crypto/openssl-0.9.7d/crypto/bn/asm/co-586.pl

   1 #!/usr/local/bin/perl
   2
   3 push(@INC,"perlasm","../../perlasm");
   4 require "x86asm.pl";
   5
   6 &asm_init($ARGV[0],$0);
   7
   8 &bn_mul_comba("bn_mul_comba8",8);
   9 &bn_mul_comba("bn_mul_comba4",4);
  10 &bn_sqr_comba("bn_sqr_comba8",8);
  11 &bn_sqr_comba("bn_sqr_comba4",4);
  12
  13 &asm_finish();
  14
  15 sub mul_add_c
  16         {
  17         local($a,$ai,$b,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
  18
  19         # pos == -1 if eax and edx are pre-loaded, 0 to load from next
  20         # words, and 1 if load return value
  21
  22         &comment("mul a[$ai]*b[$bi]");
  23
  24         # "eax" and "edx" will always be pre-loaded.
  25         # &mov("eax",&DWP($ai*4,$a,"",0)) ;
  26         # &mov("edx",&DWP($bi*4,$b,"",0));
  27
  28         &mul("edx");
  29         &add($c0,"eax");
  30          &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # laod next a
  31          &mov("eax",&wparam(0)) if $pos > 0;                    # load r[]
  32          ###
  33         &adc($c1,"edx");
  34          &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 0;        # laod next b
  35          &mov("edx",&DWP(($nb)*4,$b,"",0)) if $pos == 1;        # laod next b
  36          ###
  37         &adc($c2,0);
  38          # is pos > 1, it means it is the last loop
  39          &mov(&DWP($i*4,"eax","",0),$c0) if $pos > 0;           # save r[];
  40         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;         # laod next a
  41         }
  42
  43 sub sqr_add_c
  44         {
  45         local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
  46
  47         # pos == -1 if eax and edx are pre-loaded, 0 to load from next
  48         # words, and 1 if load return value
  49
  50         &comment("sqr a[$ai]*a[$bi]");
  51
  52         # "eax" and "edx" will always be pre-loaded.
  53         # &mov("eax",&DWP($ai*4,$a,"",0)) ;
  54         # &mov("edx",&DWP($bi*4,$b,"",0));
  55
  56         if ($ai == $bi)
  57                 { &mul("eax");}
  58         else
  59                 { &mul("edx");}
  60         &add($c0,"eax");
  61          &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # load next a
  62          ###
  63         &adc($c1,"edx");
  64          &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos == 1) && ($na != $nb);
  65          ###
  66         &adc($c2,0);
  67          # is pos > 1, it means it is the last loop
  68          &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;              # save r[];
  69         &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;         # load next b
  70         }
  71
  72 sub sqr_add_c2
  73         {
  74         local($r,$a,$ai,$bi,$c0,$c1,$c2,$pos,$i,$na,$nb)=@_;
  75
  76         # pos == -1 if eax and edx are pre-loaded, 0 to load from next
  77         # words, and 1 if load return value
  78
  79         &comment("sqr a[$ai]*a[$bi]");
  80
  81         # "eax" and "edx" will always be pre-loaded.
  82         # &mov("eax",&DWP($ai*4,$a,"",0)) ;
  83         # &mov("edx",&DWP($bi*4,$a,"",0));
  84
  85         if ($ai == $bi)
  86                 { &mul("eax");}
  87         else
  88                 { &mul("edx");}
  89         &add("eax","eax");
  90          ###
  91         &adc("edx","edx");
  92          ###
  93         &adc($c2,0);
  94          &add($c0,"eax");
  95         &adc($c1,"edx");
  96          &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 0;        # load next a
  97          &mov("eax",&DWP(($na)*4,$a,"",0)) if $pos == 1;        # load next b
  98         &adc($c2,0);
  99         &mov(&DWP($i*4,$r,"",0),$c0) if $pos > 0;               # save r[];
 100          &mov("edx",&DWP(($nb)*4,$a,"",0)) if ($pos <= 1) && ($na != $nb);
 101          ###
 102         }
 103
 104 sub bn_mul_comba
 105         {
 106         local($name,$num)=@_;
 107         local($a,$b,$c0,$c1,$c2);
 108         local($i,$as,$ae,$bs,$be,$ai,$bi);
 109         local($tot,$end);
 110
 111         &function_begin_B($name,"");
 112
 113         $c0="ebx";
 114         $c1="ecx";
 115         $c2="ebp";
 116         $a="esi";
 117         $b="edi";
 118
 119         $as=0;
 120         $ae=0;
 121         $bs=0;
 122         $be=0;
 123         $tot=$num+$num-1;
 124
 125         &push("esi");
 126          &mov($a,&wparam(1));
 127         &push("edi");
 128          &mov($b,&wparam(2));
 129         &push("ebp");
 130          &push("ebx");
 131
 132         &xor($c0,$c0);
 133          &mov("eax",&DWP(0,$a,"",0));   # load the first word
 134         &xor($c1,$c1);
 135          &mov("edx",&DWP(0,$b,"",0));   # load the first second
 136
 137         for ($i=0; $i<$tot; $i++)
 138                 {
 139                 $ai=$as;
 140                 $bi=$bs;
 141                 $end=$be+1;
 142
 143                 &comment("################## Calculate word $i");
 144
 145                 for ($j=$bs; $j<$end; $j++)
 146                         {
 147                         &xor($c2,$c2) if ($j == $bs);
 148                         if (($j+1) == $end)
 149                                 {
