xref: /aosp_15_r20/external/boringssl/src/crypto/fipsmodule/sha/asm/sha1-armv4-large.pl (revision 8fb009dc861624b67b6cdb62ea21f0f22d0c584b)
1#! /usr/bin/env perl
2# Copyright 2007-2016 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 <[email protected]> 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# sha1_block procedure for ARMv4.
18#
19# January 2007.
20
21# Size/performance trade-off
22# ====================================================================
23# impl		size in bytes	comp cycles[*]	measured performance
24# ====================================================================
25# thumb		304		3212		4420
26# armv4-small	392/+29%	1958/+64%	2250/+96%
27# armv4-compact	740/+89%	1552/+26%	1840/+22%
28# armv4-large	1420/+92%	1307/+19%	1370/+34%[***]
29# full unroll	~5100/+260%	~1260/+4%	~1300/+5%
30# ====================================================================
31# thumb		= same as 'small' but in Thumb instructions[**] and
32#		  with recurring code in two private functions;
33# small		= detached Xload/update, loops are folded;
34# compact	= detached Xload/update, 5x unroll;
35# large		= interleaved Xload/update, 5x unroll;
36# full unroll	= interleaved Xload/update, full unroll, estimated[!];
37#
38# [*]	Manually counted instructions in "grand" loop body. Measured
39#	performance is affected by prologue and epilogue overhead,
40#	i-cache availability, branch penalties, etc.
41# [**]	While each Thumb instruction is twice smaller, they are not as
42#	diverse as ARM ones: e.g., there are only two arithmetic
43#	instructions with 3 arguments, no [fixed] rotate, addressing
44#	modes are limited. As result it takes more instructions to do
45#	the same job in Thumb, therefore the code is never twice as
46#	small and always slower.
47# [***]	which is also ~35% better than compiler generated code. Dual-
48#	issue Cortex A8 core was measured to process input block in
49#	~990 cycles.
50
51# August 2010.
52#
53# Rescheduling for dual-issue pipeline resulted in 13% improvement on
54# Cortex A8 core and in absolute terms ~870 cycles per input block
55# [or 13.6 cycles per byte].
56
57# February 2011.
58#
59# Profiler-assisted and platform-specific optimization resulted in 10%
60# improvement on Cortex A8 core and 12.2 cycles per byte.
61
62# September 2013.
63#
64# Add NEON implementation (see sha1-586.pl for background info). On
65# Cortex A8 it was measured to process one byte in 6.7 cycles or >80%
66# faster than integer-only code. Because [fully unrolled] NEON code
67# is ~2.5x larger and there are some redundant instructions executed
68# when processing last block, improvement is not as big for smallest
69# blocks, only ~30%. Snapdragon S4 is a tad faster, 6.4 cycles per
70# byte, which is also >80% faster than integer-only code. Cortex-A15
71# is even faster spending 5.6 cycles per byte outperforming integer-
72# only code by factor of 2.
73
74# May 2014.
75#
76# Add ARMv8 code path performing at 2.35 cpb on Apple A7.
77
78$flavour = shift;
79if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
80else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} }
81
82if ($flavour && $flavour ne "void") {
83    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
84    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
85    ( $xlate="${dir}../../../perlasm/arm-xlate.pl" and -f $xlate) or
86    die "can't locate arm-xlate.pl";
87
88    open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\"";
89    *STDOUT=*OUT;
90} else {
91    open OUT,">$output";
92    *STDOUT=*OUT;
93}
94
95$ctx="r0";
96$inp="r1";
97$len="r2";
98$a="r3";
99$b="r4";
100$c="r5";
101$d="r6";
102$e="r7";
103$K="r8";
104$t0="r9";
105$t1="r10";
106$t2="r11";
107$t3="r12";
108$Xi="r14";
109@V=($a,$b,$c,$d,$e);
110
111sub Xupdate {
112my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_;
113$code.=<<___;
114	ldr	$t0,[$Xi,#15*4]
115	ldr	$t1,[$Xi,#13*4]
116	ldr	$t2,[$Xi,#7*4]
117	add	$e,$K,$e,ror#2			@ E+=K_xx_xx
118	ldr	$t3,[$Xi,#2*4]
119	eor	$t0,$t0,$t1
120	eor	$t2,$t2,$t3			@ 1 cycle stall
121	eor	$t1,$c,$d			@ F_xx_xx
122	mov	$t0,$t0,ror#31
123	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
124	eor	$t0,$t0,$t2,ror#31
125	str	$t0,[$Xi,#-4]!
