crypto/des/asm/des-586.pl

   1 #!/usr/local/bin/perl
   2 #
   3 # The inner loop instruction sequence and the IP/FP modifications are from
   4 # Svend Olaf Mikkelsen <svolaf@inet.uni-c.dk>
   5 #
   6
   7 push(@INC,"perlasm","../../perlasm");
   8 require "x86asm.pl";
   9 require "cbc.pl";
  10 require "desboth.pl";
  11
  12 # base code is in microsft
  13 # op dest, source
  14 # format.
  15 #
  16
  17 &asm_init($ARGV[0],"des-586.pl");
  18
  19 $L="edi";
  20 $R="esi";
  21 $trans="ebp";
  22 $small_footprint=1 if (grep(/\-DOPENSSL_SMALL_FOOTPRINT/,@ARGV));
  23 # one can discuss setting this variable to 1 unconditionally, as
  24 # the folded loop is only 3% slower than unrolled, but >7 times smaller
  25
  26 &DES_encrypt_internal();
  27 &DES_decrypt_internal();
  28 &DES_encrypt("DES_encrypt1",1);
  29 &DES_encrypt("DES_encrypt2",0);
  30 &DES_encrypt3("DES_encrypt3",1);
  31 &DES_encrypt3("DES_decrypt3",0);
  32 &cbc("DES_ncbc_encrypt","DES_encrypt1","DES_encrypt1",0,4,5,3,5,-1);
  33 &cbc("DES_ede3_cbc_encrypt","DES_encrypt3","DES_decrypt3",0,6,7,3,4,5);
  34 &DES_SPtrans();
  35
  36 &asm_finish();
  37
  38 sub DES_encrypt_internal()
  39         {
  40         &function_begin_B("_x86_DES_encrypt");
  41
  42         if ($small_footprint)
  43             {
  44             &lea("edx",&DWP(128,"ecx"));
  45             &push("edx");
  46             &push("ecx");
  47             &set_label("eloop");
  48                 &D_ENCRYPT(0,$L,$R,0,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  49                 &comment("");
  50                 &D_ENCRYPT(1,$R,$L,2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  51                 &comment("");
  52                 &add("ecx",16);
  53                 &cmp("ecx",&swtmp(1));
  54                 &mov(&swtmp(0),"ecx");
  55                 &jb(&label("eloop"));
  56             &add("esp",8);
  57             }
  58         else
  59             {
  60             &push("ecx");
  61             for ($i=0; $i<16; $i+=2)
  62                 {
  63                 &comment("Round $i");
  64                 &D_ENCRYPT($i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  65                 &comment("Round ".sprintf("%d",$i+1));
  66                 &D_ENCRYPT($i+1,$R,$L,($i+1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  67                 }
  68             &add("esp",4);
  69         }
  70         &ret();
  71
  72         &function_end_B("_x86_DES_encrypt");
  73         }
  74
  75 sub DES_decrypt_internal()
  76         {
  77         &function_begin_B("_x86_DES_decrypt");
  78
  79         if ($small_footprint)
  80             {
  81             &push("ecx");
  82             &lea("ecx",&DWP(128,"ecx"));
  83             &push("ecx");
  84             &set_label("dloop");
  85                 &D_ENCRYPT(0,$L,$R,-2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  86                 &comment("");
  87                 &D_ENCRYPT(1,$R,$L,-4,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  88                 &comment("");
  89                 &sub("ecx",16);
  90                 &cmp("ecx",&swtmp(1));
  91                 &mov(&swtmp(0),"ecx");
  92                 &ja(&label("dloop"));
  93             &add("esp",8);
  94             }
  95         else
  96             {
  97             &push("ecx");
  98             for ($i=15; $i>0; $i-=2)
  99                 {
