crypto/aes/asm/aes-mips.pl

   1 #!/usr/bin/env perl
   2
   3 # ====================================================================
   4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
   5 # project. The module is, however, dual licensed under OpenSSL and
   6 # CRYPTOGAMS licenses depending on where you obtain it. For further
   7 # details see http://www.openssl.org/~appro/cryptogams/.
   8 # ====================================================================
   9
  10 # AES for MIPS
  11
  12 # October 2010
  13 #
  14 # Code uses 1K[+256B] S-box and on single-issue core [such as R5000]
  15 # spends ~68 cycles per byte processed with 128-bit key. This is ~16%
  16 # faster than gcc-generated code, which is not very impressive. But
  17 # recall that compressed S-box requires extra processing, namely
  18 # additional rotations. Rotations are implemented with lwl/lwr pairs,
  19 # which is normally used for loading unaligned data. Another cool
  20 # thing about this module is its endian neutrality, which means that
  21 # it processes data without ever changing byte order...
  22
  23 ######################################################################
  24 # There is a number of MIPS ABI in use, O32 and N32/64 are most
  25 # widely used. Then there is a new contender: NUBI. It appears that if
  26 # one picks the latter, it's possible to arrange code in ABI neutral
  27 # manner. Therefore let's stick to NUBI register layout:
  28 #
  29 ($zero,$at,$t0,$t1,$t2)=map("\$$_",(0..2,24,25));
  30 ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
  31 ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7,$s8,$s9,$s10,$s11)=map("\$$_",(12..23));
  32 ($gp,$tp,$sp,$fp,$ra)=map("\$$_",(3,28..31));
  33 #
  34 # The return value is placed in $a0. Following coding rules facilitate
  35 # interoperability:
  36 #
  37 # - never ever touch $tp, "thread pointer", former $gp;
  38 # - copy return value to $t0, former $v0 [or to $a0 if you're adapting
  39 #   old code];
  40 # - on O32 populate $a4-$a7 with 'lw $aN,4*N($sp)' if necessary;
  41 #
  42 # For reference here is register layout for N32/64 MIPS ABIs:
  43 #
  44 # ($zero,$at,$v0,$v1)=map("\$$_",(0..3));
  45 # ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
  46 # ($t0,$t1,$t2,$t3,$t8,$t9)=map("\$$_",(12..15,24,25));
  47 # ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7)=map("\$$_",(16..23));
  48 # ($gp,$sp,$fp,$ra)=map("\$$_",(28..31));
  49 #
  50 $flavour = shift; # supported flavours are o32,n32,64,nubi32,nubi64
  51
  52 if ($flavour =~ /64|n32/i) {
  53         $PTR_ADD="dadd";        # incidentally works even on n32
  54         $PTR_SUB="dsub";        # incidentally works even on n32
  55         $REG_S="sd";
  56         $REG_L="ld";
  57         $PTR_SLL="dsll";        # incidentally works even on n32
  58         $SZREG=8;
  59 } else {
  60         $PTR_ADD="add";
  61         $PTR_SUB="sub";
  62         $REG_S="sw";
  63         $REG_L="lw";
  64         $PTR_SLL="sll";
  65         $SZREG=4;
  66 }
  67 $pf = ($flavour =~ /nubi/i) ? $t0 : $t2;
  68 #
  69 # <appro@openssl.org>
  70 #
  71 ######################################################################
  72
  73 for (@ARGV) {   $big_endian=1 if (/\-DB_ENDIAN/);
  74                 $big_endian=0 if (/\-DL_ENDIAN/);
  75                 $output=$_ if (/^\w[\w\-]*\.\w+$/);     }
  76 open STDOUT,">$output";
  77
  78 if (!defined($big_endian))
  79 {    $big_endian=(unpack('L',pack('N',1))==1);   }
  80
  81 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
  82 open STDOUT,">$output";
  83
  84 my ($MSB,$LSB)=(0,3);   # automatically converted to little-endian
  85
  86 $code.=<<___;
  87 .text
  88
  89 .option pic2
  90 .set    noat
  91 ___
  92 \f
  93 {{{
  94 my $FRAMESIZE=16*$SZREG;
  95 my $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? 0xc0fff008 : 0xc0ff0000;
  96
  97 my ($inp,$out,$key,$Tbl,$s0,$s1,$s2,$s3)=($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7);
  98 my ($i0,$i1,$i2,$i3)=($at,$t0,$t1,$t2);
  99 my ($t0,$t1,$t2,$t3,$t4,$t5,$t6,$t7,$t8,$t9,$t10,$t11) = map("\$$_",(12..23));
 100 my ($key0,$cnt)=($gp,$fp);
 101
 102 # instuction ordering is "stolen" from output from MIPSpro assembler
 103 # invoked with -mips3 -O3 arguments...
 104 $code.=<<___;
 105 .align  5
 106 .ent    _mips_AES_encrypt
 107 _mips_AES_encrypt:
 108         .frame  $sp,0,$ra
 109         .set    reorder
 110         lw      $t0,0($key)
 111         lw      $t1,4($key)
 112         lw      $t2,8($key)
 113         lw      $t3,12($key)
 114         lw      $cnt,240($key)
 115         $PTR_ADD $key0,$key,16
 116
 117         xor     $s0,$t0
 118         xor     $s1,$t1
 119         xor     $s2,$t2
 120         xor     $s3,$t3
 121
 122         sub     $cnt,1
 123         _xtr    $i0,$s1,16-2
 124 .Loop_enc:
 125         _xtr    $i1,$s2,16-2
 126         _xtr    $i2,$s3,16-2
 127         _xtr    $i3,$s0,16-2
 128         and     $i0,0x3fc
 129         and     $i1,0x3fc
 130         and     $i2,0x3fc
 131         and     $i3,0x3fc
 132         $PTR_ADD $i0,$Tbl
 133         $PTR_ADD $i1,$Tbl
 134         $PTR_ADD $i2,$Tbl
 135         $PTR_ADD $i3,$Tbl
 136         lwl     $t0,3($i0)              # Te1[s1>>16]
 137         lwl     $t1,3($i1)              # Te1[s2>>16]
 138         lwl     $t2,3($i2)              # Te1[s3>>16]
 139         lwl     $t3,3($i3)              # Te1[s0>>16]
 140         lwr     $t0,2($i0)              # Te1[s1>>16]
 141         lwr     $t1,2($i1)              # Te1[s2>>16]
 142         lwr     $t2,2($i2)              # Te1[s3>>16]
 143         lwr     $t3,2($i3)              # Te1[s0>>16]
 144
 145         _xtr    $i0,$s2,8-2
 146         _xtr    $i1,$s3,8-2
 147         _xtr    $i2,$s0,8-2
 148         _xtr    $i3,$s1,8-2
