xref: /freebsd/crypto/openssl/crypto/bn/asm/x86_64-mont5.pl (revision 273c26a3c3bea87a241d6879abd4f991db180bf0) 
1#!/usr/bin/env perl
2
3# ====================================================================
4# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9
10# August 2011.
11#
12# Companion to x86_64-mont.pl that optimizes cache-timing attack
13# countermeasures. The subroutines are produced by replacing bp[i]
14# references in their x86_64-mont.pl counterparts with cache-neutral
15# references to powers table computed in BN_mod_exp_mont_consttime.
16# In addition subroutine that scatters elements of the powers table
17# is implemented, so that scatter-/gathering can be tuned without
18# bn_exp.c modifications.
19
20# August 2013.
21#
22# Add MULX/AD*X code paths and additional interfaces to optimize for
23# branch prediction unit. For input lengths that are multiples of 8
24# the np argument is not just modulus value, but one interleaved
25# with 0. This is to optimize post-condition...
26
27$flavour = shift;
28$output  = shift;
29if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
30
31$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
32
33$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
34( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
35( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
36die "can't locate x86_64-xlate.pl";
37
38open OUT,"| \"$^X\" $xlate $flavour $output";
39*STDOUT=*OUT;
40
41if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
42		=~ /GNU assembler version ([2-9]\.[0-9]+)/) {
43	$addx = ($1>=2.23);
44}
45
46if (!$addx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
47	    `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
48	$addx = ($1>=2.10);
49}
50
51if (!$addx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
52	    `ml64 2>&1` =~ /Version ([0-9]+)\./) {
53	$addx = ($1>=12);
54}
55
56if (!$addx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9])\.([0-9]+)/) {
57	my $ver = $2 + $3/100.0;	# 3.1->3.01, 3.10->3.10
58	$addx = ($ver>=3.03);
59}
60
61# int bn_mul_mont_gather5(
62$rp="%rdi";	# BN_ULONG *rp,
63$ap="%rsi";	# const BN_ULONG *ap,
64$bp="%rdx";	# const BN_ULONG *bp,
65$np="%rcx";	# const BN_ULONG *np,
66$n0="%r8";	# const BN_ULONG *n0,
67$num="%r9";	# int num,
68		# int idx);	# 0 to 2^5-1, "index" in $bp holding
69				# pre-computed powers of a', interlaced
70				# in such manner that b[0] is $bp[idx],
71				# b[1] is [2^5+idx], etc.
72$lo0="%r10";
73$hi0="%r11";
74$hi1="%r13";
75$i="%r14";
76$j="%r15";
77$m0="%rbx";
78$m1="%rbp";
79
80$code=<<___;
81.text
82
83.extern	OPENSSL_ia32cap_P
84
85.globl	bn_mul_mont_gather5
86.type	bn_mul_mont_gather5,\@function,6
87.align	64
88bn_mul_mont_gather5:
89	mov	${num}d,${num}d
90	mov	%rsp,%rax
91	test	\$7,${num}d
92	jnz	.Lmul_enter
93___
94$code.=<<___ if ($addx);
95	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
96___
97$code.=<<___;
98	jmp	.Lmul4x_enter
99
100.align	16
101.Lmul_enter:
102	movd	`($win64?56:8)`(%rsp),%xmm5	# load 7th argument
103	push	%rbx
104	push	%rbp
105	push	%r12
106	push	%r13
107	push	%r14
108	push	%r15
109
110	neg	$num
111	mov	%rsp,%r11
112	lea	-280(%rsp,$num,8),%r10	# future alloca(8*(num+2)+256+8)
113	neg	$num			# restore $num
114	and	\$-1024,%r10		# minimize TLB usage
115
116	# Some OSes, *cough*-dows, insist on stack being "wired" to
117	# physical memory in strictly sequential manner, i.e. if stack
118	# allocation spans two pages, then reference to farmost one can
119	# be punishable by SEGV. But page walking can do good even on
120	# other OSes, because it guarantees that villain thread hits
121	# the guard page before it can make damage to innocent one...
122	sub	%r10,%r11
123	and	\$-4096,%r11
124	lea	(%r10,%r11),%rsp
125	mov	(%rsp),%r11
126	cmp	%r10,%rsp
127	ja	.Lmul_page_walk
128	jmp	.Lmul_page_walk_done
129
130.Lmul_page_walk:
131	lea	-4096(%rsp),%rsp
132	mov	(%rsp),%r11
133	cmp	%r10,%rsp
134	ja	.Lmul_page_walk
135.Lmul_page_walk_done:
136
137	lea	.Linc(%rip),%r10
138	mov	%rax,8(%rsp,$num,8)	# tp[num+1]=%rsp
139.Lmul_body:
140
141	lea	128($bp),%r12		# reassign $bp (+size optimization)
142___
143		$bp="%r12";
144		$STRIDE=2**5*8;		# 5 is "window size"
145		$N=$STRIDE/4;		# should match cache line size
146$code.=<<___;
147	movdqa	0(%r10),%xmm0		# 00000001000000010000000000000000
148	movdqa	16(%r10),%xmm1		# 00000002000000020000000200000002
149	lea	24-112(%rsp,$num,8),%r10# place the mask after tp[num+3] (+ICache optimization)
150	and	\$-16,%r10
151
152	pshufd	\$0,%xmm5,%xmm5		# broadcast index
153	movdqa	%xmm1,%xmm4
154	movdqa	%xmm1,%xmm2
155___
156########################################################################
157# calculate mask by comparing 0..31 to index and save result to stack
158#
159$code.=<<___;
160	paddd	%xmm0,%xmm1
161	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
162	.byte	0x67
163	movdqa	%xmm4,%xmm3
164___
165for($k=0;$k<$STRIDE/16-4;$k+=4) {
166$code.=<<___;
167	paddd	%xmm1,%xmm2
168	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
169	movdqa	%xmm0,`16*($k+0)+112`(%r10)
170	movdqa	%xmm4,%xmm0
171
172	paddd	%xmm2,%xmm3
173	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
174	movdqa	%xmm1,`16*($k+1)+112`(%r10)
175	movdqa	%xmm4,%xmm1
176
177	paddd	%xmm3,%xmm0
178	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
179	movdqa	%xmm2,`16*($k+2)+112`(%r10)
180	movdqa	%xmm4,%xmm2
181
182	paddd	%xmm0,%xmm1
183	pcmpeqd	%xmm5,%xmm0
184	movdqa	%xmm3,`16*($k+3)+112`(%r10)
185	movdqa	%xmm4,%xmm3
186___
187}
188$code.=<<___;				# last iteration can be optimized
189	paddd	%xmm1,%xmm2
190	pcmpeqd	%xmm5,%xmm1
191	movdqa	%xmm0,`16*($k+0)+112`(%r10)
192
193	paddd	%xmm2,%xmm3
194	.byte	0x67
195	pcmpeqd	%xmm5,%xmm2
196	movdqa	%xmm1,`16*($k+1)+112`(%r10)
197
198	pcmpeqd	%xmm5,%xmm3
199	movdqa	%xmm2,`16*($k+2)+112`(%r10)
200	pand	`16*($k+0)-128`($bp),%xmm0	# while it's still in register
201
202	pand	`16*($k+1)-128`($bp),%xmm1
203	pand	`16*($k+2)-128`($bp),%xmm2
204	movdqa	%xmm3,`16*($k+3)+112`(%r10)
205	pand	`16*($k+3)-128`($bp),%xmm3
206	por	%xmm2,%xmm0
207	por	%xmm3,%xmm1
208___
209for($k=0;$k<$STRIDE/16-4;$k+=4) {
210$code.=<<___;
211	movdqa	`16*($k+0)-128`($bp),%xmm4
212	movdqa	`16*($k+1)-128`($bp),%xmm5
213	movdqa	`16*($k+2)-128`($bp),%xmm2
214	pand	`16*($k+0)+112`(%r10),%xmm4
215	movdqa	`16*($k+3)-128`($bp),%xmm3
216	pand	`16*($k+1)+112`(%r10),%xmm5
217	por	%xmm4,%xmm0
218	pand	`16*($k+2)+112`(%r10),%xmm2
219	por	%xmm5,%xmm1
220	pand	`16*($k+3)+112`(%r10),%xmm3
221	por	%xmm2,%xmm0
222	por	%xmm3,%xmm1
223___
224}
225$code.=<<___;
226	por	%xmm1,%xmm0
227	pshufd	\$0x4e,%xmm0,%xmm1
228	por	%xmm1,%xmm0
229	lea	$STRIDE($bp),$bp
230	movq	%xmm0,$m0		# m0=bp[0]
231
232	mov	($n0),$n0		# pull n0[0] value
233	mov	($ap),%rax
234
235	xor	$i,$i			# i=0
236	xor	$j,$j			# j=0
237
238	mov	$n0,$m1
239	mulq	$m0			# ap[0]*bp[0]
240	mov	%rax,$lo0
241	mov	($np),%rax
242
243	imulq	$lo0,$m1		# "tp[0]"*n0
244	mov	%rdx,$hi0
245
246	mulq	$m1			# np[0]*m1
247	add	%rax,$lo0		# discarded
248	mov	8($ap),%rax
249	adc	\$0,%rdx
250	mov	%rdx,$hi1
251
252	lea	1($j),$j		# j++
253	jmp	.L1st_enter
254
255.align	16
256.L1st:
257	add	%rax,$hi1
258	mov	($ap,$j,8),%rax
259	adc	\$0,%rdx
260	add	$hi0,$hi1		# np[j]*m1+ap[j]*bp[0]
261	mov	$lo0,$hi0
262	adc	\$0,%rdx
263	mov	$hi1,-16(%rsp,$j,8)	# tp[j-1]
264	mov	%rdx,$hi1
265
266.L1st_enter:
267	mulq	$m0			# ap[j]*bp[0]
268	add	%rax,$hi0
269	mov	($np,$j,8),%rax
270	adc	\$0,%rdx
271	lea	1($j),$j		# j++
272	mov	%rdx,$lo0
273
274	mulq	$m1			# np[j]*m1
275	cmp	$num,$j
276	jne	.L1st			# note that upon exit $j==$num, so
277					# they can be used interchangeably
278
279	add	%rax,$hi1
280	adc	\$0,%rdx
281	add	$hi0,$hi1		# np[j]*m1+ap[j]*bp[0]
282	adc	\$0,%rdx
283	mov	$hi1,-16(%rsp,$num,8)	# tp[num-1]
284	mov	%rdx,$hi1
285	mov	$lo0,$hi0
286
287	xor	%rdx,%rdx
288	add	$hi0,$hi1
289	adc	\$0,%rdx
290	mov	$hi1,-8(%rsp,$num,8)
291	mov	%rdx,(%rsp,$num,8)	# store upmost overflow bit
292
293	lea	1($i),$i		# i++
294	jmp	.Louter
295.align	16
296.Louter:
297	lea	24+128(%rsp,$num,8),%rdx	# where 256-byte mask is (+size optimization)
298	and	\$-16,%rdx
299	pxor	%xmm4,%xmm4
300	pxor	%xmm5,%xmm5
301___
302for($k=0;$k<$STRIDE/16;$k+=4) {
303$code.=<<___;
304	movdqa	`16*($k+0)-128`($bp),%xmm0
305	movdqa	`16*($k+1)-128`($bp),%xmm1
306	movdqa	`16*($k+2)-128`($bp),%xmm2
307	movdqa	`16*($k+3)-128`($bp),%xmm3
308	pand	`16*($k+0)-128`(%rdx),%xmm0
309	pand	`16*($k+1)-128`(%rdx),%xmm1
310	por	%xmm0,%xmm4
311	pand	`16*($k+2)-128`(%rdx),%xmm2
312	por	%xmm1,%xmm5
313	pand	`16*($k+3)-128`(%rdx),%xmm3
314	por	%xmm2,%xmm4
315	por	%xmm3,%xmm5
316___
317}
318$code.=<<___;
319	por	%xmm5,%xmm4
320	pshufd	\$0x4e,%xmm4,%xmm0
321	por	%xmm4,%xmm0
322	lea	$STRIDE($bp),$bp
323
324	mov	($ap),%rax		# ap[0]
325	movq	%xmm0,$m0		# m0=bp[i]
326
327	xor	$j,$j			# j=0
328	mov	$n0,$m1
329	mov	(%rsp),$lo0
330
331	mulq	$m0			# ap[0]*bp[i]
332	add	%rax,$lo0		# ap[0]*bp[i]+tp[0]
333	mov	($np),%rax
334	adc	\$0,%rdx
335
336	imulq	$lo0,$m1		# tp[0]*n0
337	mov	%rdx,$hi0
338
339	mulq	$m1			# np[0]*m1
340	add	%rax,$lo0		# discarded
341	mov	8($ap),%rax
342	adc	\$0,%rdx
343	mov	8(%rsp),$lo0		# tp[1]
344	mov	%rdx,$hi1
345
346	lea	1($j),$j		# j++
347	jmp	.Linner_enter
348
349.align	16
350.Linner:
351	add	%rax,$hi1
352	mov	($ap,$j,8),%rax
353	adc	\$0,%rdx
354	add	$lo0,$hi1		# np[j]*m1+ap[j]*bp[i]+tp[j]
355	mov	(%rsp,$j,8),$lo0
356	adc	\$0,%rdx
357	mov	$hi1,-16(%rsp,$j,8)	# tp[j-1]
358	mov	%rdx,$hi1
359
360.Linner_enter:
361	mulq	$m0			# ap[j]*bp[i]
362	add	%rax,$hi0
363	mov	($np,$j,8),%rax
364	adc	\$0,%rdx
365	add	$hi0,$lo0		# ap[j]*bp[i]+tp[j]
366	mov	%rdx,$hi0
367	adc	\$0,$hi0
368	lea	1($j),$j		# j++
369
370	mulq	$m1			# np[j]*m1
371	cmp	$num,$j
372	jne	.Linner			# note that upon exit $j==$num, so
373					# they can be used interchangeably
374	add	%rax,$hi1
375	adc	\$0,%rdx
376	add	$lo0,$hi1		# np[j]*m1+ap[j]*bp[i]+tp[j]
377	mov	(%rsp,$num,8),$lo0
378	adc	\$0,%rdx
379	mov	$hi1,-16(%rsp,$num,8)	# tp[num-1]
380	mov	%rdx,$hi1
381
382	xor	%rdx,%rdx
383	add	$hi0,$hi1
384	adc	\$0,%rdx
385	add	$lo0,$hi1		# pull upmost overflow bit
386	adc	\$0,%rdx
387	mov	$hi1,-8(%rsp,$num,8)
388	mov	%rdx,(%rsp,$num,8)	# store upmost overflow bit
389
390	lea	1($i),$i		# i++
391	cmp	$num,$i
392	jb	.Louter
393
394	xor	$i,$i			# i=0 and clear CF!
395	mov	(%rsp),%rax		# tp[0]
396	lea	(%rsp),$ap		# borrow ap for tp
397	mov	$num,$j			# j=num
398	jmp	.Lsub
399.align	16
400.Lsub:	sbb	($np,$i,8),%rax
401	mov	%rax,($rp,$i,8)		# rp[i]=tp[i]-np[i]
402	mov	8($ap,$i,8),%rax	# tp[i+1]
403	lea	1($i),$i		# i++
404	dec	$j			# doesnn't affect CF!
405	jnz	.Lsub
406
407	sbb	\$0,%rax		# handle upmost overflow bit
408	xor	$i,$i
409	and	%rax,$ap
410	not	%rax
411	mov	$rp,$np
412	and	%rax,$np
413	mov	$num,$j			# j=num
414	or	$np,$ap			# ap=borrow?tp:rp
415.align	16
416.Lcopy:					# copy or in-place refresh
417	mov	($ap,$i,8),%rax
418	mov	$i,(%rsp,$i,8)		# zap temporary vector
419	mov	%rax,($rp,$i,8)		# rp[i]=tp[i]
420	lea	1($i),$i
421	sub	\$1,$j
422	jnz	.Lcopy
423
424	mov	8(%rsp,$num,8),%rsi	# restore %rsp
425	mov	\$1,%rax
426
427	mov	-48(%rsi),%r15
428	mov	-40(%rsi),%r14
429	mov	-32(%rsi),%r13
430	mov	-24(%rsi),%r12
431	mov	-16(%rsi),%rbp
432	mov	-8(%rsi),%rbx
433	lea	(%rsi),%rsp
434.Lmul_epilogue:
435	ret
436.size	bn_mul_mont_gather5,.-bn_mul_mont_gather5
437___
438{{{
439my @A=("%r10","%r11");
440my @N=("%r13","%rdi");
441$code.=<<___;
442.type	bn_mul4x_mont_gather5,\@function,6
443.align	32
444bn_mul4x_mont_gather5:
445	.byte	0x67
446	mov	%rsp,%rax
447.Lmul4x_enter:
448___
449$code.=<<___ if ($addx);
450	and	\$0x80108,%r11d
451	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
452	je	.Lmulx4x_enter
453___
454$code.=<<___;
455	push	%rbx
456	push	%rbp
457	push	%r12
458	push	%r13
459	push	%r14
460	push	%r15
461.Lmul4x_prologue:
462
463	.byte	0x67
464	shl	\$3,${num}d		# convert $num to bytes
465	lea	($num,$num,2),%r10	# 3*$num in bytes
466	neg	$num			# -$num
467
468	##############################################################
469	# Ensure that stack frame doesn't alias with $rptr+3*$num
470	# modulo 4096, which covers ret[num], am[num] and n[num]
471	# (see bn_exp.c). This is done to allow memory disambiguation
472	# logic do its magic. [Extra [num] is allocated in order
473	# to align with bn_power5's frame, which is cleansed after
474	# completing exponentiation. Extra 256 bytes is for power mask
475	# calculated from 7th argument, the index.]
