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#ifndef	__KERNEL__
194# include "arm_arch.h"
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#ifndef	__KERNEL__
206	adrp	x16,OPENSSL_armcap_P
207	ldr	w16,[x16,#:lo12:OPENSSL_armcap_P]
208___
209$code.=<<___	if ($SZ==4);
210	tst	w16,#ARMV8_SHA256
211	b.ne	.Lv8_entry
212	tst	w16,#ARMV7_NEON
213	b.ne	.Lneon_entry
214___
215$code.=<<___	if ($SZ==8);
216	tst	w16,#ARMV8_SHA512
217	b.ne	.Lv8_entry
218___
219$code.=<<___;
220#endif
221	.inst	0xd503233f				// paciasp
222	stp	x29,x30,[sp,#-128]!
223	add	x29,sp,#0
224
225	stp	x19,x20,[sp,#16]
226	stp	x21,x22,[sp,#32]
227	stp	x23,x24,[sp,#48]
228	stp	x25,x26,[sp,#64]
229	stp	x27,x28,[sp,#80]
230	sub	sp,sp,#4*$SZ
231
232	ldp	$A,$B,[$ctx]				// load context
233	ldp	$C,$D,[$ctx,#2*$SZ]
234	ldp	$E,$F,[$ctx,#4*$SZ]
235	add	$num,$inp,$num,lsl#`log(16*$SZ)/log(2)`	// end of input
236	ldp	$G,$H,[$ctx,#6*$SZ]
237	adrp	$Ktbl,.LK$BITS
238	add	$Ktbl,$Ktbl,:lo12:.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	.inst	0xd50323bf				// autiasp
285	ret
286.size	$func,.-$func
287
288.rodata
289
290.align	6
291.type	.LK$BITS,%object
292.LK$BITS:
293___
294$code.=<<___ if ($SZ==8);
295	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
296	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
297	.quad	0x3956c25bf348b538,0x59f111f1b605d019
298	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
299	.quad	0xd807aa98a3030242,0x12835b0145706fbe
300	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
301	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
302	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
303	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
304	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
305	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
306	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
307	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
308	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
309	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
310	.quad	0x06ca6351e003826f,0x142929670a0e6e70
311	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
312	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
313	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
314	.quad	0x81c2c92e47edaee6,0x92722c851482353b
315	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
316	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
317	.quad	0xd192e819d6ef5218,0xd69906245565a910
318	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
319	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
320	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
321	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
322	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
323	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
324	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
325	.quad	0x90befffa23631e28,0xa4506cebde82bde9
326	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
327	.quad	0xca273eceea26619c,0xd186b8c721c0c207
328	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
329	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
330	.quad	0x113f9804bef90dae,0x1b710b35131c471b
331	.quad	0x28db77f523047d84,0x32caab7b40c72493
332	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
333	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
334	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
335	.quad	0	// terminator
336___
337$code.=<<___ if ($SZ==4);
338	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
339	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
340	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
341	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
342	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
343	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
344	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
345	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
346	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
347	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
348	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
349	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
350	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
351	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
352	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
353	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
354	.long	0	//terminator
355___
356$code.=<<___;
357.size	.LK$BITS,.-.LK$BITS
358.asciz	"SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
359.align	2
360
361.text
362___
363
364if ($SZ==4) {
365my $Ktbl="x3";
366
367my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
368my @MSG=map("v$_.16b",(4..7));
369my ($W0,$W1)=("v16.4s","v17.4s");
370my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
371
372$code.=<<___;
373#ifndef	__KERNEL__
374.type	sha256_block_armv8,%function
375.align	6
376sha256_block_armv8:
377.Lv8_entry:
378	stp		x29,x30,[sp,#-16]!