 149         and     $i0,0x3fc
 150         and     $i1,0x3fc
 151         and     $i2,0x3fc
 152         and     $i3,0x3fc
 153         $PTR_ADD $i0,$Tbl
 154         $PTR_ADD $i1,$Tbl
 155         $PTR_ADD $i2,$Tbl
 156         $PTR_ADD $i3,$Tbl
 157         lwl     $t4,2($i0)              # Te2[s2>>8]
 158         lwl     $t5,2($i1)              # Te2[s3>>8]
 159         lwl     $t6,2($i2)              # Te2[s0>>8]
 160         lwl     $t7,2($i3)              # Te2[s1>>8]
 161         lwr     $t4,1($i0)              # Te2[s2>>8]
 162         lwr     $t5,1($i1)              # Te2[s3>>8]
 163         lwr     $t6,1($i2)              # Te2[s0>>8]
 164         lwr     $t7,1($i3)              # Te2[s1>>8]
 165
 166         _xtr    $i0,$s3,0-2
 167         _xtr    $i1,$s0,0-2
 168         _xtr    $i2,$s1,0-2
 169         _xtr    $i3,$s2,0-2
 170         and     $i0,0x3fc
 171         and     $i1,0x3fc
 172         and     $i2,0x3fc
 173         and     $i3,0x3fc
 174         $PTR_ADD $i0,$Tbl
 175         $PTR_ADD $i1,$Tbl
 176         $PTR_ADD $i2,$Tbl
 177         $PTR_ADD $i3,$Tbl
 178         lwl     $t8,1($i0)              # Te3[s3]
 179         lwl     $t9,1($i1)              # Te3[s0]
 180         lwl     $t10,1($i2)             # Te3[s1]
 181         lwl     $t11,1($i3)             # Te3[s2]
 182         lwr     $t8,0($i0)              # Te3[s3]
 183         lwr     $t9,0($i1)              # Te3[s0]
 184         lwr     $t10,0($i2)             # Te3[s1]
 185         lwr     $t11,0($i3)             # Te3[s2]
 186
 187         _xtr    $i0,$s0,24-2
 188         _xtr    $i1,$s1,24-2
 189         _xtr    $i2,$s2,24-2
 190         _xtr    $i3,$s3,24-2
 191         and     $i0,0x3fc
 192         and     $i1,0x3fc
 193         and     $i2,0x3fc
 194         and     $i3,0x3fc
 195         $PTR_ADD $i0,$Tbl
 196         $PTR_ADD $i1,$Tbl
 197         $PTR_ADD $i2,$Tbl
 198         $PTR_ADD $i3,$Tbl
 199         xor     $t0,$t4
 200         xor     $t1,$t5
 201         xor     $t2,$t6
 202         xor     $t3,$t7
 203         lw      $t4,0($i0)              # Te0[s0>>24]
 204         lw      $t5,0($i1)              # Te0[s1>>24]
 205         lw      $t6,0($i2)              # Te0[s2>>24]
 206         lw      $t7,0($i3)              # Te0[s3>>24]
 207
 208         lw      $s0,0($key0)
 209         lw      $s1,4($key0)
 210         lw      $s2,8($key0)
 211         lw      $s3,12($key0)
 212
 213         xor     $t0,$t8
 214         xor     $t1,$t9
 215         xor     $t2,$t10
 216         xor     $t3,$t11
 217
 218         xor     $t0,$t4
 219         xor     $t1,$t5
 220         xor     $t2,$t6
 221         xor     $t3,$t7
 222
 223         sub     $cnt,1
 224         $PTR_ADD $key0,16
 225         xor     $s0,$t0
 226         xor     $s1,$t1
 227         xor     $s2,$t2
 228         xor     $s3,$t3
 229         .set    noreorder
 230         bnez    $cnt,.Loop_enc
 231         _xtr    $i0,$s1,16-2
 232
 233         .set    reorder
 234         _xtr    $i1,$s2,16-2
 235         _xtr    $i2,$s3,16-2
 236         _xtr    $i3,$s0,16-2
 237         and     $i0,0x3fc
 238         and     $i1,0x3fc
 239         and     $i2,0x3fc
 240         and     $i3,0x3fc
 241         $PTR_ADD $i0,$Tbl
 242         $PTR_ADD $i1,$Tbl
 243         $PTR_ADD $i2,$Tbl
 244         $PTR_ADD $i3,$Tbl
 245         lbu     $t0,2($i0)              # Te4[s1>>16]
 246         lbu     $t1,2($i1)              # Te4[s2>>16]
 247         lbu     $t2,2($i2)              # Te4[s3>>16]
 248         lbu     $t3,2($i3)              # Te4[s0>>16]
 249
 250         _xtr    $i0,$s2,8-2
 251         _xtr    $i1,$s3,8-2
 252         _xtr    $i2,$s0,8-2
 253         _xtr    $i3,$s1,8-2
 254         and     $i0,0x3fc
 255         and     $i1,0x3fc
 256         and     $i2,0x3fc
 257         and     $i3,0x3fc
 258         $PTR_ADD $i0,$Tbl
 259         $PTR_ADD $i1,$Tbl
 260         $PTR_ADD $i2,$Tbl
 261         $PTR_ADD $i3,$Tbl
 262         lbu     $t4,2($i0)              # Te4[s2>>8]
 263         lbu     $t5,2($i1)              # Te4[s3>>8]
 264         lbu     $t6,2($i2)              # Te4[s0>>8]
 265         lbu     $t7,2($i3)              # Te4[s1>>8]
 266
 267         _xtr    $i0,$s0,24-2
 268         _xtr    $i1,$s1,24-2
 269         _xtr    $i2,$s2,24-2
 270         _xtr    $i3,$s3,24-2
 271         and     $i0,0x3fc
 272         and     $i1,0x3fc
 273         and     $i2,0x3fc
 274         and     $i3,0x3fc
 275         $PTR_ADD $i0,$Tbl
 276         $PTR_ADD $i1,$Tbl
 277         $PTR_ADD $i2,$Tbl
 278         $PTR_ADD $i3,$Tbl
 279         lbu     $t8,2($i0)              # Te4[s0>>24]
 280         lbu     $t9,2($i1)              # Te4[s1>>24]
 281         lbu     $t10,2($i2)             # Te4[s2>>24]
 282         lbu     $t11,2($i3)             # Te4[s3>>24]
 283
 284         _xtr    $i0,$s3,0-2
 285         _xtr    $i1,$s0,0-2
 286         _xtr    $i2,$s1,0-2
 287         _xtr    $i3,$s2,0-2
 288         and     $i0,0x3fc
 289         and     $i1,0x3fc
 290         and     $i2,0x3fc
 291         and     $i3,0x3fc
 292
 293         _ins    $t0,16
 294         _ins    $t1,16
 295         _ins    $t2,16
 296         _ins    $t3,16
 297
 298         _ins    $t4,8
 299         _ins    $t5,8
 300         _ins    $t6,8
 301         _ins    $t7,8
 302
 303         xor     $t0,$t4
 304         xor     $t1,$t5
 305         xor     $t2,$t6
 306         xor     $t3,$t7
 307
 308         $PTR_ADD $i0,$Tbl
 309         $PTR_ADD $i1,$Tbl
 310         $PTR_ADD $i2,$Tbl
 311         $PTR_ADD $i3,$Tbl
 312         lbu     $t4,2($i0)              # Te4[s3]
 313         lbu     $t5,2($i1)              # Te4[s0]
 314         lbu     $t6,2($i2)              # Te4[s1]
 315         lbu     $t7,2($i3)              # Te4[s2]
 316
 317         _ins    $t8,24
 318         _ins    $t9,24
