crypto/aes/asm/aes-586.pl

   1 #!/usr/bin/env perl
   2 #
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
   5 # project. Rights for redistribution and usage in source and binary
   6 # forms are granted according to the OpenSSL license.
   7 # ====================================================================
   8 #
   9 # Version 3.2.
  10 #
  11 # You might fail to appreciate this module performance from the first
  12 # try. If compared to "vanilla" linux-ia32-icc target, i.e. considered
  13 # to be *the* best Intel C compiler without -KPIC, performance appears
  14 # to be virtually identical... But try to re-configure with shared
  15 # library support... Aha! Intel compiler "suddenly" lags behind by 30%
  16 # [on P4, more on others]:-) And if compared to position-independent
  17 # code generated by GNU C, this code performs *more* than *twice* as
  18 # fast! Yes, all this buzz about PIC means that unlike other hand-
  19 # coded implementations, this one was explicitly designed to be safe
  20 # to use even in shared library context... This also means that this
  21 # code isn't necessarily absolutely fastest "ever," because in order
  22 # to achieve position independence an extra register has to be
  23 # off-loaded to stack, which affects the benchmark result.
  24 #
  25 # Special note about instruction choice. Do you recall RC4_INT code
  26 # performing poorly on P4? It might be the time to figure out why.
  27 # RC4_INT code implies effective address calculations in base+offset*4
  28 # form. Trouble is that it seems that offset scaling turned to be
  29 # critical path... At least eliminating scaling resulted in 2.8x RC4
  30 # performance improvement [as you might recall]. As AES code is hungry
  31 # for scaling too, I [try to] avoid the latter by favoring off-by-2
  32 # shifts and masking the result with 0xFF<<2 instead of "boring" 0xFF.
  33 #
  34 # As was shown by Dean Gaudet <dean@arctic.org>, the above note turned
  35 # void. Performance improvement with off-by-2 shifts was observed on
  36 # intermediate implementation, which was spilling yet another register
  37 # to stack... Final offset*4 code below runs just a tad faster on P4,
  38 # but exhibits up to 10% improvement on other cores.
  39 #
  40 # Second version is "monolithic" replacement for aes_core.c, which in
  41 # addition to AES_[de|en]crypt implements AES_set_[de|en]cryption_key.
  42 # This made it possible to implement little-endian variant of the
  43 # algorithm without modifying the base C code. Motivating factor for
  44 # the undertaken effort was that it appeared that in tight IA-32
  45 # register window little-endian flavor could achieve slightly higher
  46 # Instruction Level Parallelism, and it indeed resulted in up to 15%
  47 # better performance on most recent µ-archs...
  48 #
  49 # Third version adds AES_cbc_encrypt implementation, which resulted in
  50 # up to 40% performance imrovement of CBC benchmark results. 40% was
  51 # observed on P4 core, where "overall" imrovement coefficient, i.e. if
  52 # compared to PIC generated by GCC and in CBC mode, was observed to be
  53 # as large as 4x:-) CBC performance is virtually identical to ECB now
  54 # and on some platforms even better, e.g. 17.6 "small" cycles/byte on
  55 # Opteron, because certain function prologues and epilogues are
  56 # effectively taken out of the loop...
  57 #
  58 # Version 3.2 implements compressed tables and prefetch of these tables
  59 # in CBC[!] mode. Former means that 3/4 of table references are now
  60 # misaligned, which unfortunately has negative impact on elder IA-32
  61 # implementations, Pentium suffered 30% penalty, PIII - 10%.
  62 #
  63 # Current ECB performance numbers for 128-bit key in CPU cycles per
  64 # processed byte [measure commonly used by AES benchmarkers] are:
  65 #
  66 #               small footprint         fully unrolled
  67 # P4            24                      22
  68 # AMD K8        20                      19
  69 # PIII          25                      23
  70 # Pentium       81                      78
  71
  72 push(@INC,"perlasm","../../perlasm");
  73 require "x86asm.pl";
  74
  75 &asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386");
  76
  77 $s0="eax";
  78 $s1="ebx";
  79 $s2="ecx";
  80 $s3="edx";
  81 $key="edi";
  82 $acc="esi";
  83
  84 $small_footprint=1;     # $small_footprint=1 code is ~5% slower [on
  85                         # recent µ-archs], but ~5 times smaller!
  86                         # I favor compact code to minimize cache
  87                         # contention and in hope to "collect" 5% back
  88                         # in real-life applications...
  89 $vertical_spin=0;       # shift "verticaly" defaults to 0, because of
  90                         # its proof-of-concept status...
  91
  92 # Note that there is no decvert(), as well as last encryption round is
  93 # performed with "horizontal" shifts. This is because this "vertical"
  94 # implementation [one which groups shifts on a given $s[i] to form a
  95 # "column," unlike "horizontal" one, which groups shifts on different
  96 # $s[i] to form a "row"] is work in progress. It was observed to run
  97 # few percents faster on Intel cores, but not AMD. On AMD K8 core it's
  98 # whole 12% slower:-( So we face a trade-off... Shall it be resolved
  99 # some day? Till then the code is considered experimental and by
 100 # default remains dormant...
 101
 102 sub encvert()
 103 { my ($te,@s) = @_;
 104   my $v0 = $acc, $v1 = $key;
 105
 106         &mov    ($v0,$s[3]);                            # copy s3
 107         &mov    (&DWP(0,"esp"),$s[2]);                  # save s2
 108         &mov    ($v1,$s[0]);                            # copy s0
 109         &mov    (&DWP(4,"esp"),$s[1]);                  # save s1
 110
 111         &movz   ($s[2],&HB($s[0]));
 112         &and    ($s[0],0xFF);
 113         &mov    ($s[0],&DWP(0,$te,$s[0],8));            # s0>>0
 114         &shr    ($v1,16);
 115         &mov    ($s[3],&DWP(3,$te,$s[2],8));            # s0>>8
 116         &movz   ($s[1],&HB($v1));
 117         &and    ($v1,0xFF);
 118         &mov    ($s[2],&DWP(2,$te,$v1,8));              # s0>>16
 119          &mov   ($v1,$v0);
 120         &mov    ($s[1],&DWP(1,$te,$s[1],8));            # s0>>24
 121
 122         &and    ($v0,0xFF);
 123         &xor    ($s[3],&DWP(0,$te,$v0,8));              # s3>>0
 124         &movz   ($v0,&HB($v1));
 125         &shr    ($v1,16);
 126         &xor    ($s[2],&DWP(3,$te,$v0,8));              # s3>>8
 127         &movz   ($v0,&HB($v1));
 128         &and    ($v1,0xFF);
 129         &xor    ($s[1],&DWP(2,$te,$v1,8));              # s3>>16
 130          &mov   ($v1,&DWP(0,"esp"));                    # restore s2
 131         &xor    ($s[0],&DWP(1,$te,$v0,8));              # s3>>24
 132
 133         &mov    ($v0,$v1);
 134         &and    ($v1,0xFF);
 135         &xor    ($s[2],&DWP(0,$te,$v1,8));              # s2>>0
 136         &movz   ($v1,&HB($v0));
 137         &shr    ($v0,16);
 138         &xor    ($s[1],&DWP(3,$te,$v1,8));              # s2>>8
 139         &movz   ($v1,&HB($v0));
 140         &and    ($v0,0xFF);
 141         &xor    ($s[0],&DWP(2,$te,$v0,8));              # s2>>16
 142          &mov   ($v0,&DWP(4,"esp"));                    # restore s1
 143         &xor    ($s[3],&DWP(1,$te,$v1,8));              # s2>>24
 144
 145         &mov    ($v1,$v0);
 146         &and    ($v0,0xFF);
 147         &xor    ($s[1],&DWP(0,$te,$v0,8));              # s1>>0
 148         &movz   ($v0,&HB($v1));
 149         &shr    ($v1,16);
 150         &xor    ($s[0],&DWP(3,$te,$v0,8));              # s1>>8
 151         &movz   ($v0,&HB($v1));
 152         &and    ($v1,0xFF);
 153         &xor    ($s[3],&DWP(2,$te,$v1,8));              # s1>>16
 154          &mov   ($key,&DWP(12,"esp"));                  # reincarnate v1 as key
 155         &xor    ($s[2],&DWP(1,$te,$v0,8));              # s1>>24
 156 }
 157
 158 sub encstep()
 159 { my ($i,$te,@s) = @_;
 160   my $tmp = $key;
 161   my $out = $i==3?$s[0]:$acc;
 162
 163         # lines marked with #%e?x[i] denote "reordered" instructions...
