1#! /usr/bin/env perl
2# Copyright 2014-2016 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# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
11# project. The module is, however, dual licensed under OpenSSL and
12# CRYPTOGAMS licenses depending on where you obtain it. For further
13# details see http://www.openssl.org/~appro/cryptogams/.
14#
15# Permission to use under GPLv2 terms is granted.
16# ====================================================================
17#
18# SHA256/512 for ARMv8.
19#
20# Performance in cycles per processed byte and improvement coefficient
21# over code generated with "default" compiler:
22#
23#		SHA256-hw	SHA256(*)	SHA512
24# Apple A7	1.97		10.5 (+33%)	6.73 (-1%(**))
25# Cortex-A53	2.38		15.5 (+115%)	10.0 (+150%(***))
26# Cortex-A57	2.31		11.6 (+86%)	7.51 (+260%(***))
27# Denver	2.01		10.5 (+26%)	6.70 (+8%)
28# X-Gene			20.0 (+100%)	12.8 (+300%(***))
29# Mongoose	2.36		13.0 (+50%)	8.36 (+33%)
30#
31# (*)	Software SHA256 results are of lesser relevance, presented
32#	mostly for informational purposes.
33# (**)	The result is a trade-off: it's possible to improve it by
34#	10% (or by 1 cycle per round), but at the cost of 20% loss
35#	on Cortex-A53 (or by 4 cycles per round).
36# (***)	Super-impressive coefficients over gcc-generated code are
37#	indication of some compiler "pathology", most notably code
38#	generated with -mgeneral-regs-only is significanty faster
39#	and the gap is only 40-90%.
40
41$output=pop;
42$flavour=pop;
43
44if ($flavour && $flavour ne "void") {
45    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
46    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
47    ( $xlate="${dir}../../../perlasm/arm-xlate.pl" and -f $xlate) or
48    die "can't locate arm-xlate.pl";
49
50    open OUT,"| \"$^X\" $xlate $flavour $output";
51    *STDOUT=*OUT;
52} else {
53    open OUT,">$output";
54    *STDOUT=*OUT;
55}
56
57if ($output =~ /sha512-armv8/) {
58	$BITS=512;
59	$SZ=8;
60	@Sigma0=(28,34,39);
61	@Sigma1=(14,18,41);
62	@sigma0=(1,  8, 7);
63	@sigma1=(19,61, 6);
64	$rounds=80;
65	$reg_t="x";
66} else {
67	$BITS=256;
68	$SZ=4;
69	@Sigma0=( 2,13,22);
70	@Sigma1=( 6,11,25);
71	@sigma0=( 7,18, 3);
72	@sigma1=(17,19,10);
73	$rounds=64;
74	$reg_t="w";
75}
76
77$func="GFp_sha${BITS}_block_data_order";
78
79($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
80
81@X=map("$reg_t$_",(3..15,0..2));
82@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
83($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
84
85sub BODY_00_xx {
86my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
87my $j=($i+1)&15;
88my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
89   $T0=@X[$i+3] if ($i<11);
90
91$code.=<<___	if ($i<16);
92#ifndef	__ARMEB__
93	rev	@X[$i],@X[$i]			// $i
94#endif
95___
96$code.=<<___	if ($i<13 && ($i&1));
97	ldp	@X[$i+1],@X[$i+2],[$inp],#2*$SZ
98___
99$code.=<<___	if ($i==13);
100	ldp	@X[14],@X[15],[$inp]
101___
102$code.=<<___	if ($i>=14);
103	ldr	@X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
104___
105$code.=<<___	if ($i>0 && $i<16);
106	add	$a,$a,$t1			// h+=Sigma0(a)
107___
108$code.=<<___	if ($i>=11);
109	str	@X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
110___
111# While ARMv8 specifies merged rotate-n-logical operation such as
112# 'eor x,y,z,ror#n', it was found to negatively affect performance
113# on Apple A7. The reason seems to be that it requires even 'y' to
114# be available earlier. This means that such merged instruction is
115# not necessarily best choice on critical path... On the other hand
116# Cortex-A5x handles merged instructions much better than disjoint
117# rotate and logical... See (**) footnote above.