 319         _ins    $t10,24
 320         _ins    $t11,24
 321
 322         lw      $s0,0($key0)
 323         lw      $s1,4($key0)
 324         lw      $s2,8($key0)
 325         lw      $s3,12($key0)
 326
 327         xor     $t0,$t8
 328         xor     $t1,$t9
 329         xor     $t2,$t10
 330         xor     $t3,$t11
 331
 332         _ins    $t4,0
 333         _ins    $t5,0
 334         _ins    $t6,0
 335         _ins    $t7,0
 336
 337         xor     $t0,$t4
 338         xor     $t1,$t5
 339         xor     $t2,$t6
 340         xor     $t3,$t7
 341
 342         xor     $s0,$t0
 343         xor     $s1,$t1
 344         xor     $s2,$t2
 345         xor     $s3,$t3
 346
 347         jr      $ra
 348 .end    _mips_AES_encrypt
 349
 350 .align  5
 351 .globl  AES_encrypt
 352 .ent    AES_encrypt
 353 AES_encrypt:
 354         .frame  $sp,$FRAMESIZE,$ra
 355         .mask   $SAVED_REGS_MASK,-$SZREG
 356         .set    noreorder
 357 ___
 358 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
 359         .cpload $pf
 360 ___
 361 $code.=<<___;
 362         $PTR_SUB $sp,$FRAMESIZE
 363         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
 364         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
 365         $REG_S  $s11,$FRAMESIZE-3*$SZREG($sp)
 366         $REG_S  $s10,$FRAMESIZE-4*$SZREG($sp)
 367         $REG_S  $s9,$FRAMESIZE-5*$SZREG($sp)
 368         $REG_S  $s8,$FRAMESIZE-6*$SZREG($sp)
 369         $REG_S  $s7,$FRAMESIZE-7*$SZREG($sp)
 370         $REG_S  $s6,$FRAMESIZE-8*$SZREG($sp)
 371         $REG_S  $s5,$FRAMESIZE-9*$SZREG($sp)
 372         $REG_S  $s4,$FRAMESIZE-10*$SZREG($sp)
 373 ___
 374 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
 375         $REG_S  \$15,$FRAMESIZE-11*$SZREG($sp)
 376         $REG_S  \$14,$FRAMESIZE-12*$SZREG($sp)
 377         $REG_S  \$13,$FRAMESIZE-13*$SZREG($sp)
 378         $REG_S  \$12,$FRAMESIZE-14*$SZREG($sp)
 379         $REG_S  $gp,$FRAMESIZE-15*$SZREG($sp)
 380 ___
 381 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
 382         .cplocal        $Tbl
 383         .cpsetup        $pf,$zero,AES_encrypt
 384 ___
 385 $code.=<<___;
 386         .set    reorder
 387         la      $Tbl,AES_Te             # PIC-ified 'load address'
 388
 389         lwl     $s0,0+$MSB($inp)
 390         lwl     $s1,4+$MSB($inp)
 391         lwl     $s2,8+$MSB($inp)
 392         lwl     $s3,12+$MSB($inp)
 393         lwr     $s0,0+$LSB($inp)
 394         lwr     $s1,4+$LSB($inp)
 395         lwr     $s2,8+$LSB($inp)
 396         lwr     $s3,12+$LSB($inp)
 397
 398         bal     _mips_AES_encrypt
 399
 400         swr     $s0,0+$LSB($out)
 401         swr     $s1,4+$LSB($out)
 402         swr     $s2,8+$LSB($out)
 403         swr     $s3,12+$LSB($out)
 404         swl     $s0,0+$MSB($out)
 405         swl     $s1,4+$MSB($out)
 406         swl     $s2,8+$MSB($out)
 407         swl     $s3,12+$MSB($out)
 408
 409         .set    noreorder
 410         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
 411         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
 412         $REG_L  $s11,$FRAMESIZE-3*$SZREG($sp)
 413         $REG_L  $s10,$FRAMESIZE-4*$SZREG($sp)
 414         $REG_L  $s9,$FRAMESIZE-5*$SZREG($sp)
 415         $REG_L  $s8,$FRAMESIZE-6*$SZREG($sp)
 416         $REG_L  $s7,$FRAMESIZE-7*$SZREG($sp)
 417         $REG_L  $s6,$FRAMESIZE-8*$SZREG($sp)
 418         $REG_L  $s5,$FRAMESIZE-9*$SZREG($sp)
 419         $REG_L  $s4,$FRAMESIZE-10*$SZREG($sp)
 420 ___
 421 $code.=<<___ if ($flavour =~ /nubi/i);
 422         $REG_L  \$15,$FRAMESIZE-11*$SZREG($sp)
 423         $REG_L  \$14,$FRAMESIZE-12*$SZREG($sp)
 424         $REG_L  \$13,$FRAMESIZE-13*$SZREG($sp)
 425         $REG_L  \$12,$FRAMESIZE-14*$SZREG($sp)
 426         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
 427 ___
 428 $code.=<<___;
 429         jr      $ra
 430         $PTR_ADD $sp,$FRAMESIZE
 431 .end    AES_encrypt
 432 ___
 433 \f
 434 $code.=<<___;
 435 .align  5
 436 .ent    _mips_AES_decrypt
 437 _mips_AES_decrypt:
 438         .frame  $sp,0,$ra
 439         .set    reorder
 440         lw      $t0,0($key)
 441         lw      $t1,4($key)
 442         lw      $t2,8($key)
 443         lw      $t3,12($key)
 444         lw      $cnt,240($key)
 445         $PTR_ADD $key0,$key,16
 446
 447         xor     $s0,$t0
 448         xor     $s1,$t1
 449         xor     $s2,$t2
 450         xor     $s3,$t3
 451
 452         sub     $cnt,1
 453         _xtr    $i0,$s3,16-2
 454 .Loop_dec:
 455         _xtr    $i1,$s0,16-2
 456         _xtr    $i2,$s1,16-2
 457         _xtr    $i3,$s2,16-2
 458         and     $i0,0x3fc
 459         and     $i1,0x3fc
 460         and     $i2,0x3fc
 461         and     $i3,0x3fc
 462         $PTR_ADD $i0,$Tbl
 463         $PTR_ADD $i1,$Tbl
 464         $PTR_ADD $i2,$Tbl
 465         $PTR_ADD $i3,$Tbl
 466         lwl     $t0,3($i0)              # Td1[s3>>16]
 467         lwl     $t1,3($i1)              # Td1[s0>>16]
 468         lwl     $t2,3($i2)              # Td1[s1>>16]
 469         lwl     $t3,3($i3)              # Td1[s2>>16]
 470         lwr     $t0,2($i0)              # Td1[s3>>16]
 471         lwr     $t1,2($i1)              # Td1[s0>>16]
 472         lwr     $t2,2($i2)              # Td1[s1>>16]
 473         lwr     $t3,2($i3)              # Td1[s2>>16]
 474
 475         _xtr    $i0,$s2,8-2
 476         _xtr    $i1,$s3,8-2
 477         _xtr    $i2,$s0,8-2
 478         _xtr    $i3,$s1,8-2
 479         and     $i0,0x3fc
 480         and     $i1,0x3fc
 481         and     $i2,0x3fc
 482         and     $i3,0x3fc
 483         $PTR_ADD $i0,$Tbl
 484         $PTR_ADD $i1,$Tbl
 485         $PTR_ADD $i2,$Tbl
 486         $PTR_ADD $i3,$Tbl
 487         lwl     $t4,2($i0)              # Td2[s2>>8]
 488         lwl     $t5,2($i1)              # Td2[s3>>8]