379	add		x29,sp,#0
380
381	ld1.32		{$ABCD,$EFGH},[$ctx]
382	adrp		$Ktbl,.LK256
383	add		$Ktbl,$Ktbl,:lo12:.LK256
384
385.Loop_hw:
386	ld1		{@MSG[0]-@MSG[3]},[$inp],#64
387	sub		$num,$num,#1
388	ld1.32		{$W0},[$Ktbl],#16
389	rev32		@MSG[0],@MSG[0]
390	rev32		@MSG[1],@MSG[1]
391	rev32		@MSG[2],@MSG[2]
392	rev32		@MSG[3],@MSG[3]
393	orr		$ABCD_SAVE,$ABCD,$ABCD		// offload
394	orr		$EFGH_SAVE,$EFGH,$EFGH
395___
396for($i=0;$i<12;$i++) {
397$code.=<<___;
398	ld1.32		{$W1},[$Ktbl],#16
399	add.i32		$W0,$W0,@MSG[0]
400	sha256su0	@MSG[0],@MSG[1]
401	orr		$abcd,$ABCD,$ABCD
402	sha256h		$ABCD,$EFGH,$W0
403	sha256h2	$EFGH,$abcd,$W0
404	sha256su1	@MSG[0],@MSG[2],@MSG[3]
405___
406	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
407}
408$code.=<<___;
409	ld1.32		{$W1},[$Ktbl],#16
410	add.i32		$W0,$W0,@MSG[0]
411	orr		$abcd,$ABCD,$ABCD
412	sha256h		$ABCD,$EFGH,$W0
413	sha256h2	$EFGH,$abcd,$W0
414
415	ld1.32		{$W0},[$Ktbl],#16
416	add.i32		$W1,$W1,@MSG[1]
417	orr		$abcd,$ABCD,$ABCD
418	sha256h		$ABCD,$EFGH,$W1
419	sha256h2	$EFGH,$abcd,$W1
420
421	ld1.32		{$W1},[$Ktbl]
422	add.i32		$W0,$W0,@MSG[2]
423	sub		$Ktbl,$Ktbl,#$rounds*$SZ-16	// rewind
424	orr		$abcd,$ABCD,$ABCD
425	sha256h		$ABCD,$EFGH,$W0
426	sha256h2	$EFGH,$abcd,$W0
427
428	add.i32		$W1,$W1,@MSG[3]
429	orr		$abcd,$ABCD,$ABCD
430	sha256h		$ABCD,$EFGH,$W1
431	sha256h2	$EFGH,$abcd,$W1
432
433	add.i32		$ABCD,$ABCD,$ABCD_SAVE
434	add.i32		$EFGH,$EFGH,$EFGH_SAVE
435
436	cbnz		$num,.Loop_hw
437
438	st1.32		{$ABCD,$EFGH},[$ctx]
439
440	ldr		x29,[sp],#16
441	ret
442.size	sha256_block_armv8,.-sha256_block_armv8
443#endif
444___
445}
446
447if ($SZ==4) {	######################################### NEON stuff #
448# You'll surely note a lot of similarities with sha256-armv4 module,
449# and of course it's not a coincidence. sha256-armv4 was used as
450# initial template, but was adapted for ARMv8 instruction set and
451# extensively re-tuned for all-round performance.
452
453my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10));
454my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15));
455my $Ktbl="x16";
456my $Xfer="x17";
457my @X = map("q$_",(0..3));
458my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19));
459my $j=0;
460
461sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
462{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
463  my $arg = pop;
464    $arg = "#$arg" if ($arg*1 eq $arg);
465    $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
466}
467
468sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; }
469sub Dlo     { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; }
470sub Dhi     { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; }
471
472sub Xupdate()
473{ use integer;
474  my $body = shift;
475  my @insns = (&$body,&$body,&$body,&$body);
476  my ($a,$b,$c,$d,$e,$f,$g,$h);
477
478	&ext_8		($T0,@X[0],@X[1],4);	# X[1..4]
479	 eval(shift(@insns));
480	 eval(shift(@insns));
481	 eval(shift(@insns));
482	&ext_8		($T3,@X[2],@X[3],4);	# X[9..12]
483	 eval(shift(@insns));
484	 eval(shift(@insns));
485	&mov		(&Dscalar($T7),&Dhi(@X[3]));	# X[14..15]
486	 eval(shift(@insns));
487	 eval(shift(@insns));
488	&ushr_32	($T2,$T0,$sigma0[0]);
489	 eval(shift(@insns));
490	&ushr_32	($T1,$T0,$sigma0[2]);
491	 eval(shift(@insns));
492	&add_32 	(@X[0],@X[0],$T3);	# X[0..3] += X[9..12]
493	 eval(shift(@insns));
