xref: /freebsd/crypto/openssl/crypto/bn/asm/x86_64-mont5.pl (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41) 
1#! /usr/bin/env perl
2# Copyright 2011-2022 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16
17# August 2011.
18#
19# Companion to x86_64-mont.pl that optimizes cache-timing attack
20# countermeasures. The subroutines are produced by replacing bp[i]
21# references in their x86_64-mont.pl counterparts with cache-neutral
22# references to powers table computed in BN_mod_exp_mont_consttime.
23# In addition subroutine that scatters elements of the powers table
24# is implemented, so that scatter-/gathering can be tuned without
25# bn_exp.c modifications.
26
27# August 2013.
28#
29# Add MULX/AD*X code paths and additional interfaces to optimize for
30# branch prediction unit. For input lengths that are multiples of 8
31# the np argument is not just modulus value, but one interleaved
32# with 0. This is to optimize post-condition...
33
34$flavour = shift;
35$output  = shift;
36if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
37
38$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
39
40$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
41( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
42( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
43die "can't locate x86_64-xlate.pl";
44
45open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
46*STDOUT=*OUT;
47
48if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
49		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
50	$addx = ($1>=2.23);
51}
52
53if (!$addx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
54	    `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
55	$addx = ($1>=2.10);
56}
57
58if (!$addx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
59	    `ml64 2>&1` =~ /Version ([0-9]+)\./) {
60	$addx = ($1>=12);
61}
62
63if (!$addx && `$ENV{CC} -v 2>&1` =~ /((?:clang|LLVM) version|.*based on LLVM) ([0-9]+)\.([0-9]+)/) {
64	my $ver = $2 + $3/100.0;	# 3.1->3.01, 3.10->3.10
65	$addx = ($ver>=3.03);
66}
67
68# int bn_mul_mont_gather5(
69$rp="%rdi";	# BN_ULONG *rp,
70$ap="%rsi";	# const BN_ULONG *ap,
71$bp="%rdx";	# const BN_ULONG *bp,
72$np="%rcx";	# const BN_ULONG *np,
73$n0="%r8";	# const BN_ULONG *n0,
74$num="%r9";	# int num,
75		# int idx);	# 0 to 2^5-1, "index" in $bp holding
76				# pre-computed powers of a', interlaced
77				# in such manner that b[0] is $bp[idx],
78				# b[1] is [2^5+idx], etc.
79$lo0="%r10";
80$hi0="%r11";
81$hi1="%r13";
82$i="%r14";
83$j="%r15";
84$m0="%rbx";
85$m1="%rbp";
86
87$code=<<___;
88.text
89
90.extern	OPENSSL_ia32cap_P
91
92.globl	bn_mul_mont_gather5
93.type	bn_mul_mont_gather5,\@function,6
94.align	64
95bn_mul_mont_gather5:
96.cfi_startproc
97	mov	${num}d,${num}d
98	mov	%rsp,%rax
99.cfi_def_cfa_register	%rax
100	test	\$7,${num}d
101	jnz	.Lmul_enter
102___
103$code.=<<___ if ($addx);
104	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
105___
106$code.=<<___;
107	jmp	.Lmul4x_enter
108
109.align	16
110.Lmul_enter:
111	movd	`($win64?56:8)`(%rsp),%xmm5	# load 7th argument
112	push	%rbx
113.cfi_push	%rbx
114	push	%rbp
115.cfi_push	%rbp
116	push	%r12
117.cfi_push	%r12
118	push	%r13
119.cfi_push	%r13
120	push	%r14
121.cfi_push	%r14
122	push	%r15
123.cfi_push	%r15
124
125	neg	$num
126	mov	%rsp,%r11
127	lea	-280(%rsp,$num,8),%r10	# future alloca(8*(num+2)+256+8)
128	neg	$num			# restore $num
129	and	\$-1024,%r10		# minimize TLB usage
130
131	# An OS-agnostic version of __chkstk.
132	#
133	# Some OSes (Windows) insist on stack being "wired" to
134	# physical memory in strictly sequential manner, i.e. if stack
135	# allocation spans two pages, then reference to farmost one can
136	# be punishable by SEGV. But page walking can do good even on
137	# other OSes, because it guarantees that villain thread hits
138	# the guard page before it can make damage to innocent one...
139	sub	%r10,%r11
140	and	\$-4096,%r11
141	lea	(%r10,%r11),%rsp
142	mov	(%rsp),%r11
143	cmp	%r10,%rsp
144	ja	.Lmul_page_walk
145	jmp	.Lmul_page_walk_done
146
147.Lmul_page_walk:
148	lea	-4096(%rsp),%rsp
149	mov	(%rsp),%r11
150	cmp	%r10,%rsp
151	ja	.Lmul_page_walk
152.Lmul_page_walk_done:
153
154	lea	.Linc(%rip),%r10
155	mov	%rax,8(%rsp,$num,8)	# tp[num+1]=%rsp
156.cfi_cfa_expression	%rsp+8,$num,8,mul,plus,deref,+8
157.Lmul_body:
158
159	lea	128($bp),%r12		# reassign $bp (+size optimization)
160___
161		$bp="%r12";
162		$STRIDE=2**5*8;		# 5 is "window size"
163		$N=$STRIDE/4;		# should match cache line size
164$code.=<<___;
165	movdqa	0(%r10),%xmm0		# 00000001000000010000000000000000
166	movdqa	16(%r10),%xmm1		# 00000002000000020000000200000002
167	lea	24-112(%rsp,$num,8),%r10# place the mask after tp[num+3] (+ICache optimization)
168	and	\$-16,%r10
169
170	pshufd	\$0,%xmm5,%xmm5		# broadcast index
171	movdqa	%xmm1,%xmm4
172	movdqa	%xmm1,%xmm2
173___
174########################################################################
175# calculate mask by comparing 0..31 to index and save result to stack
176#
177$code.=<<___;
178	paddd	%xmm0,%xmm1
179	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
180	.byte	0x67
181	movdqa	%xmm4,%xmm3
182___
183for($k=0;$k<$STRIDE/16-4;$k+=4) {
184$code.=<<___;
185	paddd	%xmm1,%xmm2
186	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
187	movdqa	%xmm0,`16*($k+0)+112`(%r10)
188	movdqa	%xmm4,%xmm0
189
190	paddd	%xmm2,%xmm3
191	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
192	movdqa	%xmm1,`16*($k+1)+112`(%r10)
193	movdqa	%xmm4,%xmm1
194
195	paddd	%xmm3,%xmm0
196	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
197	movdqa	%xmm2,`16*($k+2)+112`(%r10)
198	movdqa	%xmm4,%xmm2
199
200	paddd	%xmm0,%xmm1
201	pcmpeqd	%xmm5,%xmm0
202	movdqa	%xmm3,`16*($k+3)+112`(%r10)
203	movdqa	%xmm4,%xmm3
204___
205}
206$code.=<<___;				# last iteration can be optimized
207	paddd	%xmm1,%xmm2
208	pcmpeqd	%xmm5,%xmm1
209	movdqa	%xmm0,`16*($k+0)+112`(%r10)
210
211	paddd	%xmm2,%xmm3
212	.byte	0x67
213	pcmpeqd	%xmm5,%xmm2
214	movdqa	%xmm1,`16*($k+1)+112`(%r10)
215
216	pcmpeqd	%xmm5,%xmm3
217	movdqa	%xmm2,`16*($k+2)+112`(%r10)
218	pand	`16*($k+0)-128`($bp),%xmm0	# while it's still in register
219
220	pand	`16*($k+1)-128`($bp),%xmm1
221	pand	`16*($k+2)-128`($bp),%xmm2
222	movdqa	%xmm3,`16*($k+3)+112`(%r10)
223	pand	`16*($k+3)-128`($bp),%xmm3
224	por	%xmm2,%xmm0
225	por	%xmm3,%xmm1
226___
227for($k=0;$k<$STRIDE/16-4;$k+=4) {
228$code.=<<___;
229	movdqa	`16*($k+0)-128`($bp),%xmm4
230	movdqa	`16*($k+1)-128`($bp),%xmm5
231	movdqa	`16*($k+2)-128`($bp),%xmm2
232	pand	`16*($k+0)+112`(%r10),%xmm4
233	movdqa	`16*($k+3)-128`($bp),%xmm3
234	pand	`16*($k+1)+112`(%r10),%xmm5
235	por	%xmm4,%xmm0
236	pand	`16*($k+2)+112`(%r10),%xmm2
237	por	%xmm5,%xmm1
238	pand	`16*($k+3)+112`(%r10),%xmm3
239	por	%xmm2,%xmm0
240	por	%xmm3,%xmm1
241___
242}
243$code.=<<___;
244	por	%xmm1,%xmm0
245	pshufd	\$0x4e,%xmm0,%xmm1
246	por	%xmm1,%xmm0
247	lea	$STRIDE($bp),$bp
248	movq	%xmm0,$m0		# m0=bp[0]
249
250	mov	($n0),$n0		# pull n0[0] value
251	mov	($ap),%rax
252
253	xor	$i,$i			# i=0
254	xor	$j,$j			# j=0
255
256	mov	$n0,$m1
257	mulq	$m0			# ap[0]*bp[0]
258	mov	%rax,$lo0
259	mov	($np),%rax
260
261	imulq	$lo0,$m1		# "tp[0]"*n0
262	mov	%rdx,$hi0
263
264	mulq	$m1			# np[0]*m1
265	add	%rax,$lo0		# discarded
266	mov	8($ap),%rax
267	adc	\$0,%rdx
268	mov	%rdx,$hi1
269
270	lea	1($j),$j		# j++
271	jmp	.L1st_enter
272
273.align	16
274.L1st:
275	add	%rax,$hi1
276	mov	($ap,$j,8),%rax
277	adc	\$0,%rdx
278	add	$hi0,$hi1		# np[j]*m1+ap[j]*bp[0]
279	mov	$lo0,$hi0
280	adc	\$0,%rdx
281	mov	$hi1,-16(%rsp,$j,8)	# tp[j-1]
282	mov	%rdx,$hi1
283
284.L1st_enter:
285	mulq	$m0			# ap[j]*bp[0]
286	add	%rax,$hi0
287	mov	($np,$j,8),%rax
288	adc	\$0,%rdx
289	lea	1($j),$j		# j++
290	mov	%rdx,$lo0
291
292	mulq	$m1			# np[j]*m1
293	cmp	$num,$j
294	jne	.L1st			# note that upon exit $j==$num, so
295					# they can be used interchangeably
296
297	add	%rax,$hi1
298	adc	\$0,%rdx
299	add	$hi0,$hi1		# np[j]*m1+ap[j]*bp[0]
300	adc	\$0,%rdx
301	mov	$hi1,-16(%rsp,$num,8)	# tp[num-1]
302	mov	%rdx,$hi1
303	mov	$lo0,$hi0
304
305	xor	%rdx,%rdx
306	add	$hi0,$hi1
307	adc	\$0,%rdx
308	mov	$hi1,-8(%rsp,$num,8)
309	mov	%rdx,(%rsp,$num,8)	# store upmost overflow bit
310
311	lea	1($i),$i		# i++
312	jmp	.Louter
313.align	16
314.Louter:
315	lea	24+128(%rsp,$num,8),%rdx	# where 256-byte mask is (+size optimization)
316	and	\$-16,%rdx
317	pxor	%xmm4,%xmm4
318	pxor	%xmm5,%xmm5
319___
320for($k=0;$k<$STRIDE/16;$k+=4) {
321$code.=<<___;
322	movdqa	`16*($k+0)-128`($bp),%xmm0
323	movdqa	`16*($k+1)-128`($bp),%xmm1
324	movdqa	`16*($k+2)-128`($bp),%xmm2
325	movdqa	`16*($k+3)-128`($bp),%xmm3
326	pand	`16*($k+0)-128`(%rdx),%xmm0
327	pand	`16*($k+1)-128`(%rdx),%xmm1
328	por	%xmm0,%xmm4
329	pand	`16*($k+2)-128`(%rdx),%xmm2
330	por	%xmm1,%xmm5
331	pand	`16*($k+3)-128`(%rdx),%xmm3
332	por	%xmm2,%xmm4
333	por	%xmm3,%xmm5
334___
335}
336$code.=<<___;
337	por	%xmm5,%xmm4
338	pshufd	\$0x4e,%xmm4,%xmm0
339	por	%xmm4,%xmm0
340	lea	$STRIDE($bp),$bp
341
342	mov	($ap),%rax		# ap[0]
343	movq	%xmm0,$m0		# m0=bp[i]
344
345	xor	$j,$j			# j=0
346	mov	$n0,$m1
347	mov	(%rsp),$lo0
348
349	mulq	$m0			# ap[0]*bp[i]
350	add	%rax,$lo0		# ap[0]*bp[i]+tp[0]
351	mov	($np),%rax
352	adc	\$0,%rdx
353
354	imulq	$lo0,$m1		# tp[0]*n0
355	mov	%rdx,$hi0
356
357	mulq	$m1			# np[0]*m1
358	add	%rax,$lo0		# discarded
359	mov	8($ap),%rax
360	adc	\$0,%rdx
361	mov	8(%rsp),$lo0		# tp[1]
362	mov	%rdx,$hi1
363
364	lea	1($j),$j		# j++
365	jmp	.Linner_enter
366
367.align	16
368.Linner:
369	add	%rax,$hi1
370	mov	($ap,$j,8),%rax
371	adc	\$0,%rdx
372	add	$lo0,$hi1		# np[j]*m1+ap[j]*bp[i]+tp[j]
373	mov	(%rsp,$j,8),$lo0
374	adc	\$0,%rdx
375	mov	$hi1,-16(%rsp,$j,8)	# tp[j-1]
376	mov	%rdx,$hi1
377
378.Linner_enter:
379	mulq	$m0			# ap[j]*bp[i]
380	add	%rax,$hi0
381	mov	($np,$j,8),%rax
382	adc	\$0,%rdx
383	add	$hi0,$lo0		# ap[j]*bp[i]+tp[j]
384	mov	%rdx,$hi0
385	adc	\$0,$hi0
386	lea	1($j),$j		# j++
387
388	mulq	$m1			# np[j]*m1
389	cmp	$num,$j
390	jne	.Linner			# note that upon exit $j==$num, so
391					# they can be used interchangeably
392	add	%rax,$hi1
393	adc	\$0,%rdx
394	add	$lo0,$hi1		# np[j]*m1+ap[j]*bp[i]+tp[j]
395	mov	(%rsp,$num,8),$lo0
396	adc	\$0,%rdx
397	mov	$hi1,-16(%rsp,$num,8)	# tp[num-1]
398	mov	%rdx,$hi1
399
400	xor	%rdx,%rdx
401	add	$hi0,$hi1
402	adc	\$0,%rdx
403	add	$lo0,$hi1		# pull upmost overflow bit
404	adc	\$0,%rdx
405	mov	$hi1,-8(%rsp,$num,8)
406	mov	%rdx,(%rsp,$num,8)	# store upmost overflow bit
407
408	lea	1($i),$i		# i++
409	cmp	$num,$i
410	jb	.Louter
411
412	xor	$i,$i			# i=0 and clear CF!
413	mov	(%rsp),%rax		# tp[0]
414	lea	(%rsp),$ap		# borrow ap for tp
415	mov	$num,$j			# j=num
416	jmp	.Lsub
417.align	16
418.Lsub:	sbb	($np,$i,8),%rax
419	mov	%rax,($rp,$i,8)		# rp[i]=tp[i]-np[i]
420	mov	8($ap,$i,8),%rax	# tp[i+1]
421	lea	1($i),$i		# i++
422	dec	$j			# doesn't affect CF!
423	jnz	.Lsub
424
425	sbb	\$0,%rax		# handle upmost overflow bit
426	mov	\$-1,%rbx
427	xor	%rax,%rbx
428	xor	$i,$i
429	mov	$num,$j			# j=num
430
431.Lcopy:					# conditional copy
432	mov	($rp,$i,8),%rcx
433	mov	(%rsp,$i,8),%rdx
434	and	%rbx,%rcx
435	and	%rax,%rdx
436	mov	$i,(%rsp,$i,8)		# zap temporary vector
437	or	%rcx,%rdx
438	mov	%rdx,($rp,$i,8)		# rp[i]=tp[i]
439	lea	1($i),$i
440	sub	\$1,$j
441	jnz	.Lcopy
442
443	mov	8(%rsp,$num,8),%rsi	# restore %rsp
444.cfi_def_cfa	%rsi,8
445	mov	\$1,%rax
446
447	mov	-48(%rsi),%r15
448.cfi_restore	%r15
449	mov	-40(%rsi),%r14
450.cfi_restore	%r14
451	mov	-32(%rsi),%r13
452.cfi_restore	%r13
453	mov	-24(%rsi),%r12
454.cfi_restore	%r12
455	mov	-16(%rsi),%rbp
456.cfi_restore	%rbp
457	mov	-8(%rsi),%rbx
458.cfi_restore	%rbx
459	lea	(%rsi),%rsp
460.cfi_def_cfa_register	%rsp
461.Lmul_epilogue:
462	ret
463.cfi_endproc
464.size	bn_mul_mont_gather5,.-bn_mul_mont_gather5
465___
466{{{
467my @A=("%r10","%r11");
468my @N=("%r13","%rdi");
469$code.=<<___;
470.type	bn_mul4x_mont_gather5,\@function,6
471.align	32
472bn_mul4x_mont_gather5:
473.cfi_startproc
474	.byte	0x67
475	mov	%rsp,%rax
476.cfi_def_cfa_register	%rax
477.Lmul4x_enter:
478___
479$code.=<<___ if ($addx);
480	and	\$0x80108,%r11d
481	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
482	je	.Lmulx4x_enter
483___
484$code.=<<___;
485	push	%rbx
486.cfi_push	%rbx
487	push	%rbp
488.cfi_push	%rbp
489	push	%r12
490.cfi_push	%r12
491	push	%r13
492.cfi_push	%r13
493	push	%r14
494.cfi_push	%r14
495	push	%r15
496.cfi_push	%r15
497.Lmul4x_prologue:
498
499	.byte	0x67
500	shl	\$3,${num}d		# convert $num to bytes
501	lea	($num,$num,2),%r10	# 3*$num in bytes
502	neg	$num			# -$num
503
504	##############################################################
505	# Ensure that stack frame doesn't alias with $rptr+3*$num
506	# modulo 4096, which covers ret[num], am[num] and n[num]
507	# (see bn_exp.c). This is done to allow memory disambiguation
508	# logic do its magic. [Extra [num] is allocated in order
509	# to align with bn_power5's frame, which is cleansed after
510	# completing exponentiation. Extra 256 bytes is for power mask
511	# calculated from 7th argument, the index.]