 150                                 $v=1;
 151                                 $v=2 if (($i+1) == $tot);
 152                                 }
 153                         else
 154                                 { $v=0; }
 155                         if (($j+1) != $end)
 156                                 {
 157                                 $na=($ai-1);
 158                                 $nb=($bi+1);
 159                                 }
 160                         else
 161                                 {
 162                                 $na=$as+($i < ($num-1));
 163                                 $nb=$bs+($i >= ($num-1));
 164                                 }
 165 #printf STDERR "[$ai,$bi] -> [$na,$nb]\n";
 166                         &mul_add_c($a,$ai,$b,$bi,$c0,$c1,$c2,$v,$i,$na,$nb);
 167                         if ($v)
 168                                 {
 169                                 &comment("saved r[$i]");
 170                                 # &mov("eax",&wparam(0));
 171                                 # &mov(&DWP($i*4,"eax","",0),$c0);
 172                                 ($c0,$c1,$c2)=($c1,$c2,$c0);
 173                                 }
 174                         $ai--;
 175                         $bi++;
 176                         }
 177                 $as++ if ($i < ($num-1));
 178                 $ae++ if ($i >= ($num-1));
 179
 180                 $bs++ if ($i >= ($num-1));
 181                 $be++ if ($i < ($num-1));
 182                 }
 183         &comment("save r[$i]");
 184         # &mov("eax",&wparam(0));
 185         &mov(&DWP($i*4,"eax","",0),$c0);
 186
 187         &pop("ebx");
 188         &pop("ebp");
 189         &pop("edi");
 190         &pop("esi");
 191         &ret();
 192         &function_end_B($name);
 193         }
 194
 195 sub bn_sqr_comba
 196         {
 197         local($name,$num)=@_;
 198         local($r,$a,$c0,$c1,$c2)=@_;
 199         local($i,$as,$ae,$bs,$be,$ai,$bi);
 200         local($b,$tot,$end,$half);
 201
 202         &function_begin_B($name,"");
 203
 204         $c0="ebx";
 205         $c1="ecx";
 206         $c2="ebp";
 207         $a="esi";
 208         $r="edi";
 209
 210         &push("esi");
 211          &push("edi");
 212         &push("ebp");
 213          &push("ebx");
 214         &mov($r,&wparam(0));
 215          &mov($a,&wparam(1));
 216         &xor($c0,$c0);
 217          &xor($c1,$c1);
 218         &mov("eax",&DWP(0,$a,"",0)); # load the first word
 219
 220         $as=0;
 221         $ae=0;
 222         $bs=0;
 223         $be=0;
 224         $tot=$num+$num-1;
 225
 226         for ($i=0; $i<$tot; $i++)
 227                 {
 228                 $ai=$as;
 229                 $bi=$bs;
 230                 $end=$be+1;
 231
 232                 &comment("############### Calculate word $i");
 233                 for ($j=$bs; $j<$end; $j++)
 234                         {
 235                         &xor($c2,$c2) if ($j == $bs);
 236                         if (($ai-1) < ($bi+1))
 237                                 {
 238                                 $v=1;
 239                                 $v=2 if ($i+1) == $tot;
 240                                 }
 241                         else
 242                                 { $v=0; }
 243                         if (!$v)
 244                                 {
 245                                 $na=$ai-1;
 246                                 $nb=$bi+1;
 247                                 }
 248                         else
 249                                 {
 250                                 $na=$as+($i < ($num-1));
 251                                 $nb=$bs+($i >= ($num-1));
 252                                 }
 253                         if ($ai == $bi)
 254                                 {
 255                                 &sqr_add_c($r,$a,$ai,$bi,
 256                                         $c0,$c1,$c2,$v,$i,$na,$nb);
 257                                 }
 258                         else
 259                                 {
 260                                 &sqr_add_c2($r,$a,$ai,$bi,
 261                                         $c0,$c1,$c2,$v,$i,$na,$nb);
 262                                 }
 263                         if ($v)
 264                                 {
 265                                 &comment("saved r[$i]");
 266                                 #&mov(&DWP($i*4,$r,"",0),$c0);
 267                                 ($c0,$c1,$c2)=($c1,$c2,$c0);
 268                                 last;
 269                                 }
 270                         $ai--;
 271                         $bi++;
 272                         }
 273                 $as++ if ($i < ($num-1));
 274                 $ae++ if ($i >= ($num-1));
 275
 276                 $bs++ if ($i >= ($num-1));
 277                 $be++ if ($i < ($num-1));
 278                 }
 279         &mov(&DWP($i*4,$r,"",0),$c0);
 280         &pop("ebx");
 281         &pop("ebp");
 282         &pop("edi");
 283         &pop("esi");
 284         &ret();
 285         &function_end_B($name);
 286         }