xref: /freebsd/crypto/openssl/crypto/bn/asm/x86_64-mont5.pl (revision e9b1dc32c9bd2ebae5f9e140bfa0e0321bc366b5) 
1#! /usr/bin/env perl
2# Copyright 2011-2018 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) ([3-9])\.([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	$num
550
551	mov	%rax,40(%rsp)
552.cfi_cfa_expression	%rsp+40,deref,+8
553.Lmul4x_body:
554
555	call	mul4x_internal
556
557	mov	40(%rsp),%rsi		# restore %rsp
558.cfi_def_cfa	%rsi,8
559	mov	\$1,%rax
560
561	mov	-48(%rsi),%r15
562.cfi_restore	%r15
563	mov	-40(%rsi),%r14
564.cfi_restore	%r14
565	mov	-32(%rsi),%r13
566.cfi_restore	%r13
567	mov	-24(%rsi),%r12
568.cfi_restore	%r12
569	mov	-16(%rsi),%rbp
570.cfi_restore	%rbp
571	mov	-8(%rsi),%rbx
572.cfi_restore	%rbx
573	lea	(%rsi),%rsp
574.cfi_def_cfa_register	%rsp
575.Lmul4x_epilogue:
576	ret
577.cfi_endproc
578.size	bn_mul4x_mont_gather5,.-bn_mul4x_mont_gather5
579
580.type	mul4x_internal,\@abi-omnipotent
581.align	32
582mul4x_internal:
583	shl	\$5,$num		# $num was in bytes
584	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument, index
585	lea	.Linc(%rip),%rax
586	lea	128(%rdx,$num),%r13	# end of powers table (+size optimization)
587	shr	\$5,$num		# restore $num
588___
589		$bp="%r12";
590		$STRIDE=2**5*8;		# 5 is "window size"
591		$N=$STRIDE/4;		# should match cache line size
592		$tp=$i;
593$code.=<<___;
594	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
595	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
596	lea	88-112(%rsp,$num),%r10	# place the mask after tp[num+1] (+ICache optimization)
597	lea	128(%rdx),$bp		# size optimization
598
599	pshufd	\$0,%xmm5,%xmm5		# broadcast index
600	movdqa	%xmm1,%xmm4
601	.byte	0x67,0x67
602	movdqa	%xmm1,%xmm2
603___
604########################################################################
605# calculate mask by comparing 0..31 to index and save result to stack
606#
607$code.=<<___;
608	paddd	%xmm0,%xmm1
609	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
610	.byte	0x67
611	movdqa	%xmm4,%xmm3
612___
613for($i=0;$i<$STRIDE/16-4;$i+=4) {
614$code.=<<___;
615	paddd	%xmm1,%xmm2
616	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
617	movdqa	%xmm0,`16*($i+0)+112`(%r10)
618	movdqa	%xmm4,%xmm0
619
620	paddd	%xmm2,%xmm3
621	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
622	movdqa	%xmm1,`16*($i+1)+112`(%r10)
623	movdqa	%xmm4,%xmm1
624
625	paddd	%xmm3,%xmm0
626	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
627	movdqa	%xmm2,`16*($i+2)+112`(%r10)
628	movdqa	%xmm4,%xmm2
629
630	paddd	%xmm0,%xmm1
631	pcmpeqd	%xmm5,%xmm0
632	movdqa	%xmm3,`16*($i+3)+112`(%r10)
633	movdqa	%xmm4,%xmm3
634___
635}
636$code.=<<___;				# last iteration can be optimized
637	paddd	%xmm1,%xmm2
638	pcmpeqd	%xmm5,%xmm1
639	movdqa	%xmm0,`16*($i+0)+112`(%r10)
640
641	paddd	%xmm2,%xmm3
642	.byte	0x67
643	pcmpeqd	%xmm5,%xmm2
644	movdqa	%xmm1,`16*($i+1)+112`(%r10)
645
646	pcmpeqd	%xmm5,%xmm3
647	movdqa	%xmm2,`16*($i+2)+112`(%r10)
648	pand	`16*($i+0)-128`($bp),%xmm0	# while it's still in register
649
650	pand	`16*($i+1)-128`($bp),%xmm1
651	pand	`16*($i+2)-128`($bp),%xmm2
652	movdqa	%xmm3,`16*($i+3)+112`(%r10)
653	pand	`16*($i+3)-128`($bp),%xmm3
654	por	%xmm2,%xmm0
655	por	%xmm3,%xmm1
656___
657for($i=0;$i<$STRIDE/16-4;$i+=4) {
658$code.=<<___;
659	movdqa	`16*($i+0)-128`($bp),%xmm4
660	movdqa	`16*($i+1)-128`($bp),%xmm5
661	movdqa	`16*($i+2)-128`($bp),%xmm2
662	pand	`16*($i+0)+112`(%r10),%xmm4
663	movdqa	`16*($i+3)-128`($bp),%xmm3
664	pand	`16*($i+1)+112`(%r10),%xmm5
665	por	%xmm4,%xmm0
666	pand	`16*($i+2)+112`(%r10),%xmm2
667	por	%xmm5,%xmm1
668	pand	`16*($i+3)+112`(%r10),%xmm3
669	por	%xmm2,%xmm0
670	por	%xmm3,%xmm1
671___
672}
673$code.=<<___;
674	por	%xmm1,%xmm0
675	pshufd	\$0x4e,%xmm0,%xmm1
676	por	%xmm1,%xmm0
677	lea	$STRIDE($bp),$bp
678	movq	%xmm0,$m0		# m0=bp[0]
679
680	mov	%r13,16+8(%rsp)		# save end of b[num]
681	mov	$rp, 56+8(%rsp)		# save $rp
682
683	mov	($n0),$n0		# pull n0[0] value
684	mov	($ap),%rax
685	lea	($ap,$num),$ap		# end of a[num]
686	neg	$num
687
688	mov	$n0,$m1
689	mulq	$m0			# ap[0]*bp[0]
690	mov	%rax,$A[0]
691	mov	($np),%rax
692
693	imulq	$A[0],$m1		# "tp[0]"*n0
694	lea	64+8(%rsp),$tp
695	mov	%rdx,$A[1]
696
697	mulq	$m1			# np[0]*m1
698	add	%rax,$A[0]		# discarded
699	mov	8($ap,$num),%rax
700	adc	\$0,%rdx
701	mov	%rdx,$N[1]
702
703	mulq	$m0
704	add	%rax,$A[1]
705	mov	8*1($np),%rax
706	adc	\$0,%rdx
707	mov	%rdx,$A[0]
708
709	mulq	$m1
710	add	%rax,$N[1]
711	mov	16($ap,$num),%rax
712	adc	\$0,%rdx
713	add	$A[1],$N[1]
714	lea	4*8($num),$j		# j=4
715	lea	8*4($np),$np
716	adc	\$0,%rdx
717	mov	$N[1],($tp)
718	mov	%rdx,$N[0]
719	jmp	.L1st4x
720
721.align	32
722.L1st4x:
723	mulq	$m0			# ap[j]*bp[0]
724	add	%rax,$A[0]
725	mov	-8*2($np),%rax
726	lea	32($tp),$tp
727	adc	\$0,%rdx
728	mov	%rdx,$A[1]
729
730	mulq	$m1			# np[j]*m1
731	add	%rax,$N[0]
732	mov	-8($ap,$j),%rax
733	adc	\$0,%rdx
734	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
735	adc	\$0,%rdx
736	mov	$N[0],-24($tp)		# tp[j-1]
737	mov	%rdx,$N[1]
738
739	mulq	$m0			# ap[j]*bp[0]
740	add	%rax,$A[1]
741	mov	-8*1($np),%rax
742	adc	\$0,%rdx
743	mov	%rdx,$A[0]
744
745	mulq	$m1			# np[j]*m1
746	add	%rax,$N[1]
747	mov	($ap,$j),%rax
748	adc	\$0,%rdx
749	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
750	adc	\$0,%rdx
751	mov	$N[1],-16($tp)		# tp[j-1]
752	mov	%rdx,$N[0]
753
754	mulq	$m0			# ap[j]*bp[0]
755	add	%rax,$A[0]
756	mov	8*0($np),%rax
757	adc	\$0,%rdx
758	mov	%rdx,$A[1]
759
760	mulq	$m1			# np[j]*m1
761	add	%rax,$N[0]
762	mov	8($ap,$j),%rax
763	adc	\$0,%rdx
764	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
765	adc	\$0,%rdx
766	mov	$N[0],-8($tp)		# tp[j-1]
767	mov	%rdx,$N[1]
768
769	mulq	$m0			# ap[j]*bp[0]
770	add	%rax,$A[1]
771	mov	8*1($np),%rax
772	adc	\$0,%rdx
773	mov	%rdx,$A[0]
774
775	mulq	$m1			# np[j]*m1
776	add	%rax,$N[1]
777	mov	16($ap,$j),%rax
778	adc	\$0,%rdx
779	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
780	lea	8*4($np),$np
781	adc	\$0,%rdx
782	mov	$N[1],($tp)		# tp[j-1]
783	mov	%rdx,$N[0]
784
785	add	\$32,$j			# j+=4
786	jnz	.L1st4x
787
788	mulq	$m0			# ap[j]*bp[0]
789	add	%rax,$A[0]
790	mov	-8*2($np),%rax
791	lea	32($tp),$tp
792	adc	\$0,%rdx
793	mov	%rdx,$A[1]
794
795	mulq	$m1			# np[j]*m1
796	add	%rax,$N[0]
797	mov	-8($ap),%rax
798	adc	\$0,%rdx
799	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
800	adc	\$0,%rdx
801	mov	$N[0],-24($tp)		# tp[j-1]
802	mov	%rdx,$N[1]
803
804	mulq	$m0			# ap[j]*bp[0]
805	add	%rax,$A[1]
806	mov	-8*1($np),%rax
807	adc	\$0,%rdx
808	mov	%rdx,$A[0]
809
810	mulq	$m1			# np[j]*m1
811	add	%rax,$N[1]
812	mov	($ap,$num),%rax		# ap[0]
813	adc	\$0,%rdx
814	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
815	adc	\$0,%rdx
816	mov	$N[1],-16($tp)		# tp[j-1]
817	mov	%rdx,$N[0]
818
819	lea	($np,$num),$np		# rewind $np
820
821	xor	$N[1],$N[1]
822	add	$A[0],$N[0]
823	adc	\$0,$N[1]
824	mov	$N[0],-8($tp)
825
826	jmp	.Louter4x
827
828.align	32
829.Louter4x:
830	lea	16+128($tp),%rdx	# where 256-byte mask is (+size optimization)
831	pxor	%xmm4,%xmm4
832	pxor	%xmm5,%xmm5
833___
834for($i=0;$i<$STRIDE/16;$i+=4) {
835$code.=<<___;
836	movdqa	`16*($i+0)-128`($bp),%xmm0
837	movdqa	`16*($i+1)-128`($bp),%xmm1
838	movdqa	`16*($i+2)-128`($bp),%xmm2
839	movdqa	`16*($i+3)-128`($bp),%xmm3
840	pand	`16*($i+0)-128`(%rdx),%xmm0
841	pand	`16*($i+1)-128`(%rdx),%xmm1
842	por	%xmm0,%xmm4
843	pand	`16*($i+2)-128`(%rdx),%xmm2
844	por	%xmm1,%xmm5
845	pand	`16*($i+3)-128`(%rdx),%xmm3
846	por	%xmm2,%xmm4
847	por	%xmm3,%xmm5
848___
849}
850$code.=<<___;
851	por	%xmm5,%xmm4
852	pshufd	\$0x4e,%xmm4,%xmm0
853	por	%xmm4,%xmm0
854	lea	$STRIDE($bp),$bp
855	movq	%xmm0,$m0		# m0=bp[i]
856
857	mov	($tp,$num),$A[0]
858	mov	$n0,$m1
859	mulq	$m0			# ap[0]*bp[i]
860	add	%rax,$A[0]		# ap[0]*bp[i]+tp[0]
861	mov	($np),%rax
862	adc	\$0,%rdx
863
864	imulq	$A[0],$m1		# tp[0]*n0
865	mov	%rdx,$A[1]
866	mov	$N[1],($tp)		# store upmost overflow bit
867
868	lea	($tp,$num),$tp		# rewind $tp
869
870	mulq	$m1			# np[0]*m1
871	add	%rax,$A[0]		# "$N[0]", discarded
872	mov	8($ap,$num),%rax
873	adc	\$0,%rdx
874	mov	%rdx,$N[1]
875
876	mulq	$m0			# ap[j]*bp[i]
877	add	%rax,$A[1]
878	mov	8*1($np),%rax
879	adc	\$0,%rdx
880	add	8($tp),$A[1]		# +tp[1]
881	adc	\$0,%rdx
882	mov	%rdx,$A[0]
883
884	mulq	$m1			# np[j]*m1
885	add	%rax,$N[1]
886	mov	16($ap,$num),%rax
887	adc	\$0,%rdx
888	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[i]+tp[j]
889	lea	4*8($num),$j		# j=4
890	lea	8*4($np),$np
891	adc	\$0,%rdx
892	mov	%rdx,$N[0]
893	jmp	.Linner4x
894
895.align	32
896.Linner4x:
897	mulq	$m0			# ap[j]*bp[i]
898	add	%rax,$A[0]
899	mov	-8*2($np),%rax
900	adc	\$0,%rdx
901	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
902	lea	32($tp),$tp
903	adc	\$0,%rdx
904	mov	%rdx,$A[1]
905
906	mulq	$m1			# np[j]*m1
907	add	%rax,$N[0]
908	mov	-8($ap,$j),%rax
909	adc	\$0,%rdx
910	add	$A[0],$N[0]
911	adc	\$0,%rdx
912	mov	$N[1],-32($tp)		# tp[j-1]
913	mov	%rdx,$N[1]
914
915	mulq	$m0			# ap[j]*bp[i]
916	add	%rax,$A[1]
917	mov	-8*1($np),%rax
918	adc	\$0,%rdx
919	add	-8($tp),$A[1]
920	adc	\$0,%rdx
921	mov	%rdx,$A[0]
922
923	mulq	$m1			# np[j]*m1
924	add	%rax,$N[1]
925	mov	($ap,$j),%rax
926	adc	\$0,%rdx
927	add	$A[1],$N[1]
928	adc	\$0,%rdx
929	mov	$N[0],-24($tp)		# tp[j-1]
930	mov	%rdx,$N[0]
931
932	mulq	$m0			# ap[j]*bp[i]
933	add	%rax,$A[0]
934	mov	8*0($np),%rax
935	adc	\$0,%rdx
936	add	($tp),$A[0]		# ap[j]*bp[i]+tp[j]
937	adc	\$0,%rdx
938	mov	%rdx,$A[1]
939
940	mulq	$m1			# np[j]*m1
941	add	%rax,$N[0]
942	mov	8($ap,$j),%rax
943	adc	\$0,%rdx
944	add	$A[0],$N[0]
945	adc	\$0,%rdx
946	mov	$N[1],-16($tp)		# tp[j-1]
947	mov	%rdx,$N[1]
948
949	mulq	$m0			# ap[j]*bp[i]
950	add	%rax,$A[1]
951	mov	8*1($np),%rax
952	adc	\$0,%rdx
953	add	8($tp),$A[1]
954	adc	\$0,%rdx
955	mov	%rdx,$A[0]
956
957	mulq	$m1			# np[j]*m1
958	add	%rax,$N[1]
959	mov	16($ap,$j),%rax
960	adc	\$0,%rdx
961	add	$A[1],$N[1]
962	lea	8*4($np),$np
963	adc	\$0,%rdx
964	mov	$N[0],-8($tp)		# tp[j-1]
965	mov	%rdx,$N[0]
966
967	add	\$32,$j			# j+=4
968	jnz	.Linner4x
969
970	mulq	$m0			# ap[j]*bp[i]
971	add	%rax,$A[0]
972	mov	-8*2($np),%rax
973	adc	\$0,%rdx
974	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
975	lea	32($tp),$tp
976	adc	\$0,%rdx
977	mov	%rdx,$A[1]
978
979	mulq	$m1			# np[j]*m1
980	add	%rax,$N[0]
981	mov	-8($ap),%rax
982	adc	\$0,%rdx
983	add	$A[0],$N[0]
984	adc	\$0,%rdx
985	mov	$N[1],-32($tp)		# tp[j-1]
986	mov	%rdx,$N[1]
987
988	mulq	$m0			# ap[j]*bp[i]
989	add	%rax,$A[1]
990	mov	$m1,%rax
991	mov	-8*1($np),$m1
992	adc	\$0,%rdx
993	add	-8($tp),$A[1]
994	adc	\$0,%rdx
995	mov	%rdx,$A[0]
996
997	mulq	$m1			# np[j]*m1
998	add	%rax,$N[1]
999	mov	($ap,$num),%rax		# ap[0]
1000	adc	\$0,%rdx
1001	add	$A[1],$N[1]
1002	adc	\$0,%rdx
1003	mov	$N[0],-24($tp)		# tp[j-1]
1004	mov	%rdx,$N[0]
1005
1006	mov	$N[1],-16($tp)		# tp[j-1]
1007	lea	($np,$num),$np		# rewind $np
1008
1009	xor	$N[1],$N[1]
1010	add	$A[0],$N[0]
1011	adc	\$0,$N[1]
1012	add	($tp),$N[0]		# pull upmost overflow bit
1013	adc	\$0,$N[1]		# upmost overflow bit
1014	mov	$N[0],-8($tp)
1015
1016	cmp	16+8(%rsp),$bp
1017	jb	.Louter4x
1018___
1019if (1) {
1020$code.=<<___;
1021	xor	%rax,%rax
1022	sub	$N[0],$m1		# compare top-most words
1023	adc	$j,$j			# $j is zero
1024	or	$j,$N[1]
1025	sub	$N[1],%rax		# %rax=-$N[1]
1026	lea	($tp,$num),%rbx		# tptr in .sqr4x_sub
1027	mov	($np),%r12
1028	lea	($np),%rbp		# nptr in .sqr4x_sub
1029	mov	%r9,%rcx
1030	sar	\$3+2,%rcx
1031	mov	56+8(%rsp),%rdi		# rptr in .sqr4x_sub
1032	dec	%r12			# so that after 'not' we get -n[0]
1033	xor	%r10,%r10
1034	mov	8*1(%rbp),%r13
1035	mov	8*2(%rbp),%r14
1036	mov	8*3(%rbp),%r15
1037	jmp	.Lsqr4x_sub_entry
1038___
1039} else {
1040my @ri=("%rax",$bp,$m0,$m1);
1041my $rp="%rdx";
1042$code.=<<___
1043	xor	\$1,$N[1]
