1#! /usr/bin/env perl
2# Copyright 2011-2016 The OpenSSL Project Authors. All Rights Reserved.
3#
4# Licensed under the OpenSSL license (the "License").  You may not use
5# this file except in compliance with the License.  You can obtain a copy
6# in the file LICENSE in the source distribution or at
7# https://www.openssl.org/source/license.html
8
9#
10# ====================================================================
11# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12# project. The module is, however, dual licensed under OpenSSL and
13# CRYPTOGAMS licenses depending on where you obtain it. For further
14# details see http://www.openssl.org/~appro/cryptogams/.
15# ====================================================================
16#
17# May 2011
18#
19# The module implements bn_GF2m_mul_2x2 polynomial multiplication
20# used in bn_gf2m.c. It's kind of low-hanging mechanical port from
21# C for the time being... Except that it has two code paths: pure
22# integer code suitable for any ARMv4 and later CPU and NEON code
23# suitable for ARMv7. Pure integer 1x1 multiplication subroutine runs
24# in ~45 cycles on dual-issue core such as Cortex A8, which is ~50%
25# faster than compiler-generated code. For ECDH and ECDSA verify (but
26# not for ECDSA sign) it means 25%-45% improvement depending on key
27# length, more for longer keys. Even though NEON 1x1 multiplication
28# runs in even less cycles, ~30, improvement is measurable only on
29# longer keys. One has to optimize code elsewhere to get NEON glow...
30#
31# April 2014
32#
33# Double bn_GF2m_mul_2x2 performance by using algorithm from paper
34# referred below, which improves ECDH and ECDSA verify benchmarks
35# by 18-40%.
36#
37# Câmara, D.; Gouvêa, C. P. L.; López, J. & Dahab, R.: Fast Software
38# Polynomial Multiplication on ARM Processors using the NEON Engine.
39#
40# http://conradoplg.cryptoland.net/files/2010/12/mocrysen13.pdf
41
42$flavour = shift;
43if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
44else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} }
45
46if ($flavour && $flavour ne "void") {
47    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
48    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
49    ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
50    die "can't locate arm-xlate.pl";
51
52    open STDOUT,"| \"$^X\" $xlate $flavour $output";
53} else {
54    open STDOUT,">$output";
55}
56
57$code=<<___;
58#include "arm_arch.h"
59
60.text
61#if defined(__thumb2__)
62.syntax	unified
63.thumb
64#else
65.code	32
66#endif
67___
68################
69# private interface to mul_1x1_ialu
70#
71$a="r1";
72$b="r0";
73
74($a0,$a1,$a2,$a12,$a4,$a14)=
75($hi,$lo,$t0,$t1, $i0,$i1 )=map("r$_",(4..9),12);
76
77$mask="r12";
78
79$code.=<<___;
80.type	mul_1x1_ialu,%function
81.align	5
82mul_1x1_ialu:
83	mov	$a0,#0
84	bic	$a1,$a,#3<<30		@ a1=a&0x3fffffff
85	str	$a0,[sp,#0]		@ tab[0]=0
86	add	$a2,$a1,$a1		@ a2=a1<<1
87	str	$a1,[sp,#4]		@ tab[1]=a1
88	eor	$a12,$a1,$a2		@ a1^a2
89	str	$a2,[sp,#8]		@ tab[2]=a2
90	mov	$a4,$a1,lsl#2		@ a4=a1<<2
