| 1 | #! /usr/bin/env perl
|
|---|
| 2 | # Copyright 1995-2020 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 | $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
|
|---|
| 10 | push(@INC,"${dir}","${dir}../../perlasm");
|
|---|
| 11 | require "x86asm.pl";
|
|---|
| 12 |
|
|---|
| 13 | $output = pop;
|
|---|
| 14 | open STDOUT,">$output";
|
|---|
| 15 |
|
|---|
| 16 | &asm_init($ARGV[0]);
|
|---|
| 17 |
|
|---|
| 18 | &bn_mul_comba("bn_mul_comba8",8);
|
|---|
| 19 | &bn_mul_comba("bn_mul_comba4",4);
|
|---|
| 20 | &bn_sqr_comba("bn_sqr_comba8",8);
|
|---|
| 21 | &bn_sqr_comba("bn_sqr_comba4",4);
|
|---|
| 22 |
|
|---|
| 23 | &asm_finish();
|
|---|
| 24 |
|
|---|
| 25 | close STDOUT or die "error closing STDOUT: $!";
|
|---|
| 26 |
|
|---|
| 27 | sub mul_add_c
|
|---|
| 28 | {
|
|---|
| 29 | local($a,$ai,$b,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
|
|---|
| 30 |
|
|---|
| 31 | # pos == -1 if eax and edx are pre-loaded, 0 to load from next
|
|---|
| 32 | # words, and 1 if load return value
|
|---|
| 33 |
|
|---|
| 34 | &comment("mul a[$ai]*b[$bi]");
|
|---|
| 35 |
|
|---|
| 36 | # "eax" and "edx" will always be pre-loaded.
|
|---|
| 37 | # &mov("eax",&DWP($ai*4,$a,"",0)) ;
|
|---|
| 38 | # &mov("edx",&DWP($bi*4,$b,"",0));
|
|---|
| 39 |
|
|---|
| 40 | &mul("edx");
|
|---|
| 41 | &add($c0,"eax");
|
|---|
| 42 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0; # load next a
|
|---|
| 43 | &mov("eax",&wparam(0)) if $pos > 0; # load r[]
|
|---|
| 44 | ###
|
|---|
| 45 | &adc($c1,"edx");
|
|---|
| 46 | &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 0; # load next b
|
|---|
| 47 | &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 1; # load next b
|
|---|
| 48 | ###
|
|---|
| 49 | &adc($c2,0);
|
|---|
| 50 | # is pos > 1, it means it is the last loop
|
|---|
| 51 | &mov(&DWP($i*4,"eax","",0),$c0) if $pos > 0; # save r[];
|
|---|
| 52 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1; # load next a
|
|---|
| 53 | }
|
|---|
| 54 |
|
|---|
| 55 | sub sqr_add_c
|
|---|
| 56 | {
|
|---|
| 57 | local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
|
|---|
| 58 |
|
|---|
| 59 | # pos == -1 if eax and edx are pre-loaded, 0 to load from next
|
|---|
| 60 | # words, and 1 if load return value
|
|---|
| 61 |
|
|---|
| 62 | &comment("sqr a[$ai]*a[$bi]");
|
|---|
| 63 |
|
|---|
| 64 | # "eax" and "edx" will always be pre-loaded.
