1#! /usr/bin/env perl 2# Copyright 2011-2020 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# 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$flavour = shift; 36 37if ($flavour =~ /3[12]/) { 38 $SIZE_T=4; 39 $g=""; 40} else { 41 $SIZE_T=8; 42 $g="g"; 43} 44 45while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} 46open STDOUT,">$output"; 47 48$stdframe=16*$SIZE_T+4*8; 49 50$rp="%r2"; 51$a1="%r3"; 52$a0="%r4"; 53$b1="%r5"; 54$b0="%r6"; 55 56$ra="%r14"; 57$sp="%r15"; 58 59@T=("%r0","%r1"); 60@i=("%r12","%r13"); 61 62($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(6..11)); 63($lo,$hi,$b)=map("%r$_",(3..5)); $a=$lo; $mask=$a8; 64 65$code.=<<___; 66.text 67 68.type _mul_1x1,\@function 69.align 16 70_mul_1x1: 71 lgr $a1,$a 72 sllg $a2,$a,1 73 sllg $a4,$a,2 74 sllg $a8,$a,3 75 76 srag $lo,$a1,63 # broadcast 63rd bit 77 nihh $a1,0x1fff 78 srag @i[0],$a2,63 # broadcast 62nd bit 79 nihh $a2,0x3fff 80 srag @i[1],$a4,63 # broadcast 61st bit 81 nihh $a4,0x7fff 82 ngr $lo,$b 83 ngr @i[0],$b 84 ngr @i[1],$b 85 86 lghi @T[0],0 87 lgr $a12,$a1 88 stg @T[0],`$stdframe+0*8`($sp) # tab[0]=0 89 xgr $a12,$a2 90 stg $a1,`$stdframe+1*8`($sp) # tab[1]=a1 91 lgr $a48,$a4 92 stg $a2,`$stdframe+2*8`($sp) # tab[2]=a2 93 xgr $a48,$a8 94 stg $a12,`$stdframe+3*8`($sp) # tab[3]=a1^a2 95 xgr $a1,$a4 96 97 stg $a4,`$stdframe+4*8`($sp) # tab[4]=a4 98 xgr $a2,$a4 99 stg $a1,`$stdframe+5*8`($sp) # tab[5]=a1^a4 100 xgr $a12,$a4 101 stg $a2,`$stdframe+6*8`($sp) # tab[6]=a2^a4 102 xgr $a1,$a48 103 stg $a12,`$stdframe+7*8`($sp) # tab[7]=a1^a2^a4 104 xgr $a2,$a48 105 106 stg $a8,`$stdframe+8*8`($sp) # tab[8]=a8 107 xgr $a12,$a48 108 stg $a1,`$stdframe+9*8`($sp) # tab[9]=a1^a8 109 xgr $a1,$a4 110 stg $a2,`$stdframe+10*8`($sp) # tab[10]=a2^a8 111 xgr $a2,$a4 112 stg $a12,`$stdframe+11*8`($sp) # tab[11]=a1^a2^a8 113 114 xgr $a12,$a4 115 stg $a48,`$stdframe+12*8`($sp) # tab[12]=a4^a8 116 srlg $hi,$lo,1 117 stg $a1,`$stdframe+13*8`($sp) # tab[13]=a1^a4^a8 118 sllg $lo,$lo,63 119 stg $a2,`$stdframe+14*8`($sp) # tab[14]=a2^a4^a8 120 srlg @T[0],@i[0],2 121 stg $a12,`$stdframe+15*8`($sp) # tab[15]=a1^a2^a4^a8 122 123 lghi $mask,`0xf<<3` 124 sllg $a1,@i[0],62 125 sllg @i[0],$b,3 126 srlg @T[1],@i[1],3 127 ngr @i[0],$mask 128 sllg $a2,@i[1],61 129 srlg @i[1],$b,4-3 130 xgr $hi,@T[0] 131 ngr @i[1],$mask 132 xgr $lo,$a1 133 xgr $hi,@T[1] 134 xgr $lo,$a2 135 136 xg $lo,$stdframe(@i[0],$sp) 137 srlg @i[0],$b,8-3 138 ngr @i[0],$mask 139___ 140for($n=1;$n<14;$n++) { 141$code.=<<___; 142 lg @T[1],$stdframe(@i[1],$sp) 143 srlg @i[1],$b,`($n+2)*4`-3 144 sllg @T[0],@T[1],`$n*4` 145 ngr @i[1],$mask 146 srlg @T[1],@T[1],`64-$n*4` 147 xgr $lo,@T[0] 148 xgr $hi,@T[1] 149___ 150 push(@i,shift(@i)); push(@T,shift(@T)); 151} 152$code.=<<___; 153 lg @T[1],$stdframe(@i[1],$sp) 154 sllg @T[0],@T[1],`$n*4` 155 srlg @T[1],@T[1],`64-$n*4` 156 xgr $lo,@T[0] 157 xgr $hi,@T[1] 158 159 lg @T[0],$stdframe(@i[0],$sp) 160 sllg @T[1],@T[0],`($n+1)*4` 161 srlg @T[0],@T[0],`64-($n+1)*4` 162 xgr $lo,@T[1] 163 xgr $hi,@T[0] 164 165 br $ra 166.size _mul_1x1,.-_mul_1x1 167 168.globl bn_GF2m_mul_2x2 169.type bn_GF2m_mul_2x2,\@function 170.align 16 171bn_GF2m_mul_2x2: 172 stm${g} %r3,%r15,3*$SIZE_T($sp) 173 174 lghi %r1,-$stdframe-128 175 la %r0,0($sp) 176 la $sp,0(%r1,$sp) # alloca 177 st${g} %r0,0($sp) # back chain 178___ 179if ($SIZE_T==8) { 180my @r=map("%r$_",(6..9)); 181$code.=<<___; 182 bras $ra,_mul_1x1 # a1·b1 183 stmg $lo,$hi,16($rp) 184 185 lg $a,`$stdframe+128+4*$SIZE_T`($sp) 186 lg $b,`$stdframe+128+6*$SIZE_T`($sp) 187 bras $ra,_mul_1x1 # a0·b0 188 stmg $lo,$hi,0($rp) 189 190 lg $a,`$stdframe+128+3*$SIZE_T`($sp) 191 lg $b,`$stdframe+128+5*$SIZE_T`($sp) 192 xg $a,`$stdframe+128+4*$SIZE_T`($sp) 193 xg $b,`$stdframe+128+6*$SIZE_T`($sp) 194 bras $ra,_mul_1x1 # (a0+a1)·(b0+b1) 195 lmg @r[0],@r[3],0($rp) 196 197 xgr $lo,$hi 198 xgr $hi,@r[1] 199 xgr $lo,@r[0] 200 xgr $hi,@r[2] 201 xgr $lo,@r[3] 202 xgr $hi,@r[3] 203 xgr $lo,$hi 204 stg $hi,16($rp) 205 stg $lo,8($rp) 206___ 207} else { 208$code.=<<___; 209 sllg %r3,%r3,32 210 sllg %r5,%r5,32 211 or %r3,%r4 212 or %r5,%r6 213 bras $ra,_mul_1x1 214 rllg $lo,$lo,32 215 rllg $hi,$hi,32 216 stmg $lo,$hi,0($rp) 217___ 218} 219$code.=<<___; 220 lm${g} %r6,%r15,`$stdframe+128+6*$SIZE_T`($sp) 221 br $ra 222.size bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2 223.string "GF(2^m) Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>" 224___ 225 226$code =~ s/\`([^\`]*)\`/eval($1)/gem; 227print $code; 228close STDOUT or die "error closing STDOUT: $!"; 229