1#! /usr/bin/env perl
2# Copyright 2014-2016 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 <appro@openssl.org> 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# GHASH for ARMv8 Crypto Extension, 64-bit polynomial multiplication.
18#
19# June 2014
20# Initial version was developed in tight cooperation with Ard Biesheuvel
21# of Linaro from bits-n-pieces from other assembly modules. Just like
22# aesv8-armx.pl this module supports both AArch32 and AArch64 execution modes.
23#
24# July 2014
25# Implement 2x aggregated reduction [see ghash-x86.pl for background
26# information].
27#
28# Current performance in cycles per processed byte:
29#
30#		PMULL[2]	32-bit NEON(*)
31# Apple A7	0.92		5.62
32# Cortex-A53	1.01		8.39
33# Cortex-A57	1.17		7.61
34# Denver	0.71		6.02
35# Mongoose	1.10		8.06
36# Kryo		1.16		8.00
37#
38# (*)	presented for reference/comparison purposes;
39
40$flavour = shift;
41$output  = shift;
42
43$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
44( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
45( $xlate="${dir}../../../perlasm/arm-xlate.pl" and -f $xlate) or
46die "can't locate arm-xlate.pl";
47
48open OUT,"| \"$^X\" $xlate $flavour $output";
49*STDOUT=*OUT;
50
51$Xi="x0";	# argument block
52$Htbl="x1";
53$inp="x2";
54$len="x3";
55
56$inc="x12";
57
58{
59my ($Xl,$Xm,$Xh,$IN)=map("q$_",(0..3));
60my ($t0,$t1,$t2,$xC2,$H,$Hhl,$H2)=map("q$_",(8..14));
61
62$code=<<___;
63#include <openssl/arm_arch.h>
64
65.text
66___
67$code.=".arch	armv8-a+crypto\n"	if ($flavour =~ /64/);
68$code.=<<___				if ($flavour !~ /64/);
69.fpu	neon
70.code	32
71#undef	__thumb2__
72___
73
74################################################################################
75# void gcm_init_v8(u128 Htable[16],const u64 H[2]);
76#
77# input:	128-bit H - secret parameter E(K,0^128)
78# output:	precomputed table filled with degrees of twisted H;
79#		H is twisted to handle reverse bitness of GHASH;
80#		only few of 16 slots of Htable[16] are used;
81#		data is opaque to outside world (which allows to
82#		optimize the code independently);
83#
84$code.=<<___;
85.global	gcm_init_v8
86.type	gcm_init_v8,%function
87.align	4
88gcm_init_v8:
89	vld1.64		{$t1},[x1]		@ load input H
90	vmov.i8		$xC2,#0xe1
91	vshl.i64	$xC2,$xC2,#57		@ 0xc2.0
92	vext.8		$IN,$t1,$t1,#8
93	vshr.u64	$t2,$xC2,#63
94	vdup.32		$t1,${t1}[1]
95	vext.8		$t0,$t2,$xC2,#8		@ t0=0xc2....01
96	vshr.u64	$t2,$IN,#63
97	vshr.s32	$t1,$t1,#31		@ broadcast carry bit
98	vand		$t2,$t2,$t0
99	vshl.i64	$IN,$IN,#1
100	vext.8		$t2,$t2,$t2,#8
101	vand		$t0,$t0,$t1
102	vorr		$IN,$IN,$t2		@ H<<<=1
103	veor		$H,$IN,$t0		@ twisted H
104	vst1.64		{$H},[x0],#16		@ store Htable[0]
105
106	@ calculate H^2
107	vext.8		$t0,$H,$H,#8		@ Karatsuba pre-processing
108	vpmull.p64	$Xl,$H,$H
109	veor		$t0,$t0,$H
110	vpmull2.p64	$Xh,$H,$H
111	vpmull.p64	$Xm,$t0,$t0
112
113	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
114	veor		$t2,$Xl,$Xh
115	veor		$Xm,$Xm,$t1
116	veor		$Xm,$Xm,$t2
117	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase
118
119	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
120	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
121	veor		$Xl,$Xm,$t2
122
123	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase
124	vpmull.p64	$Xl,$Xl,$xC2
125	veor		$t2,$t2,$Xh
126	veor		$H2,$Xl,$t2
127
128	vext.8		$t1,$H2,$H2,#8		@ Karatsuba pre-processing
129	veor		$t1,$t1,$H2
130	vext.8		$Hhl,$t0,$t1,#8		@ pack Karatsuba pre-processed
131	vst1.64		{$Hhl-$H2},[x0]		@ store Htable[1..2]
132
133	ret
134.size	gcm_init_v8,.-gcm_init_v8
135___
136################################################################################
137# void gcm_gmult_v8(u64 Xi[2],const u128 Htable[16]);
138#
139# input:	Xi - current hash value;
140#		Htable - table precomputed in gcm_init_v8;
141# output:	Xi - next hash value Xi;
142#
143$code.=<<___;
144.global	gcm_gmult_v8
145.type	gcm_gmult_v8,%function
146.align	4
147gcm_gmult_v8:
148	vld1.64		{$t1},[$Xi]		@ load Xi
149	vmov.i8		$xC2,#0xe1
150	vld1.64		{$H-$Hhl},[$Htbl]	@ load twisted H, ...