|
|---|
| 65 | # &mov("eax",&DWP($ai*4,$a,"",0)) ;
|
|---|
| 66 | # &mov("edx",&DWP($bi*4,$b,"",0));
|
|---|
| 67 |
|
|---|
| 68 | if ($ai == $bi)
|
|---|
| 69 | { &mul("eax");}
|
|---|
| 70 | else
|
|---|
| 71 | { &mul("edx");}
|
|---|
| 72 | &add($c0,"eax");
|
|---|
| 73 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0; # load next a
|
|---|
| 74 | ###
|
|---|
| 75 | &adc($c1,"edx");
|
|---|
| 76 | &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos == 1) && ($na != $nb);
|
|---|
| 77 | ###
|
|---|
| 78 | &adc($c2,0);
|
|---|
| 79 | # is pos > 1, it means it is the last loop
|
|---|
| 80 | &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0; # save r[];
|
|---|
| 81 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1; # load next b
|
|---|
| 82 | }
|
|---|
| 83 |
|
|---|
| 84 | sub sqr_add_c2
|
|---|
| 85 | {
|
|---|
| 86 | local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
|
|---|
| 87 |
|
|---|
| 88 | # pos == -1 if eax and edx are pre-loaded, 0 to load from next
|
|---|
| 89 | # words, and 1 if load return value
|
|---|
| 90 |
|
|---|
| 91 | &comment("sqr a[$ai]*a[$bi]");
|
|---|
| 92 |
|
|---|
| 93 | # "eax" and "edx" will always be pre-loaded.
|
|---|
| 94 | # &mov("eax",&DWP($ai*4,$a,"",0)) ;
|
|---|
| 95 | # &mov("edx",&DWP($bi*4,$a,"",0));
|
|---|
| 96 |
|
|---|
| 97 | if ($ai == $bi)
|
|---|
| 98 | { &mul("eax");}
|
|---|
| 99 | else
|
|---|
| 100 | { &mul("edx");}
|
|---|
| 101 | &add("eax","eax");
|
|---|
| 102 | ###
|
|---|
| 103 | &adc("edx","edx");
|
|---|
| 104 | ###
|
|---|
| 105 | &adc($c2,0);
|
|---|
| 106 | &add($c0,"eax");
|
|---|
| 107 | &adc($c1,"edx");
|
|---|
| 108 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0; # load next a
|
|---|
| 109 | &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1; # load next b
|
|---|
| 110 | &adc($c2,0);
|
|---|
| 111 | &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0; # save r[];
|
|---|
| 112 | &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos <= 1) && ($na != $nb);
|
|---|
| 113 | ###
|
|---|
| 114 | }
|
|---|
| 115 |
|
|---|
| 116 | sub bn_mul_comba
|
|---|
| 117 | {
|
|---|
| 118 | local($name,$num)=@_;
|
|---|
| 119 | local($a,$b,$c0,$c1,$c2);
|
|---|
| 120 | local($i,$as,$ae,$bs,$be,$ai,$bi);
|
|---|
| 121 | local($tot,$end);
|
|---|
| 122 |
|
|---|
| 123 | &function_begin_B($name,"");
|
|---|
| 124 |
|
|---|
| 125 | $c0="ebx";
|
|---|
| 126 | $c1="ecx";
|
|---|
| 127 | $c2="ebp";
|
|---|
| 128 | $a="esi";
|
|---|
| 129 | $b="edi";
|
|---|
| 130 |
|
|---|
| 131 | $as=0;
|
