xref: /openssl/crypto/bn/asm/armv4-mont.pl (revision da1c088f)
1#! /usr/bin/env perl
2# Copyright 2007-2023 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the Apache License 2.0 (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16
17# January 2007.
18
19# Montgomery multiplication for ARMv4.
20#
21# Performance improvement naturally varies among CPU implementations
22# and compilers. The code was observed to provide +65-35% improvement
23# [depending on key length, less for longer keys] on ARM920T, and
24# +115-80% on Intel IXP425. This is compared to pre-bn_mul_mont code
25# base and compiler generated code with in-lined umull and even umlal
26# instructions. The latter means that this code didn't really have an
27# "advantage" of utilizing some "secret" instruction.
28#
29# The code is interoperable with Thumb ISA and is rather compact, less
30# than 1/2KB. Windows CE port would be trivial, as it's exclusively
31# about decorations, ABI and instruction syntax are identical.
32
33# November 2013
34#
35# Add NEON code path, which handles lengths divisible by 8. RSA/DSA
36# performance improvement on Cortex-A8 is ~45-100% depending on key
37# length, more for longer keys. On Cortex-A15 the span is ~10-105%.
38# On Snapdragon S4 improvement was measured to vary from ~70% to
39# incredible ~380%, yes, 4.8x faster, for RSA4096 sign. But this is
40# rather because original integer-only code seems to perform
41# suboptimally on S4. Situation on Cortex-A9 is unfortunately
42# different. It's being looked into, but the trouble is that
43# performance for vectors longer than 256 bits is actually couple
44# of percent worse than for integer-only code. The code is chosen
45# for execution on all NEON-capable processors, because gain on
46# others outweighs the marginal loss on Cortex-A9.
47
48# September 2015
49#
50# Align Cortex-A9 performance with November 2013 improvements, i.e.
51# NEON code is now ~20-105% faster than integer-only one on this
52# processor. But this optimization further improved performance even
53# on other processors: NEON code path is ~45-180% faster than original
54# integer-only on Cortex-A8, ~10-210% on Cortex-A15, ~70-450% on
55# Snapdragon S4.
56
57# $output is the last argument if it looks like a file (it has an extension)
58# $flavour is the first argument if it doesn't look like a file
59my $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
60my $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
61
62if ($flavour && $flavour ne "void") {
63    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
64    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
65    ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
66    die "can't locate arm-xlate.pl";
67
68    open STDOUT,"| \"$^X\" $xlate $flavour \"$output\""
69        or die "can't call $xlate: $1";
70} else {
71    $output and open STDOUT,">$output";
72}
73
74$num="r0";	# starts as num argument, but holds &tp[num-1]
75$ap="r1";
76$bp="r2"; $bi="r2"; $rp="r2";
77$np="r3";
78$tp="r4";
79$aj="r5";
80$nj="r6";
81$tj="r7";
82$n0="r8";
83###########	# r9 is reserved by ELF as platform specific, e.g. TLS pointer
84$alo="r10";	# sl, gcc uses it to keep @GOT
85$ahi="r11";	# fp
86$nlo="r12";	# ip
87###########	# r13 is stack pointer
88$nhi="r14";	# lr
89###########	# r15 is program counter
90
91#### argument block layout relative to &tp[num-1], a.k.a. $num
92$_rp="$num,#12*4";
93# ap permanently resides in r1
94$_bp="$num,#13*4";
95# np permanently resides in r3
96$_n0="$num,#14*4";
97$_num="$num,#15*4";	$_bpend=$_num;
98
99$code=<<___;
100#include "arm_arch.h"
101
102#if defined(__thumb2__)
103.syntax	unified
104.thumb
105#else
106.code	32
107#endif
108
109.text
110
111#if __ARM_MAX_ARCH__>=7
112.align	5
113.LOPENSSL_armcap:
114# ifdef	_WIN32
115.word	OPENSSL_armcap_P
116# else
117.word	OPENSSL_armcap_P-.Lbn_mul_mont
118# endif
119#endif
120
121.global	bn_mul_mont
122.type	bn_mul_mont,%function
123
124.align	5
125bn_mul_mont:
126.Lbn_mul_mont:
127	ldr	ip,[sp,#4]		@ load num
128	stmdb	sp!,{r0,r2}		@ sp points at argument block
129#if __ARM_MAX_ARCH__>=7
130	tst	ip,#7
131	bne	.Lialu
132	ldr	r0,.LOPENSSL_armcap
133#if !defined(_WIN32)
134	adr	r2,.Lbn_mul_mont
135	ldr	r0,[r0,r2]
136# endif
137# if defined(__APPLE__) || defined(_WIN32)
138	ldr	r0,[r0]
139# endif
140	tst	r0,#ARMV7_NEON		@ NEON available?
