s390x-gf2m.pl revision 291721 
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# May 2011
11#
12# The module implements bn_GF2m_mul_2x2 polynomial multiplication used
13# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for
14# the time being... gcc 4.3 appeared to generate poor code, therefore
15# the effort. And indeed, the module delivers 55%-90%(*) improvement
16# on haviest ECDSA verify and ECDH benchmarks for 163- and 571-bit
17# key lengths on z990, 30%-55%(*) - on z10, and 70%-110%(*) - on z196.
18# This is for 64-bit build. In 32-bit "highgprs" case improvement is
19# even higher, for example on z990 it was measured 80%-150%. ECDSA
20# sign is modest 9%-12% faster. Keep in mind that these coefficients
21# are not ones for bn_GF2m_mul_2x2 itself, as not all CPU time is
22# burnt in it...
23#
24# (*)	gcc 4.1 was observed to deliver better results than gcc 4.3,
25#	so that improvement coefficients can vary from one specific
26#	setup to another.
27
28$flavour = shift;
29
30if ($flavour =~ /3[12]/) {
31        $SIZE_T=4;
32        $g="";
33} else {
34        $SIZE_T=8;
35        $g="g";
36}
37
38while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
39open STDOUT,">$output";
40
41$stdframe=16*$SIZE_T+4*8;
42
43$rp="%r2";
44$a1="%r3";
45$a0="%r4";
46$b1="%r5";
47$b0="%r6";
48
49$ra="%r14";
50$sp="%r15";
51
52@T=("%r0","%r1");
53@i=("%r12","%r13");
54
55($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(6..11));
56($lo,$hi,$b)=map("%r$_",(3..5)); $a=$lo; $mask=$a8;
57
58$code.=<<___;
59.text
60
61.type	_mul_1x1,\@function
62.align	16
63_mul_1x1:
64	lgr	$a1,$a
65	sllg	$a2,$a,1
66	sllg	$a4,$a,2
67	sllg	$a8,$a,3
68
69	srag	$lo,$a1,63			# broadcast 63rd bit
70	nihh	$a1,0x1fff
71	srag	@i[0],$a2,63			# broadcast 62nd bit
72	nihh	$a2,0x3fff
73	srag	@i[1],$a4,63			# broadcast 61st bit
74	nihh	$a4,0x7fff
75	ngr	$lo,$b
76	ngr	@i[0],$b
77	ngr	@i[1],$b
78
79	lghi	@T[0],0
80	lgr	$a12,$a1
81	stg	@T[0],`$stdframe+0*8`($sp)	# tab[0]=0
82	xgr	$a12,$a2
83	stg	$a1,`$stdframe+1*8`($sp)	# tab[1]=a1
84	 lgr	$a48,$a4
85	stg	$a2,`$stdframe+2*8`($sp)	# tab[2]=a2
86	 xgr	$a48,$a8
87	stg	$a12,`$stdframe+3*8`($sp)	# tab[3]=a1^a2
88	 xgr	$a1,$a4
89
90	stg	$a4,`$stdframe+4*8`($sp)	# tab[4]=a4
91	xgr	$a2,$a4
92	stg	$a1,`$stdframe+5*8`($sp)	# tab[5]=a1^a4
93	xgr	$a12,$a4
94	stg	$a2,`$stdframe+6*8`($sp)	# tab[6]=a2^a4
95	 xgr	$a1,$a48
96	stg	$a12,`$stdframe+7*8`($sp)	# tab[7]=a1^a2^a4
97	 xgr	$a2,$a48
98
99	stg	$a8,`$stdframe+8*8`($sp)	# tab[8]=a8
100	xgr	$a12,$a48
101	stg	$a1,`$stdframe+9*8`($sp)	# tab[9]=a1^a8
102	 xgr	$a1,$a4
103	stg	$a2,`$stdframe+10*8`($sp)	# tab[10]=a2^a8
104	 xgr	$a2,$a4
105	stg	$a12,`$stdframe+11*8`($sp)	# tab[11]=a1^a2^a8