 100                 &comment("Round $i");
 101                 &D_ENCRYPT(15-$i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 102                 &comment("Round ".sprintf("%d",$i-1));
 103                 &D_ENCRYPT(15-$i+1,$R,$L,($i-1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 104                 }
 105             &add("esp",4);
 106             }
 107         &ret();
 108
 109         &function_end_B("_x86_DES_decrypt");
 110         }
 111
 112 sub DES_encrypt
 113         {
 114         local($name,$do_ip)=@_;
 115
 116         &function_begin_B($name);
 117
 118         &push("esi");
 119         &push("edi");
 120
 121         &comment("");
 122         &comment("Load the 2 words");
 123
 124         if ($do_ip)
 125                 {
 126                 &mov($R,&wparam(0));
 127                  &xor(  "ecx",          "ecx"           );
 128
 129                 &push("ebx");
 130                 &push("ebp");
 131
 132                 &mov("eax",&DWP(0,$R,"",0));
 133                  &mov("ebx",&wparam(2));        # get encrypt flag
 134                 &mov($L,&DWP(4,$R,"",0));
 135                 &comment("");
 136                 &comment("IP");
 137                 &IP_new("eax",$L,$R,3);
 138                 }
 139         else
 140                 {
 141                 &mov("eax",&wparam(0));
 142                  &xor(  "ecx",          "ecx"           );
 143
 144                 &push("ebx");
 145                 &push("ebp");
 146
 147                 &mov($R,&DWP(0,"eax","",0));
 148                  &mov("ebx",&wparam(2));        # get encrypt flag
 149                 &rotl($R,3);
 150                 &mov($L,&DWP(4,"eax","",0));
 151                 &rotl($L,3);
 152                 }
 153
 154         # PIC-ification:-)
 155         &picmeup($trans,"DES_SPtrans");
 156
 157         &mov(   "ecx",  &wparam(1)      );
 158
 159         &cmp("ebx","0");
 160         &je(&label("decrypt"));
 161         &call("_x86_DES_encrypt");
 162         &jmp(&label("done"));
 163         &set_label("decrypt");
 164         &call("_x86_DES_decrypt");
 165         &set_label("done");
 166
 167         if ($do_ip)
 168                 {
 169                 &comment("");
 170                 &comment("FP");
 171                 &mov("edx",&wparam(0));
 172                 &FP_new($L,$R,"eax",3);
 173
 174                 &mov(&DWP(0,"edx","",0),"eax");
 175                 &mov(&DWP(4,"edx","",0),$R);
 176                 }
 177         else
 178                 {
 179                 &comment("");
 180                 &comment("Fixup");
 181                 &rotr($L,3);            # r
 182                  &mov("eax",&wparam(0));
 183                 &rotr($R,3);            # l
 184                  &mov(&DWP(0,"eax","",0),$L);
 185                  &mov(&DWP(4,"eax","",0),$R);
 186                 }
 187
 188         &pop("ebp");
 189         &pop("ebx");
 190         &pop("edi");
 191         &pop("esi");
 192         &ret();
 193
 194         &function_end_B($name);
 195         }
 196
 197 sub D_ENCRYPT
 198         {
 199         local($r,$L,$R,$S,$trans,$u,$tmp1,$tmp2,$t,$wp1)=@_;
 200
 201          &mov(  $u,             &DWP(&n2a($S*4),$tmp2,"",0));
 202         &xor(   $tmp1,          $tmp1);
 203          &mov(  $t,             &DWP(&n2a(($S+1)*4),$tmp2,"",0));
 204         &xor(   $u,             $R);
