xref: /openssl/crypto/sha/asm/sha512-armv8.pl (revision 19e277dd)
1#! /usr/bin/env perl
2# Copyright 2014-2020 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# 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# Kryo		1.92		17.4 (+30%)	11.2 (+8%)
31# ThunderX2	2.54		13.2 (+40%)	8.40 (+18%)
32#
33# (*)	Software SHA256 results are of lesser relevance, presented
34#	mostly for informational purposes.
35# (**)	The result is a trade-off: it's possible to improve it by
36#	10% (or by 1 cycle per round), but at the cost of 20% loss
37#	on Cortex-A53 (or by 4 cycles per round).
38# (***)	Super-impressive coefficients over gcc-generated code are
39#	indication of some compiler "pathology", most notably code
40#	generated with -mgeneral-regs-only is significantly faster
41#	and the gap is only 40-90%.
42#
43# October 2016.
44#
45# Originally it was reckoned that it makes no sense to implement NEON
46# version of SHA256 for 64-bit processors. This is because performance
47# improvement on most wide-spread Cortex-A5x processors was observed
48# to be marginal, same on Cortex-A53 and ~10% on A57. But then it was
49# observed that 32-bit NEON SHA256 performs significantly better than
50# 64-bit scalar version on *some* of the more recent processors. As
51# result 64-bit NEON version of SHA256 was added to provide best
52# all-round performance. For example it executes ~30% faster on X-Gene
53# and Mongoose. [For reference, NEON version of SHA512 is bound to
54# deliver much less improvement, likely *negative* on Cortex-A5x.
55# Which is why NEON support is limited to SHA256.]
56
57# $output is the last argument if it looks like a file (it has an extension)
58# $flavour is the first argument if it doesn't look like a file
59$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
60$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 $xlate) or
66    die "can't locate arm-xlate.pl";
67
68    open OUT,"| \"$^X\" $xlate $flavour \"$output\""
69        or die "can't call $xlate: $!";
70    *STDOUT=*OUT;
71} else {
72    $output and open STDOUT,">$output";
73}
74
75if ($output =~ /512/) {
76	$BITS=512;
77	$SZ=8;
78	@Sigma0=(28,34,39);
79	@Sigma1=(14,18,41);
80	@sigma0=(1,  8, 7);
81	@sigma1=(19,61, 6);
82	$rounds=80;
83	$reg_t="x";
84} else {
85	$BITS=256;
86	$SZ=4;
87	@Sigma0=( 2,13,22);
88	@Sigma1=( 6,11,25);
89	@sigma0=( 7,18, 3);
90	@sigma1=(17,19,10);
91	$rounds=64;
92	$reg_t="w";
93}
94
95$func="sha${BITS}_block_data_order";
96
97($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
98
99@X=map("$reg_t$_",(3..15,0..2));
100@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
101($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
102
103sub BODY_00_xx {
104my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
105my $j=($i+1)&15;
106my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
107   $T0=@X[$i+3] if ($i<11);
108
109$code.=<<___	if ($i<16);
110#ifndef	__AARCH64EB__
111	rev	@X[$i],@X[$i]			// $i
112#endif
113___
114$code.=<<___	if ($i<13 && ($i&1));
115	ldp	@X[$i+1],@X[$i+2],[$inp],#2*$SZ
116___
117$code.=<<___	if ($i==13);
118	ldp	@X[14],@X[15],[$inp]
119___
120$code.=<<___	if ($i>=14);
121	ldr	@X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
122___
123$code.=<<___	if ($i>0 && $i<16);
124	add	$a,$a,$t1			// h+=Sigma0(a)
125___
126$code.=<<___	if ($i>=11);
127	str	@X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
128___
129# While ARMv8 specifies merged rotate-n-logical operation such as
130# 'eor x,y,z,ror#n', it was found to negatively affect performance
131# on Apple A7. The reason seems to be that it requires even 'y' to
132# be available earlier. This means that such merged instruction is
133# not necessarily best choice on critical path... On the other hand
134# Cortex-A5x handles merged instructions much better than disjoint
135# rotate and logical... See (**) footnote above.