 164         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
 165         else        {   &mov    ($out,$s[0]);
 166                         &and    ($out,0xFF);                    }
 167         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
 168         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
 169                         &mov    ($out,&DWP(0,$te,$out,8));
 170
 171         if ($i==3)  {   $tmp=$s[1];                             }##%eax
 172                         &movz   ($tmp,&HB($s[1]));
 173                         &xor    ($out,&DWP(3,$te,$tmp,8));
 174
 175         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(0,"esp")); }##%ebx
 176         else        {   &mov    ($tmp,$s[2]);
 177                         &shr    ($tmp,16);                      }
 178         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
 179                         &and    ($tmp,0xFF);
 180                         &xor    ($out,&DWP(2,$te,$tmp,8));
 181
 182         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }##%ecx
 183         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
 184         else        {   &mov    ($tmp,$s[3]);
 185                         &shr    ($tmp,24)                       }
 186                         &xor    ($out,&DWP(1,$te,$tmp,8));
 187         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
 188         if ($i==3)  {   &mov    ($s[3],$acc);                   }
 189                         &comment();
 190 }
 191
 192 sub enclast()
 193 { my ($i,$te,@s)=@_;
 194   my $tmp = $key;
 195   my $out = $i==3?$s[0]:$acc;
 196
 197         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
 198         else        {   &mov    ($out,$s[0]);                   }
 199                         &and    ($out,0xFF);
 200         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
 201         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
 202                         &mov    ($out,&DWP(2,$te,$out,8));
 203                         &and    ($out,0x000000ff);
 204
 205         if ($i==3)  {   $tmp=$s[1];                             }##%eax
 206                         &movz   ($tmp,&HB($s[1]));
 207                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
 208                         &and    ($tmp,0x0000ff00);
 209                         &xor    ($out,$tmp);
 210
 211         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(0,"esp")); }##%ebx
 212         else        {   mov     ($tmp,$s[2]);
 213                         &shr    ($tmp,16);                      }
 214         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
 215                         &and    ($tmp,0xFF);
 216                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
 217                         &and    ($tmp,0x00ff0000);
 218                         &xor    ($out,$tmp);
 219
 220         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }##%ecx
 221         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
 222         else        {   &mov    ($tmp,$s[3]);
 223                         &shr    ($tmp,24);                      }
 224                         &mov    ($tmp,&DWP(2,$te,$tmp,8));
 225                         &and    ($tmp,0xff000000);
 226                         &xor    ($out,$tmp);
 227         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
 228         if ($i==3)  {   &mov    ($s[3],$acc);                   }
 229 }
 230
 231 sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 232
 233 &public_label("AES_Te");
 234 &function_begin_B("_x86_AES_encrypt");
 235         if ($vertical_spin) {
 236                 # I need high parts of volatile registers to be accessible...
 237                 &exch   ($s1="edi",$key="ebx");
 238                 &mov    ($s2="esi",$acc="ecx");
 239         }
 240
 241         # allocate aligned stack frame
 242         &mov    ($acc,"esp");
 243         &sub    ("esp",20);
 244         &and    ("esp",-16);
 245
 246         &mov    (&DWP(12,"esp"),$key);          # save key
 247         &mov    (&DWP(16,"esp"),$acc);          # save %esp
 248
 249         &xor    ($s0,&DWP(0,$key));             # xor with key
 250         &xor    ($s1,&DWP(4,$key));
 251         &xor    ($s2,&DWP(8,$key));
 252         &xor    ($s3,&DWP(12,$key));
 253
 254         &mov    ($acc,&DWP(240,$key));          # load key->rounds
 255
 256         if ($small_footprint) {
 257             &lea        ($acc,&DWP(-2,$acc,$acc));
 258             &lea        ($acc,&DWP(0,$key,$acc,8));
 259             &mov        (&DWP(8,"esp"),$acc);   # end of key schedule
 260             &align      (4);
 261             &set_label("loop");
 262                 if ($vertical_spin) {
 263                     &encvert("ebp",$s0,$s1,$s2,$s3);
 264                 } else {
 265                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 266                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 267                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 268                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 269                 }
 270                 &add    ($key,16);                      # advance rd_key
 271                 &xor    ($s0,&DWP(0,$key));
 272                 &xor    ($s1,&DWP(4,$key));
 273                 &xor    ($s2,&DWP(8,$key));
 274                 &xor    ($s3,&DWP(12,$key));
 275             &cmp        ($key,&DWP(8,"esp"));
 276             &mov        (&DWP(12,"esp"),$key);
 277             &jb         (&label("loop"));
 278         }
 279         else {
 280             &cmp        ($acc,10);
 281             &jle        (&label("10rounds"));
 282             &cmp        ($acc,12);
 283             &jle        (&label("12rounds"));
 284
 285         &set_label("14rounds");
 286             for ($i=1;$i<3;$i++) {
 287                 if ($vertical_spin) {
 288                     &encvert("ebp",$s0,$s1,$s2,$s3);
 289                 } else {
 290                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 291                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 292                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 293                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 294                 }
 295                 &xor    ($s0,&DWP(16*$i+0,$key));
 296                 &xor    ($s1,&DWP(16*$i+4,$key));
 297                 &xor    ($s2,&DWP(16*$i+8,$key));
 298                 &xor    ($s3,&DWP(16*$i+12,$key));
 299             }
 300             &add        ($key,32);
 301             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
 302         &set_label("12rounds");
 303             for ($i=1;$i<3;$i++) {
 304                 if ($vertical_spin) {
 305                     &encvert("ebp",$s0,$s1,$s2,$s3);
 306                 } else {
 307                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 308                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 309                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 310                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 311                 }
 312                 &xor    ($s0,&DWP(16*$i+0,$key));
 313                 &xor    ($s1,&DWP(16*$i+4,$key));
 314                 &xor    ($s2,&DWP(16*$i+8,$key));
 315                 &xor    ($s3,&DWP(16*$i+12,$key));
 316             }
 317             &add        ($key,32);
 318             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
 319         &set_label("10rounds");
 320             for ($i=1;$i<10;$i++) {
 321                 if ($vertical_spin) {
 322                     &encvert("ebp",$s0,$s1,$s2,$s3);
 323                 } else {
 324                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 325                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 326                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 327                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 328                 }
 329                 &xor    ($s0,&DWP(16*$i+0,$key));
 330                 &xor    ($s1,&DWP(16*$i+4,$key));
 331                 &xor    ($s2,&DWP(16*$i+8,$key));
 332                 &xor    ($s3,&DWP(16*$i+12,$key));
 333             }
 334         }
 335
 336         if ($vertical_spin) {
 337             # "reincarnate" some registers for "horizontal" spin...
