crypto/des/asm/des-586.pl

   1 #! /usr/bin/env perl
   2 # Copyright 1995-2016 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 # The inner loop instruction sequence and the IP/FP modifications are from
  10 # Svend Olaf Mikkelsen.
  11
  12 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  13 push(@INC,"${dir}","${dir}../../perlasm");
  14 require "x86asm.pl";
  15 require "cbc.pl";
  16 require "desboth.pl";
  17
  18 # base code is in Microsoft
  19 # op dest, source
  20 # format.
  21 #
  22
  23 $output=pop;
  24 open STDOUT,">$output";
  25
  26 &asm_init($ARGV[0]);
  27
  28 $L="edi";
  29 $R="esi";
  30 $trans="ebp";
  31 $small_footprint=1 if (grep(/\-DOPENSSL_SMALL_FOOTPRINT/,@ARGV));
  32 # one can discuss setting this variable to 1 unconditionally, as
  33 # the folded loop is only 3% slower than unrolled, but >7 times smaller
  34
  35 &public_label("DES_SPtrans");
  36 &static_label("des_sptrans");
  37
  38 &DES_encrypt_internal();
  39 &DES_decrypt_internal();
  40 &DES_encrypt("DES_encrypt1",1);
  41 &DES_encrypt("DES_encrypt2",0);
  42 &DES_encrypt3("DES_encrypt3",1);
  43 &DES_encrypt3("DES_decrypt3",0);
  44 &cbc("DES_ncbc_encrypt","DES_encrypt1","DES_encrypt1",0,4,5,3,5,-1);
  45 &cbc("DES_ede3_cbc_encrypt","DES_encrypt3","DES_decrypt3",0,6,7,3,4,5);
  46 &DES_SPtrans();
  47
  48 &asm_finish();
  49
  50 close STDOUT;
  51
  52 sub DES_encrypt_internal()
  53         {
  54         &function_begin_B("_x86_DES_encrypt");
  55
  56         if ($small_footprint)
  57             {
  58             &lea("edx",&DWP(128,"ecx"));
  59             &push("edx");
  60             &push("ecx");
  61             &set_label("eloop");
  62                 &D_ENCRYPT(0,$L,$R,0,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  63                 &comment("");
  64                 &D_ENCRYPT(1,$R,$L,2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  65                 &comment("");
  66                 &add("ecx",16);
  67                 &cmp("ecx",&swtmp(1));
  68                 &mov(&swtmp(0),"ecx");
  69                 &jb(&label("eloop"));
  70             &add("esp",8);
  71             }
  72         else
  73             {
  74             &push("ecx");
  75             for ($i=0; $i<16; $i+=2)
  76                 {
  77                 &comment("Round $i");
  78                 &D_ENCRYPT($i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  79                 &comment("Round ".sprintf("%d",$i+1));
  80                 &D_ENCRYPT($i+1,$R,$L,($i+1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
  81                 }
  82             &add("esp",4);
  83         }
  84         &ret();
  85
  86         &function_end_B("_x86_DES_encrypt");
  87         }
  88
  89 sub DES_decrypt_internal()
  90         {
  91         &function_begin_B("_x86_DES_decrypt");
  92
  93         if ($small_footprint)
  94             {
  95             &push("ecx");
  96             &lea("ecx",&DWP(128,"ecx"));
  97             &push("ecx");
  98             &set_label("dloop");
  99                 &D_ENCRYPT(0,$L,$R,-2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 100                 &comment("");
 101                 &D_ENCRYPT(1,$R,$L,-4,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 102                 &comment("");
 103                 &sub("ecx",16);
 104                 &cmp("ecx",&swtmp(1));
 105                 &mov(&swtmp(0),"ecx");
 106                 &ja(&label("dloop"));