141	ldmia	sp, {r0,r2}
142	beq	.Lialu
143	add	sp,sp,#8
144	b	bn_mul8x_mont_neon
145.align	4
146.Lialu:
147#endif
148	cmp	ip,#2
149	mov	$num,ip			@ load num
150#ifdef	__thumb2__
151	ittt	lt
152#endif
153	movlt	r0,#0
154	addlt	sp,sp,#2*4
155	blt	.Labrt
156
157	stmdb	sp!,{r4-r12,lr}		@ save 10 registers
158
159	mov	$num,$num,lsl#2		@ rescale $num for byte count
160	sub	sp,sp,$num		@ alloca(4*num)
161	sub	sp,sp,#4		@ +extra dword
162	sub	$num,$num,#4		@ "num=num-1"
163	add	$tp,$bp,$num		@ &bp[num-1]
164
165	add	$num,sp,$num		@ $num to point at &tp[num-1]
166	ldr	$n0,[$_n0]		@ &n0
167	ldr	$bi,[$bp]		@ bp[0]
168	ldr	$aj,[$ap],#4		@ ap[0],ap++
169	ldr	$nj,[$np],#4		@ np[0],np++
170	ldr	$n0,[$n0]		@ *n0
171	str	$tp,[$_bpend]		@ save &bp[num]
172
173	umull	$alo,$ahi,$aj,$bi	@ ap[0]*bp[0]
174	str	$n0,[$_n0]		@ save n0 value
175	mul	$n0,$alo,$n0		@ "tp[0]"*n0
176	mov	$nlo,#0
177	umlal	$alo,$nlo,$nj,$n0	@ np[0]*n0+"t[0]"
178	mov	$tp,sp
179
180.L1st:
181	ldr	$aj,[$ap],#4		@ ap[j],ap++
182	mov	$alo,$ahi
183	ldr	$nj,[$np],#4		@ np[j],np++
184	mov	$ahi,#0
185	umlal	$alo,$ahi,$aj,$bi	@ ap[j]*bp[0]
186	mov	$nhi,#0
187	umlal	$nlo,$nhi,$nj,$n0	@ np[j]*n0
188	adds	$nlo,$nlo,$alo
189	str	$nlo,[$tp],#4		@ tp[j-1]=,tp++
190	adc	$nlo,$nhi,#0
191	cmp	$tp,$num
192	bne	.L1st
193
194	adds	$nlo,$nlo,$ahi
195	ldr	$tp,[$_bp]		@ restore bp
196	mov	$nhi,#0
197	ldr	$n0,[$_n0]		@ restore n0
198	adc	$nhi,$nhi,#0
199	str	$nlo,[$num]		@ tp[num-1]=
200	mov	$tj,sp
201	str	$nhi,[$num,#4]		@ tp[num]=
202
203.Louter:
204	sub	$tj,$num,$tj		@ "original" $num-1 value
205	sub	$ap,$ap,$tj		@ "rewind" ap to &ap[1]
206	ldr	$bi,[$tp,#4]!		@ *(++bp)
207	sub	$np,$np,$tj		@ "rewind" np to &np[1]
208	ldr	$aj,[$ap,#-4]		@ ap[0]
209	ldr	$alo,[sp]		@ tp[0]
210	ldr	$nj,[$np,#-4]		@ np[0]
211	ldr	$tj,[sp,#4]		@ tp[1]
212
213	mov	$ahi,#0
214	umlal	$alo,$ahi,$aj,$bi	@ ap[0]*bp[i]+tp[0]
215	str	$tp,[$_bp]		@ save bp
216	mul	$n0,$alo,$n0
217	mov	$nlo,#0
218	umlal	$alo,$nlo,$nj,$n0	@ np[0]*n0+"tp[0]"
219	mov	$tp,sp
220
221.Linner:
222	ldr	$aj,[$ap],#4		@ ap[j],ap++
223	adds	$alo,$ahi,$tj		@ +=tp[j]
224	ldr	$nj,[$np],#4		@ np[j],np++
225	mov	$ahi,#0
226	umlal	$alo,$ahi,$aj,$bi	@ ap[j]*bp[i]
227	mov	$nhi,#0
228	umlal	$nlo,$nhi,$nj,$n0	@ np[j]*n0
229	adc	$ahi,$ahi,#0
230	ldr	$tj,[$tp,#8]		@ tp[j+1]
231	adds	$nlo,$nlo,$alo
232	str	$nlo,[$tp],#4		@ tp[j-1]=,tp++
233	adc	$nlo,$nhi,#0
234	cmp	$tp,$num
235	bne	.Linner
236