91	str	$a12,[sp,#12]		@ tab[3]=a1^a2
92	eor	$a14,$a1,$a4		@ a1^a4
93	str	$a4,[sp,#16]		@ tab[4]=a4
94	eor	$a0,$a2,$a4		@ a2^a4
95	str	$a14,[sp,#20]		@ tab[5]=a1^a4
96	eor	$a12,$a12,$a4		@ a1^a2^a4
97	str	$a0,[sp,#24]		@ tab[6]=a2^a4
98	and	$i0,$mask,$b,lsl#2
99	str	$a12,[sp,#28]		@ tab[7]=a1^a2^a4
100
101	and	$i1,$mask,$b,lsr#1
102	ldr	$lo,[sp,$i0]		@ tab[b       & 0x7]
103	and	$i0,$mask,$b,lsr#4
104	ldr	$t1,[sp,$i1]		@ tab[b >>  3 & 0x7]
105	and	$i1,$mask,$b,lsr#7
106	ldr	$t0,[sp,$i0]		@ tab[b >>  6 & 0x7]
107	eor	$lo,$lo,$t1,lsl#3	@ stall
108	mov	$hi,$t1,lsr#29
109	ldr	$t1,[sp,$i1]		@ tab[b >>  9 & 0x7]
110
111	and	$i0,$mask,$b,lsr#10
112	eor	$lo,$lo,$t0,lsl#6
113	eor	$hi,$hi,$t0,lsr#26
114	ldr	$t0,[sp,$i0]		@ tab[b >> 12 & 0x7]
115
116	and	$i1,$mask,$b,lsr#13
117	eor	$lo,$lo,$t1,lsl#9
118	eor	$hi,$hi,$t1,lsr#23
119	ldr	$t1,[sp,$i1]		@ tab[b >> 15 & 0x7]
120
121	and	$i0,$mask,$b,lsr#16
122	eor	$lo,$lo,$t0,lsl#12
123	eor	$hi,$hi,$t0,lsr#20
124	ldr	$t0,[sp,$i0]		@ tab[b >> 18 & 0x7]
125
126	and	$i1,$mask,$b,lsr#19
127	eor	$lo,$lo,$t1,lsl#15
128	eor	$hi,$hi,$t1,lsr#17
129	ldr	$t1,[sp,$i1]		@ tab[b >> 21 & 0x7]
130
131	and	$i0,$mask,$b,lsr#22
132	eor	$lo,$lo,$t0,lsl#18
133	eor	$hi,$hi,$t0,lsr#14
134	ldr	$t0,[sp,$i0]		@ tab[b >> 24 & 0x7]
135
136	and	$i1,$mask,$b,lsr#25
137	eor	$lo,$lo,$t1,lsl#21
138	eor	$hi,$hi,$t1,lsr#11
139	ldr	$t1,[sp,$i1]		@ tab[b >> 27 & 0x7]
140
141	tst	$a,#1<<30
142	and	$i0,$mask,$b,lsr#28
143	eor	$lo,$lo,$t0,lsl#24
144	eor	$hi,$hi,$t0,lsr#8
145	ldr	$t0,[sp,$i0]		@ tab[b >> 30      ]
146
147#ifdef	__thumb2__
148	itt	ne
149#endif
150	eorne	$lo,$lo,$b,lsl#30
151	eorne	$hi,$hi,$b,lsr#2
152	tst	$a,#1<<31
153	eor	$lo,$lo,$t1,lsl#27
154	eor	$hi,$hi,$t1,lsr#5
155#ifdef	__thumb2__
156	itt	ne
157#endif
158	eorne	$lo,$lo,$b,lsl#31
159	eorne	$hi,$hi,$b,lsr#1
160	eor	$lo,$lo,$t0,lsl#30
161	eor	$hi,$hi,$t0,lsr#2
162
163	mov	pc,lr
164.size	mul_1x1_ialu,.-mul_1x1_ialu
165___
166################
167# void	bn_GF2m_mul_2x2(BN_ULONG *r,
168#	BN_ULONG a1,BN_ULONG a0,
169#	BN_ULONG b1,BN_ULONG b0);	# r[3..0]=a1a0·b1b0
170{
171$code.=<<___;
172.global	bn_GF2m_mul_2x2
173.type	bn_GF2m_mul_2x2,%function
174.align	5
175bn_GF2m_mul_2x2:
176#if __ARM_MAX_ARCH__>=7
177	stmdb	sp!,{r10,lr}
178	ldr	r12,.LOPENSSL_armcap
179	adr	r10,.LOPENSSL_armcap
180	ldr	r12,[r12,r10]
181#ifdef	__APPLE__
182	ldr	r12,[r12]
183#endif
184	tst	r12,#ARMV7_NEON
185	itt	ne
186	ldrne	r10,[sp],#8
187	bne	.LNEON
188	stmdb	sp!,{r4-r9}
189#else
190	stmdb	sp!,{r4-r10,lr}
191#endif
192___
193$ret="r10";	# reassigned 1st argument
194$code.=<<___;
195	mov	$ret,r0			@ reassign 1st argument
196	mov	$b,r3			@ $b=b1
197	sub	r7,sp,#36
198	mov	r8,sp
199	and	r7,r7,#-32
200	ldr	r3,[sp,#32]		@ load b0
201	mov	$mask,#7<<2
202	mov	sp,r7			@ allocate tab[8]
203	str	r8,[r7,#32]
204
205	bl	mul_1x1_ialu		@ a1·b1
206	str	$lo,[$ret,#8]
207	str	$hi,[$ret,#12]
208
209	eor	$b,$b,r3		@ flip b0 and b1
210	 eor	$a,$a,r2		@ flip a0 and a1
211	eor	r3,r3,$b
212	 eor	r2,r2,$a
213	eor	$b,$b,r3
214	 eor	$a,$a,r2
215	bl	mul_1x1_ialu		@ a0·b0
216	str	$lo,[$ret]
217	str	$hi,[$ret,#4]
218
219	eor	$a,$a,r2
