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# sha1_block procedure for ARMv4.
11#
12# January 2007.
13
14# Size/performance trade-off
15# ====================================================================
16# impl		size in bytes	comp cycles[*]	measured performance
17# ====================================================================
18# thumb		304		3212		4420
19# armv4-small	392/+29%	1958/+64%	2250/+96%
20# armv4-compact	740/+89%	1552/+26%	1840/+22%
21# armv4-large	1420/+92%	1307/+19%	1370/+34%[***]
22# full unroll	~5100/+260%	~1260/+4%	~1300/+5%
23# ====================================================================
24# thumb		= same as 'small' but in Thumb instructions[**] and
25#		  with recurring code in two private functions;
26# small		= detached Xload/update, loops are folded;
27# compact	= detached Xload/update, 5x unroll;
28# large		= interleaved Xload/update, 5x unroll;
29# full unroll	= interleaved Xload/update, full unroll, estimated[!];
30#
31# [*]	Manually counted instructions in "grand" loop body. Measured
32#	performance is affected by prologue and epilogue overhead,
33#	i-cache availability, branch penalties, etc.
34# [**]	While each Thumb instruction is twice smaller, they are not as
35#	diverse as ARM ones: e.g., there are only two arithmetic
36#	instructions with 3 arguments, no [fixed] rotate, addressing
37#	modes are limited. As result it takes more instructions to do
38#	the same job in Thumb, therefore the code is never twice as
39#	small and always slower.
40# [***]	which is also ~35% better than compiler generated code. Dual-
41#	issue Cortex A8 core was measured to process input block in
42#	~990 cycles.
43
44# August 2010.
45#
46# Rescheduling for dual-issue pipeline resulted in 13% improvement on
47# Cortex A8 core and in absolute terms ~870 cycles per input block
48# [or 13.6 cycles per byte].
49
50# February 2011.
51#
52# Profiler-assisted and platform-specific optimization resulted in 10%
53# improvement on Cortex A8 core and 12.2 cycles per byte.
54
55# September 2013.
56#
57# Add NEON implementation (see sha1-586.pl for background info). On
58# Cortex A8 it was measured to process one byte in 6.7 cycles or >80%
59# faster than integer-only code. Because [fully unrolled] NEON code
60# is ~2.5x larger and there are some redundant instructions executed
61# when processing last block, improvement is not as big for smallest
62# blocks, only ~30%. Snapdragon S4 is a tad faster, 6.4 cycles per
63# byte, which is also >80% faster than integer-only code. Cortex-A15
64# is even faster spending 5.6 cycles per byte outperforming integer-
65# only code by factor of 2.
66
67# May 2014.
68#
69# Add ARMv8 code path performing at 2.35 cpb on Apple A7.
70
71$flavour = shift;
72if ($flavour=~/^\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
73else { while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {} }
74
75if ($flavour && $flavour ne "void") {
76    $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
77    ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
78    ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
79    die "can't locate arm-xlate.pl";
80
81    open STDOUT,"| \"$^X\" $xlate $flavour $output";
82} else {
83    open STDOUT,">$output";
84}
85
86$ctx="r0";
87$inp="r1";
88$len="r2";
89$a="r3";
90$b="r4";
91$c="r5";
92$d="r6";
93$e="r7";
94$K="r8";
95$t0="r9";
96$t1="r10";
97$t2="r11";
98$t3="r12";
99$Xi="r14";
100@V=($a,$b,$c,$d,$e);
101
102sub Xupdate {
103my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_;
104$code.=<<___;
105	ldr	$t0,[$Xi,#15*4]
106	ldr	$t1,[$Xi,#13*4]
107	ldr	$t2,[$Xi,#7*4]
108	add	$e,$K,$e,ror#2			@ E+=K_xx_xx
109	ldr	$t3,[$Xi,#2*4]
110	eor	$t0,$t0,$t1
111	eor	$t2,$t2,$t3			@ 1 cycle stall
112	eor	$t1,$c,$d			@ F_xx_xx
113	mov	$t0,$t0,ror#31
114	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
115	eor	$t0,$t0,$t2,ror#31
116	str	$t0,[$Xi,#-4]!