237	adds	$nlo,$nlo,$ahi
238	mov	$nhi,#0
239	ldr	$tp,[$_bp]		@ restore bp
240	adc	$nhi,$nhi,#0
241	ldr	$n0,[$_n0]		@ restore n0
242	adds	$nlo,$nlo,$tj
243	ldr	$tj,[$_bpend]		@ restore &bp[num]
244	adc	$nhi,$nhi,#0
245	str	$nlo,[$num]		@ tp[num-1]=
246	str	$nhi,[$num,#4]		@ tp[num]=
247
248	cmp	$tp,$tj
249#ifdef	__thumb2__
250	itt	ne
251#endif
252	movne	$tj,sp
253	bne	.Louter
254
255	ldr	$rp,[$_rp]		@ pull rp
256	mov	$aj,sp
257	add	$num,$num,#4		@ $num to point at &tp[num]
258	sub	$aj,$num,$aj		@ "original" num value
259	mov	$tp,sp			@ "rewind" $tp
260	mov	$ap,$tp			@ "borrow" $ap
261	sub	$np,$np,$aj		@ "rewind" $np to &np[0]
262
263	subs	$tj,$tj,$tj		@ "clear" carry flag
264.Lsub:	ldr	$tj,[$tp],#4
265	ldr	$nj,[$np],#4
266	sbcs	$tj,$tj,$nj		@ tp[j]-np[j]
267	str	$tj,[$rp],#4		@ rp[j]=
268	teq	$tp,$num		@ preserve carry
269	bne	.Lsub
270	sbcs	$nhi,$nhi,#0		@ upmost carry
271	mov	$tp,sp			@ "rewind" $tp
272	sub	$rp,$rp,$aj		@ "rewind" $rp
273
274.Lcopy:	ldr	$tj,[$tp]		@ conditional copy
275	ldr	$aj,[$rp]
276	str	sp,[$tp],#4		@ zap tp
277#ifdef	__thumb2__
278	it	cc
279#endif
280	movcc	$aj,$tj
281	str	$aj,[$rp],#4
282	teq	$tp,$num		@ preserve carry
283	bne	.Lcopy
284
285	mov	sp,$num
286	add	sp,sp,#4		@ skip over tp[num+1]
287	ldmia	sp!,{r4-r12,lr}		@ restore registers
288	add	sp,sp,#2*4		@ skip over {r0,r2}
289	mov	r0,#1
290.Labrt:
291#if __ARM_ARCH__>=5
292	ret				@ bx lr
293#else
294	tst	lr,#1
295	moveq	pc,lr			@ be binary compatible with V4, yet
296	bx	lr			@ interoperable with Thumb ISA:-)
297#endif
298.size	bn_mul_mont,.-bn_mul_mont
299___
300{
301my ($A0,$A1,$A2,$A3)=map("d$_",(0..3));
302my ($N0,$N1,$N2,$N3)=map("d$_",(4..7));
303my ($Z,$Temp)=("q4","q5");
304my @ACC=map("q$_",(6..13));
305my ($Bi,$Ni,$M0)=map("d$_",(28..31));
306my $zero="$Z#lo";
307my $temp="$Temp#lo";
308
309my ($rptr,$aptr,$bptr,$nptr,$n0,$num)=map("r$_",(0..5));
310my ($tinptr,$toutptr,$inner,$outer,$bnptr)=map("r$_",(6..11));
311
312$code.=<<___;
313#if __ARM_MAX_ARCH__>=7
314.arch	armv7-a
315.fpu	neon
316
317.type	bn_mul8x_mont_neon,%function
318.align	5
319bn_mul8x_mont_neon:
320	mov	ip,sp
321	stmdb	sp!,{r4-r11}
322	vstmdb	sp!,{d8-d15}		@ ABI specification says so
323	ldmia	ip,{r4-r5}		@ load rest of parameter block
324	mov	ip,sp
325
326	cmp	$num,#8
327	bhi	.LNEON_8n
328
329	@ special case for $num==8, everything is in register bank...
330
331	vld1.32		{${Bi}[0]}, [$bptr,:32]!