136$code.=<<___	if ($i<15);
137	ror	$t0,$e,#$Sigma1[0]
138	add	$h,$h,$t2			// h+=K[i]
139	eor	$T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
140	and	$t1,$f,$e
141	bic	$t2,$g,$e
142	add	$h,$h,@X[$i&15]			// h+=X[i]
143	orr	$t1,$t1,$t2			// Ch(e,f,g)
144	eor	$t2,$a,$b			// a^b, b^c in next round
145	eor	$t0,$t0,$T0,ror#$Sigma1[1]	// Sigma1(e)
146	ror	$T0,$a,#$Sigma0[0]
147	add	$h,$h,$t1			// h+=Ch(e,f,g)
148	eor	$t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
149	add	$h,$h,$t0			// h+=Sigma1(e)
150	and	$t3,$t3,$t2			// (b^c)&=(a^b)
151	add	$d,$d,$h			// d+=h
152	eor	$t3,$t3,$b			// Maj(a,b,c)
153	eor	$t1,$T0,$t1,ror#$Sigma0[1]	// Sigma0(a)
154	add	$h,$h,$t3			// h+=Maj(a,b,c)
155	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
156	//add	$h,$h,$t1			// h+=Sigma0(a)
157___
158$code.=<<___	if ($i>=15);
159	ror	$t0,$e,#$Sigma1[0]
160	add	$h,$h,$t2			// h+=K[i]
161	ror	$T1,@X[($j+1)&15],#$sigma0[0]
162	and	$t1,$f,$e
163	ror	$T2,@X[($j+14)&15],#$sigma1[0]
164	bic	$t2,$g,$e
165	ror	$T0,$a,#$Sigma0[0]
166	add	$h,$h,@X[$i&15]			// h+=X[i]
167	eor	$t0,$t0,$e,ror#$Sigma1[1]
168	eor	$T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
169	orr	$t1,$t1,$t2			// Ch(e,f,g)
170	eor	$t2,$a,$b			// a^b, b^c in next round
171	eor	$t0,$t0,$e,ror#$Sigma1[2]	// Sigma1(e)
172	eor	$T0,$T0,$a,ror#$Sigma0[1]
173	add	$h,$h,$t1			// h+=Ch(e,f,g)
174	and	$t3,$t3,$t2			// (b^c)&=(a^b)
175	eor	$T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
176	eor	$T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]	// sigma0(X[i+1])
177	add	$h,$h,$t0			// h+=Sigma1(e)
178	eor	$t3,$t3,$b			// Maj(a,b,c)
179	eor	$t1,$T0,$a,ror#$Sigma0[2]	// Sigma0(a)
180	eor	$T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]	// sigma1(X[i+14])
181	add	@X[$j],@X[$j],@X[($j+9)&15]
182	add	$d,$d,$h			// d+=h
183	add	$h,$h,$t3			// h+=Maj(a,b,c)
184	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
185	add	@X[$j],@X[$j],$T1
186	add	$h,$h,$t1			// h+=Sigma0(a)
187	add	@X[$j],@X[$j],$T2
188___
189	($t2,$t3)=($t3,$t2);
190}
191
192$code.=<<___;
193#include "arm_arch.h"
194#ifndef	__KERNEL__
195.extern	OPENSSL_armcap_P
196.hidden	OPENSSL_armcap_P
197#endif
198
199.text
200
201.globl	$func
202.type	$func,%function
203.align	6
204$func:
205	AARCH64_VALID_CALL_TARGET
206#ifndef	__KERNEL__
207	adrp	x16,OPENSSL_armcap_P
208	ldr	w16,[x16,#:lo12:OPENSSL_armcap_P]
209___
210$code.=<<___	if ($SZ==4);
211	tst	w16,#ARMV8_SHA256
212	b.ne	.Lv8_entry
213	tst	w16,#ARMV7_NEON
214	b.ne	.Lneon_entry
215___
216$code.=<<___	if ($SZ==8);
217	tst	w16,#ARMV8_SHA512
218	b.ne	.Lv8_entry
219___
220$code.=<<___;
221#endif
222	AARCH64_SIGN_LINK_REGISTER
223	stp	x29,x30,[sp,#-128]!