117	$opt1					@ F_xx_xx
118	$opt2					@ F_xx_xx
119	add	$e,$e,$t0			@ E+=X[i]
120___
121}
122
123sub BODY_00_15 {
124my ($a,$b,$c,$d,$e)=@_;
125$code.=<<___;
126#if __ARM_ARCH__<7
127	ldrb	$t1,[$inp,#2]
128	ldrb	$t0,[$inp,#3]
129	ldrb	$t2,[$inp,#1]
130	add	$e,$K,$e,ror#2			@ E+=K_00_19
131	ldrb	$t3,[$inp],#4
132	orr	$t0,$t0,$t1,lsl#8
133	eor	$t1,$c,$d			@ F_xx_xx
134	orr	$t0,$t0,$t2,lsl#16
135	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
136	orr	$t0,$t0,$t3,lsl#24
137#else
138	ldr	$t0,[$inp],#4			@ handles unaligned
139	add	$e,$K,$e,ror#2			@ E+=K_00_19
140	eor	$t1,$c,$d			@ F_xx_xx
141	add	$e,$e,$a,ror#27			@ E+=ROR(A,27)
142#ifdef __ARMEL__
143	rev	$t0,$t0				@ byte swap
144#endif
145#endif
146	and	$t1,$b,$t1,ror#2
147	add	$e,$e,$t0			@ E+=X[i]
148	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
149	str	$t0,[$Xi,#-4]!
150	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
151___
152}
153
154sub BODY_16_19 {
155my ($a,$b,$c,$d,$e)=@_;
156	&Xupdate(@_,"and $t1,$b,$t1,ror#2");
157$code.=<<___;
158	eor	$t1,$t1,$d,ror#2		@ F_00_19(B,C,D)
159	add	$e,$e,$t1			@ E+=F_00_19(B,C,D)
160___
161}
162
163sub BODY_20_39 {
164my ($a,$b,$c,$d,$e)=@_;
165	&Xupdate(@_,"eor $t1,$b,$t1,ror#2");
166$code.=<<___;
167	add	$e,$e,$t1			@ E+=F_20_39(B,C,D)
168___
169}
170
171sub BODY_40_59 {
172my ($a,$b,$c,$d,$e)=@_;
173	&Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
174$code.=<<___;
175	add	$e,$e,$t1			@ E+=F_40_59(B,C,D)
176	add	$e,$e,$t2,ror#2
177___
178}
179
180$code=<<___;
181#include "arm_arch.h"
182
183.text
184.code	32
185
186.global	sha1_block_data_order
187.type	sha1_block_data_order,%function
188
189.align	5
190sha1_block_data_order:
191#if __ARM_MAX_ARCH__>=7
192	sub	r3,pc,#8		@ sha1_block_data_order
193	ldr	r12,.LOPENSSL_armcap
194	ldr	r12,[r3,r12]		@ OPENSSL_armcap_P
195#ifdef	__APPLE__
196	ldr	r12,[r12]
197#endif
198	tst	r12,#ARMV8_SHA1
199	bne	.LARMv8
200	tst	r12,#ARMV7_NEON
201	bne	.LNEON
202#endif
203	stmdb	sp!,{r4-r12,lr}
204	add	$len,$inp,$len,lsl#6	@ $len to point at the end of $inp
205	ldmia	$ctx,{$a,$b,$c,$d,$e}
206.Lloop:
207	ldr	$K,.LK_00_19
208	mov	$Xi,sp
209	sub	sp,sp,#15*4
210	mov	$c,$c,ror#30
211	mov	$d,$d,ror#30
212	mov	$e,$e,ror#30		@ [6]
213.L_00_15:
214___
215for($i=0;$i<5;$i++) {
216	&BODY_00_15(@V);	unshift(@V,pop(@V));
217}
218$code.=<<___;
219	teq	$Xi,sp
220	bne	.L_00_15		@ [((11+4)*5+2)*3]
221	sub	sp,sp,#25*4
222___
223	&BODY_00_15(@V);	unshift(@V,pop(@V));
224	&BODY_16_19(@V);	unshift(@V,pop(@V));
225	&BODY_16_19(@V);	unshift(@V,pop(@V));
226	&BODY_16_19(@V);	unshift(@V,pop(@V));
227	&BODY_16_19(@V);	unshift(@V,pop(@V));
228$code.=<<___;
229
230	ldr	$K,.LK_20_39		@ [+15+16*4]
231	cmn	sp,#0			@ [+3], clear carry to denote 20_39
232.L_20_39_or_60_79:
233___
234for($i=0;$i<5;$i++) {
235	&BODY_20_39(@V);	unshift(@V,pop(@V));
236}
237$code.=<<___;
238	teq	$Xi,sp			@ preserve carry
239	bne	.L_20_39_or_60_79	@ [+((12+3)*5+2)*4]
