sha512-armv8.pl revision 3f9e6ada2c9f7183a41081263585e6a70bbd9f59 
1#!/usr/bin/env perl
2#
3# ====================================================================
4# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5# project. The module is, however, dual licensed under OpenSSL and
6# CRYPTOGAMS licenses depending on where you obtain it. For further
7# details see http://www.openssl.org/~appro/cryptogams/.
8# ====================================================================
9#
10# SHA256/512 for ARMv8.
11#
12# Performance in cycles per processed byte and improvement coefficient
13# over code generated with "default" compiler:
14#
15#		SHA256-hw	SHA256(*)	SHA512
16# Apple A7	1.97		10.5 (+33%)	6.73 (-1%(**))
17# Cortex-A5x	n/a		n/a		n/a
18#
19# (*)	Software SHA256 results are of lesser relevance, presented
20#	mostly for informational purposes.
21# (**)	The result is a trade-off: it's possible to improve it by
22#	10%, but at the cost of 20% loss on Cortex-A5x.
23
24$flavour=shift;
25$output=shift;
26open STDOUT,">$output";
27
28if ($output =~ /512/) {
29	$BITS=512;
30	$SZ=8;
31	@Sigma0=(28,34,39);
32	@Sigma1=(14,18,41);
33	@sigma0=(1,  8, 7);
34	@sigma1=(19,61, 6);
35	$rounds=80;
36	$reg_t="x";
37} else {
38	$BITS=256;
39	$SZ=4;
40	@Sigma0=( 2,13,22);
41	@Sigma1=( 6,11,25);
42	@sigma0=( 7,18, 3);
43	@sigma1=(17,19,10);
44	$rounds=64;
45	$reg_t="w";
46}
47
48$func="sha${BITS}_block_data_order";
49
50($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30));
51
52@X=map("$reg_t$_",(3..15,0..2));
53@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27));
54($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28));
55
56sub BODY_00_xx {
57my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_;
58my $j=($i+1)&15;
59my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]);
60   $T0=@X[$i+3] if ($i<11);
61
62$code.=<<___	if ($i<16);
63#ifndef	__ARMEB__
64	rev	@X[$i],@X[$i]			// $i
65#endif
66___
67$code.=<<___	if ($i<13 && ($i&1));
68	ldp	@X[$i+1],@X[$i+2],[$inp],#2*$SZ
69___
70$code.=<<___	if ($i==13);
71	ldp	@X[14],@X[15],[$inp]
72___
73$code.=<<___	if ($i>=14);
74	ldr	@X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`]
75___
76$code.=<<___	if ($i>0 && $i<16);
77	add	$a,$a,$t1			// h+=Sigma0(a)
78___
79$code.=<<___	if ($i>=11);
80	str	@X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`]
81___
82# While ARMv8 specifies merged rotate-n-logical operation such as
83# 'eor x,y,z,ror#n', it was found to negatively affect performance
84# on Apple A7. The reason seems to be that it requires even 'y' to
85# be available earlier. This means that such merged instruction is
86# not necessarily best choice on critical path... On the other hand
87# Cortex-A5x handles merged instructions much better than disjoint
88# rotate and logical... See (**) footnote above.
89$code.=<<___	if ($i<15);
90	ror	$t0,$e,#$Sigma1[0]
91	add	$h,$h,$t2			// h+=K[i]
92	eor	$T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]`
93	and	$t1,$f,$e
94	bic	$t2,$g,$e
95	add	$h,$h,@X[$i&15]			// h+=X[i]
96	orr	$t1,$t1,$t2			// Ch(e,f,g)
97	eor	$t2,$a,$b			// a^b, b^c in next round
98	eor	$t0,$t0,$T0,ror#$Sigma1[1]	// Sigma1(e)
99	ror	$T0,$a,#$Sigma0[0]
100	add	$h,$h,$t1			// h+=Ch(e,f,g)
101	eor	$t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]`
102	add	$h,$h,$t0			// h+=Sigma1(e)