224	add	x29,sp,#0
225
226	stp	x19,x20,[sp,#16]
227	stp	x21,x22,[sp,#32]
228	stp	x23,x24,[sp,#48]
229	stp	x25,x26,[sp,#64]
230	stp	x27,x28,[sp,#80]
231	sub	sp,sp,#4*$SZ
232
233	ldp	$A,$B,[$ctx]				// load context
234	ldp	$C,$D,[$ctx,#2*$SZ]
235	ldp	$E,$F,[$ctx,#4*$SZ]
236	add	$num,$inp,$num,lsl#`log(16*$SZ)/log(2)`	// end of input
237	ldp	$G,$H,[$ctx,#6*$SZ]
238	adr	$Ktbl,.LK$BITS
239	stp	$ctx,$num,[x29,#96]
240
241.Loop:
242	ldp	@X[0],@X[1],[$inp],#2*$SZ
243	ldr	$t2,[$Ktbl],#$SZ			// *K++
244	eor	$t3,$B,$C				// magic seed
245	str	$inp,[x29,#112]
246___
247for ($i=0;$i<16;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
248$code.=".Loop_16_xx:\n";
249for (;$i<32;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
250$code.=<<___;
251	cbnz	$t2,.Loop_16_xx
252
253	ldp	$ctx,$num,[x29,#96]
254	ldr	$inp,[x29,#112]
255	sub	$Ktbl,$Ktbl,#`$SZ*($rounds+1)`		// rewind
256
257	ldp	@X[0],@X[1],[$ctx]
258	ldp	@X[2],@X[3],[$ctx,#2*$SZ]
259	add	$inp,$inp,#14*$SZ			// advance input pointer
260	ldp	@X[4],@X[5],[$ctx,#4*$SZ]
261	add	$A,$A,@X[0]
262	ldp	@X[6],@X[7],[$ctx,#6*$SZ]
263	add	$B,$B,@X[1]
264	add	$C,$C,@X[2]
265	add	$D,$D,@X[3]
266	stp	$A,$B,[$ctx]
267	add	$E,$E,@X[4]
268	add	$F,$F,@X[5]
269	stp	$C,$D,[$ctx,#2*$SZ]
270	add	$G,$G,@X[6]
271	add	$H,$H,@X[7]
272	cmp	$inp,$num
273	stp	$E,$F,[$ctx,#4*$SZ]
274	stp	$G,$H,[$ctx,#6*$SZ]
275	b.ne	.Loop
276
277	ldp	x19,x20,[x29,#16]
278	add	sp,sp,#4*$SZ
279	ldp	x21,x22,[x29,#32]
280	ldp	x23,x24,[x29,#48]
281	ldp	x25,x26,[x29,#64]
282	ldp	x27,x28,[x29,#80]
283	ldp	x29,x30,[sp],#128
284	AARCH64_VALIDATE_LINK_REGISTER
285	ret
286.size	$func,.-$func
287
288.align	6
289.type	.LK$BITS,%object
290.LK$BITS:
291___
292$code.=<<___ if ($SZ==8);
293	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
294	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
295	.quad	0x3956c25bf348b538,0x59f111f1b605d019
296	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
297	.quad	0xd807aa98a3030242,0x12835b0145706fbe
298	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
299	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
300	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
301	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
302	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
303	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
304	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
305	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
306	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
307	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
308	.quad	0x06ca6351e003826f,0x142929670a0e6e70
309	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
310	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
311	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
312	.quad	0x81c2c92e47edaee6,0x92722c851482353b
313	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
314	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
315	.quad	0xd192e819d6ef5218,0xd69906245565a910
316	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
317	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
318	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
319	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
320	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
321	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
322	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
323	.quad	0x90befffa23631e28,0xa4506cebde82bde9
324	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
325	.quad	0xca273eceea26619c,0xd186b8c721c0c207
326	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
327	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
328	.quad	0x113f9804bef90dae,0x1b710b35131c471b
329	.quad	0x28db77f523047d84,0x32caab7b40c72493
330	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
331	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
332	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
333	.quad	0	// terminator
334___
335$code.=<<___ if ($SZ==4);
336	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
337	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
338	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
339	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
340	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
341	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
342	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
343	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
344	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
345	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
346	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
347	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
348	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
349	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
350	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
351	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
352	.long	0	//terminator
353___
354$code.=<<___;
355.size	.LK$BITS,.-.LK$BITS
356.asciz	"SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
357.align	2
358___
359
360if ($SZ==4) {
361my $Ktbl="x3";
362
363my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
364my @MSG=map("v$_.16b",(4..7));
365my ($W0,$W1)=("v16.4s","v17.4s");
366my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
367
368$code.=<<___;
369#ifndef	__KERNEL__
370.type	sha256_block_armv8,%function
371.align	6
372sha256_block_armv8:
373.Lv8_entry:
374	// Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later.