 338             &mov        ($s1="ebx",$key="edi");
 339             &mov        ($s2="ecx",$acc="esi");
 340         }
 341         &enclast(0,"ebp",$s0,$s1,$s2,$s3);
 342         &enclast(1,"ebp",$s1,$s2,$s3,$s0);
 343         &enclast(2,"ebp",$s2,$s3,$s0,$s1);
 344         &enclast(3,"ebp",$s3,$s0,$s1,$s2);
 345
 346         &mov    ("esp",&DWP(16,"esp"));         # restore %esp
 347         &add    ($key,$small_footprint?16:160);
 348         &xor    ($s0,&DWP(0,$key));
 349         &xor    ($s1,&DWP(4,$key));
 350         &xor    ($s2,&DWP(8,$key));
 351         &xor    ($s3,&DWP(12,$key));
 352
 353         &ret    ();
 354
 355 &set_label("AES_Te",64);        # Yes! I keep it in the code segment!
 356         &_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
 357         &_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
 358         &_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
 359         &_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
 360         &_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
 361         &_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
 362         &_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
 363         &_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
 364         &_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
 365         &_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
 366         &_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
 367         &_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
 368         &_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
 369         &_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
 370         &_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
 371         &_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
 372         &_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
 373         &_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
 374         &_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
 375         &_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
 376         &_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
 377         &_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
 378         &_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
 379         &_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
 380         &_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
 381         &_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
 382         &_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
 383         &_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
 384         &_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
 385         &_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
 386         &_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
 387         &_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
 388         &_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
 389         &_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
 390         &_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
 391         &_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
 392         &_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
 393         &_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
 394         &_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
 395         &_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
 396         &_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
 397         &_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
 398         &_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
 399         &_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
 400         &_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
 401         &_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
 402         &_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
 403         &_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
 404         &_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
 405         &_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
 406         &_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
 407         &_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
 408         &_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
 409         &_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
 410         &_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
 411         &_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
 412         &_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
 413         &_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
 414         &_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
 415         &_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
 416         &_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
 417         &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
 418         &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
 419         &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
 420 #rcon:
 421         &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
 422         &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
 423         &data_word(0x0000001b, 0x00000036);
 424 &function_end_B("_x86_AES_encrypt");
 425
 426 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
 427 &public_label("AES_Te");
 428 &function_begin("AES_encrypt");
 429         &mov    ($acc,&wparam(0));              # load inp
 430         &mov    ($key,&wparam(2));              # load key
 431
 432         &call   (&label("pic_point"));          # make it PIC!
 433         &set_label("pic_point");
 434         &blindpop("ebp");
 435         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
 436
 437         &mov    ($s0,&DWP(0,$acc));             # load input data
 438         &mov    ($s1,&DWP(4,$acc));
 439         &mov    ($s2,&DWP(8,$acc));
 440         &mov    ($s3,&DWP(12,$acc));
 441
 442         &call   ("_x86_AES_encrypt");
 443
 444         &mov    ($acc,&wparam(1));              # load out
 445         &mov    (&DWP(0,$acc),$s0);             # write output data
 446         &mov    (&DWP(4,$acc),$s1);
 447         &mov    (&DWP(8,$acc),$s2);
 448         &mov    (&DWP(12,$acc),$s3);
 449 &function_end("AES_encrypt");
 450
 451 #------------------------------------------------------------------#
 452
 453 sub decstep()
 454 { my ($i,$td,@s) = @_;
 455   my $tmp = $key;
 456   my $out = $i==3?$s[0]:$acc;
 457
 458         # no instructions are reordered, as performance appears
 459         # optimal... or rather that all attempts to reorder didn't
 460         # result in better performance [which by the way is not a
 461         # bit lower than ecryption].
 462         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
 463         else        {   &mov    ($out,$s[0]);                   }
 464                         &and    ($out,0xFF);
 465                         &mov    ($out,&DWP(0,$td,$out,8));
 466
 467         if ($i==3)  {   $tmp=$s[1];                             }
 468                         &movz   ($tmp,&HB($s[1]));
 469                         &xor    ($out,&DWP(3,$td,$tmp,8));
 470
 471         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
 472         else        {   &mov    ($tmp,$s[2]);                   }
 473                         &shr    ($tmp,16);
 474                         &and    ($tmp,0xFF);
 475                         &xor    ($out,&DWP(2,$td,$tmp,8));
 476
 477         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }
 478         else        {   &mov    ($tmp,$s[3]);                   }
 479                         &shr    ($tmp,24);
 480                         &xor    ($out,&DWP(1,$td,$tmp,8));
 481         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
 482         if ($i==3)  {   &mov    ($s[3],&DWP(0,"esp"));          }
 483                         &comment();
 484 }
 485
 486 sub declast()
 487 { my ($i,$td,@s)=@_;
 488   my $tmp = $key;
 489   my $out = $i==3?$s[0]:$acc;
 490
 491         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
 492         else        {   &mov    ($out,$s[0]);                   }
 493                         &and    ($out,0xFF);
 494                         &mov    ($out,&DWP(2048,$td,$out,4));
 495                         &and    ($out,0x000000ff);
 496
 497         if ($i==3)  {   $tmp=$s[1];                             }
 498                         &movz   ($tmp,&HB($s[1]));
 499                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 500                         &and    ($tmp,0x0000ff00);
 501                         &xor    ($out,$tmp);
 502
 503         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
 504         else        {   mov     ($tmp,$s[2]);                   }
 505                         &shr    ($tmp,16);
 506                         &and    ($tmp,0xFF);
 507                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 508                         &and    ($tmp,0x00ff0000);
 509                         &xor    ($out,$tmp);
 510
 511         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(4,"esp")); }
 512         else        {   &mov    ($tmp,$s[3]);                   }
 513                         &shr    ($tmp,24);
 514                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 515                         &and    ($tmp,0xff000000);
 516                         &xor    ($out,$tmp);
 517         if ($i<2)   {   &mov    (&DWP(4*$i,"esp"),$out);        }
 518         if ($i==3)  {   &mov    ($s[3],&DWP(0,"esp"));          }
 519 }
 520
 521 &public_label("AES_Td");
 522 &function_begin_B("_x86_AES_decrypt");
 523         # allocate aligned stack frame
 524         &mov    ($acc,"esp");
 525         &sub    ("esp",20);
 526         &and    ("esp",-16);
 527
 528         &mov    (&DWP(12,"esp"),$key);          # save key
 529         &mov    (&DWP(16,"esp"),$acc);          # save %esp
 530
 531         &xor    ($s0,&DWP(0,$key));             # xor with key
 532         &xor    ($s1,&DWP(4,$key));
 533         &xor    ($s2,&DWP(8,$key));
 534         &xor    ($s3,&DWP(12,$key));
 535
 536         &mov    ($acc,&DWP(240,$key));          # load key->rounds
 537
 538         if ($small_footprint) {
 539             &lea        ($acc,&DWP(-2,$acc,$acc));
 540             &lea        ($acc,&DWP(0,$key,$acc,8));
 541             &mov        (&DWP(8,"esp"),$acc);   # end of key schedule
 542             &align      (4);
 543             &set_label("loop");
 544                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 545                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 546                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 547                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 548                 &add    ($key,16);                      # advance rd_key
 549                 &xor    ($s0,&DWP(0,$key));
 550                 &xor    ($s1,&DWP(4,$key));
 551                 &xor    ($s2,&DWP(8,$key));
 552                 &xor    ($s3,&DWP(12,$key));
 553             &cmp        ($key,&DWP(8,"esp"));
 554             &mov        (&DWP(12,"esp"),$key);
 555             &jb         (&label("loop"));
 556         }
 557         else {
 558             &cmp        ($acc,10);
 559             &jle        (&label("10rounds"));
 560             &cmp        ($acc,12);
 561             &jle        (&label("12rounds"));
 562
 563         &set_label("14rounds");
 564             for ($i=1;$i<3;$i++) {
 565                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 566                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 567                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 568                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 569                 &xor    ($s0,&DWP(16*$i+0,$key));
 570                 &xor    ($s1,&DWP(16*$i+4,$key));
 571                 &xor    ($s2,&DWP(16*$i+8,$key));
 572                 &xor    ($s3,&DWP(16*$i+12,$key));