103	and	$t3,$t3,$t2			// (b^c)&=(a^b)
104	add	$d,$d,$h			// d+=h
105	eor	$t3,$t3,$b			// Maj(a,b,c)
106	eor	$t1,$T0,$t1,ror#$Sigma0[1]	// Sigma0(a)
107	add	$h,$h,$t3			// h+=Maj(a,b,c)
108	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
109	//add	$h,$h,$t1			// h+=Sigma0(a)
110___
111$code.=<<___	if ($i>=15);
112	ror	$t0,$e,#$Sigma1[0]
113	add	$h,$h,$t2			// h+=K[i]
114	ror	$T1,@X[($j+1)&15],#$sigma0[0]
115	and	$t1,$f,$e
116	ror	$T2,@X[($j+14)&15],#$sigma1[0]
117	bic	$t2,$g,$e
118	ror	$T0,$a,#$Sigma0[0]
119	add	$h,$h,@X[$i&15]			// h+=X[i]
120	eor	$t0,$t0,$e,ror#$Sigma1[1]
121	eor	$T1,$T1,@X[($j+1)&15],ror#$sigma0[1]
122	orr	$t1,$t1,$t2			// Ch(e,f,g)
123	eor	$t2,$a,$b			// a^b, b^c in next round
124	eor	$t0,$t0,$e,ror#$Sigma1[2]	// Sigma1(e)
125	eor	$T0,$T0,$a,ror#$Sigma0[1]
126	add	$h,$h,$t1			// h+=Ch(e,f,g)
127	and	$t3,$t3,$t2			// (b^c)&=(a^b)
128	eor	$T2,$T2,@X[($j+14)&15],ror#$sigma1[1]
129	eor	$T1,$T1,@X[($j+1)&15],lsr#$sigma0[2]	// sigma0(X[i+1])
130	add	$h,$h,$t0			// h+=Sigma1(e)
131	eor	$t3,$t3,$b			// Maj(a,b,c)
132	eor	$t1,$T0,$a,ror#$Sigma0[2]	// Sigma0(a)
133	eor	$T2,$T2,@X[($j+14)&15],lsr#$sigma1[2]	// sigma1(X[i+14])
134	add	@X[$j],@X[$j],@X[($j+9)&15]
135	add	$d,$d,$h			// d+=h
136	add	$h,$h,$t3			// h+=Maj(a,b,c)
137	ldr	$t3,[$Ktbl],#$SZ		// *K++, $t2 in next round
138	add	@X[$j],@X[$j],$T1
139	add	$h,$h,$t1			// h+=Sigma0(a)
140	add	@X[$j],@X[$j],$T2
141___
142	($t2,$t3)=($t3,$t2);
143}
144
145$code.=<<___;
146#include "arm_arch.h"
147
148.text
149
150.globl	$func
151.type	$func,%function
152.align	6
153$func:
154___
155$code.=<<___	if ($SZ==4);
156	ldr	x16,.LOPENSSL_armcap_P
157	adr	x17,.LOPENSSL_armcap_P
158	add	x16,x16,x17
159	ldr	w16,[x16]
160	tst	w16,#ARMV8_SHA256
161	b.ne	.Lv8_entry
162___
163$code.=<<___;
164	stp	x29,x30,[sp,#-128]!
165	add	x29,sp,#0
166
167	stp	x19,x20,[sp,#16]
168	stp	x21,x22,[sp,#32]
169	stp	x23,x24,[sp,#48]
170	stp	x25,x26,[sp,#64]
171	stp	x27,x28,[sp,#80]
172	sub	sp,sp,#4*$SZ
173
174	ldp	$A,$B,[$ctx]				// load context
175	ldp	$C,$D,[$ctx,#2*$SZ]
176	ldp	$E,$F,[$ctx,#4*$SZ]
177	add	$num,$inp,$num,lsl#`log(16*$SZ)/log(2)`	// end of input
178	ldp	$G,$H,[$ctx,#6*$SZ]
179	adr	$Ktbl,K$BITS
180	stp	$ctx,$num,[x29,#96]
181
182.Loop:
183	ldp	@X[0],@X[1],[$inp],#2*$SZ
184	ldr	$t2,[$Ktbl],#$SZ			// *K++
185	eor	$t3,$B,$C				// magic seed
186	str	$inp,[x29,#112]
187___
188for ($i=0;$i<16;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
189$code.=".Loop_16_xx:\n";
190for (;$i<32;$i++)	{ &BODY_00_xx($i,@V); unshift(@V,pop(@V)); }
191$code.=<<___;
192	cbnz	$t2,.Loop_16_xx
193
194	ldp	$ctx,$num,[x29,#96]
195	ldr	$inp,[x29,#112]
196	sub	$Ktbl,$Ktbl,#`$SZ*($rounds+1)`		// rewind
197
198	ldp	@X[0],@X[1],[$ctx]
199	ldp	@X[2],@X[3],[$ctx,#2*$SZ]
200	add	$inp,$inp,#14*$SZ			// advance input pointer
201	ldp	@X[4],@X[5],[$ctx,#4*$SZ]
202	add	$A,$A,@X[0]
203	ldp	@X[6],@X[7],[$ctx,#6*$SZ]
204	add	$B,$B,@X[1]
205	add	$C,$C,@X[2]
206	add	$D,$D,@X[3]
207	stp	$A,$B,[$ctx]
208	add	$E,$E,@X[4]
209	add	$F,$F,@X[5]
210	stp	$C,$D,[$ctx,#2*$SZ]
211	add	$G,$G,@X[6]
212	add	$H,$H,@X[7]
213	cmp	$inp,$num
214	stp	$E,$F,[$ctx,#4*$SZ]
215	stp	$G,$H,[$ctx,#6*$SZ]
216	b.ne	.Loop
217
218	ldp	x19,x20,[x29,#16]
219	add	sp,sp,#4*$SZ
220	ldp	x21,x22,[x29,#32]
221	ldp	x23,x24,[x29,#48]
222	ldp	x25,x26,[x29,#64]
223	ldp	x27,x28,[x29,#80]
224	ldp	x29,x30,[sp],#128
225	ret
226.size	$func,.-$func