375	stp		x29,x30,[sp,#-16]!
376	add		x29,sp,#0
377
378	ld1.32		{$ABCD,$EFGH},[$ctx]
379	adr		$Ktbl,.LK256
380
381.Loop_hw:
382	ld1		{@MSG[0]-@MSG[3]},[$inp],#64
383	sub		$num,$num,#1
384	ld1.32		{$W0},[$Ktbl],#16
385	rev32		@MSG[0],@MSG[0]
386	rev32		@MSG[1],@MSG[1]
387	rev32		@MSG[2],@MSG[2]
388	rev32		@MSG[3],@MSG[3]
389	orr		$ABCD_SAVE,$ABCD,$ABCD		// offload
390	orr		$EFGH_SAVE,$EFGH,$EFGH
391___
392for($i=0;$i<12;$i++) {
393$code.=<<___;
394	ld1.32		{$W1},[$Ktbl],#16
395	add.i32		$W0,$W0,@MSG[0]
396	sha256su0	@MSG[0],@MSG[1]
397	orr		$abcd,$ABCD,$ABCD
398	sha256h		$ABCD,$EFGH,$W0
399	sha256h2	$EFGH,$abcd,$W0
400	sha256su1	@MSG[0],@MSG[2],@MSG[3]
401___
402	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
403}
404$code.=<<___;
405	ld1.32		{$W1},[$Ktbl],#16
406	add.i32		$W0,$W0,@MSG[0]
407	orr		$abcd,$ABCD,$ABCD
408	sha256h		$ABCD,$EFGH,$W0
409	sha256h2	$EFGH,$abcd,$W0
410
411	ld1.32		{$W0},[$Ktbl],#16
412	add.i32		$W1,$W1,@MSG[1]
413	orr		$abcd,$ABCD,$ABCD
414	sha256h		$ABCD,$EFGH,$W1
415	sha256h2	$EFGH,$abcd,$W1
416
417	ld1.32		{$W1},[$Ktbl]
418	add.i32		$W0,$W0,@MSG[2]
419	sub		$Ktbl,$Ktbl,#$rounds*$SZ-16	// rewind
420	orr		$abcd,$ABCD,$ABCD
421	sha256h		$ABCD,$EFGH,$W0
422	sha256h2	$EFGH,$abcd,$W0
423
424	add.i32		$W1,$W1,@MSG[3]
425	orr		$abcd,$ABCD,$ABCD
426	sha256h		$ABCD,$EFGH,$W1
427	sha256h2	$EFGH,$abcd,$W1
428
429	add.i32		$ABCD,$ABCD,$ABCD_SAVE
430	add.i32		$EFGH,$EFGH,$EFGH_SAVE
431
432	cbnz		$num,.Loop_hw
433
434	st1.32		{$ABCD,$EFGH},[$ctx]
435
436	ldr		x29,[sp],#16
437	ret
438.size	sha256_block_armv8,.-sha256_block_armv8
439#endif
440___
441}
442
443if ($SZ==4) {	######################################### NEON stuff #
444# You'll surely note a lot of similarities with sha256-armv4 module,
445# and of course it's not a coincidence. sha256-armv4 was used as
446# initial template, but was adapted for ARMv8 instruction set and
447# extensively re-tuned for all-round performance.