494	&sli_32		($T2,$T0,32-$sigma0[0]);
495	 eval(shift(@insns));
496	 eval(shift(@insns));
497	&ushr_32	($T3,$T0,$sigma0[1]);
498	 eval(shift(@insns));
499	 eval(shift(@insns));
500	&eor_8		($T1,$T1,$T2);
501	 eval(shift(@insns));
502	 eval(shift(@insns));
503	&sli_32		($T3,$T0,32-$sigma0[1]);
504	 eval(shift(@insns));
505	 eval(shift(@insns));
506	  &ushr_32	($T4,$T7,$sigma1[0]);
507	 eval(shift(@insns));
508	 eval(shift(@insns));
509	&eor_8		($T1,$T1,$T3);		# sigma0(X[1..4])
510	 eval(shift(@insns));
511	 eval(shift(@insns));
512	  &sli_32	($T4,$T7,32-$sigma1[0]);
513	 eval(shift(@insns));
514	 eval(shift(@insns));
515	  &ushr_32	($T5,$T7,$sigma1[2]);
516	 eval(shift(@insns));
517	 eval(shift(@insns));
518	  &ushr_32	($T3,$T7,$sigma1[1]);
519	 eval(shift(@insns));
520	 eval(shift(@insns));
521	&add_32		(@X[0],@X[0],$T1);	# X[0..3] += sigma0(X[1..4])
522	 eval(shift(@insns));
523	 eval(shift(@insns));
524	  &sli_u32	($T3,$T7,32-$sigma1[1]);
525	 eval(shift(@insns));
526	 eval(shift(@insns));
527	  &eor_8	($T5,$T5,$T4);
528	 eval(shift(@insns));
529	 eval(shift(@insns));
530	 eval(shift(@insns));
531	  &eor_8	($T5,$T5,$T3);		# sigma1(X[14..15])
532	 eval(shift(@insns));
533	 eval(shift(@insns));
534	 eval(shift(@insns));
535	&add_32		(@X[0],@X[0],$T5);	# X[0..1] += sigma1(X[14..15])
536	 eval(shift(@insns));
537	 eval(shift(@insns));
538	 eval(shift(@insns));
539	  &ushr_32	($T6,@X[0],$sigma1[0]);
540	 eval(shift(@insns));
541	  &ushr_32	($T7,@X[0],$sigma1[2]);
542	 eval(shift(@insns));
543	 eval(shift(@insns));
544	  &sli_32	($T6,@X[0],32-$sigma1[0]);
545	 eval(shift(@insns));
546	  &ushr_32	($T5,@X[0],$sigma1[1]);
547	 eval(shift(@insns));
548	 eval(shift(@insns));
549	  &eor_8	($T7,$T7,$T6);
550	 eval(shift(@insns));
551	 eval(shift(@insns));
552	  &sli_32	($T5,@X[0],32-$sigma1[1]);
553	 eval(shift(@insns));
554	 eval(shift(@insns));
555	&ld1_32		("{$T0}","[$Ktbl], #16");
556	 eval(shift(@insns));
557	  &eor_8	($T7,$T7,$T5);		# sigma1(X[16..17])
558	 eval(shift(@insns));
559	 eval(shift(@insns));
560	&eor_8		($T5,$T5,$T5);
561	 eval(shift(@insns));
562	 eval(shift(@insns));
563	&mov		(&Dhi($T5), &Dlo($T7));
564	 eval(shift(@insns));
565	 eval(shift(@insns));
566	 eval(shift(@insns));
567	&add_32		(@X[0],@X[0],$T5);	# X[2..3] += sigma1(X[16..17])
568	 eval(shift(@insns));
569	 eval(shift(@insns));
570	 eval(shift(@insns));
571	&add_32		($T0,$T0,@X[0]);
572	 while($#insns>=1) { eval(shift(@insns)); }
573	&st1_32		("{$T0}","[$Xfer], #16");
574	 eval(shift(@insns));
575
576	push(@X,shift(@X));		# "rotate" X[]
577}
578
579sub Xpreload()
580{ use integer;
581  my $body = shift;
582  my @insns = (&$body,&$body,&$body,&$body);
583  my ($a,$b,$c,$d,$e,$f,$g,$h);
584
585	 eval(shift(@insns));
586	 eval(shift(@insns));
587	&ld1_8		("{@X[0]}","[$inp],#16");
588	 eval(shift(@insns));
589	 eval(shift(@insns));
590	&ld1_32		("{$T0}","[$Ktbl],#16");
591	 eval(shift(@insns));
592	 eval(shift(@insns));
593	 eval(shift(@insns));
594	 eval(shift(@insns));
595	&rev32		(@X[0],@X[0]);
596	 eval(shift(@insns));
597	 eval(shift(@insns));
598	 eval(shift(@insns));
599	 eval(shift(@insns));
600	&add_32		($T0,$T0,@X[0]);
601	 foreach (@insns) { eval; }	# remaining instructions
602	&st1_32		("{$T0}","[$Xfer], #16");
603
604	push(@X,shift(@X));		# "rotate" X[]
605}
606
607sub body_00_15 () {
608	(
609	'($a,$b,$c,$d,$e,$f,$g,$h)=@V;'.