 205         &xor(   $tmp2,          $tmp2);
 206          &xor(  $t,             $R);
 207         &and(   $u,             "0xfcfcfcfc"    );
 208          &and(  $t,             "0xcfcfcfcf"    );
 209         &movb(  &LB($tmp1),     &LB($u) );
 210          &movb( &LB($tmp2),     &HB($u) );
 211         &rotr(  $t,             4               );
 212         &xor(   $L,             &DWP("     ",$trans,$tmp1,0));
 213          &movb( &LB($tmp1),     &LB($t) );
 214          &xor(  $L,             &DWP("0x200",$trans,$tmp2,0));
 215          &movb( &LB($tmp2),     &HB($t) );
 216         &shr(   $u,             16);
 217          &xor(  $L,             &DWP("0x100",$trans,$tmp1,0));
 218          &movb( &LB($tmp1),     &HB($u) );
 219         &shr(   $t,             16);
 220          &xor(  $L,             &DWP("0x300",$trans,$tmp2,0));
 221         &movb(  &LB($tmp2),     &HB($t) );
 222          &and(  $u,             "0xff"  );
 223         &and(   $t,             "0xff"  );
 224          &xor(  $L,             &DWP("0x600",$trans,$tmp1,0));
 225          &xor(  $L,             &DWP("0x700",$trans,$tmp2,0));
 226         &mov(   $tmp2,          $wp1    );
 227          &xor(  $L,             &DWP("0x400",$trans,$u,0));
 228          &xor(  $L,             &DWP("0x500",$trans,$t,0));
 229         }
 230
 231 sub n2a
 232         {
 233         sprintf("%d",$_[0]);
 234         }
 235
 236 # now has a side affect of rotating $a by $shift
 237 sub R_PERM_OP
 238         {
 239         local($a,$b,$tt,$shift,$mask,$last)=@_;
 240
 241         &rotl(  $a,             $shift          ) if ($shift != 0);
 242         &mov(   $tt,            $a              );
 243         &xor(   $a,             $b              );
 244         &and(   $a,             $mask           );
 245         # This can never succeed, and besides it is difficult to see what the
 246         # idea was - Ben 13 Feb 99
 247         if (!$last eq $b)
 248                 {
 249                 &xor(   $b,             $a              );
 250                 &xor(   $tt,            $a              );
 251                 }
 252         else
 253                 {
 254                 &xor(   $tt,            $a              );
 255                 &xor(   $b,             $a              );
 256                 }
 257         &comment("");
 258         }
 259
 260 sub IP_new
 261         {
 262         local($l,$r,$tt,$lr)=@_;
 263
 264         &R_PERM_OP($l,$r,$tt, 4,"0xf0f0f0f0",$l);
 265         &R_PERM_OP($r,$tt,$l,20,"0xfff0000f",$l);
 266         &R_PERM_OP($l,$tt,$r,14,"0x33333333",$r);
 267         &R_PERM_OP($tt,$r,$l,22,"0x03fc03fc",$r);
 268         &R_PERM_OP($l,$r,$tt, 9,"0xaaaaaaaa",$r);
 269
 270         if ($lr != 3)
 271                 {
 272                 if (($lr-3) < 0)
 273                         { &rotr($tt,    3-$lr); }
 274                 else    { &rotl($tt,    $lr-3); }
 275                 }
 276         if ($lr != 2)
 277                 {
 278                 if (($lr-2) < 0)
 279                         { &rotr($r,     2-$lr); }
 280                 else    { &rotl($r,     $lr-2); }
 281                 }
 282         }
 283
 284 sub FP_new
 285         {
 286         local($l,$r,$tt,$lr)=@_;
 287
 288         if ($lr != 2)
 289                 {
 290                 if (($lr-2) < 0)
 291                         { &rotl($r,     2-$lr); }