106
107	xgr	$a12,$a4
108	stg	$a48,`$stdframe+12*8`($sp)	# tab[12]=a4^a8
109	 srlg	$hi,$lo,1
110	stg	$a1,`$stdframe+13*8`($sp)	# tab[13]=a1^a4^a8
111	 sllg	$lo,$lo,63
112	stg	$a2,`$stdframe+14*8`($sp)	# tab[14]=a2^a4^a8
113	 srlg	@T[0],@i[0],2
114	stg	$a12,`$stdframe+15*8`($sp)	# tab[15]=a1^a2^a4^a8
115
116	lghi	$mask,`0xf<<3`
117	sllg	$a1,@i[0],62
118	 sllg	@i[0],$b,3
119	srlg	@T[1],@i[1],3
120	 ngr	@i[0],$mask
121	sllg	$a2,@i[1],61
122	 srlg	@i[1],$b,4-3
123	xgr	$hi,@T[0]
124	 ngr	@i[1],$mask
125	xgr	$lo,$a1
126	xgr	$hi,@T[1]
127	xgr	$lo,$a2
128
129	xg	$lo,$stdframe(@i[0],$sp)
130	srlg	@i[0],$b,8-3
131	ngr	@i[0],$mask
132___
133for($n=1;$n<14;$n++) {
134$code.=<<___;
135	lg	@T[1],$stdframe(@i[1],$sp)
136	srlg	@i[1],$b,`($n+2)*4`-3
137	sllg	@T[0],@T[1],`$n*4`
138	ngr	@i[1],$mask
139	srlg	@T[1],@T[1],`64-$n*4`
140	xgr	$lo,@T[0]
141	xgr	$hi,@T[1]
142___
143	push(@i,shift(@i)); push(@T,shift(@T));
144}
145$code.=<<___;
146	lg	@T[1],$stdframe(@i[1],$sp)
147	sllg	@T[0],@T[1],`$n*4`
148	srlg	@T[1],@T[1],`64-$n*4`
149	xgr	$lo,@T[0]
150	xgr	$hi,@T[1]
151
152	lg	@T[0],$stdframe(@i[0],$sp)
153	sllg	@T[1],@T[0],`($n+1)*4`
154	srlg	@T[0],@T[0],`64-($n+1)*4`
155	xgr	$lo,@T[1]
156	xgr	$hi,@T[0]
157
158	br	$ra
159.size	_mul_1x1,.-_mul_1x1
160
161.globl	bn_GF2m_mul_2x2
162.type	bn_GF2m_mul_2x2,\@function
163.align	16
164bn_GF2m_mul_2x2:
165	stm${g}	%r3,%r15,3*$SIZE_T($sp)
166
167	lghi	%r1,-$stdframe-128
168	la	%r0,0($sp)
169	la	$sp,0(%r1,$sp)			# alloca
170	st${g}	%r0,0($sp)			# back chain
171___
172if ($SIZE_T==8) {
173my @r=map("%r$_",(6..9));
174$code.=<<___;
175	bras	$ra,_mul_1x1			# a1��b1
176	stmg	$lo,$hi,16($rp)
177
178	lg	$a,`$stdframe+128+4*$SIZE_T`($sp)
179	lg	$b,`$stdframe+128+6*$SIZE_T`($sp)
180	bras	$ra,_mul_1x1			# a0��b0
181	stmg	$lo,$hi,0($rp)
182
183	lg	$a,`$stdframe+128+3*$SIZE_T`($sp)
184	lg	$b,`$stdframe+128+5*$SIZE_T`($sp)
185	xg	$a,`$stdframe+128+4*$SIZE_T`($sp)
186	xg	$b,`$stdframe+128+6*$SIZE_T`($sp)
187	bras	$ra,_mul_1x1			# (a0+a1)��(b0+b1)
188	lmg	@r[0],@r[3],0($rp)
189
190	xgr	$lo,$hi
191	xgr	$hi,@r[1]
192	xgr	$lo,@r[0]
193	xgr	$hi,@r[2]
194	xgr	$lo,@r[3]
195	xgr	$hi,@r[3]
196	xgr	$lo,$hi
197	stg	$hi,16($rp)
198	stg	$lo,8($rp)
199___
200} else {
201$code.=<<___;
202	sllg	%r3,%r3,32
203	sllg	%r5,%r5,32
204	or	%r3,%r4
205	or	%r5,%r6
206	bras	$ra,_mul_1x1
207	rllg	$lo,$lo,32
208	rllg	$hi,$hi,32
209	stmg	$lo,$hi,0($rp)
210___
211}
212$code.=<<___;
213	lm${g}	%r6,%r15,`$stdframe+128+6*$SIZE_T`($sp)
214	br	$ra
215.size	bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2
216.string	"GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
217___
218
219$code =~ s/\`([^\`]*)\`/eval($1)/gem;
220print $code;
221close STDOUT;
222