 489         lwl     $t6,2($i2)              # Td2[s0>>8]
 490         lwl     $t7,2($i3)              # Td2[s1>>8]
 491         lwr     $t4,1($i0)              # Td2[s2>>8]
 492         lwr     $t5,1($i1)              # Td2[s3>>8]
 493         lwr     $t6,1($i2)              # Td2[s0>>8]
 494         lwr     $t7,1($i3)              # Td2[s1>>8]
 495
 496         _xtr    $i0,$s1,0-2
 497         _xtr    $i1,$s2,0-2
 498         _xtr    $i2,$s3,0-2
 499         _xtr    $i3,$s0,0-2
 500         and     $i0,0x3fc
 501         and     $i1,0x3fc
 502         and     $i2,0x3fc
 503         and     $i3,0x3fc
 504         $PTR_ADD $i0,$Tbl
 505         $PTR_ADD $i1,$Tbl
 506         $PTR_ADD $i2,$Tbl
 507         $PTR_ADD $i3,$Tbl
 508         lwl     $t8,1($i0)              # Td3[s1]
 509         lwl     $t9,1($i1)              # Td3[s2]
 510         lwl     $t10,1($i2)             # Td3[s3]
 511         lwl     $t11,1($i3)             # Td3[s0]
 512         lwr     $t8,0($i0)              # Td3[s1]
 513         lwr     $t9,0($i1)              # Td3[s2]
 514         lwr     $t10,0($i2)             # Td3[s3]
 515         lwr     $t11,0($i3)             # Td3[s0]
 516
 517         _xtr    $i0,$s0,24-2
 518         _xtr    $i1,$s1,24-2
 519         _xtr    $i2,$s2,24-2
 520         _xtr    $i3,$s3,24-2
 521         and     $i0,0x3fc
 522         and     $i1,0x3fc
 523         and     $i2,0x3fc
 524         and     $i3,0x3fc
 525         $PTR_ADD $i0,$Tbl
 526         $PTR_ADD $i1,$Tbl
 527         $PTR_ADD $i2,$Tbl
 528         $PTR_ADD $i3,$Tbl
 529
 530         xor     $t0,$t4
 531         xor     $t1,$t5
 532         xor     $t2,$t6
 533         xor     $t3,$t7
 534
 535
 536         lw      $t4,0($i0)              # Td0[s0>>24]
 537         lw      $t5,0($i1)              # Td0[s1>>24]
 538         lw      $t6,0($i2)              # Td0[s2>>24]
 539         lw      $t7,0($i3)              # Td0[s3>>24]
 540
 541         lw      $s0,0($key0)
 542         lw      $s1,4($key0)
 543         lw      $s2,8($key0)
 544         lw      $s3,12($key0)
 545
 546         xor     $t0,$t8
 547         xor     $t1,$t9
 548         xor     $t2,$t10
 549         xor     $t3,$t11
 550
 551         xor     $t0,$t4
 552         xor     $t1,$t5
 553         xor     $t2,$t6
 554         xor     $t3,$t7
 555
 556         sub     $cnt,1
 557         $PTR_ADD $key0,16
 558         xor     $s0,$t0
 559         xor     $s1,$t1
 560         xor     $s2,$t2
 561         xor     $s3,$t3
 562         .set    noreorder
 563         bnez    $cnt,.Loop_dec
 564         _xtr    $i0,$s3,16-2
 565
 566         .set    reorder
 567         lw      $t4,1024($Tbl)          # prefetch Td4
 568         lw      $t5,1024+32($Tbl)
 569         lw      $t6,1024+64($Tbl)
 570         lw      $t7,1024+96($Tbl)
 571         lw      $t8,1024+128($Tbl)
 572         lw      $t9,1024+160($Tbl)
 573         lw      $t10,1024+192($Tbl)
 574         lw      $t11,1024+224($Tbl)
 575
 576         _xtr    $i0,$s3,16
 577         _xtr    $i1,$s0,16
 578         _xtr    $i2,$s1,16
 579         _xtr    $i3,$s2,16
 580         and     $i0,0xff
 581         and     $i1,0xff
 582         and     $i2,0xff
 583         and     $i3,0xff
 584         $PTR_ADD $i0,$Tbl
 585         $PTR_ADD $i1,$Tbl
 586         $PTR_ADD $i2,$Tbl
 587         $PTR_ADD $i3,$Tbl
 588         lbu     $t0,1024($i0)           # Td4[s3>>16]
 589         lbu     $t1,1024($i1)           # Td4[s0>>16]
 590         lbu     $t2,1024($i2)           # Td4[s1>>16]
 591         lbu     $t3,1024($i3)           # Td4[s2>>16]
 592
 593         _xtr    $i0,$s2,8
 594         _xtr    $i1,$s3,8
 595         _xtr    $i2,$s0,8
 596         _xtr    $i3,$s1,8
 597         and     $i0,0xff
 598         and     $i1,0xff
 599         and     $i2,0xff
 600         and     $i3,0xff
 601         $PTR_ADD $i0,$Tbl
 602         $PTR_ADD $i1,$Tbl
 603         $PTR_ADD $i2,$Tbl
 604         $PTR_ADD $i3,$Tbl
 605         lbu     $t4,1024($i0)           # Td4[s2>>8]
 606         lbu     $t5,1024($i1)           # Td4[s3>>8]
 607         lbu     $t6,1024($i2)           # Td4[s0>>8]
 608         lbu     $t7,1024($i3)           # Td4[s1>>8]
 609
 610         _xtr    $i0,$s0,24
 611         _xtr    $i1,$s1,24
 612         _xtr    $i2,$s2,24
 613         _xtr    $i3,$s3,24
 614         $PTR_ADD $i0,$Tbl
 615         $PTR_ADD $i1,$Tbl
 616         $PTR_ADD $i2,$Tbl
 617         $PTR_ADD $i3,$Tbl
 618         lbu     $t8,1024($i0)           # Td4[s0>>24]
 619         lbu     $t9,1024($i1)           # Td4[s1>>24]
 620         lbu     $t10,1024($i2)          # Td4[s2>>24]
 621         lbu     $t11,1024($i3)          # Td4[s3>>24]
 622
 623         _xtr    $i0,$s1,0
 624         _xtr    $i1,$s2,0
 625         _xtr    $i2,$s3,0
 626         _xtr    $i3,$s0,0
 627
 628         _ins    $t0,16
 629         _ins    $t1,16
 630         _ins    $t2,16
 631         _ins    $t3,16
 632
 633         _ins    $t4,8
 634         _ins    $t5,8
 635         _ins    $t6,8
 636         _ins    $t7,8
 637
 638         xor     $t0,$t4
 639         xor     $t1,$t5
 640         xor     $t2,$t6
 641         xor     $t3,$t7
 642
 643         $PTR_ADD $i0,$Tbl
 644         $PTR_ADD $i1,$Tbl
 645         $PTR_ADD $i2,$Tbl
 646         $PTR_ADD $i3,$Tbl
 647         lbu     $t4,1024($i0)           # Td4[s1]
 648         lbu     $t5,1024($i1)           # Td4[s2]
 649         lbu     $t6,1024($i2)           # Td4[s3]
 650         lbu     $t7,1024($i3)           # Td4[s0]
 651
 652         _ins    $t8,24
 653         _ins    $t9,24
 654         _ins    $t10,24
 655         _ins    $t11,24
 656
 657         lw      $s0,0($key0)
 658         lw      $s1,4($key0)
 659         lw      $s2,8($key0)
 660         lw      $s3,12($key0)
 661
 662         _ins    $t4,0
