1#!/usr/bin/env perl 2 3# ==================================================================== 4# Written by Andy Polyakov <appro@fy.chalmers.se> 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# April 2007. 11# 12# Performance improvement over vanilla C code varies from 85% to 45% 13# depending on key length and benchmark. Unfortunately in this context 14# these are not very impressive results [for code that utilizes "wide" 15# 64x64=128-bit multiplication, which is not commonly available to C 16# programmers], at least hand-coded bn_asm.c replacement is known to 17# provide 30-40% better results for longest keys. Well, on a second 18# thought it's not very surprising, because z-CPUs are single-issue 19# and _strictly_ in-order execution, while bn_mul_mont is more or less 20# dependent on CPU ability to pipe-line instructions and have several 21# of them "in-flight" at the same time. I mean while other methods, 22# for example Karatsuba, aim to minimize amount of multiplications at 23# the cost of other operations increase, bn_mul_mont aim to neatly 24# "overlap" multiplications and the other operations [and on most 25# platforms even minimize the amount of the other operations, in 26# particular references to memory]. But it's possible to improve this 27# module performance by implementing dedicated squaring code-path and 28# possibly by unrolling loops... 29 30# January 2009. 31# 32# Reschedule to minimize/avoid Address Generation Interlock hazard, 33# make inner loops counter-based. 34 35# November 2010. 36# 37# Adapt for -m31 build. If kernel supports what's called "highgprs" 38# feature on Linux [see /proc/cpuinfo], it's possible to use 64-bit 39# instructions and achieve "64-bit" performance even in 31-bit legacy 40# application context. The feature is not specific to any particular 41# processor, as long as it's "z-CPU". Latter implies that the code 42# remains z/Architecture specific. Compatibility with 32-bit BN_ULONG 43# is achieved by swapping words after 64-bit loads, follow _dswap-s. 44# On z990 it was measured to perform 2.6-2.2 times better than 45# compiler-generated code, less for longer keys... 46 47$flavour = shift; 48 49if ($flavour =~ /3[12]/) { 50 $SIZE_T=4; 51 $g=""; 52} else { 53 $SIZE_T=8; 54 $g="g"; 55} 56 57while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {} 58open STDOUT,">$output"; 59 60$stdframe=16*$SIZE_T+4*8; 61 62$mn0="%r0"; 63$num="%r1"; 64 65# int bn_mul_mont( 66$rp="%r2"; # BN_ULONG *rp, 67$ap="%r3"; # const BN_ULONG *ap, 68$bp="%r4"; # const BN_ULONG *bp, 69$np="%r5"; # const BN_ULONG *np, 70$n0="%r6"; # const BN_ULONG *n0, 71#$num="160(%r15)" # int num); 72 73$bi="%r2"; # zaps rp 74$j="%r7"; 75 76$ahi="%r8"; 77$alo="%r9"; 78$nhi="%r10"; 79$nlo="%r11"; 80$AHI="%r12"; 81$NHI="%r13"; 82$count="%r14"; 83$sp="%r15"; 84 85$code.=<<___; 86.text 87.globl bn_mul_mont 88.type bn_mul_mont,\@function 89bn_mul_mont: 90 lgf $num,`$stdframe+$SIZE_T-4`($sp) # pull $num 91 sla $num,`log($SIZE_T)/log(2)` # $num to enumerate bytes 92 la $bp,0($num,$bp) 93 94 st${g} %r2,2*$SIZE_T($sp) 95 96 cghi $num,16 # 97 lghi %r2,0 # 98 blr %r14 # if($num<16) return 0; 99___ 100$code.=<<___ if ($flavour =~ /3[12]/); 101 tmll $num,4 102 bnzr %r14 # if ($num&1) return 0; 103___ 104$code.=<<___ if ($flavour !~ /3[12]/); 105 cghi $num,96 # 106 bhr %r14 # if($num>96) return 0; 107___ 108$code.