227
228.align	6
229.type	K$BITS,%object
230K$BITS:
231___
232$code.=<<___ if ($SZ==8);
233	.quad	0x428a2f98d728ae22,0x7137449123ef65cd
234	.quad	0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc
235	.quad	0x3956c25bf348b538,0x59f111f1b605d019
236	.quad	0x923f82a4af194f9b,0xab1c5ed5da6d8118
237	.quad	0xd807aa98a3030242,0x12835b0145706fbe
238	.quad	0x243185be4ee4b28c,0x550c7dc3d5ffb4e2
239	.quad	0x72be5d74f27b896f,0x80deb1fe3b1696b1
240	.quad	0x9bdc06a725c71235,0xc19bf174cf692694
241	.quad	0xe49b69c19ef14ad2,0xefbe4786384f25e3
242	.quad	0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65
243	.quad	0x2de92c6f592b0275,0x4a7484aa6ea6e483
244	.quad	0x5cb0a9dcbd41fbd4,0x76f988da831153b5
245	.quad	0x983e5152ee66dfab,0xa831c66d2db43210
246	.quad	0xb00327c898fb213f,0xbf597fc7beef0ee4
247	.quad	0xc6e00bf33da88fc2,0xd5a79147930aa725
248	.quad	0x06ca6351e003826f,0x142929670a0e6e70
249	.quad	0x27b70a8546d22ffc,0x2e1b21385c26c926
250	.quad	0x4d2c6dfc5ac42aed,0x53380d139d95b3df
251	.quad	0x650a73548baf63de,0x766a0abb3c77b2a8
252	.quad	0x81c2c92e47edaee6,0x92722c851482353b
253	.quad	0xa2bfe8a14cf10364,0xa81a664bbc423001
254	.quad	0xc24b8b70d0f89791,0xc76c51a30654be30
255	.quad	0xd192e819d6ef5218,0xd69906245565a910
256	.quad	0xf40e35855771202a,0x106aa07032bbd1b8
257	.quad	0x19a4c116b8d2d0c8,0x1e376c085141ab53
258	.quad	0x2748774cdf8eeb99,0x34b0bcb5e19b48a8
259	.quad	0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb
260	.quad	0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3
261	.quad	0x748f82ee5defb2fc,0x78a5636f43172f60
262	.quad	0x84c87814a1f0ab72,0x8cc702081a6439ec
263	.quad	0x90befffa23631e28,0xa4506cebde82bde9
264	.quad	0xbef9a3f7b2c67915,0xc67178f2e372532b
265	.quad	0xca273eceea26619c,0xd186b8c721c0c207
266	.quad	0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178
267	.quad	0x06f067aa72176fba,0x0a637dc5a2c898a6
268	.quad	0x113f9804bef90dae,0x1b710b35131c471b
269	.quad	0x28db77f523047d84,0x32caab7b40c72493
270	.quad	0x3c9ebe0a15c9bebc,0x431d67c49c100d4c
271	.quad	0x4cc5d4becb3e42b6,0x597f299cfc657e2a
272	.quad	0x5fcb6fab3ad6faec,0x6c44198c4a475817
273	.quad	0	// terminator
274___
275$code.=<<___ if ($SZ==4);
276	.long	0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5
277	.long	0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5
278	.long	0xd807aa98,0x12835b01,0x243185be,0x550c7dc3
279	.long	0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174
280	.long	0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc
281	.long	0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da
282	.long	0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7
283	.long	0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967
284	.long	0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13
285	.long	0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85
286	.long	0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3
287	.long	0xd192e819,0xd6990624,0xf40e3585,0x106aa070
288	.long	0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5
289	.long	0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3
290	.long	0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208
291	.long	0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2
292	.long	0	//terminator
293___
294$code.=<<___;
295.size	K$BITS,.-K$BITS
296.align	3
297.LOPENSSL_armcap_P:
298	.quad	OPENSSL_armcap_P-.