126	$opt1					@ F_xx_xx
127	$opt2					@ F_xx_xx
128	add	$e,$e,$t0			@ E+=X[i]
129___
130}
131
132sub BODY_00_15 {
133my ($a,$b,$c,$d,$e)=@_;
134$code.=<<___;
135#if __ARM_ARCH<7
136	ldrb	$t1,[$inp,#2]
137	ldrb	$t0,[$inp,#3]
138	ldrb	$t2,[$inp,#1]
139	add	$e,$K,$e,ror#2			@ E+=K_00_19
140	ldrb	$t3,[$inp],#4
141	orr	$t0,$t0,$t1,lsl#8
142	eor	$t1,$c,$d			@ F_xx_xx
143	orr	$t0,$t0,$t2,lsl#16
144	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
145	orr	$t0,$t0,$t3,lsl#24
146#else
147	ldr	$t0,[$inp],#4			@ handles unaligned
148	add	$e,$K,$e,ror#2			@ E+=K_00_19
149	eor	$t1,$c,$d			@ F_xx_xx
150	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
151#ifdef __ARMEL__
152	rev	$t0,$t0				@ byte swap
153#endif
154#endif
155	and	$t1,$b,$t1,ror#2
156	add	$e,$e,$t0			@ E+=X[i]
157	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
158	str	$t0,[$Xi,#-4]!
159	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
160___
161}
162
163sub BODY_16_19 {
164my ($a,$b,$c,$d,$e)=@_;
165	&Xupdate(@_,"and $t1,$b,$t1,ror#2");
166$code.=<<___;
167	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
168	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
169___
170}
171
172sub BODY_20_39 {
173my ($a,$b,$c,$d,$e)=@_;
174	&Xupdate(@_,"eor $t1,$b,$t1,ror#2");
175$code.=<<___;
176	add	$e,$e,$t1			@ E+=F_20_39(B,C,D)
177___
178}
179
180sub BODY_40_59 {
181my ($a,$b,$c,$d,$e)=@_;
182	&Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
183$code.=<<___;
184	add	$e,$e,$t1			@ E+=F_40_59(B,C,D)
185	add	$e,$e,$t2,ror#2
186___
187}
188
189$code=<<___;
190#include <openssl/arm_arch.h>
191
192.text
193#if defined(__thumb2__)
194.syntax	unified
195.thumb
196#else
197.code	32
198#endif
199
200.global	sha1_block_data_order_nohw
201.type	sha1_block_data_order_nohw,%function
202
203.align	5
204sha1_block_data_order_nohw:
205	stmdb	sp!,{r4-r12,lr}
206	add	$len,$inp,$len,lsl#6	@ $len to point at the end of $inp
207	ldmia	$ctx,{$a,$b,$c,$d,$e}
208.Lloop:
209	ldr	$K,.LK_00_19
210	mov	$Xi,sp
211	sub	sp,sp,#15*4
212	mov	$c,$c,ror#30
213	mov	$d,$d,ror#30
214	mov	$e,$e,ror#30		@ [6]
215.L_00_15:
216___
217for($i=0;$i<5;$i++) {
218	&BODY_00_15(@V);	unshift(@V,pop(@V));
219}
220$code.=<<___;
221#if defined(__thumb2__)
222	mov	$t3,sp
223	teq	$Xi,$t3
224#else
225	teq	$Xi,sp
226#endif
227	bne	.L_00_15		@ [((11+4)*5+2)*3]
228	sub	sp,sp,#25*4
229___
230	&BODY_00_15(@V);	unshift(@V,pop(@V));
231	&BODY_16_19(@V);	unshift(@V,pop(@V));
232	&BODY_16_19(@V);	unshift(@V,pop(@V));
233	&BODY_16_19(@V);	unshift(@V,pop(@V));
234	&BODY_16_19(@V);	unshift(@V,pop(@V));
235$code.=<<___;
236
237	ldr	$K,.LK_20_39		@ [+15+16*4]
238	cmn	sp,#0			@ [+3], clear carry to denote 20_39
239.L_20_39_or_60_79:
240___
241for($i=0;$i<5;$i++) {
242	&BODY_20_39(@V);	unshift(@V,pop(@V));
243}
244$code.=<<___;
245#if defined(__thumb2__)
246	mov	$t3,sp
247	teq	$Xi,$t3
248#else
249	teq	$Xi,sp			@ preserve carry