 107             &add("esp",8);
 108             }
 109         else
 110             {
 111             &push("ecx");
 112             for ($i=15; $i>0; $i-=2)
 113                 {
 114                 &comment("Round $i");
 115                 &D_ENCRYPT(15-$i,$L,$R,$i*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 116                 &comment("Round ".sprintf("%d",$i-1));
 117                 &D_ENCRYPT(15-$i+1,$R,$L,($i-1)*2,$trans,"eax","ebx","ecx","edx",&swtmp(0));
 118                 }
 119             &add("esp",4);
 120             }
 121         &ret();
 122
 123         &function_end_B("_x86_DES_decrypt");
 124         }
 125
 126 sub DES_encrypt
 127         {
 128         local($name,$do_ip)=@_;
 129
 130         &function_begin_B($name);
 131
 132         &push("esi");
 133         &push("edi");
 134
 135         &comment("");
 136         &comment("Load the 2 words");
 137
 138         if ($do_ip)
 139                 {
 140                 &mov($R,&wparam(0));
 141                  &xor(  "ecx",          "ecx"           );
 142
 143                 &push("ebx");
 144                 &push("ebp");
 145
 146                 &mov("eax",&DWP(0,$R,"",0));
 147                  &mov("ebx",&wparam(2));        # get encrypt flag
 148                 &mov($L,&DWP(4,$R,"",0));
 149                 &comment("");
 150                 &comment("IP");
 151                 &IP_new("eax",$L,$R,3);
 152                 }
 153         else
 154                 {
 155                 &mov("eax",&wparam(0));
 156                  &xor(  "ecx",          "ecx"           );
 157
 158                 &push("ebx");
 159                 &push("ebp");
 160
 161                 &mov($R,&DWP(0,"eax","",0));
 162                  &mov("ebx",&wparam(2));        # get encrypt flag
 163                 &rotl($R,3);
 164                 &mov($L,&DWP(4,"eax","",0));
 165                 &rotl($L,3);
 166                 }
 167
 168         # PIC-ification:-)
 169         &call   (&label("pic_point"));
 170         &set_label("pic_point");
 171         &blindpop($trans);
 172         &lea    ($trans,&DWP(&label("des_sptrans")."-".&label("pic_point"),$trans));
 173
 174         &mov(   "ecx",  &wparam(1)      );
 175
 176         &cmp("ebx","0");
 177         &je(&label("decrypt"));
 178         &call("_x86_DES_encrypt");
 179         &jmp(&label("done"));
 180         &set_label("decrypt");
 181         &call("_x86_DES_decrypt");
 182         &set_label("done");
 183
 184         if ($do_ip)
 185                 {
 186                 &comment("");
 187                 &comment("FP");
 188                 &mov("edx",&wparam(0));
 189                 &FP_new($L,$R,"eax",3);
 190
 191                 &mov(&DWP(0,"edx","",0),"eax");
 192                 &mov(&DWP(4,"edx","",0),$R);
 193                 }
 194         else
 195                 {
 196                 &comment("");
 197                 &comment("Fixup");
 198                 &rotr($L,3);            # r
 199                  &mov("eax",&wparam(0));
 200                 &rotr($R,3);            # l
 201                  &mov(&DWP(0,"eax","",0),$L);
 202                  &mov(&DWP(4,"eax","",0),$R);
 203                 }
 204
 205         &pop("ebp");
 206         &pop("ebx");
 207         &pop("edi");
 208         &pop("esi");
 209         &ret();
 210
 211         &function_end_B($name);
 212         }
 213
 214 sub D_ENCRYPT
 215         {