 663         _ins    $t5,0
 664         _ins    $t6,0
 665         _ins    $t7,0
 666
 667
 668         xor     $t0,$t8
 669         xor     $t1,$t9
 670         xor     $t2,$t10
 671         xor     $t3,$t11
 672
 673         xor     $t0,$t4
 674         xor     $t1,$t5
 675         xor     $t2,$t6
 676         xor     $t3,$t7
 677
 678         xor     $s0,$t0
 679         xor     $s1,$t1
 680         xor     $s2,$t2
 681         xor     $s3,$t3
 682
 683         jr      $ra
 684 .end    _mips_AES_decrypt
 685
 686 .align  5
 687 .globl  AES_decrypt
 688 .ent    AES_decrypt
 689 AES_decrypt:
 690         .frame  $sp,$FRAMESIZE,$ra
 691         .mask   $SAVED_REGS_MASK,-$SZREG
 692         .set    noreorder
 693 ___
 694 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
 695         .cpload $pf
 696 ___
 697 $code.=<<___;
 698         $PTR_SUB $sp,$FRAMESIZE
 699         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
 700         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
 701         $REG_S  $s11,$FRAMESIZE-3*$SZREG($sp)
 702         $REG_S  $s10,$FRAMESIZE-4*$SZREG($sp)
 703         $REG_S  $s9,$FRAMESIZE-5*$SZREG($sp)
 704         $REG_S  $s8,$FRAMESIZE-6*$SZREG($sp)
 705         $REG_S  $s7,$FRAMESIZE-7*$SZREG($sp)
 706         $REG_S  $s6,$FRAMESIZE-8*$SZREG($sp)
 707         $REG_S  $s5,$FRAMESIZE-9*$SZREG($sp)
 708         $REG_S  $s4,$FRAMESIZE-10*$SZREG($sp)
 709 ___
 710 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
 711         $REG_S  \$15,$FRAMESIZE-11*$SZREG($sp)
 712         $REG_S  \$14,$FRAMESIZE-12*$SZREG($sp)
 713         $REG_S  \$13,$FRAMESIZE-13*$SZREG($sp)
 714         $REG_S  \$12,$FRAMESIZE-14*$SZREG($sp)
 715         $REG_S  $gp,$FRAMESIZE-15*$SZREG($sp)
 716 ___
 717 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
 718         .cplocal        $Tbl
 719         .cpsetup        $pf,$zero,AES_decrypt
 720 ___
 721 $code.=<<___;
 722         .set    reorder
 723         la      $Tbl,AES_Td             # PIC-ified 'load address'
 724
 725         lwl     $s0,0+$MSB($inp)
 726         lwl     $s1,4+$MSB($inp)
 727         lwl     $s2,8+$MSB($inp)
 728         lwl     $s3,12+$MSB($inp)
 729         lwr     $s0,0+$LSB($inp)
 730         lwr     $s1,4+$LSB($inp)
 731         lwr     $s2,8+$LSB($inp)
 732         lwr     $s3,12+$LSB($inp)
 733
 734         bal     _mips_AES_decrypt
 735
 736         swr     $s0,0+$LSB($out)
 737         swr     $s1,4+$LSB($out)
 738         swr     $s2,8+$LSB($out)
 739         swr     $s3,12+$LSB($out)
 740         swl     $s0,0+$MSB($out)
 741         swl     $s1,4+$MSB($out)
 742         swl     $s2,8+$MSB($out)
 743         swl     $s3,12+$MSB($out)
 744
 745         .set    noreorder
 746         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
 747         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
 748         $REG_L  $s11,$FRAMESIZE-3*$SZREG($sp)
 749         $REG_L  $s10,$FRAMESIZE-4*$SZREG($sp)
 750         $REG_L  $s9,$FRAMESIZE-5*$SZREG($sp)
 751         $REG_L  $s8,$FRAMESIZE-6*$SZREG($sp)
 752         $REG_L  $s7,$FRAMESIZE-7*$SZREG($sp)
 753         $REG_L  $s6,$FRAMESIZE-8*$SZREG($sp)
 754         $REG_L  $s5,$FRAMESIZE-9*$SZREG($sp)
 755         $REG_L  $s4,$FRAMESIZE-10*$SZREG($sp)
 756 ___
 757 $code.=<<___ if ($flavour =~ /nubi/i);
 758         $REG_L  \$15,$FRAMESIZE-11*$SZREG($sp)
 759         $REG_L  \$14,$FRAMESIZE-12*$SZREG($sp)
 760         $REG_L  \$13,$FRAMESIZE-13*$SZREG($sp)
 761         $REG_L  \$12,$FRAMESIZE-14*$SZREG($sp)
 762         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
 763 ___
 764 $code.=<<___;
 765         jr      $ra
 766         $PTR_ADD $sp,$FRAMESIZE
 767 .end    AES_decrypt
 768 ___
 769 }}}
 770 \f
 771 {{{
 772 my $FRAMESIZE=8*$SZREG;
 773 my $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? 0xc000f008 : 0xc0000000;
 774
 775 my ($inp,$bits,$key,$Tbl)=($a0,$a1,$a2,$a3);
 776 my ($rk0,$rk1,$rk2,$rk3,$rk4,$rk5,$rk6,$rk7)=($a4,$a5,$a6,$a7,$s0,$s1,$s2,$s3);
 777 my ($i0,$i1,$i2,$i3)=($at,$t0,$t1,$t2);
 778 my ($rcon,$cnt)=($gp,$fp);
 779
 780 $code.=<<___;
 781 .align  5
 782 .ent    _mips_AES_set_encrypt_key
 783 _mips_AES_set_encrypt_key:
 784         .frame  $sp,0,$ra
 785         .set    noreorder
 786         beqz    $inp,.Lekey_done
 787         li      $t0,-1
 788         beqz    $key,.Lekey_done
 789         $PTR_ADD $rcon,$Tbl,1024+256
 790
 791         .set    reorder
 792         lwl     $rk0,0+$MSB($inp)       # load 128 bits
 793         lwl     $rk1,4+$MSB($inp)
 794         lwl     $rk2,8+$MSB($inp)
 795         lwl     $rk3,12+$MSB($inp)
 796         li      $at,128
 797         lwr     $rk0,0+$LSB($inp)
 798         lwr     $rk1,4+$LSB($inp)
 799         lwr     $rk2,8+$LSB($inp)
 800         lwr     $rk3,12+$LSB($inp)
 801         .set    noreorder
 802         beq     $bits,$at,.L128bits
 803         li      $cnt,10
 804
 805         .set    reorder
 806         lwl     $rk4,16+$MSB($inp)      # load 192 bits
 807         lwl     $rk5,20+$MSB($inp)
 808         li      $at,192
 809         lwr     $rk4,16+$LSB($inp)
 810         lwr     $rk5,20+$LSB($inp)
 811         .set    noreorder
 812         beq     $bits,$at,.L192bits
 813         li      $cnt,8
 814
 815         .set    reorder
 816         lwl     $rk6,24+$MSB($inp)      # load 256 bits
 817         lwl     $rk7,28+$MSB($inp)
 818         li      $at,256
 819         lwr     $rk6,24+$LSB($inp)
 820         lwr     $rk7,28+$LSB($inp)
 821         .set    noreorder
 822         beq     $bits,$at,.L256bits
 823         li      $cnt,7
 824
 825         b       .Lekey_done
 826         li      $t0,-2
 827
 828 .align  4
 829 .L128bits:
 830         .set    reorder
 831         srl     $i0,$rk3,16
 832         srl     $i1,$rk3,8
 833         and     $i0,0xff