512	#
513	lea	-320(%rsp,$num,2),%r11
514	mov	%rsp,%rbp
515	sub	$rp,%r11
516	and	\$4095,%r11
517	cmp	%r11,%r10
518	jb	.Lmul4xsp_alt
519	sub	%r11,%rbp		# align with $rp
520	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
521	jmp	.Lmul4xsp_done
522
523.align	32
524.Lmul4xsp_alt:
525	lea	4096-320(,$num,2),%r10
526	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
527	sub	%r10,%r11
528	mov	\$0,%r10
529	cmovc	%r10,%r11
530	sub	%r11,%rbp
531.Lmul4xsp_done:
532	and	\$-64,%rbp
533	mov	%rsp,%r11
534	sub	%rbp,%r11
535	and	\$-4096,%r11
536	lea	(%rbp,%r11),%rsp
537	mov	(%rsp),%r10
538	cmp	%rbp,%rsp
539	ja	.Lmul4x_page_walk
540	jmp	.Lmul4x_page_walk_done
541
542.Lmul4x_page_walk:
543	lea	-4096(%rsp),%rsp
544	mov	(%rsp),%r10
545	cmp	%rbp,%rsp
546	ja	.Lmul4x_page_walk
547.Lmul4x_page_walk_done:
548
549	neg	$num
550
551	mov	%rax,40(%rsp)
552.cfi_cfa_expression	%rsp+40,deref,+8
553.Lmul4x_body:
554
555	call	mul4x_internal
556
557	mov	40(%rsp),%rsi		# restore %rsp
558.cfi_def_cfa	%rsi,8
559	mov	\$1,%rax
560
561	mov	-48(%rsi),%r15
562.cfi_restore	%r15
563	mov	-40(%rsi),%r14
564.cfi_restore	%r14
565	mov	-32(%rsi),%r13
566.cfi_restore	%r13
567	mov	-24(%rsi),%r12
568.cfi_restore	%r12
569	mov	-16(%rsi),%rbp
570.cfi_restore	%rbp
571	mov	-8(%rsi),%rbx
572.cfi_restore	%rbx
573	lea	(%rsi),%rsp
574.cfi_def_cfa_register	%rsp
575.Lmul4x_epilogue:
576	ret
577.cfi_endproc
578.size	bn_mul4x_mont_gather5,.-bn_mul4x_mont_gather5
579
580.type	mul4x_internal,\@abi-omnipotent
581.align	32
582mul4x_internal:
583.cfi_startproc
584	shl	\$5,$num		# $num was in bytes
585	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument, index
586	lea	.Linc(%rip),%rax
587	lea	128(%rdx,$num),%r13	# end of powers table (+size optimization)
588	shr	\$5,$num		# restore $num
589___
590		$bp="%r12";
591		$STRIDE=2**5*8;		# 5 is "window size"
592		$N=$STRIDE/4;		# should match cache line size
593		$tp=$i;
594$code.=<<___;
595	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
596	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
597	lea	88-112(%rsp,$num),%r10	# place the mask after tp[num+1] (+ICache optimization)
598	lea	128(%rdx),$bp		# size optimization
599
600	pshufd	\$0,%xmm5,%xmm5		# broadcast index
601	movdqa	%xmm1,%xmm4
602	.byte	0x67,0x67
603	movdqa	%xmm1,%xmm2
604___
605########################################################################
606# calculate mask by comparing 0..31 to index and save result to stack
607#
608$code.=<<___;
609	paddd	%xmm0,%xmm1
610	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
611	.byte	0x67
612	movdqa	%xmm4,%xmm3
613___
614for($i=0;$i<$STRIDE/16-4;$i+=4) {
615$code.=<<___;
616	paddd	%xmm1,%xmm2
617	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
618	movdqa	%xmm0,`16*($i+0)+112`(%r10)
619	movdqa	%xmm4,%xmm0
620
621	paddd	%xmm2,%xmm3
622	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
623	movdqa	%xmm1,`16*($i+1)+112`(%r10)
624	movdqa	%xmm4,%xmm1
625
626	paddd	%xmm3,%xmm0
627	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
628	movdqa	%xmm2,`16*($i+2)+112`(%r10)
629	movdqa	%xmm4,%xmm2
630
631	paddd	%xmm0,%xmm1
632	pcmpeqd	%xmm5,%xmm0
633	movdqa	%xmm3,`16*($i+3)+112`(%r10)
634	movdqa	%xmm4,%xmm3
635___
636}
637$code.=<<___;				# last iteration can be optimized
638	paddd	%xmm1,%xmm2
639	pcmpeqd	%xmm5,%xmm1
640	movdqa	%xmm0,`16*($i+0)+112`(%r10)
641
642	paddd	%xmm2,%xmm3
643	.byte	0x67
644	pcmpeqd	%xmm5,%xmm2
645	movdqa	%xmm1,`16*($i+1)+112`(%r10)
646
647	pcmpeqd	%xmm5,%xmm3
648	movdqa	%xmm2,`16*($i+2)+112`(%r10)
649	pand	`16*($i+0)-128`($bp),%xmm0	# while it's still in register
650
651	pand	`16*($i+1)-128`($bp),%xmm1
652	pand	`16*($i+2)-128`($bp),%xmm2
653	movdqa	%xmm3,`16*($i+3)+112`(%r10)
654	pand	`16*($i+3)-128`($bp),%xmm3
655	por	%xmm2,%xmm0
656	por	%xmm3,%xmm1
657___
658for($i=0;$i<$STRIDE/16-4;$i+=4) {
659$code.=<<___;
660	movdqa	`16*($i+0)-128`($bp),%xmm4
661	movdqa	`16*($i+1)-128`($bp),%xmm5
662	movdqa	`16*($i+2)-128`($bp),%xmm2
663	pand	`16*($i+0)+112`(%r10),%xmm4
664	movdqa	`16*($i+3)-128`($bp),%xmm3
665	pand	`16*($i+1)+112`(%r10),%xmm5
666	por	%xmm4,%xmm0
667	pand	`16*($i+2)+112`(%r10),%xmm2
668	por	%xmm5,%xmm1
669	pand	`16*($i+3)+112`(%r10),%xmm3
670	por	%xmm2,%xmm0
671	por	%xmm3,%xmm1
672___
673}
674$code.=<<___;
675	por	%xmm1,%xmm0
676	pshufd	\$0x4e,%xmm0,%xmm1
677	por	%xmm1,%xmm0
678	lea	$STRIDE($bp),$bp
679	movq	%xmm0,$m0		# m0=bp[0]
680
681	mov	%r13,16+8(%rsp)		# save end of b[num]
682	mov	$rp, 56+8(%rsp)		# save $rp
683
684	mov	($n0),$n0		# pull n0[0] value
685	mov	($ap),%rax
686	lea	($ap,$num),$ap		# end of a[num]
687	neg	$num
688
689	mov	$n0,$m1
690	mulq	$m0			# ap[0]*bp[0]
691	mov	%rax,$A[0]
692	mov	($np),%rax
693
694	imulq	$A[0],$m1		# "tp[0]"*n0
695	lea	64+8(%rsp),$tp
696	mov	%rdx,$A[1]
697
698	mulq	$m1			# np[0]*m1
699	add	%rax,$A[0]		# discarded
700	mov	8($ap,$num),%rax
701	adc	\$0,%rdx
702	mov	%rdx,$N[1]
703
704	mulq	$m0
705	add	%rax,$A[1]
706	mov	8*1($np),%rax
707	adc	\$0,%rdx
708	mov	%rdx,$A[0]
709
710	mulq	$m1
711	add	%rax,$N[1]
712	mov	16($ap,$num),%rax
713	adc	\$0,%rdx
714	add	$A[1],$N[1]
715	lea	4*8($num),$j		# j=4
716	lea	8*4($np),$np
717	adc	\$0,%rdx
718	mov	$N[1],($tp)
719	mov	%rdx,$N[0]
720	jmp	.L1st4x
721
722.align	32
723.L1st4x:
724	mulq	$m0			# ap[j]*bp[0]
725	add	%rax,$A[0]
726	mov	-8*2($np),%rax
727	lea	32($tp),$tp
728	adc	\$0,%rdx
729	mov	%rdx,$A[1]
730
731	mulq	$m1			# np[j]*m1
732	add	%rax,$N[0]
733	mov	-8($ap,$j),%rax
734	adc	\$0,%rdx
735	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
736	adc	\$0,%rdx
737	mov	$N[0],-24($tp)		# tp[j-1]
738	mov	%rdx,$N[1]
739
740	mulq	$m0			# ap[j]*bp[0]
741	add	%rax,$A[1]
742	mov	-8*1($np),%rax
743	adc	\$0,%rdx
744	mov	%rdx,$A[0]
745
746	mulq	$m1			# np[j]*m1
747	add	%rax,$N[1]
748	mov	($ap,$j),%rax
749	adc	\$0,%rdx
750	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
751	adc	\$0,%rdx
752	mov	$N[1],-16($tp)		# tp[j-1]
753	mov	%rdx,$N[0]
754
755	mulq	$m0			# ap[j]*bp[0]
756	add	%rax,$A[0]
757	mov	8*0($np),%rax
758	adc	\$0,%rdx
759	mov	%rdx,$A[1]
760
761	mulq	$m1			# np[j]*m1
762	add	%rax,$N[0]
763	mov	8($ap,$j),%rax
764	adc	\$0,%rdx
765	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
766	adc	\$0,%rdx
767	mov	$N[0],-8($tp)		# tp[j-1]
768	mov	%rdx,$N[1]
769
770	mulq	$m0			# ap[j]*bp[0]
771	add	%rax,$A[1]
772	mov	8*1($np),%rax
773	adc	\$0,%rdx
774	mov	%rdx,$A[0]
775
776	mulq	$m1			# np[j]*m1
777	add	%rax,$N[1]
778	mov	16($ap,$j),%rax
779	adc	\$0,%rdx
780	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
781	lea	8*4($np),$np
782	adc	\$0,%rdx
783	mov	$N[1],($tp)		# tp[j-1]
784	mov	%rdx,$N[0]
785
786	add	\$32,$j			# j+=4
787	jnz	.L1st4x
788
789	mulq	$m0			# ap[j]*bp[0]
790	add	%rax,$A[0]
791	mov	-8*2($np),%rax
792	lea	32($tp),$tp
793	adc	\$0,%rdx
794	mov	%rdx,$A[1]
795
796	mulq	$m1			# np[j]*m1
797	add	%rax,$N[0]
798	mov	-8($ap),%rax
799	adc	\$0,%rdx
800	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
801	adc	\$0,%rdx
802	mov	$N[0],-24($tp)		# tp[j-1]
803	mov	%rdx,$N[1]
804
805	mulq	$m0			# ap[j]*bp[0]
806	add	%rax,$A[1]
807	mov	-8*1($np),%rax
808	adc	\$0,%rdx
809	mov	%rdx,$A[0]
810
811	mulq	$m1			# np[j]*m1
812	add	%rax,$N[1]
813	mov	($ap,$num),%rax		# ap[0]
814	adc	\$0,%rdx
815	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
816	adc	\$0,%rdx
817	mov	$N[1],-16($tp)		# tp[j-1]
818	mov	%rdx,$N[0]
819
820	lea	($np,$num),$np		# rewind $np
821
822	xor	$N[1],$N[1]
823	add	$A[0],$N[0]
824	adc	\$0,$N[1]
825	mov	$N[0],-8($tp)
826
827	jmp	.Louter4x
828
829.align	32
830.Louter4x:
831	lea	16+128($tp),%rdx	# where 256-byte mask is (+size optimization)
832	pxor	%xmm4,%xmm4
833	pxor	%xmm5,%xmm5
834___
835for($i=0;$i<$STRIDE/16;$i+=4) {
836$code.=<<___;
837	movdqa	`16*($i+0)-128`($bp),%xmm0
838	movdqa	`16*($i+1)-128`($bp),%xmm1
839	movdqa	`16*($i+2)-128`($bp),%xmm2
840	movdqa	`16*($i+3)-128`($bp),%xmm3
841	pand	`16*($i+0)-128`(%rdx),%xmm0
842	pand	`16*($i+1)-128`(%rdx),%xmm1
843	por	%xmm0,%xmm4
844	pand	`16*($i+2)-128`(%rdx),%xmm2
845	por	%xmm1,%xmm5
846	pand	`16*($i+3)-128`(%rdx),%xmm3
847	por	%xmm2,%xmm4
848	por	%xmm3,%xmm5
849___
850}
851$code.=<<___;
852	por	%xmm5,%xmm4
853	pshufd	\$0x4e,%xmm4,%xmm0
854	por	%xmm4,%xmm0
855	lea	$STRIDE($bp),$bp
856	movq	%xmm0,$m0		# m0=bp[i]
857
858	mov	($tp,$num),$A[0]
859	mov	$n0,$m1
860	mulq	$m0			# ap[0]*bp[i]
861	add	%rax,$A[0]		# ap[0]*bp[i]+tp[0]
862	mov	($np),%rax
863	adc	\$0,%rdx
864
865	imulq	$A[0],$m1		# tp[0]*n0
866	mov	%rdx,$A[1]
867	mov	$N[1],($tp)		# store upmost overflow bit
868
869	lea	($tp,$num),$tp		# rewind $tp
870
871	mulq	$m1			# np[0]*m1
872	add	%rax,$A[0]		# "$N[0]", discarded
873	mov	8($ap,$num),%rax
874	adc	\$0,%rdx
875	mov	%rdx,$N[1]
876
877	mulq	$m0			# ap[j]*bp[i]
878	add	%rax,$A[1]
879	mov	8*1($np),%rax
880	adc	\$0,%rdx
881	add	8($tp),$A[1]		# +tp[1]
882	adc	\$0,%rdx
883	mov	%rdx,$A[0]
884
885	mulq	$m1			# np[j]*m1
886	add	%rax,$N[1]
887	mov	16($ap,$num),%rax
888	adc	\$0,%rdx
889	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[i]+tp[j]
890	lea	4*8($num),$j		# j=4
891	lea	8*4($np),$np
892	adc	\$0,%rdx
893	mov	%rdx,$N[0]
894	jmp	.Linner4x
895
896.align	32
897.Linner4x:
898	mulq	$m0			# ap[j]*bp[i]
899	add	%rax,$A[0]
900	mov	-8*2($np),%rax
901	adc	\$0,%rdx
902	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
903	lea	32($tp),$tp
904	adc	\$0,%rdx
905	mov	%rdx,$A[1]
906
907	mulq	$m1			# np[j]*m1
908	add	%rax,$N[0]
909	mov	-8($ap,$j),%rax
910	adc	\$0,%rdx
911	add	$A[0],$N[0]
912	adc	\$0,%rdx
913	mov	$N[1],-32($tp)		# tp[j-1]
914	mov	%rdx,$N[1]
915
916	mulq	$m0			# ap[j]*bp[i]
917	add	%rax,$A[1]
918	mov	-8*1($np),%rax
919	adc	\$0,%rdx
920	add	-8($tp),$A[1]
921	adc	\$0,%rdx
922	mov	%rdx,$A[0]
923
924	mulq	$m1			# np[j]*m1
925	add	%rax,$N[1]
926	mov	($ap,$j),%rax
927	adc	\$0,%rdx
928	add	$A[1],$N[1]
929	adc	\$0,%rdx
930	mov	$N[0],-24($tp)		# tp[j-1]
931	mov	%rdx,$N[0]
932
933	mulq	$m0			# ap[j]*bp[i]
934	add	%rax,$A[0]
935	mov	8*0($np),%rax
936	adc	\$0,%rdx
937	add	($tp),$A[0]		# ap[j]*bp[i]+tp[j]
938	adc	\$0,%rdx
939	mov	%rdx,$A[1]
940
941	mulq	$m1			# np[j]*m1
942	add	%rax,$N[0]
943	mov	8($ap,$j),%rax
944	adc	\$0,%rdx
945	add	$A[0],$N[0]
946	adc	\$0,%rdx
947	mov	$N[1],-16($tp)		# tp[j-1]
948	mov	%rdx,$N[1]
949
950	mulq	$m0			# ap[j]*bp[i]
951	add	%rax,$A[1]
952	mov	8*1($np),%rax
953	adc	\$0,%rdx
954	add	8($tp),$A[1]
955	adc	\$0,%rdx
956	mov	%rdx,$A[0]
957
958	mulq	$m1			# np[j]*m1
959	add	%rax,$N[1]
960	mov	16($ap,$j),%rax
961	adc	\$0,%rdx
962	add	$A[1],$N[1]
963	lea	8*4($np),$np
964	adc	\$0,%rdx
965	mov	$N[0],-8($tp)		# tp[j-1]
966	mov	%rdx,$N[0]
967
968	add	\$32,$j			# j+=4
969	jnz	.Linner4x
970
971	mulq	$m0			# ap[j]*bp[i]
972	add	%rax,$A[0]
973	mov	-8*2($np),%rax
974	adc	\$0,%rdx
975	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
976	lea	32($tp),$tp
977	adc	\$0,%rdx
978	mov	%rdx,$A[1]
979
980	mulq	$m1			# np[j]*m1
981	add	%rax,$N[0]
982	mov	-8($ap),%rax
983	adc	\$0,%rdx
984	add	$A[0],$N[0]
985	adc	\$0,%rdx
986	mov	$N[1],-32($tp)		# tp[j-1]
987	mov	%rdx,$N[1]
988
989	mulq	$m0			# ap[j]*bp[i]
990	add	%rax,$A[1]
991	mov	$m1,%rax
992	mov	-8*1($np),$m1
993	adc	\$0,%rdx
994	add	-8($tp),$A[1]
995	adc	\$0,%rdx
996	mov	%rdx,$A[0]
997
998	mulq	$m1			# np[j]*m1
999	add	%rax,$N[1]
1000	mov	($ap,$num),%rax		# ap[0]
1001	adc	\$0,%rdx
1002	add	$A[1],$N[1]
1003	adc	\$0,%rdx
1004	mov	$N[0],-24($tp)		# tp[j-1]
1005	mov	%rdx,$N[0]
1006
1007	mov	$N[1],-16($tp)		# tp[j-1]
1008	lea	($np,$num),$np		# rewind $np
1009
1010	xor	$N[1],$N[1]
1011	add	$A[0],$N[0]
1012	adc	\$0,$N[1]
1013	add	($tp),$N[0]		# pull upmost overflow bit
1014	adc	\$0,$N[1]		# upmost overflow bit
1015	mov	$N[0],-8($tp)
1016
1017	cmp	16+8(%rsp),$bp
1018	jb	.Louter4x
1019___
1020if (1) {
1021$code.=<<___;
1022	xor	%rax,%rax
1023	sub	$N[0],$m1		# compare top-most words
1024	adc	$j,$j			# $j is zero
1025	or	$j,$N[1]
1026	sub	$N[1],%rax		# %rax=-$N[1]
1027	lea	($tp,$num),%rbx		# tptr in .sqr4x_sub
1028	mov	($np),%r12
1029	lea	($np),%rbp		# nptr in .sqr4x_sub
1030	mov	%r9,%rcx
1031	sar	\$3+2,%rcx
1032	mov	56+8(%rsp),%rdi		# rptr in .sqr4x_sub
1033	dec	%r12			# so that after 'not' we get -n[0]
1034	xor	%r10,%r10
1035	mov	8*1(%rbp),%r13
1036	mov	8*2(%rbp),%r14
1037	mov	8*3(%rbp),%r15
1038	jmp	.Lsqr4x_sub_entry
1039___
1040} else {
1041my @ri=("%rax",$bp,$m0,$m1);
1042my $rp="%rdx";
1043$code.=<<___
1044	xor	\$1,$N[1]
1045	lea	($tp,$num),$tp		# rewind $tp
1046	sar	\$5,$num		# cf=0
1047	lea	($np,$N[1],8),$np
1048	mov	56+8(%rsp),$rp		# restore $rp
1049	jmp	.Lsub4x
1050
1051.align	32
1052.Lsub4x:
1053	.byte	0x66
1054	mov	8*0($tp),@ri[0]
1055	mov	8*1($tp),@ri[1]
1056	.byte	0x66
1057	sbb	16*0($np),@ri[0]
1058	mov	8*2($tp),@ri[2]
1059	sbb	16*1($np),@ri[1]
1060	mov	3*8($tp),@ri[3]
1061	lea	4*8($tp),$tp
1062	sbb	16*2($np),@ri[2]
1063	mov	@ri[0],8*0($rp)
1064	sbb	16*3($np),@ri[3]
1065	lea	16*4($np),$np
1066	mov	@ri[1],8*1($rp)
1067	mov	@ri[2],8*2($rp)
1068	mov	@ri[3],8*3($rp)
1069	lea	8*4($rp),$rp
1070
1071	inc	$num
1072	jnz	.Lsub4x
1073
1074	ret
1075___
1076}
1077$code.=<<___;
1078.cfi_endproc
1079.size	mul4x_internal,.-mul4x_internal
1080___
1081}}}
1082{{{
1083######################################################################
1084# void bn_power5(
1085my $rptr="%rdi";	# BN_ULONG *rptr,
1086my $aptr="%rsi";	# const BN_ULONG *aptr,
1087my $bptr="%rdx";	# const void *table,
1088my $nptr="%rcx";	# const BN_ULONG *nptr,
1089my $n0  ="%r8";		# const BN_ULONG *n0);
1090my $num ="%r9";		# int num, has to be divisible by 8
1091			# int pwr
1092
1093my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
1094my @A0=("%r10","%r11");
1095my @A1=("%r12","%r13");
1096my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
1097
1098$code.=<<___;
1099.globl	bn_power5
1100.type	bn_power5,\@function,6
1101.align	32
1102bn_power5:
1103.cfi_startproc
1104	mov	%rsp,%rax
1105.cfi_def_cfa_register	%rax
1106___
1107$code.=<<___ if ($addx);
1108	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
1109	and	\$0x80108,%r11d
1110	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
1111	je	.Lpowerx5_enter
1112___
1113$code.=<<___;
1114	push	%rbx
1115.cfi_push	%rbx
1116	push	%rbp
1117.cfi_push	%rbp
1118	push	%r12
1119.cfi_push	%r12
1120	push	%r13
1121.cfi_push	%r13
1122	push	%r14
1123.cfi_push	%r14
1124	push	%r15
1125.cfi_push	%r15
1126.Lpower5_prologue:
1127
1128	shl	\$3,${num}d		# convert $num to bytes
1129	lea	($num,$num,2),%r10d	# 3*$num
1130	neg	$num
1131	mov	($n0),$n0		# *n0
1132
1133	##############################################################
1134	# Ensure that stack frame doesn't alias with $rptr+3*$num
1135	# modulo 4096, which covers ret[num], am[num] and n[num]
1136	# (see bn_exp.c). This is done to allow memory disambiguation
1137	# logic do its magic. [Extra 256 bytes is for power mask
1138	# calculated from 7th argument, the index.]