448
449my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
450my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
451my $Ktbl="x16";
452my $Xfer="x17";
453my @X = map("q$_",(0..3));
454my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
455my $j=0;
456
457sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
458{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
459  my $arg = pop;
460    $arg = "#$arg" if ($arg*1 eq $arg);
461    $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
462}
463
464sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
465sub Dlo     { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
466sub Dhi     { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
467
468sub Xupdate()
469{ use integer;
470  my $body = shift;
471  my @insns = (&$body,&$body,&$body,&$body);
472  my ($a,$b,$c,$d,$e,$f,$g,$h);
473
474	&ext_8		($T0,@X[0],@X[1],4);	# X[1..4]
475	 eval(shift(@insns));
476	 eval(shift(@insns));
477	 eval(shift(@insns));
478	&ext_8		($T3,@X[2],@X[3],4);	# X[9..12]
479	 eval(shift(@insns));
480	 eval(shift(@insns));
481	&mov		(&Dscalar($T7),&Dhi(@X[3]));	# X[14..15]
482	 eval(shift(@insns));
483	 eval(shift(@insns));
484	&ushr_32	($T2,$T0,$sigma0[0]);
485	 eval(shift(@insns));
486	&ushr_32	($T1,$T0,$sigma0[2]);
487	 eval(shift(@insns));
488	&add_32 	(@X[0],@X[0],$T3);	# X[0..3] += X[9..12]
489	 eval(shift(@insns));
490	&sli_32		($T2,$T0,32-$sigma0[0]);
491	 eval(shift(@insns));
492	 eval(shift(@insns));
493	&ushr_32	($T3,$T0,$sigma0[1]);
494	 eval(shift(@insns));
495	 eval(shift(@insns));
496	&eor_8		($T1,$T1,$T2);
497	 eval(shift(@insns));
498	 eval(shift(@insns));
499	&sli_32		($T3,$T0,32-$sigma0[1]);
500	 eval(shift(@insns));
501	 eval(shift(@insns));
502	  &ushr_32	($T4,$T7,$sigma1[0]);
503	 eval(shift(@insns));
504	 eval(shift(@insns));
505	&eor_8		($T1,$T1,$T3);		# sigma0(X[1..4])
506	 eval(shift(@insns));
507	 eval(shift(@insns));
508	  &sli_32	($T4,$T7,32-$sigma1[0]);
509	 eval(shift(@insns));
510	 eval(shift(@insns));
511	  &ushr_32	($T5,$T7,$sigma1[2]);
512	 eval(shift(@insns));
513	 eval(shift(@insns));
514	  &ushr_32	($T3,$T7,$sigma1[1]);
515	 eval(shift(@insns));
516	 eval(shift(@insns));
517	&add_32		(@X[0],@X[0],$T1);	# X[0..3] += sigma0(X[1..4])
518	 eval(shift(@insns));
519	 eval(shift(@insns));
520	  &sli_u32	($T3,$T7,32-$sigma1[1]);
521	 eval(shift(@insns));
522	 eval(shift(@insns));
523	  &eor_8	($T5,$T5,$T4);
524	 eval(shift(@insns));
525	 eval(shift(@insns));
526	 eval(shift(@insns));
527	  &eor_8	($T5,$T5,$T3);		# sigma1(X[14..15])
528	 eval(shift(@insns));
529	 eval(shift(@insns));
530	 eval(shift(@insns));
531	&add_32		(@X[0],@X[0],$T5);	# X[0..1] += sigma1(X[14..15])
532	 eval(shift(@insns));
533	 eval(shift(@insns));
534	 eval(shift(@insns));
535	  &ushr_32	($T6,@X[0],$sigma1[0]);
536	 eval(shift(@insns));
537	  &ushr_32	($T7,@X[0],$sigma1[2]);
538	 eval(shift(@insns));
539	 eval(shift(@insns));