1044	lea	($tp,$num),$tp		# rewind $tp
1045	sar	\$5,$num		# cf=0
1046	lea	($np,$N[1],8),$np
1047	mov	56+8(%rsp),$rp		# restore $rp
1048	jmp	.Lsub4x
1049
1050.align	32
1051.Lsub4x:
1052	.byte	0x66
1053	mov	8*0($tp),@ri[0]
1054	mov	8*1($tp),@ri[1]
1055	.byte	0x66
1056	sbb	16*0($np),@ri[0]
1057	mov	8*2($tp),@ri[2]
1058	sbb	16*1($np),@ri[1]
1059	mov	3*8($tp),@ri[3]
1060	lea	4*8($tp),$tp
1061	sbb	16*2($np),@ri[2]
1062	mov	@ri[0],8*0($rp)
1063	sbb	16*3($np),@ri[3]
1064	lea	16*4($np),$np
1065	mov	@ri[1],8*1($rp)
1066	mov	@ri[2],8*2($rp)
1067	mov	@ri[3],8*3($rp)
1068	lea	8*4($rp),$rp
1069
1070	inc	$num
1071	jnz	.Lsub4x
1072
1073	ret
1074___
1075}
1076$code.=<<___;
1077.size	mul4x_internal,.-mul4x_internal
1078___
1079}}}
1080{{{
1081######################################################################
1082# void bn_power5(
1083my $rptr="%rdi";	# BN_ULONG *rptr,
1084my $aptr="%rsi";	# const BN_ULONG *aptr,
1085my $bptr="%rdx";	# const void *table,
1086my $nptr="%rcx";	# const BN_ULONG *nptr,
1087my $n0  ="%r8";		# const BN_ULONG *n0);
1088my $num ="%r9";		# int num, has to be divisible by 8
1089			# int pwr
1090
1091my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
1092my @A0=("%r10","%r11");
1093my @A1=("%r12","%r13");
1094my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
1095
1096$code.=<<___;
1097.globl	bn_power5
1098.type	bn_power5,\@function,6
1099.align	32
1100bn_power5:
1101.cfi_startproc
1102	mov	%rsp,%rax
1103.cfi_def_cfa_register	%rax
1104___
1105$code.=<<___ if ($addx);
1106	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
1107	and	\$0x80108,%r11d
1108	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
1109	je	.Lpowerx5_enter
1110___
1111$code.=<<___;
1112	push	%rbx
1113.cfi_push	%rbx
1114	push	%rbp
1115.cfi_push	%rbp
1116	push	%r12
1117.cfi_push	%r12
1118	push	%r13
1119.cfi_push	%r13
1120	push	%r14
1121.cfi_push	%r14
1122	push	%r15
1123.cfi_push	%r15
1124.Lpower5_prologue:
1125
1126	shl	\$3,${num}d		# convert $num to bytes
1127	lea	($num,$num,2),%r10d	# 3*$num
1128	neg	$num
1129	mov	($n0),$n0		# *n0
1130
1131	##############################################################
1132	# Ensure that stack frame doesn't alias with $rptr+3*$num
1133	# modulo 4096, which covers ret[num], am[num] and n[num]
1134	# (see bn_exp.c). This is done to allow memory disambiguation
1135	# logic do its magic. [Extra 256 bytes is for power mask
1136	# calculated from 7th argument, the index.]
1137	#
1138	lea	-320(%rsp,$num,2),%r11
1139	mov	%rsp,%rbp
1140	sub	$rptr,%r11
1141	and	\$4095,%r11
1142	cmp	%r11,%r10
1143	jb	.Lpwr_sp_alt
1144	sub	%r11,%rbp		# align with $aptr
1145	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1146	jmp	.Lpwr_sp_done
1147
1148.align	32
1149.Lpwr_sp_alt:
1150	lea	4096-320(,$num,2),%r10
1151	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1152	sub	%r10,%r11
1153	mov	\$0,%r10
1154	cmovc	%r10,%r11
1155	sub	%r11,%rbp
1156.Lpwr_sp_done:
1157	and	\$-64,%rbp
1158	mov	%rsp,%r11
1159	sub	%rbp,%r11
1160	and	\$-4096,%r11
1161	lea	(%rbp,%r11),%rsp
1162	mov	(%rsp),%r10
1163	cmp	%rbp,%rsp
1164	ja	.Lpwr_page_walk
1165	jmp	.Lpwr_page_walk_done
1166
1167.Lpwr_page_walk:
1168	lea	-4096(%rsp),%rsp
1169	mov	(%rsp),%r10
1170	cmp	%rbp,%rsp
1171	ja	.Lpwr_page_walk
1172.Lpwr_page_walk_done:
1173
1174	mov	$num,%r10
1175	neg	$num
1176
1177	##############################################################
1178	# Stack layout
1179	#
1180	# +0	saved $num, used in reduction section
1181	# +8	&t[2*$num], used in reduction section
1182	# +32	saved *n0
1183	# +40	saved %rsp
1184	# +48	t[2*$num]
1185	#
1186	mov	$n0,  32(%rsp)
1187	mov	%rax, 40(%rsp)		# save original %rsp
1188.cfi_cfa_expression	%rsp+40,deref,+8
1189.Lpower5_body:
1190	movq	$rptr,%xmm1		# save $rptr, used in sqr8x
1191	movq	$nptr,%xmm2		# save $nptr
1192	movq	%r10, %xmm3		# -$num, used in sqr8x
1193	movq	$bptr,%xmm4
1194
1195	call	__bn_sqr8x_internal
1196	call	__bn_post4x_internal
1197	call	__bn_sqr8x_internal
1198	call	__bn_post4x_internal
1199	call	__bn_sqr8x_internal
1200	call	__bn_post4x_internal
1201	call	__bn_sqr8x_internal
1202	call	__bn_post4x_internal
1203	call	__bn_sqr8x_internal
1204	call	__bn_post4x_internal
1205
1206	movq	%xmm2,$nptr
1207	movq	%xmm4,$bptr
1208	mov	$aptr,$rptr
1209	mov	40(%rsp),%rax
1210	lea	32(%rsp),$n0
1211
1212	call	mul4x_internal
1213
1214	mov	40(%rsp),%rsi		# restore %rsp
1215.cfi_def_cfa	%rsi,8
1216	mov	\$1,%rax
1217	mov	-48(%rsi),%r15
1218.cfi_restore	%r15
1219	mov	-40(%rsi),%r14
1220.cfi_restore	%r14
1221	mov	-32(%rsi),%r13
1222.cfi_restore	%r13
1223	mov	-24(%rsi),%r12
1224.cfi_restore	%r12
1225	mov	-16(%rsi),%rbp
1226.cfi_restore	%rbp
1227	mov	-8(%rsi),%rbx
1228.cfi_restore	%rbx
1229	lea	(%rsi),%rsp
1230.cfi_def_cfa_register	%rsp
1231.Lpower5_epilogue:
1232	ret
1233.cfi_endproc
1234.size	bn_power5,.-bn_power5
1235
1236.globl	bn_sqr8x_internal
1237.hidden	bn_sqr8x_internal
1238.type	bn_sqr8x_internal,\@abi-omnipotent
1239.align	32
1240bn_sqr8x_internal:
1241__bn_sqr8x_internal:
1242	##############################################################
1243	# Squaring part:
1244	#
1245	# a) multiply-n-add everything but a[i]*a[i];
1246	# b) shift result of a) by 1 to the left and accumulate
1247	#    a[i]*a[i] products;
1248	#
1249	##############################################################
1250	#                                                     a[1]a[0]
1251	#                                                 a[2]a[0]
1252	#                                             a[3]a[0]
1253	#                                             a[2]a[1]
1254	#                                         a[4]a[0]
1255	#                                         a[3]a[1]
1256	#                                     a[5]a[0]
1257	#                                     a[4]a[1]
1258	#                                     a[3]a[2]
1259	#                                 a[6]a[0]
1260	#                                 a[5]a[1]
1261	#                                 a[4]a[2]
1262	#                             a[7]a[0]
1263	#                             a[6]a[1]
1264	#                             a[5]a[2]
1265	#                             a[4]a[3]
1266	#                         a[7]a[1]
1267	#                         a[6]a[2]
1268	#                         a[5]a[3]
1269	#                     a[7]a[2]
1270	#                     a[6]a[3]
1271	#                     a[5]a[4]
1272	#                 a[7]a[3]
1273	#                 a[6]a[4]
1274	#             a[7]a[4]
1275	#             a[6]a[5]
1276	#         a[7]a[5]
1277	#     a[7]a[6]
1278	#                                                     a[1]a[0]
1279	#                                                 a[2]a[0]
1280	#                                             a[3]a[0]
1281	#                                         a[4]a[0]
1282	#                                     a[5]a[0]
1283	#                                 a[6]a[0]
1284	#                             a[7]a[0]
1285	#                                             a[2]a[1]
1286	#                                         a[3]a[1]
1287	#                                     a[4]a[1]
1288	#                                 a[5]a[1]
1289	#                             a[6]a[1]
1290	#                         a[7]a[1]
1291	#                                     a[3]a[2]
1292	#                                 a[4]a[2]
1293	#                             a[5]a[2]
1294	#                         a[6]a[2]
1295	#                     a[7]a[2]
1296	#                             a[4]a[3]
1297	#                         a[5]a[3]
1298	#                     a[6]a[3]
1299	#                 a[7]a[3]
1300	#                     a[5]a[4]
1301	#                 a[6]a[4]
1302	#             a[7]a[4]
1303	#             a[6]a[5]
1304	#         a[7]a[5]
1305	#     a[7]a[6]
1306	#                                                         a[0]a[0]
1307	#                                                 a[1]a[1]
1308	#                                         a[2]a[2]
1309	#                                 a[3]a[3]
1310	#                         a[4]a[4]
1311	#                 a[5]a[5]
1312	#         a[6]a[6]
1313	# a[7]a[7]
1314
1315	lea	32(%r10),$i		# $i=-($num-32)
1316	lea	($aptr,$num),$aptr	# end of a[] buffer, ($aptr,$i)=&ap[2]
1317
1318	mov	$num,$j			# $j=$num
1319
1320					# comments apply to $num==8 case
1321	mov	-32($aptr,$i),$a0	# a[0]
1322	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1323	mov	-24($aptr,$i),%rax	# a[1]
1324	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1325	mov	-16($aptr,$i),$ai	# a[2]
1326	mov	%rax,$a1
1327
1328	mul	$a0			# a[1]*a[0]
1329	mov	%rax,$A0[0]		# a[1]*a[0]
1330	 mov	$ai,%rax		# a[2]
1331	mov	%rdx,$A0[1]
1332	mov	$A0[0],-24($tptr,$i)	# t[1]
1333
1334	mul	$a0			# a[2]*a[0]
1335	add	%rax,$A0[1]
1336	 mov	$ai,%rax
1337	adc	\$0,%rdx
1338	mov	$A0[1],-16($tptr,$i)	# t[2]
1339	mov	%rdx,$A0[0]
1340
1341
1342	 mov	-8($aptr,$i),$ai	# a[3]
1343	mul	$a1			# a[2]*a[1]
1344	mov	%rax,$A1[0]		# a[2]*a[1]+t[3]
1345	 mov	$ai,%rax
1346	mov	%rdx,$A1[1]
1347
1348	 lea	($i),$j
1349	mul	$a0			# a[3]*a[0]
1350	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1351	 mov	$ai,%rax
1352	mov	%rdx,$A0[1]
1353	adc	\$0,$A0[1]
1354	add	$A1[0],$A0[0]
1355	adc	\$0,$A0[1]
1356	mov	$A0[0],-8($tptr,$j)	# t[3]
1357	jmp	.Lsqr4x_1st
1358
1359.align	32
1360.Lsqr4x_1st:
1361	 mov	($aptr,$j),$ai		# a[4]
1362	mul	$a1			# a[3]*a[1]
1363	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1364	 mov	$ai,%rax
1365	mov	%rdx,$A1[0]
1366	adc	\$0,$A1[0]
1367
1368	mul	$a0			# a[4]*a[0]
1369	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1370	 mov	$ai,%rax		# a[3]
1371	 mov	8($aptr,$j),$ai		# a[5]
1372	mov	%rdx,$A0[0]
1373	adc	\$0,$A0[0]
1374	add	$A1[1],$A0[1]
1375	adc	\$0,$A0[0]
1376
1377
1378	mul	$a1			# a[4]*a[3]
1379	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1380	 mov	$ai,%rax
1381	 mov	$A0[1],($tptr,$j)	# t[4]
1382	mov	%rdx,$A1[1]
1383	adc	\$0,$A1[1]
1384
1385	mul	$a0			# a[5]*a[2]
1386	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1387	 mov	$ai,%rax
1388	 mov	16($aptr,$j),$ai	# a[6]
1389	mov	%rdx,$A0[1]
1390	adc	\$0,$A0[1]
1391	add	$A1[0],$A0[0]
1392	adc	\$0,$A0[1]
1393
1394	mul	$a1			# a[5]*a[3]
1395	add	%rax,$A1[1]		# a[5]*a[3]+t[6]
1396	 mov	$ai,%rax
1397	 mov	$A0[0],8($tptr,$j)	# t[5]
1398	mov	%rdx,$A1[0]
1399	adc	\$0,$A1[0]
1400
1401	mul	$a0			# a[6]*a[2]
1402	add	%rax,$A0[1]		# a[6]*a[2]+a[5]*a[3]+t[6]
1403	 mov	$ai,%rax		# a[3]
1404	 mov	24($aptr,$j),$ai	# a[7]
1405	mov	%rdx,$A0[0]
1406	adc	\$0,$A0[0]
1407	add	$A1[1],$A0[1]
1408	adc	\$0,$A0[0]
1409
1410
1411	mul	$a1			# a[6]*a[5]
1412	add	%rax,$A1[0]		# a[6]*a[5]+t[7]
1413	 mov	$ai,%rax
1414	 mov	$A0[1],16($tptr,$j)	# t[6]
1415	mov	%rdx,$A1[1]
1416	adc	\$0,$A1[1]
1417	 lea	32($j),$j
1418
1419	mul	$a0			# a[7]*a[4]
1420	add	%rax,$A0[0]		# a[7]*a[4]+a[6]*a[5]+t[6]
1421	 mov	$ai,%rax
1422	mov	%rdx,$A0[1]
1423	adc	\$0,$A0[1]
1424	add	$A1[0],$A0[0]
1425	adc	\$0,$A0[1]
1426	mov	$A0[0],-8($tptr,$j)	# t[7]
1427
1428	cmp	\$0,$j
1429	jne	.Lsqr4x_1st
1430
1431	mul	$a1			# a[7]*a[5]
1432	add	%rax,$A1[1]
1433	lea	16($i),$i
1434	adc	\$0,%rdx
1435	add	$A0[1],$A1[1]
1436	adc	\$0,%rdx
1437
1438	mov	$A1[1],($tptr)		# t[8]
1439	mov	%rdx,$A1[0]
1440	mov	%rdx,8($tptr)		# t[9]
1441	jmp	.Lsqr4x_outer
1442
1443.align	32
1444.Lsqr4x_outer:				# comments apply to $num==6 case
1445	mov	-32($aptr,$i),$a0	# a[0]
1446	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1447	mov	-24($aptr,$i),%rax	# a[1]
1448	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1449	mov	-16($aptr,$i),$ai	# a[2]
1450	mov	%rax,$a1
1451
1452	mul	$a0			# a[1]*a[0]
1453	mov	-24($tptr,$i),$A0[0]	# t[1]
1454	add	%rax,$A0[0]		# a[1]*a[0]+t[1]