1139	#
1140	lea	-320(%rsp,$num,2),%r11
1141	mov	%rsp,%rbp
1142	sub	$rptr,%r11
1143	and	\$4095,%r11
1144	cmp	%r11,%r10
1145	jb	.Lpwr_sp_alt
1146	sub	%r11,%rbp		# align with $aptr
1147	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1148	jmp	.Lpwr_sp_done
1149
1150.align	32
1151.Lpwr_sp_alt:
1152	lea	4096-320(,$num,2),%r10
1153	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1154	sub	%r10,%r11
1155	mov	\$0,%r10
1156	cmovc	%r10,%r11
1157	sub	%r11,%rbp
1158.Lpwr_sp_done:
1159	and	\$-64,%rbp
1160	mov	%rsp,%r11
1161	sub	%rbp,%r11
1162	and	\$-4096,%r11
1163	lea	(%rbp,%r11),%rsp
1164	mov	(%rsp),%r10
1165	cmp	%rbp,%rsp
1166	ja	.Lpwr_page_walk
1167	jmp	.Lpwr_page_walk_done
1168
1169.Lpwr_page_walk:
1170	lea	-4096(%rsp),%rsp
1171	mov	(%rsp),%r10
1172	cmp	%rbp,%rsp
1173	ja	.Lpwr_page_walk
1174.Lpwr_page_walk_done:
1175
1176	mov	$num,%r10
1177	neg	$num
1178
1179	##############################################################
1180	# Stack layout
1181	#
1182	# +0	saved $num, used in reduction section
1183	# +8	&t[2*$num], used in reduction section
1184	# +32	saved *n0
1185	# +40	saved %rsp
1186	# +48	t[2*$num]
1187	#
1188	mov	$n0,  32(%rsp)
1189	mov	%rax, 40(%rsp)		# save original %rsp
1190.cfi_cfa_expression	%rsp+40,deref,+8
1191.Lpower5_body:
1192	movq	$rptr,%xmm1		# save $rptr, used in sqr8x
1193	movq	$nptr,%xmm2		# save $nptr
1194	movq	%r10, %xmm3		# -$num, used in sqr8x
1195	movq	$bptr,%xmm4
1196
1197	call	__bn_sqr8x_internal
1198	call	__bn_post4x_internal
1199	call	__bn_sqr8x_internal
1200	call	__bn_post4x_internal
1201	call	__bn_sqr8x_internal
1202	call	__bn_post4x_internal
1203	call	__bn_sqr8x_internal
1204	call	__bn_post4x_internal
1205	call	__bn_sqr8x_internal
1206	call	__bn_post4x_internal
1207
1208	movq	%xmm2,$nptr
1209	movq	%xmm4,$bptr
1210	mov	$aptr,$rptr
1211	mov	40(%rsp),%rax
1212	lea	32(%rsp),$n0
1213
1214	call	mul4x_internal
1215
1216	mov	40(%rsp),%rsi		# restore %rsp
1217.cfi_def_cfa	%rsi,8
1218	mov	\$1,%rax
1219	mov	-48(%rsi),%r15
1220.cfi_restore	%r15
1221	mov	-40(%rsi),%r14
1222.cfi_restore	%r14
1223	mov	-32(%rsi),%r13
1224.cfi_restore	%r13
1225	mov	-24(%rsi),%r12
1226.cfi_restore	%r12
1227	mov	-16(%rsi),%rbp
1228.cfi_restore	%rbp
1229	mov	-8(%rsi),%rbx
1230.cfi_restore	%rbx
1231	lea	(%rsi),%rsp
1232.cfi_def_cfa_register	%rsp
1233.Lpower5_epilogue:
1234	ret
1235.cfi_endproc
1236.size	bn_power5,.-bn_power5
1237
1238.globl	bn_sqr8x_internal
1239.hidden	bn_sqr8x_internal
1240.type	bn_sqr8x_internal,\@abi-omnipotent
1241.align	32
1242bn_sqr8x_internal:
1243__bn_sqr8x_internal:
1244.cfi_startproc
1245	##############################################################
1246	# Squaring part:
1247	#
1248	# a) multiply-n-add everything but a[i]*a[i];
1249	# b) shift result of a) by 1 to the left and accumulate
1250	#    a[i]*a[i] products;
1251	#
1252	##############################################################
1253	#                                                     a[1]a[0]
1254	#                                                 a[2]a[0]
1255	#                                             a[3]a[0]
1256	#                                             a[2]a[1]
1257	#                                         a[4]a[0]
1258	#                                         a[3]a[1]
1259	#                                     a[5]a[0]
1260	#                                     a[4]a[1]
1261	#                                     a[3]a[2]
1262	#                                 a[6]a[0]
1263	#                                 a[5]a[1]
1264	#                                 a[4]a[2]
1265	#                             a[7]a[0]
1266	#                             a[6]a[1]
1267	#                             a[5]a[2]
1268	#                             a[4]a[3]
1269	#                         a[7]a[1]
1270	#                         a[6]a[2]
1271	#                         a[5]a[3]
1272	#                     a[7]a[2]
1273	#                     a[6]a[3]
1274	#                     a[5]a[4]
1275	#                 a[7]a[3]
1276	#                 a[6]a[4]
1277	#             a[7]a[4]
1278	#             a[6]a[5]
1279	#         a[7]a[5]
1280	#     a[7]a[6]
1281	#                                                     a[1]a[0]
1282	#                                                 a[2]a[0]
1283	#                                             a[3]a[0]
1284	#                                         a[4]a[0]
1285	#                                     a[5]a[0]
1286	#                                 a[6]a[0]
1287	#                             a[7]a[0]
1288	#                                             a[2]a[1]
1289	#                                         a[3]a[1]
1290	#                                     a[4]a[1]
1291	#                                 a[5]a[1]
1292	#                             a[6]a[1]
1293	#                         a[7]a[1]
1294	#                                     a[3]a[2]
1295	#                                 a[4]a[2]
1296	#                             a[5]a[2]
1297	#                         a[6]a[2]
1298	#                     a[7]a[2]
1299	#                             a[4]a[3]
1300	#                         a[5]a[3]
1301	#                     a[6]a[3]
1302	#                 a[7]a[3]
1303	#                     a[5]a[4]
1304	#                 a[6]a[4]
1305	#             a[7]a[4]
1306	#             a[6]a[5]
1307	#         a[7]a[5]
1308	#     a[7]a[6]
1309	#                                                         a[0]a[0]
1310	#                                                 a[1]a[1]
1311	#                                         a[2]a[2]
1312	#                                 a[3]a[3]
1313	#                         a[4]a[4]
1314	#                 a[5]a[5]
1315	#         a[6]a[6]
1316	# a[7]a[7]
1317
1318	lea	32(%r10),$i		# $i=-($num-32)
1319	lea	($aptr,$num),$aptr	# end of a[] buffer, ($aptr,$i)=&ap[2]
1320
1321	mov	$num,$j			# $j=$num
1322
1323					# comments apply to $num==8 case
1324	mov	-32($aptr,$i),$a0	# a[0]
1325	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1326	mov	-24($aptr,$i),%rax	# a[1]
1327	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1328	mov	-16($aptr,$i),$ai	# a[2]
1329	mov	%rax,$a1
1330
1331	mul	$a0			# a[1]*a[0]
1332	mov	%rax,$A0[0]		# a[1]*a[0]
1333	 mov	$ai,%rax		# a[2]
1334	mov	%rdx,$A0[1]
1335	mov	$A0[0],-24($tptr,$i)	# t[1]
1336
1337	mul	$a0			# a[2]*a[0]
1338	add	%rax,$A0[1]
1339	 mov	$ai,%rax
1340	adc	\$0,%rdx
1341	mov	$A0[1],-16($tptr,$i)	# t[2]
1342	mov	%rdx,$A0[0]
1343
1344
1345	 mov	-8($aptr,$i),$ai	# a[3]
1346	mul	$a1			# a[2]*a[1]
1347	mov	%rax,$A1[0]		# a[2]*a[1]+t[3]
1348	 mov	$ai,%rax
1349	mov	%rdx,$A1[1]
1350
1351	 lea	($i),$j
1352	mul	$a0			# a[3]*a[0]
1353	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1354	 mov	$ai,%rax
1355	mov	%rdx,$A0[1]
1356	adc	\$0,$A0[1]
1357	add	$A1[0],$A0[0]
1358	adc	\$0,$A0[1]
1359	mov	$A0[0],-8($tptr,$j)	# t[3]
1360	jmp	.Lsqr4x_1st
1361
1362.align	32
1363.Lsqr4x_1st:
1364	 mov	($aptr,$j),$ai		# a[4]
1365	mul	$a1			# a[3]*a[1]
1366	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1367	 mov	$ai,%rax
1368	mov	%rdx,$A1[0]
1369	adc	\$0,$A1[0]
1370
1371	mul	$a0			# a[4]*a[0]
1372	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1373	 mov	$ai,%rax		# a[3]
1374	 mov	8($aptr,$j),$ai		# a[5]
1375	mov	%rdx,$A0[0]
1376	adc	\$0,$A0[0]
1377	add	$A1[1],$A0[1]
1378	adc	\$0,$A0[0]
1379
1380
1381	mul	$a1			# a[4]*a[3]
1382	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1383	 mov	$ai,%rax
1384	 mov	$A0[1],($tptr,$j)	# t[4]
1385	mov	%rdx,$A1[1]
1386	adc	\$0,$A1[1]
1387
1388	mul	$a0			# a[5]*a[2]
1389	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1390	 mov	$ai,%rax
1391	 mov	16($aptr,$j),$ai	# a[6]
1392	mov	%rdx,$A0[1]
1393	adc	\$0,$A0[1]
1394	add	$A1[0],$A0[0]
1395	adc	\$0,$A0[1]
1396
1397	mul	$a1			# a[5]*a[3]
1398	add	%rax,$A1[1]		# a[5]*a[3]+t[6]
1399	 mov	$ai,%rax
1400	 mov	$A0[0],8($tptr,$j)	# t[5]
1401	mov	%rdx,$A1[0]
1402	adc	\$0,$A1[0]
1403
1404	mul	$a0			# a[6]*a[2]
1405	add	%rax,$A0[1]		# a[6]*a[2]+a[5]*a[3]+t[6]
1406	 mov	$ai,%rax		# a[3]
1407	 mov	24($aptr,$j),$ai	# a[7]
1408	mov	%rdx,$A0[0]
1409	adc	\$0,$A0[0]
1410	add	$A1[1],$A0[1]
1411	adc	\$0,$A0[0]
1412
1413
1414	mul	$a1			# a[6]*a[5]
1415	add	%rax,$A1[0]		# a[6]*a[5]+t[7]
1416	 mov	$ai,%rax
1417	 mov	$A0[1],16($tptr,$j)	# t[6]
1418	mov	%rdx,$A1[1]
1419	adc	\$0,$A1[1]
1420	 lea	32($j),$j
1421
1422	mul	$a0			# a[7]*a[4]
1423	add	%rax,$A0[0]		# a[7]*a[4]+a[6]*a[5]+t[6]
1424	 mov	$ai,%rax
1425	mov	%rdx,$A0[1]
1426	adc	\$0,$A0[1]
1427	add	$A1[0],$A0[0]
1428	adc	\$0,$A0[1]
1429	mov	$A0[0],-8($tptr,$j)	# t[7]
1430
1431	cmp	\$0,$j
1432	jne	.Lsqr4x_1st
1433
1434	mul	$a1			# a[7]*a[5]
1435	add	%rax,$A1[1]
1436	lea	16($i),$i
1437	adc	\$0,%rdx
1438	add	$A0[1],$A1[1]
1439	adc	\$0,%rdx
1440
1441	mov	$A1[1],($tptr)		# t[8]
1442	mov	%rdx,$A1[0]
1443	mov	%rdx,8($tptr)		# t[9]
1444	jmp	.Lsqr4x_outer
1445
1446.align	32
1447.Lsqr4x_outer:				# comments apply to $num==6 case
1448	mov	-32($aptr,$i),$a0	# a[0]
1449	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1450	mov	-24($aptr,$i),%rax	# a[1]
1451	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1452	mov	-16($aptr,$i),$ai	# a[2]
1453	mov	%rax,$a1
1454
1455	mul	$a0			# a[1]*a[0]
1456	mov	-24($tptr,$i),$A0[0]	# t[1]
1457	add	%rax,$A0[0]		# a[1]*a[0]+t[1]
1458	 mov	$ai,%rax		# a[2]
1459	adc	\$0,%rdx
1460	mov	$A0[0],-24($tptr,$i)	# t[1]
1461	mov	%rdx,$A0[1]
1462
1463	mul	$a0			# a[2]*a[0]
1464	add	%rax,$A0[1]
1465	 mov	$ai,%rax
1466	adc	\$0,%rdx
1467	add	-16($tptr,$i),$A0[1]	# a[2]*a[0]+t[2]
1468	mov	%rdx,$A0[0]
1469	adc	\$0,$A0[0]
1470	mov	$A0[1],-16($tptr,$i)	# t[2]
1471
1472	xor	$A1[0],$A1[0]
1473
1474	 mov	-8($aptr,$i),$ai	# a[3]
1475	mul	$a1			# a[2]*a[1]
1476	add	%rax,$A1[0]		# a[2]*a[1]+t[3]
1477	 mov	$ai,%rax
1478	adc	\$0,%rdx
1479	add	-8($tptr,$i),$A1[0]
1480	mov	%rdx,$A1[1]
1481	adc	\$0,$A1[1]
1482
1483	mul	$a0			# a[3]*a[0]
1484	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1485	 mov	$ai,%rax
1486	adc	\$0,%rdx
1487	add	$A1[0],$A0[0]
1488	mov	%rdx,$A0[1]
1489	adc	\$0,$A0[1]
1490	mov	$A0[0],-8($tptr,$i)	# t[3]
1491
1492	lea	($i),$j
1493	jmp	.Lsqr4x_inner
1494
1495.align	32
1496.Lsqr4x_inner:
1497	 mov	($aptr,$j),$ai		# a[4]
1498	mul	$a1			# a[3]*a[1]
1499	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1500	 mov	$ai,%rax
1501	mov	%rdx,$A1[0]
1502	adc	\$0,$A1[0]
1503	add	($tptr,$j),$A1[1]
1504	adc	\$0,$A1[0]
1505
1506	.byte	0x67
1507	mul	$a0			# a[4]*a[0]
1508	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1509	 mov	$ai,%rax		# a[3]
1510	 mov	8($aptr,$j),$ai		# a[5]
1511	mov	%rdx,$A0[0]
1512	adc	\$0,$A0[0]
1513	add	$A1[1],$A0[1]
1514	adc	\$0,$A0[0]
1515
1516	mul	$a1			# a[4]*a[3]
1517	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1518	mov	$A0[1],($tptr,$j)	# t[4]
1519	 mov	$ai,%rax
1520	mov	%rdx,$A1[1]
1521	adc	\$0,$A1[1]
1522	add	8($tptr,$j),$A1[0]
1523	lea	16($j),$j		# j++
1524	adc	\$0,$A1[1]
1525
1526	mul	$a0			# a[5]*a[2]
1527	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1528	 mov	$ai,%rax
1529	adc	\$0,%rdx
1530	add	$A1[0],$A0[0]
1531	mov	%rdx,$A0[1]
1532	adc	\$0,$A0[1]
1533	mov	$A0[0],-8($tptr,$j)	# t[5], "preloaded t[1]" below
1534
1535	cmp	\$0,$j
1536	jne	.Lsqr4x_inner
1537
1538	.byte	0x67
1539	mul	$a1			# a[5]*a[3]
1540	add	%rax,$A1[1]
1541	adc	\$0,%rdx
1542	add	$A0[1],$A1[1]
1543	adc	\$0,%rdx
1544
1545	mov	$A1[1],($tptr)		# t[6], "preloaded t[2]" below
1546	mov	%rdx,$A1[0]
