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 3.5.
  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 # Version 3.3 avoids L1 cache aliasing between stack frame and
  64 # S-boxes, and 3.4 - L1 cache aliasing even between key schedule. The
  65 # latter is achieved by copying the key schedule to controlled place in
  66 # stack. This unfortunately has rather strong impact on small block CBC
  67 # performance, ~2x deterioration on 16-byte block if compared to 3.3.
  68 #
  69 # Version 3.5 checks if there L1 cache aliasing between user-supplied
  70 # key schedule and S-boxes and abstains from copying the former if
  71 # there no. This allows end-user to consciously retain small block
  72 # performance by aligning key schedule in specific manner.
  73 #
  74 # Current ECB performance numbers for 128-bit key in CPU cycles per
  75 # processed byte [measure commonly used by AES benchmarkers] are:
  76 #
  77 #               small footprint         fully unrolled
  78 # P4            24                      22
  79 # AMD K8        20                      19
  80 # PIII          25                      23
  81 # Pentium       81                      78
  82
  83 push(@INC,"perlasm","../../perlasm");
  84 require "x86asm.pl";
  85
  86 &asm_init($ARGV[0],"aes-586.pl",$ARGV[$#ARGV] eq "386");
  87
  88 $s0="eax";
  89 $s1="ebx";
  90 $s2="ecx";
  91 $s3="edx";
  92 $key="edi";
  93 $acc="esi";
  94
  95 $compromise=0;          # $compromise=128 abstains from copying key
  96                         # schedule to stack when encrypting inputs
  97                         # shorter than 128 bytes at the cost of
  98                         # risksing aliasing with S-boxes. In return
  99                         # you get way better, up to +70%, small block
 100                         # performance.
 101 $small_footprint=1;     # $small_footprint=1 code is ~5% slower [on
 102                         # recent µ-archs], but ~5 times smaller!
 103                         # I favor compact code to minimize cache
 104                         # contention and in hope to "collect" 5% back
 105                         # in real-life applications...
 106 $vertical_spin=0;       # shift "verticaly" defaults to 0, because of
 107                         # its proof-of-concept status...
 108
 109 # Note that there is no decvert(), as well as last encryption round is
 110 # performed with "horizontal" shifts. This is because this "vertical"
 111 # implementation [one which groups shifts on a given $s[i] to form a
 112 # "column," unlike "horizontal" one, which groups shifts on different
 113 # $s[i] to form a "row"] is work in progress. It was observed to run
 114 # few percents faster on Intel cores, but not AMD. On AMD K8 core it's
 115 # whole 12% slower:-( So we face a trade-off... Shall it be resolved
 116 # some day? Till then the code is considered experimental and by
 117 # default remains dormant...
 118
 119 sub encvert()
 120 { my ($te,@s) = @_;
 121   my $v0 = $acc, $v1 = $key;
 122
 123         &mov    ($v0,$s[3]);                            # copy s3
 124         &mov    (&DWP(4,"esp"),$s[2]);                  # save s2
 125         &mov    ($v1,$s[0]);                            # copy s0
 126         &mov    (&DWP(8,"esp"),$s[1]);                  # save s1
 127
 128         &movz   ($s[2],&HB($s[0]));
 129         &and    ($s[0],0xFF);
 130         &mov    ($s[0],&DWP(0,$te,$s[0],8));            # s0>>0
 131         &shr    ($v1,16);
 132         &mov    ($s[3],&DWP(3,$te,$s[2],8));            # s0>>8
 133         &movz   ($s[1],&HB($v1));
 134         &and    ($v1,0xFF);
 135         &mov    ($s[2],&DWP(2,$te,$v1,8));              # s0>>16
 136          &mov   ($v1,$v0);
 137         &mov    ($s[1],&DWP(1,$te,$s[1],8));            # s0>>24
 138
 139         &and    ($v0,0xFF);
 140         &xor    ($s[3],&DWP(0,$te,$v0,8));              # s3>>0
 141         &movz   ($v0,&HB($v1));
 142         &shr    ($v1,16);
 143         &xor    ($s[2],&DWP(3,$te,$v0,8));              # s3>>8
 144         &movz   ($v0,&HB($v1));
 145         &and    ($v1,0xFF);
 146         &xor    ($s[1],&DWP(2,$te,$v1,8));              # s3>>16
 147          &mov   ($v1,&DWP(4,"esp"));                    # restore s2
 148         &xor    ($s[0],&DWP(1,$te,$v0,8));              # s3>>24
 149
 150         &mov    ($v0,$v1);
 151         &and    ($v1,0xFF);
 152         &xor    ($s[2],&DWP(0,$te,$v1,8));              # s2>>0
 153         &movz   ($v1,&HB($v0));
 154         &shr    ($v0,16);
 155         &xor    ($s[1],&DWP(3,$te,$v1,8));              # s2>>8
 156         &movz   ($v1,&HB($v0));
 157         &and    ($v0,0xFF);
 158         &xor    ($s[0],&DWP(2,$te,$v0,8));              # s2>>16
 159          &mov   ($v0,&DWP(8,"esp"));                    # restore s1
 160         &xor    ($s[3],&DWP(1,$te,$v1,8));              # s2>>24
 161
 162         &mov    ($v1,$v0);
 163         &and    ($v0,0xFF);
 164         &xor    ($s[1],&DWP(0,$te,$v0,8));              # s1>>0
 165         &movz   ($v0,&HB($v1));
 166         &shr    ($v1,16);
 167         &xor    ($s[0],&DWP(3,$te,$v0,8));              # s1>>8
 168         &movz   ($v0,&HB($v1));
 169         &and    ($v1,0xFF);
 170         &xor    ($s[3],&DWP(2,$te,$v1,8));              # s1>>16
 171          &mov   ($key,&DWP(12,"esp"));                  # reincarnate v1 as key
 172         &xor    ($s[2],&DWP(1,$te,$v0,8));              # s1>>24
 173 }
 174
 175 sub encstep()
 176 { my ($i,$te,@s) = @_;
 177   my $tmp = $key;
 178   my $out = $i==3?$s[0]:$acc;
 179
 180         # lines marked with #%e?x[i] denote "reordered" instructions...
 181         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
 182         else        {   &mov    ($out,$s[0]);
 183                         &and    ($out,0xFF);                    }
 184         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
 185         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
 186                         &mov    ($out,&DWP(0,$te,$out,8));
 187
 188         if ($i==3)  {   $tmp=$s[1];                             }##%eax
 189                         &movz   ($tmp,&HB($s[1]));
 190                         &xor    ($out,&DWP(3,$te,$tmp,8));
 191
 192         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx
 193         else        {   &mov    ($tmp,$s[2]);
 194                         &shr    ($tmp,16);                      }
 195         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
 196                         &and    ($tmp,0xFF);
 197                         &xor    ($out,&DWP(2,$te,$tmp,8));
 198
 199         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx
 200         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
 201         else        {   &mov    ($tmp,$s[3]);
 202                         &shr    ($tmp,24)                       }
 203                         &xor    ($out,&DWP(1,$te,$tmp,8));
 204         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
 205         if ($i==3)  {   &mov    ($s[3],$acc);                   }
 206                         &comment();
 207 }
 208
 209 sub enclast()
 210 { my ($i,$te,@s)=@_;
 211   my $tmp = $key;
 212   my $out = $i==3?$s[0]:$acc;
 213
 214         if ($i==3)  {   &mov    ($key,&DWP(12,"esp"));          }##%edx
 215         else        {   &mov    ($out,$s[0]);                   }
 216                         &and    ($out,0xFF);
 217         if ($i==1)  {   &shr    ($s[0],16);                     }#%ebx[1]
 218         if ($i==2)  {   &shr    ($s[0],24);                     }#%ecx[2]
 219                         &mov    ($out,&DWP(2,$te,$out,8));
 220                         &and    ($out,0x000000ff);
 221
 222         if ($i==3)  {   $tmp=$s[1];                             }##%eax
 223                         &movz   ($tmp,&HB($s[1]));
 224                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
 225                         &and    ($tmp,0x0000ff00);
 226                         &xor    ($out,$tmp);
 227
 228         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],&DWP(4,"esp")); }##%ebx
 229         else        {   mov     ($tmp,$s[2]);
 230                         &shr    ($tmp,16);                      }
 231         if ($i==2)  {   &and    ($s[1],0xFF);                   }#%edx[2]
 232                         &and    ($tmp,0xFF);
 233                         &mov    ($tmp,&DWP(0,$te,$tmp,8));
 234                         &and    ($tmp,0x00ff0000);
 235                         &xor    ($out,$tmp);
 236
 237         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }##%ecx
 238         elsif($i==2){   &movz   ($tmp,&HB($s[3]));              }#%ebx[2]
 239         else        {   &mov    ($tmp,$s[3]);
 240                         &shr    ($tmp,24);                      }
 241                         &mov    ($tmp,&DWP(2,$te,$tmp,8));
 242                         &and    ($tmp,0xff000000);
 243                         &xor    ($out,$tmp);
 244         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
 245         if ($i==3)  {   &mov    ($s[3],$acc);                   }
 246 }
 247
 248 sub _data_word() { my $i; while(defined($i=shift)) { &data_word($i,$i); } }
 249
 250 &public_label("AES_Te");
 251 &function_begin_B("_x86_AES_encrypt");
 252         if ($vertical_spin) {
 253                 # I need high parts of volatile registers to be accessible...
 254                 &exch   ($s1="edi",$key="ebx");
 255                 &mov    ($s2="esi",$acc="ecx");
 256         }
 257
 258         # note that caller is expected to allocate stack frame for me!