|---|
| 132 | $ae=0;
|
|---|
| 133 | $bs=0;
|
|---|
| 134 | $be=0;
|
|---|
| 135 | $tot=$num+$num-1;
|
|---|
| 136 |
|
|---|
| 137 | &push("esi");
|
|---|
| 138 | &mov($a,&wparam(1));
|
|---|
| 139 | &push("edi");
|
|---|
| 140 | &mov($b,&wparam(2));
|
|---|
| 141 | &push("ebp");
|
|---|
| 142 | &push("ebx");
|
|---|
| 143 |
|
|---|
| 144 | &xor($c0,$c0);
|
|---|
| 145 | &mov("eax",&DWP(0,$a,"",0)); # load the first word
|
|---|
| 146 | &xor($c1,$c1);
|
|---|
| 147 | &mov("edx",&DWP(0,$b,"",0)); # load the first second
|
|---|
| 148 |
|
|---|
| 149 | for ($i=0; $i<$tot; $i++)
|
|---|
| 150 | {
|
|---|
| 151 | $ai=$as;
|
|---|
| 152 | $bi=$bs;
|
|---|
| 153 | $end=$be+1;
|
|---|
| 154 |
|
|---|
| 155 | &comment("################## Calculate word $i");
|
|---|
| 156 |
|
|---|
| 157 | for ($j=$bs; $j<$end; $j++)
|
|---|
| 158 | {
|
|---|
| 159 | &xor($c2,$c2) if ($j == $bs);
|
|---|
| 160 | if (($j+1) == $end)
|
|---|
| 161 | {
|
|---|
| 162 | $v=1;
|
|---|
| 163 | $v=2 if (($i+1) == $tot);
|
|---|
| 164 | }
|
|---|
| 165 | else
|
|---|
| 166 | { $v=0; }
|
|---|
| 167 | if (($j+1) != $end)
|
|---|
| 168 | {
|
|---|
| 169 | $na=($ai-1);
|
|---|
| 170 | $nb=($bi+1);
|
|---|
| 171 | }
|
|---|
| 172 | else
|
|---|
| 173 | {
|
|---|
| 174 | $na=$as+($i < ($num-1));
|
|---|
| 175 | $nb=$bs+($i >= ($num-1));
|
|---|
| 176 | }
|
|---|
| 177 | #printf STDERR "[$ai,$bi] -> [$na,$nb]\n";
|
|---|
| 178 | &mul_add_c($a,$ai,$b,$bi,$c0,$c1,$c2,$v,$i,$na,$nb);
|
|---|
| 179 | if ($v)
|
|---|
| 180 | {
|
|---|
| 181 | &comment("saved r[$i]");
|
|---|
| 182 | # &mov("eax",&wparam(0));
|
|---|
| 183 | # &mov(&DWP($i*4,"eax","",0),$c0);
|
|---|
| 184 | ($c0,$c1,$c2)=($c1,$c2,$c0);
|
|---|
| 185 | }
|
|---|
| 186 | $ai--;
|
|---|
| 187 | $bi++;
|
|---|
| 188 | }
|
|---|
| 189 | $as++ if ($i < ($num-1));
|
|---|
| 190 | $ae++ if ($i >= ($num-1));
|
|---|
| 191 |
|
|---|
| 192 | $bs++ if ($i >= ($num-1));
|
|---|
| 193 | $be++ if ($i < ($num-1));
|
|---|
| 194 | }
|
|---|
| 195 | &comment("save r[$i]");
|
|---|
| 196 | # &mov("eax",&wparam(0));
|
|---|
| 197 | &mov(&DWP($i*4,"eax","",0),$c0);
|
|---|
| 198 |
|
|---|
| 199 | &pop("ebx");
|
|---|
| 200 | &pop("ebp");
|
|---|
| 201 | &pop("edi");
|
|---|
| 202 | &pop("esi");
|
|---|
| 203 | &ret();
|
|---|
| 204 | &function_end_B($name);
|
|---|
| 205 | }
|
|---|
| 206 |
|
|---|
| 207 | sub bn_sqr_comba
|
|---|
| 208 | {
|
|---|
| 209 | local($name,$num)=@_;
|
|---|
| 210 | local($r,$a,$c0,$c1,$c2)=@_;
|
|---|
| 211 | local($i,$as,$ae,$bs,$be,$ai,$bi);
|
|---|