240	bcs	.L_done			@ [+((12+3)*5+2)*4], spare 300 bytes
241
242	ldr	$K,.LK_40_59
243	sub	sp,sp,#20*4		@ [+2]
244.L_40_59:
245___
246for($i=0;$i<5;$i++) {
247	&BODY_40_59(@V);	unshift(@V,pop(@V));
248}
249$code.=<<___;
250	teq	$Xi,sp
251	bne	.L_40_59		@ [+((12+5)*5+2)*4]
252
253	ldr	$K,.LK_60_79
254	sub	sp,sp,#20*4
255	cmp	sp,#0			@ set carry to denote 60_79
256	b	.L_20_39_or_60_79	@ [+4], spare 300 bytes
257.L_done:
258	add	sp,sp,#80*4		@ "deallocate" stack frame
259	ldmia	$ctx,{$K,$t0,$t1,$t2,$t3}
260	add	$a,$K,$a
261	add	$b,$t0,$b
262	add	$c,$t1,$c,ror#2
263	add	$d,$t2,$d,ror#2
264	add	$e,$t3,$e,ror#2
265	stmia	$ctx,{$a,$b,$c,$d,$e}
266	teq	$inp,$len
267	bne	.Lloop			@ [+18], total 1307
268
269#if __ARM_ARCH__>=5
270	ldmia	sp!,{r4-r12,pc}
271#else
272	ldmia	sp!,{r4-r12,lr}
273	tst	lr,#1
274	moveq	pc,lr			@ be binary compatible with V4, yet
275	bx	lr			@ interoperable with Thumb ISA:-)
276#endif
277.size	sha1_block_data_order,.-sha1_block_data_order
278
279.align	5
280.LK_00_19:	.word	0x5a827999
281.LK_20_39:	.word	0x6ed9eba1
282.LK_40_59:	.word	0x8f1bbcdc
283.LK_60_79:	.word	0xca62c1d6
284#if __ARM_MAX_ARCH__>=7
285.LOPENSSL_armcap:
286.word	OPENSSL_armcap_P-sha1_block_data_order
287#endif
288.asciz	"SHA1 block transform for ARMv4/NEON/ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
289.align	5
290___
291#####################################################################
292# NEON stuff
293#
294{{{
295my @V=($a,$b,$c,$d,$e);
296my ($K_XX_XX,$Ki,$t0,$t1,$Xfer,$saved_sp)=map("r$_",(8..12,14));
297my $Xi=4;
298my @X=map("q$_",(8..11,0..3));
299my @Tx=("q12","q13");
300my ($K,$zero)=("q14","q15");
301my $j=0;
302
303sub AUTOLOAD()          # thunk [simplified] x86-style perlasm
304{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
305  my $arg = pop;
306    $arg = "#$arg" if ($arg*1 eq $arg);
307    $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
308}
309
310sub body_00_19 () {
311	(
312	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
313	'&bic	($t0,$d,$b)',
314	'&add	($e,$e,$Ki)',		# e+=X[i]+K
315	'&and	($t1,$c,$b)',
316	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
317	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
318	'&eor	($t1,$t1,$t0)',		# F_00_19
319	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
320	'&add	($e,$e,$t1);'.		# e+=F_00_19
321	'$j++;	unshift(@V,pop(@V));'
322	)
323}
324sub body_20_39 () {
325	(
326	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
327	'&eor	($t0,$b,$d)',
328	'&add	($e,$e,$Ki)',		# e+=X[i]+K
329	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15)) if ($j<79)',
330	'&eor	($t1,$t0,$c)',		# F_20_39
331	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
332	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
333	'&add	($e,$e,$t1);'.		# e+=F_20_39
334	'$j++;	unshift(@V,pop(@V));'
335	)
336}
337sub body_40_59 () {
338	(
339	'($a,$b,$c,$d,$e)=@V;'.		# '$code.="@ $j\n";'.