1455	 mov	$ai,%rax		# a[2]
1456	adc	\$0,%rdx
1457	mov	$A0[0],-24($tptr,$i)	# t[1]
1458	mov	%rdx,$A0[1]
1459
1460	mul	$a0			# a[2]*a[0]
1461	add	%rax,$A0[1]
1462	 mov	$ai,%rax
1463	adc	\$0,%rdx
1464	add	-16($tptr,$i),$A0[1]	# a[2]*a[0]+t[2]
1465	mov	%rdx,$A0[0]
1466	adc	\$0,$A0[0]
1467	mov	$A0[1],-16($tptr,$i)	# t[2]
1468
1469	xor	$A1[0],$A1[0]
1470
1471	 mov	-8($aptr,$i),$ai	# a[3]
1472	mul	$a1			# a[2]*a[1]
1473	add	%rax,$A1[0]		# a[2]*a[1]+t[3]
1474	 mov	$ai,%rax
1475	adc	\$0,%rdx
1476	add	-8($tptr,$i),$A1[0]
1477	mov	%rdx,$A1[1]
1478	adc	\$0,$A1[1]
1479
1480	mul	$a0			# a[3]*a[0]
1481	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1482	 mov	$ai,%rax
1483	adc	\$0,%rdx
1484	add	$A1[0],$A0[0]
1485	mov	%rdx,$A0[1]
1486	adc	\$0,$A0[1]
1487	mov	$A0[0],-8($tptr,$i)	# t[3]
1488
1489	lea	($i),$j
1490	jmp	.Lsqr4x_inner
1491
1492.align	32
1493.Lsqr4x_inner:
1494	 mov	($aptr,$j),$ai		# a[4]
1495	mul	$a1			# a[3]*a[1]
1496	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1497	 mov	$ai,%rax
1498	mov	%rdx,$A1[0]
1499	adc	\$0,$A1[0]
1500	add	($tptr,$j),$A1[1]
1501	adc	\$0,$A1[0]
1502
1503	.byte	0x67
1504	mul	$a0			# a[4]*a[0]
1505	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1506	 mov	$ai,%rax		# a[3]
1507	 mov	8($aptr,$j),$ai		# a[5]
1508	mov	%rdx,$A0[0]
1509	adc	\$0,$A0[0]
1510	add	$A1[1],$A0[1]
1511	adc	\$0,$A0[0]
1512
1513	mul	$a1			# a[4]*a[3]
1514	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1515	mov	$A0[1],($tptr,$j)	# t[4]
1516	 mov	$ai,%rax
1517	mov	%rdx,$A1[1]
1518	adc	\$0,$A1[1]
1519	add	8($tptr,$j),$A1[0]
1520	lea	16($j),$j		# j++
1521	adc	\$0,$A1[1]
1522
1523	mul	$a0			# a[5]*a[2]
1524	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1525	 mov	$ai,%rax
1526	adc	\$0,%rdx
1527	add	$A1[0],$A0[0]
1528	mov	%rdx,$A0[1]
1529	adc	\$0,$A0[1]
1530	mov	$A0[0],-8($tptr,$j)	# t[5], "preloaded t[1]" below
1531
1532	cmp	\$0,$j
1533	jne	.Lsqr4x_inner
1534
1535	.byte	0x67
1536	mul	$a1			# a[5]*a[3]
1537	add	%rax,$A1[1]
1538	adc	\$0,%rdx
1539	add	$A0[1],$A1[1]
1540	adc	\$0,%rdx
1541
1542	mov	$A1[1],($tptr)		# t[6], "preloaded t[2]" below
1543	mov	%rdx,$A1[0]
1544	mov	%rdx,8($tptr)		# t[7], "preloaded t[3]" below
1545
1546	add	\$16,$i
1547	jnz	.Lsqr4x_outer
1548
1549					# comments apply to $num==4 case
1550	mov	-32($aptr),$a0		# a[0]
1551	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1552	mov	-24($aptr),%rax		# a[1]
1553	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1554	mov	-16($aptr),$ai		# a[2]
1555	mov	%rax,$a1
1556
1557	mul	$a0			# a[1]*a[0]
1558	add	%rax,$A0[0]		# a[1]*a[0]+t[1], preloaded t[1]
1559	 mov	$ai,%rax		# a[2]
1560	mov	%rdx,$A0[1]
1561	adc	\$0,$A0[1]
1562
1563	mul	$a0			# a[2]*a[0]
1564	add	%rax,$A0[1]
1565	 mov	$ai,%rax
1566	 mov	$A0[0],-24($tptr)	# t[1]
1567	mov	%rdx,$A0[0]
1568	adc	\$0,$A0[0]
1569	add	$A1[1],$A0[1]		# a[2]*a[0]+t[2], preloaded t[2]
1570	 mov	-8($aptr),$ai		# a[3]
1571	adc	\$0,$A0[0]
1572
1573	mul	$a1			# a[2]*a[1]
1574	add	%rax,$A1[0]		# a[2]*a[1]+t[3], preloaded t[3]
1575	 mov	$ai,%rax
1576	 mov	$A0[1],-16($tptr)	# t[2]
1577	mov	%rdx,$A1[1]
1578	adc	\$0,$A1[1]
1579
1580	mul	$a0			# a[3]*a[0]
1581	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1582	 mov	$ai,%rax
1583	mov	%rdx,$A0[1]
1584	adc	\$0,$A0[1]
1585	add	$A1[0],$A0[0]
1586	adc	\$0,$A0[1]
1587	mov	$A0[0],-8($tptr)	# t[3]
1588
1589	mul	$a1			# a[3]*a[1]
1590	add	%rax,$A1[1]
1591	 mov	-16($aptr),%rax		# a[2]
1592	adc	\$0,%rdx
1593	add	$A0[1],$A1[1]
1594	adc	\$0,%rdx
1595
1596	mov	$A1[1],($tptr)		# t[4]
1597	mov	%rdx,$A1[0]
1598	mov	%rdx,8($tptr)		# t[5]
1599
1600	mul	$ai			# a[2]*a[3]
1601___
1602{
1603my ($shift,$carry)=($a0,$a1);
1604my @S=(@A1,$ai,$n0);
1605$code.=<<___;
1606	 add	\$16,$i
1607	 xor	$shift,$shift
1608	 sub	$num,$i			# $i=16-$num
1609	 xor	$carry,$carry
1610
1611	add	$A1[0],%rax		# t[5]
1612	adc	\$0,%rdx
1613	mov	%rax,8($tptr)		# t[5]
1614	mov	%rdx,16($tptr)		# t[6]
1615	mov	$carry,24($tptr)	# t[7]
1616
1617	 mov	-16($aptr,$i),%rax	# a[0]
1618	lea	48+8(%rsp),$tptr
1619	 xor	$A0[0],$A0[0]		# t[0]
1620	 mov	8($tptr),$A0[1]		# t[1]
1621
1622	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1623	shr	\$63,$A0[0]
1624	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1625	shr	\$63,$A0[1]
1626	or	$A0[0],$S[1]		# | t[2*i]>>63
1627	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1628	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1629	mul	%rax			# a[i]*a[i]
1630	neg	$carry			# mov $carry,cf
1631	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1632	adc	%rax,$S[0]
1633	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1634	mov	$S[0],($tptr)
1635	adc	%rdx,$S[1]
1636
1637	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1638	 mov	$S[1],8($tptr)
1639	 sbb	$carry,$carry		# mov cf,$carry
1640	shr	\$63,$A0[0]
1641	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1642	shr	\$63,$A0[1]
1643	or	$A0[0],$S[3]		# | t[2*i]>>63
1644	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1645	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1646	mul	%rax			# a[i]*a[i]
1647	neg	$carry			# mov $carry,cf
1648	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1649	adc	%rax,$S[2]
1650	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1651	mov	$S[2],16($tptr)
1652	adc	%rdx,$S[3]
1653	lea	16($i),$i
1654	mov	$S[3],24($tptr)
1655	sbb	$carry,$carry		# mov cf,$carry
1656	lea	64($tptr),$tptr
1657	jmp	.Lsqr4x_shift_n_add
1658
1659.align	32
1660.Lsqr4x_shift_n_add:
1661	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1662	shr	\$63,$A0[0]
1663	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1664	shr	\$63,$A0[1]
1665	or	$A0[0],$S[1]		# | t[2*i]>>63
1666	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1667	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1668	mul	%rax			# a[i]*a[i]
1669	neg	$carry			# mov $carry,cf
1670	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1671	adc	%rax,$S[0]
1672	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1673	mov	$S[0],-32($tptr)
1674	adc	%rdx,$S[1]
1675
1676	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1677	 mov	$S[1],-24($tptr)
1678	 sbb	$carry,$carry		# mov cf,$carry
1679	shr	\$63,$A0[0]
1680	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1681	shr	\$63,$A0[1]
1682	or	$A0[0],$S[3]		# | t[2*i]>>63
1683	 mov	0($tptr),$A0[0]		# t[2*i+2]	# prefetch
1684	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1685	mul	%rax			# a[i]*a[i]
1686	neg	$carry			# mov $carry,cf
1687	 mov	8($tptr),$A0[1]		# t[2*i+2+1]	# prefetch
1688	adc	%rax,$S[2]
1689	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1690	mov	$S[2],-16($tptr)
1691	adc	%rdx,$S[3]
1692
1693	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1694	 mov	$S[3],-8($tptr)
1695	 sbb	$carry,$carry		# mov cf,$carry
1696	shr	\$63,$A0[0]
1697	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1698	shr	\$63,$A0[1]
1699	or	$A0[0],$S[1]		# | t[2*i]>>63
1700	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1701	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1702	mul	%rax			# a[i]*a[i]
1703	neg	$carry			# mov $carry,cf
1704	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1705	adc	%rax,$S[0]
1706	 mov	8($aptr,$i),%rax	# a[i+1]	# prefetch
1707	mov	$S[0],0($tptr)
1708	adc	%rdx,$S[1]
1709
1710	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1711	 mov	$S[1],8($tptr)
1712	 sbb	$carry,$carry		# mov cf,$carry
1713	shr	\$63,$A0[0]
1714	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1715	shr	\$63,$A0[1]
1716	or	$A0[0],$S[3]		# | t[2*i]>>63
1717	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1718	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1719	mul	%rax			# a[i]*a[i]
1720	neg	$carry			# mov $carry,cf
1721	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1722	adc	%rax,$S[2]
1723	 mov	16($aptr,$i),%rax	# a[i+1]	# prefetch
1724	mov	$S[2],16($tptr)
1725	adc	%rdx,$S[3]
1726	mov	$S[3],24($tptr)
1727	sbb	$carry,$carry		# mov cf,$carry
1728	lea	64($tptr),$tptr
1729	add	\$32,$i
1730	jnz	.Lsqr4x_shift_n_add
1731
1732	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1733	.byte	0x67
1734	shr	\$63,$A0[0]
1735	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1736	shr	\$63,$A0[1]
1737	or	$A0[0],$S[1]		# | t[2*i]>>63
1738	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1739	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1740	mul	%rax			# a[i]*a[i]
1741	neg	$carry			# mov $carry,cf
1742	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1743	adc	%rax,$S[0]
1744	 mov	-8($aptr),%rax		# a[i+1]	# prefetch
1745	mov	$S[0],-32($tptr)
1746	adc	%rdx,$S[1]
1747
1748	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1|shift
1749	 mov	$S[1],-24($tptr)
1750	 sbb	$carry,$carry		# mov cf,$carry
1751	shr	\$63,$A0[0]
1752	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1753	shr	\$63,$A0[1]
1754	or	$A0[0],$S[3]		# | t[2*i]>>63
1755	mul	%rax			# a[i]*a[i]
1756	neg	$carry			# mov $carry,cf
1757	adc	%rax,$S[2]
1758	adc	%rdx,$S[3]
1759	mov	$S[2],-16($tptr)
1760	mov	$S[3],-8($tptr)
1761___
1762}
1763######################################################################
1764# Montgomery reduction part, "word-by-word" algorithm.
1765#
1766# This new path is inspired by multiple submissions from Intel, by
1767# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
1768# Vinodh Gopal...
1769{
1770my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx");
1771
1772$code.=<<___;
1773	movq	%xmm2,$nptr
1774__bn_sqr8x_reduction:
1775	xor	%rax,%rax
1776	lea	($nptr,$num),%rcx	# end of n[]
1777	lea	48+8(%rsp,$num,2),%rdx	# end of t[] buffer
1778	mov	%rcx,0+8(%rsp)
1779	lea	48+8(%rsp,$num),$tptr	# end of initial t[] window
1780	mov	%rdx,8+8(%rsp)
1781	neg	$num
1782	jmp	.L8x_reduction_loop
1783
1784.align	32
1785.L8x_reduction_loop:
1786	lea	($tptr,$num),$tptr	# start of current t[] window
1787	.byte	0x66
1788	mov	8*0($tptr),$m0
1789	mov	8*1($tptr),%r9
1790	mov	8*2($tptr),%r10
1791	mov	8*3($tptr),%r11
1792	mov	8*4($tptr),%r12
1793	mov	8*5($tptr),%r13
1794	mov	8*6($tptr),%r14
1795	mov	8*7($tptr),%r15
1796	mov	%rax,(%rdx)		# store top-most carry bit
1797	lea	8*8($tptr),$tptr
1798
1799	.byte	0x67
1800	mov	$m0,%r8
1801	imulq	32+8(%rsp),$m0		# n0*a[0]
1802	mov	8*0($nptr),%rax		# n[0]
1803	mov	\$8,%ecx
1804	jmp	.L8x_reduce
1805
1806.align	32
1807.L8x_reduce:
1808	mulq	$m0
1809	 mov	8*1($nptr),%rax		# n[1]
1810	neg	%r8
1811	mov	%rdx,%r8
1812	adc	\$0,%r8
1813
1814	mulq	$m0
1815	add	%rax,%r9
1816	 mov	8*2($nptr),%rax
1817	adc	\$0,%rdx
1818	add	%r9,%r8
1819	 mov	$m0,48-8+8(%rsp,%rcx,8)	# put aside n0*a[i]
1820	mov	%rdx,%r9
1821	adc	\$0,%r9
1822
1823	mulq	$m0
1824	add	%rax,%r10
1825	 mov	8*3($nptr),%rax
1826	adc	\$0,%rdx
1827	add	%r10,%r9
1828	 mov	32+8(%rsp),$carry	# pull n0, borrow $carry
1829	mov	%rdx,%r10
1830	adc	\$0,%r10
1831
1832	mulq	$m0
1833	add	%rax,%r11
1834	 mov	8*4($nptr),%rax
1835	adc	\$0,%rdx
1836	 imulq	%r8,$carry		# modulo-scheduled
1837	add	%r11,%r10
1838	mov	%rdx,%r11
1839	adc	\$0,%r11
1840
1841	mulq	$m0
1842	add	%rax,%r12
1843	 mov	8*5($nptr),%rax
1844	adc	\$0,%rdx
1845	add	%r12,%r11
1846	mov	%rdx,%r12
1847	adc	\$0,%r12
1848
1849	mulq	$m0
1850	add	%rax,%r13
1851	 mov	8*6($nptr),%rax
1852	adc	\$0,%rdx
1853	add	%r13,%r12
1854	mov	%rdx,%r13
1855	adc	\$0,%r13
1856
1857	mulq	$m0
1858	add	%rax,%r14
1859	 mov	8*7($nptr),%rax
1860	adc	\$0,%rdx
1861	add	%r14,%r13
1862	mov	%rdx,%r14
1863	adc	\$0,%r14
1864
1865	mulq	$m0
1866	 mov	$carry,$m0		# n0*a[i]
1867	add	%rax,%r15
1868	 mov	8*0($nptr),%rax		# n[0]
1869	adc	\$0,%rdx
1870	add	%r15,%r14
1871	mov	%rdx,%r15
1872	adc	\$0,%r15
1873
1874	dec	%ecx
1875	jnz	.L8x_reduce
1876
1877	lea	8*8($nptr),$nptr
1878	xor	%rax,%rax
1879	mov	8+8(%rsp),%rdx		# pull end of t[]
1880	cmp	0+8(%rsp),$nptr		# end of n[]?