 216         local($r,$L,$R,$S,$trans,$u,$tmp1,$tmp2,$t,$wp1)=@_;
 217
 218          &mov(  $u,             &DWP(&n2a($S*4),$tmp2,"",0));
 219         &xor(   $tmp1,          $tmp1);
 220          &mov(  $t,             &DWP(&n2a(($S+1)*4),$tmp2,"",0));
 221         &xor(   $u,             $R);
 222         &xor(   $tmp2,          $tmp2);
 223          &xor(  $t,             $R);
 224         &and(   $u,             "0xfcfcfcfc"    );
 225          &and(  $t,             "0xcfcfcfcf"    );
 226         &movb(  &LB($tmp1),     &LB($u) );
 227          &movb( &LB($tmp2),     &HB($u) );
 228         &rotr(  $t,             4               );
 229         &xor(   $L,             &DWP("     ",$trans,$tmp1,0));
 230          &movb( &LB($tmp1),     &LB($t) );
 231          &xor(  $L,             &DWP("0x200",$trans,$tmp2,0));
 232          &movb( &LB($tmp2),     &HB($t) );
 233         &shr(   $u,             16);
 234          &xor(  $L,             &DWP("0x100",$trans,$tmp1,0));
 235          &movb( &LB($tmp1),     &HB($u) );
 236         &shr(   $t,             16);
 237          &xor(  $L,             &DWP("0x300",$trans,$tmp2,0));
 238         &movb(  &LB($tmp2),     &HB($t) );
 239          &and(  $u,             "0xff"  );
 240         &and(   $t,             "0xff"  );
 241          &xor(  $L,             &DWP("0x600",$trans,$tmp1,0));
 242          &xor(  $L,             &DWP("0x700",$trans,$tmp2,0));
 243         &mov(   $tmp2,          $wp1    );
 244          &xor(  $L,             &DWP("0x400",$trans,$u,0));
 245          &xor(  $L,             &DWP("0x500",$trans,$t,0));
 246         }
 247
 248 sub n2a
 249         {
 250         sprintf("%d",$_[0]);
 251         }
 252
 253 # now has a side affect of rotating $a by $shift
 254 sub R_PERM_OP
 255         {
 256         local($a,$b,$tt,$shift,$mask,$last)=@_;
 257
 258         &rotl(  $a,             $shift          ) if ($shift != 0);
 259         &mov(   $tt,            $a              );
 260         &xor(   $a,             $b              );
 261         &and(   $a,             $mask           );
 262         # This can never succeed, and besides it is difficult to see what the
 263         # idea was - Ben 13 Feb 99
 264         if (!$last eq $b)
 265                 {
 266                 &xor(   $b,             $a              );
 267                 &xor(   $tt,            $a              );
 268                 }
 269         else
 270                 {
 271                 &xor(   $tt,            $a              );
 272                 &xor(   $b,             $a              );
 273                 }
 274         &comment("");
 275         }
 276
 277 sub IP_new
 278         {
 279         local($l,$r,$tt,$lr)=@_;
 280
 281         &R_PERM_OP($l,$r,$tt, 4,"0xf0f0f0f0",$l);
 282         &R_PERM_OP($r,$tt,$l,20,"0xfff0000f",$l);
 283         &R_PERM_OP($l,$tt,$r,14,"0x33333333",$r);
 284         &R_PERM_OP($tt,$r,$l,22,"0x03fc03fc",$r);
 285         &R_PERM_OP($l,$r,$tt, 9,"0xaaaaaaaa",$r);
 286
 287         if ($lr != 3)
 288                 {
 289                 if (($lr-3) < 0)
 290                         { &rotr($tt,    3-$lr); }
 291                 else    { &rotl($tt,    $lr-3); }
 292                 }
 293         if ($lr != 2)
 294                 {
 295                 if (($lr-2) < 0)
 296                         { &rotr($r,     2-$lr); }
 297                 else    { &rotl($r,     $lr-2); }
 298                 }
 299         }
 300
 301 sub FP_new
 302         {