| 212 | local($b,$tot,$end,$half);
|
|---|
| 213 |
|
|---|
| 214 | &function_begin_B($name,"");
|
|---|
| 215 |
|
|---|
| 216 | $c0="ebx";
|
|---|
| 217 | $c1="ecx";
|
|---|
| 218 | $c2="ebp";
|
|---|
| 219 | $a="esi";
|
|---|
| 220 | $r="edi";
|
|---|
| 221 |
|
|---|
| 222 | &push("esi");
|
|---|
| 223 | &push("edi");
|
|---|
| 224 | &push("ebp");
|
|---|
| 225 | &push("ebx");
|
|---|
| 226 | &mov($r,&wparam(0));
|
|---|
| 227 | &mov($a,&wparam(1));
|
|---|
| 228 | &xor($c0,$c0);
|
|---|
| 229 | &xor($c1,$c1);
|
|---|
| 230 | &mov("eax",&DWP(0,$a,"",0)); # load the first word
|
|---|
| 231 |
|
|---|
| 232 | $as=0;
|
|---|
| 233 | $ae=0;
|
|---|
| 234 | $bs=0;
|
|---|
| 235 | $be=0;
|
|---|
| 236 | $tot=$num+$num-1;
|
|---|
| 237 |
|
|---|
| 238 | for ($i=0; $i<$tot; $i++)
|
|---|
| 239 | {
|
|---|
| 240 | $ai=$as;
|
|---|
| 241 | $bi=$bs;
|
|---|
| 242 | $end=$be+1;
|
|---|
| 243 |
|
|---|
| 244 | &comment("############### Calculate word $i");
|
|---|
| 245 | for ($j=$bs; $j<$end; $j++)
|
|---|
| 246 | {
|
|---|
| 247 | &xor($c2,$c2) if ($j == $bs);
|
|---|
| 248 | if (($ai-1) < ($bi+1))
|
|---|
| 249 | {
|
|---|
| 250 | $v=1;
|
|---|
| 251 | $v=2 if ($i+1) == $tot;
|
|---|
| 252 | }
|
|---|
| 253 | else
|
|---|
| 254 | { $v=0; }
|
|---|
| 255 | if (!$v)
|
|---|
| 256 | {
|
|---|
| 257 | $na=$ai-1;
|
|---|
| 258 | $nb=$bi+1;
|
|---|
| 259 | }
|
|---|
| 260 | else
|
|---|
| 261 | {
|
|---|
| 262 | $na=$as+($i < ($num-1));
|
|---|
| 263 | $nb=$bs+($i >= ($num-1));
|
|---|
| 264 | }
|
|---|
| 265 | if ($ai == $bi)
|
|---|
| 266 | {
|
|---|
| 267 | &sqr_add_c($r,$a,$ai,$bi,
|
|---|
| 268 | $c0,$c1,$c2,$v,$i,$na,$nb);
|
|---|
| 269 | }
|
|---|
| 270 | else
|
|---|
| 271 | {
|
|---|
| 272 | &sqr_add_c2($r,$a,$ai,$bi,
|
|---|
| 273 | $c0,$c1,$c2,$v,$i,$na,$nb);
|
|---|
| 274 | }
|
|---|
| 275 | if ($v)
|
|---|
| 276 | {
|
|---|
| 277 | &comment("saved r[$i]");
|
|---|
| 278 | #&mov(&DWP($i*4,$r,"",0),$c0);
|
|---|
| 279 | ($c0,$c1,$c2)=($c1,$c2,$c0);
|
|---|
| 280 | last;
|
|---|
| 281 | }
|
|---|
| 282 | $ai--;
|
|---|
| 283 | $bi++;
|
|---|
| 284 | }
|
|---|
| 285 | $as++ if ($i < ($num-1));
|
|---|
| 286 | $ae++ if ($i >= ($num-1));
|
|---|
| 287 |
|
|---|
| 288 | $bs++ if ($i >= ($num-1));
|
|---|
| 289 | $be++ if ($i < ($num-1));
|
|---|
| 290 | }
|
|---|
| 291 | &mov(&DWP($i*4,$r,"",0),$c0);
|
|---|
| 292 | &pop("ebx");
|
|---|
| 293 | &pop("ebp");
|
|---|
| 294 | &pop("edi");
|
|---|
| 295 | &pop("esi");
|
|---|
| 296 | &ret();
|
|---|
| 297 | &function_end_B($name);
|
|---|
| 298 | }
|
|---|