 259         &mov    (&DWP(12,"esp"),$key);          # save key
 260
 261         &xor    ($s0,&DWP(0,$key));             # xor with key
 262         &xor    ($s1,&DWP(4,$key));
 263         &xor    ($s2,&DWP(8,$key));
 264         &xor    ($s3,&DWP(12,$key));
 265
 266         &mov    ($acc,&DWP(240,$key));          # load key->rounds
 267
 268         if ($small_footprint) {
 269             &lea        ($acc,&DWP(-2,$acc,$acc));
 270             &lea        ($acc,&DWP(0,$key,$acc,8));
 271             &mov        (&DWP(16,"esp"),$acc);  # end of key schedule
 272             &align      (4);
 273             &set_label("loop");
 274                 if ($vertical_spin) {
 275                     &encvert("ebp",$s0,$s1,$s2,$s3);
 276                 } else {
 277                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 278                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 279                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 280                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 281                 }
 282                 &add    ($key,16);              # advance rd_key
 283                 &xor    ($s0,&DWP(0,$key));
 284                 &xor    ($s1,&DWP(4,$key));
 285                 &xor    ($s2,&DWP(8,$key));
 286                 &xor    ($s3,&DWP(12,$key));
 287             &cmp        ($key,&DWP(16,"esp"));
 288             &mov        (&DWP(12,"esp"),$key);
 289             &jb         (&label("loop"));
 290         }
 291         else {
 292             &cmp        ($acc,10);
 293             &jle        (&label("10rounds"));
 294             &cmp        ($acc,12);
 295             &jle        (&label("12rounds"));
 296
 297         &set_label("14rounds");
 298             for ($i=1;$i<3;$i++) {
 299                 if ($vertical_spin) {
 300                     &encvert("ebp",$s0,$s1,$s2,$s3);
 301                 } else {
 302                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 303                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 304                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 305                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 306                 }
 307                 &xor    ($s0,&DWP(16*$i+0,$key));
 308                 &xor    ($s1,&DWP(16*$i+4,$key));
 309                 &xor    ($s2,&DWP(16*$i+8,$key));
 310                 &xor    ($s3,&DWP(16*$i+12,$key));
 311             }
 312             &add        ($key,32);
 313             &mov        (&DWP(12,"esp"),$key);  # advance rd_key
 314         &set_label("12rounds");
 315             for ($i=1;$i<3;$i++) {
 316                 if ($vertical_spin) {
 317                     &encvert("ebp",$s0,$s1,$s2,$s3);
 318                 } else {
 319                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 320                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 321                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 322                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 323                 }
 324                 &xor    ($s0,&DWP(16*$i+0,$key));
 325                 &xor    ($s1,&DWP(16*$i+4,$key));
 326                 &xor    ($s2,&DWP(16*$i+8,$key));
 327                 &xor    ($s3,&DWP(16*$i+12,$key));
 328             }
 329             &add        ($key,32);
 330             &mov        (&DWP(12,"esp"),$key);  # advance rd_key
 331         &set_label("10rounds");
 332             for ($i=1;$i<10;$i++) {
 333                 if ($vertical_spin) {
 334                     &encvert("ebp",$s0,$s1,$s2,$s3);
 335                 } else {
 336                     &encstep(0,"ebp",$s0,$s1,$s2,$s3);
 337                     &encstep(1,"ebp",$s1,$s2,$s3,$s0);
 338                     &encstep(2,"ebp",$s2,$s3,$s0,$s1);
 339                     &encstep(3,"ebp",$s3,$s0,$s1,$s2);
 340                 }
 341                 &xor    ($s0,&DWP(16*$i+0,$key));
 342                 &xor    ($s1,&DWP(16*$i+4,$key));
 343                 &xor    ($s2,&DWP(16*$i+8,$key));
 344                 &xor    ($s3,&DWP(16*$i+12,$key));
 345             }
 346         }
 347
 348         if ($vertical_spin) {
 349             # "reincarnate" some registers for "horizontal" spin...
 350             &mov        ($s1="ebx",$key="edi");
 351             &mov        ($s2="ecx",$acc="esi");
 352         }
 353         &enclast(0,"ebp",$s0,$s1,$s2,$s3);
 354         &enclast(1,"ebp",$s1,$s2,$s3,$s0);
 355         &enclast(2,"ebp",$s2,$s3,$s0,$s1);
 356         &enclast(3,"ebp",$s3,$s0,$s1,$s2);
 357
 358         &add    ($key,$small_footprint?16:160);
 359         &xor    ($s0,&DWP(0,$key));
 360         &xor    ($s1,&DWP(4,$key));
 361         &xor    ($s2,&DWP(8,$key));
 362         &xor    ($s3,&DWP(12,$key));
 363
 364         &ret    ();
 365
 366 &set_label("AES_Te",64);        # Yes! I keep it in the code segment!
 367         &_data_word(0xa56363c6, 0x847c7cf8, 0x997777ee, 0x8d7b7bf6);
 368         &_data_word(0x0df2f2ff, 0xbd6b6bd6, 0xb16f6fde, 0x54c5c591);
 369         &_data_word(0x50303060, 0x03010102, 0xa96767ce, 0x7d2b2b56);
 370         &_data_word(0x19fefee7, 0x62d7d7b5, 0xe6abab4d, 0x9a7676ec);
 371         &_data_word(0x45caca8f, 0x9d82821f, 0x40c9c989, 0x877d7dfa);
 372         &_data_word(0x15fafaef, 0xeb5959b2, 0xc947478e, 0x0bf0f0fb);
 373         &_data_word(0xecadad41, 0x67d4d4b3, 0xfda2a25f, 0xeaafaf45);
 374         &_data_word(0xbf9c9c23, 0xf7a4a453, 0x967272e4, 0x5bc0c09b);
 375         &_data_word(0xc2b7b775, 0x1cfdfde1, 0xae93933d, 0x6a26264c);
 376         &_data_word(0x5a36366c, 0x413f3f7e, 0x02f7f7f5, 0x4fcccc83);
 377         &_data_word(0x5c343468, 0xf4a5a551, 0x34e5e5d1, 0x08f1f1f9);
 378         &_data_word(0x937171e2, 0x73d8d8ab, 0x53313162, 0x3f15152a);
 379         &_data_word(0x0c040408, 0x52c7c795, 0x65232346, 0x5ec3c39d);
 380         &_data_word(0x28181830, 0xa1969637, 0x0f05050a, 0xb59a9a2f);
 381         &_data_word(0x0907070e, 0x36121224, 0x9b80801b, 0x3de2e2df);
 382         &_data_word(0x26ebebcd, 0x6927274e, 0xcdb2b27f, 0x9f7575ea);
 383         &_data_word(0x1b090912, 0x9e83831d, 0x742c2c58, 0x2e1a1a34);
 384         &_data_word(0x2d1b1b36, 0xb26e6edc, 0xee5a5ab4, 0xfba0a05b);
 385         &_data_word(0xf65252a4, 0x4d3b3b76, 0x61d6d6b7, 0xceb3b37d);
 386         &_data_word(0x7b292952, 0x3ee3e3dd, 0x712f2f5e, 0x97848413);
 387         &_data_word(0xf55353a6, 0x68d1d1b9, 0x00000000, 0x2cededc1);
 388         &_data_word(0x60202040, 0x1ffcfce3, 0xc8b1b179, 0xed5b5bb6);
 389         &_data_word(0xbe6a6ad4, 0x46cbcb8d, 0xd9bebe67, 0x4b393972);
 390         &_data_word(0xde4a4a94, 0xd44c4c98, 0xe85858b0, 0x4acfcf85);
 391         &_data_word(0x6bd0d0bb, 0x2aefefc5, 0xe5aaaa4f, 0x16fbfbed);
 392         &_data_word(0xc5434386, 0xd74d4d9a, 0x55333366, 0x94858511);
 393         &_data_word(0xcf45458a, 0x10f9f9e9, 0x06020204, 0x817f7ffe);
 394         &_data_word(0xf05050a0, 0x443c3c78, 0xba9f9f25, 0xe3a8a84b);
 395         &_data_word(0xf35151a2, 0xfea3a35d, 0xc0404080, 0x8a8f8f05);
 396         &_data_word(0xad92923f, 0xbc9d9d21, 0x48383870, 0x04f5f5f1);
 397         &_data_word(0xdfbcbc63, 0xc1b6b677, 0x75dadaaf, 0x63212142);
 398         &_data_word(0x30101020, 0x1affffe5, 0x0ef3f3fd, 0x6dd2d2bf);
 399         &_data_word(0x4ccdcd81, 0x140c0c18, 0x35131326, 0x2fececc3);
 400         &_data_word(0xe15f5fbe, 0xa2979735, 0xcc444488, 0x3917172e);
 401         &_data_word(0x57c4c493, 0xf2a7a755, 0x827e7efc, 0x473d3d7a);
 402         &_data_word(0xac6464c8, 0xe75d5dba, 0x2b191932, 0x957373e6);
 403         &_data_word(0xa06060c0, 0x98818119, 0xd14f4f9e, 0x7fdcdca3);
 404         &_data_word(0x66222244, 0x7e2a2a54, 0xab90903b, 0x8388880b);
 405         &_data_word(0xca46468c, 0x29eeeec7, 0xd3b8b86b, 0x3c141428);
 406         &_data_word(0x79dedea7, 0xe25e5ebc, 0x1d0b0b16, 0x76dbdbad);
 407         &_data_word(0x3be0e0db, 0x56323264, 0x4e3a3a74, 0x1e0a0a14);
 408         &_data_word(0xdb494992, 0x0a06060c, 0x6c242448, 0xe45c5cb8);
 409         &_data_word(0x5dc2c29f, 0x6ed3d3bd, 0xefacac43, 0xa66262c4);
 410         &_data_word(0xa8919139, 0xa4959531, 0x37e4e4d3, 0x8b7979f2);
 411         &_data_word(0x32e7e7d5, 0x43c8c88b, 0x5937376e, 0xb76d6dda);
 412         &_data_word(0x8c8d8d01, 0x64d5d5b1, 0xd24e4e9c, 0xe0a9a949);
 413         &_data_word(0xb46c6cd8, 0xfa5656ac, 0x07f4f4f3, 0x25eaeacf);
 414         &_data_word(0xaf6565ca, 0x8e7a7af4, 0xe9aeae47, 0x18080810);
 415         &_data_word(0xd5baba6f, 0x887878f0, 0x6f25254a, 0x722e2e5c);
 416         &_data_word(0x241c1c38, 0xf1a6a657, 0xc7b4b473, 0x51c6c697);
 417         &_data_word(0x23e8e8cb, 0x7cdddda1, 0x9c7474e8, 0x211f1f3e);
 418         &_data_word(0xdd4b4b96, 0xdcbdbd61, 0x868b8b0d, 0x858a8a0f);
 419         &_data_word(0x907070e0, 0x423e3e7c, 0xc4b5b571, 0xaa6666cc);
 420         &_data_word(0xd8484890, 0x05030306, 0x01f6f6f7, 0x120e0e1c);
 421         &_data_word(0xa36161c2, 0x5f35356a, 0xf95757ae, 0xd0b9b969);
 422         &_data_word(0x91868617, 0x58c1c199, 0x271d1d3a, 0xb99e9e27);
 423         &_data_word(0x38e1e1d9, 0x13f8f8eb, 0xb398982b, 0x33111122);
 424         &_data_word(0xbb6969d2, 0x70d9d9a9, 0x898e8e07, 0xa7949433);
 425         &_data_word(0xb69b9b2d, 0x221e1e3c, 0x92878715, 0x20e9e9c9);
 426         &_data_word(0x49cece87, 0xff5555aa, 0x78282850, 0x7adfdfa5);
 427         &_data_word(0x8f8c8c03, 0xf8a1a159, 0x80898909, 0x170d0d1a);
 428         &_data_word(0xdabfbf65, 0x31e6e6d7, 0xc6424284, 0xb86868d0);
 429         &_data_word(0xc3414182, 0xb0999929, 0x772d2d5a, 0x110f0f1e);
 430         &_data_word(0xcbb0b07b, 0xfc5454a8, 0xd6bbbb6d, 0x3a16162c);
 431 #rcon:
 432         &data_word(0x00000001, 0x00000002, 0x00000004, 0x00000008);
 433         &data_word(0x00000010, 0x00000020, 0x00000040, 0x00000080);
 434         &data_word(0x0000001b, 0x00000036, 0, 0, 0, 0, 0, 0);
 435 &function_end_B("_x86_AES_encrypt");
 436
 437 # void AES_encrypt (const void *inp,void *out,const AES_KEY *key);
 438 &public_label("AES_Te");
 439 &function_begin("AES_encrypt");
 440         &mov    ($acc,&wparam(0));              # load inp
 441         &mov    ($key,&wparam(2));              # load key
 442
 443         &mov    ($s0,"esp");
 444         &sub    ("esp",24);
 445         &and    ("esp",-64);
 446         &add    ("esp",4);
 447         &mov    (&DWP(16,"esp"),$s0);
 448
 449         &call   (&label("pic_point"));          # make it PIC!