1547	mov	%rdx,8($tptr)		# t[7], "preloaded t[3]" below
1548
1549	add	\$16,$i
1550	jnz	.Lsqr4x_outer
1551
1552					# comments apply to $num==4 case
1553	mov	-32($aptr),$a0		# a[0]
1554	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1555	mov	-24($aptr),%rax		# a[1]
1556	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1557	mov	-16($aptr),$ai		# a[2]
1558	mov	%rax,$a1
1559
1560	mul	$a0			# a[1]*a[0]
1561	add	%rax,$A0[0]		# a[1]*a[0]+t[1], preloaded t[1]
1562	 mov	$ai,%rax		# a[2]
1563	mov	%rdx,$A0[1]
1564	adc	\$0,$A0[1]
1565
1566	mul	$a0			# a[2]*a[0]
1567	add	%rax,$A0[1]
1568	 mov	$ai,%rax
1569	 mov	$A0[0],-24($tptr)	# t[1]
1570	mov	%rdx,$A0[0]
1571	adc	\$0,$A0[0]
1572	add	$A1[1],$A0[1]		# a[2]*a[0]+t[2], preloaded t[2]
1573	 mov	-8($aptr),$ai		# a[3]
1574	adc	\$0,$A0[0]
1575
1576	mul	$a1			# a[2]*a[1]
1577	add	%rax,$A1[0]		# a[2]*a[1]+t[3], preloaded t[3]
1578	 mov	$ai,%rax
1579	 mov	$A0[1],-16($tptr)	# t[2]
1580	mov	%rdx,$A1[1]
1581	adc	\$0,$A1[1]
1582
1583	mul	$a0			# a[3]*a[0]
1584	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1585	 mov	$ai,%rax
1586	mov	%rdx,$A0[1]
1587	adc	\$0,$A0[1]
1588	add	$A1[0],$A0[0]
1589	adc	\$0,$A0[1]
1590	mov	$A0[0],-8($tptr)	# t[3]
1591
1592	mul	$a1			# a[3]*a[1]
1593	add	%rax,$A1[1]
1594	 mov	-16($aptr),%rax		# a[2]
1595	adc	\$0,%rdx
1596	add	$A0[1],$A1[1]
1597	adc	\$0,%rdx
1598
1599	mov	$A1[1],($tptr)		# t[4]
1600	mov	%rdx,$A1[0]
1601	mov	%rdx,8($tptr)		# t[5]
1602
1603	mul	$ai			# a[2]*a[3]
1604___
1605{
1606my ($shift,$carry)=($a0,$a1);
1607my @S=(@A1,$ai,$n0);
1608$code.=<<___;
1609	 add	\$16,$i
1610	 xor	$shift,$shift
1611	 sub	$num,$i			# $i=16-$num
1612	 xor	$carry,$carry
1613
1614	add	$A1[0],%rax		# t[5]
1615	adc	\$0,%rdx
1616	mov	%rax,8($tptr)		# t[5]
1617	mov	%rdx,16($tptr)		# t[6]
1618	mov	$carry,24($tptr)	# t[7]
1619
1620	 mov	-16($aptr,$i),%rax	# a[0]
1621	lea	48+8(%rsp),$tptr
1622	 xor	$A0[0],$A0[0]		# t[0]
1623	 mov	8($tptr),$A0[1]		# t[1]
1624
1625	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1626	shr	\$63,$A0[0]
1627	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1628	shr	\$63,$A0[1]
1629	or	$A0[0],$S[1]		# | t[2*i]>>63
1630	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1631	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1632	mul	%rax			# a[i]*a[i]
1633	neg	$carry			# mov $carry,cf
1634	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1635	adc	%rax,$S[0]
1636	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1637	mov	$S[0],($tptr)
1638	adc	%rdx,$S[1]
1639
1640	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1641	 mov	$S[1],8($tptr)
1642	 sbb	$carry,$carry		# mov cf,$carry
1643	shr	\$63,$A0[0]
1644	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1645	shr	\$63,$A0[1]
1646	or	$A0[0],$S[3]		# | t[2*i]>>63
1647	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1648	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1649	mul	%rax			# a[i]*a[i]
1650	neg	$carry			# mov $carry,cf
1651	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1652	adc	%rax,$S[2]
1653	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1654	mov	$S[2],16($tptr)
1655	adc	%rdx,$S[3]
1656	lea	16($i),$i
1657	mov	$S[3],24($tptr)
1658	sbb	$carry,$carry		# mov cf,$carry
1659	lea	64($tptr),$tptr
1660	jmp	.Lsqr4x_shift_n_add
1661
1662.align	32
1663.Lsqr4x_shift_n_add:
1664	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1665	shr	\$63,$A0[0]
1666	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1667	shr	\$63,$A0[1]
1668	or	$A0[0],$S[1]		# | t[2*i]>>63
1669	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1670	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1671	mul	%rax			# a[i]*a[i]
1672	neg	$carry			# mov $carry,cf
1673	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1674	adc	%rax,$S[0]
1675	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1676	mov	$S[0],-32($tptr)
1677	adc	%rdx,$S[1]
1678
1679	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1680	 mov	$S[1],-24($tptr)
1681	 sbb	$carry,$carry		# mov cf,$carry
1682	shr	\$63,$A0[0]
1683	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1684	shr	\$63,$A0[1]
1685	or	$A0[0],$S[3]		# | t[2*i]>>63
1686	 mov	0($tptr),$A0[0]		# t[2*i+2]	# prefetch
1687	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1688	mul	%rax			# a[i]*a[i]
1689	neg	$carry			# mov $carry,cf
1690	 mov	8($tptr),$A0[1]		# t[2*i+2+1]	# prefetch
1691	adc	%rax,$S[2]
1692	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1693	mov	$S[2],-16($tptr)
1694	adc	%rdx,$S[3]
1695
1696	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1697	 mov	$S[3],-8($tptr)
1698	 sbb	$carry,$carry		# mov cf,$carry
1699	shr	\$63,$A0[0]
1700	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1701	shr	\$63,$A0[1]
1702	or	$A0[0],$S[1]		# | t[2*i]>>63
1703	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1704	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1705	mul	%rax			# a[i]*a[i]
1706	neg	$carry			# mov $carry,cf
1707	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1708	adc	%rax,$S[0]
1709	 mov	8($aptr,$i),%rax	# a[i+1]	# prefetch
1710	mov	$S[0],0($tptr)
1711	adc	%rdx,$S[1]
1712
1713	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1714	 mov	$S[1],8($tptr)
1715	 sbb	$carry,$carry		# mov cf,$carry
1716	shr	\$63,$A0[0]
1717	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1718	shr	\$63,$A0[1]
1719	or	$A0[0],$S[3]		# | t[2*i]>>63
1720	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1721	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1722	mul	%rax			# a[i]*a[i]
1723	neg	$carry			# mov $carry,cf
1724	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1725	adc	%rax,$S[2]
1726	 mov	16($aptr,$i),%rax	# a[i+1]	# prefetch
1727	mov	$S[2],16($tptr)
1728	adc	%rdx,$S[3]
1729	mov	$S[3],24($tptr)
1730	sbb	$carry,$carry		# mov cf,$carry
1731	lea	64($tptr),$tptr
1732	add	\$32,$i
1733	jnz	.Lsqr4x_shift_n_add
1734
1735	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1736	.byte	0x67
1737	shr	\$63,$A0[0]
1738	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1739	shr	\$63,$A0[1]
1740	or	$A0[0],$S[1]		# | t[2*i]>>63
1741	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1742	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1743	mul	%rax			# a[i]*a[i]
1744	neg	$carry			# mov $carry,cf
1745	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1746	adc	%rax,$S[0]
1747	 mov	-8($aptr),%rax		# a[i+1]	# prefetch
1748	mov	$S[0],-32($tptr)
1749	adc	%rdx,$S[1]
1750
1751	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1|shift
1752	 mov	$S[1],-24($tptr)
1753	 sbb	$carry,$carry		# mov cf,$carry
1754	shr	\$63,$A0[0]
1755	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1756	shr	\$63,$A0[1]
1757	or	$A0[0],$S[3]		# | t[2*i]>>63
1758	mul	%rax			# a[i]*a[i]
1759	neg	$carry			# mov $carry,cf
1760	adc	%rax,$S[2]
1761	adc	%rdx,$S[3]
1762	mov	$S[2],-16($tptr)
1763	mov	$S[3],-8($tptr)
1764___
1765}
1766######################################################################
1767# Montgomery reduction part, "word-by-word" algorithm.
1768#
1769# This new path is inspired by multiple submissions from Intel, by
1770# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
1771# Vinodh Gopal...
1772{
1773my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx");
1774
1775$code.=<<___;
1776	movq	%xmm2,$nptr
1777__bn_sqr8x_reduction:
1778	xor	%rax,%rax
1779	lea	($nptr,$num),%rcx	# end of n[]
1780	lea	48+8(%rsp,$num,2),%rdx	# end of t[] buffer
1781	mov	%rcx,0+8(%rsp)
1782	lea	48+8(%rsp,$num),$tptr	# end of initial t[] window
1783	mov	%rdx,8+8(%rsp)
1784	neg	$num
1785	jmp	.L8x_reduction_loop
1786
1787.align	32
1788.L8x_reduction_loop:
1789	lea	($tptr,$num),$tptr	# start of current t[] window
1790	.byte	0x66
1791	mov	8*0($tptr),$m0
1792	mov	8*1($tptr),%r9
1793	mov	8*2($tptr),%r10
1794	mov	8*3($tptr),%r11
1795	mov	8*4($tptr),%r12
1796	mov	8*5($tptr),%r13
1797	mov	8*6($tptr),%r14
1798	mov	8*7($tptr),%r15
1799	mov	%rax,(%rdx)		# store top-most carry bit
1800	lea	8*8($tptr),$tptr
1801
1802	.byte	0x67
1803	mov	$m0,%r8
1804	imulq	32+8(%rsp),$m0		# n0*a[0]
1805	mov	8*0($nptr),%rax		# n[0]
1806	mov	\$8,%ecx
1807	jmp	.L8x_reduce
1808
1809.align	32
1810.L8x_reduce:
1811	mulq	$m0
1812	 mov	8*1($nptr),%rax		# n[1]
1813	neg	%r8
1814	mov	%rdx,%r8
1815	adc	\$0,%r8
1816
1817	mulq	$m0
1818	add	%rax,%r9
1819	 mov	8*2($nptr),%rax
1820	adc	\$0,%rdx
1821	add	%r9,%r8
1822	 mov	$m0,48-8+8(%rsp,%rcx,8)	# put aside n0*a[i]
1823	mov	%rdx,%r9
1824	adc	\$0,%r9
1825
1826	mulq	$m0
1827	add	%rax,%r10
1828	 mov	8*3($nptr),%rax
1829	adc	\$0,%rdx
1830	add	%r10,%r9
1831	 mov	32+8(%rsp),$carry	# pull n0, borrow $carry
1832	mov	%rdx,%r10
1833	adc	\$0,%r10
1834
1835	mulq	$m0
1836	add	%rax,%r11
1837	 mov	8*4($nptr),%rax
1838	adc	\$0,%rdx
1839	 imulq	%r8,$carry		# modulo-scheduled
1840	add	%r11,%r10
1841	mov	%rdx,%r11
1842	adc	\$0,%r11
1843
1844	mulq	$m0
1845	add	%rax,%r12
1846	 mov	8*5($nptr),%rax
1847	adc	\$0,%rdx
1848	add	%r12,%r11
1849	mov	%rdx,%r12
1850	adc	\$0,%r12
1851
1852	mulq	$m0
1853	add	%rax,%r13
1854	 mov	8*6($nptr),%rax
1855	adc	\$0,%rdx
1856	add	%r13,%r12
1857	mov	%rdx,%r13
1858	adc	\$0,%r13
1859
1860	mulq	$m0
1861	add	%rax,%r14
1862	 mov	8*7($nptr),%rax
1863	adc	\$0,%rdx
1864	add	%r14,%r13
1865	mov	%rdx,%r14
1866	adc	\$0,%r14
1867
1868	mulq	$m0
1869	 mov	$carry,$m0		# n0*a[i]
1870	add	%rax,%r15
1871	 mov	8*0($nptr),%rax		# n[0]
1872	adc	\$0,%rdx
1873	add	%r15,%r14
1874	mov	%rdx,%r15
1875	adc	\$0,%r15
1876
1877	dec	%ecx
1878	jnz	.L8x_reduce
1879
1880	lea	8*8($nptr),$nptr
1881	xor	%rax,%rax
1882	mov	8+8(%rsp),%rdx		# pull end of t[]
1883	cmp	0+8(%rsp),$nptr		# end of n[]?
1884	jae	.L8x_no_tail
1885
1886	.byte	0x66
1887	add	8*0($tptr),%r8
1888	adc	8*1($tptr),%r9
1889	adc	8*2($tptr),%r10
1890	adc	8*3($tptr),%r11
1891	adc	8*4($tptr),%r12
1892	adc	8*5($tptr),%r13
1893	adc	8*6($tptr),%r14
1894	adc	8*7($tptr),%r15
1895	sbb	$carry,$carry		# top carry
1896
1897	mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1898	mov	\$8,%ecx
1899	mov	8*0($nptr),%rax
1900	jmp	.L8x_tail
1901
1902.align	32
1903.L8x_tail:
1904	mulq	$m0
1905	add	%rax,%r8
1906	 mov	8*1($nptr),%rax
1907	 mov	%r8,($tptr)		# save result
1908	mov	%rdx,%r8
1909	adc	\$0,%r8
1910
1911	mulq	$m0
1912	add	%rax,%r9
1913	 mov	8*2($nptr),%rax
1914	adc	\$0,%rdx
1915	add	%r9,%r8
1916	 lea	8($tptr),$tptr		# $tptr++
1917	mov	%rdx,%r9
1918	adc	\$0,%r9
1919
1920	mulq	$m0
1921	add	%rax,%r10
1922	 mov	8*3($nptr),%rax
1923	adc	\$0,%rdx
1924	add	%r10,%r9
1925	mov	%rdx,%r10
1926	adc	\$0,%r10
1927
1928	mulq	$m0
1929	add	%rax,%r11
1930	 mov	8*4($nptr),%rax
1931	adc	\$0,%rdx
1932	add	%r11,%r10
1933	mov	%rdx,%r11
1934	adc	\$0,%r11
1935
1936	mulq	$m0
1937	add	%rax,%r12
1938	 mov	8*5($nptr),%rax
1939	adc	\$0,%rdx
1940	add	%r12,%r11
1941	mov	%rdx,%r12
1942	adc	\$0,%r12
1943
1944	mulq	$m0
1945	add	%rax,%r13
1946	 mov	8*6($nptr),%rax
1947	adc	\$0,%rdx
1948	add	%r13,%r12
1949	mov	%rdx,%r13
1950	adc	\$0,%r13
1951
1952	mulq	$m0
1953	add	%rax,%r14
1954	 mov	8*7($nptr),%rax
1955	adc	\$0,%rdx
1956	add	%r14,%r13
1957	mov	%rdx,%r14
1958	adc	\$0,%r14
1959
1960	mulq	$m0
1961	 mov	48-16+8(%rsp,%rcx,8),$m0# pull n0*a[i]
1962	add	%rax,%r15
1963	adc	\$0,%rdx
1964	add	%r15,%r14
1965	 mov	8*0($nptr),%rax		# pull n[0]
1966	mov	%rdx,%r15
1967	adc	\$0,%r15
1968
1969	dec	%ecx
1970	jnz	.L8x_tail
1971
1972	lea	8*8($nptr),$nptr
1973	mov	8+8(%rsp),%rdx		# pull end of t[]
1974	cmp	0+8(%rsp),$nptr		# end of n[]?