1881	jae	.L8x_no_tail
1882
1883	.byte	0x66
1884	add	8*0($tptr),%r8
1885	adc	8*1($tptr),%r9
1886	adc	8*2($tptr),%r10
1887	adc	8*3($tptr),%r11
1888	adc	8*4($tptr),%r12
1889	adc	8*5($tptr),%r13
1890	adc	8*6($tptr),%r14
1891	adc	8*7($tptr),%r15
1892	sbb	$carry,$carry		# top carry
1893
1894	mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1895	mov	\$8,%ecx
1896	mov	8*0($nptr),%rax
1897	jmp	.L8x_tail
1898
1899.align	32
1900.L8x_tail:
1901	mulq	$m0
1902	add	%rax,%r8
1903	 mov	8*1($nptr),%rax
1904	 mov	%r8,($tptr)		# save result
1905	mov	%rdx,%r8
1906	adc	\$0,%r8
1907
1908	mulq	$m0
1909	add	%rax,%r9
1910	 mov	8*2($nptr),%rax
1911	adc	\$0,%rdx
1912	add	%r9,%r8
1913	 lea	8($tptr),$tptr		# $tptr++
1914	mov	%rdx,%r9
1915	adc	\$0,%r9
1916
1917	mulq	$m0
1918	add	%rax,%r10
1919	 mov	8*3($nptr),%rax
1920	adc	\$0,%rdx
1921	add	%r10,%r9
1922	mov	%rdx,%r10
1923	adc	\$0,%r10
1924
1925	mulq	$m0
1926	add	%rax,%r11
1927	 mov	8*4($nptr),%rax
1928	adc	\$0,%rdx
1929	add	%r11,%r10
1930	mov	%rdx,%r11
1931	adc	\$0,%r11
1932
1933	mulq	$m0
1934	add	%rax,%r12
1935	 mov	8*5($nptr),%rax
1936	adc	\$0,%rdx
1937	add	%r12,%r11
1938	mov	%rdx,%r12
1939	adc	\$0,%r12
1940
1941	mulq	$m0
1942	add	%rax,%r13
1943	 mov	8*6($nptr),%rax
1944	adc	\$0,%rdx
1945	add	%r13,%r12
1946	mov	%rdx,%r13
1947	adc	\$0,%r13
1948
1949	mulq	$m0
1950	add	%rax,%r14
1951	 mov	8*7($nptr),%rax
1952	adc	\$0,%rdx
1953	add	%r14,%r13
1954	mov	%rdx,%r14
1955	adc	\$0,%r14
1956
1957	mulq	$m0
1958	 mov	48-16+8(%rsp,%rcx,8),$m0# pull n0*a[i]
1959	add	%rax,%r15
1960	adc	\$0,%rdx
1961	add	%r15,%r14
1962	 mov	8*0($nptr),%rax		# pull n[0]
1963	mov	%rdx,%r15
1964	adc	\$0,%r15
1965
1966	dec	%ecx
1967	jnz	.L8x_tail
1968
1969	lea	8*8($nptr),$nptr
1970	mov	8+8(%rsp),%rdx		# pull end of t[]
1971	cmp	0+8(%rsp),$nptr		# end of n[]?
1972	jae	.L8x_tail_done		# break out of loop
1973
1974	 mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1975	neg	$carry
1976	 mov	8*0($nptr),%rax		# pull n[0]
1977	adc	8*0($tptr),%r8
1978	adc	8*1($tptr),%r9
1979	adc	8*2($tptr),%r10
1980	adc	8*3($tptr),%r11
1981	adc	8*4($tptr),%r12
1982	adc	8*5($tptr),%r13
1983	adc	8*6($tptr),%r14
1984	adc	8*7($tptr),%r15
1985	sbb	$carry,$carry		# top carry
1986
1987	mov	\$8,%ecx
1988	jmp	.L8x_tail
1989
1990.align	32
1991.L8x_tail_done:
1992	xor	%rax,%rax
1993	add	(%rdx),%r8		# can this overflow?
1994	adc	\$0,%r9
1995	adc	\$0,%r10
1996	adc	\$0,%r11
1997	adc	\$0,%r12
1998	adc	\$0,%r13
1999	adc	\$0,%r14
2000	adc	\$0,%r15
2001	adc	\$0,%rax
2002
2003	neg	$carry
2004.L8x_no_tail:
2005	adc	8*0($tptr),%r8
2006	adc	8*1($tptr),%r9
2007	adc	8*2($tptr),%r10
2008	adc	8*3($tptr),%r11
2009	adc	8*4($tptr),%r12
2010	adc	8*5($tptr),%r13
2011	adc	8*6($tptr),%r14
2012	adc	8*7($tptr),%r15
2013	adc	\$0,%rax		# top-most carry
2014	 mov	-8($nptr),%rcx		# np[num-1]
2015	 xor	$carry,$carry
2016
2017	movq	%xmm2,$nptr		# restore $nptr
2018
2019	mov	%r8,8*0($tptr)		# store top 512 bits
2020	mov	%r9,8*1($tptr)
2021	 movq	%xmm3,$num		# $num is %r9, can't be moved upwards
2022	mov	%r10,8*2($tptr)
2023	mov	%r11,8*3($tptr)
2024	mov	%r12,8*4($tptr)
2025	mov	%r13,8*5($tptr)
2026	mov	%r14,8*6($tptr)
2027	mov	%r15,8*7($tptr)
2028	lea	8*8($tptr),$tptr
2029
2030	cmp	%rdx,$tptr		# end of t[]?
2031	jb	.L8x_reduction_loop
2032	ret
2033.size	bn_sqr8x_internal,.-bn_sqr8x_internal
2034___
2035}
2036##############################################################
2037# Post-condition, 4x unrolled
2038#
2039{
2040my ($tptr,$nptr)=("%rbx","%rbp");
2041$code.=<<___;
2042.type	__bn_post4x_internal,\@abi-omnipotent
2043.align	32
2044__bn_post4x_internal:
2045	mov	8*0($nptr),%r12
2046	lea	(%rdi,$num),$tptr	# %rdi was $tptr above
2047	mov	$num,%rcx
2048	movq	%xmm1,$rptr		# restore $rptr
2049	neg	%rax
2050	movq	%xmm1,$aptr		# prepare for back-to-back call
2051	sar	\$3+2,%rcx
2052	dec	%r12			# so that after 'not' we get -n[0]
2053	xor	%r10,%r10
2054	mov	8*1($nptr),%r13
2055	mov	8*2($nptr),%r14
2056	mov	8*3($nptr),%r15
2057	jmp	.Lsqr4x_sub_entry
2058
2059.align	16
2060.Lsqr4x_sub:
2061	mov	8*0($nptr),%r12
2062	mov	8*1($nptr),%r13
2063	mov	8*2($nptr),%r14
2064	mov	8*3($nptr),%r15
2065.Lsqr4x_sub_entry:
2066	lea	8*4($nptr),$nptr
2067	not	%r12
2068	not	%r13
2069	not	%r14
2070	not	%r15
2071	and	%rax,%r12
2072	and	%rax,%r13
2073	and	%rax,%r14
2074	and	%rax,%r15
2075
2076	neg	%r10			# mov %r10,%cf
2077	adc	8*0($tptr),%r12
2078	adc	8*1($tptr),%r13
2079	adc	8*2($tptr),%r14
2080	adc	8*3($tptr),%r15
2081	mov	%r12,8*0($rptr)
2082	lea	8*4($tptr),$tptr
2083	mov	%r13,8*1($rptr)
2084	sbb	%r10,%r10		# mov %cf,%r10
2085	mov	%r14,8*2($rptr)
2086	mov	%r15,8*3($rptr)
2087	lea	8*4($rptr),$rptr
2088
2089	inc	%rcx			# pass %cf
2090	jnz	.Lsqr4x_sub
2091
2092	mov	$num,%r10		# prepare for back-to-back call
2093	neg	$num			# restore $num
2094	ret
2095.size	__bn_post4x_internal,.-__bn_post4x_internal
2096___
2097}
2098{
2099$code.=<<___;
2100.globl	bn_from_montgomery
2101.type	bn_from_montgomery,\@abi-omnipotent
2102.align	32
2103bn_from_montgomery:
2104	testl	\$7,`($win64?"48(%rsp)":"%r9d")`
2105	jz	bn_from_mont8x
2106	xor	%eax,%eax
2107	ret
2108.size	bn_from_montgomery,.-bn_from_montgomery
2109
2110.type	bn_from_mont8x,\@function,6
2111.align	32
2112bn_from_mont8x:
2113.cfi_startproc
2114	.byte	0x67
2115	mov	%rsp,%rax
2116.cfi_def_cfa_register	%rax
2117	push	%rbx
2118.cfi_push	%rbx
2119	push	%rbp
2120.cfi_push	%rbp
2121	push	%r12
2122.cfi_push	%r12
2123	push	%r13
2124.cfi_push	%r13
2125	push	%r14
2126.cfi_push	%r14
2127	push	%r15
2128.cfi_push	%r15
2129.Lfrom_prologue:
2130
2131	shl	\$3,${num}d		# convert $num to bytes
2132	lea	($num,$num,2),%r10	# 3*$num in bytes
2133	neg	$num
2134	mov	($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). The stack is allocated to aligned with
2140	# bn_power5's frame, and as bn_from_montgomery happens to be
2141	# last operation, we use the opportunity to cleanse it.
2142	#
2143	lea	-320(%rsp,$num,2),%r11
2144	mov	%rsp,%rbp
2145	sub	$rptr,%r11
2146	and	\$4095,%r11
2147	cmp	%r11,%r10
2148	jb	.Lfrom_sp_alt
2149	sub	%r11,%rbp		# align with $aptr
2150	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2151	jmp	.Lfrom_sp_done
2152
2153.align	32
2154.Lfrom_sp_alt:
2155	lea	4096-320(,$num,2),%r10
2156	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2157	sub	%r10,%r11
2158	mov	\$0,%r10
2159	cmovc	%r10,%r11
2160	sub	%r11,%rbp
2161.Lfrom_sp_done:
2162	and	\$-64,%rbp
2163	mov	%rsp,%r11
2164	sub	%rbp,%r11
2165	and	\$-4096,%r11
2166	lea	(%rbp,%r11),%rsp
2167	mov	(%rsp),%r10
2168	cmp	%rbp,%rsp
2169	ja	.Lfrom_page_walk
2170	jmp	.Lfrom_page_walk_done
2171
2172.Lfrom_page_walk:
2173	lea	-4096(%rsp),%rsp
2174	mov	(%rsp),%r10
2175	cmp	%rbp,%rsp
2176	ja	.Lfrom_page_walk
2177.Lfrom_page_walk_done:
2178
2179	mov	$num,%r10
2180	neg	$num
2181
2182	##############################################################
2183	# Stack layout
2184	#
2185	# +0	saved $num, used in reduction section
2186	# +8	&t[2*$num], used in reduction section
2187	# +32	saved *n0
2188	# +40	saved %rsp
2189	# +48	t[2*$num]
2190	#
2191	mov	$n0,  32(%rsp)
2192	mov	%rax, 40(%rsp)		# save original %rsp
2193.cfi_cfa_expression	%rsp+40,deref,+8
2194.Lfrom_body:
2195	mov	$num,%r11
2196	lea	48(%rsp),%rax
2197	pxor	%xmm0,%xmm0
2198	jmp	.Lmul_by_1
2199
2200.align	32
2201.Lmul_by_1:
2202	movdqu	($aptr),%xmm1
2203	movdqu	16($aptr),%xmm2
2204	movdqu	32($aptr),%xmm3
2205	movdqa	%xmm0,(%rax,$num)
2206	movdqu	48($aptr),%xmm4
2207	movdqa	%xmm0,16(%rax,$num)
2208	.byte	0x48,0x8d,0xb6,0x40,0x00,0x00,0x00	# lea	64($aptr),$aptr
2209	movdqa	%xmm1,(%rax)
2210	movdqa	%xmm0,32(%rax,$num)
2211	movdqa	%xmm2,16(%rax)
2212	movdqa	%xmm0,48(%rax,$num)
2213	movdqa	%xmm3,32(%rax)
2214	movdqa	%xmm4,48(%rax)
2215	lea	64(%rax),%rax
2216	sub	\$64,%r11
2217	jnz	.Lmul_by_1
2218
2219	movq	$rptr,%xmm1
2220	movq	$nptr,%xmm2
2221	.byte	0x67
2222	mov	$nptr,%rbp
2223	movq	%r10, %xmm3		# -num
2224___
2225$code.=<<___ if ($addx);
2226	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
2227	and	\$0x80108,%r11d
2228	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
2229	jne	.Lfrom_mont_nox
2230
2231	lea	(%rax,$num),$rptr
2232	call	__bn_sqrx8x_reduction
2233	call	__bn_postx4x_internal
2234
2235	pxor	%xmm0,%xmm0
2236	lea	48(%rsp),%rax
2237	jmp	.Lfrom_mont_zero
2238
2239.align	32
2240.Lfrom_mont_nox:
2241___
2242$code.=<<___;
2243	call	__bn_sqr8x_reduction
2244	call	__bn_post4x_internal
2245
2246	pxor	%xmm0,%xmm0
2247	lea	48(%rsp),%rax
2248	jmp	.Lfrom_mont_zero
2249
2250.align	32
2251.Lfrom_mont_zero:
2252	mov	40(%rsp),%rsi		# restore %rsp
2253.cfi_def_cfa	%rsi,8
2254	movdqa	%xmm0,16*0(%rax)
2255	movdqa	%xmm0,16*1(%rax)
2256	movdqa	%xmm0,16*2(%rax)
2257	movdqa	%xmm0,16*3(%rax)
2258	lea	16*4(%rax),%rax
2259	sub	\$32,$num
2260	jnz	.Lfrom_mont_zero
2261
2262	mov	\$1,%rax
2263	mov	-48(%rsi),%r15
2264.cfi_restore	%r15
2265	mov	-40(%rsi),%r14
2266.cfi_restore	%r14
2267	mov	-32(%rsi),%r13
2268.cfi_restore	%r13
2269	mov	-24(%rsi),%r12
2270.cfi_restore	%r12
2271	mov	-16(%rsi),%rbp
2272.cfi_restore	%rbp
2273	mov	-8(%rsi),%rbx
2274.cfi_restore	%rbx
2275	lea	(%rsi),%rsp
2276.cfi_def_cfa_register	%rsp
2277.Lfrom_epilogue:
2278	ret
2279.cfi_endproc
2280.size	bn_from_mont8x,.-bn_from_mont8x
2281___
2282}
2283}}}
2284
2285if ($addx) {{{
2286my $bp="%rdx";	# restore original value
2287
2288$code.=<<___;
2289.type	bn_mulx4x_mont_gather5,\@function,6
2290.align	32
2291bn_mulx4x_mont_gather5:
2292.cfi_startproc
2293	mov	%rsp,%rax
2294.cfi_def_cfa_register	%rax
2295.Lmulx4x_enter:
2296	push	%rbx
2297.cfi_push	%rbx
2298	push	%rbp
2299.cfi_push	%rbp
2300	push	%r12
2301.cfi_push	%r12
2302	push	%r13
2303.cfi_push	%r13
2304	push	%r14
2305.cfi_push	%r14
2306	push	%r15
2307.cfi_push	%r15
2308.Lmulx4x_prologue:
2309
2310	shl	\$3,${num}d		# convert $num to bytes
2311	lea	($num,$num,2),%r10	# 3*$num in bytes
2312	neg	$num			# -$num
2313	mov	($n0),$n0		# *n0
2314
2315	##############################################################
2316	# Ensure that stack frame doesn't alias with $rptr+3*$num
2317	# modulo 4096, which covers ret[num], am[num] and n[num]
2318	# (see bn_exp.c). This is done to allow memory disambiguation
2319	# logic do its magic. [Extra [num] is allocated in order
2320	# to align with bn_power5's frame, which is cleansed after
2321	# completing exponentiation. Extra 256 bytes is for power mask
2322	# calculated from 7th argument, the index.]