 450         &set_label("pic_point");
 451         &blindpop("ebp");
 452         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
 453
 454         &mov    ($s0,&DWP(0,$acc));             # load input data
 455         &mov    ($s1,&DWP(4,$acc));
 456         &mov    ($s2,&DWP(8,$acc));
 457         &mov    ($s3,&DWP(12,$acc));
 458
 459         &call   ("_x86_AES_encrypt");
 460
 461         &mov    ("esp",&DWP(16,"esp"));
 462
 463         &mov    ($acc,&wparam(1));              # load out
 464         &mov    (&DWP(0,$acc),$s0);             # write output data
 465         &mov    (&DWP(4,$acc),$s1);
 466         &mov    (&DWP(8,$acc),$s2);
 467         &mov    (&DWP(12,$acc),$s3);
 468 &function_end("AES_encrypt");
 469
 470 #------------------------------------------------------------------#
 471
 472 sub decstep()
 473 { my ($i,$td,@s) = @_;
 474   my $tmp = $key;
 475   my $out = $i==3?$s[0]:$acc;
 476
 477         # no instructions are reordered, as performance appears
 478         # optimal... or rather that all attempts to reorder didn't
 479         # result in better performance [which by the way is not a
 480         # bit lower than ecryption].
 481         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
 482         else        {   &mov    ($out,$s[0]);                   }
 483                         &and    ($out,0xFF);
 484                         &mov    ($out,&DWP(0,$td,$out,8));
 485
 486         if ($i==3)  {   $tmp=$s[1];                             }
 487                         &movz   ($tmp,&HB($s[1]));
 488                         &xor    ($out,&DWP(3,$td,$tmp,8));
 489
 490         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
 491         else        {   &mov    ($tmp,$s[2]);                   }
 492                         &shr    ($tmp,16);
 493                         &and    ($tmp,0xFF);
 494                         &xor    ($out,&DWP(2,$td,$tmp,8));
 495
 496         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }
 497         else        {   &mov    ($tmp,$s[3]);                   }
 498                         &shr    ($tmp,24);
 499                         &xor    ($out,&DWP(1,$td,$tmp,8));
 500         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
 501         if ($i==3)  {   &mov    ($s[3],&DWP(4,"esp"));          }
 502                         &comment();
 503 }
 504
 505 sub declast()
 506 { my ($i,$td,@s)=@_;
 507   my $tmp = $key;
 508   my $out = $i==3?$s[0]:$acc;
 509
 510         if($i==3)   {   &mov    ($key,&DWP(12,"esp"));          }
 511         else        {   &mov    ($out,$s[0]);                   }
 512                         &and    ($out,0xFF);
 513                         &mov    ($out,&DWP(2048,$td,$out,4));
 514                         &and    ($out,0x000000ff);
 515
 516         if ($i==3)  {   $tmp=$s[1];                             }
 517                         &movz   ($tmp,&HB($s[1]));
 518                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 519                         &and    ($tmp,0x0000ff00);
 520                         &xor    ($out,$tmp);
 521
 522         if ($i==3)  {   $tmp=$s[2]; &mov ($s[1],$acc);          }
 523         else        {   mov     ($tmp,$s[2]);                   }
 524                         &shr    ($tmp,16);
 525                         &and    ($tmp,0xFF);
 526                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 527                         &and    ($tmp,0x00ff0000);
 528                         &xor    ($out,$tmp);
 529
 530         if ($i==3)  {   $tmp=$s[3]; &mov ($s[2],&DWP(8,"esp")); }
 531         else        {   &mov    ($tmp,$s[3]);                   }
 532                         &shr    ($tmp,24);
 533                         &mov    ($tmp,&DWP(2048,$td,$tmp,4));
 534                         &and    ($tmp,0xff000000);
 535                         &xor    ($out,$tmp);
 536         if ($i<2)   {   &mov    (&DWP(4+4*$i,"esp"),$out);      }
 537         if ($i==3)  {   &mov    ($s[3],&DWP(4,"esp"));          }
 538 }
 539
 540 &public_label("AES_Td");
 541 &function_begin_B("_x86_AES_decrypt");
 542         # note that caller is expected to allocate stack frame for me!
 543         &mov    (&DWP(12,"esp"),$key);          # save key
 544
 545         &xor    ($s0,&DWP(0,$key));             # xor with key
 546         &xor    ($s1,&DWP(4,$key));
 547         &xor    ($s2,&DWP(8,$key));
 548         &xor    ($s3,&DWP(12,$key));
 549
 550         &mov    ($acc,&DWP(240,$key));          # load key->rounds
 551
 552         if ($small_footprint) {
 553             &lea        ($acc,&DWP(-2,$acc,$acc));
 554             &lea        ($acc,&DWP(0,$key,$acc,8));
 555             &mov        (&DWP(16,"esp"),$acc);  # end of key schedule
 556             &align      (4);
 557             &set_label("loop");
 558                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 559                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 560                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 561                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 562                 &add    ($key,16);              # advance rd_key
 563                 &xor    ($s0,&DWP(0,$key));
 564                 &xor    ($s1,&DWP(4,$key));
 565                 &xor    ($s2,&DWP(8,$key));
 566                 &xor    ($s3,&DWP(12,$key));
 567             &cmp        ($key,&DWP(16,"esp"));
 568             &mov        (&DWP(12,"esp"),$key);
 569             &jb         (&label("loop"));
 570         }
 571         else {
 572             &cmp        ($acc,10);
 573             &jle        (&label("10rounds"));
 574             &cmp        ($acc,12);
 575             &jle        (&label("12rounds"));
 576
 577         &set_label("14rounds");
 578             for ($i=1;$i<3;$i++) {
 579                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 580                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 581                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 582                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 583                 &xor    ($s0,&DWP(16*$i+0,$key));
 584                 &xor    ($s1,&DWP(16*$i+4,$key));
 585                 &xor    ($s2,&DWP(16*$i+8,$key));
 586                 &xor    ($s3,&DWP(16*$i+12,$key));
 587             }
 588             &add        ($key,32);
 589             &mov        (&DWP(12,"esp"),$key);  # advance rd_key
 590         &set_label("12rounds");
 591             for ($i=1;$i<3;$i++) {
 592                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 593                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 594                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 595                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 596                 &xor    ($s0,&DWP(16*$i+0,$key));
 597                 &xor    ($s1,&DWP(16*$i+4,$key));
 598                 &xor    ($s2,&DWP(16*$i+8,$key));
 599                 &xor    ($s3,&DWP(16*$i+12,$key));
 600             }
 601             &add        ($key,32);
 602             &mov        (&DWP(12,"esp"),$key);  # advance rd_key
 603         &set_label("10rounds");
 604             for ($i=1;$i<10;$i++) {
 605                 &decstep(0,"ebp",$s0,$s3,$s2,$s1);
 606                 &decstep(1,"ebp",$s1,$s0,$s3,$s2);
 607                 &decstep(2,"ebp",$s2,$s1,$s0,$s3);
 608                 &decstep(3,"ebp",$s3,$s2,$s1,$s0);
 609                 &xor    ($s0,&DWP(16*$i+0,$key));
 610                 &xor    ($s1,&DWP(16*$i+4,$key));
 611                 &xor    ($s2,&DWP(16*$i+8,$key));
 612                 &xor    ($s3,&DWP(16*$i+12,$key));
 613             }
 614         }
 615
 616         &declast(0,"ebp",$s0,$s3,$s2,$s1);
 617         &declast(1,"ebp",$s1,$s0,$s3,$s2);
 618         &declast(2,"ebp",$s2,$s1,$s0,$s3);
 619         &declast(3,"ebp",$s3,$s2,$s1,$s0);
 620
 621         &add    ($key,$small_footprint?16:160);
 622         &xor    ($s0,&DWP(0,$key));
 623         &xor    ($s1,&DWP(4,$key));
 624         &xor    ($s2,&DWP(8,$key));
 625         &xor    ($s3,&DWP(12,$key));
 626
 627         &ret    ();
 628
 629 &set_label("AES_Td",64);        # Yes! I keep it in the code segment!