 292                 else    { &rotr($r,     $lr-2); }
 293                 }
 294         if ($lr != 3)
 295                 {
 296                 if (($lr-3) < 0)
 297                         { &rotl($l,     3-$lr); }
 298                 else    { &rotr($l,     $lr-3); }
 299                 }
 300
 301         &R_PERM_OP($l,$r,$tt, 0,"0xaaaaaaaa",$r);
 302         &R_PERM_OP($tt,$r,$l,23,"0x03fc03fc",$r);
 303         &R_PERM_OP($l,$r,$tt,10,"0x33333333",$l);
 304         &R_PERM_OP($r,$tt,$l,18,"0xfff0000f",$l);
 305         &R_PERM_OP($l,$tt,$r,12,"0xf0f0f0f0",$r);
 306         &rotr($tt       , 4);
 307         }
 308
 309 sub DES_SPtrans
 310         {
 311         &public_label("DES_SPtrans");
 312         &set_label("DES_SPtrans",64);
 313         &data_word(0x02080800, 0x00080000, 0x02000002, 0x02080802);
 314         &data_word(0x02000000, 0x00080802, 0x00080002, 0x02000002);
 315         &data_word(0x00080802, 0x02080800, 0x02080000, 0x00000802);
 316         &data_word(0x02000802, 0x02000000, 0x00000000, 0x00080002);
 317         &data_word(0x00080000, 0x00000002, 0x02000800, 0x00080800);
 318         &data_word(0x02080802, 0x02080000, 0x00000802, 0x02000800);
 319         &data_word(0x00000002, 0x00000800, 0x00080800, 0x02080002);
 320         &data_word(0x00000800, 0x02000802, 0x02080002, 0x00000000);
 321         &data_word(0x00000000, 0x02080802, 0x02000800, 0x00080002);
 322         &data_word(0x02080800, 0x00080000, 0x00000802, 0x02000800);
 323         &data_word(0x02080002, 0x00000800, 0x00080800, 0x02000002);
 324         &data_word(0x00080802, 0x00000002, 0x02000002, 0x02080000);
 325         &data_word(0x02080802, 0x00080800, 0x02080000, 0x02000802);
 326         &data_word(0x02000000, 0x00000802, 0x00080002, 0x00000000);
 327         &data_word(0x00080000, 0x02000000, 0x02000802, 0x02080800);
 328         &data_word(0x00000002, 0x02080002, 0x00000800, 0x00080802);
 329         # nibble 1
 330         &data_word(0x40108010, 0x00000000, 0x00108000, 0x40100000);
 331         &data_word(0x40000010, 0x00008010, 0x40008000, 0x00108000);
 332         &data_word(0x00008000, 0x40100010, 0x00000010, 0x40008000);
 333         &data_word(0x00100010, 0x40108000, 0x40100000, 0x00000010);
 334         &data_word(0x00100000, 0x40008010, 0x40100010, 0x00008000);
 335         &data_word(0x00108010, 0x40000000, 0x00000000, 0x00100010);
 336         &data_word(0x40008010, 0x00108010, 0x40108000, 0x40000010);
 337         &data_word(0x40000000, 0x00100000, 0x00008010, 0x40108010);
 338         &data_word(0x00100010, 0x40108000, 0x40008000, 0x00108010);
 339         &data_word(0x40108010, 0x00100010, 0x40000010, 0x00000000);
 340         &data_word(0x40000000, 0x00008010, 0x00100000, 0x40100010);
 341         &data_word(0x00008000, 0x40000000, 0x00108010, 0x40008010);
 342         &data_word(0x40108000, 0x00008000, 0x00000000, 0x40000010);
 343         &data_word(0x00000010, 0x40108010, 0x00108000, 0x40100000);
 344         &data_word(0x40100010, 0x00100000, 0x00008010, 0x40008000);
 345         &data_word(0x40008010, 0x00000010, 0x40100000, 0x00108000);
 346         # nibble 2
 347         &data_word(0x04000001, 0x04040100, 0x00000100, 0x04000101);
 348         &data_word(0x00040001, 0x04000000, 0x04000101, 0x00040100);
 349         &data_word(0x04000100, 0x00040000, 0x04040000, 0x00000001);
 350         &data_word(0x04040101, 0x00000101, 0x00000001, 0x04040001);
