAR $L$common_seh_tail
3914
3915	lea r10,[$L$mul_epilogue]
3916	cmp rbx,r10
3917	ja NEAR $L$body_40
3918
3919	mov r10,QWORD[192+r8]
3920	mov rax,QWORD[8+r10*8+rax]
3921
3922	jmp NEAR $L$common_pop_regs
3923
3924	$L$body_40:
3925	mov rax,QWORD[40+rax]
3926	$L$common_pop_regs:
3927	mov rbx,QWORD[((-8))+rax]
3928	mov rbp,QWORD[((-16))+rax]
3929	mov r12,QWORD[((-24))+rax]
3930	mov r13,QWORD[((-32))+rax]
3931	mov r14,QWORD[((-40))+rax]
3932	mov r15,QWORD[((-48))+rax]
3933	mov QWORD[144+r8],rbx
3934	mov QWORD[160+r8],rbp
3935	mov QWORD[216+r8],r12
3936	mov QWORD[224+r8],r13
3937	mov QWORD[232+r8],r14
3938	mov QWORD[240+r8],r15
3939
3940	$L$common_seh_tail:
3941	mov rdi,QWORD[8+rax]
3942	mov rsi,QWORD[16+rax]
3943	mov QWORD[152+r8],rax
3944	mov QWORD[168+r8],rsi
3945	mov QWORD[176+r8],rdi
3946
3947	mov rdi,QWORD[40+r9]
3948	mov rsi,r8
3949	mov ecx,154
3950	DD 0xa548f3fc
3951
3952	mov rsi,r9
3953	xor rcx,rcx
3954	mov rdx,QWORD[8+rsi]
3955	mov r8,QWORD[rsi]
3956	mov r9,QWORD[16+rsi]
3957	mov r10,QWORD[40+rsi]
3958	lea r11,[56+rsi]
3959	lea r12,[24+rsi]
3960	mov QWORD[32+rsp],r10
3961	mov QWORD[40+rsp],r11
3962	mov QWORD[48+rsp],r12
3963	mov QWORD[56+rsp],rcx
3964	call QWORD[__imp_RtlVirtualUnwind]
3965
3966	mov eax,1
3967	add rsp,64
3968	popfq
3969	pop r15
3970	pop r14
3971	pop r13
3972	pop r12
3973	pop rbp
3974	pop rbx
3975	pop rdi
3976	pop rsi
3977	DB 0F3h,0C3h ;repret
3978
3979
3980	section .pdata rdata align=4
3981	ALIGN 4
3982	DD $L$SEH_begin_bn_mul_mont_gather5 wrt ..imagebase
3983	DD $L$SEH_end_bn_mul_mont_gather5 wrt ..imagebase
3984	DD $L$SEH_info_bn_mul_mont_gather5 wrt ..imagebase
3985
3986	DD $L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase
3987	DD $L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase
3988	DD $L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase
3989
3990	DD $L$SEH_begin_bn_power5 wrt ..imagebase
3991	DD $L$SEH_end_bn_power5 wrt ..imagebase
3992	DD $L$SEH_info_bn_power5 wrt ..imagebase
3993
3994	DD $L$SEH_begin_bn_from_mont8x wrt ..imagebase
3995	DD $L$SEH_end_bn_from_mont8x wrt ..imagebase
3996	DD $L$SEH_info_bn_from_mont8x wrt ..imagebase
3997	DD $L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase
3998	DD $L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase
3999	DD $L$SEH_info_bn_mulx4x_mont_gather5 wrt ..imagebase
4000
4001	DD $L$SEH_begin_bn_powerx5 wrt ..imagebase
4002	DD $L$SEH_end_bn_powerx5 wrt ..imagebase
4003	DD $L$SEH_info_bn_powerx5 wrt ..imagebase
4004	DD $L$SEH_begin_bn_gather5 wrt ..imagebase
4005	DD $L$SEH_end_bn_gather5 wrt ..imagebase
4006	DD $L$SEH_info_bn_gather5 wrt ..imagebase
4007
4008	section .xdata rdata align=8
4009	ALIGN 8
4010	$L$SEH_info_bn_mul_mont_gather5:
4011	DB 9,0,0,0
4012	DD mul_handler wrt ..imagebase
4013	DD $L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
4014	ALIGN 8
4015	$L$SEH_info_bn_mul4x_mont_gather5:
4016	DB 9,0,0,0
4017	DD mul_handler wrt ..imagebase
4018	DD $L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase
4019	ALIGN 8
4020	$L$SEH_info_bn_power5:
4021	DB 9,0,0,0
4022	DD mul_handler wrt ..imagebase
4023	DD $L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase
4024	ALIGN 8
4025	$L$SEH_info_bn_from_mont8x:
4026	DB 9,0,0,0
4027	DD mul_handler wrt ..imagebase
4028	DD $L$from_prologue wrt ..imagebase,$L$from_body wrt ..imagebase,$L$from_epilogue wrt ..imagebase
4029	ALIGN 8
4030	$L$SEH_info_bn_mulx4x_mont_gather5:
4031	DB 9,0,0,0
4032	DD mul_handler wrt ..imagebase
4033	DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase
4034	ALIGN 8
4035	$L$SEH_info_bn_powerx5:
4036	DB 9,0,0,0
4037	DD mul_handler wrt ..imagebase
4038	DD $L$powerx5_prologue wrt ..imagebase,$L$powerx5_body wrt ..imagebase,$L$powerx5_epilogue wrt ..imagebase
4039	ALIGN 8
4040	$L$SEH_info_bn_gather5:
4041	DB 0x01,0x0b,0x03,0x0a
4042	DB 0x0b,0x01,0x21,0x00
4043	DB 0x04,0xa3,0x00,0x00
4044	ALIGN 8

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

Download in other formats:

Original Format