 630         &_data_word(0x50a7f451, 0x5365417e, 0xc3a4171a, 0x965e273a);
 631         &_data_word(0xcb6bab3b, 0xf1459d1f, 0xab58faac, 0x9303e34b);
 632         &_data_word(0x55fa3020, 0xf66d76ad, 0x9176cc88, 0x254c02f5);
 633         &_data_word(0xfcd7e54f, 0xd7cb2ac5, 0x80443526, 0x8fa362b5);
 634         &_data_word(0x495ab1de, 0x671bba25, 0x980eea45, 0xe1c0fe5d);
 635         &_data_word(0x02752fc3, 0x12f04c81, 0xa397468d, 0xc6f9d36b);
 636         &_data_word(0xe75f8f03, 0x959c9215, 0xeb7a6dbf, 0xda595295);
 637         &_data_word(0x2d83bed4, 0xd3217458, 0x2969e049, 0x44c8c98e);
 638         &_data_word(0x6a89c275, 0x78798ef4, 0x6b3e5899, 0xdd71b927);
 639         &_data_word(0xb64fe1be, 0x17ad88f0, 0x66ac20c9, 0xb43ace7d);
 640         &_data_word(0x184adf63, 0x82311ae5, 0x60335197, 0x457f5362);
 641         &_data_word(0xe07764b1, 0x84ae6bbb, 0x1ca081fe, 0x942b08f9);
 642         &_data_word(0x58684870, 0x19fd458f, 0x876cde94, 0xb7f87b52);
 643         &_data_word(0x23d373ab, 0xe2024b72, 0x578f1fe3, 0x2aab5566);
 644         &_data_word(0x0728ebb2, 0x03c2b52f, 0x9a7bc586, 0xa50837d3);
 645         &_data_word(0xf2872830, 0xb2a5bf23, 0xba6a0302, 0x5c8216ed);
 646         &_data_word(0x2b1ccf8a, 0x92b479a7, 0xf0f207f3, 0xa1e2694e);
 647         &_data_word(0xcdf4da65, 0xd5be0506, 0x1f6234d1, 0x8afea6c4);
 648         &_data_word(0x9d532e34, 0xa055f3a2, 0x32e18a05, 0x75ebf6a4);
 649         &_data_word(0x39ec830b, 0xaaef6040, 0x069f715e, 0x51106ebd);
 650         &_data_word(0xf98a213e, 0x3d06dd96, 0xae053edd, 0x46bde64d);
 651         &_data_word(0xb58d5491, 0x055dc471, 0x6fd40604, 0xff155060);
 652         &_data_word(0x24fb9819, 0x97e9bdd6, 0xcc434089, 0x779ed967);
 653         &_data_word(0xbd42e8b0, 0x888b8907, 0x385b19e7, 0xdbeec879);
 654         &_data_word(0x470a7ca1, 0xe90f427c, 0xc91e84f8, 0x00000000);
 655         &_data_word(0x83868009, 0x48ed2b32, 0xac70111e, 0x4e725a6c);
 656         &_data_word(0xfbff0efd, 0x5638850f, 0x1ed5ae3d, 0x27392d36);
 657         &_data_word(0x64d90f0a, 0x21a65c68, 0xd1545b9b, 0x3a2e3624);
 658         &_data_word(0xb1670a0c, 0x0fe75793, 0xd296eeb4, 0x9e919b1b);
 659         &_data_word(0x4fc5c080, 0xa220dc61, 0x694b775a, 0x161a121c);
 660         &_data_word(0x0aba93e2, 0xe52aa0c0, 0x43e0223c, 0x1d171b12);
 661         &_data_word(0x0b0d090e, 0xadc78bf2, 0xb9a8b62d, 0xc8a91e14);
 662         &_data_word(0x8519f157, 0x4c0775af, 0xbbdd99ee, 0xfd607fa3);
 663         &_data_word(0x9f2601f7, 0xbcf5725c, 0xc53b6644, 0x347efb5b);
 664         &_data_word(0x7629438b, 0xdcc623cb, 0x68fcedb6, 0x63f1e4b8);
 665         &_data_word(0xcadc31d7, 0x10856342, 0x40229713, 0x2011c684);
 666         &_data_word(0x7d244a85, 0xf83dbbd2, 0x1132f9ae, 0x6da129c7);
 667         &_data_word(0x4b2f9e1d, 0xf330b2dc, 0xec52860d, 0xd0e3c177);
 668         &_data_word(0x6c16b32b, 0x99b970a9, 0xfa489411, 0x2264e947);
 669         &_data_word(0xc48cfca8, 0x1a3ff0a0, 0xd82c7d56, 0xef903322);
 670         &_data_word(0xc74e4987, 0xc1d138d9, 0xfea2ca8c, 0x360bd498);
 671         &_data_word(0xcf81f5a6, 0x28de7aa5, 0x268eb7da, 0xa4bfad3f);
 672         &_data_word(0xe49d3a2c, 0x0d927850, 0x9bcc5f6a, 0x62467e54);
 673         &_data_word(0xc2138df6, 0xe8b8d890, 0x5ef7392e, 0xf5afc382);
 674         &_data_word(0xbe805d9f, 0x7c93d069, 0xa92dd56f, 0xb31225cf);
 675         &_data_word(0x3b99acc8, 0xa77d1810, 0x6e639ce8, 0x7bbb3bdb);
 676         &_data_word(0x097826cd, 0xf418596e, 0x01b79aec, 0xa89a4f83);
 677         &_data_word(0x656e95e6, 0x7ee6ffaa, 0x08cfbc21, 0xe6e815ef);
 678         &_data_word(0xd99be7ba, 0xce366f4a, 0xd4099fea, 0xd67cb029);
 679         &_data_word(0xafb2a431, 0x31233f2a, 0x3094a5c6, 0xc066a235);
 680         &_data_word(0x37bc4e74, 0xa6ca82fc, 0xb0d090e0, 0x15d8a733);
 681         &_data_word(0x4a9804f1, 0xf7daec41, 0x0e50cd7f, 0x2ff69117);
 682         &_data_word(0x8dd64d76, 0x4db0ef43, 0x544daacc, 0xdf0496e4);
 683         &_data_word(0xe3b5d19e, 0x1b886a4c, 0xb81f2cc1, 0x7f516546);
 684         &_data_word(0x04ea5e9d, 0x5d358c01, 0x737487fa, 0x2e410bfb);
 685         &_data_word(0x5a1d67b3, 0x52d2db92, 0x335610e9, 0x1347d66d);
 686         &_data_word(0x8c61d79a, 0x7a0ca137, 0x8e14f859, 0x893c13eb);
 687         &_data_word(0xee27a9ce, 0x35c961b7, 0xede51ce1, 0x3cb1477a);
 688         &_data_word(0x59dfd29c, 0x3f73f255, 0x79ce1418, 0xbf37c773);
 689         &_data_word(0xeacdf753, 0x5baafd5f, 0x146f3ddf, 0x86db4478);
 690         &_data_word(0x81f3afca, 0x3ec468b9, 0x2c342438, 0x5f40a3c2);
 691         &_data_word(0x72c31d16, 0x0c25e2bc, 0x8b493c28, 0x41950dff);
 692         &_data_word(0x7101a839, 0xdeb30c08, 0x9ce4b4d8, 0x90c15664);
 693         &_data_word(0x6184cb7b, 0x70b632d5, 0x745c6c48, 0x4257b8d0);
 694 #Td4:
 695         &data_word(0x52525252, 0x09090909, 0x6a6a6a6a, 0xd5d5d5d5);
 696         &data_word(0x30303030, 0x36363636, 0xa5a5a5a5, 0x38383838);
 697         &data_word(0xbfbfbfbf, 0x40404040, 0xa3a3a3a3, 0x9e9e9e9e);
 698         &data_word(0x81818181, 0xf3f3f3f3, 0xd7d7d7d7, 0xfbfbfbfb);
 699         &data_word(0x7c7c7c7c, 0xe3e3e3e3, 0x39393939, 0x82828282);
 700         &data_word(0x9b9b9b9b, 0x2f2f2f2f, 0xffffffff, 0x87878787);
 701         &data_word(0x34343434, 0x8e8e8e8e, 0x43434343, 0x44444444);
 702         &data_word(0xc4c4c4c4, 0xdededede, 0xe9e9e9e9, 0xcbcbcbcb);
 703         &data_word(0x54545454, 0x7b7b7b7b, 0x94949494, 0x32323232);
 704         &data_word(0xa6a6a6a6, 0xc2c2c2c2, 0x23232323, 0x3d3d3d3d);
 705         &data_word(0xeeeeeeee, 0x4c4c4c4c, 0x95959595, 0x0b0b0b0b);
 706         &data_word(0x42424242, 0xfafafafa, 0xc3c3c3c3, 0x4e4e4e4e);
 707         &data_word(0x08080808, 0x2e2e2e2e, 0xa1a1a1a1, 0x66666666);
 708         &data_word(0x28282828, 0xd9d9d9d9, 0x24242424, 0xb2b2b2b2);
 709         &data_word(0x76767676, 0x5b5b5b5b, 0xa2a2a2a2, 0x49494949);
 710         &data_word(0x6d6d6d6d, 0x8b8b8b8b, 0xd1d1d1d1, 0x25252525);
 711         &data_word(0x72727272, 0xf8f8f8f8, 0xf6f6f6f6, 0x64646464);
 712         &data_word(0x86868686, 0x68686868, 0x98989898, 0x16161616);
 713         &data_word(0xd4d4d4d4, 0xa4a4a4a4, 0x5c5c5c5c, 0xcccccccc);
 714         &data_word(0x5d5d5d5d, 0x65656565, 0xb6b6b6b6, 0x92929292);
 715         &data_word(0x6c6c6c6c, 0x70707070, 0x48484848, 0x50505050);
 716         &data_word(0xfdfdfdfd, 0xedededed, 0xb9b9b9b9, 0xdadadada);
 717         &data_word(0x5e5e5e5e, 0x15151515, 0x46464646, 0x57575757);
 718         &data_word(0xa7a7a7a7, 0x8d8d8d8d, 0x9d9d9d9d, 0x84848484);
 719         &data_word(0x90909090, 0xd8d8d8d8, 0xabababab, 0x00000000);
 720         &data_word(0x8c8c8c8c, 0xbcbcbcbc, 0xd3d3d3d3, 0x0a0a0a0a);
 721         &data_word(0xf7f7f7f7, 0xe4e4e4e4, 0x58585858, 0x05050505);
 722         &data_word(0xb8b8b8b8, 0xb3b3b3b3, 0x45454545, 0x06060606);
 723         &data_word(0xd0d0d0d0, 0x2c2c2c2c, 0x1e1e1e1e, 0x8f8f8f8f);
 724         &data_word(0xcacacaca, 0x3f3f3f3f, 0x0f0f0f0f, 0x02020202);
 725         &data_word(0xc1c1c1c1, 0xafafafaf, 0xbdbdbdbd, 0x03030303);
 726         &data_word(0x01010101, 0x13131313, 0x8a8a8a8a, 0x6b6b6b6b);
 727         &data_word(0x3a3a3a3a, 0x91919191, 0x11111111, 0x41414141);