 573             }
 574             &add        ($key,32);
 575             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
 576         &set_label("12rounds");
 577             for ($i=1;$i<3;$i++) {
 578                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 579                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 580                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 581                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 582                 &xor    ($s0,&DWP(16*$i+0,$key));
 583                 &xor    ($s1,&DWP(16*$i+4,$key));
 584                 &xor    ($s2,&DWP(16*$i+8,$key));
 585                 &xor    ($s3,&DWP(16*$i+12,$key));
 586             }
 587             &add        ($key,32);
 588             &mov        (&DWP(12,"esp"),$key);          # advance rd_key
 589         &set_label("10rounds");
 590             for ($i=1;$i<10;$i++) {
 591                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 592                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 593                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 594                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 595                 &xor    ($s0,&DWP(16*$i+0,$key));
 596                 &xor    ($s1,&DWP(16*$i+4,$key));
 597                 &xor    ($s2,&DWP(16*$i+8,$key));
 598                 &xor    ($s3,&DWP(16*$i+12,$key));
 599             }
 600         }
 601
 602         &declast(0,"ebp",$s0,$s3,$s2,$s1);
 603         &declast(1,"ebp",$s1,$s0,$s3,$s2);
 604         &declast(2,"ebp",$s2,$s1,$s0,$s3);
 605         &declast(3,"ebp",$s3,$s2,$s1,$s0);
 606
 607         &mov    ("esp",&DWP(16,"esp"));         # restore %esp
 608         &add    ($key,$small_footprint?16:160);
 609         &xor    ($s0,&DWP(0,$key));
 610         &xor    ($s1,&DWP(4,$key));
 611         &xor    ($s2,&DWP(8,$key));
 612         &xor    ($s3,&DWP(12,$key));
 613
 614         &ret    ();
 615
 616 &set_label("AES_Td",64);        # Yes! I keep it in the code segment!
 617         &_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
 618         &_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
 619         &_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
 620         &_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
 621         &_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
 622         &_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
 623         &_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
 624         &_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
 625         &_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
 626         &_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
 627         &_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
 628         &_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
 629         &_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
 630         &_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
 631         &_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
 632         &_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
 633         &_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
 634         &_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
 635         &_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
 636         &_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
 637         &_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
 638         &_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
 639         &_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
 640         &_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
 641         &_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
 642         &_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
 643         &_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
 644         &_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
 645         &_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
 646         &_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
 647         &_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
 648         &_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
 649         &_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
 650         &_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
 651         &_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
 652         &_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
 653         &_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
 654         &_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
 655         &_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
 656         &_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
 657         &_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
 658         &_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
 659         &_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
 660         &_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
 661         &_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
 662         &_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
 663         &_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
 664         &_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
 665         &_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
 666         &_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
 667         &_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
 668         &_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
 669         &_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
 670         &_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
 671         &_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
 672         &_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
 673         &_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
 674         &_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
 675         &_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
 676         &_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
 677         &_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
 678         &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
 679         &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
 680         &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
 681 #Td4:
 682         &data_word(0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5);
 683         &data_word(0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838);
 684         &data_word(0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e);
 685         &data_word(0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb);
 686         &data_word(0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282);
 687         &data_word(0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787);
 688         &data_word(0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444);
 689         &data_word(0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb);
 690         &data_word(0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232);
 691         &data_word(0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d);
 692         &data_word(0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b);
 693         &data_word(0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e);
 694         &data_word(0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666);
 695         &data_word(0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2);
 696         &data_word(0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949);
 697         &data_word(0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525);
 698         &data_word(0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464);
 699         &data_word(0x86868686, 0x68686868, 0x98989898, 0x16161616);
 700         &data_word(0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc);
 701         &data_word(0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292);
 702         &data_word(0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050);
 703         &data_word(0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada);
 704         &data_word(0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757);
 705         &data_word(0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484);
 706         &data_word(0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000);
 707         &data_word(0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a);
 708         &data_word(0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505);
 709         &data_word(0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606);
 710         &data_word(0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f);
 711         &data_word(0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202);
 712         &data_word(0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303);
 713         &data_word(0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b);
 714         &data_word(0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141);
 715         &data_word(0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea);
 716         &data_word(0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece);
 717         &data_word(0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373);
 718         &data_word(0x96969696, 0xacacacac, 0x74747474, 0x22222222);
 719         &data_word(0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585);
 720         &data_word(0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8);
 721         &data_word(0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e);
 722         &data_word(0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171);
 723         &data_word(0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989);
 724         &data_word(0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e);
 725         &data_word(0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b);
 726         &data_word(0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b);
 727         &data_word(0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020);
 728         &data_word(0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe);
 729         &data_word(0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4);
 730         &data_word(0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333);
 731         &data_word(0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131);
 732         &data_word(0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959);
 733         &data_word(0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f);
 734         &data_word(0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9);
 735         &data_word(0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d);
 736         &data_word(0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f);
 737         &data_word(0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef);
 738         &data_word(0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d);
 739         &data_word(0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0);
 740         &data_word(0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c);
 741         &data_word(0x83838383, 0x53535353, 0x99999999, 0x61616161);
 742         &data_word(0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e);
 743         &data_word(0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626);
 744         &data_word(0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363);
 745         &data_word(0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d);
 746 &function_end_B("_x86_AES_decrypt");
 747
 748 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
 749 &public_label("AES_Td");
 750 &function_begin("AES_decrypt");
 751         &mov    ($acc,&wparam(0));              # load inp
 752         &mov    ($key,&wparam(2));              # load key
 753
 754         &call   (&label("pic_point"));          # make it PIC!