2323	#
2324	lea	-320(%rsp,$num,2),%r11
2325	mov	%rsp,%rbp
2326	sub	$rp,%r11
2327	and	\$4095,%r11
2328	cmp	%r11,%r10
2329	jb	.Lmulx4xsp_alt
2330	sub	%r11,%rbp		# align with $aptr
2331	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2332	jmp	.Lmulx4xsp_done
2333
2334.Lmulx4xsp_alt:
2335	lea	4096-320(,$num,2),%r10
2336	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2337	sub	%r10,%r11
2338	mov	\$0,%r10
2339	cmovc	%r10,%r11
2340	sub	%r11,%rbp
2341.Lmulx4xsp_done:
2342	and	\$-64,%rbp		# ensure alignment
2343	mov	%rsp,%r11
2344	sub	%rbp,%r11
2345	and	\$-4096,%r11
2346	lea	(%rbp,%r11),%rsp
2347	mov	(%rsp),%r10
2348	cmp	%rbp,%rsp
2349	ja	.Lmulx4x_page_walk
2350	jmp	.Lmulx4x_page_walk_done
2351
2352.Lmulx4x_page_walk:
2353	lea	-4096(%rsp),%rsp
2354	mov	(%rsp),%r10
2355	cmp	%rbp,%rsp
2356	ja	.Lmulx4x_page_walk
2357.Lmulx4x_page_walk_done:
2358
2359	##############################################################
2360	# Stack layout
2361	# +0	-num
2362	# +8	off-loaded &b[i]
2363	# +16	end of b[num]
2364	# +24	inner counter
2365	# +32	saved n0
2366	# +40	saved %rsp
2367	# +48
2368	# +56	saved rp
2369	# +64	tmp[num+1]
2370	#
2371	mov	$n0, 32(%rsp)		# save *n0
2372	mov	%rax,40(%rsp)		# save original %rsp
2373.cfi_cfa_expression	%rsp+40,deref,+8
2374.Lmulx4x_body:
2375	call	mulx4x_internal
2376
2377	mov	40(%rsp),%rsi		# restore %rsp
2378.cfi_def_cfa	%rsi,8
2379	mov	\$1,%rax
2380
2381	mov	-48(%rsi),%r15
2382.cfi_restore	%r15
2383	mov	-40(%rsi),%r14
2384.cfi_restore	%r14
2385	mov	-32(%rsi),%r13
2386.cfi_restore	%r13
2387	mov	-24(%rsi),%r12
2388.cfi_restore	%r12
2389	mov	-16(%rsi),%rbp
2390.cfi_restore	%rbp
2391	mov	-8(%rsi),%rbx
2392.cfi_restore	%rbx
2393	lea	(%rsi),%rsp
2394.cfi_def_cfa_register	%rsp
2395.Lmulx4x_epilogue:
2396	ret
2397.cfi_endproc
2398.size	bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5
2399
2400.type	mulx4x_internal,\@abi-omnipotent
2401.align	32
2402mulx4x_internal:
2403	mov	$num,8(%rsp)		# save -$num (it was in bytes)
2404	mov	$num,%r10
2405	neg	$num			# restore $num
2406	shl	\$5,$num
2407	neg	%r10			# restore $num
2408	lea	128($bp,$num),%r13	# end of powers table (+size optimization)
2409	shr	\$5+5,$num
2410	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument
2411	sub	\$1,$num
2412	lea	.Linc(%rip),%rax
2413	mov	%r13,16+8(%rsp)		# end of b[num]
2414	mov	$num,24+8(%rsp)		# inner counter
2415	mov	$rp, 56+8(%rsp)		# save $rp
2416___
2417my ($aptr, $bptr, $nptr, $tptr, $mi,  $bi,  $zero, $num)=
2418   ("%rsi","%rdi","%rcx","%rbx","%r8","%r9","%rbp","%rax");
2419my $rptr=$bptr;
2420my $STRIDE=2**5*8;		# 5 is "window size"
2421my $N=$STRIDE/4;		# should match cache line size
2422$code.=<<___;
2423	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
2424	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
2425	lea	88-112(%rsp,%r10),%r10	# place the mask after tp[num+1] (+ICache optimization)
2426	lea	128($bp),$bptr		# size optimization
2427
2428	pshufd	\$0,%xmm5,%xmm5		# broadcast index
2429	movdqa	%xmm1,%xmm4
2430	.byte	0x67
2431	movdqa	%xmm1,%xmm2
2432___
2433########################################################################
2434# calculate mask by comparing 0..31 to index and save result to stack
2435#
2436$code.=<<___;
2437	.byte	0x67
2438	paddd	%xmm0,%xmm1
2439	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
2440	movdqa	%xmm4,%xmm3
2441___
2442for($i=0;$i<$STRIDE/16-4;$i+=4) {
2443$code.=<<___;
2444	paddd	%xmm1,%xmm2
2445	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
2446	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2447	movdqa	%xmm4,%xmm0
2448
2449	paddd	%xmm2,%xmm3
2450	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
2451	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2452	movdqa	%xmm4,%xmm1
2453
2454	paddd	%xmm3,%xmm0
2455	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
2456	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2457	movdqa	%xmm4,%xmm2
2458
2459	paddd	%xmm0,%xmm1
2460	pcmpeqd	%xmm5,%xmm0
2461	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2462	movdqa	%xmm4,%xmm3
2463___
2464}
2465$code.=<<___;				# last iteration can be optimized
2466	.byte	0x67
2467	paddd	%xmm1,%xmm2
2468	pcmpeqd	%xmm5,%xmm1
2469	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2470
2471	paddd	%xmm2,%xmm3
2472	pcmpeqd	%xmm5,%xmm2
2473	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2474
2475	pcmpeqd	%xmm5,%xmm3
2476	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2477
2478	pand	`16*($i+0)-128`($bptr),%xmm0	# while it's still in register
2479	pand	`16*($i+1)-128`($bptr),%xmm1
2480	pand	`16*($i+2)-128`($bptr),%xmm2
2481	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2482	pand	`16*($i+3)-128`($bptr),%xmm3
2483	por	%xmm2,%xmm0
2484	por	%xmm3,%xmm1
2485___
2486for($i=0;$i<$STRIDE/16-4;$i+=4) {
2487$code.=<<___;
2488	movdqa	`16*($i+0)-128`($bptr),%xmm4
2489	movdqa	`16*($i+1)-128`($bptr),%xmm5
2490	movdqa	`16*($i+2)-128`($bptr),%xmm2
2491	pand	`16*($i+0)+112`(%r10),%xmm4
2492	movdqa	`16*($i+3)-128`($bptr),%xmm3
2493	pand	`16*($i+1)+112`(%r10),%xmm5
2494	por	%xmm4,%xmm0
2495	pand	`16*($i+2)+112`(%r10),%xmm2
2496	por	%xmm5,%xmm1
2497	pand	`16*($i+3)+112`(%r10),%xmm3
2498	por	%xmm2,%xmm0
2499	por	%xmm3,%xmm1
2500___
2501}
2502$code.=<<___;
2503	pxor	%xmm1,%xmm0
2504	pshufd	\$0x4e,%xmm0,%xmm1
2505	por	%xmm1,%xmm0
2506	lea	$STRIDE($bptr),$bptr
2507	movq	%xmm0,%rdx		# bp[0]
2508	lea	64+8*4+8(%rsp),$tptr
2509
2510	mov	%rdx,$bi
2511	mulx	0*8($aptr),$mi,%rax	# a[0]*b[0]
2512	mulx	1*8($aptr),%r11,%r12	# a[1]*b[0]
2513	add	%rax,%r11
2514	mulx	2*8($aptr),%rax,%r13	# ...
2515	adc	%rax,%r12
2516	adc	\$0,%r13
2517	mulx	3*8($aptr),%rax,%r14
2518
2519	mov	$mi,%r15
2520	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2521	xor	$zero,$zero		# cf=0, of=0
2522	mov	$mi,%rdx
2523
2524	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2525
2526	lea	4*8($aptr),$aptr
2527	adcx	%rax,%r13
2528	adcx	$zero,%r14		# cf=0
2529
2530	mulx	0*8($nptr),%rax,%r10
2531	adcx	%rax,%r15		# discarded
2532	adox	%r11,%r10
2533	mulx	1*8($nptr),%rax,%r11
2534	adcx	%rax,%r10
2535	adox	%r12,%r11
2536	mulx	2*8($nptr),%rax,%r12
2537	mov	24+8(%rsp),$bptr	# counter value
2538	mov	%r10,-8*4($tptr)
2539	adcx	%rax,%r11
2540	adox	%r13,%r12
2541	mulx	3*8($nptr),%rax,%r15
2542	 mov	$bi,%rdx
2543	mov	%r11,-8*3($tptr)
2544	adcx	%rax,%r12
2545	adox	$zero,%r15		# of=0
2546	lea	4*8($nptr),$nptr
2547	mov	%r12,-8*2($tptr)
2548	jmp	.Lmulx4x_1st
2549
2550.align	32
2551.Lmulx4x_1st:
2552	adcx	$zero,%r15		# cf=0, modulo-scheduled
2553	mulx	0*8($aptr),%r10,%rax	# a[4]*b[0]
2554	adcx	%r14,%r10
2555	mulx	1*8($aptr),%r11,%r14	# a[5]*b[0]
2556	adcx	%rax,%r11
2557	mulx	2*8($aptr),%r12,%rax	# ...
2558	adcx	%r14,%r12
2559	mulx	3*8($aptr),%r13,%r14
2560	 .byte	0x67,0x67
2561	 mov	$mi,%rdx
2562	adcx	%rax,%r13
2563	adcx	$zero,%r14		# cf=0
2564	lea	4*8($aptr),$aptr
2565	lea	4*8($tptr),$tptr
2566
2567	adox	%r15,%r10
2568	mulx	0*8($nptr),%rax,%r15
2569	adcx	%rax,%r10
2570	adox	%r15,%r11
2571	mulx	1*8($nptr),%rax,%r15
2572	adcx	%rax,%r11
2573	adox	%r15,%r12
2574	mulx	2*8($nptr),%rax,%r15
2575	mov	%r10,-5*8($tptr)
2576	adcx	%rax,%r12
2577	mov	%r11,-4*8($tptr)
2578	adox	%r15,%r13
2579	mulx	3*8($nptr),%rax,%r15
2580	 mov	$bi,%rdx
2581	mov	%r12,-3*8($tptr)
2582	adcx	%rax,%r13
2583	adox	$zero,%r15
2584	lea	4*8($nptr),$nptr
2585	mov	%r13,-2*8($tptr)
2586
2587	dec	$bptr			# of=0, pass cf
2588	jnz	.Lmulx4x_1st
2589
2590	mov	8(%rsp),$num		# load -num
2591	adc	$zero,%r15		# modulo-scheduled
2592	lea	($aptr,$num),$aptr	# rewind $aptr
2593	add	%r15,%r14
2594	mov	8+8(%rsp),$bptr		# re-load &b[i]
2595	adc	$zero,$zero		# top-most carry
2596	mov	%r14,-1*8($tptr)
2597	jmp	.Lmulx4x_outer
2598
2599.align	32
2600.Lmulx4x_outer:
2601	lea	16-256($tptr),%r10	# where 256-byte mask is (+density control)
2602	pxor	%xmm4,%xmm4
2603	.byte	0x67,0x67
2604	pxor	%xmm5,%xmm5
2605___
2606for($i=0;$i<$STRIDE/16;$i+=4) {
2607$code.=<<___;
2608	movdqa	`16*($i+0)-128`($bptr),%xmm0
2609	movdqa	`16*($i+1)-128`($bptr),%xmm1
2610	movdqa	`16*($i+2)-128`($bptr),%xmm2
2611	pand	`16*($i+0)+256`(%r10),%xmm0
2612	movdqa	`16*($i+3)-128`($bptr),%xmm3
2613	pand	`16*($i+1)+256`(%r10),%xmm1
2614	por	%xmm0,%xmm4
2615	pand	`16*($i+2)+256`(%r10),%xmm2
2616	por	%xmm1,%xmm5
2617	pand	`16*($i+3)+256`(%r10),%xmm3
2618	por	%xmm2,%xmm4
2619	por	%xmm3,%xmm5
2620___
2621}
2622$code.=<<___;
2623	por	%xmm5,%xmm4
2624	pshufd	\$0x4e,%xmm4,%xmm0
2625	por	%xmm4,%xmm0
2626	lea	$STRIDE($bptr),$bptr
2627	movq	%xmm0,%rdx		# m0=bp[i]
2628
2629	mov	$zero,($tptr)		# save top-most carry
2630	lea	4*8($tptr,$num),$tptr	# rewind $tptr
2631	mulx	0*8($aptr),$mi,%r11	# a[0]*b[i]
2632	xor	$zero,$zero		# cf=0, of=0
2633	mov	%rdx,$bi
2634	mulx	1*8($aptr),%r14,%r12	# a[1]*b[i]
2635	adox	-4*8($tptr),$mi		# +t[0]
2636	adcx	%r14,%r11
2637	mulx	2*8($aptr),%r15,%r13	# ...
2638	adox	-3*8($tptr),%r11
2639	adcx	%r15,%r12
2640	mulx	3*8($aptr),%rdx,%r14
2641	adox	-2*8($tptr),%r12
2642	adcx	%rdx,%r13
2643	lea	($nptr,$num),$nptr	# rewind $nptr
2644	lea	4*8($aptr),$aptr
2645	adox	-1*8($tptr),%r13
2646	adcx	$zero,%r14
2647	adox	$zero,%r14
2648
2649	mov	$mi,%r15
2650	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2651
2652	mov	$mi,%rdx
2653	xor	$zero,$zero		# cf=0, of=0
2654	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2655
2656	mulx	0*8($nptr),%rax,%r10
2657	adcx	%rax,%r15		# discarded
2658	adox	%r11,%r10
2659	mulx	1*8($nptr),%rax,%r11
2660	adcx	%rax,%r10
2661	adox	%r12,%r11
2662	mulx	2*8($nptr),%rax,%r12
2663	adcx	%rax,%r11
2664	adox	%r13,%r12
2665	mulx	3*8($nptr),%rax,%r15
2666	 mov	$bi,%rdx
2667	mov	24+8(%rsp),$bptr	# counter value
2668	mov	%r10,-8*4($tptr)
2669	adcx	%rax,%r12
2670	mov	%r11,-8*3($tptr)
2671	adox	$zero,%r15		# of=0
2672	mov	%r12,-8*2($tptr)
2673	lea	4*8($nptr),$nptr
2674	jmp	.Lmulx4x_inner
2675
2676.align	32
2677.Lmulx4x_inner:
2678	mulx	0*8($aptr),%r10,%rax	# a[4]*b[i]
2679	adcx	$zero,%r15		# cf=0, modulo-scheduled
2680	adox	%r14,%r10
2681	mulx	1*8($aptr),%r11,%r14	# a[5]*b[i]
2682	adcx	0*8($tptr),%r10
2683	adox	%rax,%r11
2684	mulx	2*8($aptr),%r12,%rax	# ...