250#endif
251	bne	.L_20_39_or_60_79	@ [+((12+3)*5+2)*4]
252	bcs	.L_done			@ [+((12+3)*5+2)*4], spare 300 bytes
253
254	ldr	$K,.LK_40_59
255	sub	sp,sp,#20*4		@ [+2]
256.L_40_59:
257___
258for($i=0;$i<5;$i++) {
259	&BODY_40_59(@V);	unshift(@V,pop(@V));
260}
261$code.=<<___;
262#if defined(__thumb2__)
263	mov	$t3,sp
264	teq	$Xi,$t3
265#else
266	teq	$Xi,sp
267#endif
268	bne	.L_40_59		@ [+((12+5)*5+2)*4]
269
270	ldr	$K,.LK_60_79
271	sub	sp,sp,#20*4
272	cmp	sp,#0			@ set carry to denote 60_79
273	b	.L_20_39_or_60_79	@ [+4], spare 300 bytes
274.L_done:
275	add	sp,sp,#80*4		@ "deallocate" stack frame
276	ldmia	$ctx,{$K,$t0,$t1,$t2,$t3}
277	add	$a,$K,$a
278	add	$b,$t0,$b
279	add	$c,$t1,$c,ror#2
280	add	$d,$t2,$d,ror#2
281	add	$e,$t3,$e,ror#2
282	stmia	$ctx,{$a,$b,$c,$d,$e}
283	teq	$inp,$len
284	bne	.Lloop			@ [+18], total 1307
285
286#if __ARM_ARCH>=5
287	ldmia	sp!,{r4-r12,pc}
288#else
289	ldmia	sp!,{r4-r12,lr}
290	tst	lr,#1
291	moveq	pc,lr			@ be binary compatible with V4, yet
292	bx	lr			@ interoperable with Thumb ISA:-)
293#endif
294.size	sha1_block_data_order_nohw,.-sha1_block_data_order_nohw
295
296.align	5
297.LK_00_19:	.word	0x5a827999
298.LK_20_39:	.word	0x6ed9eba1
299.LK_40_59:	.word	0x8f1bbcdc
300.LK_60_79:	.word	0xca62c1d6
301.asciz	"SHA1 block transform for ARMv4/NEON/ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
302.align	5
303___
304#####################################################################
305# NEON stuff
306#
307{{{
308my @V=($a,$b,$c,$d,$e);
309my ($K_XX_XX,$Ki,$t0,$t1,$Xfer,$saved_sp)=map("r$_",(8..12,14));
310my $Xi=4;
311my @X=map("q$_",(8..11,0..3));
312my @Tx=("q12","q13");
313my ($K,$zero)=("q14","q15");
314my $j=0;
315
316sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
317{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
318  my $arg = pop;
319    $arg = "#$arg" if ($arg*1 eq $arg);
320    $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
321}
322
323sub body_00_19 () {
324	(
325	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
326	'&bic	($t0,$d,$b)',
327	'&add	($e,$e,$Ki)',		# e+=X[i]+K
328	'&and	($t1,$c,$b)',
329	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
330	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
331	'&eor	($t1,$t1,$t0)',		# F_00_19
332	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
333	'&add	($e,$e,$t1);'.		# e+=F_00_19
334	'$j++;	unshift(@V,pop(@V));'
335	)
336}
337sub body_20_39 () {
338	(
339	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
340	'&eor	($t0,$b,$d)',
341	'&add	($e,$e,$Ki)',		# e+=X[i]+K
342	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15)) if ($j<79)',
343	'&eor	($t1,$t0,$c)',		# F_20_39
344	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
345	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
346	'&add	($e,$e,$t1);'.		# e+=F_20_39
347	'$j++;	unshift(@V,pop(@V));'
348	)
349}