 755         &set_label("pic_point");
 756         &blindpop("ebp");
 757         &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
 758
 759         &mov    ($s0,&DWP(0,$acc));             # load input data
 760         &mov    ($s1,&DWP(4,$acc));
 761         &mov    ($s2,&DWP(8,$acc));
 762         &mov    ($s3,&DWP(12,$acc));
 763
 764         &call   ("_x86_AES_decrypt");
 765
 766         &mov    ($acc,&wparam(1));              # load out
 767         &mov    (&DWP(0,$acc),$s0);             # write output data
 768         &mov    (&DWP(4,$acc),$s1);
 769         &mov    (&DWP(8,$acc),$s2);
 770         &mov    (&DWP(12,$acc),$s3);
 771 &function_end("AES_decrypt");
 772
 773 # void AES_cbc_encrypt (const void char *inp, unsigned char *out,
 774 #                       size_t length, const AES_KEY *key,
 775 #                       unsigned char *ivp,const int enc);
 776 &public_label("AES_Te");
 777 &public_label("AES_Td");
 778 &function_begin("AES_cbc_encrypt");
 779         &mov    ($s2 eq "ecx"? $s2 : "",&wparam(2));    # load len
 780         &cmp    ($s2,0);
 781         &je     (&label("enc_out"));
 782
 783         &call   (&label("pic_point"));          # make it PIC!
 784         &set_label("pic_point");
 785         &blindpop("ebp");
 786
 787         &cmp    (&wparam(5),0);
 788         &je     (&label("DECRYPT"));
 789
 790         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
 791
 792         &mov    ($acc,&wparam(0));              # load inp
 793         &mov    ($key,&wparam(4));              # load ivp
 794
 795         &test   ($s2,0xFFFFFFF0);
 796         &jz     (&label("enc_tail"));           # short input...
 797
 798         # prefetch AES_Te
 799         for ($i=0;$i<2048;$i+=128)
 800         {       &mov    ($s0,&DWP($i+0,"ebp"));
 801                 &mov    ($s1,&DWP($i+32,"ebp"));
 802                 &mov    ($s2,&DWP($i+64,"ebp"));
 803                 &mov    ($s3,&DWP($i+96,"ebp"));
 804         }
 805
 806         &mov    ($s0,&DWP(0,$key));             # load iv
 807         &mov    ($s1,&DWP(4,$key));
 808
 809         &align  (4);
 810         &set_label("enc_loop");
 811                 &mov    ($s2,&DWP(8,$key));
 812                 &mov    ($s3,&DWP(12,$key));
 813
 814                 &xor    ($s0,&DWP(0,$acc));             # xor input data
 815                 &xor    ($s1,&DWP(4,$acc));
 816                 &xor    ($s2,&DWP(8,$acc));
 817                 &xor    ($s3,&DWP(12,$acc));
 818
 819                 &mov    ($key,&wparam(3));              # load key
 820                 &call   ("_x86_AES_encrypt");
 821
 822                 &mov    ($acc,&wparam(0));              # load inp
 823                 &mov    ($key,&wparam(1));              # load out
 824
 825                 &mov    (&DWP(0,$key),$s0);             # save output data
 826                 &mov    (&DWP(4,$key),$s1);
 827                 &mov    (&DWP(8,$key),$s2);
 828                 &mov    (&DWP(12,$key),$s3);
 829
 830                 &mov    ($s2,&wparam(2));               # load len
 831
 832                 &lea    ($acc,&DWP(16,$acc));
 833                 &mov    (&wparam(0),$acc);              # save inp
 834
 835                 &lea    ($s3,&DWP(16,$key));
 836                 &mov    (&wparam(1),$s3);               # save out
 837
 838                 &sub    ($s2,16);
 839                 &test   ($s2,0xFFFFFFF0);
 840                 &mov    (&wparam(2),$s2);               # save len
 841         &jnz    (&label("enc_loop"));
 842         &test   ($s2,15);
 843         &jnz    (&label("enc_tail"));
 844         &mov    ($acc,&wparam(4));              # load ivp
 845         &mov    ($s2,&DWP(8,$key));             # restore last dwords
 846         &mov    ($s3,&DWP(12,$key));
 847         &mov    (&DWP(0,$acc),$s0);             # save iv
 848         &mov    (&DWP(4,$acc),$s1);
 849         &mov    (&DWP(8,$acc),$s2);
 850         &mov    (&DWP(12,$acc),$s3);
 851     &set_label("enc_out");
 852         &function_end_A();
 853
 854     &align      (4);
 855     &set_label("enc_tail");
 856         &push   ($key eq "edi" ? $key : "");    # push ivp
 857         &pushf  ();
 858         &mov    ($key,&wparam(1));              # load out
 859         &mov    ($s1,16);
 860         &sub    ($s1,$s2);
 861         &cmp    ($key,$acc);                    # compare with inp
 862         &je     (&label("enc_in_place"));
 863         &data_word(0x90A4F3FC); # cld; rep movsb; nop   # copy input
 864         &jmp    (&label("enc_skip_in_place"));
 865     &set_label("enc_in_place");
 866         &lea    ($key,&DWP(0,$key,$s2));
 867     &set_label("enc_skip_in_place");
 868         &mov    ($s2,$s1);
 869         &xor    ($s0,$s0);
 870         &data_word(0x90AAF3FC); # cld; rep stosb; nop   # zero tail
 871         &popf   ();
 872         &pop    ($key);                         # pop ivp
 873
 874         # prefetch AES_Te
 875         for ($i=0;$i<2048;$i+=128)
 876         {       &mov    ($s0,&DWP($i+0,"ebp"));
 877                 &mov    ($s1,&DWP($i+32,"ebp"));
 878                 &mov    ($s2,&DWP($i+64,"ebp"));
 879                 &mov    ($s3,&DWP($i+96,"ebp"));
 880         }
 881
 882         &mov    ($acc,&wparam(1));              # output as input
 883         &mov    ($s0,&DWP(0,$key));
 884         &mov    ($s1,&DWP(4,$key));
 885         &mov    (&wparam(2),16);                # len=16
 886         &jmp    (&label("enc_loop"));           # one more spin...
 887
 888 #----------------------------- DECRYPT -----------------------------#
 889 &align  (4);
 890 &set_label("DECRYPT");
 891     &stack_push(5);                             # allocate temp + ivp
 892
 893         &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
 894
 895         # prefetch AES_Td
 896         for ($i=0;$i<3072;$i+=128)
 897         {       &mov    ($s0,&DWP($i+0,"ebp"));
 898                 &mov    ($s1,&DWP($i+32,"ebp"));
 899                 &mov    ($s2,&DWP($i+64,"ebp"));
 900                 &mov    ($s3,&DWP($i+96,"ebp"));
 901         }
 902
 903         &mov    ($acc,&wparam(0));              # load inp
 904         &cmp    ($acc,&wparam(1));
 905         &je     (&label("dec_in_place"));       # in-place processing...