332	veor		$zero,$zero,$zero
333	sub		$toutptr,sp,$num,lsl#4
334	vld1.32		{$A0-$A3},  [$aptr]!		@ can't specify :32 :-(
335	and		$toutptr,$toutptr,#-64
336	vld1.32		{${M0}[0]}, [$n0,:32]
337	mov		sp,$toutptr			@ alloca
338	vzip.16		$Bi,$zero
339
340	vmull.u32	@ACC[0],$Bi,${A0}[0]
341	vmull.u32	@ACC[1],$Bi,${A0}[1]
342	vmull.u32	@ACC[2],$Bi,${A1}[0]
343	vshl.i64	$Ni,@ACC[0]#hi,#16
344	vmull.u32	@ACC[3],$Bi,${A1}[1]
345
346	vadd.u64	$Ni,$Ni,@ACC[0]#lo
347	veor		$zero,$zero,$zero
348	vmul.u32	$Ni,$Ni,$M0
349
350	vmull.u32	@ACC[4],$Bi,${A2}[0]
351	 vld1.32	{$N0-$N3}, [$nptr]!
352	vmull.u32	@ACC[5],$Bi,${A2}[1]
353	vmull.u32	@ACC[6],$Bi,${A3}[0]
354	vzip.16		$Ni,$zero
355	vmull.u32	@ACC[7],$Bi,${A3}[1]
356
357	vmlal.u32	@ACC[0],$Ni,${N0}[0]
358	sub		$outer,$num,#1
359	vmlal.u32	@ACC[1],$Ni,${N0}[1]
360	vmlal.u32	@ACC[2],$Ni,${N1}[0]
361	vmlal.u32	@ACC[3],$Ni,${N1}[1]
362
363	vmlal.u32	@ACC[4],$Ni,${N2}[0]
364	vmov		$Temp,@ACC[0]
365	vmlal.u32	@ACC[5],$Ni,${N2}[1]
366	vmov		@ACC[0],@ACC[1]
367	vmlal.u32	@ACC[6],$Ni,${N3}[0]
368	vmov		@ACC[1],@ACC[2]
369	vmlal.u32	@ACC[7],$Ni,${N3}[1]
370	vmov		@ACC[2],@ACC[3]
371	vmov		@ACC[3],@ACC[4]
372	vshr.u64	$temp,$temp,#16
373	vmov		@ACC[4],@ACC[5]
374	vmov		@ACC[5],@ACC[6]
375	vadd.u64	$temp,$temp,$Temp#hi
376	vmov		@ACC[6],@ACC[7]
377	veor		@ACC[7],@ACC[7]
378	vshr.u64	$temp,$temp,#16
379
380	b	.LNEON_outer8
381
382.align	4
383.LNEON_outer8:
384	vld1.32		{${Bi}[0]}, [$bptr,:32]!
385	veor		$zero,$zero,$zero
386	vzip.16		$Bi,$zero
387	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
388
389	vmlal.u32	@ACC[0],$Bi,${A0}[0]
390	vmlal.u32	@ACC[1],$Bi,${A0}[1]
391	vmlal.u32	@ACC[2],$Bi,${A1}[0]
392	vshl.i64	$Ni,@ACC[0]#hi,#16
393	vmlal.u32	@ACC[3],$Bi,${A1}[1]
394
395	vadd.u64	$Ni,$Ni,@ACC[0]#lo
396	veor		$zero,$zero,$zero
397	subs		$outer,$outer,#1
398	vmul.u32	$Ni,$Ni,$M0
399
400	vmlal.u32	@ACC[4],$Bi,${A2}[0]
401	vmlal.u32	@ACC[5],$Bi,${A2}[1]
402	vmlal.u32	@ACC[6],$Bi,${A3}[0]
403	vzip.16		$Ni,$zero
404	vmlal.u32	@ACC[7],$Bi,${A3}[1]
405
406	vmlal.u32	@ACC[0],$Ni,${N0}[0]
407	vmlal.u32	@ACC[1],$Ni,${N0}[1]
408	vmlal.u32	@ACC[2],$Ni,${N1}[0]
409	vmlal.u32	@ACC[3],$Ni,${N1}[1]
410
411	vmlal.u32	@ACC[4],$Ni,${N2}[0]
412	vmov		$Temp,@ACC[0]
413	vmlal.u32	@ACC[5],$Ni,${N2}[1]
414	vmov		@ACC[0],@ACC[1]
415	vmlal.u32	@ACC[6],$Ni,${N3}[0]
416	vmov		@ACC[1],@ACC[2]
417	vmlal.u32	@ACC[7],$Ni,${N3}[1]
418	vmov		@ACC[2],@ACC[3]
419	vmov		@ACC[3],@ACC[4]
420	vshr.u64	$temp,$temp,#16
421	vmov		@ACC[4],@ACC[5]
422	vmov		@ACC[5],@ACC[6]
423	vadd.u64	$temp,$temp,$Temp#hi
424	vmov		@ACC[6],@ACC[7]
425	veor		@ACC[7],@ACC[7]
426	vshr.u64	$temp,$temp,#16
427
428	bne	.LNEON_outer8
429
430	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