610	'&add	($h,$h,$t1)',			# h+=X[i]+K[i]
611	'&add	($a,$a,$t4);'.			# h+=Sigma0(a) from the past
612	'&and	($t1,$f,$e)',
613	'&bic	($t4,$g,$e)',
614	'&eor	($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))',
615	'&add	($a,$a,$t2)',			# h+=Maj(a,b,c) from the past
616	'&orr	($t1,$t1,$t4)',			# Ch(e,f,g)
617	'&eor	($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))',	# Sigma1(e)
618	'&eor	($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))',
619	'&add	($h,$h,$t1)',			# h+=Ch(e,f,g)
620	'&ror	($t0,$t0,"#$Sigma1[0]")',
621	'&eor	($t2,$a,$b)',			# a^b, b^c in next round
622	'&eor	($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))',	# Sigma0(a)
623	'&add	($h,$h,$t0)',			# h+=Sigma1(e)
624	'&ldr	($t1,sprintf "[sp,#%d]",4*(($j+1)&15))	if (($j&15)!=15);'.
625	'&ldr	($t1,"[$Ktbl]")				if ($j==15);'.
626	'&and	($t3,$t3,$t2)',			# (b^c)&=(a^b)
627	'&ror	($t4,$t4,"#$Sigma0[0]")',
628	'&add	($d,$d,$h)',			# d+=h
629	'&eor	($t3,$t3,$b)',			# Maj(a,b,c)
630	'$j++;	unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);'
631	)
632}
633
634$code.=<<___;
635#ifdef	__KERNEL__
636.globl	sha256_block_neon
637#endif
638.type	sha256_block_neon,%function
639.align	4
640sha256_block_neon:
641.Lneon_entry:
642	stp	x29, x30, [sp, #-16]!
643	mov	x29, sp
644	sub	sp,sp,#16*4
645
646	adrp	$Ktbl,.LK256
647	add	$Ktbl,$Ktbl,:lo12:.LK256
648	add	$num,$inp,$num,lsl#6	// len to point at the end of inp
649
650	ld1.8	{@X[0]},[$inp], #16
651	ld1.8	{@X[1]},[$inp], #16
652	ld1.8	{@X[2]},[$inp], #16
653	ld1.8	{@X[3]},[$inp], #16
654	ld1.32	{$T0},[$Ktbl], #16
655	ld1.32	{$T1},[$Ktbl], #16
656	ld1.32	{$T2},[$Ktbl], #16
657	ld1.32	{$T3},[$Ktbl], #16
658	rev32	@X[0],@X[0]		// yes, even on
659	rev32	@X[1],@X[1]		// big-endian
660	rev32	@X[2],@X[2]
661	rev32	@X[3],@X[3]
662	mov	$Xfer,sp
663	add.32	$T0,$T0,@X[0]
664	add.32	$T1,$T1,@X[1]
665	add.32	$T2,$T2,@X[2]
666	st1.32	{$T0-$T1},[$Xfer], #32
667	add.32	$T3,$T3,@X[3]
668	st1.32	{$T2-$T3},[$Xfer]
669	sub	$Xfer,$Xfer,#32
670
671	ldp	$A,$B,[$ctx]
672	ldp	$C,$D,[$ctx,#8]
673	ldp	$E,$F,[$ctx,#16]
674	ldp	$G,$H,[$ctx,#24]
675	ldr	$t1,[sp,#0]
676	mov	$t2,wzr
677	eor	$t3,$B,$C
678	mov	$t4,wzr
679	b	.L_00_48
680
681.align	4
682.L_00_48:
683___
684	&Xupdate(\&body_00_15);
685	&Xupdate(\&body_00_15);
686	&Xupdate(\&body_00_15);
687	&Xupdate(\&body_00_15);
688$code.=<<___;
689	cmp	$t1,#0				// check for K256 terminator
690	ldr	$t1,[sp,#0]
691	sub	$Xfer,$Xfer,#64
692	bne	.L_00_48
693
694	sub	$Ktbl,$Ktbl,#256		// rewind $Ktbl
695	cmp	$inp,$num
696	mov	$Xfer, #64
697	csel	$Xfer, $Xfer, xzr, eq