 906
 907         &mov    ($key,&wparam(4));              # load ivp
 908         &mov    (&swtmp(4),$key);
 909
 910         &align  (4);
 911         &set_label("dec_loop");
 912                 &mov    ($s0,&DWP(0,$acc));             # read input
 913                 &mov    ($s1,&DWP(4,$acc));
 914                 &mov    ($s2,&DWP(8,$acc));
 915                 &mov    ($s3,&DWP(12,$acc));
 916
 917                 &mov    ($key,&wparam(3));              # load key
 918                 &call   ("_x86_AES_decrypt");
 919
 920                 &mov    ($key,&swtmp(4));               # load ivp
 921                 &mov    ($acc,&wparam(2));              # load len
 922                 &xor    ($s0,&DWP(0,$key));             # xor iv
 923                 &xor    ($s1,&DWP(4,$key));
 924                 &xor    ($s2,&DWP(8,$key));
 925                 &xor    ($s3,&DWP(12,$key));
 926
 927                 &sub    ($acc,16);
 928                 &jc     (&label("dec_partial"));
 929                 &mov    (&wparam(2),$acc);              # save len
 930                 &mov    ($acc,&wparam(0));              # load inp
 931                 &mov    ($key,&wparam(1));              # load out
 932
 933                 &mov    (&DWP(0,$key),$s0);             # write output
 934                 &mov    (&DWP(4,$key),$s1);
 935                 &mov    (&DWP(8,$key),$s2);
 936                 &mov    (&DWP(12,$key),$s3);
 937
 938                 &mov    (&swtmp(4),$acc);               # save ivp
 939                 &lea    ($acc,&DWP(16,$acc));
 940                 &mov    (&wparam(0),$acc);              # save inp
 941
 942                 &lea    ($key,&DWP(16,$key));
 943                 &mov    (&wparam(1),$key);              # save out
 944
 945         &jnz    (&label("dec_loop"));
 946         &mov    ($key,&swtmp(4));       # load temp ivp
 947     &set_label("dec_end");
 948         &mov    ($acc,&wparam(4));      # load user ivp
 949         &mov    ($s0,&DWP(0,$key));     # load iv
 950         &mov    ($s1,&DWP(4,$key));
 951         &mov    ($s2,&DWP(8,$key));
 952         &mov    ($s3,&DWP(12,$key));
 953         &mov    (&DWP(0,$acc),$s0);     # copy back to user
 954         &mov    (&DWP(4,$acc),$s1);
 955         &mov    (&DWP(8,$acc),$s2);
 956         &mov    (&DWP(12,$acc),$s3);
 957         &jmp    (&label("dec_out"));
 958
 959     &align      (4);
 960     &set_label("dec_partial");
 961         &lea    ($key,&swtmp(0));
 962         &mov    (&DWP(0,$key),$s0);     # dump output to stack
 963         &mov    (&DWP(4,$key),$s1);
 964         &mov    (&DWP(8,$key),$s2);
 965         &mov    (&DWP(12,$key),$s3);
 966         &lea    ($s2 eq "ecx" ? $s2 : "",&DWP(16,$acc));
 967         &mov    ($acc eq "esi" ? $acc : "",$key);
 968         &mov    ($key eq "edi" ? $key : "",&wparam(1));
 969         &pushf  ();
 970         &data_word(0x90A4F3FC); # cld; rep movsb; nop   # copy output
 971         &popf   ();
 972         &mov    ($key,&wparam(0));      # load temp ivp
 973         &jmp    (&label("dec_end"));
 974
 975     &align      (4);
 976     &set_label("dec_in_place");
 977         &set_label("dec_in_place_loop");
 978                 &lea    ($key,&swtmp(0));
 979                 &mov    ($s0,&DWP(0,$acc));             # read input
 980                 &mov    ($s1,&DWP(4,$acc));
 981                 &mov    ($s2,&DWP(8,$acc));
 982                 &mov    ($s3,&DWP(12,$acc));
 983
 984                 &mov    (&DWP(0,$key),$s0);             # copy to temp
 985                 &mov    (&DWP(4,$key),$s1);
 986                 &mov    (&DWP(8,$key),$s2);
 987                 &mov    (&DWP(12,$key),$s3);
 988
 989                 &mov    ($key,&wparam(3));              # load key
 990                 &call   ("_x86_AES_decrypt");
 991
 992                 &mov    ($key,&wparam(4));              # load ivp
 993                 &mov    ($acc,&wparam(1));              # load out
 994                 &xor    ($s0,&DWP(0,$key));             # xor iv
 995                 &xor    ($s1,&DWP(4,$key));
 996                 &xor    ($s2,&DWP(8,$key));
 997                 &xor    ($s3,&DWP(12,$key));
 998
 999                 &mov    (&DWP(0,$acc),$s0);             # write output
1000                 &mov    (&DWP(4,$acc),$s1);
1001                 &mov    (&DWP(8,$acc),$s2);
1002                 &mov    (&DWP(12,$acc),$s3);
1003
1004                 &lea    ($acc,&DWP(16,$acc));
1005                 &mov    (&wparam(1),$acc);              # save out
1006
1007                 &lea    ($acc,&swtmp(0));
1008                 &mov    ($s0,&DWP(0,$acc));             # read temp
1009                 &mov    ($s1,&DWP(4,$acc));
1010                 &mov    ($s2,&DWP(8,$acc));
1011                 &mov    ($s3,&DWP(12,$acc));
1012
1013                 &mov    (&DWP(0,$key),$s0);             # copy iv
1014                 &mov    (&DWP(4,$key),$s1);
1015                 &mov    (&DWP(8,$key),$s2);
1016                 &mov    (&DWP(12,$key),$s3);
1017
1018                 &mov    ($acc,&wparam(0));              # load inp
1019
1020                 &lea    ($acc,&DWP(16,$acc));
1021                 &mov    (&wparam(0),$acc);              # save inp
1022
1023                 &mov    ($s2,&wparam(2));               # load len
1024                 &sub    ($s2,16);
1025                 &jc     (&label("dec_in_place_partial"));
1026                 &mov    (&wparam(2),$s2);               # save len
1027         &jnz    (&label("dec_in_place_loop"));
1028         &jmp    (&label("dec_out"));
1029
1030     &align      (4);
1031     &set_label("dec_in_place_partial");
1032         # one can argue if this is actually required...