431	mov		$toutptr,sp
432	vshr.u64	$temp,@ACC[0]#lo,#16
433	mov		$inner,$num
434	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
435	add		$tinptr,sp,#96
436	vshr.u64	$temp,@ACC[0]#hi,#16
437	vzip.16		@ACC[0]#lo,@ACC[0]#hi
438
439	b	.LNEON_tail_entry
440
441.align	4
442.LNEON_8n:
443	veor		@ACC[0],@ACC[0],@ACC[0]
444	 sub		$toutptr,sp,#128
445	veor		@ACC[1],@ACC[1],@ACC[1]
446	 sub		$toutptr,$toutptr,$num,lsl#4
447	veor		@ACC[2],@ACC[2],@ACC[2]
448	 and		$toutptr,$toutptr,#-64
449	veor		@ACC[3],@ACC[3],@ACC[3]
450	 mov		sp,$toutptr			@ alloca
451	veor		@ACC[4],@ACC[4],@ACC[4]
452	 add		$toutptr,$toutptr,#256
453	veor		@ACC[5],@ACC[5],@ACC[5]
454	 sub		$inner,$num,#8
455	veor		@ACC[6],@ACC[6],@ACC[6]
456	veor		@ACC[7],@ACC[7],@ACC[7]
457
458.LNEON_8n_init:
459	vst1.64		{@ACC[0]-@ACC[1]},[$toutptr,:256]!
460	subs		$inner,$inner,#8
461	vst1.64		{@ACC[2]-@ACC[3]},[$toutptr,:256]!
462	vst1.64		{@ACC[4]-@ACC[5]},[$toutptr,:256]!
463	vst1.64		{@ACC[6]-@ACC[7]},[$toutptr,:256]!
464	bne		.LNEON_8n_init
465
466	add		$tinptr,sp,#256
467	vld1.32		{$A0-$A3},[$aptr]!
468	add		$bnptr,sp,#8
469	vld1.32		{${M0}[0]},[$n0,:32]
470	mov		$outer,$num
471	b		.LNEON_8n_outer
472
473.align	4
474.LNEON_8n_outer:
475	vld1.32		{${Bi}[0]},[$bptr,:32]!	@ *b++
476	veor		$zero,$zero,$zero
477	vzip.16		$Bi,$zero
478	add		$toutptr,sp,#128
479	vld1.32		{$N0-$N3},[$nptr]!
480
481	vmlal.u32	@ACC[0],$Bi,${A0}[0]
482	vmlal.u32	@ACC[1],$Bi,${A0}[1]
483	 veor		$zero,$zero,$zero
484	vmlal.u32	@ACC[2],$Bi,${A1}[0]
485	 vshl.i64	$Ni,@ACC[0]#hi,#16
486	vmlal.u32	@ACC[3],$Bi,${A1}[1]
487	 vadd.u64	$Ni,$Ni,@ACC[0]#lo
488	vmlal.u32	@ACC[4],$Bi,${A2}[0]
489	 vmul.u32	$Ni,$Ni,$M0
490	vmlal.u32	@ACC[5],$Bi,${A2}[1]
491	vst1.32		{$Bi},[sp,:64]		@ put aside smashed b[8*i+0]
492	vmlal.u32	@ACC[6],$Bi,${A3}[0]
493	 vzip.16	$Ni,$zero
494	vmlal.u32	@ACC[7],$Bi,${A3}[1]
495___
496for ($i=0; $i<7;) {
497$code.=<<___;
498	vld1.32		{${Bi}[0]},[$bptr,:32]!	@ *b++
499	vmlal.u32	@ACC[0],$Ni,${N0}[0]
500	veor		$temp,$temp,$temp
501	vmlal.u32	@ACC[1],$Ni,${N0}[1]
502	vzip.16		$Bi,$temp
503	vmlal.u32	@ACC[2],$Ni,${N1}[0]
504	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
505	vmlal.u32	@ACC[3],$Ni,${N1}[1]
506	vmlal.u32	@ACC[4],$Ni,${N2}[0]
507	 vadd.u64	@ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
508	vmlal.u32	@ACC[5],$Ni,${N2}[1]
509	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
510	vmlal.u32	@ACC[6],$Ni,${N3}[0]
511	vmlal.u32	@ACC[7],$Ni,${N3}[1]
512	 vadd.u64	@ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
513	vst1.32		{$Ni},[$bnptr,:64]!	@ put aside smashed m[8*i+$i]
514___
515	push(@ACC,shift(@ACC));	$i++;
516$code.=<<___;
517	vmlal.u32	@ACC[0],$Bi,${A0}[0]
518	vld1.64		{@ACC[7]},[$tinptr,:128]!