 351         &data_word(0x00000000, 0x00040001, 0x04040100, 0x00000100);
 352         &data_word(0x00000101, 0x04040101, 0x00040000, 0x04000001);
 353         &data_word(0x04040001, 0x04000100, 0x00040101, 0x04040000);
 354         &data_word(0x00040100, 0x00000000, 0x04000000, 0x00040101);
 355         &data_word(0x04040100, 0x00000100, 0x00000001, 0x00040000);
 356         &data_word(0x00000101, 0x00040001, 0x04040000, 0x04000101);
 357         &data_word(0x00000000, 0x04040100, 0x00040100, 0x04040001);
 358         &data_word(0x00040001, 0x04000000, 0x04040101, 0x00000001);
 359         &data_word(0x00040101, 0x04000001, 0x04000000, 0x04040101);
 360         &data_word(0x00040000, 0x04000100, 0x04000101, 0x00040100);
 361         &data_word(0x04000100, 0x00000000, 0x04040001, 0x00000101);
 362         &data_word(0x04000001, 0x00040101, 0x00000100, 0x04040000);
 363         # nibble 3
 364         &data_word(0x00401008, 0x10001000, 0x00000008, 0x10401008);
 365         &data_word(0x00000000, 0x10400000, 0x10001008, 0x00400008);
 366         &data_word(0x10401000, 0x10000008, 0x10000000, 0x00001008);
 367         &data_word(0x10000008, 0x00401008, 0x00400000, 0x10000000);
 368         &data_word(0x10400008, 0x00401000, 0x00001000, 0x00000008);
 369         &data_word(0x00401000, 0x10001008, 0x10400000, 0x00001000);
 370         &data_word(0x00001008, 0x00000000, 0x00400008, 0x10401000);
 371         &data_word(0x10001000, 0x10400008, 0x10401008, 0x00400000);
 372         &data_word(0x10400008, 0x00001008, 0x00400000, 0x10000008);
 373         &data_word(0x00401000, 0x10001000, 0x00000008, 0x10400000);
 374         &data_word(0x10001008, 0x00000000, 0x00001000, 0x00400008);
 375         &data_word(0x00000000, 0x10400008, 0x10401000, 0x00001000);
 376         &data_word(0x10000000, 0x10401008, 0x00401008, 0x00400000);
 377         &data_word(0x10401008, 0x00000008, 0x10001000, 0x00401008);
 378         &data_word(0x00400008, 0x00401000, 0x10400000, 0x10001008);
 379         &data_word(0x00001008, 0x10000000, 0x10000008, 0x10401000);
 380         # nibble 4
 381         &data_word(0x08000000, 0x00010000, 0x00000400, 0x08010420);
 382         &data_word(0x08010020, 0x08000400, 0x00010420, 0x08010000);
 383         &data_word(0x00010000, 0x00000020, 0x08000020, 0x00010400);
 384         &data_word(0x08000420, 0x08010020, 0x08010400, 0x00000000);
 385         &data_word(0x00010400, 0x08000000, 0x00010020, 0x00000420);
 386         &data_word(0x08000400, 0x00010420, 0x00000000, 0x08000020);
 387         &data_word(0x00000020, 0x08000420, 0x08010420, 0x00010020);
 388         &data_word(0x08010000, 0x00000400, 0x00000420, 0x08010400);
 389         &data_word(0x08010400, 0x08000420, 0x00010020, 0x08010000);
 390         &data_word(0x00010000, 0x00000020, 0x08000020, 0x08000400);
 391         &data_word(0x08000000, 0x00010400, 0x08010420, 0x00000000);
 392         &data_word(0x00010420, 0x08000000, 0x00000400, 0x00010020);
 393         &data_word(0x08000420, 0x00000400, 0x00000000, 0x08010420);
 394         &data_word(0x08010020, 0x08010400, 0x00000420, 0x00010000);
 395         &data_word(0x00010400, 0x08010020, 0x08000400, 0x00000420);
 396         &data_word(0x00000020, 0x00010420, 0x08010000, 0x08000020);
 397         # nibble 5
 398         &data_word(0x80000040, 0x00200040, 0x00000000, 0x80202000);
 399         &data_word(0x00200040, 0x00002000, 0x80002040, 0x00200000);