 728         &data_word(0x4f4f4f4f, 0x67676767, 0xdcdcdcdc, 0xeaeaeaea);
 729         &data_word(0x97979797, 0xf2f2f2f2, 0xcfcfcfcf, 0xcececece);
 730         &data_word(0xf0f0f0f0, 0xb4b4b4b4, 0xe6e6e6e6, 0x73737373);
 731         &data_word(0x96969696, 0xacacacac, 0x74747474, 0x22222222);
 732         &data_word(0xe7e7e7e7, 0xadadadad, 0x35353535, 0x85858585);
 733         &data_word(0xe2e2e2e2, 0xf9f9f9f9, 0x37373737, 0xe8e8e8e8);
 734         &data_word(0x1c1c1c1c, 0x75757575, 0xdfdfdfdf, 0x6e6e6e6e);
 735         &data_word(0x47474747, 0xf1f1f1f1, 0x1a1a1a1a, 0x71717171);
 736         &data_word(0x1d1d1d1d, 0x29292929, 0xc5c5c5c5, 0x89898989);
 737         &data_word(0x6f6f6f6f, 0xb7b7b7b7, 0x62626262, 0x0e0e0e0e);
 738         &data_word(0xaaaaaaaa, 0x18181818, 0xbebebebe, 0x1b1b1b1b);
 739         &data_word(0xfcfcfcfc, 0x56565656, 0x3e3e3e3e, 0x4b4b4b4b);
 740         &data_word(0xc6c6c6c6, 0xd2d2d2d2, 0x79797979, 0x20202020);
 741         &data_word(0x9a9a9a9a, 0xdbdbdbdb, 0xc0c0c0c0, 0xfefefefe);
 742         &data_word(0x78787878, 0xcdcdcdcd, 0x5a5a5a5a, 0xf4f4f4f4);
 743         &data_word(0x1f1f1f1f, 0xdddddddd, 0xa8a8a8a8, 0x33333333);
 744         &data_word(0x88888888, 0x07070707, 0xc7c7c7c7, 0x31313131);
 745         &data_word(0xb1b1b1b1, 0x12121212, 0x10101010, 0x59595959);
 746         &data_word(0x27272727, 0x80808080, 0xecececec, 0x5f5f5f5f);
 747         &data_word(0x60606060, 0x51515151, 0x7f7f7f7f, 0xa9a9a9a9);
 748         &data_word(0x19191919, 0xb5b5b5b5, 0x4a4a4a4a, 0x0d0d0d0d);
 749         &data_word(0x2d2d2d2d, 0xe5e5e5e5, 0x7a7a7a7a, 0x9f9f9f9f);
 750         &data_word(0x93939393, 0xc9c9c9c9, 0x9c9c9c9c, 0xefefefef);
 751         &data_word(0xa0a0a0a0, 0xe0e0e0e0, 0x3b3b3b3b, 0x4d4d4d4d);
 752         &data_word(0xaeaeaeae, 0x2a2a2a2a, 0xf5f5f5f5, 0xb0b0b0b0);
 753         &data_word(0xc8c8c8c8, 0xebebebeb, 0xbbbbbbbb, 0x3c3c3c3c);
 754         &data_word(0x83838383, 0x53535353, 0x99999999, 0x61616161);
 755         &data_word(0x17171717, 0x2b2b2b2b, 0x04040404, 0x7e7e7e7e);
 756         &data_word(0xbabababa, 0x77777777, 0xd6d6d6d6, 0x26262626);
 757         &data_word(0xe1e1e1e1, 0x69696969, 0x14141414, 0x63636363);
 758         &data_word(0x55555555, 0x21212121, 0x0c0c0c0c, 0x7d7d7d7d);
 759 &function_end_B("_x86_AES_decrypt");
 760
 761 # void AES_decrypt (const void *inp,void *out,const AES_KEY *key);
 762 &public_label("AES_Td");
 763 &function_begin("AES_decrypt");
 764         &mov    ($acc,&wparam(0));              # load inp
 765         &mov    ($key,&wparam(2));              # load key
 766
 767         &mov    ($s0,"esp");
 768         &sub    ("esp",24);
 769         &and    ("esp",-64);
 770         &add    ("esp",4);
 771         &mov    (&DWP(16,"esp"),$s0);
 772
 773         &call   (&label("pic_point"));          # make it PIC!
 774         &set_label("pic_point");
 775         &blindpop("ebp");
 776         &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
 777
 778         &mov    ($s0,&DWP(0,$acc));             # load input data
 779         &mov    ($s1,&DWP(4,$acc));
 780         &mov    ($s2,&DWP(8,$acc));
 781         &mov    ($s3,&DWP(12,$acc));
 782
 783         &call   ("_x86_AES_decrypt");
 784
 785         &mov    ("esp",&DWP(16,"esp"));
 786
 787         &mov    ($acc,&wparam(1));              # load out
 788         &mov    (&DWP(0,$acc),$s0);             # write output data
 789         &mov    (&DWP(4,$acc),$s1);
 790         &mov    (&DWP(8,$acc),$s2);
 791         &mov    (&DWP(12,$acc),$s3);
 792 &function_end("AES_decrypt");
 793
 794 # void AES_cbc_encrypt (const void char *inp, unsigned char *out,
 795 #                       size_t length, const AES_KEY *key,
 796 #                       unsigned char *ivp,const int enc);
 797 {
 798 # stack frame layout
 799 # -4(%esp)      0(%esp)         return address
 800 # 0(%esp)       4(%esp)         tmp1
 801 # 4(%esp)       8(%esp)         tmp2
 802 # 8(%esp)       12(%esp)        key
 803 # 12(%esp)      16(%esp)        end of key schedule
 804 my $_esp=&DWP(16,"esp");        #saved %esp
 805 my $_inp=&DWP(20,"esp");        #copy of wparam(0)
 806 my $_out=&DWP(24,"esp");        #copy of wparam(1)
 807 my $_len=&DWP(28,"esp");        #copy of wparam(2)
 808 my $_key=&DWP(32,"esp");        #copy of wparam(3)
 809 my $_ivp=&DWP(36,"esp");        #copy of wparam(4)
 810 my $_tmp=&DWP(40,"esp");        #volatile variable
 811 my $ivec=&DWP(44,"esp");        #ivec[16]
 812 my $aes_key=&DWP(60,"esp");     #copy of aes_key
 813 my $mark=&DWP(60+240,"esp");    #copy of aes_key->rounds
 814
 815 &public_label("AES_Te");
 816 &public_label("AES_Td");
 817 &function_begin("AES_cbc_encrypt");
 818         &mov    ($s2 eq "ecx"? $s2 : "",&wparam(2));    # load len
 819         &cmp    ($s2,0);
 820         &je     (&label("enc_out"));
 821
 822         &call   (&label("pic_point"));          # make it PIC!
 823         &set_label("pic_point");
 824         &blindpop("ebp");
 825
 826         &pushf  ();
 827         &cld    ();
 828
 829         &cmp    (&wparam(5),0);
 830         &je     (&label("DECRYPT"));
 831
 832         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
 833
 834         # allocate aligned stack frame...
 835         &lea    ($key,&DWP(-64-244,"esp"));
 836         &and    ($key,-64);
 837
 838         # ... and make sure it doesn't alias with AES_Te modulo 4096
 839         &mov    ($s0,"ebp");
 840         &lea    ($s1,&DWP(2048,"ebp"));
 841         &mov    ($s3,$key);
 842         &and    ($s0,0xfff);            # s = %ebp&0xfff
 843         &and    ($s1,0xfff);            # e = (%ebp+2048)&0xfff
 844         &and    ($s3,0xfff);            # p = %esp&0xfff
 845
 846         &cmp    ($s3,$s1);              # if (p>=e) %esp =- (p-e);
 847         &jb     (&label("te_break_out"));
 848         &sub    ($s3,$s1);
 849         &sub    ($key,$s3);
 850         &jmp    (&label("te_ok"));
 851         &set_label("te_break_out");     # else %esp -= (p-s)&0xfff + framesz;
 852         &sub    ($s3,$s0);
 853         &and    ($s3,0xfff);
 854         &add    ($s3,64+256);
 855         &sub    ($key,$s3);
 856         &align  (4);
 857         &set_label("te_ok");
 858
 859         &mov    ($s0,&wparam(0));       # load inp
 860         &mov    ($s1,&wparam(1));       # load out
 861         &mov    ($s3,&wparam(3));       # load key
 862         &mov    ($acc,&wparam(4));      # load ivp
 863
 864         &exch   ("esp",$key);
 865         &add    ("esp",4);              # reserve for return address!
 866         &mov    ($_esp,$key);           # save %esp
 867
 868         &mov    ($_inp,$s0);            # save copy of inp
 869         &mov    ($_out,$s1);            # save copy of out
 870         &mov    ($_len,$s2);            # save copy of len
 871         &mov    ($_key,$s3);            # save copy of key
 872         &mov    ($_ivp,$acc);           # save copy of ivp
 873
 874         &mov    ($mark,0);              # copy of aes_key->rounds = 0;
 875         if ($compromise) {
 876                 &cmp    ($s2,$compromise);
 877                 &jb     (&label("skip_ecopy"));
 878         }
 879         # do we copy key schedule to stack?