519	vmlal.u32	@ACC[1],$Bi,${A0}[1]
520	 veor		$zero,$zero,$zero
521	vmlal.u32	@ACC[2],$Bi,${A1}[0]
522	 vshl.i64	$Ni,@ACC[0]#hi,#16
523	vmlal.u32	@ACC[3],$Bi,${A1}[1]
524	 vadd.u64	$Ni,$Ni,@ACC[0]#lo
525	vmlal.u32	@ACC[4],$Bi,${A2}[0]
526	 vmul.u32	$Ni,$Ni,$M0
527	vmlal.u32	@ACC[5],$Bi,${A2}[1]
528	vst1.32		{$Bi},[$bnptr,:64]!	@ put aside smashed b[8*i+$i]
529	vmlal.u32	@ACC[6],$Bi,${A3}[0]
530	 vzip.16	$Ni,$zero
531	vmlal.u32	@ACC[7],$Bi,${A3}[1]
532___
533}
534$code.=<<___;
535	vld1.32		{$Bi},[sp,:64]		@ pull smashed b[8*i+0]
536	vmlal.u32	@ACC[0],$Ni,${N0}[0]
537	vld1.32		{$A0-$A3},[$aptr]!
538	vmlal.u32	@ACC[1],$Ni,${N0}[1]
539	vmlal.u32	@ACC[2],$Ni,${N1}[0]
540	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
541	vmlal.u32	@ACC[3],$Ni,${N1}[1]
542	vmlal.u32	@ACC[4],$Ni,${N2}[0]
543	 vadd.u64	@ACC[0]#lo,@ACC[0]#lo,@ACC[0]#hi
544	vmlal.u32	@ACC[5],$Ni,${N2}[1]
545	 vshr.u64	@ACC[0]#lo,@ACC[0]#lo,#16
546	vmlal.u32	@ACC[6],$Ni,${N3}[0]
547	vmlal.u32	@ACC[7],$Ni,${N3}[1]
548	 vadd.u64	@ACC[1]#lo,@ACC[1]#lo,@ACC[0]#lo
549	vst1.32		{$Ni},[$bnptr,:64]	@ put aside smashed m[8*i+$i]
550	add		$bnptr,sp,#8		@ rewind
551___
552	push(@ACC,shift(@ACC));
553$code.=<<___;
554	sub		$inner,$num,#8
555	b		.LNEON_8n_inner
556
557.align	4
558.LNEON_8n_inner:
559	subs		$inner,$inner,#8
560	vmlal.u32	@ACC[0],$Bi,${A0}[0]
561	vld1.64		{@ACC[7]},[$tinptr,:128]
562	vmlal.u32	@ACC[1],$Bi,${A0}[1]
563	vld1.32		{$Ni},[$bnptr,:64]!	@ pull smashed m[8*i+0]
564	vmlal.u32	@ACC[2],$Bi,${A1}[0]
565	vld1.32		{$N0-$N3},[$nptr]!
566	vmlal.u32	@ACC[3],$Bi,${A1}[1]
567	it		ne
568	addne		$tinptr,$tinptr,#16	@ don't advance in last iteration
569	vmlal.u32	@ACC[4],$Bi,${A2}[0]
570	vmlal.u32	@ACC[5],$Bi,${A2}[1]
571	vmlal.u32	@ACC[6],$Bi,${A3}[0]
572	vmlal.u32	@ACC[7],$Bi,${A3}[1]
573___
574for ($i=1; $i<8; $i++) {
575$code.=<<___;
576	vld1.32		{$Bi},[$bnptr,:64]!	@ pull smashed b[8*i+$i]
577	vmlal.u32	@ACC[0],$Ni,${N0}[0]
578	vmlal.u32	@ACC[1],$Ni,${N0}[1]
579	vmlal.u32	@ACC[2],$Ni,${N1}[0]
580	vmlal.u32	@ACC[3],$Ni,${N1}[1]
581	vmlal.u32	@ACC[4],$Ni,${N2}[0]
582	vmlal.u32	@ACC[5],$Ni,${N2}[1]
583	vmlal.u32	@ACC[6],$Ni,${N3}[0]
584	vmlal.u32	@ACC[7],$Ni,${N3}[1]
585	vst1.64		{@ACC[0]},[$toutptr,:128]!