1033         &mov    ($key eq "edi" ? $key : "",&wparam(1));
1034         &lea    ($acc eq "esi" ? $acc : "",&swtmp(0));
1035         &lea    ($key,&DWP(0,$key,$s2));
1036         &lea    ($acc,&DWP(16,$acc,$s2));
1037         &neg    ($s2 eq "ecx" ? $s2 : "");
1038         &pushf  ();
1039         &data_word(0x90A4F3FC); # cld; rep movsb; nop   # restore tail
1040         &popf   ();
1041
1042     &align      (4);
1043     &set_label("dec_out");
1044     &stack_pop(5);
1045 &function_end("AES_cbc_encrypt");
1046
1047 #------------------------------------------------------------------#
1048
1049 sub enckey()
1050 {
1051         &movz   ("esi",&LB("edx"));             # rk[i]>>0
1052         &mov    ("ebx",&DWP(2,"ebp","esi",8));
1053         &movz   ("esi",&HB("edx"));             # rk[i]>>8
1054         &and    ("ebx",0xFF000000);
1055         &xor    ("eax","ebx");
1056
1057         &mov    ("ebx",&DWP(2,"ebp","esi",8));
1058         &shr    ("edx",16);
1059         &and    ("ebx",0x000000FF);
1060         &movz   ("esi",&LB("edx"));             # rk[i]>>16
1061         &xor    ("eax","ebx");
1062
1063         &mov    ("ebx",&DWP(0,"ebp","esi",8));
1064         &movz   ("esi",&HB("edx"));             # rk[i]>>24
1065         &and    ("ebx",0x0000FF00);
1066         &xor    ("eax","ebx");
1067
1068         &mov    ("ebx",&DWP(0,"ebp","esi",8));
1069         &and    ("ebx",0x00FF0000);
1070         &xor    ("eax","ebx");
1071
1072         &xor    ("eax",&DWP(2048,"ebp","ecx",4));       # rcon
1073 }
1074
1075 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
1076 #                        AES_KEY *key)
1077 &public_label("AES_Te");
1078 &function_begin("AES_set_encrypt_key");
1079         &mov    ("esi",&wparam(0));             # user supplied key
1080         &mov    ("edi",&wparam(2));             # private key schedule
1081
1082         &test   ("esi",-1);
1083         &jz     (&label("badpointer"));
1084         &test   ("edi",-1);
1085         &jz     (&label("badpointer"));
1086
1087         &call   (&label("pic_point"));
1088         &set_label("pic_point");
1089         &blindpop("ebp");
1090         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1091
1092         &mov    ("ecx",&wparam(1));             # number of bits in key
1093         &cmp    ("ecx",128);
1094         &je     (&label("10rounds"));
1095         &cmp    ("ecx",192);
1096         &je     (&label("12rounds"));
1097         &cmp    ("ecx",256);
1098         &je     (&label("14rounds"));
1099         &mov    ("eax",-2);                     # invalid number of bits
1100         &jmp    (&label("exit"));
1101
1102     &set_label("10rounds");
1103         &mov    ("eax",&DWP(0,"esi"));          # copy first 4 dwords
1104         &mov    ("ebx",&DWP(4,"esi"));
1105         &mov    ("ecx",&DWP(8,"esi"));
1106         &mov    ("edx",&DWP(12,"esi"));
1107         &mov    (&DWP(0,"edi"),"eax");
1108         &mov    (&DWP(4,"edi"),"ebx");
1109         &mov    (&DWP(8,"edi"),"ecx");
1110         &mov    (&DWP(12,"edi"),"edx");
1111
1112         &xor    ("ecx","ecx");
1113         &jmp    (&label("10shortcut"));
1114
1115         &align  (4);
1116         &set_label("10loop");
1117                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1118                 &mov    ("edx",&DWP(12,"edi"));         # rk[3]
1119         &set_label("10shortcut");
1120                 &enckey ();
1121
1122                 &mov    (&DWP(16,"edi"),"eax");         # rk[4]
1123                 &xor    ("eax",&DWP(4,"edi"));
1124                 &mov    (&DWP(20,"edi"),"eax");         # rk[5]
1125                 &xor    ("eax",&DWP(8,"edi"));
1126                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1127                 &xor    ("eax",&DWP(12,"edi"));
1128                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1129                 &inc    ("ecx");
1130                 &add    ("edi",16);
1131                 &cmp    ("ecx",10);
1132         &jl     (&label("10loop"));
1133
1134         &mov    (&DWP(80,"edi"),10);            # setup number of rounds
1135         &xor    ("eax","eax");
1136         &jmp    (&label("exit"));
1137
1138     &set_label("12rounds");
1139         &mov    ("eax",&DWP(0,"esi"));          # copy first 6 dwords
1140         &mov    ("ebx",&DWP(4,"esi"));
1141         &mov    ("ecx",&DWP(8,"esi"));
1142         &mov    ("edx",&DWP(12,"esi"));
1143         &mov    (&DWP(0,"edi"),"eax");
1144         &mov    (&DWP(4,"edi"),"ebx");
1145         &mov    (&DWP(8,"edi"),"ecx");
1146         &mov    (&DWP(12,"edi"),"edx");
1147         &mov    ("ecx",&DWP(16,"esi"));
1148         &mov    ("edx",&DWP(20,"esi"));
1149         &mov    (&DWP(16,"edi"),"ecx");
1150         &mov    (&DWP(20,"edi"),"edx");
1151
1152         &xor    ("ecx","ecx");
1153         &jmp    (&label("12shortcut"));
1154
1155         &align  (4);
1156         &set_label("12loop");
1157                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1158                 &mov    ("edx",&DWP(20,"edi"));         # rk[5]
1159         &set_label("12shortcut");
1160                 &enckey ();
1161
1162                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1163                 &xor    ("eax",&DWP(4,"edi"));
1164                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1165                 &xor    ("eax",&DWP(8,"edi"));
1166                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1167                 &xor    ("eax",&DWP(12,"edi"));
1168                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1169
1170                 &cmp    ("ecx",7);
1171                 &je     (&label("12break"));
1172                 &inc    ("ecx");
1173
1174                 &xor    ("eax",&DWP(16,"edi"));
1175                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1176                 &xor    ("eax",&DWP(20,"edi"));
1177                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1178
1179                 &add    ("edi",24);
1180         &jmp    (&label("12loop"));
1181
1182         &set_label("12break");
1183         &mov    (&DWP(72,"edi"),12);            # setup number of rounds
1184         &xor    ("eax","eax");
1185         &jmp    (&label("exit"));
1186
1187     &set_label("14rounds");
1188         &mov    ("eax",&DWP(0,"esi"));          # copy first 8 dwords
1189         &mov    ("ebx",&DWP(4,"esi"));
1190         &mov    ("ecx",&DWP(8,"esi"));
1191         &mov    ("edx",&DWP(12,"esi"));
1192         &mov    (&DWP(0,"edi"),"eax");
1193         &mov    (&DWP(4,"edi"),"ebx");
1194         &mov    (&DWP(8,"edi"),"ecx");
1195         &mov    (&DWP(12,"edi"),"edx");
1196         &mov    ("eax",&DWP(16,"esi"));
1197         &mov    ("ebx",&DWP(20,"esi"));
1198         &mov    ("ecx",&DWP(24,"esi"));
1199         &mov    ("edx",&DWP(28,"esi"));
1200         &mov    (&DWP(16,"edi"),"eax");
1201         &mov    (&DWP(20,"edi"),"ebx");