2685	adcx	1*8($tptr),%r11
2686	adox	%r14,%r12
2687	mulx	3*8($aptr),%r13,%r14
2688	 mov	$mi,%rdx
2689	adcx	2*8($tptr),%r12
2690	adox	%rax,%r13
2691	adcx	3*8($tptr),%r13
2692	adox	$zero,%r14		# of=0
2693	lea	4*8($aptr),$aptr
2694	lea	4*8($tptr),$tptr
2695	adcx	$zero,%r14		# cf=0
2696
2697	adox	%r15,%r10
2698	mulx	0*8($nptr),%rax,%r15
2699	adcx	%rax,%r10
2700	adox	%r15,%r11
2701	mulx	1*8($nptr),%rax,%r15
2702	adcx	%rax,%r11
2703	adox	%r15,%r12
2704	mulx	2*8($nptr),%rax,%r15
2705	mov	%r10,-5*8($tptr)
2706	adcx	%rax,%r12
2707	adox	%r15,%r13
2708	mov	%r11,-4*8($tptr)
2709	mulx	3*8($nptr),%rax,%r15
2710	 mov	$bi,%rdx
2711	lea	4*8($nptr),$nptr
2712	mov	%r12,-3*8($tptr)
2713	adcx	%rax,%r13
2714	adox	$zero,%r15
2715	mov	%r13,-2*8($tptr)
2716
2717	dec	$bptr			# of=0, pass cf
2718	jnz	.Lmulx4x_inner
2719
2720	mov	0+8(%rsp),$num		# load -num
2721	adc	$zero,%r15		# modulo-scheduled
2722	sub	0*8($tptr),$bptr	# pull top-most carry to %cf
2723	mov	8+8(%rsp),$bptr		# re-load &b[i]
2724	mov	16+8(%rsp),%r10
2725	adc	%r15,%r14
2726	lea	($aptr,$num),$aptr	# rewind $aptr
2727	adc	$zero,$zero		# top-most carry
2728	mov	%r14,-1*8($tptr)
2729
2730	cmp	%r10,$bptr
2731	jb	.Lmulx4x_outer
2732
2733	mov	-8($nptr),%r10
2734	mov	$zero,%r8
2735	mov	($nptr,$num),%r12
2736	lea	($nptr,$num),%rbp	# rewind $nptr
2737	mov	$num,%rcx
2738	lea	($tptr,$num),%rdi	# rewind $tptr
2739	xor	%eax,%eax
2740	xor	%r15,%r15
2741	sub	%r14,%r10		# compare top-most words
2742	adc	%r15,%r15
2743	or	%r15,%r8
2744	sar	\$3+2,%rcx
2745	sub	%r8,%rax		# %rax=-%r8
2746	mov	56+8(%rsp),%rdx		# restore rp
2747	dec	%r12			# so that after 'not' we get -n[0]
2748	mov	8*1(%rbp),%r13
2749	xor	%r8,%r8
2750	mov	8*2(%rbp),%r14
2751	mov	8*3(%rbp),%r15
2752	jmp	.Lsqrx4x_sub_entry	# common post-condition
2753.size	mulx4x_internal,.-mulx4x_internal
2754___
2755}{
2756######################################################################
2757# void bn_power5(
2758my $rptr="%rdi";	# BN_ULONG *rptr,
2759my $aptr="%rsi";	# const BN_ULONG *aptr,
2760my $bptr="%rdx";	# const void *table,
2761my $nptr="%rcx";	# const BN_ULONG *nptr,
2762my $n0  ="%r8";		# const BN_ULONG *n0);
2763my $num ="%r9";		# int num, has to be divisible by 8
2764			# int pwr);
2765
2766my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
2767my @A0=("%r10","%r11");
2768my @A1=("%r12","%r13");
2769my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
2770
2771$code.=<<___;
2772.type	bn_powerx5,\@function,6
2773.align	32
2774bn_powerx5:
2775.cfi_startproc
2776	mov	%rsp,%rax
2777.cfi_def_cfa_register	%rax
2778.Lpowerx5_enter:
2779	push	%rbx
2780.cfi_push	%rbx
2781	push	%rbp
2782.cfi_push	%rbp
2783	push	%r12
2784.cfi_push	%r12
2785	push	%r13
2786.cfi_push	%r13
2787	push	%r14
2788.cfi_push	%r14
2789	push	%r15
2790.cfi_push	%r15
2791.Lpowerx5_prologue:
2792
2793	shl	\$3,${num}d		# convert $num to bytes
2794	lea	($num,$num,2),%r10	# 3*$num in bytes
2795	neg	$num
2796	mov	($n0),$n0		# *n0
2797
2798	##############################################################
2799	# Ensure that stack frame doesn't alias with $rptr+3*$num
2800	# modulo 4096, which covers ret[num], am[num] and n[num]
2801	# (see bn_exp.c). This is done to allow memory disambiguation
2802	# logic do its magic. [Extra 256 bytes is for power mask
2803	# calculated from 7th argument, the index.]
2804	#
2805	lea	-320(%rsp,$num,2),%r11
2806	mov	%rsp,%rbp
2807	sub	$rptr,%r11
2808	and	\$4095,%r11
2809	cmp	%r11,%r10
2810	jb	.Lpwrx_sp_alt
2811	sub	%r11,%rbp		# align with $aptr
2812	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2813	jmp	.Lpwrx_sp_done
2814
2815.align	32
2816.Lpwrx_sp_alt:
2817	lea	4096-320(,$num,2),%r10
2818	lea	-320(%rbp,$num,2),%rbp	# alloca(frame+2*$num*8+256)
2819	sub	%r10,%r11
2820	mov	\$0,%r10
2821	cmovc	%r10,%r11
2822	sub	%r11,%rbp
2823.Lpwrx_sp_done:
2824	and	\$-64,%rbp
2825	mov	%rsp,%r11
2826	sub	%rbp,%r11
2827	and	\$-4096,%r11
2828	lea	(%rbp,%r11),%rsp
2829	mov	(%rsp),%r10
2830	cmp	%rbp,%rsp
2831	ja	.Lpwrx_page_walk
2832	jmp	.Lpwrx_page_walk_done
2833
2834.Lpwrx_page_walk:
2835	lea	-4096(%rsp),%rsp
2836	mov	(%rsp),%r10
2837	cmp	%rbp,%rsp
2838	ja	.Lpwrx_page_walk
2839.Lpwrx_page_walk_done:
2840
2841	mov	$num,%r10
2842	neg	$num
2843
2844	##############################################################
2845	# Stack layout
2846	#
2847	# +0	saved $num, used in reduction section
2848	# +8	&t[2*$num], used in reduction section
2849	# +16	intermediate carry bit
2850	# +24	top-most carry bit, used in reduction section
2851	# +32	saved *n0
2852	# +40	saved %rsp
2853	# +48	t[2*$num]
2854	#
2855	pxor	%xmm0,%xmm0
2856	movq	$rptr,%xmm1		# save $rptr
2857	movq	$nptr,%xmm2		# save $nptr
2858	movq	%r10, %xmm3		# -$num
2859	movq	$bptr,%xmm4
2860	mov	$n0,  32(%rsp)
2861	mov	%rax, 40(%rsp)		# save original %rsp
2862.cfi_cfa_expression	%rsp+40,deref,+8
2863.Lpowerx5_body:
2864
2865	call	__bn_sqrx8x_internal
2866	call	__bn_postx4x_internal
2867	call	__bn_sqrx8x_internal
2868	call	__bn_postx4x_internal
2869	call	__bn_sqrx8x_internal
2870	call	__bn_postx4x_internal
2871	call	__bn_sqrx8x_internal
2872	call	__bn_postx4x_internal
2873	call	__bn_sqrx8x_internal
2874	call	__bn_postx4x_internal
2875
2876	mov	%r10,$num		# -num
2877	mov	$aptr,$rptr
2878	movq	%xmm2,$nptr
2879	movq	%xmm4,$bptr
2880	mov	40(%rsp),%rax
2881
2882	call	mulx4x_internal
2883
2884	mov	40(%rsp),%rsi		# restore %rsp
2885.cfi_def_cfa	%rsi,8
2886	mov	\$1,%rax
2887
2888	mov	-48(%rsi),%r15
2889.cfi_restore	%r15
2890	mov	-40(%rsi),%r14
2891.cfi_restore	%r14
2892	mov	-32(%rsi),%r13
2893.cfi_restore	%r13
2894	mov	-24(%rsi),%r12
2895.cfi_restore	%r12
2896	mov	-16(%rsi),%rbp
2897.cfi_restore	%rbp
2898	mov	-8(%rsi),%rbx
2899.cfi_restore	%rbx
2900	lea	(%rsi),%rsp
2901.cfi_def_cfa_register	%rsp
2902.Lpowerx5_epilogue:
2903	ret
2904.cfi_endproc
2905.size	bn_powerx5,.-bn_powerx5
2906
2907.globl	bn_sqrx8x_internal
2908.hidden	bn_sqrx8x_internal
2909.type	bn_sqrx8x_internal,\@abi-omnipotent
2910.align	32
2911bn_sqrx8x_internal:
2912__bn_sqrx8x_internal:
2913	##################################################################
2914	# Squaring part:
2915	#
2916	# a) multiply-n-add everything but a[i]*a[i];
2917	# b) shift result of a) by 1 to the left and accumulate
2918	#    a[i]*a[i] products;
2919	#
2920	##################################################################
2921	# 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]
2922	#                                                     a[1]a[0]
2923	#                                                 a[2]a[0]
2924	#                                             a[3]a[0]
2925	#                                             a[2]a[1]
2926	#                                         a[3]a[1]
2927	#                                     a[3]a[2]
2928	#
2929	#                                         a[4]a[0]
2930	#                                     a[5]a[0]
2931	#                                 a[6]a[0]
2932	#                             a[7]a[0]
2933	#                                     a[4]a[1]
2934	#                                 a[5]a[1]
2935	#                             a[6]a[1]
2936	#                         a[7]a[1]
2937	#                                 a[4]a[2]
2938	#                             a[5]a[2]
2939	#                         a[6]a[2]
2940	#                     a[7]a[2]
2941	#                             a[4]a[3]
2942	#                         a[5]a[3]
2943	#                     a[6]a[3]
2944	#                 a[7]a[3]
2945	#
2946	#                     a[5]a[4]
2947	#                 a[6]a[4]
2948	#             a[7]a[4]
2949	#             a[6]a[5]
2950	#         a[7]a[5]
2951	#     a[7]a[6]
2952	# 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]
2953___
2954{
2955my ($zero,$carry)=("%rbp","%rcx");
2956my $aaptr=$zero;
2957$code.=<<___;
2958	lea	48+8(%rsp),$tptr
2959	lea	($aptr,$num),$aaptr
2960	mov	$num,0+8(%rsp)			# save $num
2961	mov	$aaptr,8+8(%rsp)		# save end of $aptr
2962	jmp	.Lsqr8x_zero_start
2963
2964.align	32
2965.byte	0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
2966.Lsqrx8x_zero:
2967	.byte	0x3e
2968	movdqa	%xmm0,0*8($tptr)
2969	movdqa	%xmm0,2*8($tptr)
2970	movdqa	%xmm0,4*8($tptr)
2971	movdqa	%xmm0,6*8($tptr)
2972.Lsqr8x_zero_start:			# aligned at 32
2973	movdqa	%xmm0,8*8($tptr)
2974	movdqa	%xmm0,10*8($tptr)
2975	movdqa	%xmm0,12*8($tptr)
2976	movdqa	%xmm0,14*8($tptr)
2977	lea	16*8($tptr),$tptr
2978	sub	\$64,$num
2979	jnz	.Lsqrx8x_zero
2980
2981	mov	0*8($aptr),%rdx		# a[0], modulo-scheduled
2982	#xor	%r9,%r9			# t[1], ex-$num, zero already
2983	xor	%r10,%r10
2984	xor	%r11,%r11
2985	xor	%r12,%r12
2986	xor	%r13,%r13
2987	xor	%r14,%r14
2988	xor	%r15,%r15
2989	lea	48+8(%rsp),$tptr
2990	xor	$zero,$zero		# cf=0, cf=0
2991	jmp	.Lsqrx8x_outer_loop
2992
2993.align	32
2994.Lsqrx8x_outer_loop:
2995	mulx	1*8($aptr),%r8,%rax	# a[1]*a[0]
2996	adcx	%r9,%r8			# a[1]*a[0]+=t[1]
2997	adox	%rax,%r10
2998	mulx	2*8($aptr),%r9,%rax	# a[2]*a[0]
2999	adcx	%r10,%r9
3000	adox	%rax,%r11
3001	.byte	0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00	# mulx	3*8($aptr),%r10,%rax	# ...
3002	adcx	%r11,%r10
3003	adox	%rax,%r12
3004	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00	# mulx	4*8($aptr),%r11,%rax
3005	adcx	%r12,%r11
3006	adox	%rax,%r13
3007	mulx	5*8($aptr),%r12,%rax
3008	adcx	%r13,%r12
3009	adox	%rax,%r14
3010	mulx	6*8($aptr),%r13,%rax
3011	adcx	%r14,%r13
3012	adox	%r15,%rax
3013	mulx	7*8($aptr),%r14,%r15
3014	 mov	1*8($aptr),%rdx		# a[1]
3015	adcx	%rax,%r14
3016	adox	$zero,%r15
3017	adc	8*8($tptr),%r15
3018	mov	%r8,1*8($tptr)		# t[1]
3019	mov	%r9,2*8($tptr)		# t[2]
3020	sbb	$carry,$carry		# mov %cf,$carry
3021	xor	$zero,$zero		# cf=0, of=0
3022
3023
3024	mulx	2*8($aptr),%r8,%rbx	# a[2]*a[1]
3025	mulx	3*8($aptr),%r9,%rax	# a[3]*a[1]
3026	adcx	%r10,%r8
3027	adox	%rbx,%r9
3028	mulx	4*8($aptr),%r10,%rbx	# ...
3029	adcx	%r11,%r9
3030	adox	%rax,%r10
3031	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00	# mulx	5*8($aptr),%r11,%rax
3032	adcx	%r12,%r10
3033	adox	%rbx,%r11
3034	.byte	0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r12,%rbx
3035	adcx	%r13,%r11
3036	adox	%r14,%r12
3037	.byte	0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r13,%r14
3038	 mov	2*8($aptr),%rdx		# a[2]
3039	adcx	%rax,%r12
3040	adox	%rbx,%r13
3041	adcx	%r15,%r13
3042	adox	$zero,%r14		# of=0
3043	adcx	$zero,%r14		# cf=0
3044
3045	mov	%r8,3*8($tptr)		# t[3]
3046	mov	%r9,4*8($tptr)		# t[4]
3047
3048	mulx	3*8($aptr),%r8,%rbx	# a[3]*a[2]
3049	mulx	4*8($aptr),%r9,%rax	# a[4]*a[2]
3050	adcx	%r10,%r8
3051	adox	%rbx,%r9
3052	mulx	5*8($aptr),%r10,%rbx	# ...
3053	adcx	%r11,%r9
3054	adox	%rax,%r10
3055	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r11,%rax
3056	adcx	%r12,%r10
3057	adox	%r13,%r11
3058	.byte	0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r12,%r13
3059	.byte	0x3e
3060	 mov	3*8($aptr),%rdx		# a[3]
3061	adcx	%rbx,%r11
3062	adox	%rax,%r12
3063	adcx	%r14,%r12
3064	mov	%r8,5*8($tptr)		# t[5]
3065	mov	%r9,6*8($tptr)		# t[6]
3066	 mulx	4*8($aptr),%r8,%rax	# a[4]*a[3]
3067	adox	$zero,%r13		# of=0
3068	adcx	$zero,%r13		# cf=0
3069
3070	mulx	5*8($aptr),%r9,%rbx	# a[5]*a[3]
3071	adcx	%r10,%r8
3072	adox	%rax,%r9
3073	mulx	6*8($aptr),%r10,%rax	# ...
3074	adcx	%r11,%r9
3075	adox	%r12,%r10
3076	mulx	7*8($aptr),%r11,%r12
3077	 mov	4*8($aptr),%rdx		# a[4]
3078	 mov	5*8($aptr),%r14		# a[5]
3079	adcx	%rbx,%r10
3080	adox	%rax,%r11
3081	 mov	6*8($aptr),%r15		# a[6]
3082	adcx	%r13,%r11
3083	adox	$zero,%r12		# of=0
3084	adcx	$zero,%r12		# cf=0
3085
3086	mov	%r8,7*8($tptr)		# t[7]
3087	mov	%r9,8*8($tptr)		# t[8]
3088
3089	mulx	%r14,%r9,%rax		# a[5]*a[4]
3090	 mov	7*8($aptr),%r8		# a[7]
3091	adcx	%r10,%r9
3092	mulx	%r15,%r10,%rbx		# a[6]*a[4]
3093	adox	%rax,%r10
3094	adcx	%r11,%r10
3095	mulx	%r8,%r11,%rax		# a[7]*a[4]
3096	 mov	%r14,%rdx		# a[5]
3097	adox	%rbx,%r11
3098	adcx	%r12,%r11
3099	#adox	$zero,%rax		# of=0
3100	adcx	$zero,%rax		# cf=0
3101
3102	mulx	%r15,%r14,%rbx		# a[6]*a[5]
3103	mulx	%r8,%r12,%r13		# a[7]*a[5]
3104	 mov	%r15,%rdx		# a[6]
3105	 lea	8*8($aptr),$aptr
3106	adcx	%r14,%r11
3107	adox	%rbx,%r12
3108	adcx	%rax,%r12
3109	adox	$zero,%r13
3110
3111	.byte	0x67,0x67
3112	mulx	%r8,%r8,%r14		# a[7]*a[6]
3113	adcx	%r8,%r13
3114	adcx	$zero,%r14
3115
3116	cmp	8+8(%rsp),$aptr
3117	je	.Lsqrx8x_outer_break
3118
3119	neg	$carry			# mov $carry,%cf
3120	mov	\$-8,%rcx
3121	mov	$zero,%r15
3122	mov	8*8($tptr),%r8
3123	adcx	9*8($tptr),%r9		# +=t[9]
3124	adcx	10*8($tptr),%r10	# ...
3125	adcx	11*8($tptr),%r11
3126	adc	12*8($tptr),%r12
3127	adc	13*8($tptr),%r13
3128	adc	14*8($tptr),%r14
3129	adc	15*8($tptr),%r15
3130	lea	($aptr),$aaptr
3131	lea	2*64($tptr),$tptr
3132	sbb	%rax,%rax		# mov %cf,$carry
3133
3134	mov	-64($aptr),%rdx		# a[0]
3135	mov	%rax,16+8(%rsp)		# offload $carry
3136	mov	$tptr,24+8(%rsp)
3137
3138	#lea	8*8($tptr),$tptr	# see 2*8*8($tptr) above
3139	xor	%eax,%eax		# cf=0, of=0
3140	jmp	.Lsqrx8x_loop
3141
3142.align	32
3143.Lsqrx8x_loop:
3144	mov	%r8,%rbx
3145	mulx	0*8($aaptr),%rax,%r8	# a[8]*a[i]
3146	adcx	%rax,%rbx		# +=t[8]
3147	adox	%r9,%r8
3148
3149	mulx	1*8($aaptr),%rax,%r9	# ...