=<<___; 109 stm${g} %r3,%r15,3*$SIZE_T($sp) 110 111 lghi $rp,-$stdframe-8 # leave room for carry bit 112 lcgr $j,$num # -$num 113 lgr %r0,$sp 114 la $rp,0($rp,$sp) 115 la $sp,0($j,$rp) # alloca 116 st${g} %r0,0($sp) # back chain 117 118 sra $num,3 # restore $num 119 la $bp,0($j,$bp) # restore $bp 120 ahi $num,-1 # adjust $num for inner loop 121 lg $n0,0($n0) # pull n0 122 _dswap $n0 123 124 lg $bi,0($bp) 125 _dswap $bi 126 lg $alo,0($ap) 127 _dswap $alo 128 mlgr $ahi,$bi # ap[0]*bp[0] 129 lgr $AHI,$ahi 130 131 lgr $mn0,$alo # "tp[0]"*n0 132 msgr $mn0,$n0 133 134 lg $nlo,0($np) # 135 _dswap $nlo 136 mlgr $nhi,$mn0 # np[0]*m1 137 algr $nlo,$alo # +="tp[0]" 138 lghi $NHI,0 139 alcgr $NHI,$nhi 140 141 la $j,8(%r0) # j=1 142 lr $count,$num 143 144.align 16 145.L1st: 146 lg $alo,0($j,$ap) 147 _dswap $alo 148 mlgr $ahi,$bi # ap[j]*bp[0] 149 algr $alo,$AHI 150 lghi $AHI,0 151 alcgr $AHI,$ahi 152 153 lg $nlo,0($j,$np) 154 _dswap $nlo 155 mlgr $nhi,$mn0 # np[j]*m1 156 algr $nlo,$NHI 157 lghi $NHI,0 158 alcgr $nhi,$NHI # +="tp[j]" 159 algr $nlo,$alo 160 alcgr $NHI,$nhi 161 162 stg $nlo,$stdframe-8($j,$sp) # tp[j-1]= 163 la $j,8($j) # j++ 164 brct $count,.L1st 165 166 algr $NHI,$AHI 167 lghi $AHI,0 168 alcgr $AHI,$AHI # upmost overflow bit 169 stg $NHI,$stdframe-8($j,$sp) 170 stg $AHI,$stdframe($j,$sp) 171 la $bp,8($bp) # bp++ 172 173.Louter: 174 lg $bi,0($bp) # bp[i] 175 _dswap $bi 176 lg $alo,0($ap) 177 _dswap $alo 178 mlgr $ahi,$bi # ap[0]*bp[i] 179 alg $alo,$stdframe($sp) # +=tp[0] 180 lghi $AHI,0 181 alcgr $AHI,$ahi 182 183 lgr $mn0,$alo 184 msgr $mn0,$n0 # tp[0]*n0 185 186 lg $nlo,0($np) # np[0] 187 _dswap $nlo 188 mlgr $nhi,$mn0 # np[0]*m1 189 algr $nlo,$alo # +="tp[0]" 190 lghi $NHI,0 191 alcgr $NHI,$nhi 192 193 la $j,8(%r0) # j=1 194 lr $count,$num 195 196.align 16 197.Linner: 198 lg $alo,0($j,$ap) 199 _dswap $alo 200 mlgr $ahi,$bi # ap[j]*bp[i] 201 algr $alo,$AHI 202 lghi $AHI,0 203 alcgr $ahi,$AHI 204 alg $alo,$stdframe($j,$sp)# +=tp[j] 205 alcgr $AHI,$ahi 206 207 lg $nlo,0($j,$np) 208 _dswap $nlo 209 mlgr $nhi,$mn0 # np[j]*m1 210 algr $nlo,$NHI 211 lghi $NHI,0 212 alcgr $nhi,$NHI 213 algr $nlo,$alo # +="tp[j]" 214 alcgr $NHI,$nhi 215 216 stg $nlo,$stdframe-8($j,$sp) # tp[j-1]= 217 la $j,8($j) # j++ 218 brct $count,.Linner 219 220 algr $NHI,$AHI 221 lghi $AHI,0 222 alcgr $AHI,$AHI 223 alg $NHI,$stdframe($j,$sp)# accumulate previous upmost overflow bit 224 lghi $ahi,0 225 alcgr $AHI,$ahi # new upmost overflow bit 226 stg $NHI,$stdframe-8($j,$sp) 227 stg $AHI,$stdframe($j,$sp) 228 229 la $bp,8($bp) # bp++ 230 cl${g} $bp,`$stdframe+8+4*$SIZE_T`($j,$sp) # compare to &bp[num] 231 jne .Louter 232 233 l${g} $rp,`$stdframe+8+2*$SIZE_T`($j,$sp) # reincarnate rp 234 la $ap,$stdframe($sp) 235 ahi $num,1 # restore $num, incidentally clears "borrow" 236 237 la $j,0(%r0) 238 lr $count,$num 239.Lsub: lg $alo,0($j,$ap) 240 lg $nlo,0($j,$np) 241 _dswap $nlo 242 slbgr $alo,$nlo 243 stg $alo,0($j,$rp) 244 la $j,8($j) 245 brct $count,.Lsub 246 lghi $ahi,0 247 slbgr $AHI,$ahi # handle upmost carry 248 249 ngr $ap,$AHI 250 lghi $np,-1 251 xgr $np,$AHI 252 ngr $np,$rp 253 ogr $ap,$np # ap=borrow?tp:rp 254 255 la $j,0(%r0) 256 lgr $count,$num 257.Lcopy: lg $alo,0($j,$ap) # copy or in-place refresh 258 _dswap $alo 259 stg $j,$stdframe($j,$sp) # zap tp 260 stg $alo,0($j,$rp) 261 la $j,8($j) 262 brct $count,.Lcopy 263 264 la %r1,`$stdframe+8+6*$SIZE_T`($j,$sp) 265 lm${g} %r6,%r15,0(%r1) 266 lghi %r2,1 # signal "processed" 267 br %r14 268.size bn_mul_mont,.-bn_mul_mont 269.string "Montgomery Multiplication for s390x, CRYPTOGAMS by <appro\@openssl.org>" 270___ 271 272foreach (split("\n",$code)) { 273 s/\`([^\`]*)\`/eval $1/ge; 274 s/_dswap\s+(%r[0-9]+)/sprintf("rllg\t%s,%s,32",$1,$1) if($SIZE_T==4)/e; 275 print $_,"\n"; 276} 277close STDOUT; 278