 880         &mov    ($s1 eq "ebx" ? $s1 : "",$s3);
 881         &mov    ($s2 eq "ecx" ? $s2 : "",244/4);
 882         &sub    ($s1,"ebp");
 883         &mov    ("esi",$s3);
 884         &and    ($s1,0xfff);
 885         &lea    ("edi",$aes_key);
 886         &cmp    ($s1,2048);
 887         &jb     (&label("do_ecopy"));
 888         &cmp    ($s1,4096-244);
 889         &jb     (&label("skip_ecopy"));
 890         &align  (4);
 891         &set_label("do_ecopy");
 892                 &mov    ($_key,"edi");
 893                 &data_word(0xF689A5F3); # rep movsd
 894         &set_label("skip_ecopy");
 895
 896         &mov    ($acc,$s0);
 897         &mov    ($key,16);
 898         &align  (4);
 899         &set_label("prefetch_te");
 900                 &mov    ($s0,&DWP(0,"ebp"));
 901                 &mov    ($s1,&DWP(32,"ebp"));
 902                 &mov    ($s2,&DWP(64,"ebp"));
 903                 &mov    ($s3,&DWP(96,"ebp"));
 904                 &lea    ("ebp",&DWP(128,"ebp"));
 905                 &dec    ($key);
 906         &jnz    (&label("prefetch_te"));
 907         &sub    ("ebp",2048);
 908
 909         &mov    ($s2,$_len);
 910         &mov    ($key,$_ivp);
 911         &test   ($s2,0xFFFFFFF0);
 912         &jz     (&label("enc_tail"));           # short input...
 913
 914         &mov    ($s0,&DWP(0,$key));             # load iv
 915         &mov    ($s1,&DWP(4,$key));
 916
 917         &align  (4);
 918         &set_label("enc_loop");
 919                 &mov    ($s2,&DWP(8,$key));
 920                 &mov    ($s3,&DWP(12,$key));
 921
 922                 &xor    ($s0,&DWP(0,$acc));     # xor input data
 923                 &xor    ($s1,&DWP(4,$acc));
 924                 &xor    ($s2,&DWP(8,$acc));
 925                 &xor    ($s3,&DWP(12,$acc));
 926
 927                 &mov    ($key,$_key);           # load key
 928                 &call   ("_x86_AES_encrypt");
 929
 930                 &mov    ($acc,$_inp);           # load inp
 931                 &mov    ($key,$_out);           # load out
 932
 933                 &mov    (&DWP(0,$key),$s0);     # save output data
 934                 &mov    (&DWP(4,$key),$s1);
 935                 &mov    (&DWP(8,$key),$s2);
 936                 &mov    (&DWP(12,$key),$s3);
 937
 938                 &mov    ($s2,$_len);            # load len
 939
 940                 &lea    ($acc,&DWP(16,$acc));
 941                 &mov    ($_inp,$acc);           # save inp
 942
 943                 &lea    ($s3,&DWP(16,$key));
 944                 &mov    ($_out,$s3);            # save out
 945
 946                 &sub    ($s2,16);
 947                 &test   ($s2,0xFFFFFFF0);
 948                 &mov    ($_len,$s2);            # save len
 949         &jnz    (&label("enc_loop"));
 950         &test   ($s2,15);
 951         &jnz    (&label("enc_tail"));
 952         &mov    ($acc,$_ivp);           # load ivp
 953         &mov    ($s2,&DWP(8,$key));     # restore last dwords
 954         &mov    ($s3,&DWP(12,$key));
 955         &mov    (&DWP(0,$acc),$s0);     # save ivec
 956         &mov    (&DWP(4,$acc),$s1);
 957         &mov    (&DWP(8,$acc),$s2);
 958         &mov    (&DWP(12,$acc),$s3);
 959
 960         &cmp    ($mark,0);              # was the key schedule copied?
 961         &mov    ("edi",$_key);
 962         &mov    ("esp",$_esp);
 963         &je     (&label("skip_ezero"));
 964         if ($compromise) {
 965                 &cmp    (&wparam(2),$compromise);
 966                 &jb     (&label("skip_ezero"));
 967         }
 968         # zero copy of key schedule
 969         &mov    ("ecx",240/4);
 970         &xor    ("eax","eax");
 971         &align  (4);
 972         &data_word(0xF689ABF3); # rep stosd
 973         &set_label("skip_ezero")
 974         &popf   ();
 975     &set_label("enc_out");
 976         &function_end_A();
 977         &pushf  ();                     # kludge, never executed
 978
 979     &align      (4);
 980     &set_label("enc_tail");
 981         &push   ($key eq "edi" ? $key : "");    # push ivp
 982         &mov    ($key,$_out);                   # load out
 983         &mov    ($s1,16);
 984         &sub    ($s1,$s2);
 985         &cmp    ($key,$acc);                    # compare with inp
 986         &je     (&label("enc_in_place"));
 987         &align  (4);
 988         &data_word(0xF689A4F3); # rep movsb     # copy input
 989         &jmp    (&label("enc_skip_in_place"));
 990     &set_label("enc_in_place");
 991         &lea    ($key,&DWP(0,$key,$s2));
 992     &set_label("enc_skip_in_place");
 993         &mov    ($s2,$s1);
 994         &xor    ($s0,$s0);
 995         &align  (4);
 996         &data_word(0xF689AAF3); # rep stosb     # zero tail
 997         &pop    ($key);                         # pop ivp
 998
 999         &mov    ($acc,$_out);                   # output as input
1000         &mov    ($s0,&DWP(0,$key));
1001         &mov    ($s1,&DWP(4,$key));
1002         &mov    ($_len,16);                     # len=16
1003         &jmp    (&label("enc_loop"));           # one more spin...
1004
1005 #----------------------------- DECRYPT -----------------------------#
1006 &align  (4);
1007 &set_label("DECRYPT");
1008         &lea    ("ebp",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
1009
1010         # allocate aligned stack frame...
1011         &lea    ($key,&DWP(-64-244,"esp"));
1012         &and    ($key,-64);
1013
1014         # ... and make sure it doesn't alias with AES_Td modulo 4096
1015         &mov    ($s0,"ebp");
1016         &lea    ($s1,&DWP(3072,"ebp"));
1017         &mov    ($s3,$key);
1018         &and    ($s0,0xfff);            # s = %ebp&0xfff
1019         &and    ($s1,0xfff);            # e = (%ebp+3072)&0xfff
1020         &and    ($s3,0xfff);            # p = %esp&0xfff
1021
1022         &cmp    ($s3,$s1);              # if (p>=e) %esp =- (p-e);
1023         &jb     (&label("td_break_out"));
1024         &sub    ($s3,$s1);
1025         &sub    ($key,$s3);
1026         &jmp    (&label("td_ok"));
1027         &set_label("td_break_out");     # else %esp -= (p-s)&0xfff + framesz;
1028         &sub    ($s3,$s0);
1029         &and    ($s3,0xfff);
1030         &add    ($s3,64+256);
1031         &sub    ($key,$s3);
1032         &align  (4);
1033         &set_label("td_ok");
1034
1035         &mov    ($s0,&wparam(0));       # load inp
1036         &mov    ($s1,&wparam(1));       # load out
1037         &mov    ($s3,&wparam(3));       # load key
1038         &mov    ($acc,&wparam(4));      # load ivp
1039
1040         &exch   ("esp",$key);
1041         &add    ("esp",4);              # reserve for return address!
1042         &mov    ($_esp,$key);           # save %esp
1043
1044         &mov    ($_inp,$s0);            # save copy of inp
1045         &mov    ($_out,$s1);            # save copy of out
1046         &mov    ($_len,$s2);            # save copy of len
1047         &mov    ($_key,$s3);            # save copy of key
1048         &mov    ($_ivp,$acc);           # save copy of ivp
1049
1050         &mov    ($mark,0);              # copy of aes_key->rounds = 0;
1051         if ($compromise) {
1052                 &cmp    ($s2,$compromise);
1053                 &jb     (&label("skip_dcopy"));
1054         }
1055         # do we copy key schedule to stack?
1056         &mov    ($s1 eq "ebx" ? $s1 : "",$s3);
1057         &mov    ($s2 eq "ecx" ? $s2 : "",244/4);
1058         &sub    ($s1,"ebp");
1059         &mov    ("esi",$s3);
1060         &and    ($s1,0xfff);
1061         &lea    ("edi",$aes_key);
1062         &cmp    ($s1,3072);
1063         &jb     (&label("do_dcopy"));
1064         &cmp    ($s1,4096-244);
1065         &jb     (&label("skip_dcopy"));
1066         &align  (4);
1067         &set_label("do_dcopy");
1068                 &mov    ($_key,"edi");
1069                 &data_word(0xF689A5F3); # rep movsd
1070         &set_label("skip_dcopy");
1071
1072         &mov    ($acc,$s0);
1073         &mov    ($key,24);
1074         &align  (4);
1075         &set_label("prefetch_td");
1076                 &mov    ($s0,&DWP(0,"ebp"));
1077                 &mov    ($s1,&DWP(32,"ebp"));
1078                 &mov    ($s2,&DWP(64,"ebp"));
1079                 &mov    ($s3,&DWP(96,"ebp"));
1080                 &lea    ("ebp",&DWP(128,"ebp"));
1081                 &dec    ($key);
1082         &jnz    (&label("prefetch_td"));
1083         &sub    ("ebp",3072);
1084
1085         &cmp    ($acc,$_out);
1086         &je     (&label("dec_in_place"));       # in-place processing...