 834         and     $i1,0xff
 835         and     $i2,$rk3,0xff
 836         srl     $i3,$rk3,24
 837         $PTR_ADD $i0,$Tbl
 838         $PTR_ADD $i1,$Tbl
 839         $PTR_ADD $i2,$Tbl
 840         $PTR_ADD $i3,$Tbl
 841         lbu     $i0,1024($i0)
 842         lbu     $i1,1024($i1)
 843         lbu     $i2,1024($i2)
 844         lbu     $i3,1024($i3)
 845
 846         sw      $rk0,0($key)
 847         sw      $rk1,4($key)
 848         sw      $rk2,8($key)
 849         sw      $rk3,12($key)
 850         sub     $cnt,1
 851         $PTR_ADD $key,16
 852
 853         _bias   $i0,24
 854         _bias   $i1,16
 855         _bias   $i2,8
 856         _bias   $i3,0
 857
 858         xor     $rk0,$i0
 859         lw      $i0,0($rcon)
 860         xor     $rk0,$i1
 861         xor     $rk0,$i2
 862         xor     $rk0,$i3
 863         xor     $rk0,$i0
 864
 865         xor     $rk1,$rk0
 866         xor     $rk2,$rk1
 867         xor     $rk3,$rk2
 868
 869         .set    noreorder
 870         bnez    $cnt,.L128bits
 871         $PTR_ADD $rcon,4
 872
 873         sw      $rk0,0($key)
 874         sw      $rk1,4($key)
 875         sw      $rk2,8($key)
 876         li      $cnt,10
 877         sw      $rk3,12($key)
 878         li      $t0,0
 879         sw      $cnt,80($key)
 880         b       .Lekey_done
 881         $PTR_SUB $key,10*16
 882
 883 .align  4
 884 .L192bits:
 885         .set    reorder
 886         srl     $i0,$rk5,16
 887         srl     $i1,$rk5,8
 888         and     $i0,0xff
 889         and     $i1,0xff
 890         and     $i2,$rk5,0xff
 891         srl     $i3,$rk5,24
 892         $PTR_ADD $i0,$Tbl
 893         $PTR_ADD $i1,$Tbl
 894         $PTR_ADD $i2,$Tbl
 895         $PTR_ADD $i3,$Tbl
 896         lbu     $i0,1024($i0)
 897         lbu     $i1,1024($i1)
 898         lbu     $i2,1024($i2)
 899         lbu     $i3,1024($i3)
 900
 901         sw      $rk0,0($key)
 902         sw      $rk1,4($key)
 903         sw      $rk2,8($key)
 904         sw      $rk3,12($key)
 905         sw      $rk4,16($key)
 906         sw      $rk5,20($key)
 907         sub     $cnt,1
 908         $PTR_ADD $key,24
 909
 910         _bias   $i0,24
 911         _bias   $i1,16
 912         _bias   $i2,8
 913         _bias   $i3,0
 914
 915         xor     $rk0,$i0
 916         lw      $i0,0($rcon)
 917         xor     $rk0,$i1
 918         xor     $rk0,$i2
 919         xor     $rk0,$i3
 920         xor     $rk0,$i0
 921
 922         xor     $rk1,$rk0
 923         xor     $rk2,$rk1
 924         xor     $rk3,$rk2
 925         xor     $rk4,$rk3
 926         xor     $rk5,$rk4
 927
 928         .set    noreorder
 929         bnez    $cnt,.L192bits
 930         $PTR_ADD $rcon,4
 931
 932         sw      $rk0,0($key)
 933         sw      $rk1,4($key)
 934         sw      $rk2,8($key)
 935         li      $cnt,12
 936         sw      $rk3,12($key)
 937         li      $t0,0
 938         sw      $cnt,48($key)
 939         b       .Lekey_done
 940         $PTR_SUB $key,12*16
 941
 942 .align  4
 943 .L256bits:
 944         .set    reorder
 945         srl     $i0,$rk7,16
 946         srl     $i1,$rk7,8
 947         and     $i0,0xff
 948         and     $i1,0xff
 949         and     $i2,$rk7,0xff
 950         srl     $i3,$rk7,24
 951         $PTR_ADD $i0,$Tbl
 952         $PTR_ADD $i1,$Tbl
 953         $PTR_ADD $i2,$Tbl
 954         $PTR_ADD $i3,$Tbl
 955         lbu     $i0,1024($i0)
 956         lbu     $i1,1024($i1)
 957         lbu     $i2,1024($i2)
 958         lbu     $i3,1024($i3)
 959
 960         sw      $rk0,0($key)
 961         sw      $rk1,4($key)
 962         sw      $rk2,8($key)
 963         sw      $rk3,12($key)
 964         sw      $rk4,16($key)
 965         sw      $rk5,20($key)
 966         sw      $rk6,24($key)
 967         sw      $rk7,28($key)
 968         sub     $cnt,1
 969
 970         _bias   $i0,24
 971         _bias   $i1,16
 972         _bias   $i2,8
 973         _bias   $i3,0
 974
 975         xor     $rk0,$i0
 976         lw      $i0,0($rcon)
 977         xor     $rk0,$i1
 978         xor     $rk0,$i2
 979         xor     $rk0,$i3
 980         xor     $rk0,$i0
 981
 982         xor     $rk1,$rk0
 983         xor     $rk2,$rk1
 984         xor     $rk3,$rk2
 985         beqz    $cnt,.L256bits_done
 986
 987         srl     $i0,$rk3,24
 988         srl     $i1,$rk3,16
 989         srl     $i2,$rk3,8
 990         and     $i3,$rk3,0xff
 991         and     $i1,0xff
 992         and     $i2,0xff
 993         $PTR_ADD $i0,$Tbl
 994         $PTR_ADD $i1,$Tbl
 995         $PTR_ADD $i2,$Tbl
 996         $PTR_ADD $i3,$Tbl
 997         lbu     $i0,1024($i0)
 998         lbu     $i1,1024($i1)
 999         lbu     $i2,1024($i2)
1000         lbu     $i3,1024($i3)
1001         sll     $i0,24
1002         sll     $i1,16
1003         sll     $i2,8
1004
1005         xor     $rk4,$i0
1006         xor     $rk4,$i1
1007         xor     $rk4,$i2
1008         xor     $rk4,$i3
1009
1010         xor     $rk5,$rk4
1011         xor     $rk6,$rk5
1012         xor     $rk7,$rk6
1013
1014         $PTR_ADD $key,32
1015         .set    noreorder
1016         b       .L256bits
1017         $PTR_ADD $rcon,4
1018
1019 .L256bits_done:
1020         sw      $rk0,32($key)
1021         sw      $rk1,36($key)
1022         sw      $rk2,40($key)
1023         li      $cnt,14
1024         sw      $rk3,44($key)
1025         li      $t0,0
1026         sw      $cnt,48($key)
1027         $PTR_SUB $key,12*16
1028
1029 .Lekey_done:
1030         jr      $ra
1031         nop
1032 .end    _mips_AES_set_encrypt_key
1033
1034 .globl  AES_set_encrypt_key
1035 .ent    AES_set_encrypt_key
1036 AES_set_encrypt_key:
1037         .frame  $sp,$FRAMESIZE,$ra
1038         .mask   $SAVED_REGS_MASK,-$SZREG
1039         .set    noreorder
1040 ___