151	vshl.u64	$xC2,$xC2,#57
152#ifndef __ARMEB__
153	vrev64.8	$t1,$t1
154#endif
155	vext.8		$IN,$t1,$t1,#8
156
157	vpmull.p64	$Xl,$H,$IN		@ H.lo��Xi.lo
158	veor		$t1,$t1,$IN		@ Karatsuba pre-processing
159	vpmull2.p64	$Xh,$H,$IN		@ H.hi��Xi.hi
160	vpmull.p64	$Xm,$Hhl,$t1		@ (H.lo+H.hi)��(Xi.lo+Xi.hi)
161
162	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
163	veor		$t2,$Xl,$Xh
164	veor		$Xm,$Xm,$t1
165	veor		$Xm,$Xm,$t2
166	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
167
168	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
169	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
170	veor		$Xl,$Xm,$t2
171
172	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
173	vpmull.p64	$Xl,$Xl,$xC2
174	veor		$t2,$t2,$Xh
175	veor		$Xl,$Xl,$t2
176
177#ifndef __ARMEB__
178	vrev64.8	$Xl,$Xl
179#endif
180	vext.8		$Xl,$Xl,$Xl,#8
181	vst1.64		{$Xl},[$Xi]		@ write out Xi
182
183	ret
184.size	gcm_gmult_v8,.-gcm_gmult_v8
185___
186################################################################################
187# void gcm_ghash_v8(u64 Xi[2],const u128 Htable[16],const u8 *inp,size_t len);
188#
189# input:	table precomputed in gcm_init_v8;
190#		current hash value Xi;
191#		pointer to input data;
192#		length of input data in bytes, but divisible by block size;
193# output:	next hash value Xi;
194#
195$code.=<<___;
196.global	gcm_ghash_v8
197.type	gcm_ghash_v8,%function
198.align	4
199gcm_ghash_v8:
200___
201$code.=<<___		if ($flavour !~ /64/);
202	vstmdb		sp!,{d8-d15}		@ 32-bit ABI says so
203___
204$code.=<<___;
205	vld1.64		{$Xl},[$Xi]		@ load [rotated] Xi
206						@ "[rotated]" means that
207						@ loaded value would have
208						@ to be rotated in order to
209						@ make it appear as in
210						@ algorithm specification
211	subs		$len,$len,#32		@ see if $len is 32 or larger
212	mov		$inc,#16		@ $inc is used as post-
213						@ increment for input pointer;
214						@ as loop is modulo-scheduled
215						@ $inc is zeroed just in time
216						@ to preclude overstepping
217						@ inp[len], which means that
218						@ last block[s] are actually
219						@ loaded twice, but last
220						@ copy is not processed
221	vld1.64		{$H-$Hhl},[$Htbl],#32	@ load twisted H, ..., H^2
222	vmov.i8		$xC2,#0xe1
223	vld1.64		{$H2},[$Htbl]
224	cclr		$inc,eq			@ is it time to zero $inc?
225	vext.8		$Xl,$Xl,$Xl,#8		@ rotate Xi
226	vld1.64		{$t0},[$inp],#16	@ load [rotated] I[0]
227	vshl.u64	$xC2,$xC2,#57		@ compose 0xc2.0 constant
228#ifndef __ARMEB__
229	vrev64.8	$t0,$t0
230	vrev64.8	$Xl,$Xl
231#endif
232	vext.8		$IN,$t0,$t0,#8		@ rotate I[0]
233	b.lo		.Lodd_tail_v8		@ $len was less than 32
234___
235{ my ($Xln,$Xmn,$Xhn,$In) = map("q$_",(4..7));
236	#######
237	# Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
238	#	[(H*Ii+1) + (H*Xi+1)] mod P =
239	#	[(H*Ii+1) + H^2*(Ii+Xi)] mod P
240	#
241$code.=<<___;
242	vld1.64		{$t1},[$inp],$inc	@ load [rotated] I[1]
243#ifndef __ARMEB__
244	vrev64.8	$t1,$t1
245#endif
246	vext.8		$In,$t1,$t1,#8
247	veor		$IN,$IN,$Xl		@ I[i]^=Xi
248	vpmull.p64	$Xln,$H,$In		@ H��Ii+1
249	veor		$t1,$t1,$In		@ Karatsuba pre-processing
250	vpmull2.p64	$Xhn,$H,$In
251	b		.Loop_mod2x_v8
252
253.align	4
254.Loop_mod2x_v8:
255	vext.8		$t2,$IN,$IN,#8
256	subs		$len,$len,#32		@ is there more data?