299.asciz	"SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
300.align	2
301___
302
303if ($SZ==4) {
304my $Ktbl="x3";
305
306my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2));
307my @MSG=map("v$_.16b",(4..7));
308my ($W0,$W1)=("v16.4s","v17.4s");
309my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b");
310
311$code.=<<___;
312.type	sha256_block_armv8,%function
313.align	6
314sha256_block_armv8:
315.Lv8_entry:
316	stp		x29,x30,[sp,#-16]!
317	add		x29,sp,#0
318
319	ld1.32		{$ABCD,$EFGH},[$ctx]
320	adr		$Ktbl,K256
321
322.Loop_hw:
323	ld1		{@MSG[0]-@MSG[3]},[$inp],#64
324	sub		$num,$num,#1
325	ld1.32		{$W0},[$Ktbl],#16
326	rev32		@MSG[0],@MSG[0]
327	rev32		@MSG[1],@MSG[1]
328	rev32		@MSG[2],@MSG[2]
329	rev32		@MSG[3],@MSG[3]
330	orr		$ABCD_SAVE,$ABCD,$ABCD		// offload
331	orr		$EFGH_SAVE,$EFGH,$EFGH
332___
333for($i=0;$i<12;$i++) {
334$code.=<<___;
335	ld1.32		{$W1},[$Ktbl],#16
336	add.i32		$W0,$W0,@MSG[0]
337	sha256su0	@MSG[0],@MSG[1]
338	orr		$abcd,$ABCD,$ABCD
339	sha256h		$ABCD,$EFGH,$W0
340	sha256h2	$EFGH,$abcd,$W0
341	sha256su1	@MSG[0],@MSG[2],@MSG[3]
342___
343	($W0,$W1)=($W1,$W0);	push(@MSG,shift(@MSG));
344}
345$code.=<<___;
346	ld1.32		{$W1},[$Ktbl],#16
347	add.i32		$W0,$W0,@MSG[0]
348	orr		$abcd,$ABCD,$ABCD
349	sha256h		$ABCD,$EFGH,$W0
350	sha256h2	$EFGH,$abcd,$W0
351
352	ld1.32		{$W0},[$Ktbl],#16
353	add.i32		$W1,$W1,@MSG[1]
354	orr		$abcd,$ABCD,$ABCD
355	sha256h		$ABCD,$EFGH,$W1
356	sha256h2	$EFGH,$abcd,$W1
357
358	ld1.32		{$W1},[$Ktbl]
359	add.i32		$W0,$W0,@MSG[2]
360	sub		$Ktbl,$Ktbl,#$rounds*$SZ-16	// rewind
361	orr		$abcd,$ABCD,$ABCD
362	sha256h		$ABCD,$EFGH,$W0
363	sha256h2	$EFGH,$abcd,$W0
364
365	add.i32		$W1,$W1,@MSG[3]
366	orr		$abcd,$ABCD,$ABCD
367	sha256h		$ABCD,$EFGH,$W1
368	sha256h2	$EFGH,$abcd,$W1
369
370	add.i32		$ABCD,$ABCD,$ABCD_SAVE
371	add.i32		$EFGH,$EFGH,$EFGH_SAVE
372
373	cbnz		$num,.Loop_hw
374
375	st1.32		{$ABCD,$EFGH},[$ctx]
376
377	ldr		x29,[sp],#16
378	ret
379.size	sha256_block_armv8,.-sha256_block_armv8
380___
381}
382
383$code.=<<___;
384.comm	OPENSSL_armcap_P,4,4
385___
386
387{   my  %opcode = (
388	"sha256h"	=> 0x5e004000,	"sha256h2"	=> 0x5e005000,
389	"sha256su0"	=> 0x5e282800,	"sha256su1"	=> 0x5e006000	);
390
391    sub unsha256 {
392	my ($mnemonic,$arg)=@_;
393
394	$arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o
395	&&
396	sprintf ".inst\t0x%08x\t//%s %s",
397			$opcode{$mnemonic}|$1|($2<<5)|($3<<16),
398			$mnemonic,$arg;
399    }
400}
401
402foreach(split("\n",$code)) {
403
404	s/\`([^\`]*)\`/eval($1)/geo;
405
406	s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo;
407
408	s/\.\w?32\b//o		and s/\.16b/\.4s/go;
409	m/(ld|st)1[^\[]+\[0\]/o	and s/\.4s/\.s/go;
410
411	print $_,"\n";
412}
413
414close STDOUT;
415