xref: /openssl/crypto/bn/asm/s390x-gf2m.pl (revision 33388b44) 
1#! /usr/bin/env perl
2# Copyright 2011-2020 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# ====================================================================
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# May 2011
18#
19# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
20# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
21# the time being... gcc 4.3 appeared to generate poor code, therefore
22# the effort. And indeed, the module delivers 55%-90%(*) improvement
23# on heaviest ECDSA verify and ECDH benchmarks for 163- and 571-bit
24# key lengths on z990, 30%-55%(*) - on z10, and 70%-110%(*) - on z196.
25# This is for 64-bit build. In 32-bit "highgprs" case improvement is
26# even higher, for example on z990 it was measured 80%-150%. ECDSA
27# sign is modest 9%-12% faster. Keep in mind that these coefficients
28# are not ones for bn_GF2m_mul_2x2 itself, as not all CPU time is
29# burnt in it...
30#
31# (*)	gcc 4.1 was observed to deliver better results than gcc 4.3,
32#	so that improvement coefficients can vary from one specific
33#	setup to another.
34
35# $output is the last argument if it looks like a file (it has an extension)
36# $flavour is the first argument if it doesn't look like a file
37$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
38$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
39
40if ($flavour =~ /3[12]/) {
41        $SIZE_T=4;
42        $g="";
43} else {
44        $SIZE_T=8;
45        $g="g";
46}
47
48$output and open STDOUT,">$output";
49
50$stdframe=16*$SIZE_T+4*8;
51
52$rp="%r2";
53$a1="%r3";
54$a0="%r4";
55$b1="%r5";
56$b0="%r6";
57
58$ra="%r14";
59$sp="%r15";
60
61@T=("%r0","%r1");
62@i=("%r12","%r13");
63
64($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(6..11));
65($lo,$hi,$b)=map("%r$_",(3..5)); $a=$lo; $mask=$a8;
66
67$code.=<<___;
68.text
69
70.type	_mul_1x1,\@function
71.align	16
72_mul_1x1:
73	lgr	$a1,$a
74	sllg	$a2,$a,1
75	sllg	$a4,$a,2
76	sllg	$a8,$a,3
77
78	srag	$lo,$a1,63			# broadcast 63rd bit
79	nihh	$a1,0x1fff
80	srag	@i[0],$a2,63			# broadcast 62nd bit
81	nihh	$a2,0x3fff
82	srag	@i[1],$a4,63			# broadcast 61st bit
83	nihh	$a4,0x7fff
84	ngr	$lo,$b
85	ngr	@i[0],$b
86	ngr	@i[1],$b
87
88	lghi	@T[0],0
89	lgr	$a12,$a1
90	stg	@T[0],`$stdframe+0*8`($sp)	# tab[0]=0
91	xgr	$a12,$a2
92	stg	$a1,`$stdframe+1*8`($sp)	# tab[1]=a1
93	 lgr	$a48,$a4
94	stg	$a2,`$stdframe+2*8`($sp)	# tab[2]=a2
95	 xgr	$a48,$a8
96	stg	$a12,`$stdframe+3*8`($sp)	# tab[3]=a1^a2
97	 xgr	$a1,$a4
98
99	stg	$a4,`$stdframe+4*8`($sp)	# tab[4]=a4
100	xgr	$a2,$a4
101	stg	$a1,`$stdframe+5*8`($sp)	# tab[5]=a1^a4
102	xgr	$a12,$a4
103	stg	$a2,`$stdframe+6*8`($sp)	# tab[6]=a2^a4
104	 xgr	$a1,$a48
105	stg	$a12,`$stdframe+7*8`($sp)	# tab[7]=a1^a2^a4
106	 xgr	$a2,$a48
107
108	stg	$a8,`$stdframe+8*8`($sp)	# tab[8]=a8
109	xgr	$a12,$a48