257	vpmull.p64	$Xl,$H2,$IN		@ H^2.lo��Xi.lo
258	cclr		$inc,lo			@ is it time to zero $inc?
259
260	 vpmull.p64	$Xmn,$Hhl,$t1
261	veor		$t2,$t2,$IN		@ Karatsuba pre-processing
262	vpmull2.p64	$Xh,$H2,$IN		@ H^2.hi��Xi.hi
263	veor		$Xl,$Xl,$Xln		@ accumulate
264	vpmull2.p64	$Xm,$Hhl,$t2		@ (H^2.lo+H^2.hi)��(Xi.lo+Xi.hi)
265	 vld1.64	{$t0},[$inp],$inc	@ load [rotated] I[i+2]
266
267	veor		$Xh,$Xh,$Xhn
268	 cclr		$inc,eq			@ is it time to zero $inc?
269	veor		$Xm,$Xm,$Xmn
270
271	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
272	veor		$t2,$Xl,$Xh
273	veor		$Xm,$Xm,$t1
274	 vld1.64	{$t1},[$inp],$inc	@ load [rotated] I[i+3]
275#ifndef __ARMEB__
276	 vrev64.8	$t0,$t0
277#endif
278	veor		$Xm,$Xm,$t2
279	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
280
281#ifndef __ARMEB__
282	 vrev64.8	$t1,$t1
283#endif
284	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
285	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
286	 vext.8		$In,$t1,$t1,#8
287	 vext.8		$IN,$t0,$t0,#8
288	veor		$Xl,$Xm,$t2
289	 vpmull.p64	$Xln,$H,$In		@ H��Ii+1
290	veor		$IN,$IN,$Xh		@ accumulate $IN early
291
292	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
293	vpmull.p64	$Xl,$Xl,$xC2
294	veor		$IN,$IN,$t2
295	 veor		$t1,$t1,$In		@ Karatsuba pre-processing
296	veor		$IN,$IN,$Xl
297	 vpmull2.p64	$Xhn,$H,$In
298	b.hs		.Loop_mod2x_v8		@ there was at least 32 more bytes
299
300	veor		$Xh,$Xh,$t2
301	vext.8		$IN,$t0,$t0,#8		@ re-construct $IN
302	adds		$len,$len,#32		@ re-construct $len
303	veor		$Xl,$Xl,$Xh		@ re-construct $Xl
304	b.eq		.Ldone_v8		@ is $len zero?
305___
306}
307$code.=<<___;
308.Lodd_tail_v8:
309	vext.8		$t2,$Xl,$Xl,#8
310	veor		$IN,$IN,$Xl		@ inp^=Xi
311	veor		$t1,$t0,$t2		@ $t1 is rotated inp^Xi
312
313	vpmull.p64	$Xl,$H,$IN		@ H.lo��Xi.lo
314	veor		$t1,$t1,$IN		@ Karatsuba pre-processing
315	vpmull2.p64	$Xh,$H,$IN		@ H.hi��Xi.hi
316	vpmull.p64	$Xm,$Hhl,$t1		@ (H.lo+H.hi)��(Xi.lo+Xi.hi)
317
318	vext.8		$t1,$Xl,$Xh,#8		@ Karatsuba post-processing
319	veor		$t2,$Xl,$Xh
320	veor		$Xm,$Xm,$t1
321	veor		$Xm,$Xm,$t2
322	vpmull.p64	$t2,$Xl,$xC2		@ 1st phase of reduction
323
324	vmov		$Xh#lo,$Xm#hi		@ Xh|Xm - 256-bit result
325	vmov		$Xm#hi,$Xl#lo		@ Xm is rotated Xl
326	veor		$Xl,$Xm,$t2
327
328	vext.8		$t2,$Xl,$Xl,#8		@ 2nd phase of reduction
329	vpmull.p64	$Xl,$Xl,$xC2
330	veor		$t2,$t2,$Xh
331	veor		$Xl,$Xl,$t2
332
333.Ldone_v8:
334#ifndef __ARMEB__
335	vrev64.8	$Xl,$Xl
336#endif
337	vext.8		$Xl,$Xl,$Xl,#8
338	vst1.64		{$Xl},[$Xi]		@ write out Xi
339
340___
341$code.=<<___		if ($flavour !~ /64/);
342	vldmia		sp!,{d8-d15}		@ 32-bit ABI says so
343___
344$code.=<<___;
345	ret
346.size	gcm_ghash_v8,.-gcm_ghash_v8