476	#
477	lea	-320(%rsp,$num,2),%r11
478	mov	%rsp,%rbp
479	sub	$rp,%r11
480	and	\$4095,%r11
481	cmp	%r11,%r10
482	jb	.Lmul4xsp_alt
483	sub	%r11,%rbp		# align with $rp
484	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
485	jmp	.Lmul4xsp_done
486
487.align	32
488.Lmul4xsp_alt:
489	lea	4096-320(,$num,2),%r10
490	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
491	sub	%r10,%r11
492	mov	\$0,%r10
493	cmovc	%r10,%r11
494	sub	%r11,%rbp
495.Lmul4xsp_done:
496	and	\$-64,%rbp
497	mov	%rsp,%r11
498	sub	%rbp,%r11
499	and	\$-4096,%r11
500	lea	(%rbp,%r11),%rsp
501	mov	(%rsp),%r10
502	cmp	%rbp,%rsp
503	ja	.Lmul4x_page_walk
504	jmp	.Lmul4x_page_walk_done
505
506.Lmul4x_page_walk:
507	lea	-4096(%rsp),%rsp
508	mov	(%rsp),%r10
509	cmp	%rbp,%rsp
510	ja	.Lmul4x_page_walk
511.Lmul4x_page_walk_done:
512
513	neg	$num
514
515	mov	%rax,40(%rsp)
516.Lmul4x_body:
517
518	call	mul4x_internal
519
520	mov	40(%rsp),%rsi		# restore %rsp
521	mov	\$1,%rax
522
523	mov	-48(%rsi),%r15
524	mov	-40(%rsi),%r14
525	mov	-32(%rsi),%r13
526	mov	-24(%rsi),%r12
527	mov	-16(%rsi),%rbp
528	mov	-8(%rsi),%rbx
529	lea	(%rsi),%rsp
530.Lmul4x_epilogue:
531	ret
532.size	bn_mul4x_mont_gather5,.-bn_mul4x_mont_gather5
533
534.type	mul4x_internal,\@abi-omnipotent
535.align	32
536mul4x_internal:
537	shl	\$5,$num		# $num was in bytes
538	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument, index
539	lea	.Linc(%rip),%rax
540	lea	128(%rdx,$num),%r13	# end of powers table (+size optimization)
541	shr	\$5,$num		# restore $num
542___
543		$bp="%r12";
544		$STRIDE=2**5*8;		# 5 is "window size"
545		$N=$STRIDE/4;		# should match cache line size
546		$tp=$i;
547$code.=<<___;
548	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
549	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
550	lea	88-112(%rsp,$num),%r10	# place the mask after tp[num+1] (+ICache optimization)
551	lea	128(%rdx),$bp		# size optimization
552
553	pshufd	\$0,%xmm5,%xmm5		# broadcast index
554	movdqa	%xmm1,%xmm4
555	.byte	0x67,0x67
556	movdqa	%xmm1,%xmm2
557___
558########################################################################
559# calculate mask by comparing 0..31 to index and save result to stack
560#
561$code.=<<___;
562	paddd	%xmm0,%xmm1
563	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
564	.byte	0x67
565	movdqa	%xmm4,%xmm3
566___
567for($i=0;$i<$STRIDE/16-4;$i+=4) {
568$code.=<<___;
569	paddd	%xmm1,%xmm2
570	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
571	movdqa	%xmm0,`16*($i+0)+112`(%r10)
572	movdqa	%xmm4,%xmm0
573
574	paddd	%xmm2,%xmm3
575	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
576	movdqa	%xmm1,`16*($i+1)+112`(%r10)
577	movdqa	%xmm4,%xmm1
578
579	paddd	%xmm3,%xmm0
580	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
581	movdqa	%xmm2,`16*($i+2)+112`(%r10)
582	movdqa	%xmm4,%xmm2
583
584	paddd	%xmm0,%xmm1
585	pcmpeqd	%xmm5,%xmm0
586	movdqa	%xmm3,`16*($i+3)+112`(%r10)
587	movdqa	%xmm4,%xmm3
588___
589}
590$code.=<<___;				# last iteration can be optimized
591	paddd	%xmm1,%xmm2
592	pcmpeqd	%xmm5,%xmm1
593	movdqa	%xmm0,`16*($i+0)+112`(%r10)
594
595	paddd	%xmm2,%xmm3
596	.byte	0x67
597	pcmpeqd	%xmm5,%xmm2
598	movdqa	%xmm1,`16*($i+1)+112`(%r10)
599
600	pcmpeqd	%xmm5,%xmm3
601	movdqa	%xmm2,`16*($i+2)+112`(%r10)
602	pand	`16*($i+0)-128`($bp),%xmm0	# while it's still in register
603
604	pand	`16*($i+1)-128`($bp),%xmm1
605	pand	`16*($i+2)-128`($bp),%xmm2
606	movdqa	%xmm3,`16*($i+3)+112`(%r10)
607	pand	`16*($i+3)-128`($bp),%xmm3
608	por	%xmm2,%xmm0
609	por	%xmm3,%xmm1
610___
611for($i=0;$i<$STRIDE/16-4;$i+=4) {
612$code.=<<___;
613	movdqa	`16*($i+0)-128`($bp),%xmm4
614	movdqa	`16*($i+1)-128`($bp),%xmm5
615	movdqa	`16*($i+2)-128`($bp),%xmm2
616	pand	`16*($i+0)+112`(%r10),%xmm4
617	movdqa	`16*($i+3)-128`($bp),%xmm3
618	pand	`16*($i+1)+112`(%r10),%xmm5
619	por	%xmm4,%xmm0
620	pand	`16*($i+2)+112`(%r10),%xmm2
621	por	%xmm5,%xmm1
622	pand	`16*($i+3)+112`(%r10),%xmm3
623	por	%xmm2,%xmm0
624	por	%xmm3,%xmm1
625___
626}
627$code.=<<___;
628	por	%xmm1,%xmm0
629	pshufd	\$0x4e,%xmm0,%xmm1
630	por	%xmm1,%xmm0
631	lea	$STRIDE($bp),$bp
632	movq	%xmm0,$m0		# m0=bp[0]
633
634	mov	%r13,16+8(%rsp)		# save end of b[num]
635	mov	$rp, 56+8(%rsp)		# save $rp
636
637	mov	($n0),$n0		# pull n0[0] value
638	mov	($ap),%rax
639	lea	($ap,$num),$ap		# end of a[num]
640	neg	$num
641
642	mov	$n0,$m1
643	mulq	$m0			# ap[0]*bp[0]
644	mov	%rax,$A[0]
645	mov	($np),%rax
646
647	imulq	$A[0],$m1		# "tp[0]"*n0
648	lea	64+8(%rsp),$tp
649	mov	%rdx,$A[1]
650
651	mulq	$m1			# np[0]*m1
652	add	%rax,$A[0]		# discarded
653	mov	8($ap,$num),%rax
654	adc	\$0,%rdx
655	mov	%rdx,$N[1]
656
657	mulq	$m0
658	add	%rax,$A[1]
659	mov	8*1($np),%rax
660	adc	\$0,%rdx
661	mov	%rdx,$A[0]
662
663	mulq	$m1
664	add	%rax,$N[1]
665	mov	16($ap,$num),%rax
666	adc	\$0,%rdx
667	add	$A[1],$N[1]
668	lea	4*8($num),$j		# j=4
669	lea	8*4($np),$np
670	adc	\$0,%rdx
671	mov	$N[1],($tp)
672	mov	%rdx,$N[0]
673	jmp	.L1st4x
674
675.align	32
676.L1st4x:
677	mulq	$m0			# ap[j]*bp[0]
678	add	%rax,$A[0]
679	mov	-8*2($np),%rax
680	lea	32($tp),$tp
681	adc	\$0,%rdx
682	mov	%rdx,$A[1]
683
684	mulq	$m1			# np[j]*m1
685	add	%rax,$N[0]
686	mov	-8($ap,$j),%rax
687	adc	\$0,%rdx
688	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
689	adc	\$0,%rdx
690	mov	$N[0],-24($tp)		# tp[j-1]
691	mov	%rdx,$N[1]
692
693	mulq	$m0			# ap[j]*bp[0]
694	add	%rax,$A[1]
695	mov	-8*1($np),%rax
696	adc	\$0,%rdx
697	mov	%rdx,$A[0]
698
699	mulq	$m1			# np[j]*m1
700	add	%rax,$N[1]
701	mov	($ap,$j),%rax
702	adc	\$0,%rdx
703	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
704	adc	\$0,%rdx
705	mov	$N[1],-16($tp)		# tp[j-1]
706	mov	%rdx,$N[0]
707
708	mulq	$m0			# ap[j]*bp[0]
709	add	%rax,$A[0]
710	mov	8*0($np),%rax
711	adc	\$0,%rdx
712	mov	%rdx,$A[1]
713
714	mulq	$m1			# np[j]*m1
715	add	%rax,$N[0]
716	mov	8($ap,$j),%rax
717	adc	\$0,%rdx
718	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
719	adc	\$0,%rdx
720	mov	$N[0],-8($tp)		# tp[j-1]
721	mov	%rdx,$N[1]
722
723	mulq	$m0			# ap[j]*bp[0]
724	add	%rax,$A[1]
725	mov	8*1($np),%rax
726	adc	\$0,%rdx
727	mov	%rdx,$A[0]
728
729	mulq	$m1			# np[j]*m1
730	add	%rax,$N[1]
731	mov	16($ap,$j),%rax
732	adc	\$0,%rdx
733	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
734	lea	8*4($np),$np
735	adc	\$0,%rdx
736	mov	$N[1],($tp)		# tp[j-1]
737	mov	%rdx,$N[0]
738
739	add	\$32,$j			# j+=4
740	jnz	.L1st4x
741
742	mulq	$m0			# ap[j]*bp[0]
743	add	%rax,$A[0]
744	mov	-8*2($np),%rax
745	lea	32($tp),$tp
746	adc	\$0,%rdx
747	mov	%rdx,$A[1]
748
749	mulq	$m1			# np[j]*m1
750	add	%rax,$N[0]
751	mov	-8($ap),%rax
752	adc	\$0,%rdx
753	add	$A[0],$N[0]		# np[j]*m1+ap[j]*bp[0]
754	adc	\$0,%rdx
755	mov	$N[0],-24($tp)		# tp[j-1]
756	mov	%rdx,$N[1]
757
758	mulq	$m0			# ap[j]*bp[0]
759	add	%rax,$A[1]
760	mov	-8*1($np),%rax
761	adc	\$0,%rdx
762	mov	%rdx,$A[0]
763
764	mulq	$m1			# np[j]*m1
765	add	%rax,$N[1]
766	mov	($ap,$num),%rax		# ap[0]
767	adc	\$0,%rdx
768	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[0]
769	adc	\$0,%rdx
770	mov	$N[1],-16($tp)		# tp[j-1]
771	mov	%rdx,$N[0]
772
773	lea	($np,$num),$np		# rewind $np
774
775	xor	$N[1],$N[1]
776	add	$A[0],$N[0]
777	adc	\$0,$N[1]
778	mov	$N[0],-8($tp)
779
780	jmp	.Louter4x
781
782.align	32
783.Louter4x:
784	lea	16+128($tp),%rdx	# where 256-byte mask is (+size optimization)
785	pxor	%xmm4,%xmm4
786	pxor	%xmm5,%xmm5
787___
788for($i=0;$i<$STRIDE/16;$i+=4) {
789$code.=<<___;
790	movdqa	`16*($i+0)-128`($bp),%xmm0
791	movdqa	`16*($i+1)-128`($bp),%xmm1
792	movdqa	`16*($i+2)-128`($bp),%xmm2
793	movdqa	`16*($i+3)-128`($bp),%xmm3
794	pand	`16*($i+0)-128`(%rdx),%xmm0
795	pand	`16*($i+1)-128`(%rdx),%xmm1
796	por	%xmm0,%xmm4
797	pand	`16*($i+2)-128`(%rdx),%xmm2
798	por	%xmm1,%xmm5
799	pand	`16*($i+3)-128`(%rdx),%xmm3
800	por	%xmm2,%xmm4
801	por	%xmm3,%xmm5
802___
803}
804$code.=<<___;
805	por	%xmm5,%xmm4
806	pshufd	\$0x4e,%xmm4,%xmm0
807	por	%xmm4,%xmm0
808	lea	$STRIDE($bp),$bp
809	movq	%xmm0,$m0		# m0=bp[i]
810
811	mov	($tp,$num),$A[0]
812	mov	$n0,$m1
813	mulq	$m0			# ap[0]*bp[i]
814	add	%rax,$A[0]		# ap[0]*bp[i]+tp[0]
815	mov	($np),%rax
816	adc	\$0,%rdx
817
818	imulq	$A[0],$m1		# tp[0]*n0
819	mov	%rdx,$A[1]
820	mov	$N[1],($tp)		# store upmost overflow bit
821
822	lea	($tp,$num),$tp		# rewind $tp
823
824	mulq	$m1			# np[0]*m1
825	add	%rax,$A[0]		# "$N[0]", discarded
826	mov	8($ap,$num),%rax
827	adc	\$0,%rdx
828	mov	%rdx,$N[1]
829
830	mulq	$m0			# ap[j]*bp[i]
831	add	%rax,$A[1]
832	mov	8*1($np),%rax
833	adc	\$0,%rdx
834	add	8($tp),$A[1]		# +tp[1]
835	adc	\$0,%rdx
836	mov	%rdx,$A[0]
837
838	mulq	$m1			# np[j]*m1
839	add	%rax,$N[1]
840	mov	16($ap,$num),%rax
841	adc	\$0,%rdx
842	add	$A[1],$N[1]		# np[j]*m1+ap[j]*bp[i]+tp[j]
843	lea	4*8($num),$j		# j=4
844	lea	8*4($np),$np
845	adc	\$0,%rdx
846	mov	%rdx,$N[0]
847	jmp	.Linner4x
848
849.align	32
850.Linner4x:
851	mulq	$m0			# ap[j]*bp[i]
852	add	%rax,$A[0]
853	mov	-8*2($np),%rax
854	adc	\$0,%rdx
855	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
856	lea	32($tp),$tp
857	adc	\$0,%rdx
858	mov	%rdx,$A[1]
859
860	mulq	$m1			# np[j]*m1
861	add	%rax,$N[0]
862	mov	-8($ap,$j),%rax
863	adc	\$0,%rdx
864	add	$A[0],$N[0]
865	adc	\$0,%rdx
866	mov	$N[1],-32($tp)		# tp[j-1]
867	mov	%rdx,$N[1]
868
869	mulq	$m0			# ap[j]*bp[i]
870	add	%rax,$A[1]
871	mov	-8*1($np),%rax
872	adc	\$0,%rdx
873	add	-8($tp),$A[1]
874	adc	\$0,%rdx
875	mov	%rdx,$A[0]
876
877	mulq	$m1			# np[j]*m1
878	add	%rax,$N[1]
879	mov	($ap,$j),%rax
880	adc	\$0,%rdx
881	add	$A[1],$N[1]
882	adc	\$0,%rdx
883	mov	$N[0],-24($tp)		# tp[j-1]
884	mov	%rdx,$N[0]
885
886	mulq	$m0			# ap[j]*bp[i]
887	add	%rax,$A[0]
888	mov	8*0($np),%rax
889	adc	\$0,%rdx
890	add	($tp),$A[0]		# ap[j]*bp[i]+tp[j]
891	adc	\$0,%rdx
892	mov	%rdx,$A[1]
893
894	mulq	$m1			# np[j]*m1
895	add	%rax,$N[0]
896	mov	8($ap,$j),%rax
897	adc	\$0,%rdx
898	add	$A[0],$N[0]
899	adc	\$0,%rdx
900	mov	$N[1],-16($tp)		# tp[j-1]
901	mov	%rdx,$N[1]
902
903	mulq	$m0			# ap[j]*bp[i]
904	add	%rax,$A[1]
905	mov	8*1($np),%rax
906	adc	\$0,%rdx
907	add	8($tp),$A[1]
908	adc	\$0,%rdx
909	mov	%rdx,$A[0]
910
911	mulq	$m1			# np[j]*m1
912	add	%rax,$N[1]
913	mov	16($ap,$j),%rax
914	adc	\$0,%rdx
915	add	$A[1],$N[1]
916	lea	8*4($np),$np
917	adc	\$0,%rdx
918	mov	$N[0],-8($tp)		# tp[j-1]
919	mov	%rdx,$N[0]
920
921	add	\$32,$j			# j+=4
922	jnz	.Linner4x
923
924	mulq	$m0			# ap[j]*bp[i]
925	add	%rax,$A[0]
926	mov	-8*2($np),%rax
927	adc	\$0,%rdx
928	add	16($tp),$A[0]		# ap[j]*bp[i]+tp[j]
929	lea	32($tp),$tp
930	adc	\$0,%rdx
931	mov	%rdx,$A[1]
932
933	mulq	$m1			# np[j]*m1
934	add	%rax,$N[0]
935	mov	-8($ap),%rax
936	adc	\$0,%rdx
937	add	$A[0],$N[0]
938	adc	\$0,%rdx
939	mov	$N[1],-32($tp)		# tp[j-1]
940	mov	%rdx,$N[1]
941
942	mulq	$m0			# ap[j]*bp[i]
943	add	%rax,$A[1]
944	mov	$m1,%rax
945	mov	-8*1($np),$m1
946	adc	\$0,%rdx
947	add	-8($tp),$A[1]
948	adc	\$0,%rdx
949	mov	%rdx,$A[0]
950
951	mulq	$m1			# np[j]*m1
952	add	%rax,$N[1]
953	mov	($ap,$num),%rax		# ap[0]
954	adc	\$0,%rdx
955	add	$A[1],$N[1]
956	adc	\$0,%rdx
957	mov	$N[0],-24($tp)		# tp[j-1]
958	mov	%rdx,$N[0]
959
960	mov	$N[1],-16($tp)		# tp[j-1]
961	lea	($np,$num),$np		# rewind $np
962
963	xor	$N[1],$N[1]
964	add	$A[0],$N[0]
965	adc	\$0,$N[1]
966	add	($tp),$N[0]		# pull upmost overflow bit
967	adc	\$0,$N[1]		# upmost overflow bit
968	mov	$N[0],-8($tp)
969
970	cmp	16+8(%rsp),$bp
971	jb	.Louter4x
972___
973if (1) {
974$code.=<<___;
975	xor	%rax,%rax
976	sub	$N[0],$m1		# compare top-most words
977	adc	$j,$j			# $j is zero
978	or	$j,$N[1]
979	sub	$N[1],%rax		# %rax=-$N[1]
980	lea	($tp,$num),%rbx		# tptr in .sqr4x_sub
981	mov	($np),%r12
982	lea	($np),%rbp		# nptr in .sqr4x_sub
983	mov	%r9,%rcx
984	sar	\$3+2,%rcx
985	mov	56+8(%rsp),%rdi		# rptr in .sqr4x_sub
986	dec	%r12			# so that after 'not' we get -n[0]
987	xor	%r10,%r10
988	mov	8*1(%rbp),%r13
989	mov	8*2(%rbp),%r14
990	mov	8*3(%rbp),%r15
991	jmp	.Lsqr4x_sub_entry
992___
993} else {
994my @ri=("%rax",$bp,$m0,$m1);
995my $rp="%rdx";
996$code.=<<___
997	xor	\$1,$N[1]
998	lea	($tp,$num),$tp		# rewind $tp
999	sar	\$5,$num		# cf=0
1000	lea	($np,$N[1],8),$np
1001	mov	56+8(%rsp),$rp		# restore $rp
1002	jmp	.Lsub4x
1003
1004.align	32
1005.Lsub4x:
1006	.byte	0x66
1007	mov	8*0($tp),@ri[0]
1008	mov	8*1($tp),@ri[1]
1009	.byte	0x66
1010	sbb	16*0($np),@ri[0]
1011	mov	8*2($tp),@ri[2]
1012	sbb	16*1($np),@ri[1]
1013	mov	3*8($tp),@ri[3]
1014	lea	4*8($tp),$tp
1015	sbb	16*2($np),@ri[2]
1016	mov	@ri[0],8*0($rp)
1017	sbb	16*3($np),@ri[3]
1018	lea	16*4($np),$np
1019	mov	@ri[1],8*1($rp)
1020	mov	@ri[2],8*2($rp)
1021	mov	@ri[3],8*3($rp)
1022	lea	8*4($rp),$rp
1023
1024	inc	$num
1025	jnz	.Lsub4x
1026
1027	ret
1028___
1029}
1030$code.=<<___;
1031.size	mul4x_internal,.-mul4x_internal
1032___
1033}}}
1034{{{
1035######################################################################
1036# void bn_power5(
1037my $rptr="%rdi";	# BN_ULONG *rptr,
1038my $aptr="%rsi";	# const BN_ULONG *aptr,
1039my $bptr="%rdx";	# const void *table,
1040my $nptr="%rcx";	# const BN_ULONG *nptr,
1041my $n0  ="%r8";		# const BN_ULONG *n0);
1042my $num ="%r9";		# int num, has to be divisible by 8
1043			# int pwr
1044
1045my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
1046my @A0=("%r10","%r11");
1047my @A1=("%r12","%r13");
1048my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
1049
1050$code.=<<___;
1051.globl	bn_power5
1052.type	bn_power5,\@function,6
1053.align	32
1054bn_power5:
1055	mov	%rsp,%rax
1056___
1057$code.=<<___ if ($addx);
1058	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
1059	and	\$0x80108,%r11d
1060	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
1061	je	.Lpowerx5_enter
1062___
1063$code.=<<___;
1064	push	%rbx
1065	push	%rbp
1066	push	%r12
1067	push	%r13
1068	push	%r14
1069	push	%r15
1070.Lpower5_prologue:
1071
1072	shl	\$3,${num}d		# convert $num to bytes
1073	lea	($num,$num,2),%r10d	# 3*$num
1074	neg	$num
1075	mov	($n0),$n0		# *n0
1076
1077	##############################################################
1078	# Ensure that stack frame doesn't alias with $rptr+3*$num
1079	# modulo 4096, which covers ret[num], am[num] and n[num]
1080	# (see bn_exp.c). This is done to allow memory disambiguation
1081	# logic do its magic. [Extra 256 bytes is for power mask
1082	# calculated from 7th argument, the index.]