110	stg	$a1,`$stdframe+9*8`($sp)	# tab[9]=a1^a8
111	 xgr	$a1,$a4
112	stg	$a2,`$stdframe+10*8`($sp)	# tab[10]=a2^a8
113	 xgr	$a2,$a4
114	stg	$a12,`$stdframe+11*8`($sp)	# tab[11]=a1^a2^a8
115
116	xgr	$a12,$a4
117	stg	$a48,`$stdframe+12*8`($sp)	# tab[12]=a4^a8
118	 srlg	$hi,$lo,1
119	stg	$a1,`$stdframe+13*8`($sp)	# tab[13]=a1^a4^a8
120	 sllg	$lo,$lo,63
121	stg	$a2,`$stdframe+14*8`($sp)	# tab[14]=a2^a4^a8
122	 srlg	@T[0],@i[0],2
123	stg	$a12,`$stdframe+15*8`($sp)	# tab[15]=a1^a2^a4^a8
124
125	lghi	$mask,`0xf<<3`
126	sllg	$a1,@i[0],62
127	 sllg	@i[0],$b,3
128	srlg	@T[1],@i[1],3
129	 ngr	@i[0],$mask
130	sllg	$a2,@i[1],61
131	 srlg	@i[1],$b,4-3
132	xgr	$hi,@T[0]
133	 ngr	@i[1],$mask
134	xgr	$lo,$a1
135	xgr	$hi,@T[1]
136	xgr	$lo,$a2
137
138	xg	$lo,$stdframe(@i[0],$sp)
139	srlg	@i[0],$b,8-3
140	ngr	@i[0],$mask
141___
142for($n=1;$n<14;$n++) {
143$code.=<<___;
144	lg	@T[1],$stdframe(@i[1],$sp)
145	srlg	@i[1],$b,`($n+2)*4`-3
146	sllg	@T[0],@T[1],`$n*4`
147	ngr	@i[1],$mask
148	srlg	@T[1],@T[1],`64-$n*4`
149	xgr	$lo,@T[0]
150	xgr	$hi,@T[1]
151___
152	push(@i,shift(@i)); push(@T,shift(@T));
153}
154$code.=<<___;
155	lg	@T[1],$stdframe(@i[1],$sp)
156	sllg	@T[0],@T[1],`$n*4`
157	srlg	@T[1],@T[1],`64-$n*4`
158	xgr	$lo,@T[0]
159	xgr	$hi,@T[1]
160
161	lg	@T[0],$stdframe(@i[0],$sp)
162	sllg	@T[1],@T[0],`($n+1)*4`
163	srlg	@T[0],@T[0],`64-($n+1)*4`
164	xgr	$lo,@T[1]
165	xgr	$hi,@T[0]
166
167	br	$ra
168.size	_mul_1x1,.-_mul_1x1
169
170.globl	bn_GF2m_mul_2x2
171.type	bn_GF2m_mul_2x2,\@function
172.align	16
173bn_GF2m_mul_2x2:
174	stm${g}	%r3,%r15,3*$SIZE_T($sp)
175
176	lghi	%r1,-$stdframe-128
177	la	%r0,0($sp)
178	la	$sp,0(%r1,$sp)			# alloca
179	st${g}	%r0,0($sp)			# back chain
180___
181if ($SIZE_T==8) {
182my @r=map("%r$_",(6..9));
183$code.=<<___;
184	bras	$ra,_mul_1x1			# a1·b1
185	stmg	$lo,$hi,16($rp)
186
187	lg	$a,`$stdframe+128+4*$SIZE_T`($sp)
188	lg	$b,`$stdframe+128+6*$SIZE_T`($sp)
189	bras	$ra,_mul_1x1			# a0·b0
190	stmg	$lo,$hi,0($rp)
191
192	lg	$a,`$stdframe+128+3*$SIZE_T`($sp)
193	lg	$b,`$stdframe+128+5*$SIZE_T`($sp)
194	xg	$a,`$stdframe+128+4*$SIZE_T`($sp)
195	xg	$b,`$stdframe+128+6*$SIZE_T`($sp)
196	bras	$ra,_mul_1x1			# (a0+a1)·(b0+b1)
197	lmg	@r[0],@r[3],0($rp)
198
199	xgr	$lo,$hi
200	xgr	$hi,@r[1]
201	xgr	$lo,@r[0]
202	xgr	$hi,@r[2]
203	xgr	$lo,@r[3]
204	xgr	$hi,@r[3]
205	xgr	$lo,$hi
206	stg	$hi,16($rp)
207	stg	$lo,8($rp)
208___
209} else {
210$code.=<<___;
211	sllg	%r3,%r3,32
212	sllg	%r5,%r5,32
213	or	%r3,%r4
214	or	%r5,%r6
215	bras	$ra,_mul_1x1
216	rllg	$lo,$lo,32
217	rllg	$hi,$hi,32
218	stmg	$lo,$hi,0($rp)
219___
220}
221$code.=<<___;
222	lm${g}	%r6,%r15,`$stdframe+128+6*$SIZE_T`($sp)
223	br	$ra
224.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
225.string	"GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
226___
227
228$code =~ s/\`([^\`]*)\`/eval($1)/gem;
229print $code;
230close STDOUT or die "error closing STDOUT: $!";
231