118$code.=<<___	if ($i<15);
119	ror	$t0,$e,#$Sigma1[0]
120	add	$h,$h,$t2			// h+=K[i]
121	eor	$T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
122	and	$t1,$f,$e
123	bic	$t2,$g,$e
124	add	$h,$h,@X[$i&15]			// h+=X[i]
125	orr	$t1,$t1,$t2			// Ch(e,f,g)
126	eor	$t2,$a,$b			// a^b, b^c in next round
127	eor	$t0,$t0,$T0,ror#$Sigma1[1]	// Sigma1(e)
128	ror	$T0,$a,#$Sigma0[0]
129	add	$h,$h,$t1			// h+=Ch(e,f,g)
130	eor	$t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
131	add	$h,$h,$t0			// h+=Sigma1(e)
132	and	$t3,$t3,$t2			// (b^c)&=(a^b)
133	add	$d,$d,$h			// d+=h
134	eor	$t3,$t3,$b			// Maj(a,b,c)
135	eor	$t1,$T0,$t1,ror#$Sigma0[1]	// Sigma0(a)
136	add	$h,$h,$t3			// h+=Maj(a,b,c)
137	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
138	//add	$h,$h,$t1			// h+=Sigma0(a)
139___
140$code.=<<___	if ($i>=15);
141	ror	$t0,$e,#$Sigma1[0]
142	add	$h,$h,$t2			// h+=K[i]
143	ror	$T1,@X[($j+1)&15],#$sigma0[0]
144	and	$t1,$f,$e
145	ror	$T2,@X[($j+14)&15],#$sigma1[0]
146	bic	$t2,$g,$e
147	ror	$T0,$a,#$Sigma0[0]
148	add	$h,$h,@X[$i&15]			// h+=X[i]
149	eor	$t0,$t0,$e,ror#$Sigma1[1]
150	eor	$T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
151	orr	$t1,$t1,$t2			// Ch(e,f,g)
152	eor	$t2,$a,$b			// a^b, b^c in next round
153	eor	$t0,$t0,$e,ror#$Sigma1[2]	// Sigma1(e)
154	eor	$T0,$T0,$a,ror#$Sigma0[1]
155	add	$h,$h,$t1			// h+=Ch(e,f,g)
156	and	$t3,$t3,$t2			// (b^c)&=(a^b)
157	eor	$T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
158	eor	$T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]	// sigma0(X[i+1])
159	add	$h,$h,$t0			// h+=Sigma1(e)
160	eor	$t3,$t3,$b			// Maj(a,b,c)
161	eor	$t1,$T0,$a,ror#$Sigma0[2]	// Sigma0(a)
162	eor	$T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]	// sigma1(X[i+14])
163	add	@X[$j],@X[$j],@X[($j+9)&15]
164	add	$d,$d,$h			// d+=h
165	add	$h,$h,$t3			// h+=Maj(a,b,c)
166	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
167	add	@X[$j],@X[$j],$T1
168	add	$h,$h,$t1			// h+=Sigma0(a)
169	add	@X[$j],@X[$j],$T2
170___
171	($t2,$t3)=($t3,$t2);
172}
173
174$code.=<<___;
175#ifndef	__KERNEL__
176# include <GFp/arm_arch.h>
177#endif
178
179.text
180
181.extern	GFp_armcap_P
182.hidden GFp_armcap_P
183.globl	$func
184.type	$func,%function
185.align	6
186$func:
187___
188$code.=<<___	if ($SZ==4);
189	AARCH64_VALID_CALL_TARGET
190#ifndef	__KERNEL__
191#if __has_feature(hwaddress_sanitizer) && __clang_major__ >= 10
192	adrp	x16,:pg_hi21_nc:GFp_armcap_P
193#else
194	adrp	x16,:pg_hi21:GFp_armcap_P
195#endif
196	ldr	w16,[x16,:lo12:GFp_armcap_P]
197	tst	w16,#ARMV8_SHA256
198	b.ne	.Lv8_entry
199#endif
200___
201$code.=<<___;
202	AARCH64_SIGN_LINK_REGISTER
203	stp	x29,x30,[sp,#-128]!