340	'&add	($e,$e,$Ki)',		# e+=X[i]+K
341	'&and	($t0,$c,$d)',
342	'&ldr	($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
343	'&add	($e,$e,$a,"ror#27")',	# e+=ROR(A,27)
344	'&eor	($t1,$c,$d)',
345	'&add	($e,$e,$t0)',
346	'&and	($t1,$t1,$b)',
347	'&mov	($b,$b,"ror#2")',	# b=ROR(b,2)
348	'&add	($e,$e,$t1);'.		# e+=F_40_59
349	'$j++;	unshift(@V,pop(@V));'
350	)
351}
352
353sub Xupdate_16_31 ()
354{ use integer;
355  my $body = shift;
356  my @insns = (&$body,&$body,&$body,&$body);
357  my ($a,$b,$c,$d,$e);
358
359	&vext_8		(@X[0],@X[-4&7],@X[-3&7],8);	# compose "X[-14]" in "X[0]"
360	 eval(shift(@insns));
361	 eval(shift(@insns));
362	 eval(shift(@insns));
363	  &vadd_i32	(@Tx[1],@X[-1&7],$K);
364	 eval(shift(@insns));
365	  &vld1_32	("{$K\[]}","[$K_XX_XX,:32]!")	if ($Xi%5==0);
366	 eval(shift(@insns));
367	&vext_8		(@Tx[0],@X[-1&7],$zero,4);	# "X[-3]", 3 words
368	 eval(shift(@insns));
369	 eval(shift(@insns));
370	 eval(shift(@insns));
371	&veor		(@X[0],@X[0],@X[-4&7]);		# "X[0]"^="X[-16]"
372	 eval(shift(@insns));
373	 eval(shift(@insns));
374	&veor		(@Tx[0],@Tx[0],@X[-2&7]);	# "X[-3]"^"X[-8]"
375	 eval(shift(@insns));
376	 eval(shift(@insns));
377	&veor		(@Tx[0],@Tx[0],@X[0]);		# "X[0]"^="X[-3]"^"X[-8]
378	 eval(shift(@insns));
379	 eval(shift(@insns));
380	  &vst1_32	("{@Tx[1]}","[$Xfer,:128]!");	# X[]+K xfer
381	  &sub		($Xfer,$Xfer,64)		if ($Xi%4==0);
382	 eval(shift(@insns));
383	 eval(shift(@insns));
384	&vext_8		(@Tx[1],$zero,@Tx[0],4);	# "X[0]"<<96, extract one dword
385	 eval(shift(@insns));
386	 eval(shift(@insns));
387	&vadd_i32	(@X[0],@Tx[0],@Tx[0]);
388	 eval(shift(@insns));
389	 eval(shift(@insns));
390	&vsri_32	(@X[0],@Tx[0],31);		# "X[0]"<<<=1
391	 eval(shift(@insns));
392	 eval(shift(@insns));
393	 eval(shift(@insns));
394	&vshr_u32	(@Tx[0],@Tx[1],30);
395	 eval(shift(@insns));
396	 eval(shift(@insns));
397	&vshl_u32	(@Tx[1],@Tx[1],2);
398	 eval(shift(@insns));
399	 eval(shift(@insns));
400	&veor		(@X[0],@X[0],@Tx[0]);
401	 eval(shift(@insns));
402	 eval(shift(@insns));
403	&veor		(@X[0],@X[0],@Tx[1]);		# "X[0]"^=("X[0]">>96)<<<2
404
405	foreach (@insns) { eval; }	# remaining instructions [if any]
406
407  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
408}
409
410sub Xupdate_32_79 ()
411{ use integer;
412  my $body = shift;
413  my @insns = (&$body,&$body,&$body,&$body);
414  my ($a,$b,$c,$d,$e);
415
416	&vext_8		(@Tx[0],@X[-2&7],@X[-1&7],8);	# compose "X[-6]"
417	 eval(shift(@insns));
418	 eval(shift(@insns));
419	 eval(shift(@insns));
420	&veor		(@X[0],@X[0],@X[-4&7]);		# "X[0]"="X[-32]"^"X[-16]"
421	 eval(shift(@insns));
422	 eval(shift(@insns));
423	&veor		(@X[0],@X[0],@X[-7&7]);		# "X[0]"^="X[-28]"
424	 eval(shift(@insns));