698	sub	$inp,$inp,$Xfer			// avoid SEGV
699	mov	$Xfer,sp
700___
701	&Xpreload(\&body_00_15);
702	&Xpreload(\&body_00_15);
703	&Xpreload(\&body_00_15);
704	&Xpreload(\&body_00_15);
705$code.=<<___;
706	add	$A,$A,$t4			// h+=Sigma0(a) from the past
707	ldp	$t0,$t1,[$ctx,#0]
708	add	$A,$A,$t2			// h+=Maj(a,b,c) from the past
709	ldp	$t2,$t3,[$ctx,#8]
710	add	$A,$A,$t0			// accumulate
711	add	$B,$B,$t1
712	ldp	$t0,$t1,[$ctx,#16]
713	add	$C,$C,$t2
714	add	$D,$D,$t3
715	ldp	$t2,$t3,[$ctx,#24]
716	add	$E,$E,$t0
717	add	$F,$F,$t1
718	 ldr	$t1,[sp,#0]
719	stp	$A,$B,[$ctx,#0]
720	add	$G,$G,$t2
721	 mov	$t2,wzr
722	stp	$C,$D,[$ctx,#8]
723	add	$H,$H,$t3
724	stp	$E,$F,[$ctx,#16]
725	 eor	$t3,$B,$C
726	stp	$G,$H,[$ctx,#24]
727	 mov	$t4,wzr
728	 mov	$Xfer,sp
729	b.ne	.L_00_48
730
731	ldr	x29,[x29]
732	add	sp,sp,#16*4+16
733	ret
734.size	sha256_block_neon,.-sha256_block_neon
735___
736}
737
738if ($SZ==8) {
739my $Ktbl="x3";
740
741my @H = map("v$_.16b",(0..4));
742my ($fg,$de,$m9_10)=map("v$_.16b",(5..7));
743my @MSG=map("v$_.16b",(16..23));
744my ($W0,$W1)=("v24.2d","v25.2d");
745my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29));
746
747$code.=<<___;
748#ifndef	__KERNEL__
749.type	sha512_block_armv8,%function
750.align	6
751sha512_block_armv8:
752.Lv8_entry:
753	stp		x29,x30,[sp,#-16]!
754	add		x29,sp,#0
755
756	ld1		{@MSG[0]-@MSG[3]},[$inp],#64	// load input
757	ld1		{@MSG[4]-@MSG[7]},[$inp],#64
758
759	ld1.64		{@H[0]-@H[3]},[$ctx]		// load context
760	adrp		$Ktbl,.LK512
761	add		$Ktbl,$Ktbl,:lo12:.LK512
762
763	rev64		@MSG[0],@MSG[0]
764	rev64		@MSG[1],@MSG[1]
765	rev64		@MSG[2],@MSG[2]
766	rev64		@MSG[3],@MSG[3]
767	rev64		@MSG[4],@MSG[4]
768	rev64		@MSG[5],@MSG[5]
769	rev64		@MSG[6],@MSG[6]
770	rev64		@MSG[7],@MSG[7]
771	b		.Loop_hw
772
773.align	4
774.Loop_hw:
775	ld1.64		{$W0},[$Ktbl],#16
776	subs		$num,$num,#1
777	sub		x4,$inp,#128
778	orr		$AB,@H[0],@H[0]			// offload
779	orr		$CD,@H[1],@H[1]
780	orr		$EF,@H[2],@H[2]
781	orr		$GH,@H[3],@H[3]
782	csel		$inp,$inp,x4,ne			// conditional rewind
783___
784for($i=0;$i<32;$i++) {
785$code.=<<___;
786	add.i64		$W0,$W0,@MSG[0]
787	ld1.64		{$W1},[$Ktbl],#16
788	ext		$W0,$W0,$W0,#8
789	ext		$fg,@H[2],@H[3],#8
790	ext		$de,@H[1],@H[2],#8
791	add.i64		@H[3],@H[3],$W0			// "T1 + H + K512[i]"
792	 sha512su0	@MSG[0],@MSG[1]
793	 ext		$m9_10,@MSG[4],@MSG[5],#8
794	sha512h		@H[3],$fg,$de
795	 sha512su1	@MSG[0],@MSG[7],$m9_10
796	add.i64		@H[4],@H[1],@H[3]		// "D + T1"
797	sha512h2	@H[3],$H[1],@H[0]
798___
799	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