204	add	x29,sp,#0
205
206	stp	x19,x20,[sp,#16]
207	stp	x21,x22,[sp,#32]
208	stp	x23,x24,[sp,#48]
209	stp	x25,x26,[sp,#64]
210	stp	x27,x28,[sp,#80]
211	sub	sp,sp,#4*$SZ
212
213	ldp	$A,$B,[$ctx]				// load context
214	ldp	$C,$D,[$ctx,#2*$SZ]
215	ldp	$E,$F,[$ctx,#4*$SZ]
216	add	$num,$inp,$num,lsl#`log(16*$SZ)/log(2)`	// end of input
217	ldp	$G,$H,[$ctx,#6*$SZ]
218	adrp	$Ktbl,:pg_hi21:.LK$BITS
219	add	$Ktbl,$Ktbl,:lo12:.LK$BITS
220	stp	$ctx,$num,[x29,#96]
221
222.Loop:
223	ldp	@X[0],@X[1],[$inp],#2*$SZ
224	ldr	$t2,[$Ktbl],#$SZ			// *K++
225	eor	$t3,$B,$C				// magic seed
226	str	$inp,[x29,#112]
227___
228for ($i=0;$i<16;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
229$code.=".Loop_16_xx:\n";
230for (;$i<32;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
231$code.=<<___;
232	cbnz	$t2,.Loop_16_xx
233
234	ldp	$ctx,$num,[x29,#96]
235	ldr	$inp,[x29,#112]
236	sub	$Ktbl,$Ktbl,#`$SZ*($rounds+1)`		// rewind
237
238	ldp	@X[0],@X[1],[$ctx]
239	ldp	@X[2],@X[3],[$ctx,#2*$SZ]
240	add	$inp,$inp,#14*$SZ			// advance input pointer
241	ldp	@X[4],@X[5],[$ctx,#4*$SZ]
242	add	$A,$A,@X[0]
243	ldp	@X[6],@X[7],[$ctx,#6*$SZ]
244	add	$B,$B,@X[1]
245	add	$C,$C,@X[2]
246	add	$D,$D,@X[3]
247	stp	$A,$B,[$ctx]
248	add	$E,$E,@X[4]
249	add	$F,$F,@X[5]
250	stp	$C,$D,[$ctx,#2*$SZ]
251	add	$G,$G,@X[6]
252	add	$H,$H,@X[7]
253	cmp	$inp,$num
254	stp	$E,$F,[$ctx,#4*$SZ]
255	stp	$G,$H,[$ctx,#6*$SZ]
256	b.ne	.Loop
257
258	ldp	x19,x20,[x29,#16]
259	add	sp,sp,#4*$SZ
260	ldp	x21,x22,[x29,#32]
261	ldp	x23,x24,[x29,#48]
262	ldp	x25,x26,[x29,#64]
263	ldp	x27,x28,[x29,#80]
264	ldp	x29,x30,[sp],#128
265	AARCH64_VALIDATE_LINK_REGISTER
266	ret
267.size	$func,.-$func
268
269.section .rodata
270.align	6
271.type	.LK$BITS,%object
272.LK$BITS:
273___
274$code.=<<___ if ($SZ==8);
275	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
276	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
277	.quad	0x3956c25bf348b538,0x59f111f1b605d019
278	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
279	.quad	0xd807aa98a3030242,0x12835b0145706fbe
280	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
281	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
282	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
283	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
284	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
285	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
286	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
287	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
288	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
289	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
290	.quad	0x06ca6351e003826f,0x142929670a0e6e70
291	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
292	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
293	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
294	.quad	0x81c2c92e47edaee6,0x92722c851482353b
295	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
296	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
297	.quad	0xd192e819d6ef5218,0xd69906245565a910
298	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
299	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
300	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
301	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
302	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
303	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
304	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
305	.quad	0x90befffa23631e28,0xa4506cebde82bde9
306	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
307	.quad	0xca273eceea26619c,0xd186b8c721c0c207
308	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
309	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
310	.quad	0x113f9804bef90dae,0x1b710b35131c471b
311	.quad	0x28db77f523047d84,0x32caab7b40c72493
312	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
313	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
314	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
315	.quad	0	// terminator
316___
317$code.=<<___ if ($SZ==4);
318	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
319	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
320	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
321	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
322	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
323	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
324	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
325	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
326	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
327	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
328	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
329	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
330	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
331	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
332	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
333	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
334	.long	0	//terminator
335___
336$code.=<<___;
337.size	.LK$BITS,.-.LK$BITS
338.asciz	"SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
339.align	2
340___
341
342if ($SZ==4) {
343my $Ktbl="x3";
344
345my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
346my @MSG=map("v$_.16b",(4..7));
347my ($W0,$W1)=("v16.4s","v17.4s");
348my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
349
350$code.=<<___;
351.text
352#ifndef	__KERNEL__
353.type	sha256_block_armv8,%function
354.align	6
355sha256_block_armv8:
356.Lv8_entry:
357	// Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later.