350sub body_40_59 () {
351	(
352	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
353	'&add	($e,$e,$Ki)',		# e+=X[i]+K
354	'&and	($t0,$c,$d)',
355	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
356	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
357	'&eor	($t1,$c,$d)',
358	'&add	($e,$e,$t0)',
359	'&and	($t1,$t1,$b)',
360	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
361	'&add	($e,$e,$t1);'.		# e+=F_40_59
362	'$j++;	unshift(@V,pop(@V));'
363	)
364}
365
366sub Xupdate_16_31 ()
367{ use integer;
368  my $body = shift;
369  my @insns = (&$body,&$body,&$body,&$body);
370  my ($a,$b,$c,$d,$e);
371
372	&vext_8		(@X[0],@X[-4&7],@X[-3&7],8);	# compose "X[-14]" in "X[0]"
373	 eval(shift(@insns));
374	 eval(shift(@insns));
375	 eval(shift(@insns));
376	  &vadd_i32	(@Tx[1],@X[-1&7],$K);
377	 eval(shift(@insns));
378	  &vld1_32	("{$K\[]}","[$K_XX_XX,:32]!")	if ($Xi%5==0);
379	 eval(shift(@insns));
380	&vext_8		(@Tx[0],@X[-1&7],$zero,4);	# "X[-3]", 3 words
381	 eval(shift(@insns));
382	 eval(shift(@insns));
383	 eval(shift(@insns));
384	&veor		(@X[0],@X[0],@X[-4&7]);		# "X[0]"^="X[-16]"
385	 eval(shift(@insns));
386	 eval(shift(@insns));
387	&veor		(@Tx[0],@Tx[0],@X[-2&7]);	# "X[-3]"^"X[-8]"
388	 eval(shift(@insns));
389	 eval(shift(@insns));
390	&veor		(@Tx[0],@Tx[0],@X[0]);		# "X[0]"^="X[-3]"^"X[-8]
391	 eval(shift(@insns));
392	 eval(shift(@insns));
393	  &vst1_32	("{@Tx[1]}","[$Xfer,:128]!");	# X[]+K xfer
394	  &sub		($Xfer,$Xfer,64)		if ($Xi%4==0);
395	 eval(shift(@insns));
396	 eval(shift(@insns));
397	&vext_8		(@Tx[1],$zero,@Tx[0],4);	# "X[0]"<<96, extract one dword
398	 eval(shift(@insns));
399	 eval(shift(@insns));
400	&vadd_i32	(@X[0],@Tx[0],@Tx[0]);
401	 eval(shift(@insns));
402	 eval(shift(@insns));
403	&vsri_32	(@X[0],@Tx[0],31);		# "X[0]"<<<=1
404	 eval(shift(@insns));
405	 eval(shift(@insns));
406	 eval(shift(@insns));
407	&vshr_u32	(@Tx[0],@Tx[1],30);
408	 eval(shift(@insns));
409	 eval(shift(@insns));
410	&vshl_u32	(@Tx[1],@Tx[1],2);
411	 eval(shift(@insns));
412	 eval(shift(@insns));
413	&veor		(@X[0],@X[0],@Tx[0]);
414	 eval(shift(@insns));
415	 eval(shift(@insns));
416	&veor		(@X[0],@X[0],@Tx[1]);		# "X[0]"^=("X[0]">>96)<<<2
417
418	foreach (@insns) { eval; }	# remaining instructions [if any]
419
420  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
421}
422
423sub Xupdate_32_79 ()
424{ use integer;
425  my $body = shift;
426  my @insns = (&$body,&$body,&$body,&$body);
427  my ($a,$b,$c,$d,$e);
428
429	&vext_8		(@Tx[0],@X[-2&7],@X[-1&7],8);	# compose "X[-6]"
430	 eval(shift(@insns));
431	 eval(shift(@insns));
432	 eval(shift(@insns));
433	&veor		(@X[0],@X[0],@X[-4&7]);		# "X[0]"="X[-32]"^"X[-16]"
434	 eval(shift(@insns));
435	 eval(shift(@insns));
436	&veor		(@X[0],@X[0],@X[-7&7]);		# "X[0]"^="X[-28]"
437	 eval(shift(@insns));
438	 eval(shift(@insns));
439	  &vadd_i32	(@Tx[1],@X[-1&7],$K);