800	@H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
801}
802for(;$i<40;$i++) {
803$code.=<<___	if ($i<39);
804	ld1.64		{$W1},[$Ktbl],#16
805___
806$code.=<<___	if ($i==39);
807	sub		$Ktbl,$Ktbl,#$rounds*$SZ	// rewind
808___
809$code.=<<___;
810	add.i64		$W0,$W0,@MSG[0]
811	 ld1		{@MSG[0]},[$inp],#16		// load next input
812	ext		$W0,$W0,$W0,#8
813	ext		$fg,@H[2],@H[3],#8
814	ext		$de,@H[1],@H[2],#8
815	add.i64		@H[3],@H[3],$W0			// "T1 + H + K512[i]"
816	sha512h		@H[3],$fg,$de
817	 rev64		@MSG[0],@MSG[0]
818	add.i64		@H[4],@H[1],@H[3]		// "D + T1"
819	sha512h2	@H[3],$H[1],@H[0]
820___
821	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
822	@H = (@H[3],@H[0],@H[4],@H[2],@H[1]);
823}
824$code.=<<___;
825	add.i64		@H[0],@H[0],$AB			// accumulate
826	add.i64		@H[1],@H[1],$CD
827	add.i64		@H[2],@H[2],$EF
828	add.i64		@H[3],@H[3],$GH
829
830	cbnz		$num,.Loop_hw
831
832	st1.64		{@H[0]-@H[3]},[$ctx]		// store context
833
834	ldr		x29,[sp],#16
835	ret
836.size	sha512_block_armv8,.-sha512_block_armv8
837#endif
838___
839}
840
841{   my  %opcode = (
842	"sha256h"	=> 0x5e004000,	"sha256h2"	=> 0x5e005000,
843	"sha256su0"	=> 0x5e282800,	"sha256su1"	=> 0x5e006000	);
844
845    sub unsha256 {
846	my ($mnemonic,$arg)=@_;
847
848	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
849	&&
850	sprintf ".inst\t0x%08x\t//%s %s",
851			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
852			$mnemonic,$arg;
853    }
854}
855
856{   my  %opcode = (
857	"sha512h"	=> 0xce608000,	"sha512h2"	=> 0xce608400,
858	"sha512su0"	=> 0xcec08000,	"sha512su1"	=> 0xce608800	);
859
860    sub unsha512 {
861	my ($mnemonic,$arg)=@_;
862
863	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
864	&&
865	sprintf ".inst\t0x%08x\t//%s %s",
866			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
867			$mnemonic,$arg;
868    }
869}
870
871open SELF,$0;
872while(<SELF>) {
873        next if (/^#!/);
874        last if (!s/^#/\/\// and !/^$/);
875        print;
876}
877close SELF;
878
879foreach(split("\n",$code)) {
880
881	s/\`([^\`]*)\`/eval($1)/ge;
882
883	s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge	or
884	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge;
885
886	s/\bq([0-9]+)\b/v$1.16b/g;		# old->new registers
887
888	s/\.[ui]?8(\s)/$1/;
889	s/\.\w?64\b//		and s/\.16b/\.2d/g	or
890	s/\.\w?32\b//		and s/\.16b/\.4s/g;
891	m/\bext\b/		and s/\.2d/\.16b/g	or
892	m/(ld|st)1[^\[]+\[0\]/	and s/\.4s/\.s/g;
893
894	print $_,"\n";
895}
896
897close STDOUT or die "error closing STDOUT: $!";
898