1975	jae	.L8x_tail_done		# break out of loop
1976
1977	 mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1978	neg	$carry
1979	 mov	8*0($nptr),%rax		# pull n[0]
1980	adc	8*0($tptr),%r8
1981	adc	8*1($tptr),%r9
1982	adc	8*2($tptr),%r10
1983	adc	8*3($tptr),%r11
1984	adc	8*4($tptr),%r12
1985	adc	8*5($tptr),%r13
1986	adc	8*6($tptr),%r14
1987	adc	8*7($tptr),%r15
1988	sbb	$carry,$carry		# top carry
1989
1990	mov	\$8,%ecx
1991	jmp	.L8x_tail
1992
1993.align	32
1994.L8x_tail_done:
1995	xor	%rax,%rax
1996	add	(%rdx),%r8		# can this overflow?
1997	adc	\$0,%r9
1998	adc	\$0,%r10
1999	adc	\$0,%r11
2000	adc	\$0,%r12
2001	adc	\$0,%r13
2002	adc	\$0,%r14
2003	adc	\$0,%r15
2004	adc	\$0,%rax
2005
2006	neg	$carry
2007.L8x_no_tail:
2008	adc	8*0($tptr),%r8
2009	adc	8*1($tptr),%r9
2010	adc	8*2($tptr),%r10
2011	adc	8*3($tptr),%r11
2012	adc	8*4($tptr),%r12
2013	adc	8*5($tptr),%r13
2014	adc	8*6($tptr),%r14
2015	adc	8*7($tptr),%r15
2016	adc	\$0,%rax		# top-most carry
2017	 mov	-8($nptr),%rcx		# np[num-1]
2018	 xor	$carry,$carry
2019
2020	movq	%xmm2,$nptr		# restore $nptr
2021
2022	mov	%r8,8*0($tptr)		# store top 512 bits
2023	mov	%r9,8*1($tptr)
2024	 movq	%xmm3,$num		# $num is %r9, can't be moved upwards
2025	mov	%r10,8*2($tptr)
2026	mov	%r11,8*3($tptr)
2027	mov	%r12,8*4($tptr)
2028	mov	%r13,8*5($tptr)
2029	mov	%r14,8*6($tptr)
2030	mov	%r15,8*7($tptr)
2031	lea	8*8($tptr),$tptr
2032
2033	cmp	%rdx,$tptr		# end of t[]?
2034	jb	.L8x_reduction_loop
2035	ret
2036.cfi_endproc
2037.size	bn_sqr8x_internal,.-bn_sqr8x_internal
2038___
2039}
2040##############################################################
2041# Post-condition, 4x unrolled
2042#
2043{
2044my ($tptr,$nptr)=("%rbx","%rbp");
2045$code.=<<___;
2046.type	__bn_post4x_internal,\@abi-omnipotent
2047.align	32
2048__bn_post4x_internal:
2049.cfi_startproc
2050	mov	8*0($nptr),%r12
2051	lea	(%rdi,$num),$tptr	# %rdi was $tptr above
2052	mov	$num,%rcx
2053	movq	%xmm1,$rptr		# restore $rptr
2054	neg	%rax
2055	movq	%xmm1,$aptr		# prepare for back-to-back call
2056	sar	\$3+2,%rcx
2057	dec	%r12			# so that after 'not' we get -n[0]
2058	xor	%r10,%r10
2059	mov	8*1($nptr),%r13
2060	mov	8*2($nptr),%r14
2061	mov	8*3($nptr),%r15
2062	jmp	.Lsqr4x_sub_entry
2063
2064.align	16
2065.Lsqr4x_sub:
2066	mov	8*0($nptr),%r12
2067	mov	8*1($nptr),%r13
2068	mov	8*2($nptr),%r14
2069	mov	8*3($nptr),%r15
2070.Lsqr4x_sub_entry:
2071	lea	8*4($nptr),$nptr
2072	not	%r12
2073	not	%r13
2074	not	%r14
2075	not	%r15
2076	and	%rax,%r12
2077	and	%rax,%r13
2078	and	%rax,%r14
2079	and	%rax,%r15
2080
2081	neg	%r10			# mov %r10,%cf
2082	adc	8*0($tptr),%r12
2083	adc	8*1($tptr),%r13
2084	adc	8*2($tptr),%r14
2085	adc	8*3($tptr),%r15
2086	mov	%r12,8*0($rptr)
2087	lea	8*4($tptr),$tptr
2088	mov	%r13,8*1($rptr)
2089	sbb	%r10,%r10		# mov %cf,%r10
2090	mov	%r14,8*2($rptr)
2091	mov	%r15,8*3($rptr)
2092	lea	8*4($rptr),$rptr
2093
2094	inc	%rcx			# pass %cf
2095	jnz	.Lsqr4x_sub
2096
2097	mov	$num,%r10		# prepare for back-to-back call
2098	neg	$num			# restore $num
2099	ret
2100.cfi_endproc
2101.size	__bn_post4x_internal,.-__bn_post4x_internal
2102___
2103}
2104}}}
2105
2106if ($addx) {{{
2107my $bp="%rdx";	# restore original value
2108
2109$code.=<<___;
2110.type	bn_mulx4x_mont_gather5,\@function,6
2111.align	32
2112bn_mulx4x_mont_gather5:
2113.cfi_startproc
2114	mov	%rsp,%rax
2115.cfi_def_cfa_register	%rax
2116.Lmulx4x_enter:
2117	push	%rbx
2118.cfi_push	%rbx
2119	push	%rbp
2120.cfi_push	%rbp
2121	push	%r12
2122.cfi_push	%r12
2123	push	%r13
2124.cfi_push	%r13
2125	push	%r14
2126.cfi_push	%r14
2127	push	%r15
2128.cfi_push	%r15
2129.Lmulx4x_prologue:
2130
2131	shl	\$3,${num}d		# convert $num to bytes
2132	lea	($num,$num,2),%r10	# 3*$num in bytes
2133	neg	$num			# -$num
2134	mov	($n0),$n0		# *n0
2135
2136	##############################################################
2137	# Ensure that stack frame doesn't alias with $rptr+3*$num
2138	# modulo 4096, which covers ret[num], am[num] and n[num]
2139	# (see bn_exp.c). This is done to allow memory disambiguation
2140	# logic do its magic. [Extra [num] is allocated in order
2141	# to align with bn_power5's frame, which is cleansed after
2142	# completing exponentiation. Extra 256 bytes is for power mask
2143	# calculated from 7th argument, the index.]
2144	#
2145	lea	-320(%rsp,$num,2),%r11
2146	mov	%rsp,%rbp
2147	sub	$rp,%r11
2148	and	\$4095,%r11
2149	cmp	%r11,%r10
2150	jb	.Lmulx4xsp_alt
2151	sub	%r11,%rbp		# align with $aptr
2152	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2153	jmp	.Lmulx4xsp_done
2154
2155.Lmulx4xsp_alt:
2156	lea	4096-320(,$num,2),%r10
2157	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2158	sub	%r10,%r11
2159	mov	\$0,%r10
2160	cmovc	%r10,%r11
2161	sub	%r11,%rbp
2162.Lmulx4xsp_done:
2163	and	\$-64,%rbp		# ensure alignment
2164	mov	%rsp,%r11
2165	sub	%rbp,%r11
2166	and	\$-4096,%r11
2167	lea	(%rbp,%r11),%rsp
2168	mov	(%rsp),%r10
2169	cmp	%rbp,%rsp
2170	ja	.Lmulx4x_page_walk
2171	jmp	.Lmulx4x_page_walk_done
2172
2173.Lmulx4x_page_walk:
2174	lea	-4096(%rsp),%rsp
2175	mov	(%rsp),%r10
2176	cmp	%rbp,%rsp
2177	ja	.Lmulx4x_page_walk
2178.Lmulx4x_page_walk_done:
2179
2180	##############################################################
2181	# Stack layout
2182	# +0	-num
2183	# +8	off-loaded &b[i]
2184	# +16	end of b[num]
2185	# +24	inner counter
2186	# +32	saved n0
2187	# +40	saved %rsp
2188	# +48
2189	# +56	saved rp
2190	# +64	tmp[num+1]
2191	#
2192	mov	$n0, 32(%rsp)		# save *n0
2193	mov	%rax,40(%rsp)		# save original %rsp
2194.cfi_cfa_expression	%rsp+40,deref,+8
2195.Lmulx4x_body:
2196	call	mulx4x_internal
2197
2198	mov	40(%rsp),%rsi		# restore %rsp
2199.cfi_def_cfa	%rsi,8
2200	mov	\$1,%rax
2201
2202	mov	-48(%rsi),%r15
2203.cfi_restore	%r15
2204	mov	-40(%rsi),%r14
2205.cfi_restore	%r14
2206	mov	-32(%rsi),%r13
2207.cfi_restore	%r13
2208	mov	-24(%rsi),%r12
2209.cfi_restore	%r12
2210	mov	-16(%rsi),%rbp
2211.cfi_restore	%rbp
2212	mov	-8(%rsi),%rbx
2213.cfi_restore	%rbx
2214	lea	(%rsi),%rsp
2215.cfi_def_cfa_register	%rsp
2216.Lmulx4x_epilogue:
2217	ret
2218.cfi_endproc
2219.size	bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5
2220
2221.type	mulx4x_internal,\@abi-omnipotent
2222.align	32
2223mulx4x_internal:
2224.cfi_startproc
2225	mov	$num,8(%rsp)		# save -$num (it was in bytes)
2226	mov	$num,%r10
2227	neg	$num			# restore $num
2228	shl	\$5,$num
2229	neg	%r10			# restore $num
2230	lea	128($bp,$num),%r13	# end of powers table (+size optimization)
2231	shr	\$5+5,$num
2232	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument
2233	sub	\$1,$num
2234	lea	.Linc(%rip),%rax
2235	mov	%r13,16+8(%rsp)		# end of b[num]
2236	mov	$num,24+8(%rsp)		# inner counter
2237	mov	$rp, 56+8(%rsp)		# save $rp
2238___
2239my ($aptr, $bptr, $nptr, $tptr, $mi,  $bi,  $zero, $num)=
2240   ("%rsi","%rdi","%rcx","%rbx","%r8","%r9","%rbp","%rax");
2241my $rptr=$bptr;
2242my $STRIDE=2**5*8;		# 5 is "window size"
2243my $N=$STRIDE/4;		# should match cache line size
2244$code.=<<___;
2245	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
2246	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
2247	lea	88-112(%rsp,%r10),%r10	# place the mask after tp[num+1] (+ICache optimization)
2248	lea	128($bp),$bptr		# size optimization
2249
2250	pshufd	\$0,%xmm5,%xmm5		# broadcast index
2251	movdqa	%xmm1,%xmm4
2252	.byte	0x67
2253	movdqa	%xmm1,%xmm2
2254___
2255########################################################################
2256# calculate mask by comparing 0..31 to index and save result to stack
2257#
2258$code.=<<___;
2259	.byte	0x67
2260	paddd	%xmm0,%xmm1
2261	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
2262	movdqa	%xmm4,%xmm3
2263___
2264for($i=0;$i<$STRIDE/16-4;$i+=4) {
2265$code.=<<___;
2266	paddd	%xmm1,%xmm2
2267	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
2268	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2269	movdqa	%xmm4,%xmm0
2270
2271	paddd	%xmm2,%xmm3
2272	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
2273	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2274	movdqa	%xmm4,%xmm1
2275
2276	paddd	%xmm3,%xmm0
2277	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
2278	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2279	movdqa	%xmm4,%xmm2
2280
2281	paddd	%xmm0,%xmm1
2282	pcmpeqd	%xmm5,%xmm0
2283	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2284	movdqa	%xmm4,%xmm3
2285___
2286}
2287$code.=<<___;				# last iteration can be optimized
2288	.byte	0x67
2289	paddd	%xmm1,%xmm2
2290	pcmpeqd	%xmm5,%xmm1
2291	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2292
2293	paddd	%xmm2,%xmm3
2294	pcmpeqd	%xmm5,%xmm2
2295	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2296
2297	pcmpeqd	%xmm5,%xmm3
2298	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2299
2300	pand	`16*($i+0)-128`($bptr),%xmm0	# while it's still in register
2301	pand	`16*($i+1)-128`($bptr),%xmm1
2302	pand	`16*($i+2)-128`($bptr),%xmm2
2303	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2304	pand	`16*($i+3)-128`($bptr),%xmm3
2305	por	%xmm2,%xmm0
2306	por	%xmm3,%xmm1
2307___
2308for($i=0;$i<$STRIDE/16-4;$i+=4) {
2309$code.=<<___;
2310	movdqa	`16*($i+0)-128`($bptr),%xmm4
2311	movdqa	`16*($i+1)-128`($bptr),%xmm5
2312	movdqa	`16*($i+2)-128`($bptr),%xmm2
2313	pand	`16*($i+0)+112`(%r10),%xmm4
2314	movdqa	`16*($i+3)-128`($bptr),%xmm3
2315	pand	`16*($i+1)+112`(%r10),%xmm5
2316	por	%xmm4,%xmm0
2317	pand	`16*($i+2)+112`(%r10),%xmm2
2318	por	%xmm5,%xmm1
2319	pand	`16*($i+3)+112`(%r10),%xmm3
2320	por	%xmm2,%xmm0
2321	por	%xmm3,%xmm1
2322___
2323}
2324$code.=<<___;
2325	pxor	%xmm1,%xmm0
2326	pshufd	\$0x4e,%xmm0,%xmm1
2327	por	%xmm1,%xmm0
2328	lea	$STRIDE($bptr),$bptr
2329	movq	%xmm0,%rdx		# bp[0]
2330	lea	64+8*4+8(%rsp),$tptr
2331
2332	mov	%rdx,$bi
2333	mulx	0*8($aptr),$mi,%rax	# a[0]*b[0]
2334	mulx	1*8($aptr),%r11,%r12	# a[1]*b[0]
2335	add	%rax,%r11
2336	mulx	2*8($aptr),%rax,%r13	# ...
2337	adc	%rax,%r12
2338	adc	\$0,%r13
2339	mulx	3*8($aptr),%rax,%r14
2340
2341	mov	$mi,%r15
2342	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2343	xor	$zero,$zero		# cf=0, of=0
2344	mov	$mi,%rdx
2345
2346	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2347
2348	lea	4*8($aptr),$aptr
2349	adcx	%rax,%r13
2350	adcx	$zero,%r14		# cf=0
2351
2352	mulx	0*8($nptr),%rax,%r10
2353	adcx	%rax,%r15		# discarded
2354	adox	%r11,%r10
2355	mulx	1*8($nptr),%rax,%r11
2356	adcx	%rax,%r10
2357	adox	%r12,%r11
2358	mulx	2*8($nptr),%rax,%r12
2359	mov	24+8(%rsp),$bptr	# counter value
2360	mov	%r10,-8*4($tptr)
2361	adcx	%rax,%r11
2362	adox	%r13,%r12
2363	mulx	3*8($nptr),%rax,%r15
2364	 mov	$bi,%rdx
2365	mov	%r11,-8*3($tptr)
2366	adcx	%rax,%r12
2367	adox	$zero,%r15		# of=0
2368	lea	4*8($nptr),$nptr
2369	mov	%r12,-8*2($tptr)
2370	jmp	.Lmulx4x_1st
2371
2372.align	32
2373.Lmulx4x_1st:
2374	adcx	$zero,%r15		# cf=0, modulo-scheduled
2375	mulx	0*8($aptr),%r10,%rax	# a[4]*b[0]
2376	adcx	%r14,%r10
2377	mulx	1*8($aptr),%r11,%r14	# a[5]*b[0]
2378	adcx	%rax,%r11
2379	mulx	2*8($aptr),%r12,%rax	# ...
2380	adcx	%r14,%r12
2381	mulx	3*8($aptr),%r13,%r14
2382	 .byte	0x67,0x67
2383	 mov	$mi,%rdx
2384	adcx	%rax,%r13
2385	adcx	$zero,%r14		# cf=0
2386	lea	4*8($aptr),$aptr
2387	lea	4*8($tptr),$tptr
2388
2389	adox	%r15,%r10
2390	mulx	0*8($nptr),%rax,%r15
2391	adcx	%rax,%r10
2392	adox	%r15,%r11
2393	mulx	1*8($nptr),%rax,%r15
2394	adcx	%rax,%r11
2395	adox	%r15,%r12
2396	mulx	2*8($nptr),%rax,%r15
2397	mov	%r10,-5*8($tptr)
2398	adcx	%rax,%r12
2399	mov	%r11,-4*8($tptr)
2400	adox	%r15,%r13
2401	mulx	3*8($nptr),%rax,%r15
2402	 mov	$bi,%rdx
2403	mov	%r12,-3*8($tptr)
2404	adcx	%rax,%r13
2405	adox	$zero,%r15
2406	lea	4*8($nptr),$nptr
2407	mov	%r13,-2*8($tptr)
2408
2409	dec	$bptr			# of=0, pass cf
2410	jnz	.Lmulx4x_1st
2411
2412	mov	8(%rsp),$num		# load -num
2413	adc	$zero,%r15		# modulo-scheduled
2414	lea	($aptr,$num),$aptr	# rewind $aptr
2415	add	%r15,%r14
2416	mov	8+8(%rsp),$bptr		# re-load &b[i]
2417	adc	$zero,$zero		# top-most carry
2418	mov	%r14,-1*8($tptr)
2419	jmp	.Lmulx4x_outer
2420
2421.align	32
2422.Lmulx4x_outer:
2423	lea	16-256($tptr),%r10	# where 256-byte mask is (+density control)
2424	pxor	%xmm4,%xmm4
2425	.byte	0x67,0x67
2426	pxor	%xmm5,%xmm5
2427___
2428for($i=0;$i<$STRIDE/16;$i+=4) {
2429$code.=<<___;
2430	movdqa	`16*($i+0)-128`($bptr),%xmm0
2431	movdqa	`16*($i+1)-128`($bptr),%xmm1
2432	movdqa	`16*($i+2)-128`($bptr),%xmm2
2433	pand	`16*($i+0)+256`(%r10),%xmm0
2434	movdqa	`16*($i+3)-128`($bptr),%xmm3
2435	pand	`16*($i+1)+256`(%r10),%xmm1
2436	por	%xmm0,%xmm4
2437	pand	`16*($i+2)+256`(%r10),%xmm2
2438	por	%xmm1,%xmm5
2439	pand	`16*($i+3)+256`(%r10),%xmm3
2440	por	%xmm2,%xmm4
2441	por	%xmm3,%xmm5
2442___
2443}
2444$code.=<<___;
2445	por	%xmm5,%xmm4
2446	pshufd	\$0x4e,%xmm4,%xmm0
2447	por	%xmm4,%xmm0
2448	lea	$STRIDE($bptr),$bptr
2449	movq	%xmm0,%rdx		# m0=bp[i]
2450
2451	mov	$zero,($tptr)		# save top-most carry
2452	lea	4*8($tptr,$num),$tptr	# rewind $tptr
2453	mulx	0*8($aptr),$mi,%r11	# a[0]*b[i]
2454	xor	$zero,$zero		# cf=0, of=0
2455	mov	%rdx,$bi
2456	mulx	1*8($aptr),%r14,%r12	# a[1]*b[i]
2457	adox	-4*8($tptr),$mi		# +t[0]
2458	adcx	%r14,%r11
2459	mulx	2*8($aptr),%r15,%r13	# ...