1083	#
1084	lea	-320(%rsp,$num,2),%r11
1085	mov	%rsp,%rbp
1086	sub	$rptr,%r11
1087	and	\$4095,%r11
1088	cmp	%r11,%r10
1089	jb	.Lpwr_sp_alt
1090	sub	%r11,%rbp		# align with $aptr
1091	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1092	jmp	.Lpwr_sp_done
1093
1094.align	32
1095.Lpwr_sp_alt:
1096	lea	4096-320(,$num,2),%r10
1097	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*num*8+256)
1098	sub	%r10,%r11
1099	mov	\$0,%r10
1100	cmovc	%r10,%r11
1101	sub	%r11,%rbp
1102.Lpwr_sp_done:
1103	and	\$-64,%rbp
1104	mov	%rsp,%r11
1105	sub	%rbp,%r11
1106	and	\$-4096,%r11
1107	lea	(%rbp,%r11),%rsp
1108	mov	(%rsp),%r10
1109	cmp	%rbp,%rsp
1110	ja	.Lpwr_page_walk
1111	jmp	.Lpwr_page_walk_done
1112
1113.Lpwr_page_walk:
1114	lea	-4096(%rsp),%rsp
1115	mov	(%rsp),%r10
1116	cmp	%rbp,%rsp
1117	ja	.Lpwr_page_walk
1118.Lpwr_page_walk_done:
1119
1120	mov	$num,%r10
1121	neg	$num
1122
1123	##############################################################
1124	# Stack layout
1125	#
1126	# +0	saved $num, used in reduction section
1127	# +8	&t[2*$num], used in reduction section
1128	# +32	saved *n0
1129	# +40	saved %rsp
1130	# +48	t[2*$num]
1131	#
1132	mov	$n0,  32(%rsp)
1133	mov	%rax, 40(%rsp)		# save original %rsp
1134.Lpower5_body:
1135	movq	$rptr,%xmm1		# save $rptr, used in sqr8x
1136	movq	$nptr,%xmm2		# save $nptr
1137	movq	%r10, %xmm3		# -$num, used in sqr8x
1138	movq	$bptr,%xmm4
1139
1140	call	__bn_sqr8x_internal
1141	call	__bn_post4x_internal
1142	call	__bn_sqr8x_internal
1143	call	__bn_post4x_internal
1144	call	__bn_sqr8x_internal
1145	call	__bn_post4x_internal
1146	call	__bn_sqr8x_internal
1147	call	__bn_post4x_internal
1148	call	__bn_sqr8x_internal
1149	call	__bn_post4x_internal
1150
1151	movq	%xmm2,$nptr
1152	movq	%xmm4,$bptr
1153	mov	$aptr,$rptr
1154	mov	40(%rsp),%rax
1155	lea	32(%rsp),$n0
1156
1157	call	mul4x_internal
1158
1159	mov	40(%rsp),%rsi		# restore %rsp
1160	mov	\$1,%rax
1161	mov	-48(%rsi),%r15
1162	mov	-40(%rsi),%r14
1163	mov	-32(%rsi),%r13
1164	mov	-24(%rsi),%r12
1165	mov	-16(%rsi),%rbp
1166	mov	-8(%rsi),%rbx
1167	lea	(%rsi),%rsp
1168.Lpower5_epilogue:
1169	ret
1170.size	bn_power5,.-bn_power5
1171
1172.globl	bn_sqr8x_internal
1173.hidden	bn_sqr8x_internal
1174.type	bn_sqr8x_internal,\@abi-omnipotent
1175.align	32
1176bn_sqr8x_internal:
1177__bn_sqr8x_internal:
1178	##############################################################
1179	# Squaring part:
1180	#
1181	# a) multiply-n-add everything but a[i]*a[i];
1182	# b) shift result of a) by 1 to the left and accumulate
1183	#    a[i]*a[i] products;
1184	#
1185	##############################################################
1186	#                                                     a[1]a[0]
1187	#                                                 a[2]a[0]
1188	#                                             a[3]a[0]
1189	#                                             a[2]a[1]
1190	#                                         a[4]a[0]
1191	#                                         a[3]a[1]
1192	#                                     a[5]a[0]
1193	#                                     a[4]a[1]
1194	#                                     a[3]a[2]
1195	#                                 a[6]a[0]
1196	#                                 a[5]a[1]
1197	#                                 a[4]a[2]
1198	#                             a[7]a[0]
1199	#                             a[6]a[1]
1200	#                             a[5]a[2]
1201	#                             a[4]a[3]
1202	#                         a[7]a[1]
1203	#                         a[6]a[2]
1204	#                         a[5]a[3]
1205	#                     a[7]a[2]
1206	#                     a[6]a[3]
1207	#                     a[5]a[4]
1208	#                 a[7]a[3]
1209	#                 a[6]a[4]
1210	#             a[7]a[4]
1211	#             a[6]a[5]
1212	#         a[7]a[5]
1213	#     a[7]a[6]
1214	#                                                     a[1]a[0]
1215	#                                                 a[2]a[0]
1216	#                                             a[3]a[0]
1217	#                                         a[4]a[0]
1218	#                                     a[5]a[0]
1219	#                                 a[6]a[0]
1220	#                             a[7]a[0]
1221	#                                             a[2]a[1]
1222	#                                         a[3]a[1]
1223	#                                     a[4]a[1]
1224	#                                 a[5]a[1]
1225	#                             a[6]a[1]
1226	#                         a[7]a[1]
1227	#                                     a[3]a[2]
1228	#                                 a[4]a[2]
1229	#                             a[5]a[2]
1230	#                         a[6]a[2]
1231	#                     a[7]a[2]
1232	#                             a[4]a[3]
1233	#                         a[5]a[3]
1234	#                     a[6]a[3]
1235	#                 a[7]a[3]
1236	#                     a[5]a[4]
1237	#                 a[6]a[4]
1238	#             a[7]a[4]
1239	#             a[6]a[5]
1240	#         a[7]a[5]
1241	#     a[7]a[6]
1242	#                                                         a[0]a[0]
1243	#                                                 a[1]a[1]
1244	#                                         a[2]a[2]
1245	#                                 a[3]a[3]
1246	#                         a[4]a[4]
1247	#                 a[5]a[5]
1248	#         a[6]a[6]
1249	# a[7]a[7]
1250
1251	lea	32(%r10),$i		# $i=-($num-32)
1252	lea	($aptr,$num),$aptr	# end of a[] buffer, ($aptr,$i)=&ap[2]
1253
1254	mov	$num,$j			# $j=$num
1255
1256					# comments apply to $num==8 case
1257	mov	-32($aptr,$i),$a0	# a[0]
1258	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1259	mov	-24($aptr,$i),%rax	# a[1]
1260	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1261	mov	-16($aptr,$i),$ai	# a[2]
1262	mov	%rax,$a1
1263
1264	mul	$a0			# a[1]*a[0]
1265	mov	%rax,$A0[0]		# a[1]*a[0]
1266	 mov	$ai,%rax		# a[2]
1267	mov	%rdx,$A0[1]
1268	mov	$A0[0],-24($tptr,$i)	# t[1]
1269
1270	mul	$a0			# a[2]*a[0]
1271	add	%rax,$A0[1]
1272	 mov	$ai,%rax
1273	adc	\$0,%rdx
1274	mov	$A0[1],-16($tptr,$i)	# t[2]
1275	mov	%rdx,$A0[0]
1276
1277
1278	 mov	-8($aptr,$i),$ai	# a[3]
1279	mul	$a1			# a[2]*a[1]
1280	mov	%rax,$A1[0]		# a[2]*a[1]+t[3]
1281	 mov	$ai,%rax
1282	mov	%rdx,$A1[1]
1283
1284	 lea	($i),$j
1285	mul	$a0			# a[3]*a[0]
1286	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1287	 mov	$ai,%rax
1288	mov	%rdx,$A0[1]
1289	adc	\$0,$A0[1]
1290	add	$A1[0],$A0[0]
1291	adc	\$0,$A0[1]
1292	mov	$A0[0],-8($tptr,$j)	# t[3]
1293	jmp	.Lsqr4x_1st
1294
1295.align	32
1296.Lsqr4x_1st:
1297	 mov	($aptr,$j),$ai		# a[4]
1298	mul	$a1			# a[3]*a[1]
1299	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1300	 mov	$ai,%rax
1301	mov	%rdx,$A1[0]
1302	adc	\$0,$A1[0]
1303
1304	mul	$a0			# a[4]*a[0]
1305	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1306	 mov	$ai,%rax		# a[3]
1307	 mov	8($aptr,$j),$ai		# a[5]
1308	mov	%rdx,$A0[0]
1309	adc	\$0,$A0[0]
1310	add	$A1[1],$A0[1]
1311	adc	\$0,$A0[0]
1312
1313
1314	mul	$a1			# a[4]*a[3]
1315	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1316	 mov	$ai,%rax
1317	 mov	$A0[1],($tptr,$j)	# t[4]
1318	mov	%rdx,$A1[1]
1319	adc	\$0,$A1[1]
1320
1321	mul	$a0			# a[5]*a[2]
1322	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1323	 mov	$ai,%rax
1324	 mov	16($aptr,$j),$ai	# a[6]
1325	mov	%rdx,$A0[1]
1326	adc	\$0,$A0[1]
1327	add	$A1[0],$A0[0]
1328	adc	\$0,$A0[1]
1329
1330	mul	$a1			# a[5]*a[3]
1331	add	%rax,$A1[1]		# a[5]*a[3]+t[6]
1332	 mov	$ai,%rax
1333	 mov	$A0[0],8($tptr,$j)	# t[5]
1334	mov	%rdx,$A1[0]
1335	adc	\$0,$A1[0]
1336
1337	mul	$a0			# a[6]*a[2]
1338	add	%rax,$A0[1]		# a[6]*a[2]+a[5]*a[3]+t[6]
1339	 mov	$ai,%rax		# a[3]
1340	 mov	24($aptr,$j),$ai	# a[7]
1341	mov	%rdx,$A0[0]
1342	adc	\$0,$A0[0]
1343	add	$A1[1],$A0[1]
1344	adc	\$0,$A0[0]
1345
1346
1347	mul	$a1			# a[6]*a[5]
1348	add	%rax,$A1[0]		# a[6]*a[5]+t[7]
1349	 mov	$ai,%rax
1350	 mov	$A0[1],16($tptr,$j)	# t[6]
1351	mov	%rdx,$A1[1]
1352	adc	\$0,$A1[1]
1353	 lea	32($j),$j
1354
1355	mul	$a0			# a[7]*a[4]
1356	add	%rax,$A0[0]		# a[7]*a[4]+a[6]*a[5]+t[6]
1357	 mov	$ai,%rax
1358	mov	%rdx,$A0[1]
1359	adc	\$0,$A0[1]
1360	add	$A1[0],$A0[0]
1361	adc	\$0,$A0[1]
1362	mov	$A0[0],-8($tptr,$j)	# t[7]
1363
1364	cmp	\$0,$j
1365	jne	.Lsqr4x_1st
1366
1367	mul	$a1			# a[7]*a[5]
1368	add	%rax,$A1[1]
1369	lea	16($i),$i
1370	adc	\$0,%rdx
1371	add	$A0[1],$A1[1]
1372	adc	\$0,%rdx
1373
1374	mov	$A1[1],($tptr)		# t[8]
1375	mov	%rdx,$A1[0]
1376	mov	%rdx,8($tptr)		# t[9]
1377	jmp	.Lsqr4x_outer
1378
1379.align	32
1380.Lsqr4x_outer:				# comments apply to $num==6 case
1381	mov	-32($aptr,$i),$a0	# a[0]
1382	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1383	mov	-24($aptr,$i),%rax	# a[1]
1384	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1385	mov	-16($aptr,$i),$ai	# a[2]
1386	mov	%rax,$a1
1387
1388	mul	$a0			# a[1]*a[0]
1389	mov	-24($tptr,$i),$A0[0]	# t[1]
1390	add	%rax,$A0[0]		# a[1]*a[0]+t[1]
1391	 mov	$ai,%rax		# a[2]
1392	adc	\$0,%rdx
1393	mov	$A0[0],-24($tptr,$i)	# t[1]
1394	mov	%rdx,$A0[1]
1395
1396	mul	$a0			# a[2]*a[0]
1397	add	%rax,$A0[1]
1398	 mov	$ai,%rax
1399	adc	\$0,%rdx
1400	add	-16($tptr,$i),$A0[1]	# a[2]*a[0]+t[2]
1401	mov	%rdx,$A0[0]
1402	adc	\$0,$A0[0]
1403	mov	$A0[1],-16($tptr,$i)	# t[2]
1404
1405	xor	$A1[0],$A1[0]
1406
1407	 mov	-8($aptr,$i),$ai	# a[3]
1408	mul	$a1			# a[2]*a[1]
1409	add	%rax,$A1[0]		# a[2]*a[1]+t[3]
1410	 mov	$ai,%rax
1411	adc	\$0,%rdx
1412	add	-8($tptr,$i),$A1[0]
1413	mov	%rdx,$A1[1]
1414	adc	\$0,$A1[1]
1415
1416	mul	$a0			# a[3]*a[0]
1417	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1418	 mov	$ai,%rax
1419	adc	\$0,%rdx
1420	add	$A1[0],$A0[0]
1421	mov	%rdx,$A0[1]
1422	adc	\$0,$A0[1]
1423	mov	$A0[0],-8($tptr,$i)	# t[3]
1424
1425	lea	($i),$j
1426	jmp	.Lsqr4x_inner
1427
1428.align	32
1429.Lsqr4x_inner:
1430	 mov	($aptr,$j),$ai		# a[4]
1431	mul	$a1			# a[3]*a[1]
1432	add	%rax,$A1[1]		# a[3]*a[1]+t[4]
1433	 mov	$ai,%rax
1434	mov	%rdx,$A1[0]
1435	adc	\$0,$A1[0]
1436	add	($tptr,$j),$A1[1]
1437	adc	\$0,$A1[0]
1438
1439	.byte	0x67
1440	mul	$a0			# a[4]*a[0]
1441	add	%rax,$A0[1]		# a[4]*a[0]+a[3]*a[1]+t[4]
1442	 mov	$ai,%rax		# a[3]
1443	 mov	8($aptr,$j),$ai		# a[5]
1444	mov	%rdx,$A0[0]
1445	adc	\$0,$A0[0]
1446	add	$A1[1],$A0[1]
1447	adc	\$0,$A0[0]
1448
1449	mul	$a1			# a[4]*a[3]
1450	add	%rax,$A1[0]		# a[4]*a[3]+t[5]
1451	mov	$A0[1],($tptr,$j)	# t[4]
1452	 mov	$ai,%rax
1453	mov	%rdx,$A1[1]
1454	adc	\$0,$A1[1]
1455	add	8($tptr,$j),$A1[0]
1456	lea	16($j),$j		# j++
1457	adc	\$0,$A1[1]
1458
1459	mul	$a0			# a[5]*a[2]
1460	add	%rax,$A0[0]		# a[5]*a[2]+a[4]*a[3]+t[5]
1461	 mov	$ai,%rax
1462	adc	\$0,%rdx
1463	add	$A1[0],$A0[0]
1464	mov	%rdx,$A0[1]
1465	adc	\$0,$A0[1]
1466	mov	$A0[0],-8($tptr,$j)	# t[5], "preloaded t[1]" below
1467
1468	cmp	\$0,$j
1469	jne	.Lsqr4x_inner
1470
1471	.byte	0x67
1472	mul	$a1			# a[5]*a[3]
1473	add	%rax,$A1[1]
1474	adc	\$0,%rdx
1475	add	$A0[1],$A1[1]
1476	adc	\$0,%rdx
1477
1478	mov	$A1[1],($tptr)		# t[6], "preloaded t[2]" below
1479	mov	%rdx,$A1[0]
1480	mov	%rdx,8($tptr)		# t[7], "preloaded t[3]" below
1481
1482	add	\$16,$i
1483	jnz	.Lsqr4x_outer
1484
1485					# comments apply to $num==4 case
1486	mov	-32($aptr),$a0		# a[0]
1487	lea	48+8(%rsp,$num,2),$tptr	# end of tp[] buffer, &tp[2*$num]
1488	mov	-24($aptr),%rax		# a[1]
1489	lea	-32($tptr,$i),$tptr	# end of tp[] window, &tp[2*$num-"$i"]
1490	mov	-16($aptr),$ai		# a[2]
1491	mov	%rax,$a1
1492
1493	mul	$a0			# a[1]*a[0]
1494	add	%rax,$A0[0]		# a[1]*a[0]+t[1], preloaded t[1]
1495	 mov	$ai,%rax		# a[2]
1496	mov	%rdx,$A0[1]
1497	adc	\$0,$A0[1]
1498
1499	mul	$a0			# a[2]*a[0]
1500	add	%rax,$A0[1]
1501	 mov	$ai,%rax
1502	 mov	$A0[0],-24($tptr)	# t[1]
1503	mov	%rdx,$A0[0]
1504	adc	\$0,$A0[0]
1505	add	$A1[1],$A0[1]		# a[2]*a[0]+t[2], preloaded t[2]
1506	 mov	-8($aptr),$ai		# a[3]
1507	adc	\$0,$A0[0]
1508
1509	mul	$a1			# a[2]*a[1]