 303         local($l,$r,$tt,$lr)=@_;
 304
 305         if ($lr != 2)
 306                 {
 307                 if (($lr-2) < 0)
 308                         { &rotl($r,     2-$lr); }
 309                 else    { &rotr($r,     $lr-2); }
 310                 }
 311         if ($lr != 3)
 312                 {
 313                 if (($lr-3) < 0)
 314                         { &rotl($l,     3-$lr); }
 315                 else    { &rotr($l,     $lr-3); }
 316                 }
 317
 318         &R_PERM_OP($l,$r,$tt, 0,"0xaaaaaaaa",$r);
 319         &R_PERM_OP($tt,$r,$l,23,"0x03fc03fc",$r);
 320         &R_PERM_OP($l,$r,$tt,10,"0x33333333",$l);
 321         &R_PERM_OP($r,$tt,$l,18,"0xfff0000f",$l);
 322         &R_PERM_OP($l,$tt,$r,12,"0xf0f0f0f0",$r);
 323         &rotr($tt       , 4);
 324         }
 325
 326 sub DES_SPtrans
 327         {
 328         &set_label("DES_SPtrans",64);
 329         &set_label("des_sptrans");
 330         &data_word(0x02080800, 0x00080000, 0x02000002, 0x02080802);
 331         &data_word(0x02000000, 0x00080802, 0x00080002, 0x02000002);
 332         &data_word(0x00080802, 0x02080800, 0x02080000, 0x00000802);
 333         &data_word(0x02000802, 0x02000000, 0x00000000, 0x00080002);
 334         &data_word(0x00080000, 0x00000002, 0x02000800, 0x00080800);
 335         &data_word(0x02080802, 0x02080000, 0x00000802, 0x02000800);
 336         &data_word(0x00000002, 0x00000800, 0x00080800, 0x02080002);
 337         &data_word(0x00000800, 0x02000802, 0x02080002, 0x00000000);
 338         &data_word(0x00000000, 0x02080802, 0x02000800, 0x00080002);
 339         &data_word(0x02080800, 0x00080000, 0x00000802, 0x02000800);
 340         &data_word(0x02080002, 0x00000800, 0x00080800, 0x02000002);
 341         &data_word(0x00080802, 0x00000002, 0x02000002, 0x02080000);
 342         &data_word(0x02080802, 0x00080800, 0x02080000, 0x02000802);
 343         &data_word(0x02000000, 0x00000802, 0x00080002, 0x00000000);
 344         &data_word(0x00080000, 0x02000000, 0x02000802, 0x02080800);
 345         &data_word(0x00000002, 0x02080002, 0x00000800, 0x00080802);
 346         # nibble 1
 347         &data_word(0x40108010, 0x00000000, 0x00108000, 0x40100000);
 348         &data_word(0x40000010, 0x00008010, 0x40008000, 0x00108000);
 349         &data_word(0x00008000, 0x40100010, 0x00000010, 0x40008000);
 350         &data_word(0x00100010, 0x40108000, 0x40100000, 0x00000010);
 351         &data_word(0x00100000, 0x40008010, 0x40100010, 0x00008000);
 352         &data_word(0x00108010, 0x40000000, 0x00000000, 0x00100010);
 353         &data_word(0x40008010, 0x00108010, 0x40108000, 0x40000010);
 354         &data_word(0x40000000, 0x00100000, 0x00008010, 0x40108010);
 355         &data_word(0x00100010, 0x40108000, 0x40008000, 0x00108010);
 356         &data_word(0x40108010, 0x00100010, 0x40000010, 0x00000000);
 357         &data_word(0x40000000, 0x00008010, 0x00100000, 0x40100010);
 358         &data_word(0x00008000, 0x40000000, 0x00108010, 0x40008010);
 359         &data_word(0x40108000, 0x00008000, 0x00000000, 0x40000010);
 360         &data_word(0x00000010, 0x40108010, 0x00108000, 0x40100000);
 361         &data_word(0x40100010, 0x00100000, 0x00008010, 0x40008000);
 362         &data_word(0x40008010, 0x00000010, 0x40100000, 0x00108000);
 363         # nibble 2
 364         &data_word(0x04000001, 0x04040100, 0x00000100, 0x04000101);
 365         &data_word(0x00040001, 0x04000000, 0x04000101, 0x00040100);