2460	adox	-3*8($tptr),%r11
2461	adcx	%r15,%r12
2462	mulx	3*8($aptr),%rdx,%r14
2463	adox	-2*8($tptr),%r12
2464	adcx	%rdx,%r13
2465	lea	($nptr,$num),$nptr	# rewind $nptr
2466	lea	4*8($aptr),$aptr
2467	adox	-1*8($tptr),%r13
2468	adcx	$zero,%r14
2469	adox	$zero,%r14
2470
2471	mov	$mi,%r15
2472	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2473
2474	mov	$mi,%rdx
2475	xor	$zero,$zero		# cf=0, of=0
2476	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2477
2478	mulx	0*8($nptr),%rax,%r10
2479	adcx	%rax,%r15		# discarded
2480	adox	%r11,%r10
2481	mulx	1*8($nptr),%rax,%r11
2482	adcx	%rax,%r10
2483	adox	%r12,%r11
2484	mulx	2*8($nptr),%rax,%r12
2485	adcx	%rax,%r11
2486	adox	%r13,%r12
2487	mulx	3*8($nptr),%rax,%r15
2488	 mov	$bi,%rdx
2489	mov	24+8(%rsp),$bptr	# counter value
2490	mov	%r10,-8*4($tptr)
2491	adcx	%rax,%r12
2492	mov	%r11,-8*3($tptr)
2493	adox	$zero,%r15		# of=0
2494	mov	%r12,-8*2($tptr)
2495	lea	4*8($nptr),$nptr
2496	jmp	.Lmulx4x_inner
2497
2498.align	32
2499.Lmulx4x_inner:
2500	mulx	0*8($aptr),%r10,%rax	# a[4]*b[i]
2501	adcx	$zero,%r15		# cf=0, modulo-scheduled
2502	adox	%r14,%r10
2503	mulx	1*8($aptr),%r11,%r14	# a[5]*b[i]
2504	adcx	0*8($tptr),%r10
2505	adox	%rax,%r11
2506	mulx	2*8($aptr),%r12,%rax	# ...
2507	adcx	1*8($tptr),%r11
2508	adox	%r14,%r12
2509	mulx	3*8($aptr),%r13,%r14
2510	 mov	$mi,%rdx
2511	adcx	2*8($tptr),%r12
2512	adox	%rax,%r13
2513	adcx	3*8($tptr),%r13
2514	adox	$zero,%r14		# of=0
2515	lea	4*8($aptr),$aptr
2516	lea	4*8($tptr),$tptr
2517	adcx	$zero,%r14		# cf=0
2518
2519	adox	%r15,%r10
2520	mulx	0*8($nptr),%rax,%r15
2521	adcx	%rax,%r10
2522	adox	%r15,%r11
2523	mulx	1*8($nptr),%rax,%r15
2524	adcx	%rax,%r11
2525	adox	%r15,%r12
2526	mulx	2*8($nptr),%rax,%r15
2527	mov	%r10,-5*8($tptr)
2528	adcx	%rax,%r12
2529	adox	%r15,%r13
2530	mov	%r11,-4*8($tptr)
2531	mulx	3*8($nptr),%rax,%r15
2532	 mov	$bi,%rdx
2533	lea	4*8($nptr),$nptr
2534	mov	%r12,-3*8($tptr)
2535	adcx	%rax,%r13
2536	adox	$zero,%r15
2537	mov	%r13,-2*8($tptr)
2538
2539	dec	$bptr			# of=0, pass cf
2540	jnz	.Lmulx4x_inner
2541
2542	mov	0+8(%rsp),$num		# load -num
2543	adc	$zero,%r15		# modulo-scheduled
2544	sub	0*8($tptr),$bptr	# pull top-most carry to %cf
2545	mov	8+8(%rsp),$bptr		# re-load &b[i]
2546	mov	16+8(%rsp),%r10
2547	adc	%r15,%r14
2548	lea	($aptr,$num),$aptr	# rewind $aptr
2549	adc	$zero,$zero		# top-most carry
2550	mov	%r14,-1*8($tptr)
2551
2552	cmp	%r10,$bptr
2553	jb	.Lmulx4x_outer
2554
2555	mov	-8($nptr),%r10
2556	mov	$zero,%r8
2557	mov	($nptr,$num),%r12
2558	lea	($nptr,$num),%rbp	# rewind $nptr
2559	mov	$num,%rcx
2560	lea	($tptr,$num),%rdi	# rewind $tptr
2561	xor	%eax,%eax
2562	xor	%r15,%r15
2563	sub	%r14,%r10		# compare top-most words
2564	adc	%r15,%r15
2565	or	%r15,%r8
2566	sar	\$3+2,%rcx
2567	sub	%r8,%rax		# %rax=-%r8
2568	mov	56+8(%rsp),%rdx		# restore rp
2569	dec	%r12			# so that after 'not' we get -n[0]
2570	mov	8*1(%rbp),%r13
2571	xor	%r8,%r8
2572	mov	8*2(%rbp),%r14
2573	mov	8*3(%rbp),%r15
2574	jmp	.Lsqrx4x_sub_entry	# common post-condition
2575.cfi_endproc
2576.size	mulx4x_internal,.-mulx4x_internal
2577___
2578}{
2579######################################################################
2580# void bn_power5(
2581my $rptr="%rdi";	# BN_ULONG *rptr,
2582my $aptr="%rsi";	# const BN_ULONG *aptr,
2583my $bptr="%rdx";	# const void *table,
2584my $nptr="%rcx";	# const BN_ULONG *nptr,
2585my $n0  ="%r8";		# const BN_ULONG *n0);
2586my $num ="%r9";		# int num, has to be divisible by 8
2587			# int pwr);
2588
2589my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
2590my @A0=("%r10","%r11");
2591my @A1=("%r12","%r13");
2592my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
2593
2594$code.=<<___;
2595.type	bn_powerx5,\@function,6
2596.align	32
2597bn_powerx5:
2598.cfi_startproc
2599	mov	%rsp,%rax
2600.cfi_def_cfa_register	%rax
2601.Lpowerx5_enter:
2602	push	%rbx
2603.cfi_push	%rbx
2604	push	%rbp
2605.cfi_push	%rbp
2606	push	%r12
2607.cfi_push	%r12
2608	push	%r13
2609.cfi_push	%r13
2610	push	%r14
2611.cfi_push	%r14
2612	push	%r15
2613.cfi_push	%r15
2614.Lpowerx5_prologue:
2615
2616	shl	\$3,${num}d		# convert $num to bytes
2617	lea	($num,$num,2),%r10	# 3*$num in bytes
2618	neg	$num
2619	mov	($n0),$n0		# *n0
2620
2621	##############################################################
2622	# Ensure that stack frame doesn't alias with $rptr+3*$num
2623	# modulo 4096, which covers ret[num], am[num] and n[num]
2624	# (see bn_exp.c). This is done to allow memory disambiguation
2625	# logic do its magic. [Extra 256 bytes is for power mask
2626	# calculated from 7th argument, the index.]
2627	#
2628	lea	-320(%rsp,$num,2),%r11
2629	mov	%rsp,%rbp
2630	sub	$rptr,%r11
2631	and	\$4095,%r11
2632	cmp	%r11,%r10
2633	jb	.Lpwrx_sp_alt
2634	sub	%r11,%rbp		# align with $aptr
2635	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2636	jmp	.Lpwrx_sp_done
2637
2638.align	32
2639.Lpwrx_sp_alt:
2640	lea	4096-320(,$num,2),%r10
2641	lea	-320(%rbp,$num,2),%rbp	# alloca(frame+2*$num*8+256)
2642	sub	%r10,%r11
2643	mov	\$0,%r10
2644	cmovc	%r10,%r11
2645	sub	%r11,%rbp
2646.Lpwrx_sp_done:
2647	and	\$-64,%rbp
2648	mov	%rsp,%r11
2649	sub	%rbp,%r11
2650	and	\$-4096,%r11
2651	lea	(%rbp,%r11),%rsp
2652	mov	(%rsp),%r10
2653	cmp	%rbp,%rsp
2654	ja	.Lpwrx_page_walk
2655	jmp	.Lpwrx_page_walk_done
2656
2657.Lpwrx_page_walk:
2658	lea	-4096(%rsp),%rsp
2659	mov	(%rsp),%r10
2660	cmp	%rbp,%rsp
2661	ja	.Lpwrx_page_walk
2662.Lpwrx_page_walk_done:
2663
2664	mov	$num,%r10
2665	neg	$num
2666
2667	##############################################################
2668	# Stack layout
2669	#
2670	# +0	saved $num, used in reduction section
2671	# +8	&t[2*$num], used in reduction section
2672	# +16	intermediate carry bit
2673	# +24	top-most carry bit, used in reduction section
2674	# +32	saved *n0
2675	# +40	saved %rsp
2676	# +48	t[2*$num]
2677	#
2678	pxor	%xmm0,%xmm0
2679	movq	$rptr,%xmm1		# save $rptr
2680	movq	$nptr,%xmm2		# save $nptr
2681	movq	%r10, %xmm3		# -$num
2682	movq	$bptr,%xmm4
2683	mov	$n0,  32(%rsp)
2684	mov	%rax, 40(%rsp)		# save original %rsp
2685.cfi_cfa_expression	%rsp+40,deref,+8
2686.Lpowerx5_body:
2687
2688	call	__bn_sqrx8x_internal
2689	call	__bn_postx4x_internal
2690	call	__bn_sqrx8x_internal
2691	call	__bn_postx4x_internal
2692	call	__bn_sqrx8x_internal
2693	call	__bn_postx4x_internal
2694	call	__bn_sqrx8x_internal
2695	call	__bn_postx4x_internal
2696	call	__bn_sqrx8x_internal
2697	call	__bn_postx4x_internal
2698
2699	mov	%r10,$num		# -num
2700	mov	$aptr,$rptr
2701	movq	%xmm2,$nptr
2702	movq	%xmm4,$bptr
2703	mov	40(%rsp),%rax
2704
2705	call	mulx4x_internal
2706
2707	mov	40(%rsp),%rsi		# restore %rsp
2708.cfi_def_cfa	%rsi,8
2709	mov	\$1,%rax
2710
2711	mov	-48(%rsi),%r15
2712.cfi_restore	%r15
2713	mov	-40(%rsi),%r14
2714.cfi_restore	%r14
2715	mov	-32(%rsi),%r13
2716.cfi_restore	%r13
2717	mov	-24(%rsi),%r12
2718.cfi_restore	%r12
2719	mov	-16(%rsi),%rbp
2720.cfi_restore	%rbp
2721	mov	-8(%rsi),%rbx
2722.cfi_restore	%rbx
2723	lea	(%rsi),%rsp
2724.cfi_def_cfa_register	%rsp
2725.Lpowerx5_epilogue:
2726	ret
2727.cfi_endproc
2728.size	bn_powerx5,.-bn_powerx5
2729
2730.globl	bn_sqrx8x_internal
2731.hidden	bn_sqrx8x_internal
2732.type	bn_sqrx8x_internal,\@abi-omnipotent
2733.align	32
2734bn_sqrx8x_internal:
2735__bn_sqrx8x_internal:
2736.cfi_startproc
2737	##################################################################
2738	# Squaring part:
2739	#
2740	# a) multiply-n-add everything but a[i]*a[i];
2741	# b) shift result of a) by 1 to the left and accumulate
2742	#    a[i]*a[i] products;
2743	#
2744	##################################################################
2745	# a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0]
2746	#                                                     a[1]a[0]
2747	#                                                 a[2]a[0]
2748	#                                             a[3]a[0]
2749	#                                             a[2]a[1]
2750	#                                         a[3]a[1]
2751	#                                     a[3]a[2]
2752	#
2753	#                                         a[4]a[0]
2754	#                                     a[5]a[0]
2755	#                                 a[6]a[0]
2756	#                             a[7]a[0]
2757	#                                     a[4]a[1]
2758	#                                 a[5]a[1]
2759	#                             a[6]a[1]
2760	#                         a[7]a[1]
2761	#                                 a[4]a[2]
2762	#                             a[5]a[2]
2763	#                         a[6]a[2]
2764	#                     a[7]a[2]
2765	#                             a[4]a[3]
2766	#                         a[5]a[3]
2767	#                     a[6]a[3]
2768	#                 a[7]a[3]
2769	#
2770	#                     a[5]a[4]
2771	#                 a[6]a[4]
2772	#             a[7]a[4]
2773	#             a[6]a[5]
2774	#         a[7]a[5]
2775	#     a[7]a[6]
2776	# a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0]
2777___
2778{
2779my ($zero,$carry)=("%rbp","%rcx");
2780my $aaptr=$zero;
2781$code.=<<___;
2782	lea	48+8(%rsp),$tptr
2783	lea	($aptr,$num),$aaptr
2784	mov	$num,0+8(%rsp)			# save $num
2785	mov	$aaptr,8+8(%rsp)		# save end of $aptr
2786	jmp	.Lsqr8x_zero_start
2787
2788.align	32
2789.byte	0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
2790.Lsqrx8x_zero:
2791	.byte	0x3e
2792	movdqa	%xmm0,0*8($tptr)
2793	movdqa	%xmm0,2*8($tptr)
2794	movdqa	%xmm0,4*8($tptr)
2795	movdqa	%xmm0,6*8($tptr)
2796.Lsqr8x_zero_start:			# aligned at 32
2797	movdqa	%xmm0,8*8($tptr)
2798	movdqa	%xmm0,10*8($tptr)
2799	movdqa	%xmm0,12*8($tptr)
2800	movdqa	%xmm0,14*8($tptr)
2801	lea	16*8($tptr),$tptr
2802	sub	\$64,$num
2803	jnz	.Lsqrx8x_zero
2804
2805	mov	0*8($aptr),%rdx		# a[0], modulo-scheduled
2806	#xor	%r9,%r9			# t[1], ex-$num, zero already
2807	xor	%r10,%r10
2808	xor	%r11,%r11
2809	xor	%r12,%r12
2810	xor	%r13,%r13
2811	xor	%r14,%r14
2812	xor	%r15,%r15
2813	lea	48+8(%rsp),$tptr
2814	xor	$zero,$zero		# cf=0, cf=0
2815	jmp	.Lsqrx8x_outer_loop
2816
2817.align	32
2818.Lsqrx8x_outer_loop:
2819	mulx	1*8($aptr),%r8,%rax	# a[1]*a[0]
2820	adcx	%r9,%r8			# a[1]*a[0]+=t[1]
2821	adox	%rax,%r10
2822	mulx	2*8($aptr),%r9,%rax	# a[2]*a[0]
2823	adcx	%r10,%r9
2824	adox	%rax,%r11
2825	.byte	0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00	# mulx	3*8($aptr),%r10,%rax	# ...
2826	adcx	%r11,%r10
2827	adox	%rax,%r12
2828	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00	# mulx	4*8($aptr),%r11,%rax
2829	adcx	%r12,%r11
2830	adox	%rax,%r13
2831	mulx	5*8($aptr),%r12,%rax
2832	adcx	%r13,%r12
2833	adox	%rax,%r14
2834	mulx	6*8($aptr),%r13,%rax
2835	adcx	%r14,%r13
2836	adox	%r15,%rax
2837	mulx	7*8($aptr),%r14,%r15
2838	 mov	1*8($aptr),%rdx		# a[1]
2839	adcx	%rax,%r14
2840	adox	$zero,%r15
2841	adc	8*8($tptr),%r15
2842	mov	%r8,1*8($tptr)		# t[1]
2843	mov	%r9,2*8($tptr)		# t[2]
2844	sbb	$carry,$carry		# mov %cf,$carry
2845	xor	$zero,$zero		# cf=0, of=0
2846
2847
2848	mulx	2*8($aptr),%r8,%rbx	# a[2]*a[1]
2849	mulx	3*8($aptr),%r9,%rax	# a[3]*a[1]
2850	adcx	%r10,%r8
2851	adox	%rbx,%r9
2852	mulx	4*8($aptr),%r10,%rbx	# ...
2853	adcx	%r11,%r9
2854	adox	%rax,%r10
2855	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00	# mulx	5*8($aptr),%r11,%rax
2856	adcx	%r12,%r10
2857	adox	%rbx,%r11
2858	.byte	0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r12,%rbx
2859	adcx	%r13,%r11
2860	adox	%r14,%r12
2861	.byte	0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r13,%r14
2862	 mov	2*8($aptr),%rdx		# a[2]
2863	adcx	%rax,%r12
2864	adox	%rbx,%r13
2865	adcx	%r15,%r13
2866	adox	$zero,%r14		# of=0
2867	adcx	$zero,%r14		# cf=0
2868
2869	mov	%r8,3*8($tptr)		# t[3]
2870	mov	%r9,4*8($tptr)		# t[4]
2871
2872	mulx	3*8($aptr),%r8,%rbx	# a[3]*a[2]
2873	mulx	4*8($aptr),%r9,%rax	# a[4]*a[2]
2874	adcx	%r10,%r8
2875	adox	%rbx,%r9
2876	mulx	5*8($aptr),%r10,%rbx	# ...