1510	add	%rax,$A1[0]		# a[2]*a[1]+t[3], preloaded t[3]
1511	 mov	$ai,%rax
1512	 mov	$A0[1],-16($tptr)	# t[2]
1513	mov	%rdx,$A1[1]
1514	adc	\$0,$A1[1]
1515
1516	mul	$a0			# a[3]*a[0]
1517	add	%rax,$A0[0]		# a[3]*a[0]+a[2]*a[1]+t[3]
1518	 mov	$ai,%rax
1519	mov	%rdx,$A0[1]
1520	adc	\$0,$A0[1]
1521	add	$A1[0],$A0[0]
1522	adc	\$0,$A0[1]
1523	mov	$A0[0],-8($tptr)	# t[3]
1524
1525	mul	$a1			# a[3]*a[1]
1526	add	%rax,$A1[1]
1527	 mov	-16($aptr),%rax		# a[2]
1528	adc	\$0,%rdx
1529	add	$A0[1],$A1[1]
1530	adc	\$0,%rdx
1531
1532	mov	$A1[1],($tptr)		# t[4]
1533	mov	%rdx,$A1[0]
1534	mov	%rdx,8($tptr)		# t[5]
1535
1536	mul	$ai			# a[2]*a[3]
1537___
1538{
1539my ($shift,$carry)=($a0,$a1);
1540my @S=(@A1,$ai,$n0);
1541$code.=<<___;
1542	 add	\$16,$i
1543	 xor	$shift,$shift
1544	 sub	$num,$i			# $i=16-$num
1545	 xor	$carry,$carry
1546
1547	add	$A1[0],%rax		# t[5]
1548	adc	\$0,%rdx
1549	mov	%rax,8($tptr)		# t[5]
1550	mov	%rdx,16($tptr)		# t[6]
1551	mov	$carry,24($tptr)	# t[7]
1552
1553	 mov	-16($aptr,$i),%rax	# a[0]
1554	lea	48+8(%rsp),$tptr
1555	 xor	$A0[0],$A0[0]		# t[0]
1556	 mov	8($tptr),$A0[1]		# t[1]
1557
1558	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1559	shr	\$63,$A0[0]
1560	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1561	shr	\$63,$A0[1]
1562	or	$A0[0],$S[1]		# | t[2*i]>>63
1563	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1564	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1565	mul	%rax			# a[i]*a[i]
1566	neg	$carry			# mov $carry,cf
1567	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1568	adc	%rax,$S[0]
1569	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1570	mov	$S[0],($tptr)
1571	adc	%rdx,$S[1]
1572
1573	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1574	 mov	$S[1],8($tptr)
1575	 sbb	$carry,$carry		# mov cf,$carry
1576	shr	\$63,$A0[0]
1577	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1578	shr	\$63,$A0[1]
1579	or	$A0[0],$S[3]		# | t[2*i]>>63
1580	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1581	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1582	mul	%rax			# a[i]*a[i]
1583	neg	$carry			# mov $carry,cf
1584	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1585	adc	%rax,$S[2]
1586	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1587	mov	$S[2],16($tptr)
1588	adc	%rdx,$S[3]
1589	lea	16($i),$i
1590	mov	$S[3],24($tptr)
1591	sbb	$carry,$carry		# mov cf,$carry
1592	lea	64($tptr),$tptr
1593	jmp	.Lsqr4x_shift_n_add
1594
1595.align	32
1596.Lsqr4x_shift_n_add:
1597	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1598	shr	\$63,$A0[0]
1599	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1600	shr	\$63,$A0[1]
1601	or	$A0[0],$S[1]		# | t[2*i]>>63
1602	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1603	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1604	mul	%rax			# a[i]*a[i]
1605	neg	$carry			# mov $carry,cf
1606	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1607	adc	%rax,$S[0]
1608	 mov	-8($aptr,$i),%rax	# a[i+1]	# prefetch
1609	mov	$S[0],-32($tptr)
1610	adc	%rdx,$S[1]
1611
1612	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1613	 mov	$S[1],-24($tptr)
1614	 sbb	$carry,$carry		# mov cf,$carry
1615	shr	\$63,$A0[0]
1616	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1617	shr	\$63,$A0[1]
1618	or	$A0[0],$S[3]		# | t[2*i]>>63
1619	 mov	0($tptr),$A0[0]		# t[2*i+2]	# prefetch
1620	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1621	mul	%rax			# a[i]*a[i]
1622	neg	$carry			# mov $carry,cf
1623	 mov	8($tptr),$A0[1]		# t[2*i+2+1]	# prefetch
1624	adc	%rax,$S[2]
1625	 mov	0($aptr,$i),%rax	# a[i+1]	# prefetch
1626	mov	$S[2],-16($tptr)
1627	adc	%rdx,$S[3]
1628
1629	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1630	 mov	$S[3],-8($tptr)
1631	 sbb	$carry,$carry		# mov cf,$carry
1632	shr	\$63,$A0[0]
1633	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1634	shr	\$63,$A0[1]
1635	or	$A0[0],$S[1]		# | t[2*i]>>63
1636	 mov	16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1637	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1638	mul	%rax			# a[i]*a[i]
1639	neg	$carry			# mov $carry,cf
1640	 mov	24($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1641	adc	%rax,$S[0]
1642	 mov	8($aptr,$i),%rax	# a[i+1]	# prefetch
1643	mov	$S[0],0($tptr)
1644	adc	%rdx,$S[1]
1645
1646	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1 | shift
1647	 mov	$S[1],8($tptr)
1648	 sbb	$carry,$carry		# mov cf,$carry
1649	shr	\$63,$A0[0]
1650	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1651	shr	\$63,$A0[1]
1652	or	$A0[0],$S[3]		# | t[2*i]>>63
1653	 mov	32($tptr),$A0[0]	# t[2*i+2]	# prefetch
1654	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1655	mul	%rax			# a[i]*a[i]
1656	neg	$carry			# mov $carry,cf
1657	 mov	40($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1658	adc	%rax,$S[2]
1659	 mov	16($aptr,$i),%rax	# a[i+1]	# prefetch
1660	mov	$S[2],16($tptr)
1661	adc	%rdx,$S[3]
1662	mov	$S[3],24($tptr)
1663	sbb	$carry,$carry		# mov cf,$carry
1664	lea	64($tptr),$tptr
1665	add	\$32,$i
1666	jnz	.Lsqr4x_shift_n_add
1667
1668	lea	($shift,$A0[0],2),$S[0]	# t[2*i]<<1 | shift
1669	.byte	0x67
1670	shr	\$63,$A0[0]
1671	lea	($j,$A0[1],2),$S[1]	# t[2*i+1]<<1 |
1672	shr	\$63,$A0[1]
1673	or	$A0[0],$S[1]		# | t[2*i]>>63
1674	 mov	-16($tptr),$A0[0]	# t[2*i+2]	# prefetch
1675	mov	$A0[1],$shift		# shift=t[2*i+1]>>63
1676	mul	%rax			# a[i]*a[i]
1677	neg	$carry			# mov $carry,cf
1678	 mov	-8($tptr),$A0[1]	# t[2*i+2+1]	# prefetch
1679	adc	%rax,$S[0]
1680	 mov	-8($aptr),%rax		# a[i+1]	# prefetch
1681	mov	$S[0],-32($tptr)
1682	adc	%rdx,$S[1]
1683
1684	lea	($shift,$A0[0],2),$S[2]	# t[2*i]<<1|shift
1685	 mov	$S[1],-24($tptr)
1686	 sbb	$carry,$carry		# mov cf,$carry
1687	shr	\$63,$A0[0]
1688	lea	($j,$A0[1],2),$S[3]	# t[2*i+1]<<1 |
1689	shr	\$63,$A0[1]
1690	or	$A0[0],$S[3]		# | t[2*i]>>63
1691	mul	%rax			# a[i]*a[i]
1692	neg	$carry			# mov $carry,cf
1693	adc	%rax,$S[2]
1694	adc	%rdx,$S[3]
1695	mov	$S[2],-16($tptr)
1696	mov	$S[3],-8($tptr)
1697___
1698}
1699######################################################################
1700# Montgomery reduction part, "word-by-word" algorithm.
1701#
1702# This new path is inspired by multiple submissions from Intel, by
1703# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
1704# Vinodh Gopal...
1705{
1706my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx");
1707
1708$code.=<<___;
1709	movq	%xmm2,$nptr
1710__bn_sqr8x_reduction:
1711	xor	%rax,%rax
1712	lea	($nptr,$num),%rcx	# end of n[]
1713	lea	48+8(%rsp,$num,2),%rdx	# end of t[] buffer
1714	mov	%rcx,0+8(%rsp)
1715	lea	48+8(%rsp,$num),$tptr	# end of initial t[] window
1716	mov	%rdx,8+8(%rsp)
1717	neg	$num
1718	jmp	.L8x_reduction_loop
1719
1720.align	32
1721.L8x_reduction_loop:
1722	lea	($tptr,$num),$tptr	# start of current t[] window
1723	.byte	0x66
1724	mov	8*0($tptr),$m0
1725	mov	8*1($tptr),%r9
1726	mov	8*2($tptr),%r10
1727	mov	8*3($tptr),%r11
1728	mov	8*4($tptr),%r12
1729	mov	8*5($tptr),%r13
1730	mov	8*6($tptr),%r14
1731	mov	8*7($tptr),%r15
1732	mov	%rax,(%rdx)		# store top-most carry bit
1733	lea	8*8($tptr),$tptr
1734
1735	.byte	0x67
1736	mov	$m0,%r8
1737	imulq	32+8(%rsp),$m0		# n0*a[0]
1738	mov	8*0($nptr),%rax		# n[0]
1739	mov	\$8,%ecx
1740	jmp	.L8x_reduce
1741
1742.align	32
1743.L8x_reduce:
1744	mulq	$m0
1745	 mov	8*1($nptr),%rax		# n[1]
1746	neg	%r8
1747	mov	%rdx,%r8
1748	adc	\$0,%r8
1749
1750	mulq	$m0
1751	add	%rax,%r9
1752	 mov	8*2($nptr),%rax
1753	adc	\$0,%rdx
1754	add	%r9,%r8
1755	 mov	$m0,48-8+8(%rsp,%rcx,8)	# put aside n0*a[i]
1756	mov	%rdx,%r9
1757	adc	\$0,%r9
1758
1759	mulq	$m0
1760	add	%rax,%r10
1761	 mov	8*3($nptr),%rax
1762	adc	\$0,%rdx
1763	add	%r10,%r9
1764	 mov	32+8(%rsp),$carry	# pull n0, borrow $carry
1765	mov	%rdx,%r10
1766	adc	\$0,%r10
1767
1768	mulq	$m0
1769	add	%rax,%r11
1770	 mov	8*4($nptr),%rax
1771	adc	\$0,%rdx
1772	 imulq	%r8,$carry		# modulo-scheduled
1773	add	%r11,%r10
1774	mov	%rdx,%r11
1775	adc	\$0,%r11
1776
1777	mulq	$m0
1778	add	%rax,%r12
1779	 mov	8*5($nptr),%rax
1780	adc	\$0,%rdx
1781	add	%r12,%r11
1782	mov	%rdx,%r12
1783	adc	\$0,%r12
1784
1785	mulq	$m0
1786	add	%rax,%r13
1787	 mov	8*6($nptr),%rax
1788	adc	\$0,%rdx
1789	add	%r13,%r12
1790	mov	%rdx,%r13
1791	adc	\$0,%r13
1792
1793	mulq	$m0
1794	add	%rax,%r14
1795	 mov	8*7($nptr),%rax
1796	adc	\$0,%rdx
1797	add	%r14,%r13
1798	mov	%rdx,%r14
1799	adc	\$0,%r14
1800
1801	mulq	$m0
1802	 mov	$carry,$m0		# n0*a[i]
1803	add	%rax,%r15
1804	 mov	8*0($nptr),%rax		# n[0]
1805	adc	\$0,%rdx
1806	add	%r15,%r14
1807	mov	%rdx,%r15
1808	adc	\$0,%r15
1809
1810	dec	%ecx
1811	jnz	.L8x_reduce
1812
1813	lea	8*8($nptr),$nptr
1814	xor	%rax,%rax
1815	mov	8+8(%rsp),%rdx		# pull end of t[]
1816	cmp	0+8(%rsp),$nptr		# end of n[]?
1817	jae	.L8x_no_tail
1818
1819	.byte	0x66
1820	add	8*0($tptr),%r8
1821	adc	8*1($tptr),%r9
1822	adc	8*2($tptr),%r10
1823	adc	8*3($tptr),%r11
1824	adc	8*4($tptr),%r12
1825	adc	8*5($tptr),%r13
1826	adc	8*6($tptr),%r14
1827	adc	8*7($tptr),%r15
1828	sbb	$carry,$carry		# top carry
1829
1830	mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1831	mov	\$8,%ecx
1832	mov	8*0($nptr),%rax
1833	jmp	.L8x_tail
1834
1835.align	32
1836.L8x_tail:
1837	mulq	$m0
1838	add	%rax,%r8
1839	 mov	8*1($nptr),%rax
1840	 mov	%r8,($tptr)		# save result
1841	mov	%rdx,%r8
1842	adc	\$0,%r8
1843
1844	mulq	$m0
1845	add	%rax,%r9
1846	 mov	8*2($nptr),%rax
1847	adc	\$0,%rdx
1848	add	%r9,%r8
1849	 lea	8($tptr),$tptr		# $tptr++
1850	mov	%rdx,%r9
1851	adc	\$0,%r9
1852
1853	mulq	$m0
1854	add	%rax,%r10
1855	 mov	8*3($nptr),%rax
1856	adc	\$0,%rdx
1857	add	%r10,%r9
1858	mov	%rdx,%r10
1859	adc	\$0,%r10
1860
1861	mulq	$m0
1862	add	%rax,%r11
1863	 mov	8*4($nptr),%rax
1864	adc	\$0,%rdx
1865	add	%r11,%r10
1866	mov	%rdx,%r11
1867	adc	\$0,%r11
1868
1869	mulq	$m0
1870	add	%rax,%r12
1871	 mov	8*5($nptr),%rax
1872	adc	\$0,%rdx
1873	add	%r12,%r11
1874	mov	%rdx,%r12
1875	adc	\$0,%r12
1876
1877	mulq	$m0
1878	add	%rax,%r13
1879	 mov	8*6($nptr),%rax
1880	adc	\$0,%rdx
1881	add	%r13,%r12
1882	mov	%rdx,%r13
1883	adc	\$0,%r13
1884
1885	mulq	$m0
1886	add	%rax,%r14
1887	 mov	8*7($nptr),%rax
1888	adc	\$0,%rdx
1889	add	%r14,%r13
1890	mov	%rdx,%r14
1891	adc	\$0,%r14
1892
1893	mulq	$m0
1894	 mov	48-16+8(%rsp,%rcx,8),$m0# pull n0*a[i]
1895	add	%rax,%r15
1896	adc	\$0,%rdx
1897	add	%r15,%r14
1898	 mov	8*0($nptr),%rax		# pull n[0]
1899	mov	%rdx,%r15
1900	adc	\$0,%r15
1901
1902	dec	%ecx
1903	jnz	.L8x_tail
1904
1905	lea	8*8($nptr),$nptr
1906	mov	8+8(%rsp),%rdx		# pull end of t[]
1907	cmp	0+8(%rsp),$nptr		# end of n[]?
1908	jae	.L8x_tail_done		# break out of loop
1909
1910	 mov	48+56+8(%rsp),$m0	# pull n0*a[0]
1911	neg	$carry
1912	 mov	8*0($nptr),%rax		# pull n[0]
1913	adc	8*0($tptr),%r8
1914	adc	8*1($tptr),%r9
1915	adc	8*2($tptr),%r10
1916	adc	8*3($tptr),%r11
1917	adc	8*4($tptr),%r12
1918	adc	8*5($tptr),%r13
1919	adc	8*6($tptr),%r14
1920	adc	8*7($tptr),%r15
1921	sbb	$carry,$carry		# top carry
1922
1923	mov	\$8,%ecx
1924	jmp	.L8x_tail
1925
1926.align	32
1927.L8x_tail_done:
1928	add	(%rdx),%r8		# can this overflow?
1929	adc	\$0,%r9
1930	adc	\$0,%r10
1931	adc	\$0,%r11
1932	adc	\$0,%r12
1933	adc	\$0,%r13
1934	adc	\$0,%r14
1935	adc	\$0,%r15		# can't overflow, because we
1936					# started with "overhung" part
1937					# of multiplication
1938	xor	%rax,%rax
1939
1940	neg	$carry
1941.L8x_no_tail:
1942	adc	8*0($tptr),%r8
1943	adc	8*1($tptr),%r9
1944	adc	8*2($tptr),%r10
1945	adc	8*3($tptr),%r11
1946	adc	8*4($tptr),%r12
1947	adc	8*5($tptr),%r13
1948	adc	8*6($tptr),%r14
1949	adc	8*7($tptr),%r15
1950	adc	\$0,%rax		# top-most carry
1951	 mov	-8($nptr),%rcx		# np[num-1]
1952	 xor	$carry,$carry
1953
1954	movq	%xmm2,$nptr		# restore $nptr
1955
1956	mov	%r8,8*0($tptr)		# store top 512 bits
1957	mov	%r9,8*1($tptr)
1958	 movq	%xmm3,$num		# $num is %r9, can't be moved upwards
1959	mov	%r10,8*2($tptr)
1960	mov	%r11,8*3($tptr)
1961	mov	%r12,8*4($tptr)
1962	mov	%r13,8*5($tptr)
1963	mov	%r14,8*6($tptr)
1964	mov	%r15,8*7($tptr)
1965	lea	8*8($tptr),$tptr
1966
1967	cmp	%rdx,$tptr		# end of t[]?