440	 eval(shift(@insns));
441	  &vld1_32	("{$K\[]}","[$K_XX_XX,:32]!")	if ($Xi%5==0);
442	 eval(shift(@insns));
443	&veor		(@Tx[0],@Tx[0],@X[0]);		# "X[-6]"^="X[0]"
444	 eval(shift(@insns));
445	 eval(shift(@insns));
446	&vshr_u32	(@X[0],@Tx[0],30);
447	 eval(shift(@insns));
448	 eval(shift(@insns));
449	  &vst1_32	("{@Tx[1]}","[$Xfer,:128]!");	# X[]+K xfer
450	  &sub		($Xfer,$Xfer,64)		if ($Xi%4==0);
451	 eval(shift(@insns));
452	 eval(shift(@insns));
453	&vsli_32	(@X[0],@Tx[0],2);		# "X[0]"="X[-6]"<<<2
454
455	foreach (@insns) { eval; }	# remaining instructions [if any]
456
457  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
458}
459
460sub Xuplast_80 ()
461{ use integer;
462  my $body = shift;
463  my @insns = (&$body,&$body,&$body,&$body);
464  my ($a,$b,$c,$d,$e);
465
466	&vadd_i32	(@Tx[1],@X[-1&7],$K);
467	 eval(shift(@insns));
468	 eval(shift(@insns));
469	&vst1_32	("{@Tx[1]}","[$Xfer,:128]!");
470	&sub		($Xfer,$Xfer,64);
471
472	&teq		($inp,$len);
473	&sub		($K_XX_XX,$K_XX_XX,16);	# rewind $K_XX_XX
474	&it		("eq");
475	&subeq		($inp,$inp,64);		# reload last block to avoid SEGV
476	&vld1_8		("{@X[-4&7]-@X[-3&7]}","[$inp]!");
477	 eval(shift(@insns));
478	 eval(shift(@insns));
479	&vld1_8		("{@X[-2&7]-@X[-1&7]}","[$inp]!");
480	 eval(shift(@insns));
481	 eval(shift(@insns));
482	&vld1_32	("{$K\[]}","[$K_XX_XX,:32]!");	# load K_00_19
483	 eval(shift(@insns));
484	 eval(shift(@insns));
485	&vrev32_8	(@X[-4&7],@X[-4&7]);
486
487	foreach (@insns) { eval; }		# remaining instructions
488
489   $Xi=0;
490}
491
492sub Xloop()
493{ use integer;
494  my $body = shift;
495  my @insns = (&$body,&$body,&$body,&$body);
496  my ($a,$b,$c,$d,$e);
497
498	&vrev32_8	(@X[($Xi-3)&7],@X[($Xi-3)&7]);
499	 eval(shift(@insns));
500	 eval(shift(@insns));
501	&vadd_i32	(@X[$Xi&7],@X[($Xi-4)&7],$K);
502	 eval(shift(@insns));
503	 eval(shift(@insns));
504	&vst1_32	("{@X[$Xi&7]}","[$Xfer,:128]!");# X[]+K xfer to IALU
505
506	foreach (@insns) { eval; }
507
508  $Xi++;
509}
510
511$code.=<<___;
512#if __ARM_MAX_ARCH__>=7
513.arch	armv7-a
514.fpu	neon
515
516.global	sha1_block_data_order_neon
517.type	sha1_block_data_order_neon,%function
518.align	4
519sha1_block_data_order_neon:
520	stmdb	sp!,{r4-r12,lr}
521	add	$len,$inp,$len,lsl#6	@ $len to point at the end of $inp
522	@ dmb				@ errata #451034 on early Cortex A8
523	@ vstmdb	sp!,{d8-d15}	@ ABI specification says so
524	mov	$saved_sp,sp
525	sub	$Xfer,sp,#64
526	adr	$K_XX_XX,.LK_00_19
527	bic	$Xfer,$Xfer,#15		@ align for 128-bit stores
528
529	ldmia	$ctx,{$a,$b,$c,$d,$e}	@ load context
530	mov	sp,$Xfer		@ alloca
531
532	vld1.8		{@X[-4&7]-@X[-3&7]},[$inp]!	@ handles unaligned
533	veor		$zero,$zero,$zero
534	vld1.8		{@X[-2&7]-@X[-1&7]},[$inp]!
535	vld1.32		{${K}\[]},[$K_XX_XX,:32]!	@ load K_00_19
536	vrev32.8	@X[-4&7],@X[-4&7]		@ yes, even on
537	vrev32.8	@X[-3&7],@X[-3&7]		@ big-endian...