1202         &mov    (&DWP(24,"edi"),"ecx");
1203         &mov    (&DWP(28,"edi"),"edx");
1204
1205         &xor    ("ecx","ecx");
1206         &jmp    (&label("14shortcut"));
1207
1208         &align  (4);
1209         &set_label("14loop");
1210                 &mov    ("edx",&DWP(28,"edi"));         # rk[7]
1211         &set_label("14shortcut");
1212                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1213
1214                 &enckey ();
1215
1216                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1217                 &xor    ("eax",&DWP(4,"edi"));
1218                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1219                 &xor    ("eax",&DWP(8,"edi"));
1220                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1221                 &xor    ("eax",&DWP(12,"edi"));
1222                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1223
1224                 &cmp    ("ecx",6);
1225                 &je     (&label("14break"));
1226                 &inc    ("ecx");
1227
1228                 &mov    ("edx","eax");
1229                 &mov    ("eax",&DWP(16,"edi"));         # rk[4]
1230                 &movz   ("esi",&LB("edx"));             # rk[11]>>0
1231                 &mov    ("ebx",&DWP(2,"ebp","esi",8));
1232                 &movz   ("esi",&HB("edx"));             # rk[11]>>8
1233                 &and    ("ebx",0x000000FF);
1234                 &xor    ("eax","ebx");
1235
1236                 &mov    ("ebx",&DWP(0,"ebp","esi",8));
1237                 &shr    ("edx",16);
1238                 &and    ("ebx",0x0000FF00);
1239                 &movz   ("esi",&LB("edx"));             # rk[11]>>16
1240                 &xor    ("eax","ebx");
1241
1242                 &mov    ("ebx",&DWP(0,"ebp","esi",8));
1243                 &movz   ("esi",&HB("edx"));             # rk[11]>>24
1244                 &and    ("ebx",0x00FF0000);
1245                 &xor    ("eax","ebx");
1246
1247                 &mov    ("ebx",&DWP(2,"ebp","esi",8));
1248                 &and    ("ebx",0xFF000000);
1249                 &xor    ("eax","ebx");
1250
1251                 &mov    (&DWP(48,"edi"),"eax");         # rk[12]
1252                 &xor    ("eax",&DWP(20,"edi"));
1253                 &mov    (&DWP(52,"edi"),"eax");         # rk[13]
1254                 &xor    ("eax",&DWP(24,"edi"));
1255                 &mov    (&DWP(56,"edi"),"eax");         # rk[14]
1256                 &xor    ("eax",&DWP(28,"edi"));
1257                 &mov    (&DWP(60,"edi"),"eax");         # rk[15]
1258
1259                 &add    ("edi",32);
1260         &jmp    (&label("14loop"));
1261
1262         &set_label("14break");
1263         &mov    (&DWP(48,"edi"),14);            # setup number of rounds
1264         &xor    ("eax","eax");
1265         &jmp    (&label("exit"));
1266
1267     &set_label("badpointer");
1268         &mov    ("eax",-1);
1269     &set_label("exit");
1270 &function_end("AES_set_encrypt_key");
1271
1272 sub deckey()
1273 { my ($i,$ptr,$te,$td) = @_;
1274
1275         &mov    ("eax",&DWP($i,$ptr));
1276         &mov    ("edx","eax");
1277         &movz   ("ebx",&HB("eax"));
1278         &shr    ("edx",16);
1279         &and    ("eax",0xFF);
1280         &movz   ("eax",&BP(2,$te,"eax",8));
1281         &movz   ("ebx",&BP(2,$te,"ebx",8));
1282         &mov    ("eax",&DWP(0,$td,"eax",8));
1283         &xor    ("eax",&DWP(3,$td,"ebx",8));
1284         &movz   ("ebx",&HB("edx"));
1285         &and    ("edx",0xFF);
1286         &movz   ("edx",&BP(2,$te,"edx",8));
1287         &movz   ("ebx",&BP(2,$te,"ebx",8));
1288         &xor    ("eax",&DWP(2,$td,"edx",8));
1289         &xor    ("eax",&DWP(1,$td,"ebx",8));
1290         &mov    (&DWP($i,$ptr),"eax");
1291 }
1292
1293 # int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
1294 #                        AES_KEY *key)
1295 &public_label("AES_Td");
1296 &public_label("AES_Te");
1297 &function_begin_B("AES_set_decrypt_key");
1298         &mov    ("eax",&wparam(0));
1299         &mov    ("ecx",&wparam(1));
1300         &mov    ("edx",&wparam(2));
1301         &sub    ("esp",12);
1302         &mov    (&DWP(0,"esp"),"eax");
1303         &mov    (&DWP(4,"esp"),"ecx");
1304         &mov    (&DWP(8,"esp"),"edx");
1305         &call   ("AES_set_encrypt_key");
1306         &add    ("esp",12);
1307         &cmp    ("eax",0);
1308         &je     (&label("proceed"));
1309         &ret    ();
1310
1311     &set_label("proceed");
1312         &push   ("ebp");
1313         &push   ("ebx");
1314         &push   ("esi");
1315         &push   ("edi");
1316
1317         &mov    ("esi",&wparam(2));
1318         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1319         &lea    ("ecx",&DWP(0,"","ecx",4));
1320         &lea    ("edi",&DWP(0,"esi","ecx",4));  # pointer to last chunk
1321
1322         &align  (4);
1323         &set_label("invert");                   # invert order of chunks
1324                 &mov    ("eax",&DWP(0,"esi"));
1325                 &mov    ("ebx",&DWP(4,"esi"));
1326                 &mov    ("ecx",&DWP(0,"edi"));
1327                 &mov    ("edx",&DWP(4,"edi"));
1328                 &mov    (&DWP(0,"edi"),"eax");
1329                 &mov    (&DWP(4,"edi"),"ebx");
1330                 &mov    (&DWP(0,"esi"),"ecx");
1331                 &mov    (&DWP(4,"esi"),"edx");
1332                 &mov    ("eax",&DWP(8,"esi"));
1333                 &mov    ("ebx",&DWP(12,"esi"));
1334                 &mov    ("ecx",&DWP(8,"edi"));
1335                 &mov    ("edx",&DWP(12,"edi"));
1336                 &mov    (&DWP(8,"edi"),"eax");
1337                 &mov    (&DWP(12,"edi"),"ebx");
1338                 &mov    (&DWP(8,"esi"),"ecx");
1339                 &mov    (&DWP(12,"esi"),"edx");
1340                 &add    ("esi",16);
1341                 &sub    ("edi",16);
1342                 &cmp    ("esi","edi");
1343         &jne    (&label("invert"));
1344
1345         &call   (&label("pic_point"));
1346         &set_label("pic_point");
1347         blindpop("ebp");
1348         &lea    ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
1349         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1350
1351         &mov    ("esi",&wparam(2));
1352         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1353         &dec    ("ecx");
1354         &align  (4);
1355         &set_label("permute");                  # permute the key schedule
1356                 &add    ("esi",16);
1357                 &deckey (0,"esi","ebp","edi");
1358                 &deckey (4,"esi","ebp","edi");
1359                 &deckey (8,"esi","ebp","edi");
1360                 &deckey (12,"esi","ebp","edi");
1361                 &dec    ("ecx");
1362         &jnz    (&label("permute"));
1363
1364         &xor    ("eax","eax");                  # return success
1365 &function_end("AES_set_decrypt_key");
1366
1367 &asm_finish();