1968	jb	.L8x_reduction_loop
1969	ret
1970.size	bn_sqr8x_internal,.-bn_sqr8x_internal
1971___
1972}
1973##############################################################
1974# Post-condition, 4x unrolled
1975#
1976{
1977my ($tptr,$nptr)=("%rbx","%rbp");
1978$code.=<<___;
1979.type	__bn_post4x_internal,\@abi-omnipotent
1980.align	32
1981__bn_post4x_internal:
1982	mov	8*0($nptr),%r12
1983	lea	(%rdi,$num),$tptr	# %rdi was $tptr above
1984	mov	$num,%rcx
1985	movq	%xmm1,$rptr		# restore $rptr
1986	neg	%rax
1987	movq	%xmm1,$aptr		# prepare for back-to-back call
1988	sar	\$3+2,%rcx
1989	dec	%r12			# so that after 'not' we get -n[0]
1990	xor	%r10,%r10
1991	mov	8*1($nptr),%r13
1992	mov	8*2($nptr),%r14
1993	mov	8*3($nptr),%r15
1994	jmp	.Lsqr4x_sub_entry
1995
1996.align	16
1997.Lsqr4x_sub:
1998	mov	8*0($nptr),%r12
1999	mov	8*1($nptr),%r13
2000	mov	8*2($nptr),%r14
2001	mov	8*3($nptr),%r15
2002.Lsqr4x_sub_entry:
2003	lea	8*4($nptr),$nptr
2004	not	%r12
2005	not	%r13
2006	not	%r14
2007	not	%r15
2008	and	%rax,%r12
2009	and	%rax,%r13
2010	and	%rax,%r14
2011	and	%rax,%r15
2012
2013	neg	%r10			# mov %r10,%cf
2014	adc	8*0($tptr),%r12
2015	adc	8*1($tptr),%r13
2016	adc	8*2($tptr),%r14
2017	adc	8*3($tptr),%r15
2018	mov	%r12,8*0($rptr)
2019	lea	8*4($tptr),$tptr
2020	mov	%r13,8*1($rptr)
2021	sbb	%r10,%r10		# mov %cf,%r10
2022	mov	%r14,8*2($rptr)
2023	mov	%r15,8*3($rptr)
2024	lea	8*4($rptr),$rptr
2025
2026	inc	%rcx			# pass %cf
2027	jnz	.Lsqr4x_sub
2028
2029	mov	$num,%r10		# prepare for back-to-back call
2030	neg	$num			# restore $num
2031	ret
2032.size	__bn_post4x_internal,.-__bn_post4x_internal
2033___
2034}
2035{
2036$code.=<<___;
2037.globl	bn_from_montgomery
2038.type	bn_from_montgomery,\@abi-omnipotent
2039.align	32
2040bn_from_montgomery:
2041	testl	\$7,`($win64?"48(%rsp)":"%r9d")`
2042	jz	bn_from_mont8x
2043	xor	%eax,%eax
2044	ret
2045.size	bn_from_montgomery,.-bn_from_montgomery
2046
2047.type	bn_from_mont8x,\@function,6
2048.align	32
2049bn_from_mont8x:
2050	.byte	0x67
2051	mov	%rsp,%rax
2052	push	%rbx
2053	push	%rbp
2054	push	%r12
2055	push	%r13
2056	push	%r14
2057	push	%r15
2058.Lfrom_prologue:
2059
2060	shl	\$3,${num}d		# convert $num to bytes
2061	lea	($num,$num,2),%r10	# 3*$num in bytes
2062	neg	$num
2063	mov	($n0),$n0		# *n0
2064
2065	##############################################################
2066	# Ensure that stack frame doesn't alias with $rptr+3*$num
2067	# modulo 4096, which covers ret[num], am[num] and n[num]
2068	# (see bn_exp.c). The stack is allocated to aligned with
2069	# bn_power5's frame, and as bn_from_montgomery happens to be
2070	# last operation, we use the opportunity to cleanse it.
2071	#
2072	lea	-320(%rsp,$num,2),%r11
2073	mov	%rsp,%rbp
2074	sub	$rptr,%r11
2075	and	\$4095,%r11
2076	cmp	%r11,%r10
2077	jb	.Lfrom_sp_alt
2078	sub	%r11,%rbp		# align with $aptr
2079	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2080	jmp	.Lfrom_sp_done
2081
2082.align	32
2083.Lfrom_sp_alt:
2084	lea	4096-320(,$num,2),%r10
2085	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2086	sub	%r10,%r11
2087	mov	\$0,%r10
2088	cmovc	%r10,%r11
2089	sub	%r11,%rbp
2090.Lfrom_sp_done:
2091	and	\$-64,%rbp
2092	mov	%rsp,%r11
2093	sub	%rbp,%r11
2094	and	\$-4096,%r11
2095	lea	(%rbp,%r11),%rsp
2096	mov	(%rsp),%r10
2097	cmp	%rbp,%rsp
2098	ja	.Lfrom_page_walk
2099	jmp	.Lfrom_page_walk_done
2100
2101.Lfrom_page_walk:
2102	lea	-4096(%rsp),%rsp
2103	mov	(%rsp),%r10
2104	cmp	%rbp,%rsp
2105	ja	.Lfrom_page_walk
2106.Lfrom_page_walk_done:
2107
2108	mov	$num,%r10
2109	neg	$num
2110
2111	##############################################################
2112	# Stack layout
2113	#
2114	# +0	saved $num, used in reduction section
2115	# +8	&t[2*$num], used in reduction section
2116	# +32	saved *n0
2117	# +40	saved %rsp
2118	# +48	t[2*$num]
2119	#
2120	mov	$n0,  32(%rsp)
2121	mov	%rax, 40(%rsp)		# save original %rsp
2122.Lfrom_body:
2123	mov	$num,%r11
2124	lea	48(%rsp),%rax
2125	pxor	%xmm0,%xmm0
2126	jmp	.Lmul_by_1
2127
2128.align	32
2129.Lmul_by_1:
2130	movdqu	($aptr),%xmm1
2131	movdqu	16($aptr),%xmm2
2132	movdqu	32($aptr),%xmm3
2133	movdqa	%xmm0,(%rax,$num)
2134	movdqu	48($aptr),%xmm4
2135	movdqa	%xmm0,16(%rax,$num)
2136	.byte	0x48,0x8d,0xb6,0x40,0x00,0x00,0x00	# lea	64($aptr),$aptr
2137	movdqa	%xmm1,(%rax)
2138	movdqa	%xmm0,32(%rax,$num)
2139	movdqa	%xmm2,16(%rax)
2140	movdqa	%xmm0,48(%rax,$num)
2141	movdqa	%xmm3,32(%rax)
2142	movdqa	%xmm4,48(%rax)
2143	lea	64(%rax),%rax
2144	sub	\$64,%r11
2145	jnz	.Lmul_by_1
2146
2147	movq	$rptr,%xmm1
2148	movq	$nptr,%xmm2
2149	.byte	0x67
2150	mov	$nptr,%rbp
2151	movq	%r10, %xmm3		# -num
2152___
2153$code.=<<___ if ($addx);
2154	mov	OPENSSL_ia32cap_P+8(%rip),%r11d
2155	and	\$0x80108,%r11d
2156	cmp	\$0x80108,%r11d		# check for AD*X+BMI2+BMI1
2157	jne	.Lfrom_mont_nox
2158
2159	lea	(%rax,$num),$rptr
2160	call	__bn_sqrx8x_reduction
2161	call	__bn_postx4x_internal
2162
2163	pxor	%xmm0,%xmm0
2164	lea	48(%rsp),%rax
2165	mov	40(%rsp),%rsi		# restore %rsp
2166	jmp	.Lfrom_mont_zero
2167
2168.align	32
2169.Lfrom_mont_nox:
2170___
2171$code.=<<___;
2172	call	__bn_sqr8x_reduction
2173	call	__bn_post4x_internal
2174
2175	pxor	%xmm0,%xmm0
2176	lea	48(%rsp),%rax
2177	mov	40(%rsp),%rsi		# restore %rsp
2178	jmp	.Lfrom_mont_zero
2179
2180.align	32
2181.Lfrom_mont_zero:
2182	movdqa	%xmm0,16*0(%rax)
2183	movdqa	%xmm0,16*1(%rax)
2184	movdqa	%xmm0,16*2(%rax)
2185	movdqa	%xmm0,16*3(%rax)
2186	lea	16*4(%rax),%rax
2187	sub	\$32,$num
2188	jnz	.Lfrom_mont_zero
2189
2190	mov	\$1,%rax
2191	mov	-48(%rsi),%r15
2192	mov	-40(%rsi),%r14
2193	mov	-32(%rsi),%r13
2194	mov	-24(%rsi),%r12
2195	mov	-16(%rsi),%rbp
2196	mov	-8(%rsi),%rbx
2197	lea	(%rsi),%rsp
2198.Lfrom_epilogue:
2199	ret
2200.size	bn_from_mont8x,.-bn_from_mont8x
2201___
2202}
2203}}}
2204
2205if ($addx) {{{
2206my $bp="%rdx";	# restore original value
2207
2208$code.=<<___;
2209.type	bn_mulx4x_mont_gather5,\@function,6
2210.align	32
2211bn_mulx4x_mont_gather5:
2212	mov	%rsp,%rax
2213.Lmulx4x_enter:
2214	push	%rbx
2215	push	%rbp
2216	push	%r12
2217	push	%r13
2218	push	%r14
2219	push	%r15
2220.Lmulx4x_prologue:
2221
2222	shl	\$3,${num}d		# convert $num to bytes
2223	lea	($num,$num,2),%r10	# 3*$num in bytes
2224	neg	$num			# -$num
2225	mov	($n0),$n0		# *n0
2226
2227	##############################################################
2228	# Ensure that stack frame doesn't alias with $rptr+3*$num
2229	# modulo 4096, which covers ret[num], am[num] and n[num]
2230	# (see bn_exp.c). This is done to allow memory disambiguation
2231	# logic do its magic. [Extra [num] is allocated in order
2232	# to align with bn_power5's frame, which is cleansed after
2233	# completing exponentiation. Extra 256 bytes is for power mask
2234	# calculated from 7th argument, the index.]
2235	#
2236	lea	-320(%rsp,$num,2),%r11
2237	mov	%rsp,%rbp
2238	sub	$rp,%r11
2239	and	\$4095,%r11
2240	cmp	%r11,%r10
2241	jb	.Lmulx4xsp_alt
2242	sub	%r11,%rbp		# align with $aptr
2243	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2244	jmp	.Lmulx4xsp_done
2245
2246.Lmulx4xsp_alt:
2247	lea	4096-320(,$num,2),%r10
2248	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2249	sub	%r10,%r11
2250	mov	\$0,%r10
2251	cmovc	%r10,%r11
2252	sub	%r11,%rbp
2253.Lmulx4xsp_done:
2254	and	\$-64,%rbp		# ensure alignment
2255	mov	%rsp,%r11
2256	sub	%rbp,%r11
2257	and	\$-4096,%r11
2258	lea	(%rbp,%r11),%rsp
2259	mov	(%rsp),%r10
2260	cmp	%rbp,%rsp
2261	ja	.Lmulx4x_page_walk
2262	jmp	.Lmulx4x_page_walk_done
2263
2264.Lmulx4x_page_walk:
2265	lea	-4096(%rsp),%rsp
2266	mov	(%rsp),%r10
2267	cmp	%rbp,%rsp
2268	ja	.Lmulx4x_page_walk
2269.Lmulx4x_page_walk_done:
2270
2271	##############################################################
2272	# Stack layout
2273	# +0	-num
2274	# +8	off-loaded &b[i]
2275	# +16	end of b[num]
2276	# +24	inner counter
2277	# +32	saved n0
2278	# +40	saved %rsp
2279	# +48
2280	# +56	saved rp
2281	# +64	tmp[num+1]
2282	#
2283	mov	$n0, 32(%rsp)		# save *n0
2284	mov	%rax,40(%rsp)		# save original %rsp
2285.Lmulx4x_body:
2286	call	mulx4x_internal
2287
2288	mov	40(%rsp),%rsi		# restore %rsp
2289	mov	\$1,%rax
2290
2291	mov	-48(%rsi),%r15
2292	mov	-40(%rsi),%r14
2293	mov	-32(%rsi),%r13
2294	mov	-24(%rsi),%r12
2295	mov	-16(%rsi),%rbp
2296	mov	-8(%rsi),%rbx
2297	lea	(%rsi),%rsp
2298.Lmulx4x_epilogue:
2299	ret
2300.size	bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5
2301
2302.type	mulx4x_internal,\@abi-omnipotent
2303.align	32
2304mulx4x_internal:
2305	mov	$num,8(%rsp)		# save -$num (it was in bytes)
2306	mov	$num,%r10
2307	neg	$num			# restore $num
2308	shl	\$5,$num
2309	neg	%r10			# restore $num
2310	lea	128($bp,$num),%r13	# end of powers table (+size optimization)
2311	shr	\$5+5,$num
2312	movd	`($win64?56:8)`(%rax),%xmm5	# load 7th argument
2313	sub	\$1,$num
2314	lea	.Linc(%rip),%rax
2315	mov	%r13,16+8(%rsp)		# end of b[num]
2316	mov	$num,24+8(%rsp)		# inner counter
2317	mov	$rp, 56+8(%rsp)		# save $rp
2318___
2319my ($aptr, $bptr, $nptr, $tptr, $mi,  $bi,  $zero, $num)=
2320   ("%rsi","%rdi","%rcx","%rbx","%r8","%r9","%rbp","%rax");
2321my $rptr=$bptr;
2322my $STRIDE=2**5*8;		# 5 is "window size"
2323my $N=$STRIDE/4;		# should match cache line size
2324$code.=<<___;
2325	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
2326	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
2327	lea	88-112(%rsp,%r10),%r10	# place the mask after tp[num+1] (+ICache optimizaton)
2328	lea	128($bp),$bptr		# size optimization
2329
2330	pshufd	\$0,%xmm5,%xmm5		# broadcast index
2331	movdqa	%xmm1,%xmm4
2332	.byte	0x67
2333	movdqa	%xmm1,%xmm2
2334___
2335########################################################################
2336# calculate mask by comparing 0..31 to index and save result to stack
2337#
2338$code.=<<___;
2339	.byte	0x67
2340	paddd	%xmm0,%xmm1
2341	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
2342	movdqa	%xmm4,%xmm3
2343___
2344for($i=0;$i<$STRIDE/16-4;$i+=4) {
2345$code.=<<___;
2346	paddd	%xmm1,%xmm2
2347	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
2348	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2349	movdqa	%xmm4,%xmm0
2350
2351	paddd	%xmm2,%xmm3
2352	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
2353	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2354	movdqa	%xmm4,%xmm1
2355
2356	paddd	%xmm3,%xmm0
2357	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
2358	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2359	movdqa	%xmm4,%xmm2
2360
2361	paddd	%xmm0,%xmm1
2362	pcmpeqd	%xmm5,%xmm0
2363	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2364	movdqa	%xmm4,%xmm3
2365___
2366}
2367$code.=<<___;				# last iteration can be optimized
2368	.byte	0x67
2369	paddd	%xmm1,%xmm2
2370	pcmpeqd	%xmm5,%xmm1
2371	movdqa	%xmm0,`16*($i+0)+112`(%r10)
2372
2373	paddd	%xmm2,%xmm3
2374	pcmpeqd	%xmm5,%xmm2
2375	movdqa	%xmm1,`16*($i+1)+112`(%r10)
2376
2377	pcmpeqd	%xmm5,%xmm3
2378	movdqa	%xmm2,`16*($i+2)+112`(%r10)
2379
2380	pand	`16*($i+0)-128`($bptr),%xmm0	# while it's still in register
2381	pand	`16*($i+1)-128`($bptr),%xmm1
2382	pand	`16*($i+2)-128`($bptr),%xmm2
2383	movdqa	%xmm3,`16*($i+3)+112`(%r10)
2384	pand	`16*($i+3)-128`($bptr),%xmm3
2385	por	%xmm2,%xmm0
2386	por	%xmm3,%xmm1
2387___
2388for($i=0;$i<$STRIDE/16-4;$i+=4) {
2389$code.=<<___;
2390	movdqa	`16*($i+0)-128`($bptr),%xmm4
2391	movdqa	`16*($i+1)-128`($bptr),%xmm5
2392	movdqa	`16*($i+2)-128`($bptr),%xmm2
2393	pand	`16*($i+0)+112`(%r10),%xmm4
2394	movdqa	`16*($i+3)-128`($bptr),%xmm3
2395	pand	`16*($i+1)+112`(%r10),%xmm5
2396	por	%xmm4,%xmm0
2397	pand	`16*($i+2)+112`(%r10),%xmm2
2398	por	%xmm5,%xmm1
2399	pand	`16*($i+3)+112`(%r10),%xmm3
2400	por	%xmm2,%xmm0
2401	por	%xmm3,%xmm1
2402___
2403}
2404$code.=<<___;
2405	pxor	%xmm1,%xmm0
2406	pshufd	\$0x4e,%xmm0,%xmm1
2407	por	%xmm1,%xmm0
2408	lea	$STRIDE($bptr),$bptr
2409	movq	%xmm0,%rdx		# bp[0]
2410	lea	64+8*4+8(%rsp),$tptr
2411
2412	mov	%rdx,$bi
2413	mulx	0*8($aptr),$mi,%rax	# a[0]*b[0]
2414	mulx	1*8($aptr),%r11,%r12	# a[1]*b[0]
2415	add	%rax,%r11
2416	mulx	2*8($aptr),%rax,%r13	# ...
2417	adc	%rax,%r12
2418	adc	\$0,%r13
2419	mulx	3*8($aptr),%rax,%r14
2420
2421	mov	$mi,%r15
2422	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2423	xor	$zero,$zero		# cf=0, of=0
2424	mov	$mi,%rdx
2425
2426	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2427
2428	lea	4*8($aptr),$aptr
2429	adcx	%rax,%r13
2430	adcx	$zero,%r14		# cf=0
2431
2432	mulx	0*8($nptr),%rax,%r10
2433	adcx	%rax,%r15		# discarded
2434	adox	%r11,%r10
2435	mulx	1*8($nptr),%rax,%r11
2436	adcx	%rax,%r10
2437	adox	%r12,%r11
2438	mulx	2*8($nptr),%rax,%r12
2439	mov	24+8(%rsp),$bptr	# counter value
2440	mov	%r10,-8*4($tptr)
2441	adcx	%rax,%r11
2442	adox	%r13,%r12
2443	mulx	3*8($nptr),%rax,%r15
2444	 mov	$bi,%rdx
2445	mov	%r11,-8*3($tptr)
2446	adcx	%rax,%r12
2447	adox	$zero,%r15		# of=0
2448	lea	4*8($nptr),$nptr
2449	mov	%r12,-8*2($tptr)
2450	jmp	.Lmulx4x_1st
2451
2452.align	32
2453.Lmulx4x_1st:
2454	adcx	$zero,%r15		# cf=0, modulo-scheduled
2455	mulx	0*8($aptr),%r10,%rax	# a[4]*b[0]
2456	adcx	%r14,%r10
2457	mulx	1*8($aptr),%r11,%r14	# a[5]*b[0]
2458	adcx	%rax,%r11
2459	mulx	2*8($aptr),%r12,%rax	# ...
