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 # ====================================================================
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
24 # (*) gcc 4.1 was observed to deliver better results than gcc 4.3,
25 # so that improvement coefficients can vary from one specific
30 if ($flavour =~ /3[12]/) {
38 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
39 open STDOUT
,">$output";
41 $stdframe=16*$SIZE_T+4*8;
55 ($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(6..11));
56 ($lo,$hi,$b)=map("%r$_",(3..5)); $a=$lo; $mask=$a8;
61 .type _mul_1x1
,\
@function
69 srag
$lo,$a1,63 # broadcast 63rd bit
71 srag
@i[0],$a2,63 # broadcast 62nd bit
73 srag
@i[1],$a4,63 # broadcast 61st bit
81 stg
@T[0],`$stdframe+0*8`($sp) # tab[0]=0
83 stg
$a1,`$stdframe+1*8`($sp) # tab[1]=a1
85 stg
$a2,`$stdframe+2*8`($sp) # tab[2]=a2
87 stg
$a12,`$stdframe+3*8`($sp) # tab[3]=a1^a2
90 stg
$a4,`$stdframe+4*8`($sp) # tab[4]=a4
92 stg
$a1,`$stdframe+5*8`($sp) # tab[5]=a1^a4
94 stg
$a2,`$stdframe+6*8`($sp) # tab[6]=a2^a4
96 stg
$a12,`$stdframe+7*8`($sp) # tab[7]=a1^a2^a4
99 stg
$a8,`$stdframe+8*8`($sp) # tab[8]=a8
101 stg
$a1,`$stdframe+9*8`($sp) # tab[9]=a1^a8
103 stg
$a2,`$stdframe+10*8`($sp) # tab[10]=a2^a8
105 stg
$a12,`$stdframe+11*8`($sp) # tab[11]=a1^a2^a8
108 stg
$a48,`$stdframe+12*8`($sp) # tab[12]=a4^a8
110 stg
$a1,`$stdframe+13*8`($sp) # tab[13]=a1^a4^a8
112 stg
$a2,`$stdframe+14*8`($sp) # tab[14]=a2^a4^a8
114 stg
$a12,`$stdframe+15*8`($sp) # tab[15]=a1^a2^a4^a8
129 xg
$lo,$stdframe(@i[0],$sp)
133 for($n=1;$n<14;$n++) {
135 lg
@T[1],$stdframe(@i[1],$sp)
136 srlg
@i[1],$b,`($n+2)*4`-3
137 sllg
@T[0],@T[1],`$n*4`
139 srlg
@T[1],@T[1],`64-$n*4`
143 push(@i,shift(@i)); push(@T,shift(@T));
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`
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`
159 .size _mul_1x1
,.-_mul_1x1
161 .globl bn_GF2m_mul_2x2
162 .type bn_GF2m_mul_2x2
,\
@function
165 stm
${g
} %r3,%r15,3*$SIZE_T($sp)
167 lghi
%r1,-$stdframe-128
169 la
$sp,0(%r1,$sp) # alloca
170 st
${g
} %r0,0($sp) # back chain
173 my @r=map("%r$_",(6..9));
175 bras
$ra,_mul_1x1
# a1·b1
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
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)
213 lm
${g
} %r6,%r15,`$stdframe+128+6*$SIZE_T`($sp)
215 .size bn_GF2m_mul_2x2
,.-bn_GF2m_mul_2x2
216 .string
"GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>"
219 $code =~ s/\`([^\`]*)\`/eval($1)/gem;