 366         &data_word(0x04000100, 0x00040000, 0x04040000, 0x00000001);
 367         &data_word(0x04040101, 0x00000101, 0x00000001, 0x04040001);
 368         &data_word(0x00000000, 0x00040001, 0x04040100, 0x00000100);
 369         &data_word(0x00000101, 0x04040101, 0x00040000, 0x04000001);
 370         &data_word(0x04040001, 0x04000100, 0x00040101, 0x04040000);
 371         &data_word(0x00040100, 0x00000000, 0x04000000, 0x00040101);
 372         &data_word(0x04040100, 0x00000100, 0x00000001, 0x00040000);
 373         &data_word(0x00000101, 0x00040001, 0x04040000, 0x04000101);
 374         &data_word(0x00000000, 0x04040100, 0x00040100, 0x04040001);
 375         &data_word(0x00040001, 0x04000000, 0x04040101, 0x00000001);
 376         &data_word(0x00040101, 0x04000001, 0x04000000, 0x04040101);
 377         &data_word(0x00040000, 0x04000100, 0x04000101, 0x00040100);
 378         &data_word(0x04000100, 0x00000000, 0x04040001, 0x00000101);
 379         &data_word(0x04000001, 0x00040101, 0x00000100, 0x04040000);
 380         # nibble 3
 381         &data_word(0x00401008, 0x10001000, 0x00000008, 0x10401008);
 382         &data_word(0x00000000, 0x10400000, 0x10001008, 0x00400008);
 383         &data_word(0x10401000, 0x10000008, 0x10000000, 0x00001008);
 384         &data_word(0x10000008, 0x00401008, 0x00400000, 0x10000000);
 385         &data_word(0x10400008, 0x00401000, 0x00001000, 0x00000008);
 386         &data_word(0x00401000, 0x10001008, 0x10400000, 0x00001000);
 387         &data_word(0x00001008, 0x00000000, 0x00400008, 0x10401000);
 388         &data_word(0x10001000, 0x10400008, 0x10401008, 0x00400000);
 389         &data_word(0x10400008, 0x00001008, 0x00400000, 0x10000008);
 390         &data_word(0x00401000, 0x10001000, 0x00000008, 0x10400000);
 391         &data_word(0x10001008, 0x00000000, 0x00001000, 0x00400008);
 392         &data_word(0x00000000, 0x10400008, 0x10401000, 0x00001000);
 393         &data_word(0x10000000, 0x10401008, 0x00401008, 0x00400000);
 394         &data_word(0x10401008, 0x00000008, 0x10001000, 0x00401008);
 395         &data_word(0x00400008, 0x00401000, 0x10400000, 0x10001008);
 396         &data_word(0x00001008, 0x10000000, 0x10000008, 0x10401000);
 397         # nibble 4
 398         &data_word(0x08000000, 0x00010000, 0x00000400, 0x08010420);
 399         &data_word(0x08010020, 0x08000400, 0x00010420, 0x08010000);
 400         &data_word(0x00010000, 0x00000020, 0x08000020, 0x00010400);
 401         &data_word(0x08000420, 0x08010020, 0x08010400, 0x00000000);
 402         &data_word(0x00010400, 0x08000000, 0x00010020, 0x00000420);
 403         &data_word(0x08000400, 0x00010420, 0x00000000, 0x08000020);
 404         &data_word(0x00000020, 0x08000420, 0x08010420, 0x00010020);
 405         &data_word(0x08010000, 0x00000400, 0x00000420, 0x08010400);
 406         &data_word(0x08010400, 0x08000420, 0x00010020, 0x08010000);
 407         &data_word(0x00010000, 0x00000020, 0x08000020, 0x08000400);
 408         &data_word(0x08000000, 0x00010400, 0x08010420, 0x00000000);
 409         &data_word(0x00010420, 0x08000000, 0x00000400, 0x00010020);
 410         &data_word(0x08000420, 0x00000400, 0x00000000, 0x08010420);
 411         &data_word(0x08010020, 0x08010400, 0x00000420, 0x00010000);
 412         &data_word(0x00010400, 0x08010020, 0x08000400, 0x00000420);
 413         &data_word(0x00000020, 0x00010420, 0x08010000, 0x08000020);
 414         # nibble 5
 415         &data_word(0x80000040, 0x00200040, 0x00000000, 0x80202000);