2460	adcx	%r14,%r12
2461	mulx	3*8($aptr),%r13,%r14
2462	 .byte	0x67,0x67
2463	 mov	$mi,%rdx
2464	adcx	%rax,%r13
2465	adcx	$zero,%r14		# cf=0
2466	lea	4*8($aptr),$aptr
2467	lea	4*8($tptr),$tptr
2468
2469	adox	%r15,%r10
2470	mulx	0*8($nptr),%rax,%r15
2471	adcx	%rax,%r10
2472	adox	%r15,%r11
2473	mulx	1*8($nptr),%rax,%r15
2474	adcx	%rax,%r11
2475	adox	%r15,%r12
2476	mulx	2*8($nptr),%rax,%r15
2477	mov	%r10,-5*8($tptr)
2478	adcx	%rax,%r12
2479	mov	%r11,-4*8($tptr)
2480	adox	%r15,%r13
2481	mulx	3*8($nptr),%rax,%r15
2482	 mov	$bi,%rdx
2483	mov	%r12,-3*8($tptr)
2484	adcx	%rax,%r13
2485	adox	$zero,%r15
2486	lea	4*8($nptr),$nptr
2487	mov	%r13,-2*8($tptr)
2488
2489	dec	$bptr			# of=0, pass cf
2490	jnz	.Lmulx4x_1st
2491
2492	mov	8(%rsp),$num		# load -num
2493	adc	$zero,%r15		# modulo-scheduled
2494	lea	($aptr,$num),$aptr	# rewind $aptr
2495	add	%r15,%r14
2496	mov	8+8(%rsp),$bptr		# re-load &b[i]
2497	adc	$zero,$zero		# top-most carry
2498	mov	%r14,-1*8($tptr)
2499	jmp	.Lmulx4x_outer
2500
2501.align	32
2502.Lmulx4x_outer:
2503	lea	16-256($tptr),%r10	# where 256-byte mask is (+density control)
2504	pxor	%xmm4,%xmm4
2505	.byte	0x67,0x67
2506	pxor	%xmm5,%xmm5
2507___
2508for($i=0;$i<$STRIDE/16;$i+=4) {
2509$code.=<<___;
2510	movdqa	`16*($i+0)-128`($bptr),%xmm0
2511	movdqa	`16*($i+1)-128`($bptr),%xmm1
2512	movdqa	`16*($i+2)-128`($bptr),%xmm2
2513	pand	`16*($i+0)+256`(%r10),%xmm0
2514	movdqa	`16*($i+3)-128`($bptr),%xmm3
2515	pand	`16*($i+1)+256`(%r10),%xmm1
2516	por	%xmm0,%xmm4
2517	pand	`16*($i+2)+256`(%r10),%xmm2
2518	por	%xmm1,%xmm5
2519	pand	`16*($i+3)+256`(%r10),%xmm3
2520	por	%xmm2,%xmm4
2521	por	%xmm3,%xmm5
2522___
2523}
2524$code.=<<___;
2525	por	%xmm5,%xmm4
2526	pshufd	\$0x4e,%xmm4,%xmm0
2527	por	%xmm4,%xmm0
2528	lea	$STRIDE($bptr),$bptr
2529	movq	%xmm0,%rdx		# m0=bp[i]
2530
2531	mov	$zero,($tptr)		# save top-most carry
2532	lea	4*8($tptr,$num),$tptr	# rewind $tptr
2533	mulx	0*8($aptr),$mi,%r11	# a[0]*b[i]
2534	xor	$zero,$zero		# cf=0, of=0
2535	mov	%rdx,$bi
2536	mulx	1*8($aptr),%r14,%r12	# a[1]*b[i]
2537	adox	-4*8($tptr),$mi		# +t[0]
2538	adcx	%r14,%r11
2539	mulx	2*8($aptr),%r15,%r13	# ...
2540	adox	-3*8($tptr),%r11
2541	adcx	%r15,%r12
2542	mulx	3*8($aptr),%rdx,%r14
2543	adox	-2*8($tptr),%r12
2544	adcx	%rdx,%r13
2545	lea	($nptr,$num),$nptr	# rewind $nptr
2546	lea	4*8($aptr),$aptr
2547	adox	-1*8($tptr),%r13
2548	adcx	$zero,%r14
2549	adox	$zero,%r14
2550
2551	mov	$mi,%r15
2552	imulq	32+8(%rsp),$mi		# "t[0]"*n0
2553
2554	mov	$mi,%rdx
2555	xor	$zero,$zero		# cf=0, of=0
2556	mov	$bptr,8+8(%rsp)		# off-load &b[i]
2557
2558	mulx	0*8($nptr),%rax,%r10
2559	adcx	%rax,%r15		# discarded
2560	adox	%r11,%r10
2561	mulx	1*8($nptr),%rax,%r11
2562	adcx	%rax,%r10
2563	adox	%r12,%r11
2564	mulx	2*8($nptr),%rax,%r12
2565	adcx	%rax,%r11
2566	adox	%r13,%r12
2567	mulx	3*8($nptr),%rax,%r15
2568	 mov	$bi,%rdx
2569	mov	24+8(%rsp),$bptr	# counter value
2570	mov	%r10,-8*4($tptr)
2571	adcx	%rax,%r12
2572	mov	%r11,-8*3($tptr)
2573	adox	$zero,%r15		# of=0
2574	mov	%r12,-8*2($tptr)
2575	lea	4*8($nptr),$nptr
2576	jmp	.Lmulx4x_inner
2577
2578.align	32
2579.Lmulx4x_inner:
2580	mulx	0*8($aptr),%r10,%rax	# a[4]*b[i]
2581	adcx	$zero,%r15		# cf=0, modulo-scheduled
2582	adox	%r14,%r10
2583	mulx	1*8($aptr),%r11,%r14	# a[5]*b[i]
2584	adcx	0*8($tptr),%r10
2585	adox	%rax,%r11
2586	mulx	2*8($aptr),%r12,%rax	# ...
2587	adcx	1*8($tptr),%r11
2588	adox	%r14,%r12
2589	mulx	3*8($aptr),%r13,%r14
2590	 mov	$mi,%rdx
2591	adcx	2*8($tptr),%r12
2592	adox	%rax,%r13
2593	adcx	3*8($tptr),%r13
2594	adox	$zero,%r14		# of=0
2595	lea	4*8($aptr),$aptr
2596	lea	4*8($tptr),$tptr
2597	adcx	$zero,%r14		# cf=0
2598
2599	adox	%r15,%r10
2600	mulx	0*8($nptr),%rax,%r15
2601	adcx	%rax,%r10
2602	adox	%r15,%r11
2603	mulx	1*8($nptr),%rax,%r15
2604	adcx	%rax,%r11
2605	adox	%r15,%r12
2606	mulx	2*8($nptr),%rax,%r15
2607	mov	%r10,-5*8($tptr)
2608	adcx	%rax,%r12
2609	adox	%r15,%r13
2610	mov	%r11,-4*8($tptr)
2611	mulx	3*8($nptr),%rax,%r15
2612	 mov	$bi,%rdx
2613	lea	4*8($nptr),$nptr
2614	mov	%r12,-3*8($tptr)
2615	adcx	%rax,%r13
2616	adox	$zero,%r15
2617	mov	%r13,-2*8($tptr)
2618
2619	dec	$bptr			# of=0, pass cf
2620	jnz	.Lmulx4x_inner
2621
2622	mov	0+8(%rsp),$num		# load -num
2623	adc	$zero,%r15		# modulo-scheduled
2624	sub	0*8($tptr),$bptr	# pull top-most carry to %cf
2625	mov	8+8(%rsp),$bptr		# re-load &b[i]
2626	mov	16+8(%rsp),%r10
2627	adc	%r15,%r14
2628	lea	($aptr,$num),$aptr	# rewind $aptr
2629	adc	$zero,$zero		# top-most carry
2630	mov	%r14,-1*8($tptr)
2631
2632	cmp	%r10,$bptr
2633	jb	.Lmulx4x_outer
2634
2635	mov	-8($nptr),%r10
2636	mov	$zero,%r8
2637	mov	($nptr,$num),%r12
2638	lea	($nptr,$num),%rbp	# rewind $nptr
2639	mov	$num,%rcx
2640	lea	($tptr,$num),%rdi	# rewind $tptr
2641	xor	%eax,%eax
2642	xor	%r15,%r15
2643	sub	%r14,%r10		# compare top-most words
2644	adc	%r15,%r15
2645	or	%r15,%r8
2646	sar	\$3+2,%rcx
2647	sub	%r8,%rax		# %rax=-%r8
2648	mov	56+8(%rsp),%rdx		# restore rp
2649	dec	%r12			# so that after 'not' we get -n[0]
2650	mov	8*1(%rbp),%r13
2651	xor	%r8,%r8
2652	mov	8*2(%rbp),%r14
2653	mov	8*3(%rbp),%r15
2654	jmp	.Lsqrx4x_sub_entry	# common post-condition
2655.size	mulx4x_internal,.-mulx4x_internal
2656___
2657}{
2658######################################################################
2659# void bn_power5(
2660my $rptr="%rdi";	# BN_ULONG *rptr,
2661my $aptr="%rsi";	# const BN_ULONG *aptr,
2662my $bptr="%rdx";	# const void *table,
2663my $nptr="%rcx";	# const BN_ULONG *nptr,
2664my $n0  ="%r8";		# const BN_ULONG *n0);
2665my $num ="%r9";		# int num, has to be divisible by 8
2666			# int pwr);
2667
2668my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
2669my @A0=("%r10","%r11");
2670my @A1=("%r12","%r13");
2671my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
2672
2673$code.=<<___;
2674.type	bn_powerx5,\@function,6
2675.align	32
2676bn_powerx5:
2677	mov	%rsp,%rax
2678.Lpowerx5_enter:
2679	push	%rbx
2680	push	%rbp
2681	push	%r12
2682	push	%r13
2683	push	%r14
2684	push	%r15
2685.Lpowerx5_prologue:
2686
2687	shl	\$3,${num}d		# convert $num to bytes
2688	lea	($num,$num,2),%r10	# 3*$num in bytes
2689	neg	$num
2690	mov	($n0),$n0		# *n0
2691
2692	##############################################################
2693	# Ensure that stack frame doesn't alias with $rptr+3*$num
2694	# modulo 4096, which covers ret[num], am[num] and n[num]
2695	# (see bn_exp.c). This is done to allow memory disambiguation
2696	# logic do its magic. [Extra 256 bytes is for power mask
2697	# calculated from 7th argument, the index.]
2698	#
2699	lea	-320(%rsp,$num,2),%r11
2700	mov	%rsp,%rbp
2701	sub	$rptr,%r11
2702	and	\$4095,%r11
2703	cmp	%r11,%r10
2704	jb	.Lpwrx_sp_alt
2705	sub	%r11,%rbp		# align with $aptr
2706	lea	-320(%rbp,$num,2),%rbp	# future alloca(frame+2*$num*8+256)
2707	jmp	.Lpwrx_sp_done
2708
2709.align	32
2710.Lpwrx_sp_alt:
2711	lea	4096-320(,$num,2),%r10
2712	lea	-320(%rbp,$num,2),%rbp	# alloca(frame+2*$num*8+256)
2713	sub	%r10,%r11
2714	mov	\$0,%r10
2715	cmovc	%r10,%r11
2716	sub	%r11,%rbp
2717.Lpwrx_sp_done:
2718	and	\$-64,%rbp
2719	mov	%rsp,%r11
2720	sub	%rbp,%r11
2721	and	\$-4096,%r11
2722	lea	(%rbp,%r11),%rsp
2723	mov	(%rsp),%r10
2724	cmp	%rbp,%rsp
2725	ja	.Lpwrx_page_walk
2726	jmp	.Lpwrx_page_walk_done
2727
2728.Lpwrx_page_walk:
2729	lea	-4096(%rsp),%rsp
2730	mov	(%rsp),%r10
2731	cmp	%rbp,%rsp
2732	ja	.Lpwrx_page_walk
2733.Lpwrx_page_walk_done:
2734
2735	mov	$num,%r10
2736	neg	$num
2737
2738	##############################################################
2739	# Stack layout
2740	#
2741	# +0	saved $num, used in reduction section
2742	# +8	&t[2*$num], used in reduction section
2743	# +16	intermediate carry bit
2744	# +24	top-most carry bit, used in reduction section
2745	# +32	saved *n0
2746	# +40	saved %rsp
2747	# +48	t[2*$num]
2748	#
2749	pxor	%xmm0,%xmm0
2750	movq	$rptr,%xmm1		# save $rptr
2751	movq	$nptr,%xmm2		# save $nptr
2752	movq	%r10, %xmm3		# -$num
2753	movq	$bptr,%xmm4
2754	mov	$n0,  32(%rsp)
2755	mov	%rax, 40(%rsp)		# save original %rsp
2756.Lpowerx5_body:
2757
2758	call	__bn_sqrx8x_internal
2759	call	__bn_postx4x_internal
2760	call	__bn_sqrx8x_internal
2761	call	__bn_postx4x_internal
2762	call	__bn_sqrx8x_internal
2763	call	__bn_postx4x_internal
2764	call	__bn_sqrx8x_internal
2765	call	__bn_postx4x_internal
2766	call	__bn_sqrx8x_internal
2767	call	__bn_postx4x_internal
2768
2769	mov	%r10,$num		# -num
2770	mov	$aptr,$rptr
2771	movq	%xmm2,$nptr
2772	movq	%xmm4,$bptr
2773	mov	40(%rsp),%rax
2774
2775	call	mulx4x_internal
2776
2777	mov	40(%rsp),%rsi		# restore %rsp
2778	mov	\$1,%rax
2779
2780	mov	-48(%rsi),%r15
2781	mov	-40(%rsi),%r14
2782	mov	-32(%rsi),%r13
2783	mov	-24(%rsi),%r12
2784	mov	-16(%rsi),%rbp
2785	mov	-8(%rsi),%rbx
2786	lea	(%rsi),%rsp
2787.Lpowerx5_epilogue:
2788	ret
2789.size	bn_powerx5,.-bn_powerx5
2790
2791.globl	bn_sqrx8x_internal
2792.hidden	bn_sqrx8x_internal
2793.type	bn_sqrx8x_internal,\@abi-omnipotent
2794.align	32
2795bn_sqrx8x_internal:
2796__bn_sqrx8x_internal:
2797	##################################################################
2798	# Squaring part:
2799	#
2800	# a) multiply-n-add everything but a[i]*a[i];
2801	# b) shift result of a) by 1 to the left and accumulate
2802	#    a[i]*a[i] products;
2803	#
2804	##################################################################
2805	# 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]
2806	#                                                     a[1]a[0]
2807	#                                                 a[2]a[0]
2808	#                                             a[3]a[0]
2809	#                                             a[2]a[1]
2810	#                                         a[3]a[1]
2811	#                                     a[3]a[2]
2812	#
2813	#                                         a[4]a[0]
2814	#                                     a[5]a[0]
2815	#                                 a[6]a[0]
2816	#                             a[7]a[0]
2817	#                                     a[4]a[1]
2818	#                                 a[5]a[1]
2819	#                             a[6]a[1]
2820	#                         a[7]a[1]
2821	#                                 a[4]a[2]
2822	#                             a[5]a[2]
2823	#                         a[6]a[2]
2824	#                     a[7]a[2]
2825	#                             a[4]a[3]
2826	#                         a[5]a[3]
2827	#                     a[6]a[3]
2828	#                 a[7]a[3]
2829	#
2830	#                     a[5]a[4]
2831	#                 a[6]a[4]
2832	#             a[7]a[4]
2833	#             a[6]a[5]
2834	#         a[7]a[5]
2835	#     a[7]a[6]
2836	# 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]
2837___
2838{
2839my ($zero,$carry)=("%rbp","%rcx");
2840my $aaptr=$zero;
2841$code.=<<___;
2842	lea	48+8(%rsp),$tptr
2843	lea	($aptr,$num),$aaptr
2844	mov	$num,0+8(%rsp)			# save $num
2845	mov	$aaptr,8+8(%rsp)		# save end of $aptr
2846	jmp	.Lsqr8x_zero_start
2847
2848.align	32
2849.byte	0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
2850.Lsqrx8x_zero:
2851	.byte	0x3e
2852	movdqa	%xmm0,0*8($tptr)
2853	movdqa	%xmm0,2*8($tptr)
2854	movdqa	%xmm0,4*8($tptr)
2855	movdqa	%xmm0,6*8($tptr)
2856.Lsqr8x_zero_start:			# aligned at 32
2857	movdqa	%xmm0,8*8($tptr)
2858	movdqa	%xmm0,10*8($tptr)
2859	movdqa	%xmm0,12*8($tptr)
2860	movdqa	%xmm0,14*8($tptr)
2861	lea	16*8($tptr),$tptr
2862	sub	\$64,$num
2863	jnz	.Lsqrx8x_zero
2864
2865	mov	0*8($aptr),%rdx		# a[0], modulo-scheduled
2866	#xor	%r9,%r9			# t[1], ex-$num, zero already
2867	xor	%r10,%r10
2868	xor	%r11,%r11
2869	xor	%r12,%r12
2870	xor	%r13,%r13
2871	xor	%r14,%r14
2872	xor	%r15,%r15
2873	lea	48+8(%rsp),$tptr
2874	xor	$zero,$zero		# cf=0, cf=0
2875	jmp	.Lsqrx8x_outer_loop
2876
2877.align	32
2878.Lsqrx8x_outer_loop:
2879	mulx	1*8($aptr),%r8,%rax	# a[1]*a[0]
2880	adcx	%r9,%r8			# a[1]*a[0]+=t[1]
2881	adox	%rax,%r10
2882	mulx	2*8($aptr),%r9,%rax	# a[2]*a[0]
2883	adcx	%r10,%r9
2884	adox	%rax,%r11
2885	.byte	0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00	# mulx	3*8($aptr),%r10,%rax	# ...
2886	adcx	%r11,%r10
2887	adox	%rax,%r12
2888	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00	# mulx	4*8($aptr),%r11,%rax
2889	adcx	%r12,%r11
2890	adox	%rax,%r13
2891	mulx	5*8($aptr),%r12,%rax
2892	adcx	%r13,%r12
2893	adox	%rax,%r14
2894	mulx	6*8($aptr),%r13,%rax
2895	adcx	%r14,%r13
2896	adox	%r15,%rax
2897	mulx	7*8($aptr),%r14,%r15
2898	 mov	1*8($aptr),%rdx		# a[1]
2899	adcx	%rax,%r14
2900	adox	$zero,%r15
2901	adc	8*8($tptr),%r15
2902	mov	%r8,1*8($tptr)		# t[1]
2903	mov	%r9,2*8($tptr)		# t[2]
2904	sbb	$carry,$carry		# mov %cf,$carry
2905	xor	$zero,$zero		# cf=0, of=0
2906
2907
2908	mulx	2*8($aptr),%r8,%rbx	# a[2]*a[1]
2909	mulx	3*8($aptr),%r9,%rax	# a[3]*a[1]
2910	adcx	%r10,%r8
2911	adox	%rbx,%r9
2912	mulx	4*8($aptr),%r10,%rbx	# ...
2913	adcx	%r11,%r9
2914	adox	%rax,%r10
2915	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00	# mulx	5*8($aptr),%r11,%rax
2916	adcx	%r12,%r10
2917	adox	%rbx,%r11
2918	.byte	0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r12,%rbx
2919	adcx	%r13,%r11
2920	adox	%r14,%r12
2921	.byte	0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r13,%r14
2922	 mov	2*8($aptr),%rdx		# a[2]
2923	adcx	%rax,%r12
2924	adox	%rbx,%r13
2925	adcx	%r15,%r13
2926	adox	$zero,%r14		# of=0
2927	adcx	$zero,%r14		# cf=0
2928
2929	mov	%r8,3*8($tptr)		# t[3]
2930	mov	%r9,4*8($tptr)		# t[4]
2931
2932	mulx	3*8($aptr),%r8,%rbx	# a[3]*a[2]
2933	mulx	4*8($aptr),%r9,%rax	# a[4]*a[2]
2934	adcx	%r10,%r8
2935	adox	%rbx,%r9
2936	mulx	5*8($aptr),%r10,%rbx	# ...