 400         &data_word(0x00002040, 0x80202040, 0x00202000, 0x80000000);
 401         &data_word(0x80002000, 0x80000040, 0x80200000, 0x00202040);
 402         &data_word(0x00200000, 0x80002040, 0x80200040, 0x00000000);
 403         &data_word(0x00002000, 0x00000040, 0x80202000, 0x80200040);
 404         &data_word(0x80202040, 0x80200000, 0x80000000, 0x00002040);
 405         &data_word(0x00000040, 0x00202000, 0x00202040, 0x80002000);
 406         &data_word(0x00002040, 0x80000000, 0x80002000, 0x00202040);
 407         &data_word(0x80202000, 0x00200040, 0x00000000, 0x80002000);
 408         &data_word(0x80000000, 0x00002000, 0x80200040, 0x00200000);
 409         &data_word(0x00200040, 0x80202040, 0x00202000, 0x00000040);
 410         &data_word(0x80202040, 0x00202000, 0x00200000, 0x80002040);
 411         &data_word(0x80000040, 0x80200000, 0x00202040, 0x00000000);
 412         &data_word(0x00002000, 0x80000040, 0x80002040, 0x80202000);
 413         &data_word(0x80200000, 0x00002040, 0x00000040, 0x80200040);
 414         # nibble 6
 415         &data_word(0x00004000, 0x00000200, 0x01000200, 0x01000004);
 416         &data_word(0x01004204, 0x00004004, 0x00004200, 0x00000000);
 417         &data_word(0x01000000, 0x01000204, 0x00000204, 0x01004000);
 418         &data_word(0x00000004, 0x01004200, 0x01004000, 0x00000204);
 419         &data_word(0x01000204, 0x00004000, 0x00004004, 0x01004204);
 420         &data_word(0x00000000, 0x01000200, 0x01000004, 0x00004200);
 421         &data_word(0x01004004, 0x00004204, 0x01004200, 0x00000004);
 422         &data_word(0x00004204, 0x01004004, 0x00000200, 0x01000000);
 423         &data_word(0x00004204, 0x01004000, 0x01004004, 0x00000204);
 424         &data_word(0x00004000, 0x00000200, 0x01000000, 0x01004004);
 425         &data_word(0x01000204, 0x00004204, 0x00004200, 0x00000000);
 426         &data_word(0x00000200, 0x01000004, 0x00000004, 0x01000200);
 427         &data_word(0x00000000, 0x01000204, 0x01000200, 0x00004200);
 428         &data_word(0x00000204, 0x00004000, 0x01004204, 0x01000000);
 429         &data_word(0x01004200, 0x00000004, 0x00004004, 0x01004204);
 430         &data_word(0x01000004, 0x01004200, 0x01004000, 0x00004004);
 431         # nibble 7
 432         &data_word(0x20800080, 0x20820000, 0x00020080, 0x00000000);
 433         &data_word(0x20020000, 0x00800080, 0x20800000, 0x20820080);
 434         &data_word(0x00000080, 0x20000000, 0x00820000, 0x00020080);
 435         &data_word(0x00820080, 0x20020080, 0x20000080, 0x20800000);
 436         &data_word(0x00020000, 0x00820080, 0x00800080, 0x20020000);
 437         &data_word(0x20820080, 0x20000080, 0x00000000, 0x00820000);
 438         &data_word(0x20000000, 0x00800000, 0x20020080, 0x20800080);
 439         &data_word(0x00800000, 0x00020000, 0x20820000, 0x00000080);
 440         &data_word(0x00800000, 0x00020000, 0x20000080, 0x20820080);
 441         &data_word(0x00020080, 0x20000000, 0x00000000, 0x00820000);
 442         &data_word(0x20800080, 0x20020080, 0x20020000, 0x00800080);
 443         &data_word(0x20820000, 0x00000080, 0x00800080, 0x20020000);
 444         &data_word(0x20820080, 0x00800000, 0x20800000, 0x20000080);
 445         &data_word(0x00820000, 0x00020080, 0x20020080, 0x20800000);
 446         &data_word(0x00000080, 0x20820000, 0x00820080, 0x00000000);
 447         &data_word(0x20000000, 0x20800080, 0x00020000, 0x00820080);
 448         }