220	eor	$b,$b,r3
221	bl	mul_1x1_ialu		@ (a1+a0)·(b1+b0)
222___
223@r=map("r$_",(6..9));
224$code.=<<___;
225	ldmia	$ret,{@r[0]-@r[3]}
226	eor	$lo,$lo,$hi
227	ldr	sp,[sp,#32]		@ destroy tab[8]
228	eor	$hi,$hi,@r[1]
229	eor	$lo,$lo,@r[0]
230	eor	$hi,$hi,@r[2]
231	eor	$lo,$lo,@r[3]
232	eor	$hi,$hi,@r[3]
233	str	$hi,[$ret,#8]
234	eor	$lo,$lo,$hi
235	str	$lo,[$ret,#4]
236
237#if __ARM_ARCH__>=5
238	ldmia	sp!,{r4-r10,pc}
239#else
240	ldmia	sp!,{r4-r10,lr}
241	tst	lr,#1
242	moveq	pc,lr			@ be binary compatible with V4, yet
243	bx	lr			@ interoperable with Thumb ISA:-)
244#endif
245___
246}
247{
248my ($r,$t0,$t1,$t2,$t3)=map("q$_",(0..3,8..12));
249my ($a,$b,$k48,$k32,$k16)=map("d$_",(26..31));
250
251$code.=<<___;
252#if __ARM_MAX_ARCH__>=7
253.arch	armv7-a
254.fpu	neon
255
256.align	5
257.LNEON:
258	ldr		r12, [sp]		@ 5th argument
259	vmov		$a, r2, r1
260	vmov		$b, r12, r3
261	vmov.i64	$k48, #0x0000ffffffffffff
262	vmov.i64	$k32, #0x00000000ffffffff
263	vmov.i64	$k16, #0x000000000000ffff
264
265	vext.8		$t0#lo, $a, $a, #1	@ A1
266	vmull.p8	$t0, $t0#lo, $b		@ F = A1*B
267	vext.8		$r#lo, $b, $b, #1	@ B1
268	vmull.p8	$r, $a, $r#lo		@ E = A*B1
269	vext.8		$t1#lo, $a, $a, #2	@ A2
270	vmull.p8	$t1, $t1#lo, $b		@ H = A2*B
271	vext.8		$t3#lo, $b, $b, #2	@ B2
272	vmull.p8	$t3, $a, $t3#lo		@ G = A*B2
273	vext.8		$t2#lo, $a, $a, #3	@ A3
274	veor		$t0, $t0, $r		@ L = E + F
275	vmull.p8	$t2, $t2#lo, $b		@ J = A3*B
276	vext.8		$r#lo, $b, $b, #3	@ B3
277	veor		$t1, $t1, $t3		@ M = G + H
278	vmull.p8	$r, $a, $r#lo		@ I = A*B3
279	veor		$t0#lo, $t0#lo, $t0#hi	@ t0 = (L) (P0 + P1) << 8
280	vand		$t0#hi, $t0#hi, $k48
281	vext.8		$t3#lo, $b, $b, #4	@ B4
282	veor		$t1#lo, $t1#lo, $t1#hi	@ t1 = (M) (P2 + P3) << 16
283	vand		$t1#hi, $t1#hi, $k32
284	vmull.p8	$t3, $a, $t3#lo		@ K = A*B4
285	veor		$t2, $t2, $r		@ N = I + J
286	veor		$t0#lo, $t0#lo, $t0#hi
287	veor		$t1#lo, $t1#lo, $t1#hi
288	veor		$t2#lo, $t2#lo, $t2#hi	@ t2 = (N) (P4 + P5) << 24
289	vand		$t2#hi, $t2#hi, $k16
290	vext.8		$t0, $t0, $t0, #15
291	veor		$t3#lo, $t3#lo, $t3#hi	@ t3 = (K) (P6 + P7) << 32
292	vmov.i64	$t3#hi, #0
293	vext.8		$t1, $t1, $t1, #14
294	veor		$t2#lo, $t2#lo, $t2#hi
295	vmull.p8	$r, $a, $b		@ D = A*B
296	vext.8		$t3, $t3, $t3, #12
297	vext.8		$t2, $t2, $t2, #13
298	veor		$t0, $t0, $t1
299	veor		$t2, $t2, $t3
300	veor		$r, $r, $t0
301	veor		$r, $r, $t2
302
303	vst1.32		{$r}, [r0]
304	ret		@ bx lr
305#endif
306___
307}
308$code.=<<___;
309.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
310#if __ARM_MAX_ARCH__>=7
311.align	5
312.LOPENSSL_armcap:
313.word	OPENSSL_armcap_P-.
314#endif
315.asciz	"GF(2^m) Multiplication for ARMv4/NEON, CRYPTOGAMS by <appro\@openssl.org>"
316.align	5
317
318#if __ARM_MAX_ARCH__>=7
319.comm	OPENSSL_armcap_P,4,4
320#endif
321___
322
323foreach (split("\n",$code)) {
324	s/\`([^\`]*)\`/eval $1/geo;
325
326	s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo	or
327	s/\bret\b/bx	lr/go		or
328	s/\bbx\s+lr\b/.word\t0xe12fff1e/go;    # make it possible to compile with -march=armv4
329
330	print $_,"\n";
331}
332close STDOUT;   # enforce flush
333