 416         &data_word(0x00200040, 0x00002000, 0x80002040, 0x00200000);
 417         &data_word(0x00002040, 0x80202040, 0x00202000, 0x80000000);
 418         &data_word(0x80002000, 0x80000040, 0x80200000, 0x00202040);
 419         &data_word(0x00200000, 0x80002040, 0x80200040, 0x00000000);
 420         &data_word(0x00002000, 0x00000040, 0x80202000, 0x80200040);
 421         &data_word(0x80202040, 0x80200000, 0x80000000, 0x00002040);
 422         &data_word(0x00000040, 0x00202000, 0x00202040, 0x80002000);
 423         &data_word(0x00002040, 0x80000000, 0x80002000, 0x00202040);
 424         &data_word(0x80202000, 0x00200040, 0x00000000, 0x80002000);
 425         &data_word(0x80000000, 0x00002000, 0x80200040, 0x00200000);
 426         &data_word(0x00200040, 0x80202040, 0x00202000, 0x00000040);
 427         &data_word(0x80202040, 0x00202000, 0x00200000, 0x80002040);
 428         &data_word(0x80000040, 0x80200000, 0x00202040, 0x00000000);
 429         &data_word(0x00002000, 0x80000040, 0x80002040, 0x80202000);
 430         &data_word(0x80200000, 0x00002040, 0x00000040, 0x80200040);
 431         # nibble 6
 432         &data_word(0x00004000, 0x00000200, 0x01000200, 0x01000004);
 433         &data_word(0x01004204, 0x00004004, 0x00004200, 0x00000000);
 434         &data_word(0x01000000, 0x01000204, 0x00000204, 0x01004000);
 435         &data_word(0x00000004, 0x01004200, 0x01004000, 0x00000204);
 436         &data_word(0x01000204, 0x00004000, 0x00004004, 0x01004204);
 437         &data_word(0x00000000, 0x01000200, 0x01000004, 0x00004200);
 438         &data_word(0x01004004, 0x00004204, 0x01004200, 0x00000004);
 439         &data_word(0x00004204, 0x01004004, 0x00000200, 0x01000000);
 440         &data_word(0x00004204, 0x01004000, 0x01004004, 0x00000204);
 441         &data_word(0x00004000, 0x00000200, 0x01000000, 0x01004004);
 442         &data_word(0x01000204, 0x00004204, 0x00004200, 0x00000000);
 443         &data_word(0x00000200, 0x01000004, 0x00000004, 0x01000200);
 444         &data_word(0x00000000, 0x01000204, 0x01000200, 0x00004200);
 445         &data_word(0x00000204, 0x00004000, 0x01004204, 0x01000000);
 446         &data_word(0x01004200, 0x00000004, 0x00004004, 0x01004204);
 447         &data_word(0x01000004, 0x01004200, 0x01004000, 0x00004004);
 448         # nibble 7
 449         &data_word(0x20800080, 0x20820000, 0x00020080, 0x00000000);
 450         &data_word(0x20020000, 0x00800080, 0x20800000, 0x20820080);
 451         &data_word(0x00000080, 0x20000000, 0x00820000, 0x00020080);
 452         &data_word(0x00820080, 0x20020080, 0x20000080, 0x20800000);
 453         &data_word(0x00020000, 0x00820080, 0x00800080, 0x20020000);
 454         &data_word(0x20820080, 0x20000080, 0x00000000, 0x00820000);
 455         &data_word(0x20000000, 0x00800000, 0x20020080, 0x20800080);
 456         &data_word(0x00800000, 0x00020000, 0x20820000, 0x00000080);
 457         &data_word(0x00800000, 0x00020000, 0x20000080, 0x20820080);
 458         &data_word(0x00020080, 0x20000000, 0x00000000, 0x00820000);
 459         &data_word(0x20800080, 0x20020080, 0x20020000, 0x00800080);
 460         &data_word(0x20820000, 0x00000080, 0x00800080, 0x20020000);
 461         &data_word(0x20820080, 0x00800000, 0x20800000, 0x20000080);
 462         &data_word(0x00820000, 0x00020080, 0x20020080, 0x20800000);
 463         &data_word(0x00000080, 0x20820000, 0x00820080, 0x00000000);
 464         &data_word(0x20000000, 0x20800080, 0x00020000, 0x00820080);
 465         }