3150	adcx	%rax,%r8
3151	adox	%r10,%r9
3152
3153	mulx	2*8($aaptr),%rax,%r10
3154	adcx	%rax,%r9
3155	adox	%r11,%r10
3156
3157	mulx	3*8($aaptr),%rax,%r11
3158	adcx	%rax,%r10
3159	adox	%r12,%r11
3160
3161	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	4*8($aaptr),%rax,%r12
3162	adcx	%rax,%r11
3163	adox	%r13,%r12
3164
3165	mulx	5*8($aaptr),%rax,%r13
3166	adcx	%rax,%r12
3167	adox	%r14,%r13
3168
3169	mulx	6*8($aaptr),%rax,%r14
3170	 mov	%rbx,($tptr,%rcx,8)	# store t[8+i]
3171	 mov	\$0,%ebx
3172	adcx	%rax,%r13
3173	adox	%r15,%r14
3174
3175	.byte	0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00	# mulx	7*8($aaptr),%rax,%r15
3176	 mov	8($aptr,%rcx,8),%rdx	# a[i]
3177	adcx	%rax,%r14
3178	adox	%rbx,%r15		# %rbx is 0, of=0
3179	adcx	%rbx,%r15		# cf=0
3180
3181	.byte	0x67
3182	inc	%rcx			# of=0
3183	jnz	.Lsqrx8x_loop
3184
3185	lea	8*8($aaptr),$aaptr
3186	mov	\$-8,%rcx
3187	cmp	8+8(%rsp),$aaptr	# done?
3188	je	.Lsqrx8x_break
3189
3190	sub	16+8(%rsp),%rbx		# mov 16(%rsp),%cf
3191	.byte	0x66
3192	mov	-64($aptr),%rdx
3193	adcx	0*8($tptr),%r8
3194	adcx	1*8($tptr),%r9
3195	adc	2*8($tptr),%r10
3196	adc	3*8($tptr),%r11
3197	adc	4*8($tptr),%r12
3198	adc	5*8($tptr),%r13
3199	adc	6*8($tptr),%r14
3200	adc	7*8($tptr),%r15
3201	lea	8*8($tptr),$tptr
3202	.byte	0x67
3203	sbb	%rax,%rax		# mov %cf,%rax
3204	xor	%ebx,%ebx		# cf=0, of=0
3205	mov	%rax,16+8(%rsp)		# offload carry
3206	jmp	.Lsqrx8x_loop
3207
3208.align	32
3209.Lsqrx8x_break:
3210	xor	$zero,$zero
3211	sub	16+8(%rsp),%rbx		# mov 16(%rsp),%cf
3212	adcx	$zero,%r8
3213	mov	24+8(%rsp),$carry	# initial $tptr, borrow $carry
3214	adcx	$zero,%r9
3215	mov	0*8($aptr),%rdx		# a[8], modulo-scheduled
3216	adc	\$0,%r10
3217	mov	%r8,0*8($tptr)
3218	adc	\$0,%r11
3219	adc	\$0,%r12
3220	adc	\$0,%r13
3221	adc	\$0,%r14
3222	adc	\$0,%r15
3223	cmp	$carry,$tptr		# cf=0, of=0
3224	je	.Lsqrx8x_outer_loop
3225
3226	mov	%r9,1*8($tptr)
3227	 mov	1*8($carry),%r9
3228	mov	%r10,2*8($tptr)
3229	 mov	2*8($carry),%r10
3230	mov	%r11,3*8($tptr)
3231	 mov	3*8($carry),%r11
3232	mov	%r12,4*8($tptr)
3233	 mov	4*8($carry),%r12
3234	mov	%r13,5*8($tptr)
3235	 mov	5*8($carry),%r13
3236	mov	%r14,6*8($tptr)
3237	 mov	6*8($carry),%r14
3238	mov	%r15,7*8($tptr)
3239	 mov	7*8($carry),%r15
3240	mov	$carry,$tptr
3241	jmp	.Lsqrx8x_outer_loop
3242
3243.align	32
3244.Lsqrx8x_outer_break:
3245	mov	%r9,9*8($tptr)		# t[9]
3246	 movq	%xmm3,%rcx		# -$num
3247	mov	%r10,10*8($tptr)	# ...
3248	mov	%r11,11*8($tptr)
3249	mov	%r12,12*8($tptr)
3250	mov	%r13,13*8($tptr)
3251	mov	%r14,14*8($tptr)
3252___
3253}{
3254my $i="%rcx";
3255$code.=<<___;
3256	lea	48+8(%rsp),$tptr
3257	mov	($aptr,$i),%rdx		# a[0]
3258
3259	mov	8($tptr),$A0[1]		# t[1]
3260	xor	$A0[0],$A0[0]		# t[0], of=0, cf=0
3261	mov	0+8(%rsp),$num		# restore $num
3262	adox	$A0[1],$A0[1]
3263	 mov	16($tptr),$A1[0]	# t[2]	# prefetch
3264	 mov	24($tptr),$A1[1]	# t[3]	# prefetch
3265	#jmp	.Lsqrx4x_shift_n_add	# happens to be aligned
3266
3267.align	32
3268.Lsqrx4x_shift_n_add:
3269	mulx	%rdx,%rax,%rbx
3270	 adox	$A1[0],$A1[0]
3271	adcx	$A0[0],%rax
3272	 .byte	0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00	# mov	8($aptr,$i),%rdx	# a[i+1]	# prefetch
3273	 .byte	0x4c,0x8b,0x97,0x20,0x00,0x00,0x00	# mov	32($tptr),$A0[0]	# t[2*i+4]	# prefetch
3274	 adox	$A1[1],$A1[1]
3275	adcx	$A0[1],%rbx
3276	 mov	40($tptr),$A0[1]		# t[2*i+4+1]	# prefetch
3277	mov	%rax,0($tptr)
3278	mov	%rbx,8($tptr)
3279
3280	mulx	%rdx,%rax,%rbx
3281	 adox	$A0[0],$A0[0]
3282	adcx	$A1[0],%rax
3283	 mov	16($aptr,$i),%rdx	# a[i+2]	# prefetch
3284	 mov	48($tptr),$A1[0]	# t[2*i+6]	# prefetch
3285	 adox	$A0[1],$A0[1]
3286	adcx	$A1[1],%rbx
3287	 mov	56($tptr),$A1[1]	# t[2*i+6+1]	# prefetch
3288	mov	%rax,16($tptr)
3289	mov	%rbx,24($tptr)
3290
3291	mulx	%rdx,%rax,%rbx
3292	 adox	$A1[0],$A1[0]
3293	adcx	$A0[0],%rax
3294	 mov	24($aptr,$i),%rdx	# a[i+3]	# prefetch
3295	 lea	32($i),$i
3296	 mov	64($tptr),$A0[0]	# t[2*i+8]	# prefetch
3297	 adox	$A1[1],$A1[1]
3298	adcx	$A0[1],%rbx
3299	 mov	72($tptr),$A0[1]	# t[2*i+8+1]	# prefetch
3300	mov	%rax,32($tptr)
3301	mov	%rbx,40($tptr)
3302
3303	mulx	%rdx,%rax,%rbx
3304	 adox	$A0[0],$A0[0]
3305	adcx	$A1[0],%rax
3306	jrcxz	.Lsqrx4x_shift_n_add_break
3307	 .byte	0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00	# mov	0($aptr,$i),%rdx	# a[i+4]	# prefetch
3308	 adox	$A0[1],$A0[1]
3309	adcx	$A1[1],%rbx
3310	 mov	80($tptr),$A1[0]	# t[2*i+10]	# prefetch
3311	 mov	88($tptr),$A1[1]	# t[2*i+10+1]	# prefetch
3312	mov	%rax,48($tptr)
3313	mov	%rbx,56($tptr)
3314	lea	64($tptr),$tptr
3315	nop
3316	jmp	.Lsqrx4x_shift_n_add
3317
3318.align	32
3319.Lsqrx4x_shift_n_add_break:
3320	adcx	$A1[1],%rbx
3321	mov	%rax,48($tptr)
3322	mov	%rbx,56($tptr)
3323	lea	64($tptr),$tptr		# end of t[] buffer
3324___
3325}
3326######################################################################
3327# Montgomery reduction part, "word-by-word" algorithm.
3328#
3329# This new path is inspired by multiple submissions from Intel, by
3330# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
3331# Vinodh Gopal...
3332{
3333my ($nptr,$carry,$m0)=("%rbp","%rsi","%rdx");
3334
3335$code.=<<___;
3336	movq	%xmm2,$nptr
3337__bn_sqrx8x_reduction:
3338	xor	%eax,%eax		# initial top-most carry bit
3339	mov	32+8(%rsp),%rbx		# n0
3340	mov	48+8(%rsp),%rdx		# "%r8", 8*0($tptr)
3341	lea	-8*8($nptr,$num),%rcx	# end of n[]
3342	#lea	48+8(%rsp,$num,2),$tptr	# end of t[] buffer
3343	mov	%rcx, 0+8(%rsp)		# save end of n[]
3344	mov	$tptr,8+8(%rsp)		# save end of t[]
3345
3346	lea	48+8(%rsp),$tptr		# initial t[] window
3347	jmp	.Lsqrx8x_reduction_loop
3348
3349.align	32
3350.Lsqrx8x_reduction_loop:
3351	mov	8*1($tptr),%r9
3352	mov	8*2($tptr),%r10
3353	mov	8*3($tptr),%r11
3354	mov	8*4($tptr),%r12
3355	mov	%rdx,%r8
3356	imulq	%rbx,%rdx		# n0*a[i]
3357	mov	8*5($tptr),%r13
3358	mov	8*6($tptr),%r14
3359	mov	8*7($tptr),%r15
3360	mov	%rax,24+8(%rsp)		# store top-most carry bit
3361
3362	lea	8*8($tptr),$tptr
3363	xor	$carry,$carry		# cf=0,of=0
3364	mov	\$-8,%rcx
3365	jmp	.Lsqrx8x_reduce
3366
3367.align	32
3368.Lsqrx8x_reduce:
3369	mov	%r8, %rbx
3370	mulx	8*0($nptr),%rax,%r8	# n[0]
3371	adcx	%rbx,%rax		# discarded
3372	adox	%r9,%r8
3373
3374	mulx	8*1($nptr),%rbx,%r9	# n[1]
3375	adcx	%rbx,%r8
3376	adox	%r10,%r9
3377
3378	mulx	8*2($nptr),%rbx,%r10
3379	adcx	%rbx,%r9
3380	adox	%r11,%r10
3381
3382	mulx	8*3($nptr),%rbx,%r11
3383	adcx	%rbx,%r10
3384	adox	%r12,%r11
3385
3386	.byte	0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rbx,%r12
3387	 mov	%rdx,%rax
3388	 mov	%r8,%rdx
3389	adcx	%rbx,%r11
3390	adox	%r13,%r12
3391
3392	 mulx	32+8(%rsp),%rbx,%rdx	# %rdx discarded
3393	 mov	%rax,%rdx
3394	 mov	%rax,64+48+8(%rsp,%rcx,8)	# put aside n0*a[i]
3395
3396	mulx	8*5($nptr),%rax,%r13
3397	adcx	%rax,%r12
3398	adox	%r14,%r13
3399
3400	mulx	8*6($nptr),%rax,%r14
3401	adcx	%rax,%r13
3402	adox	%r15,%r14
3403
3404	mulx	8*7($nptr),%rax,%r15
3405	 mov	%rbx,%rdx
3406	adcx	%rax,%r14
3407	adox	$carry,%r15		# $carry is 0
3408	adcx	$carry,%r15		# cf=0
3409
3410	.byte	0x67,0x67,0x67
3411	inc	%rcx			# of=0
3412	jnz	.Lsqrx8x_reduce
3413
3414	mov	$carry,%rax		# xor	%rax,%rax
3415	cmp	0+8(%rsp),$nptr		# end of n[]?
3416	jae	.Lsqrx8x_no_tail
3417
3418	mov	48+8(%rsp),%rdx		# pull n0*a[0]
3419	add	8*0($tptr),%r8
3420	lea	8*8($nptr),$nptr
3421	mov	\$-8,%rcx
3422	adcx	8*1($tptr),%r9
3423	adcx	8*2($tptr),%r10
3424	adc	8*3($tptr),%r11
3425	adc	8*4($tptr),%r12
3426	adc	8*5($tptr),%r13
3427	adc	8*6($tptr),%r14
3428	adc	8*7($tptr),%r15
3429	lea	8*8($tptr),$tptr
3430	sbb	%rax,%rax		# top carry
3431
3432	xor	$carry,$carry		# of=0, cf=0
3433	mov	%rax,16+8(%rsp)
3434	jmp	.Lsqrx8x_tail
3435
3436.align	32
3437.Lsqrx8x_tail:
3438	mov	%r8,%rbx
3439	mulx	8*0($nptr),%rax,%r8
3440	adcx	%rax,%rbx
3441	adox	%r9,%r8
3442
3443	mulx	8*1($nptr),%rax,%r9
3444	adcx	%rax,%r8
3445	adox	%r10,%r9
3446
3447	mulx	8*2($nptr),%rax,%r10
3448	adcx	%rax,%r9
3449	adox	%r11,%r10
3450
3451	mulx	8*3($nptr),%rax,%r11
3452	adcx	%rax,%r10
3453	adox	%r12,%r11
3454
3455	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rax,%r12
3456	adcx	%rax,%r11
3457	adox	%r13,%r12
3458
3459	mulx	8*5($nptr),%rax,%r13
3460	adcx	%rax,%r12
3461	adox	%r14,%r13
3462
3463	mulx	8*6($nptr),%rax,%r14
3464	adcx	%rax,%r13
3465	adox	%r15,%r14
3466
3467	mulx	8*7($nptr),%rax,%r15
3468	 mov	72+48+8(%rsp,%rcx,8),%rdx	# pull n0*a[i]
3469	adcx	%rax,%r14
3470	adox	$carry,%r15
3471	 mov	%rbx,($tptr,%rcx,8)	# save result
3472	 mov	%r8,%rbx
3473	adcx	$carry,%r15		# cf=0
3474
3475	inc	%rcx			# of=0
3476	jnz	.Lsqrx8x_tail
3477
3478	cmp	0+8(%rsp),$nptr		# end of n[]?
3479	jae	.Lsqrx8x_tail_done	# break out of loop
3480
3481	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3482	 mov	48+8(%rsp),%rdx		# pull n0*a[0]
3483	 lea	8*8($nptr),$nptr
3484	adc	8*0($tptr),%r8
3485	adc	8*1($tptr),%r9
3486	adc	8*2($tptr),%r10
3487	adc	8*3($tptr),%r11
3488	adc	8*4($tptr),%r12
3489	adc	8*5($tptr),%r13
3490	adc	8*6($tptr),%r14
3491	adc	8*7($tptr),%r15
3492	lea	8*8($tptr),$tptr
3493	sbb	%rax,%rax
3494	sub	\$8,%rcx		# mov	\$-8,%rcx
3495
3496	xor	$carry,$carry		# of=0, cf=0
3497	mov	%rax,16+8(%rsp)
3498	jmp	.Lsqrx8x_tail
3499
3500.align	32
3501.Lsqrx8x_tail_done:
3502	xor	%rax,%rax
3503	add	24+8(%rsp),%r8		# can this overflow?
3504	adc	\$0,%r9
3505	adc	\$0,%r10
3506	adc	\$0,%r11
3507	adc	\$0,%r12
3508	adc	\$0,%r13
3509	adc	\$0,%r14
3510	adc	\$0,%r15
3511	adc	\$0,%rax
3512
3513	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3514.Lsqrx8x_no_tail:			# %cf is 0 if jumped here
3515	adc	8*0($tptr),%r8
3516	 movq	%xmm3,%rcx
3517	adc	8*1($tptr),%r9
3518	 mov	8*7($nptr),$carry
3519	 movq	%xmm2,$nptr		# restore $nptr
3520	adc	8*2($tptr),%r10
3521	adc	8*3($tptr),%r11
3522	adc	8*4($tptr),%r12
3523	adc	8*5($tptr),%r13
3524	adc	8*6($tptr),%r14
3525	adc	8*7($tptr),%r15
3526	adc	\$0,%rax		# top-most carry
3527
3528	mov	32+8(%rsp),%rbx		# n0
3529	mov	8*8($tptr,%rcx),%rdx	# modulo-scheduled "%r8"
3530
3531	mov	%r8,8*0($tptr)		# store top 512 bits
3532	 lea	8*8($tptr),%r8		# borrow %r8
3533	mov	%r9,8*1($tptr)
3534	mov	%r10,8*2($tptr)
3535	mov	%r11,8*3($tptr)
3536	mov	%r12,8*4($tptr)
3537	mov	%r13,8*5($tptr)
3538	mov	%r14,8*6($tptr)
3539	mov	%r15,8*7($tptr)
3540
3541	lea	8*8($tptr,%rcx),$tptr	# start of current t[] window
3542	cmp	8+8(%rsp),%r8		# end of t[]?