347___
348}
349$code.=<<___;
350.asciz  "GHASH for ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
351.align  2
352___
353
354if ($flavour =~ /64/) {			######## 64-bit code
355    sub unvmov {
356	my $arg=shift;
357
358	$arg =~ m/q([0-9]+)#(lo|hi),\s*q([0-9]+)#(lo|hi)/o &&
359	sprintf	"ins	v%d.d[%d],v%d.d[%d]",$1,($2 eq "lo")?0:1,$3,($4 eq "lo")?0:1;
360    }
361    foreach(split("\n",$code)) {
362	s/cclr\s+([wx])([^,]+),\s*([a-z]+)/csel	$1$2,$1zr,$1$2,$3/o	or
363	s/vmov\.i8/movi/o		or	# fix up legacy mnemonics
364	s/vmov\s+(.*)/unvmov($1)/geo	or
365	s/vext\.8/ext/o			or
366	s/vshr\.s/sshr\.s/o		or
367	s/vshr/ushr/o			or
368	s/^(\s+)v/$1/o			or	# strip off v prefix
369	s/\bbx\s+lr\b/ret/o;
370
371	s/\bq([0-9]+)\b/"v".($1<8?$1:$1+8).".16b"/geo;	# old->new registers
372	s/@\s/\/\//o;				# old->new style commentary
373
374	# fix up remaining legacy suffixes
375	s/\.[ui]?8(\s)/$1/o;
376	s/\.[uis]?32//o and s/\.16b/\.4s/go;
377	m/\.p64/o and s/\.16b/\.1q/o;		# 1st pmull argument
378	m/l\.p64/o and s/\.16b/\.1d/go;		# 2nd and 3rd pmull arguments
379	s/\.[uisp]?64//o and s/\.16b/\.2d/go;
380	s/\.[42]([sd])\[([0-3])\]/\.$1\[$2\]/o;
381
382	print $_,"\n";
383    }
384} else {				######## 32-bit code
385    sub unvdup32 {
386	my $arg=shift;
387
388	$arg =~ m/q([0-9]+),\s*q([0-9]+)\[([0-3])\]/o &&
389	sprintf	"vdup.32	q%d,d%d[%d]",$1,2*$2+($3>>1),$3&1;
390    }
391    sub unvpmullp64 {
392	my ($mnemonic,$arg)=@_;
393
394	if ($arg =~ m/q([0-9]+),\s*q([0-9]+),\s*q([0-9]+)/o) {
395	    my $word = 0xf2a00e00|(($1&7)<<13)|(($1&8)<<19)
396				 |(($2&7)<<17)|(($2&8)<<4)
397				 |(($3&7)<<1) |(($3&8)<<2);
398	    $word |= 0x00010001	 if ($mnemonic =~ "2");
399	    # since ARMv7 instructions are always encoded little-endian.
400	    # correct solution is to use .inst directive, but older
401	    # assemblers don't implement it:-(
402	    sprintf ".byte\t0x%02x,0x%02x,0x%02x,0x%02x\t@ %s %s",
403			$word&0xff,($word>>8)&0xff,
404			($word>>16)&0xff,($word>>24)&0xff,
405			$mnemonic,$arg;
406	}
407    }
408
409    foreach(split("\n",$code)) {
410	s/\b[wx]([0-9]+)\b/r$1/go;		# new->old registers
411	s/\bv([0-9])\.[12468]+[bsd]\b/q$1/go;	# new->old registers
412	s/\/\/\s?/@ /o;				# new->old style commentary
413
414	# fix up remaining new-style suffixes
415	s/\],#[0-9]+/]!/o;
416
417	s/cclr\s+([^,]+),\s*([a-z]+)/mov$2	$1,#0/o			or
418	s/vdup\.32\s+(.*)/unvdup32($1)/geo				or
419	s/v?(pmull2?)\.p64\s+(.*)/unvpmullp64($1,$2)/geo		or
420	s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo	or
421	s/^(\s+)b\./$1b/o						or
422	s/^(\s+)ret/$1bx\tlr/o;
423
424	print $_,"\n";
425    }
426}
427
428close STDOUT; # enforce flush
429