1087
1088         &mov    ($key,$_ivp);           # load ivp
1089         &mov    ($_tmp,$key);
1090
1091         &align  (4);
1092         &set_label("dec_loop");
1093                 &mov    ($s0,&DWP(0,$acc));     # read input
1094                 &mov    ($s1,&DWP(4,$acc));
1095                 &mov    ($s2,&DWP(8,$acc));
1096                 &mov    ($s3,&DWP(12,$acc));
1097
1098                 &mov    ($key,$_key);           # load key
1099                 &call   ("_x86_AES_decrypt");
1100
1101                 &mov    ($key,$_tmp);           # load ivp
1102                 &mov    ($acc,$_len);           # load len
1103                 &xor    ($s0,&DWP(0,$key));     # xor iv
1104                 &xor    ($s1,&DWP(4,$key));
1105                 &xor    ($s2,&DWP(8,$key));
1106                 &xor    ($s3,&DWP(12,$key));
1107
1108                 &sub    ($acc,16);
1109                 &jc     (&label("dec_partial"));
1110                 &mov    ($_len,$acc);           # save len
1111                 &mov    ($acc,$_inp);           # load inp
1112                 &mov    ($key,$_out);           # load out
1113
1114                 &mov    (&DWP(0,$key),$s0);     # write output
1115                 &mov    (&DWP(4,$key),$s1);
1116                 &mov    (&DWP(8,$key),$s2);
1117                 &mov    (&DWP(12,$key),$s3);
1118
1119                 &mov    ($_tmp,$acc);           # save ivp
1120                 &lea    ($acc,&DWP(16,$acc));
1121                 &mov    ($_inp,$acc);           # save inp
1122
1123                 &lea    ($key,&DWP(16,$key));
1124                 &mov    ($_out,$key);           # save out
1125
1126         &jnz    (&label("dec_loop"));
1127         &mov    ($key,$_tmp);           # load temp ivp
1128     &set_label("dec_end");
1129         &mov    ($acc,$_ivp);           # load user ivp
1130         &mov    ($s0,&DWP(0,$key));     # load iv
1131         &mov    ($s1,&DWP(4,$key));
1132         &mov    ($s2,&DWP(8,$key));
1133         &mov    ($s3,&DWP(12,$key));
1134         &mov    (&DWP(0,$acc),$s0);     # copy back to user
1135         &mov    (&DWP(4,$acc),$s1);
1136         &mov    (&DWP(8,$acc),$s2);
1137         &mov    (&DWP(12,$acc),$s3);
1138         &jmp    (&label("dec_out"));
1139
1140     &align      (4);
1141     &set_label("dec_partial");
1142         &lea    ($key,$ivec);
1143         &mov    (&DWP(0,$key),$s0);     # dump output to stack
1144         &mov    (&DWP(4,$key),$s1);
1145         &mov    (&DWP(8,$key),$s2);
1146         &mov    (&DWP(12,$key),$s3);
1147         &lea    ($s2 eq "ecx" ? $s2 : "",&DWP(16,$acc));
1148         &mov    ($acc eq "esi" ? $acc : "",$key);
1149         &mov    ($key eq "edi" ? $key : "",$_out);      # load out
1150         &data_word(0xF689A4F3); # rep movsb             # copy output
1151         &mov    ($key,$_inp);                           # use inp as temp ivp
1152         &jmp    (&label("dec_end"));
1153
1154     &align      (4);
1155     &set_label("dec_in_place");
1156         &set_label("dec_in_place_loop");
1157                 &lea    ($key,$ivec);
1158                 &mov    ($s0,&DWP(0,$acc));     # read input
1159                 &mov    ($s1,&DWP(4,$acc));
1160                 &mov    ($s2,&DWP(8,$acc));
1161                 &mov    ($s3,&DWP(12,$acc));
1162
1163                 &mov    (&DWP(0,$key),$s0);     # copy to temp
1164                 &mov    (&DWP(4,$key),$s1);
1165                 &mov    (&DWP(8,$key),$s2);
1166                 &mov    (&DWP(12,$key),$s3);
1167
1168                 &mov    ($key,$_key);           # load key
1169                 &call   ("_x86_AES_decrypt");
1170
1171                 &mov    ($key,$_ivp);           # load ivp
1172                 &mov    ($acc,$_out);           # load out
1173                 &xor    ($s0,&DWP(0,$key));     # xor iv
1174                 &xor    ($s1,&DWP(4,$key));
1175                 &xor    ($s2,&DWP(8,$key));
1176                 &xor    ($s3,&DWP(12,$key));
1177
1178                 &mov    (&DWP(0,$acc),$s0);     # write output
1179                 &mov    (&DWP(4,$acc),$s1);
1180                 &mov    (&DWP(8,$acc),$s2);
1181                 &mov    (&DWP(12,$acc),$s3);
1182
1183                 &lea    ($acc,&DWP(16,$acc));
1184                 &mov    ($_out,$acc);           # save out
1185
1186                 &lea    ($acc,$ivec);
1187                 &mov    ($s0,&DWP(0,$acc));     # read temp
1188                 &mov    ($s1,&DWP(4,$acc));
1189                 &mov    ($s2,&DWP(8,$acc));
1190                 &mov    ($s3,&DWP(12,$acc));
1191
1192                 &mov    (&DWP(0,$key),$s0);     # copy iv
1193                 &mov    (&DWP(4,$key),$s1);
1194                 &mov    (&DWP(8,$key),$s2);
1195                 &mov    (&DWP(12,$key),$s3);
1196
1197                 &mov    ($acc,$_inp);           # load inp
1198
1199                 &lea    ($acc,&DWP(16,$acc));
1200                 &mov    ($_inp,$acc);           # save inp
1201
1202                 &mov    ($s2,$_len);            # load len
1203                 &sub    ($s2,16);
1204                 &jc     (&label("dec_in_place_partial"));
1205                 &mov    ($_len,$s2);            # save len
1206         &jnz    (&label("dec_in_place_loop"));
1207         &jmp    (&label("dec_out"));
1208
1209     &align      (4);
1210     &set_label("dec_in_place_partial");
1211         # one can argue if this is actually required...
1212         &mov    ($key eq "edi" ? $key : "",$_out);
1213         &lea    ($acc eq "esi" ? $acc : "",$ivec);
1214         &lea    ($key,&DWP(0,$key,$s2));
1215         &lea    ($acc,&DWP(16,$acc,$s2));
1216         &neg    ($s2 eq "ecx" ? $s2 : "");
1217         &data_word(0xF689A4F3); # rep movsb     # restore tail
1218
1219     &align      (4);
1220     &set_label("dec_out");
1221     &cmp        ($mark,0);              # was the key schedule copied?
1222     &mov        ("edi",$_key);
1223     &mov        ("esp",$_esp);
1224     &je         (&label("skip_dzero"));
1225     if ($compromise) {
1226         &cmp    (&wparam(2),$compromise);
1227         &jb     (&label("skip_dzero"));
1228     }
1229     # zero copy of key schedule
1230     &mov        ("ecx",240/4);
1231     &xor        ("eax","eax");
1232     &align      (4);
1233     &data_word(0xF689ABF3);     # rep stosd
1234     &set_label("skip_dzero")
1235     &popf       ();
1236 &function_end("AES_cbc_encrypt");
1237 }
1238
1239 #------------------------------------------------------------------#
1240
1241 sub enckey()
1242 {
1243         &movz   ("esi",&LB("edx"));             # rk[i]>>0
1244         &mov    ("ebx",&DWP(2,"ebp","esi",8));
1245         &movz   ("esi",&HB("edx"));             # rk[i]>>8
1246         &and    ("ebx",0xFF000000);
1247         &xor    ("eax","ebx");
1248
1249         &mov    ("ebx",&DWP(2,"ebp","esi",8));
1250         &shr    ("edx",16);
1251         &and    ("ebx",0x000000FF);
1252         &movz   ("esi",&LB("edx"));             # rk[i]>>16
1253         &xor    ("eax","ebx");
1254
1255         &mov    ("ebx",&DWP(0,"ebp","esi",8));
1256         &movz   ("esi",&HB("edx"));             # rk[i]>>24
1257         &and    ("ebx",0x0000FF00);
1258         &xor    ("eax","ebx");
1259
1260         &mov    ("ebx",&DWP(0,"ebp","esi",8));
1261         &and    ("ebx",0x00FF0000);
1262         &xor    ("eax","ebx");
1263
1264         &xor    ("eax",&DWP(2048,"ebp","ecx",4));       # rcon
1265 }
1266
1267 # int AES_set_encrypt_key(const unsigned char *userKey, const int bits,
1268 #                        AES_KEY *key)
1269 &public_label("AES_Te");
1270 &function_begin("AES_set_encrypt_key");
1271         &mov    ("esi",&wparam(0));             # user supplied key
1272         &mov    ("edi",&wparam(2));             # private key schedule
1273
1274         &test   ("esi",-1);
1275         &jz     (&label("badpointer"));
1276         &test   ("edi",-1);
1277         &jz     (&label("badpointer"));
1278
1279         &call   (&label("pic_point"));
1280         &set_label("pic_point");
1281         &blindpop("ebp");
1282         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1283
1284         &mov    ("ecx",&wparam(1));             # number of bits in key
1285         &cmp    ("ecx",128);
1286         &je     (&label("10rounds"));
1287         &cmp    ("ecx",192);
1288         &je     (&label("12rounds"));
1289         &cmp    ("ecx",256);
1290         &je     (&label("14rounds"));
1291         &mov    ("eax",-2);                     # invalid number of bits
1292         &jmp    (&label("exit"));
1293
1294     &set_label("10rounds");
1295         &mov    ("eax",&DWP(0,"esi"));          # copy first 4 dwords
1296         &mov    ("ebx",&DWP(4,"esi"));
1297         &mov    ("ecx",&DWP(8,"esi"));
1298         &mov    ("edx",&DWP(12,"esi"));
1299         &mov    (&DWP(0,"edi"),"eax");
1300         &mov    (&DWP(4,"edi"),"ebx");
1301         &mov    (&DWP(8,"edi"),"ecx");
1302         &mov    (&DWP(12,"edi"),"edx");
1303
1304         &xor    ("ecx","ecx");
1305         &jmp    (&label("10shortcut"));
1306
1307         &align  (4);
1308         &set_label("10loop");
1309                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1310                 &mov    ("edx",&DWP(12,"edi"));         # rk[3]
1311         &set_label("10shortcut");
1312                 &enckey ();
1313
1314                 &mov    (&DWP(16,"edi"),"eax");         # rk[4]
1315                 &xor    ("eax",&DWP(4,"edi"));
1316                 &mov    (&DWP(20,"edi"),"eax");         # rk[5]
1317                 &xor    ("eax",&DWP(8,"edi"));
1318                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1319                 &xor    ("eax",&DWP(12,"edi"));
1320                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1321                 &inc    ("ecx");
1322                 &add    ("edi",16);
1323                 &cmp    ("ecx",10);