3543	jb	.Lsqrx8x_reduction_loop
3544	ret
3545.size	bn_sqrx8x_internal,.-bn_sqrx8x_internal
3546___
3547}
3548##############################################################
3549# Post-condition, 4x unrolled
3550#
3551{
3552my ($rptr,$nptr)=("%rdx","%rbp");
3553$code.=<<___;
3554.align	32
3555__bn_postx4x_internal:
3556	mov	8*0($nptr),%r12
3557	mov	%rcx,%r10		# -$num
3558	mov	%rcx,%r9		# -$num
3559	neg	%rax
3560	sar	\$3+2,%rcx
3561	#lea	48+8(%rsp,%r9),$tptr
3562	movq	%xmm1,$rptr		# restore $rptr
3563	movq	%xmm1,$aptr		# prepare for back-to-back call
3564	dec	%r12			# so that after 'not' we get -n[0]
3565	mov	8*1($nptr),%r13
3566	xor	%r8,%r8
3567	mov	8*2($nptr),%r14
3568	mov	8*3($nptr),%r15
3569	jmp	.Lsqrx4x_sub_entry
3570
3571.align	16
3572.Lsqrx4x_sub:
3573	mov	8*0($nptr),%r12
3574	mov	8*1($nptr),%r13
3575	mov	8*2($nptr),%r14
3576	mov	8*3($nptr),%r15
3577.Lsqrx4x_sub_entry:
3578	andn	%rax,%r12,%r12
3579	lea	8*4($nptr),$nptr
3580	andn	%rax,%r13,%r13
3581	andn	%rax,%r14,%r14
3582	andn	%rax,%r15,%r15
3583
3584	neg	%r8			# mov %r8,%cf
3585	adc	8*0($tptr),%r12
3586	adc	8*1($tptr),%r13
3587	adc	8*2($tptr),%r14
3588	adc	8*3($tptr),%r15
3589	mov	%r12,8*0($rptr)
3590	lea	8*4($tptr),$tptr
3591	mov	%r13,8*1($rptr)
3592	sbb	%r8,%r8			# mov %cf,%r8
3593	mov	%r14,8*2($rptr)
3594	mov	%r15,8*3($rptr)
3595	lea	8*4($rptr),$rptr
3596
3597	inc	%rcx
3598	jnz	.Lsqrx4x_sub
3599
3600	neg	%r9			# restore $num
3601
3602	ret
3603.size	__bn_postx4x_internal,.-__bn_postx4x_internal
3604___
3605}
3606}}}
3607{
3608my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%edx","%r8", "%r9d") : # Win64 order
3609				("%rdi","%esi","%rdx","%ecx");  # Unix order
3610my $out=$inp;
3611my $STRIDE=2**5*8;
3612my $N=$STRIDE/4;
3613
3614$code.=<<___;
3615.globl	bn_get_bits5
3616.type	bn_get_bits5,\@abi-omnipotent
3617.align	16
3618bn_get_bits5:
3619	lea	0($inp),%r10
3620	lea	1($inp),%r11
3621	mov	$num,%ecx
3622	shr	\$4,$num
3623	and	\$15,%ecx
3624	lea	-8(%ecx),%eax
3625	cmp	\$11,%ecx
3626	cmova	%r11,%r10
3627	cmova	%eax,%ecx
3628	movzw	(%r10,$num,2),%eax
3629	shrl	%cl,%eax
3630	and	\$31,%eax
3631	ret
3632.size	bn_get_bits5,.-bn_get_bits5
3633
3634.globl	bn_scatter5
3635.type	bn_scatter5,\@abi-omnipotent
3636.align	16
3637bn_scatter5:
3638	cmp	\$0, $num
3639	jz	.Lscatter_epilogue
3640	lea	($tbl,$idx,8),$tbl
3641.Lscatter:
3642	mov	($inp),%rax
3643	lea	8($inp),$inp
3644	mov	%rax,($tbl)
3645	lea	32*8($tbl),$tbl
3646	sub	\$1,$num
3647	jnz	.Lscatter
3648.Lscatter_epilogue:
3649	ret
3650.size	bn_scatter5,.-bn_scatter5
3651
3652.globl	bn_gather5
3653.type	bn_gather5,\@abi-omnipotent
3654.align	32
3655bn_gather5:
3656.LSEH_begin_bn_gather5:			# Win64 thing, but harmless in other cases
3657	# I can't trust assembler to use specific encoding:-(
3658	.byte	0x4c,0x8d,0x14,0x24			#lea    (%rsp),%r10
3659	.byte	0x48,0x81,0xec,0x08,0x01,0x00,0x00	#sub	$0x108,%rsp
3660	lea	.Linc(%rip),%rax
3661	and	\$-16,%rsp		# shouldn't be formally required
3662
3663	movd	$idx,%xmm5
3664	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
3665	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
3666	lea	128($tbl),%r11		# size optimization
3667	lea	128(%rsp),%rax		# size optimization
3668
3669	pshufd	\$0,%xmm5,%xmm5		# broadcast $idx
3670	movdqa	%xmm1,%xmm4
3671	movdqa	%xmm1,%xmm2
3672___
3673########################################################################
3674# calculate mask by comparing 0..31 to $idx and save result to stack
3675#
3676for($i=0;$i<$STRIDE/16;$i+=4) {
3677$code.=<<___;
3678	paddd	%xmm0,%xmm1
3679	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
3680___
3681$code.=<<___	if ($i);
3682	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3683___
3684$code.=<<___;
3685	movdqa	%xmm4,%xmm3
3686
3687	paddd	%xmm1,%xmm2
3688	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
3689	movdqa	%xmm0,`16*($i+0)-128`(%rax)
3690	movdqa	%xmm4,%xmm0
3691
3692	paddd	%xmm2,%xmm3
3693	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
3694	movdqa	%xmm1,`16*($i+1)-128`(%rax)
3695	movdqa	%xmm4,%xmm1
3696
3697	paddd	%xmm3,%xmm0
3698	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
3699	movdqa	%xmm2,`16*($i+2)-128`(%rax)
3700	movdqa	%xmm4,%xmm2
3701___
3702}
3703$code.=<<___;
3704	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3705	jmp	.Lgather
3706
3707.align	32
3708.Lgather:
3709	pxor	%xmm4,%xmm4
3710	pxor	%xmm5,%xmm5
3711___
3712for($i=0;$i<$STRIDE/16;$i+=4) {
3713$code.=<<___;
3714	movdqa	`16*($i+0)-128`(%r11),%xmm0
3715	movdqa	`16*($i+1)-128`(%r11),%xmm1
3716	movdqa	`16*($i+2)-128`(%r11),%xmm2
3717	pand	`16*($i+0)-128`(%rax),%xmm0
3718	movdqa	`16*($i+3)-128`(%r11),%xmm3
3719	pand	`16*($i+1)-128`(%rax),%xmm1
3720	por	%xmm0,%xmm4
3721	pand	`16*($i+2)-128`(%rax),%xmm2
3722	por	%xmm1,%xmm5
3723	pand	`16*($i+3)-128`(%rax),%xmm3
3724	por	%xmm2,%xmm4
3725	por	%xmm3,%xmm5
3726___
3727}
3728$code.=<<___;
3729	por	%xmm5,%xmm4
3730	lea	$STRIDE(%r11),%r11
3731	pshufd	\$0x4e,%xmm4,%xmm0
3732	por	%xmm4,%xmm0
3733	movq	%xmm0,($out)		# m0=bp[0]
3734	lea	8($out),$out
3735	sub	\$1,$num
3736	jnz	.Lgather
3737
3738	lea	(%r10),%rsp
3739	ret
3740.LSEH_end_bn_gather5:
3741.size	bn_gather5,.-bn_gather5
3742___
3743}
3744$code.=<<___;
3745.align	64
3746.Linc:
3747	.long	0,0, 1,1
3748	.long	2,2, 2,2
3749.asciz	"Montgomery Multiplication with scatter/gather for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
3750___
3751
3752# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
3753#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
3754if ($win64) {
3755$rec="%rcx";
3756$frame="%rdx";
3757$context="%r8";
3758$disp="%r9";
3759
3760$code.=<<___;
3761.extern	__imp_RtlVirtualUnwind
3762.type	mul_handler,\@abi-omnipotent
3763.align	16
3764mul_handler:
3765	push	%rsi
3766	push	%rdi
3767	push	%rbx
3768	push	%rbp
3769	push	%r12
3770	push	%r13
3771	push	%r14
3772	push	%r15
3773	pushfq
3774	sub	\$64,%rsp
3775
3776	mov	120($context),%rax	# pull context->Rax
3777	mov	248($context),%rbx	# pull context->Rip
3778
3779	mov	8($disp),%rsi		# disp->ImageBase
3780	mov	56($disp),%r11		# disp->HandlerData
3781
3782	mov	0(%r11),%r10d		# HandlerData[0]
3783	lea	(%rsi,%r10),%r10	# end of prologue label
3784	cmp	%r10,%rbx		# context->Rip<end of prologue label
3785	jb	.Lcommon_seh_tail
3786
3787	mov	4(%r11),%r10d		# HandlerData[1]
3788	lea	(%rsi,%r10),%r10	# beginning of body label
3789	cmp	%r10,%rbx		# context->Rip<body label
3790	jb	.Lcommon_pop_regs
3791
3792	mov	152($context),%rax	# pull context->Rsp
3793
3794	mov	8(%r11),%r10d		# HandlerData[2]
3795	lea	(%rsi,%r10),%r10	# epilogue label
3796	cmp	%r10,%rbx		# context->Rip>=epilogue label
3797	jae	.Lcommon_seh_tail
3798
3799	lea	.Lmul_epilogue(%rip),%r10
3800	cmp	%r10,%rbx
3801	ja	.Lbody_40
3802
3803	mov	192($context),%r10	# pull $num
3804	mov	8(%rax,%r10,8),%rax	# pull saved stack pointer
3805
3806	jmp	.Lcommon_pop_regs
3807
3808.Lbody_40:
3809	mov	40(%rax),%rax		# pull saved stack pointer
3810.Lcommon_pop_regs:
3811	mov	-8(%rax),%rbx
3812	mov	-16(%rax),%rbp
3813	mov	-24(%rax),%r12
3814	mov	-32(%rax),%r13
3815	mov	-40(%rax),%r14
3816	mov	-48(%rax),%r15
3817	mov	%rbx,144($context)	# restore context->Rbx
3818	mov	%rbp,160($context)	# restore context->Rbp
3819	mov	%r12,216($context)	# restore context->R12
3820	mov	%r13,224($context)	# restore context->R13
3821	mov	%r14,232($context)	# restore context->R14
3822	mov	%r15,240($context)	# restore context->R15
3823
3824.Lcommon_seh_tail:
3825	mov	8(%rax),%rdi
3826	mov	16(%rax),%rsi
3827	mov	%rax,152($context)	# restore context->Rsp
3828	mov	%rsi,168($context)	# restore context->Rsi
3829	mov	%rdi,176($context)	# restore context->Rdi
3830
3831	mov	40($disp),%rdi		# disp->ContextRecord
3832	mov	$context,%rsi		# context
3833	mov	\$154,%ecx		# sizeof(CONTEXT)
3834	.long	0xa548f3fc		# cld; rep movsq
3835
3836	mov	$disp,%rsi
3837	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
3838	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
3839	mov	0(%rsi),%r8		# arg3, disp->ControlPc
3840	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
3841	mov	40(%rsi),%r10		# disp->ContextRecord
3842	lea	56(%rsi),%r11		# &disp->HandlerData
3843	lea	24(%rsi),%r12		# &disp->EstablisherFrame
3844	mov	%r10,32(%rsp)		# arg5
3845	mov	%r11,40(%rsp)		# arg6
3846	mov	%r12,48(%rsp)		# arg7
3847	mov	%rcx,56(%rsp)		# arg8, (NULL)
3848	call	*__imp_RtlVirtualUnwind(%rip)
3849
3850	mov	\$1,%eax		# ExceptionContinueSearch
3851	add	\$64,%rsp
3852	popfq
3853	pop	%r15
3854	pop	%r14
3855	pop	%r13
3856	pop	%r12
3857	pop	%rbp
3858	pop	%rbx
3859	pop	%rdi
3860	pop	%rsi
3861	ret
3862.size	mul_handler,.-mul_handler
3863
3864.section	.pdata
3865.align	4
3866	.rva	.LSEH_begin_bn_mul_mont_gather5
3867	.rva	.LSEH_end_bn_mul_mont_gather5
3868	.rva	.LSEH_info_bn_mul_mont_gather5
3869
3870	.rva	.LSEH_begin_bn_mul4x_mont_gather5
3871	.rva	.LSEH_end_bn_mul4x_mont_gather5
3872	.rva	.LSEH_info_bn_mul4x_mont_gather5
3873
3874	.rva	.LSEH_begin_bn_power5
3875	.rva	.LSEH_end_bn_power5
3876	.rva	.LSEH_info_bn_power5
3877
3878	.rva	.LSEH_begin_bn_from_mont8x
3879	.rva	.LSEH_end_bn_from_mont8x
3880	.rva	.LSEH_info_bn_from_mont8x
3881___
3882$code.=<<___ if ($addx);
3883	.rva	.LSEH_begin_bn_mulx4x_mont_gather5
3884	.rva	.LSEH_end_bn_mulx4x_mont_gather5
3885	.rva	.LSEH_info_bn_mulx4x_mont_gather5
3886
3887	.rva	.LSEH_begin_bn_powerx5
3888	.rva	.LSEH_end_bn_powerx5
3889	.rva	.LSEH_info_bn_powerx5
3890___
3891$code.=<<___;
3892	.rva	.LSEH_begin_bn_gather5
3893	.rva	.LSEH_end_bn_gather5
3894	.rva	.LSEH_info_bn_gather5
3895
3896.section	.xdata
3897.align	8
3898.LSEH_info_bn_mul_mont_gather5:
3899	.byte	9,0,0,0
3900	.rva	mul_handler
3901	.rva	.Lmul_body,.Lmul_body,.Lmul_epilogue		# HandlerData[]
3902.align	8
3903.LSEH_info_bn_mul4x_mont_gather5:
3904	.byte	9,0,0,0
3905	.rva	mul_handler
3906	.rva	.Lmul4x_prologue,.Lmul4x_body,.Lmul4x_epilogue		# HandlerData[]
3907.align	8
3908.LSEH_info_bn_power5:
3909	.byte	9,0,0,0
3910	.rva	mul_handler
3911	.rva	.Lpower5_prologue,.Lpower5_body,.Lpower5_epilogue	# HandlerData[]
3912.align	8
3913.LSEH_info_bn_from_mont8x:
3914	.byte	9,0,0,0
3915	.rva	mul_handler
3916	.rva	.Lfrom_prologue,.Lfrom_body,.Lfrom_epilogue		# HandlerData[]
3917___
3918$code.=<<___ if ($addx);
3919.align	8
3920.LSEH_info_bn_mulx4x_mont_gather5:
3921	.byte	9,0,0,0
3922	.rva	mul_handler
3923	.rva	.Lmulx4x_prologue,.Lmulx4x_body,.Lmulx4x_epilogue	# HandlerData[]
3924.align	8
3925.LSEH_info_bn_powerx5:
3926	.byte	9,0,0,0
3927	.rva	mul_handler
3928	.rva	.Lpowerx5_prologue,.Lpowerx5_body,.Lpowerx5_epilogue	# HandlerData[]
3929___
3930$code.=<<___;
3931.align	8
3932.LSEH_info_bn_gather5:
3933	.byte	0x01,0x0b,0x03,0x0a
3934	.byte	0x0b,0x01,0x21,0x00	# sub	rsp,0x108
3935	.byte	0x04,0xa3,0x00,0x00	# lea	r10,(rsp)
3936.align	8
3937___
3938}
3939
3940$code =~ s/\`([^\`]*)\`/eval($1)/gem;
3941
3942print $code;
3943close STDOUT;
3944