425	 eval(shift(@insns));
426	  &vadd_i32	(@Tx[1],@X[-1&7],$K);
427	 eval(shift(@insns));
428	  &vld1_32	("{$K\[]}","[$K_XX_XX,:32]!")	if ($Xi%5==0);
429	 eval(shift(@insns));
430	&veor		(@Tx[0],@Tx[0],@X[0]);		# "X[-6]"^="X[0]"
431	 eval(shift(@insns));
432	 eval(shift(@insns));
433	&vshr_u32	(@X[0],@Tx[0],30);
434	 eval(shift(@insns));
435	 eval(shift(@insns));
436	  &vst1_32	("{@Tx[1]}","[$Xfer,:128]!");	# X[]+K xfer
437	  &sub		($Xfer,$Xfer,64)		if ($Xi%4==0);
438	 eval(shift(@insns));
439	 eval(shift(@insns));
440	&vsli_32	(@X[0],@Tx[0],2);		# "X[0]"="X[-6]"<<<2
441
442	foreach (@insns) { eval; }	# remaining instructions [if any]
443
444  $Xi++;	push(@X,shift(@X));	# "rotate" X[]
445}
446
447sub Xuplast_80 ()
448{ use integer;
449  my $body = shift;
450  my @insns = (&$body,&$body,&$body,&$body);
451  my ($a,$b,$c,$d,$e);
452
453	&vadd_i32	(@Tx[1],@X[-1&7],$K);
454	 eval(shift(@insns));
455	 eval(shift(@insns));
456	&vst1_32	("{@Tx[1]}","[$Xfer,:128]!");
457	&sub		($Xfer,$Xfer,64);
458
459	&teq		($inp,$len);
460	&sub		($K_XX_XX,$K_XX_XX,16);	# rewind $K_XX_XX
461	&subeq		($inp,$inp,64);		# reload last block to avoid SEGV
462	&vld1_8		("{@X[-4&7]-@X[-3&7]}","[$inp]!");
463	 eval(shift(@insns));
464	 eval(shift(@insns));
465	&vld1_8		("{@X[-2&7]-@X[-1&7]}","[$inp]!");
466	 eval(shift(@insns));
467	 eval(shift(@insns));
468	&vld1_32	("{$K\[]}","[$K_XX_XX,:32]!");	# load K_00_19
469	 eval(shift(@insns));
470	 eval(shift(@insns));
471	&vrev32_8	(@X[-4&7],@X[-4&7]);
472
473	foreach (@insns) { eval; }		# remaining instructions
474
475   $Xi=0;
476}
477
478sub Xloop()
479{ use integer;
480  my $body = shift;
481  my @insns = (&$body,&$body,&$body,&$body);
482  my ($a,$b,$c,$d,$e);
483
484	&vrev32_8	(@X[($Xi-3)&7],@X[($Xi-3)&7]);
485	 eval(shift(@insns));
486	 eval(shift(@insns));
487	&vadd_i32	(@X[$Xi&7],@X[($Xi-4)&7],$K);
488	 eval(shift(@insns));
489	 eval(shift(@insns));
490	&vst1_32	("{@X[$Xi&7]}","[$Xfer,:128]!");# X[]+K xfer to IALU
491
492	foreach (@insns) { eval; }
493
494  $Xi++;
495}
496
497$code.=<<___;
498#if __ARM_MAX_ARCH__>=7
499.arch	armv7-a
500.fpu	neon
501
502.type	sha1_block_data_order_neon,%function
503.align	4
504sha1_block_data_order_neon:
505.LNEON:
506	stmdb	sp!,{r4-r12,lr}
507	add	$len,$inp,$len,lsl#6	@ $len to point at the end of $inp
508	@ dmb				@ errata #451034 on early Cortex A8
509	@ vstmdb	sp!,{d8-d15}	@ ABI specification says so
510	mov	$saved_sp,sp
511	sub	sp,sp,#64		@ alloca
512	adr	$K_XX_XX,.LK_00_19
513	bic	sp,sp,#15		@ align for 128-bit stores
514
515	ldmia	$ctx,{$a,$b,$c,$d,$e}	@ load context
516	mov	$Xfer,sp
517
518	vld1.8		{@X[-4&7]-@X[-3&7]},[$inp]!	@ handles unaligned
519	veor		$zero,$zero,$zero
520	vld1.8		{@X[-2&7]-@X[-1&7]},[$inp]!