538	vrev32.8	@X[-2&7],@X[-2&7]
539	vadd.i32	@X[0],@X[-4&7],$K
540	vrev32.8	@X[-1&7],@X[-1&7]
541	vadd.i32	@X[1],@X[-3&7],$K
542	vst1.32		{@X[0]},[$Xfer,:128]!
543	vadd.i32	@X[2],@X[-2&7],$K
544	vst1.32		{@X[1]},[$Xfer,:128]!
545	vst1.32		{@X[2]},[$Xfer,:128]!
546	ldr		$Ki,[sp]			@ big RAW stall
547
548.Loop_neon:
549___
550	&Xupdate_16_31(\&body_00_19);
551	&Xupdate_16_31(\&body_00_19);
552	&Xupdate_16_31(\&body_00_19);
553	&Xupdate_16_31(\&body_00_19);
554	&Xupdate_32_79(\&body_00_19);
555	&Xupdate_32_79(\&body_20_39);
556	&Xupdate_32_79(\&body_20_39);
557	&Xupdate_32_79(\&body_20_39);
558	&Xupdate_32_79(\&body_20_39);
559	&Xupdate_32_79(\&body_20_39);
560	&Xupdate_32_79(\&body_40_59);
561	&Xupdate_32_79(\&body_40_59);
562	&Xupdate_32_79(\&body_40_59);
563	&Xupdate_32_79(\&body_40_59);
564	&Xupdate_32_79(\&body_40_59);
565	&Xupdate_32_79(\&body_20_39);
566	&Xuplast_80(\&body_20_39);
567	&Xloop(\&body_20_39);
568	&Xloop(\&body_20_39);
569	&Xloop(\&body_20_39);
570$code.=<<___;
571	ldmia	$ctx,{$Ki,$t0,$t1,$Xfer}	@ accumulate context
572	add	$a,$a,$Ki
573	ldr	$Ki,[$ctx,#16]
574	add	$b,$b,$t0
575	add	$c,$c,$t1
576	add	$d,$d,$Xfer
577	it	eq
578	moveq	sp,$saved_sp
579	add	$e,$e,$Ki
580	it	ne
581	ldrne	$Ki,[sp]
582	stmia	$ctx,{$a,$b,$c,$d,$e}
583	itt	ne
584	addne	$Xfer,sp,#3*16
585	bne	.Loop_neon
586
587	@ vldmia	sp!,{d8-d15}
588	ldmia	sp!,{r4-r12,pc}
589.size	sha1_block_data_order_neon,.-sha1_block_data_order_neon
590#endif
591___
592}}}
593#####################################################################
594# ARMv8 stuff
595#
596{{{
597my ($ABCD,$E,$E0,$E1)=map("q$_",(0..3));
598my @MSG=map("q$_",(4..7));
599my @Kxx=map("q$_",(8..11));
600my ($W0,$W1,$ABCD_SAVE)=map("q$_",(12..14));
601
602$code.=<<___;
603#if __ARM_MAX_ARCH__>=7
604
605# if defined(__thumb2__)
606#  define INST(a,b,c,d)	.byte	c,d|0xf,a,b
607# else
608#  define INST(a,b,c,d)	.byte	a,b,c,d|0x10
609# endif
610
611.global	sha1_block_data_order_hw
612.type	sha1_block_data_order_hw,%function
613.align	5
614sha1_block_data_order_hw:
615	vstmdb	sp!,{d8-d15}		@ ABI specification says so
616
617	veor	$E,$E,$E
618	adr	r3,.LK_00_19
619	vld1.32	{$ABCD},[$ctx]!
620	vld1.32	{$E\[0]},[$ctx]
621	sub	$ctx,$ctx,#16
622	vld1.32	{@Kxx[0]\[]},[r3,:32]!
623	vld1.32	{@Kxx[1]\[]},[r3,:32]!
624	vld1.32	{@Kxx[2]\[]},[r3,:32]!
625	vld1.32	{@Kxx[3]\[]},[r3,:32]
626
627.Loop_v8:
628	vld1.8		{@MSG[0]-@MSG[1]},[$inp]!
629	vld1.8		{@MSG[2]-@MSG[3]},[$inp]!