1324         &jl     (&label("10loop"));
1325
1326         &mov    (&DWP(80,"edi"),10);            # setup number of rounds
1327         &xor    ("eax","eax");
1328         &jmp    (&label("exit"));
1329
1330     &set_label("12rounds");
1331         &mov    ("eax",&DWP(0,"esi"));          # copy first 6 dwords
1332         &mov    ("ebx",&DWP(4,"esi"));
1333         &mov    ("ecx",&DWP(8,"esi"));
1334         &mov    ("edx",&DWP(12,"esi"));
1335         &mov    (&DWP(0,"edi"),"eax");
1336         &mov    (&DWP(4,"edi"),"ebx");
1337         &mov    (&DWP(8,"edi"),"ecx");
1338         &mov    (&DWP(12,"edi"),"edx");
1339         &mov    ("ecx",&DWP(16,"esi"));
1340         &mov    ("edx",&DWP(20,"esi"));
1341         &mov    (&DWP(16,"edi"),"ecx");
1342         &mov    (&DWP(20,"edi"),"edx");
1343
1344         &xor    ("ecx","ecx");
1345         &jmp    (&label("12shortcut"));
1346
1347         &align  (4);
1348         &set_label("12loop");
1349                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1350                 &mov    ("edx",&DWP(20,"edi"));         # rk[5]
1351         &set_label("12shortcut");
1352                 &enckey ();
1353
1354                 &mov    (&DWP(24,"edi"),"eax");         # rk[6]
1355                 &xor    ("eax",&DWP(4,"edi"));
1356                 &mov    (&DWP(28,"edi"),"eax");         # rk[7]
1357                 &xor    ("eax",&DWP(8,"edi"));
1358                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1359                 &xor    ("eax",&DWP(12,"edi"));
1360                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1361
1362                 &cmp    ("ecx",7);
1363                 &je     (&label("12break"));
1364                 &inc    ("ecx");
1365
1366                 &xor    ("eax",&DWP(16,"edi"));
1367                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1368                 &xor    ("eax",&DWP(20,"edi"));
1369                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1370
1371                 &add    ("edi",24);
1372         &jmp    (&label("12loop"));
1373
1374         &set_label("12break");
1375         &mov    (&DWP(72,"edi"),12);            # setup number of rounds
1376         &xor    ("eax","eax");
1377         &jmp    (&label("exit"));
1378
1379     &set_label("14rounds");
1380         &mov    ("eax",&DWP(0,"esi"));          # copy first 8 dwords
1381         &mov    ("ebx",&DWP(4,"esi"));
1382         &mov    ("ecx",&DWP(8,"esi"));
1383         &mov    ("edx",&DWP(12,"esi"));
1384         &mov    (&DWP(0,"edi"),"eax");
1385         &mov    (&DWP(4,"edi"),"ebx");
1386         &mov    (&DWP(8,"edi"),"ecx");
1387         &mov    (&DWP(12,"edi"),"edx");
1388         &mov    ("eax",&DWP(16,"esi"));
1389         &mov    ("ebx",&DWP(20,"esi"));
1390         &mov    ("ecx",&DWP(24,"esi"));
1391         &mov    ("edx",&DWP(28,"esi"));
1392         &mov    (&DWP(16,"edi"),"eax");
1393         &mov    (&DWP(20,"edi"),"ebx");
1394         &mov    (&DWP(24,"edi"),"ecx");
1395         &mov    (&DWP(28,"edi"),"edx");
1396
1397         &xor    ("ecx","ecx");
1398         &jmp    (&label("14shortcut"));
1399
1400         &align  (4);
1401         &set_label("14loop");
1402                 &mov    ("edx",&DWP(28,"edi"));         # rk[7]
1403         &set_label("14shortcut");
1404                 &mov    ("eax",&DWP(0,"edi"));          # rk[0]
1405
1406                 &enckey ();
1407
1408                 &mov    (&DWP(32,"edi"),"eax");         # rk[8]
1409                 &xor    ("eax",&DWP(4,"edi"));
1410                 &mov    (&DWP(36,"edi"),"eax");         # rk[9]
1411                 &xor    ("eax",&DWP(8,"edi"));
1412                 &mov    (&DWP(40,"edi"),"eax");         # rk[10]
1413                 &xor    ("eax",&DWP(12,"edi"));
1414                 &mov    (&DWP(44,"edi"),"eax");         # rk[11]
1415
1416                 &cmp    ("ecx",6);
1417                 &je     (&label("14break"));
1418                 &inc    ("ecx");
1419
1420                 &mov    ("edx","eax");
1421                 &mov    ("eax",&DWP(16,"edi"));         # rk[4]
1422                 &movz   ("esi",&LB("edx"));             # rk[11]>>0
1423                 &mov    ("ebx",&DWP(2,"ebp","esi",8));
1424                 &movz   ("esi",&HB("edx"));             # rk[11]>>8
1425                 &and    ("ebx",0x000000FF);
1426                 &xor    ("eax","ebx");
1427
1428                 &mov    ("ebx",&DWP(0,"ebp","esi",8));
1429                 &shr    ("edx",16);
1430                 &and    ("ebx",0x0000FF00);
1431                 &movz   ("esi",&LB("edx"));             # rk[11]>>16
1432                 &xor    ("eax","ebx");
1433
1434                 &mov    ("ebx",&DWP(0,"ebp","esi",8));
1435                 &movz   ("esi",&HB("edx"));             # rk[11]>>24
1436                 &and    ("ebx",0x00FF0000);
1437                 &xor    ("eax","ebx");
1438
1439                 &mov    ("ebx",&DWP(2,"ebp","esi",8));
1440                 &and    ("ebx",0xFF000000);
1441                 &xor    ("eax","ebx");
1442
1443                 &mov    (&DWP(48,"edi"),"eax");         # rk[12]
1444                 &xor    ("eax",&DWP(20,"edi"));
1445                 &mov    (&DWP(52,"edi"),"eax");         # rk[13]
1446                 &xor    ("eax",&DWP(24,"edi"));
1447                 &mov    (&DWP(56,"edi"),"eax");         # rk[14]
1448                 &xor    ("eax",&DWP(28,"edi"));
1449                 &mov    (&DWP(60,"edi"),"eax");         # rk[15]
1450
1451                 &add    ("edi",32);
1452         &jmp    (&label("14loop"));
1453
1454         &set_label("14break");
1455         &mov    (&DWP(48,"edi"),14);            # setup number of rounds
1456         &xor    ("eax","eax");
1457         &jmp    (&label("exit"));
1458
1459     &set_label("badpointer");
1460         &mov    ("eax",-1);
1461     &set_label("exit");
1462 &function_end("AES_set_encrypt_key");
1463
1464 sub deckey()
1465 { my ($i,$ptr,$te,$td) = @_;
1466
1467         &mov    ("eax",&DWP($i,$ptr));
1468         &mov    ("edx","eax");
1469         &movz   ("ebx",&HB("eax"));
1470         &shr    ("edx",16);
1471         &and    ("eax",0xFF);
1472         &movz   ("eax",&BP(2,$te,"eax",8));
1473         &movz   ("ebx",&BP(2,$te,"ebx",8));
1474         &mov    ("eax",&DWP(0,$td,"eax",8));
1475         &xor    ("eax",&DWP(3,$td,"ebx",8));
1476         &movz   ("ebx",&HB("edx"));
1477         &and    ("edx",0xFF);
1478         &movz   ("edx",&BP(2,$te,"edx",8));
1479         &movz   ("ebx",&BP(2,$te,"ebx",8));
1480         &xor    ("eax",&DWP(2,$td,"edx",8));
1481         &xor    ("eax",&DWP(1,$td,"ebx",8));
1482         &mov    (&DWP($i,$ptr),"eax");
1483 }
1484
1485 # int AES_set_decrypt_key(const unsigned char *userKey, const int bits,
1486 #                        AES_KEY *key)
1487 &public_label("AES_Td");
1488 &public_label("AES_Te");
1489 &function_begin_B("AES_set_decrypt_key");
1490         &mov    ("eax",&wparam(0));
1491         &mov    ("ecx",&wparam(1));
1492         &mov    ("edx",&wparam(2));
1493         &sub    ("esp",12);
1494         &mov    (&DWP(0,"esp"),"eax");
1495         &mov    (&DWP(4,"esp"),"ecx");
1496         &mov    (&DWP(8,"esp"),"edx");
1497         &call   ("AES_set_encrypt_key");
1498         &add    ("esp",12);
1499         &cmp    ("eax",0);
1500         &je     (&label("proceed"));
1501         &ret    ();
1502
1503     &set_label("proceed");
1504         &push   ("ebp");
1505         &push   ("ebx");
1506         &push   ("esi");
1507         &push   ("edi");
1508
1509         &mov    ("esi",&wparam(2));
1510         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1511         &lea    ("ecx",&DWP(0,"","ecx",4));
1512         &lea    ("edi",&DWP(0,"esi","ecx",4));  # pointer to last chunk
1513
1514         &align  (4);
1515         &set_label("invert");                   # invert order of chunks
1516                 &mov    ("eax",&DWP(0,"esi"));
1517                 &mov    ("ebx",&DWP(4,"esi"));
1518                 &mov    ("ecx",&DWP(0,"edi"));
1519                 &mov    ("edx",&DWP(4,"edi"));
1520                 &mov    (&DWP(0,"edi"),"eax");
1521                 &mov    (&DWP(4,"edi"),"ebx");
1522                 &mov    (&DWP(0,"esi"),"ecx");
1523                 &mov    (&DWP(4,"esi"),"edx");
1524                 &mov    ("eax",&DWP(8,"esi"));
1525                 &mov    ("ebx",&DWP(12,"esi"));
1526                 &mov    ("ecx",&DWP(8,"edi"));
1527                 &mov    ("edx",&DWP(12,"edi"));
1528                 &mov    (&DWP(8,"edi"),"eax");
1529                 &mov    (&DWP(12,"edi"),"ebx");
1530                 &mov    (&DWP(8,"esi"),"ecx");
1531                 &mov    (&DWP(12,"esi"),"edx");
1532                 &add    ("esi",16);
1533                 &sub    ("edi",16);
1534                 &cmp    ("esi","edi");
1535         &jne    (&label("invert"));
1536
1537         &call   (&label("pic_point"));
1538         &set_label("pic_point");
1539         blindpop("ebp");
1540         &lea    ("edi",&DWP(&label("AES_Td")."-".&label("pic_point"),"ebp"));
1541         &lea    ("ebp",&DWP(&label("AES_Te")."-".&label("pic_point"),"ebp"));
1542
1543         &mov    ("esi",&wparam(2));
1544         &mov    ("ecx",&DWP(240,"esi"));        # pull number of rounds
1545         &dec    ("ecx");
1546         &align  (4);
1547         &set_label("permute");                  # permute the key schedule
1548                 &add    ("esi",16);
1549                 &deckey (0,"esi","ebp","edi");
1550                 &deckey (4,"esi","ebp","edi");
1551                 &deckey (8,"esi","ebp","edi");
1552                 &deckey (12,"esi","ebp","edi");
1553                 &dec    ("ecx");
1554         &jnz    (&label("permute"));
1555
1556         &xor    ("eax","eax");                  # return success
1557 &function_end("AES_set_decrypt_key");
1558
1559 &asm_finish();