2937	adcx	%r11,%r9
2938	adox	%rax,%r10
2939	.byte	0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00	# mulx	6*8($aptr),%r11,%rax
2940	adcx	%r12,%r10
2941	adox	%r13,%r11
2942	.byte	0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00	# mulx	7*8($aptr),%r12,%r13
2943	.byte	0x3e
2944	 mov	3*8($aptr),%rdx		# a[3]
2945	adcx	%rbx,%r11
2946	adox	%rax,%r12
2947	adcx	%r14,%r12
2948	mov	%r8,5*8($tptr)		# t[5]
2949	mov	%r9,6*8($tptr)		# t[6]
2950	 mulx	4*8($aptr),%r8,%rax	# a[4]*a[3]
2951	adox	$zero,%r13		# of=0
2952	adcx	$zero,%r13		# cf=0
2953
2954	mulx	5*8($aptr),%r9,%rbx	# a[5]*a[3]
2955	adcx	%r10,%r8
2956	adox	%rax,%r9
2957	mulx	6*8($aptr),%r10,%rax	# ...
2958	adcx	%r11,%r9
2959	adox	%r12,%r10
2960	mulx	7*8($aptr),%r11,%r12
2961	 mov	4*8($aptr),%rdx		# a[4]
2962	 mov	5*8($aptr),%r14		# a[5]
2963	adcx	%rbx,%r10
2964	adox	%rax,%r11
2965	 mov	6*8($aptr),%r15		# a[6]
2966	adcx	%r13,%r11
2967	adox	$zero,%r12		# of=0
2968	adcx	$zero,%r12		# cf=0
2969
2970	mov	%r8,7*8($tptr)		# t[7]
2971	mov	%r9,8*8($tptr)		# t[8]
2972
2973	mulx	%r14,%r9,%rax		# a[5]*a[4]
2974	 mov	7*8($aptr),%r8		# a[7]
2975	adcx	%r10,%r9
2976	mulx	%r15,%r10,%rbx		# a[6]*a[4]
2977	adox	%rax,%r10
2978	adcx	%r11,%r10
2979	mulx	%r8,%r11,%rax		# a[7]*a[4]
2980	 mov	%r14,%rdx		# a[5]
2981	adox	%rbx,%r11
2982	adcx	%r12,%r11
2983	#adox	$zero,%rax		# of=0
2984	adcx	$zero,%rax		# cf=0
2985
2986	mulx	%r15,%r14,%rbx		# a[6]*a[5]
2987	mulx	%r8,%r12,%r13		# a[7]*a[5]
2988	 mov	%r15,%rdx		# a[6]
2989	 lea	8*8($aptr),$aptr
2990	adcx	%r14,%r11
2991	adox	%rbx,%r12
2992	adcx	%rax,%r12
2993	adox	$zero,%r13
2994
2995	.byte	0x67,0x67
2996	mulx	%r8,%r8,%r14		# a[7]*a[6]
2997	adcx	%r8,%r13
2998	adcx	$zero,%r14
2999
3000	cmp	8+8(%rsp),$aptr
3001	je	.Lsqrx8x_outer_break
3002
3003	neg	$carry			# mov $carry,%cf
3004	mov	\$-8,%rcx
3005	mov	$zero,%r15
3006	mov	8*8($tptr),%r8
3007	adcx	9*8($tptr),%r9		# +=t[9]
3008	adcx	10*8($tptr),%r10	# ...
3009	adcx	11*8($tptr),%r11
3010	adc	12*8($tptr),%r12
3011	adc	13*8($tptr),%r13
3012	adc	14*8($tptr),%r14
3013	adc	15*8($tptr),%r15
3014	lea	($aptr),$aaptr
3015	lea	2*64($tptr),$tptr
3016	sbb	%rax,%rax		# mov %cf,$carry
3017
3018	mov	-64($aptr),%rdx		# a[0]
3019	mov	%rax,16+8(%rsp)		# offload $carry
3020	mov	$tptr,24+8(%rsp)
3021
3022	#lea	8*8($tptr),$tptr	# see 2*8*8($tptr) above
3023	xor	%eax,%eax		# cf=0, of=0
3024	jmp	.Lsqrx8x_loop
3025
3026.align	32
3027.Lsqrx8x_loop:
3028	mov	%r8,%rbx
3029	mulx	0*8($aaptr),%rax,%r8	# a[8]*a[i]
3030	adcx	%rax,%rbx		# +=t[8]
3031	adox	%r9,%r8
3032
3033	mulx	1*8($aaptr),%rax,%r9	# ...
3034	adcx	%rax,%r8
3035	adox	%r10,%r9
3036
3037	mulx	2*8($aaptr),%rax,%r10
3038	adcx	%rax,%r9
3039	adox	%r11,%r10
3040
3041	mulx	3*8($aaptr),%rax,%r11
3042	adcx	%rax,%r10
3043	adox	%r12,%r11
3044
3045	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	4*8($aaptr),%rax,%r12
3046	adcx	%rax,%r11
3047	adox	%r13,%r12
3048
3049	mulx	5*8($aaptr),%rax,%r13
3050	adcx	%rax,%r12
3051	adox	%r14,%r13
3052
3053	mulx	6*8($aaptr),%rax,%r14
3054	 mov	%rbx,($tptr,%rcx,8)	# store t[8+i]
3055	 mov	\$0,%ebx
3056	adcx	%rax,%r13
3057	adox	%r15,%r14
3058
3059	.byte	0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00	# mulx	7*8($aaptr),%rax,%r15
3060	 mov	8($aptr,%rcx,8),%rdx	# a[i]
3061	adcx	%rax,%r14
3062	adox	%rbx,%r15		# %rbx is 0, of=0
3063	adcx	%rbx,%r15		# cf=0
3064
3065	.byte	0x67
3066	inc	%rcx			# of=0
3067	jnz	.Lsqrx8x_loop
3068
3069	lea	8*8($aaptr),$aaptr
3070	mov	\$-8,%rcx
3071	cmp	8+8(%rsp),$aaptr	# done?
3072	je	.Lsqrx8x_break
3073
3074	sub	16+8(%rsp),%rbx		# mov 16(%rsp),%cf
3075	.byte	0x66
3076	mov	-64($aptr),%rdx
3077	adcx	0*8($tptr),%r8
3078	adcx	1*8($tptr),%r9
3079	adc	2*8($tptr),%r10
3080	adc	3*8($tptr),%r11
3081	adc	4*8($tptr),%r12
3082	adc	5*8($tptr),%r13
3083	adc	6*8($tptr),%r14
3084	adc	7*8($tptr),%r15
3085	lea	8*8($tptr),$tptr
3086	.byte	0x67
3087	sbb	%rax,%rax		# mov %cf,%rax
3088	xor	%ebx,%ebx		# cf=0, of=0
3089	mov	%rax,16+8(%rsp)		# offload carry
3090	jmp	.Lsqrx8x_loop
3091
3092.align	32
3093.Lsqrx8x_break:
3094	sub	16+8(%rsp),%r8		# consume last carry
3095	mov	24+8(%rsp),$carry	# initial $tptr, borrow $carry
3096	mov	0*8($aptr),%rdx		# a[8], modulo-scheduled
3097	xor	%ebp,%ebp		# xor	$zero,$zero
3098	mov	%r8,0*8($tptr)
3099	cmp	$carry,$tptr		# cf=0, of=0
3100	je	.Lsqrx8x_outer_loop
3101
3102	mov	%r9,1*8($tptr)
3103	 mov	1*8($carry),%r9
3104	mov	%r10,2*8($tptr)
3105	 mov	2*8($carry),%r10
3106	mov	%r11,3*8($tptr)
3107	 mov	3*8($carry),%r11
3108	mov	%r12,4*8($tptr)
3109	 mov	4*8($carry),%r12
3110	mov	%r13,5*8($tptr)
3111	 mov	5*8($carry),%r13
3112	mov	%r14,6*8($tptr)
3113	 mov	6*8($carry),%r14
3114	mov	%r15,7*8($tptr)
3115	 mov	7*8($carry),%r15
3116	mov	$carry,$tptr
3117	jmp	.Lsqrx8x_outer_loop
3118
3119.align	32
3120.Lsqrx8x_outer_break:
3121	mov	%r9,9*8($tptr)		# t[9]
3122	 movq	%xmm3,%rcx		# -$num
3123	mov	%r10,10*8($tptr)	# ...
3124	mov	%r11,11*8($tptr)
3125	mov	%r12,12*8($tptr)
3126	mov	%r13,13*8($tptr)
3127	mov	%r14,14*8($tptr)
3128___
3129}{
3130my $i="%rcx";
3131$code.=<<___;
3132	lea	48+8(%rsp),$tptr
3133	mov	($aptr,$i),%rdx		# a[0]
3134
3135	mov	8($tptr),$A0[1]		# t[1]
3136	xor	$A0[0],$A0[0]		# t[0], of=0, cf=0
3137	mov	0+8(%rsp),$num		# restore $num
3138	adox	$A0[1],$A0[1]
3139	 mov	16($tptr),$A1[0]	# t[2]	# prefetch
3140	 mov	24($tptr),$A1[1]	# t[3]	# prefetch
3141	#jmp	.Lsqrx4x_shift_n_add	# happens to be aligned
3142
3143.align	32
3144.Lsqrx4x_shift_n_add:
3145	mulx	%rdx,%rax,%rbx
3146	 adox	$A1[0],$A1[0]
3147	adcx	$A0[0],%rax
3148	 .byte	0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00	# mov	8($aptr,$i),%rdx	# a[i+1]	# prefetch
3149	 .byte	0x4c,0x8b,0x97,0x20,0x00,0x00,0x00	# mov	32($tptr),$A0[0]	# t[2*i+4]	# prefetch
3150	 adox	$A1[1],$A1[1]
3151	adcx	$A0[1],%rbx
3152	 mov	40($tptr),$A0[1]		# t[2*i+4+1]	# prefetch
3153	mov	%rax,0($tptr)
3154	mov	%rbx,8($tptr)
3155
3156	mulx	%rdx,%rax,%rbx
3157	 adox	$A0[0],$A0[0]
3158	adcx	$A1[0],%rax
3159	 mov	16($aptr,$i),%rdx	# a[i+2]	# prefetch
3160	 mov	48($tptr),$A1[0]	# t[2*i+6]	# prefetch
3161	 adox	$A0[1],$A0[1]
3162	adcx	$A1[1],%rbx
3163	 mov	56($tptr),$A1[1]	# t[2*i+6+1]	# prefetch
3164	mov	%rax,16($tptr)
3165	mov	%rbx,24($tptr)
3166
3167	mulx	%rdx,%rax,%rbx
3168	 adox	$A1[0],$A1[0]
3169	adcx	$A0[0],%rax
3170	 mov	24($aptr,$i),%rdx	# a[i+3]	# prefetch
3171	 lea	32($i),$i
3172	 mov	64($tptr),$A0[0]	# t[2*i+8]	# prefetch
3173	 adox	$A1[1],$A1[1]
3174	adcx	$A0[1],%rbx
3175	 mov	72($tptr),$A0[1]	# t[2*i+8+1]	# prefetch
3176	mov	%rax,32($tptr)
3177	mov	%rbx,40($tptr)
3178
3179	mulx	%rdx,%rax,%rbx
3180	 adox	$A0[0],$A0[0]
3181	adcx	$A1[0],%rax
3182	jrcxz	.Lsqrx4x_shift_n_add_break
3183	 .byte	0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00	# mov	0($aptr,$i),%rdx	# a[i+4]	# prefetch
3184	 adox	$A0[1],$A0[1]
3185	adcx	$A1[1],%rbx
3186	 mov	80($tptr),$A1[0]	# t[2*i+10]	# prefetch
3187	 mov	88($tptr),$A1[1]	# t[2*i+10+1]	# prefetch
3188	mov	%rax,48($tptr)
3189	mov	%rbx,56($tptr)
3190	lea	64($tptr),$tptr
3191	nop
3192	jmp	.Lsqrx4x_shift_n_add
3193
3194.align	32
3195.Lsqrx4x_shift_n_add_break:
3196	adcx	$A1[1],%rbx
3197	mov	%rax,48($tptr)
3198	mov	%rbx,56($tptr)
3199	lea	64($tptr),$tptr		# end of t[] buffer
3200___
3201}
3202######################################################################
3203# Montgomery reduction part, "word-by-word" algorithm.
3204#
3205# This new path is inspired by multiple submissions from Intel, by
3206# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
3207# Vinodh Gopal...
3208{
3209my ($nptr,$carry,$m0)=("%rbp","%rsi","%rdx");
3210
3211$code.=<<___;
3212	movq	%xmm2,$nptr
3213__bn_sqrx8x_reduction:
3214	xor	%eax,%eax		# initial top-most carry bit
3215	mov	32+8(%rsp),%rbx		# n0
3216	mov	48+8(%rsp),%rdx		# "%r8", 8*0($tptr)
3217	lea	-8*8($nptr,$num),%rcx	# end of n[]
3218	#lea	48+8(%rsp,$num,2),$tptr	# end of t[] buffer
3219	mov	%rcx, 0+8(%rsp)		# save end of n[]
3220	mov	$tptr,8+8(%rsp)		# save end of t[]
3221
3222	lea	48+8(%rsp),$tptr		# initial t[] window
3223	jmp	.Lsqrx8x_reduction_loop
3224
3225.align	32
3226.Lsqrx8x_reduction_loop:
3227	mov	8*1($tptr),%r9
3228	mov	8*2($tptr),%r10
3229	mov	8*3($tptr),%r11
3230	mov	8*4($tptr),%r12
3231	mov	%rdx,%r8
3232	imulq	%rbx,%rdx		# n0*a[i]
3233	mov	8*5($tptr),%r13
3234	mov	8*6($tptr),%r14
3235	mov	8*7($tptr),%r15
3236	mov	%rax,24+8(%rsp)		# store top-most carry bit
3237
3238	lea	8*8($tptr),$tptr
3239	xor	$carry,$carry		# cf=0,of=0
3240	mov	\$-8,%rcx
3241	jmp	.Lsqrx8x_reduce
3242
3243.align	32
3244.Lsqrx8x_reduce:
3245	mov	%r8, %rbx
3246	mulx	8*0($nptr),%rax,%r8	# n[0]
3247	adcx	%rbx,%rax		# discarded
3248	adox	%r9,%r8
3249
3250	mulx	8*1($nptr),%rbx,%r9	# n[1]
3251	adcx	%rbx,%r8
3252	adox	%r10,%r9
3253
3254	mulx	8*2($nptr),%rbx,%r10
3255	adcx	%rbx,%r9
3256	adox	%r11,%r10
3257
3258	mulx	8*3($nptr),%rbx,%r11
3259	adcx	%rbx,%r10
3260	adox	%r12,%r11
3261
3262	.byte	0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rbx,%r12
3263	 mov	%rdx,%rax
3264	 mov	%r8,%rdx
3265	adcx	%rbx,%r11
3266	adox	%r13,%r12
3267
3268	 mulx	32+8(%rsp),%rbx,%rdx	# %rdx discarded
3269	 mov	%rax,%rdx
3270	 mov	%rax,64+48+8(%rsp,%rcx,8)	# put aside n0*a[i]
3271
3272	mulx	8*5($nptr),%rax,%r13
3273	adcx	%rax,%r12
3274	adox	%r14,%r13
3275
3276	mulx	8*6($nptr),%rax,%r14
3277	adcx	%rax,%r13
3278	adox	%r15,%r14
3279
3280	mulx	8*7($nptr),%rax,%r15
3281	 mov	%rbx,%rdx
3282	adcx	%rax,%r14
3283	adox	$carry,%r15		# $carry is 0
3284	adcx	$carry,%r15		# cf=0
3285
3286	.byte	0x67,0x67,0x67
3287	inc	%rcx			# of=0
3288	jnz	.Lsqrx8x_reduce
3289
3290	mov	$carry,%rax		# xor	%rax,%rax
3291	cmp	0+8(%rsp),$nptr		# end of n[]?
3292	jae	.Lsqrx8x_no_tail
3293
3294	mov	48+8(%rsp),%rdx		# pull n0*a[0]
3295	add	8*0($tptr),%r8
3296	lea	8*8($nptr),$nptr
3297	mov	\$-8,%rcx
3298	adcx	8*1($tptr),%r9
3299	adcx	8*2($tptr),%r10
3300	adc	8*3($tptr),%r11
3301	adc	8*4($tptr),%r12
3302	adc	8*5($tptr),%r13
3303	adc	8*6($tptr),%r14
3304	adc	8*7($tptr),%r15
3305	lea	8*8($tptr),$tptr
3306	sbb	%rax,%rax		# top carry
3307
3308	xor	$carry,$carry		# of=0, cf=0
3309	mov	%rax,16+8(%rsp)
3310	jmp	.Lsqrx8x_tail
3311
3312.align	32
3313.Lsqrx8x_tail:
3314	mov	%r8,%rbx
3315	mulx	8*0($nptr),%rax,%r8
3316	adcx	%rax,%rbx
3317	adox	%r9,%r8
3318
3319	mulx	8*1($nptr),%rax,%r9
3320	adcx	%rax,%r8
3321	adox	%r10,%r9
3322
3323	mulx	8*2($nptr),%rax,%r10
3324	adcx	%rax,%r9
3325	adox	%r11,%r10
3326
3327	mulx	8*3($nptr),%rax,%r11
3328	adcx	%rax,%r10
3329	adox	%r12,%r11
3330
3331	.byte	0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00	# mulx	8*4($nptr),%rax,%r12
3332	adcx	%rax,%r11
3333	adox	%r13,%r12
3334
3335	mulx	8*5($nptr),%rax,%r13
3336	adcx	%rax,%r12
3337	adox	%r14,%r13
3338
3339	mulx	8*6($nptr),%rax,%r14
3340	adcx	%rax,%r13
3341	adox	%r15,%r14
3342
3343	mulx	8*7($nptr),%rax,%r15
3344	 mov	72+48+8(%rsp,%rcx,8),%rdx	# pull n0*a[i]
3345	adcx	%rax,%r14
3346	adox	$carry,%r15
3347	 mov	%rbx,($tptr,%rcx,8)	# save result
3348	 mov	%r8,%rbx
3349	adcx	$carry,%r15		# cf=0
3350
3351	inc	%rcx			# of=0
3352	jnz	.Lsqrx8x_tail
3353
3354	cmp	0+8(%rsp),$nptr		# end of n[]?