358	stp		x29,x30,[sp,#-16]!
359	add		x29,sp,#0
360
361	ld1.32		{$ABCD,$EFGH},[$ctx]
362	adrp		$Ktbl,:pg_hi21:.LK256
363	add		$Ktbl,$Ktbl,:lo12:.LK256
364
365.Loop_hw:
366	ld1		{@MSG[0]-@MSG[3]},[$inp],#64
367	sub		$num,$num,#1
368	ld1.32		{$W0},[$Ktbl],#16
369	rev32		@MSG[0],@MSG[0]
370	rev32		@MSG[1],@MSG[1]
371	rev32		@MSG[2],@MSG[2]
372	rev32		@MSG[3],@MSG[3]
373	orr		$ABCD_SAVE,$ABCD,$ABCD		// offload
374	orr		$EFGH_SAVE,$EFGH,$EFGH
375___
376for($i=0;$i<12;$i++) {
377$code.=<<___;
378	ld1.32		{$W1},[$Ktbl],#16
379	add.i32		$W0,$W0,@MSG[0]
380	sha256su0	@MSG[0],@MSG[1]
381	orr		$abcd,$ABCD,$ABCD
382	sha256h		$ABCD,$EFGH,$W0
383	sha256h2	$EFGH,$abcd,$W0
384	sha256su1	@MSG[0],@MSG[2],@MSG[3]
385___
386	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
387}
388$code.=<<___;
389	ld1.32		{$W1},[$Ktbl],#16
390	add.i32		$W0,$W0,@MSG[0]
391	orr		$abcd,$ABCD,$ABCD
392	sha256h		$ABCD,$EFGH,$W0
393	sha256h2	$EFGH,$abcd,$W0
394
395	ld1.32		{$W0},[$Ktbl],#16
396	add.i32		$W1,$W1,@MSG[1]
397	orr		$abcd,$ABCD,$ABCD
398	sha256h		$ABCD,$EFGH,$W1
399	sha256h2	$EFGH,$abcd,$W1
400
401	ld1.32		{$W1},[$Ktbl]
402	add.i32		$W0,$W0,@MSG[2]
403	sub		$Ktbl,$Ktbl,#$rounds*$SZ-16	// rewind
404	orr		$abcd,$ABCD,$ABCD
405	sha256h		$ABCD,$EFGH,$W0
406	sha256h2	$EFGH,$abcd,$W0
407
408	add.i32		$W1,$W1,@MSG[3]
409	orr		$abcd,$ABCD,$ABCD
410	sha256h		$ABCD,$EFGH,$W1
411	sha256h2	$EFGH,$abcd,$W1
412
413	add.i32		$ABCD,$ABCD,$ABCD_SAVE
414	add.i32		$EFGH,$EFGH,$EFGH_SAVE
415
416	cbnz		$num,.Loop_hw
417
418	st1.32		{$ABCD,$EFGH},[$ctx]
419
420	ldr		x29,[sp],#16
421	ret
422.size	sha256_block_armv8,.-sha256_block_armv8
423#endif
424___
425}
426
427{   my  %opcode = (
428	"sha256h"	=> 0x5e004000,	"sha256h2"	=> 0x5e005000,
429	"sha256su0"	=> 0x5e282800,	"sha256su1"	=> 0x5e006000	);
430
431    sub unsha256 {
432	my ($mnemonic,$arg)=@_;
433
434	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
435	&&
436	sprintf ".inst\t0x%08x\t//%s %s",
437			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
438			$mnemonic,$arg;
439    }
440}
441
442open SELF,$0;
443while(<SELF>) {
444        next if (/^#!/);
445        last if (!s/^#/\/\// and !/^$/);
446        print;
447}
448close SELF;
449
450foreach(split("\n",$code)) {
451
452	s/\`([^\`]*)\`/eval($1)/geo;
453
454	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
455
456	s/\.\w?32\b//o		and s/\.16b/\.4s/go;
457	m/(ld|st)1[^\[]+\[0\]/o	and s/\.4s/\.s/go;
458
459	print $_,"\n";
460}
461
462close STDOUT or die "error closing STDOUT";
463