521	vld1.32		{${K}\[]},[$K_XX_XX,:32]!	@ load K_00_19
522	vrev32.8	@X[-4&7],@X[-4&7]		@ yes, even on
523	vrev32.8	@X[-3&7],@X[-3&7]		@ big-endian...
524	vrev32.8	@X[-2&7],@X[-2&7]
525	vadd.i32	@X[0],@X[-4&7],$K
526	vrev32.8	@X[-1&7],@X[-1&7]
527	vadd.i32	@X[1],@X[-3&7],$K
528	vst1.32		{@X[0]},[$Xfer,:128]!
529	vadd.i32	@X[2],@X[-2&7],$K
530	vst1.32		{@X[1]},[$Xfer,:128]!
531	vst1.32		{@X[2]},[$Xfer,:128]!
532	ldr		$Ki,[sp]			@ big RAW stall
533
534.Loop_neon:
535___
536	&Xupdate_16_31(\&body_00_19);
537	&Xupdate_16_31(\&body_00_19);
538	&Xupdate_16_31(\&body_00_19);
539	&Xupdate_16_31(\&body_00_19);
540	&Xupdate_32_79(\&body_00_19);
541	&Xupdate_32_79(\&body_20_39);
542	&Xupdate_32_79(\&body_20_39);
543	&Xupdate_32_79(\&body_20_39);
544	&Xupdate_32_79(\&body_20_39);
545	&Xupdate_32_79(\&body_20_39);
546	&Xupdate_32_79(\&body_40_59);
547	&Xupdate_32_79(\&body_40_59);
548	&Xupdate_32_79(\&body_40_59);
549	&Xupdate_32_79(\&body_40_59);
550	&Xupdate_32_79(\&body_40_59);
551	&Xupdate_32_79(\&body_20_39);
552	&Xuplast_80(\&body_20_39);
553	&Xloop(\&body_20_39);
554	&Xloop(\&body_20_39);
555	&Xloop(\&body_20_39);
556$code.=<<___;
557	ldmia	$ctx,{$Ki,$t0,$t1,$Xfer}	@ accumulate context
558	add	$a,$a,$Ki
559	ldr	$Ki,[$ctx,#16]
560	add	$b,$b,$t0
561	add	$c,$c,$t1
562	add	$d,$d,$Xfer
563	moveq	sp,$saved_sp
564	add	$e,$e,$Ki
565	ldrne	$Ki,[sp]
566	stmia	$ctx,{$a,$b,$c,$d,$e}
567	addne	$Xfer,sp,#3*16
568	bne	.Loop_neon
569
570	@ vldmia	sp!,{d8-d15}
571	ldmia	sp!,{r4-r12,pc}
572.size	sha1_block_data_order_neon,.-sha1_block_data_order_neon
573#endif
574___
575}}}
576#####################################################################
577# ARMv8 stuff
578#
579{{{
580my ($ABCD,$E,$E0,$E1)=map("q$_",(0..3));
581my @MSG=map("q$_",(4..7));
582my @Kxx=map("q$_",(8..11));
583my ($W0,$W1,$ABCD_SAVE)=map("q$_",(12..14));
584
585$code.=<<___;
586#if __ARM_MAX_ARCH__>=7
587.type	sha1_block_data_order_armv8,%function
588.align	5
589sha1_block_data_order_armv8:
590.LARMv8:
591	vstmdb	sp!,{d8-d15}		@ ABI specification says so
592
593	veor	$E,$E,$E
594	adr	r3,.LK_00_19
595	vld1.32	{$ABCD},[$ctx]!
596	vld1.32	{$E\[0]},[$ctx]
597	sub	$ctx,$ctx,#16
598	vld1.32	{@Kxx[0]\[]},[r3,:32]!
599	vld1.32	{@Kxx[1]\[]},[r3,:32]!
600	vld1.32	{@Kxx[2]\[]},[r3,:32]!
601	vld1.32	{@Kxx[3]\[]},[r3,:32]
602
603.Loop_v8:
604	vld1.8		{@MSG[0]-@MSG[1]},[$inp]!
605	vld1.8		{@MSG[2]-@MSG[3]},[$inp]!