540	  &sli_32	($T6,@X[0],32-$sigma1[0]);
541	 eval(shift(@insns));
542	  &ushr_32	($T5,@X[0],$sigma1[1]);
543	 eval(shift(@insns));
544	 eval(shift(@insns));
545	  &eor_8	($T7,$T7,$T6);
546	 eval(shift(@insns));
547	 eval(shift(@insns));
548	  &sli_32	($T5,@X[0],32-$sigma1[1]);
549	 eval(shift(@insns));
550	 eval(shift(@insns));
551	&ld1_32		("{$T0}","[$Ktbl], #16");
552	 eval(shift(@insns));
553	  &eor_8	($T7,$T7,$T5);		# sigma1(X[16..17])
554	 eval(shift(@insns));
555	 eval(shift(@insns));
556	&eor_8		($T5,$T5,$T5);
557	 eval(shift(@insns));
558	 eval(shift(@insns));
559	&mov		(&Dhi($T5), &Dlo($T7));
560	 eval(shift(@insns));
561	 eval(shift(@insns));
562	 eval(shift(@insns));
563	&add_32		(@X[0],@X[0],$T5);	# X[2..3] += sigma1(X[16..17])
564	 eval(shift(@insns));
565	 eval(shift(@insns));
566	 eval(shift(@insns));
567	&add_32		($T0,$T0,@X[0]);
568	 while($#insns>=1) { eval(shift(@insns)); }
569	&st1_32		("{$T0}","[$Xfer], #16");
570	 eval(shift(@insns));
571
572	push(@X,shift(@X));		# "rotate" X[]
573}
574
575sub Xpreload()
576{ use integer;
577  my $body = shift;
578  my @insns = (&$body,&$body,&$body,&$body);
579  my ($a,$b,$c,$d,$e,$f,$g,$h);
580
581	 eval(shift(@insns));
582	 eval(shift(@insns));
583	&ld1_8		("{@X[0]}","[$inp],#16");
584	 eval(shift(@insns));
585	 eval(shift(@insns));
586	&ld1_32		("{$T0}","[$Ktbl],#16");
587	 eval(shift(@insns));
588	 eval(shift(@insns));
589	 eval(shift(@insns));
590	 eval(shift(@insns));
591	&rev32		(@X[0],@X[0]);
592	 eval(shift(@insns));
593	 eval(shift(@insns));
594	 eval(shift(@insns));
595	 eval(shift(@insns));
596	&add_32		($T0,$T0,@X[0]);
597	 foreach (@insns) { eval; }	# remaining instructions
598	&st1_32		("{$T0}","[$Xfer], #16");
599
600	push(@X,shift(@X));		# "rotate" X[]
601}
602
603sub body_00_15 () {
604	(
605	'($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
606	'&add	($h,$h,$t1)',			# h+=X[i]+K[i]
607	'&add	($a,$a,$t4);'.			# h+=Sigma0(a) from the past
608	'&and	($t1,$f,$e)',
609	'&bic	($t4,$g,$e)',
610	'&eor	($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
611	'&add	($a,$a,$t2)',			# h+=Maj(a,b,c) from the past
612	'&orr	($t1,$t1,$t4)',			# Ch(e,f,g)
613	'&eor	($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))',	# Sigma1(e)
614	'&eor	($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
615	'&add	($h,$h,$t1)',			# h+=Ch(e,f,g)
616	'&ror	($t0,$t0,"#$Sigma1[0]")',
617	'&eor	($t2,$a,$b)',			# a^b, b^c in next round
618	'&eor	($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))',	# Sigma0(a)
619	'&add	($h,$h,$t0)',			# h+=Sigma1(e)
620	'&ldr	($t1,sprintf "[sp,#%d]",4*(($j+1)&15))	if (($j&15)!=15);'.
621	'&ldr	($t1,"[$Ktbl]")				if ($j==15);'.
622	'&and	($t3,$t3,$t2)',			# (b^c)&=(a^b)
623	'&ror	($t4,$t4,"#$Sigma0[0]")',
624	'&add	($d,$d,$h)',			# d+=h
625	'&eor	($t3,$t3,$b)',			# Maj(a,b,c)
626	'$j++;	unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
627	)
628}
629