2877	adcx	%r11,%r9
2878	adox	%rax,%r10
2879	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r11,%rax
2880	adcx	%r12,%r10
2881	adox	%r13,%r11
2882	.byte	0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r12,%r13
2883	.byte	0x3e
2884	 mov	3*8($aptr),%rdx		# a[3]
2885	adcx	%rbx,%r11
2886	adox	%rax,%r12
2887	adcx	%r14,%r12
2888	mov	%r8,5*8($tptr)		# t[5]
2889	mov	%r9,6*8($tptr)		# t[6]
2890	 mulx	4*8($aptr),%r8,%rax	# a[4]*a[3]
2891	adox	$zero,%r13		# of=0
2892	adcx	$zero,%r13		# cf=0
2893
2894	mulx	5*8($aptr),%r9,%rbx	# a[5]*a[3]
2895	adcx	%r10,%r8
2896	adox	%rax,%r9
2897	mulx	6*8($aptr),%r10,%rax	# ...
2898	adcx	%r11,%r9
2899	adox	%r12,%r10
2900	mulx	7*8($aptr),%r11,%r12
2901	 mov	4*8($aptr),%rdx		# a[4]
2902	 mov	5*8($aptr),%r14		# a[5]
2903	adcx	%rbx,%r10
2904	adox	%rax,%r11
2905	 mov	6*8($aptr),%r15		# a[6]
2906	adcx	%r13,%r11
2907	adox	$zero,%r12		# of=0
2908	adcx	$zero,%r12		# cf=0
2909
2910	mov	%r8,7*8($tptr)		# t[7]
2911	mov	%r9,8*8($tptr)		# t[8]
2912
2913	mulx	%r14,%r9,%rax		# a[5]*a[4]
2914	 mov	7*8($aptr),%r8		# a[7]
2915	adcx	%r10,%r9
2916	mulx	%r15,%r10,%rbx		# a[6]*a[4]
2917	adox	%rax,%r10
2918	adcx	%r11,%r10
2919	mulx	%r8,%r11,%rax		# a[7]*a[4]
2920	 mov	%r14,%rdx		# a[5]
2921	adox	%rbx,%r11
2922	adcx	%r12,%r11
2923	#adox	$zero,%rax		# of=0
2924	adcx	$zero,%rax		# cf=0
2925
2926	mulx	%r15,%r14,%rbx		# a[6]*a[5]
2927	mulx	%r8,%r12,%r13		# a[7]*a[5]
2928	 mov	%r15,%rdx		# a[6]
2929	 lea	8*8($aptr),$aptr
2930	adcx	%r14,%r11
2931	adox	%rbx,%r12
2932	adcx	%rax,%r12
2933	adox	$zero,%r13
2934
2935	.byte	0x67,0x67
2936	mulx	%r8,%r8,%r14		# a[7]*a[6]
2937	adcx	%r8,%r13
2938	adcx	$zero,%r14
2939
2940	cmp	8+8(%rsp),$aptr
2941	je	.Lsqrx8x_outer_break
2942
2943	neg	$carry			# mov $carry,%cf
2944	mov	\$-8,%rcx
2945	mov	$zero,%r15
2946	mov	8*8($tptr),%r8
2947	adcx	9*8($tptr),%r9		# +=t[9]
2948	adcx	10*8($tptr),%r10	# ...
2949	adcx	11*8($tptr),%r11
2950	adc	12*8($tptr),%r12
2951	adc	13*8($tptr),%r13
2952	adc	14*8($tptr),%r14
2953	adc	15*8($tptr),%r15
2954	lea	($aptr),$aaptr
2955	lea	2*64($tptr),$tptr
2956	sbb	%rax,%rax		# mov %cf,$carry
2957
2958	mov	-64($aptr),%rdx		# a[0]
2959	mov	%rax,16+8(%rsp)		# offload $carry
2960	mov	$tptr,24+8(%rsp)
2961
2962	#lea	8*8($tptr),$tptr	# see 2*8*8($tptr) above
2963	xor	%eax,%eax		# cf=0, of=0
2964	jmp	.Lsqrx8x_loop
2965
2966.align	32
2967.Lsqrx8x_loop:
2968	mov	%r8,%rbx
2969	mulx	0*8($aaptr),%rax,%r8	# a[8]*a[i]
2970	adcx	%rax,%rbx		# +=t[8]
2971	adox	%r9,%r8
2972
2973	mulx	1*8($aaptr),%rax,%r9	# ...
2974	adcx	%rax,%r8
2975	adox	%r10,%r9
2976
2977	mulx	2*8($aaptr),%rax,%r10
2978	adcx	%rax,%r9
2979	adox	%r11,%r10
2980
2981	mulx	3*8($aaptr),%rax,%r11
2982	adcx	%rax,%r10
2983	adox	%r12,%r11
2984
2985	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	4*8($aaptr),%rax,%r12
2986	adcx	%rax,%r11
2987	adox	%r13,%r12
2988
2989	mulx	5*8($aaptr),%rax,%r13
2990	adcx	%rax,%r12
2991	adox	%r14,%r13
2992
2993	mulx	6*8($aaptr),%rax,%r14
2994	 mov	%rbx,($tptr,%rcx,8)	# store t[8+i]
2995	 mov	\$0,%ebx
2996	adcx	%rax,%r13
2997	adox	%r15,%r14
2998
2999	.byte	0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00	# mulx	7*8($aaptr),%rax,%r15
3000	 mov	8($aptr,%rcx,8),%rdx	# a[i]
3001	adcx	%rax,%r14
3002	adox	%rbx,%r15		# %rbx is 0, of=0
3003	adcx	%rbx,%r15		# cf=0
3004
3005	.byte	0x67
3006	inc	%rcx			# of=0
3007	jnz	.Lsqrx8x_loop
3008
3009	lea	8*8($aaptr),$aaptr
3010	mov	\$-8,%rcx
3011	cmp	8+8(%rsp),$aaptr	# done?
3012	je	.Lsqrx8x_break
3013
3014	sub	16+8(%rsp),%rbx		# mov 16(%rsp),%cf
3015	.byte	0x66
3016	mov	-64($aptr),%rdx
3017	adcx	0*8($tptr),%r8
3018	adcx	1*8($tptr),%r9
3019	adc	2*8($tptr),%r10
3020	adc	3*8($tptr),%r11
3021	adc	4*8($tptr),%r12
3022	adc	5*8($tptr),%r13
3023	adc	6*8($tptr),%r14
3024	adc	7*8($tptr),%r15
3025	lea	8*8($tptr),$tptr
3026	.byte	0x67
3027	sbb	%rax,%rax		# mov %cf,%rax
3028	xor	%ebx,%ebx		# cf=0, of=0
3029	mov	%rax,16+8(%rsp)		# offload carry
3030	jmp	.Lsqrx8x_loop
3031
3032.align	32
3033.Lsqrx8x_break:
3034	xor	$zero,$zero
3035	sub	16+8(%rsp),%rbx		# mov 16(%rsp),%cf
3036	adcx	$zero,%r8
3037	mov	24+8(%rsp),$carry	# initial $tptr, borrow $carry
3038	adcx	$zero,%r9
3039	mov	0*8($aptr),%rdx		# a[8], modulo-scheduled
3040	adc	\$0,%r10
3041	mov	%r8,0*8($tptr)
3042	adc	\$0,%r11
3043	adc	\$0,%r12
3044	adc	\$0,%r13
3045	adc	\$0,%r14
3046	adc	\$0,%r15
3047	cmp	$carry,$tptr		# cf=0, of=0
3048	je	.Lsqrx8x_outer_loop
3049
3050	mov	%r9,1*8($tptr)
3051	 mov	1*8($carry),%r9
3052	mov	%r10,2*8($tptr)
3053	 mov	2*8($carry),%r10
3054	mov	%r11,3*8($tptr)
3055	 mov	3*8($carry),%r11
3056	mov	%r12,4*8($tptr)
3057	 mov	4*8($carry),%r12
3058	mov	%r13,5*8($tptr)
3059	 mov	5*8($carry),%r13
3060	mov	%r14,6*8($tptr)
3061	 mov	6*8($carry),%r14
3062	mov	%r15,7*8($tptr)
3063	 mov	7*8($carry),%r15
3064	mov	$carry,$tptr
3065	jmp	.Lsqrx8x_outer_loop
3066
3067.align	32
3068.Lsqrx8x_outer_break:
3069	mov	%r9,9*8($tptr)		# t[9]
3070	 movq	%xmm3,%rcx		# -$num
3071	mov	%r10,10*8($tptr)	# ...
3072	mov	%r11,11*8($tptr)
3073	mov	%r12,12*8($tptr)
3074	mov	%r13,13*8($tptr)
3075	mov	%r14,14*8($tptr)
3076___
3077}{
3078my $i="%rcx";
3079$code.=<<___;
3080	lea	48+8(%rsp),$tptr
3081	mov	($aptr,$i),%rdx		# a[0]
3082
3083	mov	8($tptr),$A0[1]		# t[1]
3084	xor	$A0[0],$A0[0]		# t[0], of=0, cf=0
3085	mov	0+8(%rsp),$num		# restore $num
3086	adox	$A0[1],$A0[1]
3087	 mov	16($tptr),$A1[0]	# t[2]	# prefetch
3088	 mov	24($tptr),$A1[1]	# t[3]	# prefetch
3089	#jmp	.Lsqrx4x_shift_n_add	# happens to be aligned
3090
3091.align	32
3092.Lsqrx4x_shift_n_add:
3093	mulx	%rdx,%rax,%rbx
3094	 adox	$A1[0],$A1[0]
3095	adcx	$A0[0],%rax
3096	 .byte	0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00	# mov	8($aptr,$i),%rdx	# a[i+1]	# prefetch
3097	 .byte	0x4c,0x8b,0x97,0x20,0x00,0x00,0x00	# mov	32($tptr),$A0[0]	# t[2*i+4]	# prefetch
3098	 adox	$A1[1],$A1[1]
3099	adcx	$A0[1],%rbx
3100	 mov	40($tptr),$A0[1]		# t[2*i+4+1]	# prefetch
3101	mov	%rax,0($tptr)
3102	mov	%rbx,8($tptr)
3103
3104	mulx	%rdx,%rax,%rbx
3105	 adox	$A0[0],$A0[0]
3106	adcx	$A1[0],%rax
3107	 mov	16($aptr,$i),%rdx	# a[i+2]	# prefetch
3108	 mov	48($tptr),$A1[0]	# t[2*i+6]	# prefetch
3109	 adox	$A0[1],$A0[1]
3110	adcx	$A1[1],%rbx
3111	 mov	56($tptr),$A1[1]	# t[2*i+6+1]	# prefetch
3112	mov	%rax,16($tptr)
3113	mov	%rbx,24($tptr)
3114
3115	mulx	%rdx,%rax,%rbx
3116	 adox	$A1[0],$A1[0]
3117	adcx	$A0[0],%rax
3118	 mov	24($aptr,$i),%rdx	# a[i+3]	# prefetch
3119	 lea	32($i),$i
3120	 mov	64($tptr),$A0[0]	# t[2*i+8]	# prefetch
3121	 adox	$A1[1],$A1[1]
3122	adcx	$A0[1],%rbx
3123	 mov	72($tptr),$A0[1]	# t[2*i+8+1]	# prefetch
3124	mov	%rax,32($tptr)
3125	mov	%rbx,40($tptr)
3126
3127	mulx	%rdx,%rax,%rbx
3128	 adox	$A0[0],$A0[0]
3129	adcx	$A1[0],%rax
3130	jrcxz	.Lsqrx4x_shift_n_add_break
3131	 .byte	0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00	# mov	0($aptr,$i),%rdx	# a[i+4]	# prefetch
3132	 adox	$A0[1],$A0[1]
3133	adcx	$A1[1],%rbx
3134	 mov	80($tptr),$A1[0]	# t[2*i+10]	# prefetch
3135	 mov	88($tptr),$A1[1]	# t[2*i+10+1]	# prefetch
3136	mov	%rax,48($tptr)
3137	mov	%rbx,56($tptr)
3138	lea	64($tptr),$tptr
3139	nop
3140	jmp	.Lsqrx4x_shift_n_add
3141
3142.align	32
3143.Lsqrx4x_shift_n_add_break:
3144	adcx	$A1[1],%rbx
3145	mov	%rax,48($tptr)
3146	mov	%rbx,56($tptr)
3147	lea	64($tptr),$tptr		# end of t[] buffer
3148___
3149}
3150######################################################################
3151# Montgomery reduction part, "word-by-word" algorithm.
3152#
3153# This new path is inspired by multiple submissions from Intel, by
3154# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
3155# Vinodh Gopal...
3156{
3157my ($nptr,$carry,$m0)=("%rbp","%rsi","%rdx");
3158
3159$code.=<<___;
3160	movq	%xmm2,$nptr
3161__bn_sqrx8x_reduction:
3162	xor	%eax,%eax		# initial top-most carry bit
3163	mov	32+8(%rsp),%rbx		# n0
3164	mov	48+8(%rsp),%rdx		# "%r8", 8*0($tptr)
3165	lea	-8*8($nptr,$num),%rcx	# end of n[]
3166	#lea	48+8(%rsp,$num,2),$tptr	# end of t[] buffer
3167	mov	%rcx, 0+8(%rsp)		# save end of n[]
3168	mov	$tptr,8+8(%rsp)		# save end of t[]
3169
3170	lea	48+8(%rsp),$tptr		# initial t[] window
3171	jmp	.Lsqrx8x_reduction_loop
3172
3173.align	32
3174.Lsqrx8x_reduction_loop:
3175	mov	8*1($tptr),%r9
3176	mov	8*2($tptr),%r10
3177	mov	8*3($tptr),%r11
3178	mov	8*4($tptr),%r12
3179	mov	%rdx,%r8
3180	imulq	%rbx,%rdx		# n0*a[i]
3181	mov	8*5($tptr),%r13
3182	mov	8*6($tptr),%r14
3183	mov	8*7($tptr),%r15
3184	mov	%rax,24+8(%rsp)		# store top-most carry bit
3185
3186	lea	8*8($tptr),$tptr
3187	xor	$carry,$carry		# cf=0,of=0
3188	mov	\$-8,%rcx
3189	jmp	.Lsqrx8x_reduce
3190
3191.align	32
3192.Lsqrx8x_reduce:
3193	mov	%r8, %rbx
3194	mulx	8*0($nptr),%rax,%r8	# n[0]
3195	adcx	%rbx,%rax		# discarded
3196	adox	%r9,%r8
3197
3198	mulx	8*1($nptr),%rbx,%r9	# n[1]
3199	adcx	%rbx,%r8
3200	adox	%r10,%r9
3201
3202	mulx	8*2($nptr),%rbx,%r10
3203	adcx	%rbx,%r9
3204	adox	%r11,%r10
3205
3206	mulx	8*3($nptr),%rbx,%r11
3207	adcx	%rbx,%r10
3208	adox	%r12,%r11
3209
3210	.byte	0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rbx,%r12
3211	 mov	%rdx,%rax
3212	 mov	%r8,%rdx
3213	adcx	%rbx,%r11
3214	adox	%r13,%r12
3215
3216	 mulx	32+8(%rsp),%rbx,%rdx	# %rdx discarded
3217	 mov	%rax,%rdx
3218	 mov	%rax,64+48+8(%rsp,%rcx,8)	# put aside n0*a[i]
3219
3220	mulx	8*5($nptr),%rax,%r13
3221	adcx	%rax,%r12
3222	adox	%r14,%r13
3223
3224	mulx	8*6($nptr),%rax,%r14
3225	adcx	%rax,%r13
3226	adox	%r15,%r14
3227
3228	mulx	8*7($nptr),%rax,%r15
3229	 mov	%rbx,%rdx
3230	adcx	%rax,%r14
3231	adox	$carry,%r15		# $carry is 0
3232	adcx	$carry,%r15		# cf=0
3233
3234	.byte	0x67,0x67,0x67
3235	inc	%rcx			# of=0
3236	jnz	.Lsqrx8x_reduce
3237
3238	mov	$carry,%rax		# xor	%rax,%rax
3239	cmp	0+8(%rsp),$nptr		# end of n[]?
3240	jae	.Lsqrx8x_no_tail
3241
3242	mov	48+8(%rsp),%rdx		# pull n0*a[0]
3243	add	8*0($tptr),%r8
3244	lea	8*8($nptr),$nptr
3245	mov	\$-8,%rcx
3246	adcx	8*1($tptr),%r9
3247	adcx	8*2($tptr),%r10
3248	adc	8*3($tptr),%r11
3249	adc	8*4($tptr),%r12
3250	adc	8*5($tptr),%r13
3251	adc	8*6($tptr),%r14
3252	adc	8*7($tptr),%r15
3253	lea	8*8($tptr),$tptr
3254	sbb	%rax,%rax		# top carry
3255
3256	xor	$carry,$carry		# of=0, cf=0
3257	mov	%rax,16+8(%rsp)
3258	jmp	.Lsqrx8x_tail
3259
3260.align	32
3261.Lsqrx8x_tail:
3262	mov	%r8,%rbx
3263	mulx	8*0($nptr),%rax,%r8
3264	adcx	%rax,%rbx
3265	adox	%r9,%r8
3266
3267	mulx	8*1($nptr),%rax,%r9
3268	adcx	%rax,%r8
3269	adox	%r10,%r9
3270
3271	mulx	8*2($nptr),%rax,%r10
3272	adcx	%rax,%r9
3273	adox	%r11,%r10
3274
3275	mulx	8*3($nptr),%rax,%r11
3276	adcx	%rax,%r10
3277	adox	%r12,%r11
3278
3279	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rax,%r12
3280	adcx	%rax,%r11
3281	adox	%r13,%r12
3282
3283	mulx	8*5($nptr),%rax,%r13
3284	adcx	%rax,%r12
3285	adox	%r14,%r13
3286
3287	mulx	8*6($nptr),%rax,%r14
3288	adcx	%rax,%r13
3289	adox	%r15,%r14
3290
3291	mulx	8*7($nptr),%rax,%r15
3292	 mov	72+48+8(%rsp,%rcx,8),%rdx	# pull n0*a[i]
3293	adcx	%rax,%r14
3294	adox	$carry,%r15
3295	 mov	%rbx,($tptr,%rcx,8)	# save result
3296	 mov	%r8,%rbx
3297	adcx	$carry,%r15		# cf=0
3298
3299	inc	%rcx			# of=0
3300	jnz	.Lsqrx8x_tail
3301
3302	cmp	0+8(%rsp),$nptr		# end of n[]?