630	vrev32.8	@MSG[0],@MSG[0]
631	vrev32.8	@MSG[1],@MSG[1]
632
633	vadd.i32	$W0,@Kxx[0],@MSG[0]
634	vrev32.8	@MSG[2],@MSG[2]
635	vmov		$ABCD_SAVE,$ABCD	@ offload
636	subs		$len,$len,#1
637
638	vadd.i32	$W1,@Kxx[0],@MSG[1]
639	vrev32.8	@MSG[3],@MSG[3]
640	sha1h		$E1,$ABCD		@ 0
641	sha1c		$ABCD,$E,$W0
642	vadd.i32	$W0,@Kxx[$j],@MSG[2]
643	sha1su0		@MSG[0],@MSG[1],@MSG[2]
644___
645for ($j=0,$i=1;$i<20-3;$i++) {
646my $f=("c","p","m","p")[$i/5];
647$code.=<<___;
648	sha1h		$E0,$ABCD		@ $i
649	sha1$f		$ABCD,$E1,$W1
650	vadd.i32	$W1,@Kxx[$j],@MSG[3]
651	sha1su1		@MSG[0],@MSG[3]
652___
653$code.=<<___ if ($i<20-4);
654	sha1su0		@MSG[1],@MSG[2],@MSG[3]
655___
656	($E0,$E1)=($E1,$E0);	($W0,$W1)=($W1,$W0);
657	push(@MSG,shift(@MSG));	$j++ if ((($i+3)%5)==0);
658}
659$code.=<<___;
660	sha1h		$E0,$ABCD		@ $i
661	sha1p		$ABCD,$E1,$W1
662	vadd.i32	$W1,@Kxx[$j],@MSG[3]
663
664	sha1h		$E1,$ABCD		@ 18
665	sha1p		$ABCD,$E0,$W0
666
667	sha1h		$E0,$ABCD		@ 19
668	sha1p		$ABCD,$E1,$W1
669
670	vadd.i32	$E,$E,$E0
671	vadd.i32	$ABCD,$ABCD,$ABCD_SAVE
672	bne		.Loop_v8
673
674	vst1.32		{$ABCD},[$ctx]!
675	vst1.32		{$E\[0]},[$ctx]
676
677	vldmia	sp!,{d8-d15}
678	ret					@ bx lr
679.size	sha1_block_data_order_hw,.-sha1_block_data_order_hw
680#endif
681___
682}}}
683
684{   my  %opcode = (
685	"sha1c"		=> 0xf2000c40,	"sha1p"		=> 0xf2100c40,
686	"sha1m"		=> 0xf2200c40,	"sha1su0"	=> 0xf2300c40,
687	"sha1h"		=> 0xf3b902c0,	"sha1su1"	=> 0xf3ba0380	);
688
689    sub unsha1 {
690	my ($mnemonic,$arg)=@_;
691
692	if ($arg =~ m/q([0-9]+)(?:,\s*q([0-9]+))?,\s*q([0-9]+)/o) {
693	    my $word = $opcode{$mnemonic}|(($1&7)<<13)|(($1&8)<<19)
694					 |(($2&7)<<17)|(($2&8)<<4)
695					 |(($3&7)<<1) |(($3&8)<<2);
696	    # since ARMv7 instructions are always encoded little-endian.
697	    # correct solution is to use .inst directive, but older
698	    # assemblers don't implement it:-(
699
700	    # this fix-up provides Thumb encoding in conjunction with INST
701	    $word &= ~0x10000000 if (($word & 0x0f000000) == 0x02000000);
702	    sprintf "INST(0x%02x,0x%02x,0x%02x,0x%02x)\t@ %s %s",
703			$word&0xff,($word>>8)&0xff,
704			($word>>16)&0xff,($word>>24)&0xff,
705			$mnemonic,$arg;
706	}
707    }
708}
709
710foreach (split($/,$code)) {
711	s/{q([0-9]+)\[\]}/sprintf "{d%d[],d%d[]}",2*$1,2*$1+1/eo	or
712	s/{q([0-9]+)\[0\]}/sprintf "{d%d[0]}",2*$1/eo;
713
714	s/\b(sha1\w+)\s+(q.*)/unsha1($1,$2)/geo;
715
716	s/\bret\b/bx	lr/o		or
717	s/\bbx\s+lr\b/.word\t0xe12fff1e/o;	# make it possible to compile with -march=armv4
718
719	print $_,$/;
720}
721
722close STDOUT or die "error closing STDOUT: $!"; # enforce flush
723