586___
587	push(@ACC,shift(@ACC));
588$code.=<<___;
589	vmlal.u32	@ACC[0],$Bi,${A0}[0]
590	vld1.64		{@ACC[7]},[$tinptr,:128]
591	vmlal.u32	@ACC[1],$Bi,${A0}[1]
592	vld1.32		{$Ni},[$bnptr,:64]!	@ pull smashed m[8*i+$i]
593	vmlal.u32	@ACC[2],$Bi,${A1}[0]
594	it		ne
595	addne		$tinptr,$tinptr,#16	@ don't advance in last iteration
596	vmlal.u32	@ACC[3],$Bi,${A1}[1]
597	vmlal.u32	@ACC[4],$Bi,${A2}[0]
598	vmlal.u32	@ACC[5],$Bi,${A2}[1]
599	vmlal.u32	@ACC[6],$Bi,${A3}[0]
600	vmlal.u32	@ACC[7],$Bi,${A3}[1]
601___
602}
603$code.=<<___;
604	it		eq
605	subeq		$aptr,$aptr,$num,lsl#2	@ rewind
606	vmlal.u32	@ACC[0],$Ni,${N0}[0]
607	vld1.32		{$Bi},[sp,:64]		@ pull smashed b[8*i+0]
608	vmlal.u32	@ACC[1],$Ni,${N0}[1]
609	vld1.32		{$A0-$A3},[$aptr]!
610	vmlal.u32	@ACC[2],$Ni,${N1}[0]
611	add		$bnptr,sp,#8		@ rewind
612	vmlal.u32	@ACC[3],$Ni,${N1}[1]
613	vmlal.u32	@ACC[4],$Ni,${N2}[0]
614	vmlal.u32	@ACC[5],$Ni,${N2}[1]
615	vmlal.u32	@ACC[6],$Ni,${N3}[0]
616	vst1.64		{@ACC[0]},[$toutptr,:128]!
617	vmlal.u32	@ACC[7],$Ni,${N3}[1]
618
619	bne		.LNEON_8n_inner
620___
621	push(@ACC,shift(@ACC));
622$code.=<<___;
623	add		$tinptr,sp,#128
624	vst1.64		{@ACC[0]-@ACC[1]},[$toutptr,:256]!
625	veor		q2,q2,q2		@ $N0-$N1
626	vst1.64		{@ACC[2]-@ACC[3]},[$toutptr,:256]!
627	veor		q3,q3,q3		@ $N2-$N3
628	vst1.64		{@ACC[4]-@ACC[5]},[$toutptr,:256]!
629	vst1.64		{@ACC[6]},[$toutptr,:128]
630
631	subs		$outer,$outer,#8
632	vld1.64		{@ACC[0]-@ACC[1]},[$tinptr,:256]!
633	vld1.64		{@ACC[2]-@ACC[3]},[$tinptr,:256]!
634	vld1.64		{@ACC[4]-@ACC[5]},[$tinptr,:256]!
635	vld1.64		{@ACC[6]-@ACC[7]},[$tinptr,:256]!
636
637	itt		ne
638	subne		$nptr,$nptr,$num,lsl#2	@ rewind
639	bne		.LNEON_8n_outer
640
641	add		$toutptr,sp,#128
642	vst1.64		{q2-q3}, [sp,:256]!	@ start wiping stack frame
643	vshr.u64	$temp,@ACC[0]#lo,#16
644	vst1.64		{q2-q3},[sp,:256]!
645	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
646	vst1.64		{q2-q3}, [sp,:256]!
647	vshr.u64	$temp,@ACC[0]#hi,#16
648	vst1.64		{q2-q3}, [sp,:256]!
649	vzip.16		@ACC[0]#lo,@ACC[0]#hi
650
651	mov		$inner,$num
652	b		.LNEON_tail_entry
653
654.align	4
655.LNEON_tail:
656	vadd.u64	@ACC[0]#lo,@ACC[0]#lo,$temp
657	vshr.u64	$temp,@ACC[0]#lo,#16
658	vld1.64		{@ACC[2]-@ACC[3]}, [$tinptr, :256]!
659	vadd.u64	@ACC[0]#hi,@ACC[0]#hi,$temp
660	vld1.64		{@ACC[4]-@ACC[5]}, [$tinptr, :256]!