630$code.=<<___;
631#ifdef	__KERNEL__
632.globl	sha256_block_neon
633#endif
634.type	sha256_block_neon,%function
635.align	4
636sha256_block_neon:
637	AARCH64_VALID_CALL_TARGET
638.Lneon_entry:
639	// Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later
640	stp	x29, x30, [sp, #-16]!
641	mov	x29, sp
642	sub	sp,sp,#16*4
643
644	adr	$Ktbl,.LK256
645	add	$num,$inp,$num,lsl#6	// len to point at the end of inp
646
647	ld1.8	{@X[0]},[$inp], #16
648	ld1.8	{@X[1]},[$inp], #16
649	ld1.8	{@X[2]},[$inp], #16
650	ld1.8	{@X[3]},[$inp], #16
651	ld1.32	{$T0},[$Ktbl], #16
652	ld1.32	{$T1},[$Ktbl], #16
653	ld1.32	{$T2},[$Ktbl], #16
654	ld1.32	{$T3},[$Ktbl], #16
655	rev32	@X[0],@X[0]		// yes, even on
656	rev32	@X[1],@X[1]		// big-endian
657	rev32	@X[2],@X[2]
658	rev32	@X[3],@X[3]
659	mov	$Xfer,sp
660	add.32	$T0,$T0,@X[0]
661	add.32	$T1,$T1,@X[1]
662	add.32	$T2,$T2,@X[2]
663	st1.32	{$T0-$T1},[$Xfer], #32
664	add.32	$T3,$T3,@X[3]
665	st1.32	{$T2-$T3},[$Xfer]
666	sub	$Xfer,$Xfer,#32
667
668	ldp	$A,$B,[$ctx]
669	ldp	$C,$D,[$ctx,#8]
670	ldp	$E,$F,[$ctx,#16]
671	ldp	$G,$H,[$ctx,#24]
672	ldr	$t1,[sp,#0]
673	mov	$t2,wzr
674	eor	$t3,$B,$C
675	mov	$t4,wzr
676	b	.L_00_48
677
678.align	4
679.L_00_48:
680___
681	&Xupdate(\&body_00_15);
682	&Xupdate(\&body_00_15);
683	&Xupdate(\&body_00_15);
684	&Xupdate(\&body_00_15);
685$code.=<<___;
686	cmp	$t1,#0				// check for K256 terminator
687	ldr	$t1,[sp,#0]
688	sub	$Xfer,$Xfer,#64
689	bne	.L_00_48
690
691	sub	$Ktbl,$Ktbl,#256		// rewind $Ktbl
692	cmp	$inp,$num
693	mov	$Xfer, #64
694	csel	$Xfer, $Xfer, xzr, eq
695	sub	$inp,$inp,$Xfer			// avoid SEGV
696	mov	$Xfer,sp
697___
698	&Xpreload(\&body_00_15);
699	&Xpreload(\&body_00_15);
700	&Xpreload(\&body_00_15);
701	&Xpreload(\&body_00_15);
702$code.=<<___;
703	add	$A,$A,$t4			// h+=Sigma0(a) from the past
704	ldp	$t0,$t1,[$ctx,#0]
705	add	$A,$A,$t2			// h+=Maj(a,b,c) from the past
706	ldp	$t2,$t3,[$ctx,#8]
707	add	$A,$A,$t0			// accumulate
708	add	$B,$B,$t1
709	ldp	$t0,$t1,[$ctx,#16]
710	add	$C,$C,$t2
711	add	$D,$D,$t3
712	ldp	$t2,$t3,[$ctx,#24]
713	add	$E,$E,$t0
714	add	$F,$F,$t1
715	 ldr	$t1,[sp,#0]
716	stp	$A,$B,[$ctx,#0]
717	add	$G,$G,$t2
718	 mov	$t2,wzr
719	stp	$C,$D,[$ctx,#8]
720	add	$H,$H,$t3
721	stp	$E,$F,[$ctx,#16]
722	 eor	$t3,$B,$C
723	stp	$G,$H,[$ctx,#24]
724	 mov	$t4,wzr
725	 mov	$Xfer,sp
726	b.ne	.L_00_48
727
728	ldr	x29,[x29]
729	add	sp,sp,#16*4+16
730	ret
731.size	sha256_block_neon,.-sha256_block_neon
732___
733}
734
735if ($SZ==8) {
736my $Ktbl="x3";
737
738my @H = map("v$_.16b",(0..4));
739my ($fg,$de,$m9_10)=map("v$_.16b",(5..7));
740my @MSG=map("v$_.16b",(16..23));
741my ($W0,$W1)=("v24.2d","v25.2d");
742my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29));
743
744$code.=<<___;
745#ifndef	__KERNEL__
746.type	sha512_block_armv8,%function
747.align	6
748sha512_block_armv8:
749.Lv8_entry:
750	// Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later
751	stp		x29,x30,[sp,#-16]!