3303	jae	.Lsqrx8x_tail_done	# break out of loop
3304
3305	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3306	 mov	48+8(%rsp),%rdx		# pull n0*a[0]
3307	 lea	8*8($nptr),$nptr
3308	adc	8*0($tptr),%r8
3309	adc	8*1($tptr),%r9
3310	adc	8*2($tptr),%r10
3311	adc	8*3($tptr),%r11
3312	adc	8*4($tptr),%r12
3313	adc	8*5($tptr),%r13
3314	adc	8*6($tptr),%r14
3315	adc	8*7($tptr),%r15
3316	lea	8*8($tptr),$tptr
3317	sbb	%rax,%rax
3318	sub	\$8,%rcx		# mov	\$-8,%rcx
3319
3320	xor	$carry,$carry		# of=0, cf=0
3321	mov	%rax,16+8(%rsp)
3322	jmp	.Lsqrx8x_tail
3323
3324.align	32
3325.Lsqrx8x_tail_done:
3326	xor	%rax,%rax
3327	add	24+8(%rsp),%r8		# can this overflow?
3328	adc	\$0,%r9
3329	adc	\$0,%r10
3330	adc	\$0,%r11
3331	adc	\$0,%r12
3332	adc	\$0,%r13
3333	adc	\$0,%r14
3334	adc	\$0,%r15
3335	adc	\$0,%rax
3336
3337	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3338.Lsqrx8x_no_tail:			# %cf is 0 if jumped here
3339	adc	8*0($tptr),%r8
3340	 movq	%xmm3,%rcx
3341	adc	8*1($tptr),%r9
3342	 mov	8*7($nptr),$carry
3343	 movq	%xmm2,$nptr		# restore $nptr
3344	adc	8*2($tptr),%r10
3345	adc	8*3($tptr),%r11
3346	adc	8*4($tptr),%r12
3347	adc	8*5($tptr),%r13
3348	adc	8*6($tptr),%r14
3349	adc	8*7($tptr),%r15
3350	adc	\$0,%rax		# top-most carry
3351
3352	mov	32+8(%rsp),%rbx		# n0
3353	mov	8*8($tptr,%rcx),%rdx	# modulo-scheduled "%r8"
3354
3355	mov	%r8,8*0($tptr)		# store top 512 bits
3356	 lea	8*8($tptr),%r8		# borrow %r8
3357	mov	%r9,8*1($tptr)
3358	mov	%r10,8*2($tptr)
3359	mov	%r11,8*3($tptr)
3360	mov	%r12,8*4($tptr)
3361	mov	%r13,8*5($tptr)
3362	mov	%r14,8*6($tptr)
3363	mov	%r15,8*7($tptr)
3364
3365	lea	8*8($tptr,%rcx),$tptr	# start of current t[] window
3366	cmp	8+8(%rsp),%r8		# end of t[]?
3367	jb	.Lsqrx8x_reduction_loop
3368	ret
3369.cfi_endproc
3370.size	bn_sqrx8x_internal,.-bn_sqrx8x_internal
3371___
3372}
3373##############################################################
3374# Post-condition, 4x unrolled
3375#
3376{
3377my ($rptr,$nptr)=("%rdx","%rbp");
3378$code.=<<___;
3379.align	32
3380__bn_postx4x_internal:
3381.cfi_startproc
3382	mov	8*0($nptr),%r12
3383	mov	%rcx,%r10		# -$num
3384	mov	%rcx,%r9		# -$num
3385	neg	%rax
3386	sar	\$3+2,%rcx
3387	#lea	48+8(%rsp,%r9),$tptr
3388	movq	%xmm1,$rptr		# restore $rptr
3389	movq	%xmm1,$aptr		# prepare for back-to-back call
3390	dec	%r12			# so that after 'not' we get -n[0]
3391	mov	8*1($nptr),%r13
3392	xor	%r8,%r8
3393	mov	8*2($nptr),%r14
3394	mov	8*3($nptr),%r15
3395	jmp	.Lsqrx4x_sub_entry
3396
3397.align	16
3398.Lsqrx4x_sub:
3399	mov	8*0($nptr),%r12
3400	mov	8*1($nptr),%r13
3401	mov	8*2($nptr),%r14
3402	mov	8*3($nptr),%r15
3403.Lsqrx4x_sub_entry:
3404	andn	%rax,%r12,%r12
3405	lea	8*4($nptr),$nptr
3406	andn	%rax,%r13,%r13
3407	andn	%rax,%r14,%r14
3408	andn	%rax,%r15,%r15
3409
3410	neg	%r8			# mov %r8,%cf
3411	adc	8*0($tptr),%r12
3412	adc	8*1($tptr),%r13
3413	adc	8*2($tptr),%r14
3414	adc	8*3($tptr),%r15
3415	mov	%r12,8*0($rptr)
3416	lea	8*4($tptr),$tptr
3417	mov	%r13,8*1($rptr)
3418	sbb	%r8,%r8			# mov %cf,%r8
3419	mov	%r14,8*2($rptr)
3420	mov	%r15,8*3($rptr)
3421	lea	8*4($rptr),$rptr
3422
3423	inc	%rcx
3424	jnz	.Lsqrx4x_sub
3425
3426	neg	%r9			# restore $num
3427
3428	ret
3429.cfi_endproc
3430.size	__bn_postx4x_internal,.-__bn_postx4x_internal
3431___
3432}
3433}}}
3434{
3435my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%edx","%r8", "%r9d") : # Win64 order
3436				("%rdi","%esi","%rdx","%ecx");  # Unix order
3437my $out=$inp;
3438my $STRIDE=2**5*8;
3439my $N=$STRIDE/4;
3440
3441$code.=<<___;
3442.globl	bn_get_bits5
3443.type	bn_get_bits5,\@abi-omnipotent
3444.align	16
3445bn_get_bits5:
3446.cfi_startproc
3447	lea	0($inp),%r10
3448	lea	1($inp),%r11
3449	mov	$num,%ecx
3450	shr	\$4,$num
3451	and	\$15,%ecx
3452	lea	-8(%ecx),%eax
3453	cmp	\$11,%ecx
3454	cmova	%r11,%r10
3455	cmova	%eax,%ecx
3456	movzw	(%r10,$num,2),%eax
3457	shrl	%cl,%eax
3458	and	\$31,%eax
3459	ret
3460.cfi_endproc
3461.size	bn_get_bits5,.-bn_get_bits5
3462
3463.globl	bn_scatter5
3464.type	bn_scatter5,\@abi-omnipotent
3465.align	16
3466bn_scatter5:
3467.cfi_startproc
3468	cmp	\$0, $num
3469	jz	.Lscatter_epilogue
3470	lea	($tbl,$idx,8),$tbl
3471.Lscatter:
3472	mov	($inp),%rax
3473	lea	8($inp),$inp
3474	mov	%rax,($tbl)
3475	lea	32*8($tbl),$tbl
3476	sub	\$1,$num
3477	jnz	.Lscatter
3478.Lscatter_epilogue:
3479	ret
3480.cfi_endproc
3481.size	bn_scatter5,.-bn_scatter5
3482
3483.globl	bn_gather5
3484.type	bn_gather5,\@abi-omnipotent
3485.align	32
3486bn_gather5:
3487.LSEH_begin_bn_gather5:			# Win64 thing, but harmless in other cases
3488.cfi_startproc
3489	# I can't trust assembler to use specific encoding:-(
3490	.byte	0x4c,0x8d,0x14,0x24			#lea    (%rsp),%r10
3491	.byte	0x48,0x81,0xec,0x08,0x01,0x00,0x00	#sub	$0x108,%rsp
3492	lea	.Linc(%rip),%rax
3493	and	\$-16,%rsp		# shouldn't be formally required
3494
3495	movd	$idx,%xmm5
3496	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
3497	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
3498	lea	128($tbl),%r11		# size optimization
3499	lea	128(%rsp),%rax		# size optimization
3500
3501	pshufd	\$0,%xmm5,%xmm5		# broadcast $idx
3502	movdqa	%xmm1,%xmm4
3503	movdqa	%xmm1,%xmm2
3504___
3505########################################################################
3506# calculate mask by comparing 0..31 to $idx and save result to stack
3507#
3508for($i=0;$i<$STRIDE/16;$i+=4) {
3509$code.=<<___;
3510	paddd	%xmm0,%xmm1
3511	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
3512___
3513$code.=<<___	if ($i);
3514	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3515___
3516$code.=<<___;
3517	movdqa	%xmm4,%xmm3
3518
3519	paddd	%xmm1,%xmm2
3520	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
3521	movdqa	%xmm0,`16*($i+0)-128`(%rax)
3522	movdqa	%xmm4,%xmm0
3523
3524	paddd	%xmm2,%xmm3
3525	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
3526	movdqa	%xmm1,`16*($i+1)-128`(%rax)
3527	movdqa	%xmm4,%xmm1
3528
3529	paddd	%xmm3,%xmm0
3530	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
3531	movdqa	%xmm2,`16*($i+2)-128`(%rax)
3532	movdqa	%xmm4,%xmm2
3533___
3534}
3535$code.=<<___;
3536	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3537	jmp	.Lgather
3538
3539.align	32
3540.Lgather:
3541	pxor	%xmm4,%xmm4
3542	pxor	%xmm5,%xmm5
3543___
3544for($i=0;$i<$STRIDE/16;$i+=4) {
3545$code.=<<___;
3546	movdqa	`16*($i+0)-128`(%r11),%xmm0
3547	movdqa	`16*($i+1)-128`(%r11),%xmm1
3548	movdqa	`16*($i+2)-128`(%r11),%xmm2
3549	pand	`16*($i+0)-128`(%rax),%xmm0
3550	movdqa	`16*($i+3)-128`(%r11),%xmm3
3551	pand	`16*($i+1)-128`(%rax),%xmm1
3552	por	%xmm0,%xmm4
3553	pand	`16*($i+2)-128`(%rax),%xmm2
3554	por	%xmm1,%xmm5
3555	pand	`16*($i+3)-128`(%rax),%xmm3
3556	por	%xmm2,%xmm4
3557	por	%xmm3,%xmm5
3558___
3559}
3560$code.=<<___;
3561	por	%xmm5,%xmm4
3562	lea	$STRIDE(%r11),%r11
3563	pshufd	\$0x4e,%xmm4,%xmm0
3564	por	%xmm4,%xmm0
3565	movq	%xmm0,($out)		# m0=bp[0]
3566	lea	8($out),$out
3567	sub	\$1,$num
3568	jnz	.Lgather
3569
3570	lea	(%r10),%rsp
3571	ret
3572.LSEH_end_bn_gather5:
3573.cfi_endproc
3574.size	bn_gather5,.-bn_gather5
3575___
3576}
3577$code.=<<___;
3578.align	64
3579.Linc:
3580	.long	0,0, 1,1
3581	.long	2,2, 2,2
3582.asciz	"Montgomery Multiplication with scatter/gather for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
3583___
3584
3585# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
3586#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
3587if ($win64) {
3588$rec="%rcx";
3589$frame="%rdx";
3590$context="%r8";
3591$disp="%r9";
3592
3593$code.=<<___;
3594.extern	__imp_RtlVirtualUnwind
3595.type	mul_handler,\@abi-omnipotent
3596.align	16
3597mul_handler:
3598	push	%rsi
3599	push	%rdi
3600	push	%rbx
3601	push	%rbp
3602	push	%r12
3603	push	%r13
3604	push	%r14
3605	push	%r15
3606	pushfq
3607	sub	\$64,%rsp
3608
3609	mov	120($context),%rax	# pull context->Rax
3610	mov	248($context),%rbx	# pull context->Rip
3611
3612	mov	8($disp),%rsi		# disp->ImageBase
3613	mov	56($disp),%r11		# disp->HandlerData
3614
3615	mov	0(%r11),%r10d		# HandlerData[0]
3616	lea	(%rsi,%r10),%r10	# end of prologue label
3617	cmp	%r10,%rbx		# context->Rip<end of prologue label
3618	jb	.Lcommon_seh_tail
3619
3620	mov	4(%r11),%r10d		# HandlerData[1]
3621	lea	(%rsi,%r10),%r10	# beginning of body label
3622	cmp	%r10,%rbx		# context->Rip<body label
3623	jb	.Lcommon_pop_regs
3624
3625	mov	152($context),%rax	# pull context->Rsp
3626
3627	mov	8(%r11),%r10d		# HandlerData[2]
3628	lea	(%rsi,%r10),%r10	# epilogue label
3629	cmp	%r10,%rbx		# context->Rip>=epilogue label
3630	jae	.Lcommon_seh_tail
3631
3632	lea	.Lmul_epilogue(%rip),%r10
3633	cmp	%r10,%rbx
3634	ja	.Lbody_40
3635
3636	mov	192($context),%r10	# pull $num
3637	mov	8(%rax,%r10,8),%rax	# pull saved stack pointer
3638
3639	jmp	.Lcommon_pop_regs
3640
3641.Lbody_40:
3642	mov	40(%rax),%rax		# pull saved stack pointer
3643.Lcommon_pop_regs:
3644	mov	-8(%rax),%rbx
3645	mov	-16(%rax),%rbp
3646	mov	-24(%rax),%r12
3647	mov	-32(%rax),%r13
3648	mov	-40(%rax),%r14
3649	mov	-48(%rax),%r15
3650	mov	%rbx,144($context)	# restore context->Rbx
3651	mov	%rbp,160($context)	# restore context->Rbp
3652	mov	%r12,216($context)	# restore context->R12
3653	mov	%r13,224($context)	# restore context->R13
3654	mov	%r14,232($context)	# restore context->R14
3655	mov	%r15,240($context)	# restore context->R15
3656
3657.Lcommon_seh_tail:
3658	mov	8(%rax),%rdi
3659	mov	16(%rax),%rsi
3660	mov	%rax,152($context)	# restore context->Rsp
3661	mov	%rsi,168($context)	# restore context->Rsi
3662	mov	%rdi,176($context)	# restore context->Rdi
3663
3664	mov	40($disp),%rdi		# disp->ContextRecord
3665	mov	$context,%rsi		# context
3666	mov	\$154,%ecx		# sizeof(CONTEXT)
3667	.long	0xa548f3fc		# cld; rep movsq
3668
3669	mov	$disp,%rsi
3670	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
3671	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
3672	mov	0(%rsi),%r8		# arg3, disp->ControlPc
3673	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
3674	mov	40(%rsi),%r10		# disp->ContextRecord
3675	lea	56(%rsi),%r11		# &disp->HandlerData
3676	lea	24(%rsi),%r12		# &disp->EstablisherFrame
3677	mov	%r10,32(%rsp)		# arg5
3678	mov	%r11,40(%rsp)		# arg6
3679	mov	%r12,48(%rsp)		# arg7
3680	mov	%rcx,56(%rsp)		# arg8, (NULL)
3681	call	*__imp_RtlVirtualUnwind(%rip)
3682
3683	mov	\$1,%eax		# ExceptionContinueSearch
3684	add	\$64,%rsp
3685	popfq
3686	pop	%r15
3687	pop	%r14
3688	pop	%r13
3689	pop	%r12
3690	pop	%rbp
3691	pop	%rbx
3692	pop	%rdi
3693	pop	%rsi
3694	ret
3695.size	mul_handler,.-mul_handler
3696
3697.section	.pdata
3698.align	4
3699	.rva	.LSEH_begin_bn_mul_mont_gather5
3700	.rva	.LSEH_end_bn_mul_mont_gather5
3701	.rva	.LSEH_info_bn_mul_mont_gather5
3702
3703	.rva	.LSEH_begin_bn_mul4x_mont_gather5
3704	.rva	.LSEH_end_bn_mul4x_mont_gather5
3705	.rva	.LSEH_info_bn_mul4x_mont_gather5
3706
3707	.rva	.LSEH_begin_bn_power5
3708	.rva	.LSEH_end_bn_power5
3709	.rva	.LSEH_info_bn_power5
3710___
3711$code.=<<___ if ($addx);
3712	.rva	.LSEH_begin_bn_mulx4x_mont_gather5
3713	.rva	.LSEH_end_bn_mulx4x_mont_gather5
3714	.rva	.LSEH_info_bn_mulx4x_mont_gather5
3715
3716	.rva	.LSEH_begin_bn_powerx5
3717	.rva	.LSEH_end_bn_powerx5
3718	.rva	.LSEH_info_bn_powerx5
3719___
3720$code.=<<___;
3721	.rva	.LSEH_begin_bn_gather5
3722	.rva	.LSEH_end_bn_gather5
3723	.rva	.LSEH_info_bn_gather5
3724
3725.section	.xdata
3726.align	8
3727.LSEH_info_bn_mul_mont_gather5:
3728	.byte	9,0,0,0
3729	.rva	mul_handler
3730	.rva	.Lmul_body,.Lmul_body,.Lmul_epilogue		# HandlerData[]
3731.align	8
3732.LSEH_info_bn_mul4x_mont_gather5:
3733	.byte	9,0,0,0
3734	.rva	mul_handler
3735	.rva	.Lmul4x_prologue,.Lmul4x_body,.Lmul4x_epilogue		# HandlerData[]
3736.align	8
3737.LSEH_info_bn_power5:
3738	.byte	9,0,0,0
3739	.rva	mul_handler
3740	.rva	.Lpower5_prologue,.Lpower5_body,.Lpower5_epilogue	# HandlerData[]
3741___
3742$code.=<<___ if ($addx);
3743.align	8
3744.LSEH_info_bn_mulx4x_mont_gather5:
3745	.byte	9,0,0,0
3746	.rva	mul_handler
3747	.rva	.Lmulx4x_prologue,.Lmulx4x_body,.Lmulx4x_epilogue	# HandlerData[]
3748.align	8
3749.LSEH_info_bn_powerx5:
3750	.byte	9,0,0,0
3751	.rva	mul_handler
3752	.rva	.Lpowerx5_prologue,.Lpowerx5_body,.Lpowerx5_epilogue	# HandlerData[]
3753___
3754$code.=<<___;
3755.align	8
3756.LSEH_info_bn_gather5:
3757	.byte	0x01,0x0b,0x03,0x0a
3758	.byte	0x0b,0x01,0x21,0x00	# sub	rsp,0x108
3759	.byte	0x04,0xa3,0x00,0x00	# lea	r10,(rsp)
3760.align	8
3761___
3762}
3763
3764$code =~ s/\`([^\`]*)\`/eval($1)/gem;
3765
3766print $code;
3767close STDOUT or die "error closing STDOUT: $!";
3768