661	vshr.u64	$temp,@ACC[0]#hi,#16
662	vld1.64		{@ACC[6]-@ACC[7]}, [$tinptr, :256]!
663	vzip.16		@ACC[0]#lo,@ACC[0]#hi
664
665.LNEON_tail_entry:
666___
667for ($i=1; $i<8; $i++) {
668$code.=<<___;
669	vadd.u64	@ACC[1]#lo,@ACC[1]#lo,$temp
670	vst1.32		{@ACC[0]#lo[0]}, [$toutptr, :32]!
671	vshr.u64	$temp,@ACC[1]#lo,#16
672	vadd.u64	@ACC[1]#hi,@ACC[1]#hi,$temp
673	vshr.u64	$temp,@ACC[1]#hi,#16
674	vzip.16		@ACC[1]#lo,@ACC[1]#hi
675___
676	push(@ACC,shift(@ACC));
677}
678	push(@ACC,shift(@ACC));
679$code.=<<___;
680	vld1.64		{@ACC[0]-@ACC[1]}, [$tinptr, :256]!
681	subs		$inner,$inner,#8
682	vst1.32		{@ACC[7]#lo[0]},   [$toutptr, :32]!
683	bne	.LNEON_tail
684
685	vst1.32	{${temp}[0]}, [$toutptr, :32]		@ top-most bit
686	sub	$nptr,$nptr,$num,lsl#2			@ rewind $nptr
687	subs	$aptr,sp,#0				@ clear carry flag
688	add	$bptr,sp,$num,lsl#2
689
690.LNEON_sub:
691	ldmia	$aptr!, {r4-r7}
692	ldmia	$nptr!, {r8-r11}
693	sbcs	r8, r4,r8
694	sbcs	r9, r5,r9
695	sbcs	r10,r6,r10
696	sbcs	r11,r7,r11
697	teq	$aptr,$bptr				@ preserves carry
698	stmia	$rptr!, {r8-r11}
699	bne	.LNEON_sub
700
701	ldr	r10, [$aptr]				@ load top-most bit
702	mov	r11,sp
703	veor	q0,q0,q0
704	sub	r11,$bptr,r11				@ this is num*4
705	veor	q1,q1,q1
706	mov	$aptr,sp
707	sub	$rptr,$rptr,r11				@ rewind $rptr
708	mov	$nptr,$bptr				@ second 3/4th of frame
709	sbcs	r10,r10,#0				@ result is carry flag
710
711.LNEON_copy_n_zap:
712	ldmia	$aptr!, {r4-r7}
713	ldmia	$rptr,  {r8-r11}
714	it	cc
715	movcc	r8, r4
716	vst1.64	{q0-q1}, [$nptr,:256]!			@ wipe
717	itt	cc
718	movcc	r9, r5
719	movcc	r10,r6
720	vst1.64	{q0-q1}, [$nptr,:256]!			@ wipe
721	it	cc
722	movcc	r11,r7
723	ldmia	$aptr, {r4-r7}
724	stmia	$rptr!, {r8-r11}
725	sub	$aptr,$aptr,#16
726	ldmia	$rptr, {r8-r11}
727	it	cc
728	movcc	r8, r4
729	vst1.64	{q0-q1}, [$aptr,:256]!			@ wipe
730	itt	cc
731	movcc	r9, r5
732	movcc	r10,r6
733	vst1.64	{q0-q1}, [$nptr,:256]!			@ wipe
734	it	cc
735	movcc	r11,r7
736	teq	$aptr,$bptr				@ preserves carry
737	stmia	$rptr!, {r8-r11}
738	bne	.LNEON_copy_n_zap
739
740	mov	sp,ip
741        vldmia  sp!,{d8-d15}
742        ldmia   sp!,{r4-r11}
743	ret						@ bx lr
744.size	bn_mul8x_mont_neon,.-bn_mul8x_mont_neon
745#endif
746___
747}
748$code.=<<___;
749.asciz	"Montgomery multiplication for ARMv4/NEON, CRYPTOGAMS by <appro\@openssl.org>"
750.align	2
751#if __ARM_MAX_ARCH__>=7
752.extern	OPENSSL_armcap_P
753#endif
754___
755
756foreach (split("\n",$code)) {
757	s/\`([^\`]*)\`/eval $1/ge;
758
759	s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/ge	or
760	s/\bret\b/bx    lr/g						or
761	s/\bbx\s+lr\b/.word\t0xe12fff1e/g;	# make it possible to compile with -march=armv4
762
763	print $_,"\n";
764}
765
766close STDOUT or die "error closing STDOUT: $!";
767