3355	jae	.Lsqrx8x_tail_done	# break out of loop
3356
3357	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3358	 mov	48+8(%rsp),%rdx		# pull n0*a[0]
3359	 lea	8*8($nptr),$nptr
3360	adc	8*0($tptr),%r8
3361	adc	8*1($tptr),%r9
3362	adc	8*2($tptr),%r10
3363	adc	8*3($tptr),%r11
3364	adc	8*4($tptr),%r12
3365	adc	8*5($tptr),%r13
3366	adc	8*6($tptr),%r14
3367	adc	8*7($tptr),%r15
3368	lea	8*8($tptr),$tptr
3369	sbb	%rax,%rax
3370	sub	\$8,%rcx		# mov	\$-8,%rcx
3371
3372	xor	$carry,$carry		# of=0, cf=0
3373	mov	%rax,16+8(%rsp)
3374	jmp	.Lsqrx8x_tail
3375
3376.align	32
3377.Lsqrx8x_tail_done:
3378	add	24+8(%rsp),%r8		# can this overflow?
3379	adc	\$0,%r9
3380	adc	\$0,%r10
3381	adc	\$0,%r11
3382	adc	\$0,%r12
3383	adc	\$0,%r13
3384	adc	\$0,%r14
3385	adc	\$0,%r15		# can't overflow, because we
3386					# started with "overhung" part
3387					# of multiplication
3388	mov	$carry,%rax		# xor	%rax,%rax
3389
3390	sub	16+8(%rsp),$carry	# mov 16(%rsp),%cf
3391.Lsqrx8x_no_tail:			# %cf is 0 if jumped here
3392	adc	8*0($tptr),%r8
3393	 movq	%xmm3,%rcx
3394	adc	8*1($tptr),%r9
3395	 mov	8*7($nptr),$carry
3396	 movq	%xmm2,$nptr		# restore $nptr
3397	adc	8*2($tptr),%r10
3398	adc	8*3($tptr),%r11
3399	adc	8*4($tptr),%r12
3400	adc	8*5($tptr),%r13
3401	adc	8*6($tptr),%r14
3402	adc	8*7($tptr),%r15
3403	adc	%rax,%rax		# top-most carry
3404
3405	mov	32+8(%rsp),%rbx		# n0
3406	mov	8*8($tptr,%rcx),%rdx	# modulo-scheduled "%r8"
3407
3408	mov	%r8,8*0($tptr)		# store top 512 bits
3409	 lea	8*8($tptr),%r8		# borrow %r8
3410	mov	%r9,8*1($tptr)
3411	mov	%r10,8*2($tptr)
3412	mov	%r11,8*3($tptr)
3413	mov	%r12,8*4($tptr)
3414	mov	%r13,8*5($tptr)
3415	mov	%r14,8*6($tptr)
3416	mov	%r15,8*7($tptr)
3417
3418	lea	8*8($tptr,%rcx),$tptr	# start of current t[] window
3419	cmp	8+8(%rsp),%r8		# end of t[]?
3420	jb	.Lsqrx8x_reduction_loop
3421	ret
3422.size	bn_sqrx8x_internal,.-bn_sqrx8x_internal
3423___
3424}
3425##############################################################
3426# Post-condition, 4x unrolled
3427#
3428{
3429my ($rptr,$nptr)=("%rdx","%rbp");
3430$code.=<<___;
3431.align	32
3432__bn_postx4x_internal:
3433	mov	8*0($nptr),%r12
3434	mov	%rcx,%r10		# -$num
3435	mov	%rcx,%r9		# -$num
3436	neg	%rax
3437	sar	\$3+2,%rcx
3438	#lea	48+8(%rsp,%r9),$tptr
3439	movq	%xmm1,$rptr		# restore $rptr
3440	movq	%xmm1,$aptr		# prepare for back-to-back call
3441	dec	%r12			# so that after 'not' we get -n[0]
3442	mov	8*1($nptr),%r13
3443	xor	%r8,%r8
3444	mov	8*2($nptr),%r14
3445	mov	8*3($nptr),%r15
3446	jmp	.Lsqrx4x_sub_entry
3447
3448.align	16
3449.Lsqrx4x_sub:
3450	mov	8*0($nptr),%r12
3451	mov	8*1($nptr),%r13
3452	mov	8*2($nptr),%r14
3453	mov	8*3($nptr),%r15
3454.Lsqrx4x_sub_entry:
3455	andn	%rax,%r12,%r12
3456	lea	8*4($nptr),$nptr
3457	andn	%rax,%r13,%r13
3458	andn	%rax,%r14,%r14
3459	andn	%rax,%r15,%r15
3460
3461	neg	%r8			# mov %r8,%cf
3462	adc	8*0($tptr),%r12
3463	adc	8*1($tptr),%r13
3464	adc	8*2($tptr),%r14
3465	adc	8*3($tptr),%r15
3466	mov	%r12,8*0($rptr)
3467	lea	8*4($tptr),$tptr
3468	mov	%r13,8*1($rptr)
3469	sbb	%r8,%r8			# mov %cf,%r8
3470	mov	%r14,8*2($rptr)
3471	mov	%r15,8*3($rptr)
3472	lea	8*4($rptr),$rptr
3473
3474	inc	%rcx
3475	jnz	.Lsqrx4x_sub
3476
3477	neg	%r9			# restore $num
3478
3479	ret
3480.size	__bn_postx4x_internal,.-__bn_postx4x_internal
3481___
3482}
3483}}}
3484{
3485my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%edx","%r8", "%r9d") : # Win64 order
3486				("%rdi","%esi","%rdx","%ecx");  # Unix order
3487my $out=$inp;
3488my $STRIDE=2**5*8;
3489my $N=$STRIDE/4;
3490
3491$code.=<<___;
3492.globl	bn_get_bits5
3493.type	bn_get_bits5,\@abi-omnipotent
3494.align	16
3495bn_get_bits5:
3496	lea	0($inp),%r10
3497	lea	1($inp),%r11
3498	mov	$num,%ecx
3499	shr	\$4,$num
3500	and	\$15,%ecx
3501	lea	-8(%ecx),%eax
3502	cmp	\$11,%ecx
3503	cmova	%r11,%r10
3504	cmova	%eax,%ecx
3505	movzw	(%r10,$num,2),%eax
3506	shrl	%cl,%eax
3507	and	\$31,%eax
3508	ret
3509.size	bn_get_bits5,.-bn_get_bits5
3510
3511.globl	bn_scatter5
3512.type	bn_scatter5,\@abi-omnipotent
3513.align	16
3514bn_scatter5:
3515	cmp	\$0, $num
3516	jz	.Lscatter_epilogue
3517	lea	($tbl,$idx,8),$tbl
3518.Lscatter:
3519	mov	($inp),%rax
3520	lea	8($inp),$inp
3521	mov	%rax,($tbl)
3522	lea	32*8($tbl),$tbl
3523	sub	\$1,$num
3524	jnz	.Lscatter
3525.Lscatter_epilogue:
3526	ret
3527.size	bn_scatter5,.-bn_scatter5
3528
3529.globl	bn_gather5
3530.type	bn_gather5,\@abi-omnipotent
3531.align	32
3532bn_gather5:
3533.LSEH_begin_bn_gather5:			# Win64 thing, but harmless in other cases
3534	# I can't trust assembler to use specific encoding:-(
3535	.byte	0x4c,0x8d,0x14,0x24			#lea    (%rsp),%r10
3536	.byte	0x48,0x81,0xec,0x08,0x01,0x00,0x00	#sub	$0x108,%rsp
3537	lea	.Linc(%rip),%rax
3538	and	\$-16,%rsp		# shouldn't be formally required
3539
3540	movd	$idx,%xmm5
3541	movdqa	0(%rax),%xmm0		# 00000001000000010000000000000000
3542	movdqa	16(%rax),%xmm1		# 00000002000000020000000200000002
3543	lea	128($tbl),%r11		# size optimization
3544	lea	128(%rsp),%rax		# size optimization
3545
3546	pshufd	\$0,%xmm5,%xmm5		# broadcast $idx
3547	movdqa	%xmm1,%xmm4
3548	movdqa	%xmm1,%xmm2
3549___
3550########################################################################
3551# calculate mask by comparing 0..31 to $idx and save result to stack
3552#
3553for($i=0;$i<$STRIDE/16;$i+=4) {
3554$code.=<<___;
3555	paddd	%xmm0,%xmm1
3556	pcmpeqd	%xmm5,%xmm0		# compare to 1,0
3557___
3558$code.=<<___	if ($i);
3559	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3560___
3561$code.=<<___;
3562	movdqa	%xmm4,%xmm3
3563
3564	paddd	%xmm1,%xmm2
3565	pcmpeqd	%xmm5,%xmm1		# compare to 3,2
3566	movdqa	%xmm0,`16*($i+0)-128`(%rax)
3567	movdqa	%xmm4,%xmm0
3568
3569	paddd	%xmm2,%xmm3
3570	pcmpeqd	%xmm5,%xmm2		# compare to 5,4
3571	movdqa	%xmm1,`16*($i+1)-128`(%rax)
3572	movdqa	%xmm4,%xmm1
3573
3574	paddd	%xmm3,%xmm0
3575	pcmpeqd	%xmm5,%xmm3		# compare to 7,6
3576	movdqa	%xmm2,`16*($i+2)-128`(%rax)
3577	movdqa	%xmm4,%xmm2
3578___
3579}
3580$code.=<<___;
3581	movdqa	%xmm3,`16*($i-1)-128`(%rax)
3582	jmp	.Lgather
3583
3584.align	32
3585.Lgather:
3586	pxor	%xmm4,%xmm4
3587	pxor	%xmm5,%xmm5
3588___
3589for($i=0;$i<$STRIDE/16;$i+=4) {
3590$code.=<<___;
3591	movdqa	`16*($i+0)-128`(%r11),%xmm0
3592	movdqa	`16*($i+1)-128`(%r11),%xmm1
3593	movdqa	`16*($i+2)-128`(%r11),%xmm2
3594	pand	`16*($i+0)-128`(%rax),%xmm0
3595	movdqa	`16*($i+3)-128`(%r11),%xmm3
3596	pand	`16*($i+1)-128`(%rax),%xmm1
3597	por	%xmm0,%xmm4
3598	pand	`16*($i+2)-128`(%rax),%xmm2
3599	por	%xmm1,%xmm5
3600	pand	`16*($i+3)-128`(%rax),%xmm3
3601	por	%xmm2,%xmm4
3602	por	%xmm3,%xmm5
3603___
3604}
3605$code.=<<___;
3606	por	%xmm5,%xmm4
3607	lea	$STRIDE(%r11),%r11
3608	pshufd	\$0x4e,%xmm4,%xmm0
3609	por	%xmm4,%xmm0
3610	movq	%xmm0,($out)		# m0=bp[0]
3611	lea	8($out),$out
3612	sub	\$1,$num
3613	jnz	.Lgather
3614
3615	lea	(%r10),%rsp
3616	ret
3617.LSEH_end_bn_gather5:
3618.size	bn_gather5,.-bn_gather5
3619___
3620}
3621$code.=<<___;
3622.align	64
3623.Linc:
3624	.long	0,0, 1,1
3625	.long	2,2, 2,2
3626.asciz	"Montgomery Multiplication with scatter/gather for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
3627___
3628
3629# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
3630#		CONTEXT *context,DISPATCHER_CONTEXT *disp)
3631if ($win64) {
3632$rec="%rcx";
3633$frame="%rdx";
3634$context="%r8";
3635$disp="%r9";
3636
3637$code.=<<___;
3638.extern	__imp_RtlVirtualUnwind
3639.type	mul_handler,\@abi-omnipotent
3640.align	16
3641mul_handler:
3642	push	%rsi
3643	push	%rdi
3644	push	%rbx
3645	push	%rbp
3646	push	%r12
3647	push	%r13
3648	push	%r14
3649	push	%r15
3650	pushfq
3651	sub	\$64,%rsp
3652
3653	mov	120($context),%rax	# pull context->Rax
3654	mov	248($context),%rbx	# pull context->Rip
3655
3656	mov	8($disp),%rsi		# disp->ImageBase
3657	mov	56($disp),%r11		# disp->HandlerData
3658
3659	mov	0(%r11),%r10d		# HandlerData[0]
3660	lea	(%rsi,%r10),%r10	# end of prologue label
3661	cmp	%r10,%rbx		# context->Rip<end of prologue label
3662	jb	.Lcommon_seh_tail
3663
3664	mov	4(%r11),%r10d		# HandlerData[1]
3665	lea	(%rsi,%r10),%r10	# epilogue label
3666	cmp	%r10,%rbx		# context->Rip>=epilogue label
3667	jb	.Lcommon_pop_regs
3668
3669	mov	152($context),%rax	# pull context->Rsp
3670
3671	mov	8(%r11),%r10d		# HandlerData[2]
3672	lea	(%rsi,%r10),%r10	# epilogue label
3673	cmp	%r10,%rbx		# context->Rip>=epilogue label
3674	jae	.Lcommon_seh_tail
3675
3676	lea	.Lmul_epilogue(%rip),%r10
3677	cmp	%r10,%rbx
3678	ja	.Lbody_40
3679
3680	mov	192($context),%r10	# pull $num
3681	mov	8(%rax,%r10,8),%rax	# pull saved stack pointer
3682
3683	jmp	.Lcommon_pop_regs
3684
3685.Lbody_40:
3686	mov	40(%rax),%rax		# pull saved stack pointer
3687.Lcommon_pop_regs:
3688	mov	-8(%rax),%rbx
3689	mov	-16(%rax),%rbp
3690	mov	-24(%rax),%r12
3691	mov	-32(%rax),%r13
3692	mov	-40(%rax),%r14
3693	mov	-48(%rax),%r15
3694	mov	%rbx,144($context)	# restore context->Rbx
3695	mov	%rbp,160($context)	# restore context->Rbp
3696	mov	%r12,216($context)	# restore context->R12
3697	mov	%r13,224($context)	# restore context->R13
3698	mov	%r14,232($context)	# restore context->R14
3699	mov	%r15,240($context)	# restore context->R15
3700
3701.Lcommon_seh_tail:
3702	mov	8(%rax),%rdi
3703	mov	16(%rax),%rsi
3704	mov	%rax,152($context)	# restore context->Rsp
3705	mov	%rsi,168($context)	# restore context->Rsi
3706	mov	%rdi,176($context)	# restore context->Rdi
3707
3708	mov	40($disp),%rdi		# disp->ContextRecord
3709	mov	$context,%rsi		# context
3710	mov	\$154,%ecx		# sizeof(CONTEXT)
3711	.long	0xa548f3fc		# cld; rep movsq
3712
3713	mov	$disp,%rsi
3714	xor	%rcx,%rcx		# arg1, UNW_FLAG_NHANDLER
3715	mov	8(%rsi),%rdx		# arg2, disp->ImageBase
3716	mov	0(%rsi),%r8		# arg3, disp->ControlPc
3717	mov	16(%rsi),%r9		# arg4, disp->FunctionEntry
3718	mov	40(%rsi),%r10		# disp->ContextRecord
3719	lea	56(%rsi),%r11		# &disp->HandlerData
3720	lea	24(%rsi),%r12		# &disp->EstablisherFrame
3721	mov	%r10,32(%rsp)		# arg5
3722	mov	%r11,40(%rsp)		# arg6
3723	mov	%r12,48(%rsp)		# arg7
3724	mov	%rcx,56(%rsp)		# arg8, (NULL)
3725	call	*__imp_RtlVirtualUnwind(%rip)
3726
3727	mov	\$1,%eax		# ExceptionContinueSearch
3728	add	\$64,%rsp
3729	popfq
3730	pop	%r15
3731	pop	%r14
3732	pop	%r13
3733	pop	%r12
3734	pop	%rbp
3735	pop	%rbx
3736	pop	%rdi
3737	pop	%rsi
3738	ret
3739.size	mul_handler,.-mul_handler
3740
3741.section	.pdata
3742.align	4
3743	.rva	.LSEH_begin_bn_mul_mont_gather5
3744	.rva	.LSEH_end_bn_mul_mont_gather5
3745	.rva	.LSEH_info_bn_mul_mont_gather5
3746
3747	.rva	.LSEH_begin_bn_mul4x_mont_gather5
3748	.rva	.LSEH_end_bn_mul4x_mont_gather5
3749	.rva	.LSEH_info_bn_mul4x_mont_gather5
3750
3751	.rva	.LSEH_begin_bn_power5
3752	.rva	.LSEH_end_bn_power5
3753	.rva	.LSEH_info_bn_power5
3754
3755	.rva	.LSEH_begin_bn_from_mont8x
3756	.rva	.LSEH_end_bn_from_mont8x
3757	.rva	.LSEH_info_bn_from_mont8x
3758___
3759$code.=<<___ if ($addx);
3760	.rva	.LSEH_begin_bn_mulx4x_mont_gather5
3761	.rva	.LSEH_end_bn_mulx4x_mont_gather5
3762	.rva	.LSEH_info_bn_mulx4x_mont_gather5
3763
3764	.rva	.LSEH_begin_bn_powerx5
3765	.rva	.LSEH_end_bn_powerx5
3766	.rva	.LSEH_info_bn_powerx5
3767___
3768$code.=<<___;
3769	.rva	.LSEH_begin_bn_gather5
3770	.rva	.LSEH_end_bn_gather5
3771	.rva	.LSEH_info_bn_gather5
3772
3773.section	.xdata
3774.align	8
3775.LSEH_info_bn_mul_mont_gather5:
3776	.byte	9,0,0,0
3777	.rva	mul_handler
3778	.rva	.Lmul_body,.Lmul_body,.Lmul_epilogue		# HandlerData[]
3779.align	8
3780.LSEH_info_bn_mul4x_mont_gather5:
3781	.byte	9,0,0,0
3782	.rva	mul_handler
3783	.rva	.Lmul4x_prologue,.Lmul4x_body,.Lmul4x_epilogue		# HandlerData[]
3784.align	8
3785.LSEH_info_bn_power5:
3786	.byte	9,0,0,0
3787	.rva	mul_handler
3788	.rva	.Lpower5_prologue,.Lpower5_body,.Lpower5_epilogue	# HandlerData[]
3789.align	8
3790.LSEH_info_bn_from_mont8x:
3791	.byte	9,0,0,0
3792	.rva	mul_handler
3793	.rva	.Lfrom_prologue,.Lfrom_body,.Lfrom_epilogue		# HandlerData[]
3794___
3795$code.=<<___ if ($addx);
3796.align	8
3797.LSEH_info_bn_mulx4x_mont_gather5:
3798	.byte	9,0,0,0
3799	.rva	mul_handler
3800	.rva	.Lmulx4x_prologue,.Lmulx4x_body,.Lmulx4x_epilogue	# HandlerData[]
3801.align	8
3802.LSEH_info_bn_powerx5:
3803	.byte	9,0,0,0
3804	.rva	mul_handler
3805	.rva	.Lpowerx5_prologue,.Lpowerx5_body,.Lpowerx5_epilogue	# HandlerData[]
3806___
3807$code.=<<___;
3808.align	8
3809.LSEH_info_bn_gather5:
3810	.byte	0x01,0x0b,0x03,0x0a
3811	.byte	0x0b,0x01,0x21,0x00	# sub	rsp,0x108
3812	.byte	0x04,0xa3,0x00,0x00	# lea	r10,(rsp)
3813.align	8
3814___
3815}
3816
3817$code =~ s/\`([^\`]*)\`/eval($1)/gem;
3818
3819print $code;
3820close STDOUT;
3821