606	vrev32.8	@MSG[0],@MSG[0]
607	vrev32.8	@MSG[1],@MSG[1]
608
609	vadd.i32	$W0,@Kxx[0],@MSG[0]
610	vrev32.8	@MSG[2],@MSG[2]
611	vmov		$ABCD_SAVE,$ABCD	@ offload
612	subs		$len,$len,#1
613
614	vadd.i32	$W1,@Kxx[0],@MSG[1]
615	vrev32.8	@MSG[3],@MSG[3]
616	sha1h		$E1,$ABCD		@ 0
617	sha1c		$ABCD,$E,$W0
618	vadd.i32	$W0,@Kxx[$j],@MSG[2]
619	sha1su0		@MSG[0],@MSG[1],@MSG[2]
620___
621for ($j=0,$i=1;$i<20-3;$i++) {
622my $f=("c","p","m","p")[$i/5];
623$code.=<<___;
624	sha1h		$E0,$ABCD		@ $i
625	sha1$f		$ABCD,$E1,$W1
626	vadd.i32	$W1,@Kxx[$j],@MSG[3]
627	sha1su1		@MSG[0],@MSG[3]
628___
629$code.=<<___ if ($i<20-4);
630	sha1su0		@MSG[1],@MSG[2],@MSG[3]
631___
632	($E0,$E1)=($E1,$E0);	($W0,$W1)=($W1,$W0);
633	push(@MSG,shift(@MSG));	$j++ if ((($i+3)%5)==0);
634}
635$code.=<<___;
636	sha1h		$E0,$ABCD		@ $i
637	sha1p		$ABCD,$E1,$W1
638	vadd.i32	$W1,@Kxx[$j],@MSG[3]
639
640	sha1h		$E1,$ABCD		@ 18
641	sha1p		$ABCD,$E0,$W0
642
643	sha1h		$E0,$ABCD		@ 19
644	sha1p		$ABCD,$E1,$W1
645
646	vadd.i32	$E,$E,$E0
647	vadd.i32	$ABCD,$ABCD,$ABCD_SAVE
648	bne		.Loop_v8
649
650	vst1.32		{$ABCD},[$ctx]!
651	vst1.32		{$E\[0]},[$ctx]
652
653	vldmia	sp!,{d8-d15}
654	ret					@ bx lr
655.size	sha1_block_data_order_armv8,.-sha1_block_data_order_armv8
656#endif
657___
658}}}
659$code.=<<___;
660#if __ARM_MAX_ARCH__>=7
661.comm	OPENSSL_armcap_P,4,4
662.hidden	OPENSSL_armcap_P
663#endif
664___
665
666{   my  %opcode = (
667	"sha1c"		=> 0xf2000c40,	"sha1p"		=> 0xf2100c40,
668	"sha1m"		=> 0xf2200c40,	"sha1su0"	=> 0xf2300c40,
669	"sha1h"		=> 0xf3b902c0,	"sha1su1"	=> 0xf3ba0380	);
670
671    sub unsha1 {
672	my ($mnemonic,$arg)=@_;
673
674	if ($arg =~ m/q([0-9]+)(?:,\s*q([0-9]+))?,\s*q([0-9]+)/o) {
675	    my $word = $opcode{$mnemonic}|(($1&7)<<13)|(($1&8)<<19)
676					 |(($2&7)<<17)|(($2&8)<<4)
677					 |(($3&7)<<1) |(($3&8)<<2);
678	    # since ARMv7 instructions are always encoded little-endian.
679	    # correct solution is to use .inst directive, but older
680	    # assemblers don't implement it:-(
681	    sprintf ".byte\t0x%02x,0x%02x,0x%02x,0x%02x\t@ %s %s",
682			$word&0xff,($word>>8)&0xff,
683			($word>>16)&0xff,($word>>24)&0xff,
684			$mnemonic,$arg;
685	}
686    }
687}
688
689foreach (split($/,$code)) {
690	s/{q([0-9]+)\[\]}/sprintf "{d%d[],d%d[]}",2*$1,2*$1+1/eo	or
691	s/{q([0-9]+)\[0\]}/sprintf "{d%d[0]}",2*$1/eo;
692
693	s/\b(sha1\w+)\s+(q.*)/unsha1($1,$2)/geo;
694
695	s/\bret\b/bx	lr/o		or
696	s/\bbx\s+lr\b/.word\t0xe12fff1e/o;	# make it possible to compile with -march=armv4
697
698	print $_,$/;
699}
700
701close STDOUT; # enforce flush
702