752	add		x29,sp,#0
753
754	ld1		{@MSG[0]-@MSG[3]},[$inp],#64	// load input
755	ld1		{@MSG[4]-@MSG[7]},[$inp],#64
756
757	ld1.64		{@H[0]-@H[3]},[$ctx]		// load context
758	adr		$Ktbl,.LK512
759
760	rev64		@MSG[0],@MSG[0]
761	rev64		@MSG[1],@MSG[1]
762	rev64		@MSG[2],@MSG[2]
763	rev64		@MSG[3],@MSG[3]
764	rev64		@MSG[4],@MSG[4]
765	rev64		@MSG[5],@MSG[5]
766	rev64		@MSG[6],@MSG[6]
767	rev64		@MSG[7],@MSG[7]
768	b		.Loop_hw
769
770.align	4
771.Loop_hw:
772	ld1.64		{$W0},[$Ktbl],#16
773	subs		$num,$num,#1
774	sub		x4,$inp,#128
775	orr		$AB,@H[0],@H[0]			// offload
776	orr		$CD,@H[1],@H[1]
777	orr		$EF,@H[2],@H[2]
778	orr		$GH,@H[3],@H[3]
779	csel		$inp,$inp,x4,ne			// conditional rewind
780___
781for($i=0;$i<32;$i++) {
782$code.=<<___;
783	add.i64		$W0,$W0,@MSG[0]
784	ld1.64		{$W1},[$Ktbl],#16
785	ext		$W0,$W0,$W0,#8
786	ext		$fg,@H[2],@H[3],#8
787	ext		$de,@H[1],@H[2],#8
788	add.i64		@H[3],@H[3],$W0			// "T1 + H + K512[i]"
789	 sha512su0	@MSG[0],@MSG[1]
790	 ext		$m9_10,@MSG[4],@MSG[5],#8
791	sha512h		@H[3],$fg,$de
792	 sha512su1	@MSG[0],@MSG[7],$m9_10
793	add.i64		@H[4],@H[1],@H[3]		// "D + T1"
794	sha512h2	@H[3],$H[1],@H[0]
795___
796	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
797	@H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
798}
799for(;$i<40;$i++) {
800$code.=<<___	if ($i<39);
801	ld1.64		{$W1},[$Ktbl],#16
802___
803$code.=<<___	if ($i==39);
804	sub		$Ktbl,$Ktbl,#$rounds*$SZ	// rewind
805___
806$code.=<<___;
807	add.i64		$W0,$W0,@MSG[0]
808	 ld1		{@MSG[0]},[$inp],#16		// load next input
809	ext		$W0,$W0,$W0,#8
810	ext		$fg,@H[2],@H[3],#8
811	ext		$de,@H[1],@H[2],#8
812	add.i64		@H[3],@H[3],$W0			// "T1 + H + K512[i]"
813	sha512h		@H[3],$fg,$de
814	 rev64		@MSG[0],@MSG[0]
815	add.i64		@H[4],@H[1],@H[3]		// "D + T1"
816	sha512h2	@H[3],$H[1],@H[0]
817___
818	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
819	@H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
820}
821$code.=<<___;
822	add.i64		@H[0],@H[0],$AB			// accumulate
823	add.i64		@H[1],@H[1],$CD
824	add.i64		@H[2],@H[2],$EF
825	add.i64		@H[3],@H[3],$GH
826
827	cbnz		$num,.Loop_hw
828
829	st1.64		{@H[0]-@H[3]},[$ctx]		// store context
830
831	ldr		x29,[sp],#16
832	ret
833.size	sha512_block_armv8,.-sha512_block_armv8
834#endif
835___
836}
837
838{   my  %opcode = (
839	"sha256h"	=> 0x5e004000,	"sha256h2"	=> 0x5e005000,
840	"sha256su0"	=> 0x5e282800,	"sha256su1"	=> 0x5e006000	);
841
842    sub unsha256 {
843	my ($mnemonic,$arg)=@_;
844
845	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
846	&&
847	sprintf ".inst\t0x%08x\t//%s %s",
848			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
849			$mnemonic,$arg;
850    }
851}
852
853{   my  %opcode = (
854	"sha512h"	=> 0xce608000,	"sha512h2"	=> 0xce608400,
855	"sha512su0"	=> 0xcec08000,	"sha512su1"	=> 0xce608800	);
856
857    sub unsha512 {
858	my ($mnemonic,$arg)=@_;
859
860	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
861	&&
862	sprintf ".inst\t0x%08x\t//%s %s",
863			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
864			$mnemonic,$arg;
865    }
866}
867
868open SELF,$0;
869while(<SELF>) {
870        next if (/^#!/);
871        last if (!s/^#/\/\// and !/^$/);
872        print;
873}
874close SELF;
875
876foreach(split("\n",$code)) {
877
878	s/\`([^\`]*)\`/eval($1)/ge;
879
880	s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge	or
881	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
882
883	s/\bq([0-9]+)\b/v$1.16b/g;		# old->new registers
884
885	s/\.[ui]?8(\s)/$1/;
886	s/\.\w?64\b//		and s/\.16b/\.2d/g	or
887	s/\.\w?32\b//		and s/\.16b/\.4s/g;
888	m/\bext\b/		and s/\.2d/\.16b/g	or
889	m/(ld|st)1[^\[]+\[0\]/	and s/\.4s/\.s/g;
890
891	print $_,"\n";
892}
893
894close STDOUT or die "error closing STDOUT: $!";
895