1041 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
1042         .cpload $pf
1043 ___
1044 $code.=<<___;
1045         $PTR_SUB $sp,$FRAMESIZE
1046         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
1047         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
1048 ___
1049 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
1050         $REG_S  $s3,$FRAMESIZE-3*$SZREG($sp)
1051         $REG_S  $s2,$FRAMESIZE-4*$SZREG($sp)
1052         $REG_S  $s1,$FRAMESIZE-5*$SZREG($sp)
1053         $REG_S  $s0,$FRAMESIZE-6*$SZREG($sp)
1054         $REG_S  $gp,$FRAMESIZE-7*$SZREG($sp)
1055 ___
1056 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
1057         .cplocal        $Tbl
1058         .cpsetup        $pf,$zero,AES_set_encrypt_key
1059 ___
1060 $code.=<<___;
1061         .set    reorder
1062         la      $Tbl,AES_Te             # PIC-ified 'load address'
1063
1064         bal     _mips_AES_set_encrypt_key
1065
1066         .set    noreorder
1067         move    $a0,$t0
1068         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
1069         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
1070 ___
1071 $code.=<<___ if ($flavour =~ /nubi/i);
1072         $REG_L  $s3,$FRAMESIZE-11*$SZREG($sp)
1073         $REG_L  $s2,$FRAMESIZE-12*$SZREG($sp)
1074         $REG_L  $s1,$FRAMESIZE-13*$SZREG($sp)
1075         $REG_L  $s0,$FRAMESIZE-14*$SZREG($sp)
1076         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
1077 ___
1078 $code.=<<___;
1079         jr      $ra
1080         $PTR_ADD $sp,$FRAMESIZE
1081 .end    AES_set_encrypt_key
1082 ___
1083 \f
1084 my ($head,$tail)=($inp,$bits);
1085 my ($tp1,$tp2,$tp4,$tp8,$tp9,$tpb,$tpd,$tpe)=($a4,$a5,$a6,$a7,$s0,$s1,$s2,$s3);
1086 my ($m,$x80808080,$x7f7f7f7f,$x1b1b1b1b)=($at,$t0,$t1,$t2);
1087 $code.=<<___;
1088 .align  5
1089 .globl  AES_set_decrypt_key
1090 .ent    AES_set_decrypt_key
1091 AES_set_decrypt_key:
1092         .frame  $sp,$FRAMESIZE,$ra
1093         .mask   $SAVED_REGS_MASK,-$SZREG
1094         .set    noreorder
1095 ___
1096 $code.=<<___ if ($flavour =~ /o32/i);   # o32 PIC-ification
1097         .cpload $pf
1098 ___
1099 $code.=<<___;
1100         $PTR_SUB $sp,$FRAMESIZE
1101         $REG_S  $ra,$FRAMESIZE-1*$SZREG($sp)
1102         $REG_S  $fp,$FRAMESIZE-2*$SZREG($sp)
1103 ___
1104 $code.=<<___ if ($flavour =~ /nubi/i);  # optimize non-nubi prologue
1105         $REG_S  $s3,$FRAMESIZE-3*$SZREG($sp)
1106         $REG_S  $s2,$FRAMESIZE-4*$SZREG($sp)
1107         $REG_S  $s1,$FRAMESIZE-5*$SZREG($sp)
1108         $REG_S  $s0,$FRAMESIZE-6*$SZREG($sp)
1109         $REG_S  $gp,$FRAMESIZE-7*$SZREG($sp)
1110 ___
1111 $code.=<<___ if ($flavour !~ /o32/i);   # non-o32 PIC-ification
1112         .cplocal        $Tbl
1113         .cpsetup        $pf,$zero,AES_set_decrypt_key
1114 ___
1115 $code.=<<___;
1116         .set    reorder
1117         la      $Tbl,AES_Te             # PIC-ified 'load address'
1118
1119         bal     _mips_AES_set_encrypt_key
1120
1121         bltz    $t0,.Ldkey_done
1122
1123         sll     $at,$cnt,4
1124         $PTR_ADD $head,$key,0
1125         $PTR_ADD $tail,$key,$at
1126 .align  4
1127 .Lswap:
1128         lw      $rk0,0($head)
1129         lw      $rk1,4($head)
1130         lw      $rk2,8($head)
1131         lw      $rk3,12($head)
1132         lw      $rk4,0($tail)
1133         lw      $rk5,4($tail)
1134         lw      $rk6,8($tail)
1135         lw      $rk7,12($tail)
1136         sw      $rk0,0($tail)
1137         sw      $rk1,4($tail)
1138         sw      $rk2,8($tail)
1139         sw      $rk3,12($tail)
1140         $PTR_ADD $head,16
1141         $PTR_SUB $tail,16
1142         sw      $rk4,-16($head)
1143         sw      $rk5,-12($head)
1144         sw      $rk6,-8($head)
1145         sw      $rk7,-4($head)
1146         bne     $head,$tail,.Lswap
1147
1148         lw      $tp1,16($key)           # modulo-scheduled
1149         lui     $x80808080,0x8080
1150         sub     $cnt,1
1151         or      $x80808080,0x8080
1152         sll     $cnt,2
1153         $PTR_ADD $key,16
1154         lui     $x1b1b1b1b,0x1b1b
1155         nor     $x7f7f7f7f,$zero,$x80808080
1156         or      $x1b1b1b1b,0x1b1b
1157 .align  4
1158 .Lmix:
1159         and     $m,$tp1,$x80808080
1160         and     $tp2,$tp1,$x7f7f7f7f
1161         srl     $tp4,$m,7
1162         addu    $tp2,$tp2               # tp2<<1
1163         subu    $m,$tp4
1164         and     $m,$x1b1b1b1b
1165         xor     $tp2,$m
1166
1167         and     $m,$tp2,$x80808080
1168         and     $tp4,$tp2,$x7f7f7f7f
1169         srl     $tp8,$m,7
1170         addu    $tp4,$tp4               # tp4<<1
1171         subu    $m,$tp8
1172         and     $m,$x1b1b1b1b
1173         xor     $tp4,$m
1174
1175         and     $m,$tp4,$x80808080
1176         and     $tp8,$tp4,$x7f7f7f7f
1177         srl     $tp9,$m,7
1178         addu    $tp8,$tp8               # tp8<<1
1179         subu    $m,$tp9
1180         and     $m,$x1b1b1b1b
1181         xor     $tp8,$m
1182
1183         xor     $tp9,$tp8,$tp1
1184         xor     $tpe,$tp8,$tp4
1185         xor     $tpb,$tp9,$tp2
1186         xor     $tpd,$tp9,$tp4
1187
1188         _ror    $tp1,$tpd,16
1189          xor    $tpe,$tp2
1190         _ror    $tp2,$tpd,-16
1191         xor     $tpe,$tp1
1192         _ror    $tp1,$tp9,8
1193         xor     $tpe,$tp2
1194         _ror    $tp2,$tp9,-24
1195         xor     $tpe,$tp1
1196         _ror    $tp1,$tpb,24
1197         xor     $tpe,$tp2
1198         _ror    $tp2,$tpb,-8
1199         xor     $tpe,$tp1
1200         lw      $tp1,4($key)            # modulo-scheduled
1201         xor     $tpe,$tp2
1202         sub     $cnt,1
1203         sw      $tpe,0($key)
1204         $PTR_ADD $key,4
1205         bnez    $cnt,.Lmix
1206
1207         li      $t0,0
1208 .Ldkey_done:
1209         .set    noreorder
1210         move    $a0,$t0
1211         $REG_L  $ra,$FRAMESIZE-1*$SZREG($sp)
1212         $REG_L  $fp,$FRAMESIZE-2*$SZREG($sp)
1213 ___
1214 $code.=<<___ if ($flavour =~ /nubi/i);
1215         $REG_L  $s3,$FRAMESIZE-11*$SZREG($sp)
1216         $REG_L  $s2,$FRAMESIZE-12*$SZREG($sp)
1217         $REG_L  $s1,$FRAMESIZE-13*$SZREG($sp)
1218         $REG_L  $s0,$FRAMESIZE-14*$SZREG($sp)
1219         $REG_L  $gp,$FRAMESIZE-15*$SZREG($sp)
1220 ___
1221 $code.=<<___;
1222         jr      $ra
1223         $PTR_ADD $sp,$FRAMESIZE
1224 .end    AES_set_decrypt_key
1225 ___
1226 }}}
1227
1228 ######################################################################
1229 # Tables are kept in endian-neutral manner
1230 $code.=<<___;
1231 .rdata
1232 .align  6
1233 AES_Te:
1234 .byte   0xc6,0x63,0x63,0xa5,    0xf8,0x7c,0x7c,0x84     # Te0
1235 .byte   0xee,0x77,0x77,0x99,    0xf6,0x7b,0x7b,0x8d
1236 .byte   0xff,0xf2,0xf2,0x0d,    0xd6,0x6b,0x6b,0xbd
1237 .byte   0xde,0x6f,0x6f,0xb1,    0x91,0xc5,0xc5,0x54
1238 .byte   0x60,0x30,0x30,0x50,    0x02,0x01,0x01,0x03
1239 .byte   0xce,0x67,0x67,0xa9,    0x56,0x2b,0x2b,0x7d
1240 .byte   0xe7,0xfe,0xfe,0x19,    0xb5,0xd7,0xd7,0x62
1241 .byte   0x4d,0xab,0xab,0xe6,    0xec,0x76,0x76,0x9a
1242 .byte   0x8f,0xca,0xca,0x45,    0x1f,0x82,0x82,0x9d
1243 .byte   0x89,0xc9,0xc9,0x40,    0xfa,0x7d,0x7d,0x87
1244 .byte   0xef,0xfa,0xfa,0x15,    0xb2,0x59,0x59,0xeb
1245 .byte   0x8e,0x47,0x47,0xc9,    0xfb,0xf0,0xf0,0x0b
1246 .byte   0x41,0xad,0xad,0xec,    0xb3,0xd4,0xd4,0x67
1247 .byte   0x5f,0xa2,0xa2,0xfd,    0x45,0xaf,0xaf,0xea
1248 .byte   0x23,0x9c,0x9c,0xbf,    0x53,0xa4,0xa4,0xf7
1249 .byte   0xe4,0x72,0x72,0x96,    0x9b,0xc0,0xc0,0x5b
1250 .byte   0x75,0xb7,0xb7,0xc2,    0xe1,0xfd,0xfd,0x1c
1251 .byte   0x3d,0x93,0x93,0xae,    0x4c,0x26,0x26,0x6a
1252 .byte   0x6c,0x36,0x36,0x5a,    0x7e,0x3f,0x3f,0x41
1253 .byte   0xf5,0xf7,0xf7,0x02,    0x83,0xcc,0xcc,0x4f
1254 .byte   0x68,0x34,0x34,0x5c,    0x51,0xa5,0xa5,0xf4
1255 .byte   0xd1,0xe5,0xe5,0x34,    0xf9,0xf1,0xf1,0x08
1256 .byte   0xe2,0x71,0x71,0x93,    0xab,0xd8,0xd8,0x73
1257 .byte   0x62,0x31,0x31,0x53,    0x2a,0x15,0x15,0x3f
1258 .byte   0x08,0x04,0x04,0x0c,    0x95,0xc7,0xc7,0x52
1259 .byte   0x46,0x23,0x23,0x65,    0x9d,0xc3,0xc3,0x5e
1260 .byte   0x30,0x18,0x18,0x28,    0x37,0x96,0x96,0xa1
1261 .byte   0x0a,0x05,0x05,0x0f,    0x2f,0x9a,0x9a,0xb5
1262 .byte   0x0e,0x07,0x07,0x09,    0x24,0x12,0x12,0x36
1263 .byte   0x1b,0x80,0x80,0x9b,    0xdf,0xe2,0xe2,0x3d
1264 .byte   0xcd,0xeb,0xeb,0x26,    0x4e,0x27,0x27,0x69
1265 .byte   0x7f,0xb2,0xb2,0xcd,    0xea,0x75,0x75,0x9f
1266 .byte   0x12,0x09,0x09,0x1b,    0x1d,0x83,0x83,0x9e
1267 .byte   0x58,0x2c,0x2c,0x74,    0x34,0x1a,0x1a,0x2e
1268 .byte   0x36,0x1b,0x1b,0x2d,    0xdc,0x6e,0x6e,0xb2
1269 .byte   0xb4,0x5a,0x5a,0xee,    0x5b,0xa0,0xa0,0xfb
1270 .byte   0xa4,0x52,0x52,0xf6,    0x76,0x3b,0x3b,0x4d
1271 .byte   0xb7,0xd6,0xd6,0x61,    0x7d,0xb3,0xb3,0xce
1272 .byte   0x52,0x29,0x29,0x7b,    0xdd,0xe3,0xe3,0x3e
1273 .byte   0x5e,0x2f,0x2f,0x71,    0x13,0x84,0x84,0x97
1274 .byte   0xa6,0x53,0x53,0xf5,    0xb9,0xd1,0xd1,0x68
1275 .byte   0x00,0x00,0x00,0x00,    0xc1,0xed,0xed,0x2c
1276 .byte   0x40,0x20,0x20,0x60,    0xe3,0xfc,0xfc,0x1f
1277 .byte   0x79,0xb1,0xb1,0xc8,    0xb6,0x5b,0x5b,0xed
1278 .byte   0xd4,0x6a,0x6a,0xbe,    0x8d,0xcb,0xcb,0x46
1279 .byte   0x67,0xbe,0xbe,0xd9,    0x72,0x39,0x39,0x4b
1280 .byte   0x94,0x4a,0x4a,0xde,    0x98,0x4c,0x4c,0xd4
1281 .byte   0xb0,0x58,0x58,0xe8,    0x85,0xcf,0xcf,0x4a
1282 .byte   0xbb,0xd0,0xd0,0x6b,    0xc5,0xef,0xef,0x2a
1283 .byte   0x4f,0xaa,0xaa,0xe5,    0xed,0xfb,0xfb,0x16
1284 .byte   0x86,0x43,0x43,0xc5,    0x9a,0x4d,0x4d,0xd7
1285 .byte   0x66,0x33,0x33,0x55,    0x11,0x85,0x85,0x94
1286 .byte   0x8a,0x45,0x45,0xcf,    0xe9,0xf9,0xf9,0x10
1287 .byte   0x04,0x02,0x02,0x06,    0xfe,0x7f,0x7f,0x81
1288 .byte   0xa0,0x50,0x50,0xf0,    0x78,0x3c,0x3c,0x44
1289 .byte   0x25,0x9f,0x9f,0xba,    0x4b,0xa8,0xa8,0xe3
1290 .byte   0xa2,0x51,0x51,0xf3,    0x5d,0xa3,0xa3,0xfe
1291 .byte   0x80,0x40,0x40,0xc0,    0x05,0x8f,0x8f,0x8a
1292 .byte   0x3f,0x92,0x92,0xad,    0x21,0x9d,0x9d,0xbc
1293 .byte   0x70,0x38,0x38,0x48,    0xf1,0xf5,0xf5,0x04
1294 .byte   0x63,0xbc,0xbc,0xdf,    0x77,0xb6,0xb6,0xc1
1295 .byte   0xaf,0xda,0xda,0x75,    0x42,0x21,0x21,0x63
1296 .byte   0x20,0x10,0x10,0x30,    0xe5,0xff,0xff,0x1a
1297 .byte   0xfd,0xf3,0xf3,0x0e,    0xbf,0xd2,0xd2,0x6d
1298 .byte   0x81,0xcd,0xcd,0x4c,    0x18,0x0c,0x0c,0x14
1299 .byte   0x26,0x13,0x13,0x35,    0xc3,0xec,0xec,0x2f
1300 .byte   0xbe,0x5f,0x5f,0xe1,    0x35,0x97,0x97,0xa2
1301 .byte   0x88,0x44,0x44,0xcc,    0x2e,0x17,0x17,0x39
1302 .byte   0x93,0xc4,0xc4,0x57,    0x55,0xa7,0xa7,0xf2
1303 .byte   0xfc,0x7e,0x7e,0x82,    0x7a,0x3d,0x3d,0x47
1304 .byte   0xc8,0x64,0x64,0xac,    0xba,0x5d,0x5d,0xe7
1305 .byte   0x32,0x19,0x19,0x2b,    0xe6,0x73,0x73,0x95
1306 .byte   0xc0,0x60,0x60,0xa0,    0x19,0x81,0x81,0x98
1307 .byte   0x9e,0x4f,0x4f,0xd1,    0xa3,0xdc,0xdc,0x7f
1308 .byte   0x44,0x22,0x22,0x66,    0x54,0x2a,0x2a,0x7e
1309 .byte   0x3b,0x90,0x90,0xab,    0x0b,0x88,0x88,0x83
1310 .byte   0x8c,0x46,0x46,0xca,    0xc7,0xee,0xee,0x29
1311 .byte   0x6b,0xb8,0xb8,0xd3,    0x28,0x14,0x14,0x3c
1312 .byte   0xa7,0xde,0xde,0x79,    0xbc,0x5e,0x5e,0xe2
1313 .byte   0x16,0x0b,0x0b,0x1d,    0xad,0xdb,0xdb,0x76
1314 .byte   0xdb,0xe0,0xe0,0x3b,    0x64,0x32,0x32,0x56
1315 .byte   0x74,0x3a,0x3a,0x4e,    0x14,0x0a,0x0a,0x1e
1316 .byte   0x92,0x49,0x49,0xdb,    0x0c,0x06,0x06,0x0a
1317 .byte   0x48,0x24,0x24,0x6c,    0xb8,0x5c,0x5c,0xe4
1318 .byte   0x9f,0xc2,0xc2,0x5d,    0xbd,0xd3,0xd3,0x6e
1319 .byte   0x43,0xac,0xac,0xef,    0xc4,0x62,0x62,0xa6
1320 .byte   0x39,0x91,0x91,0xa8,    0x31,0x95,0x95,0xa4
1321 .byte   0xd3,0xe4,0xe4,0x37,    0xf2,0x79,0x79,0x8b
1322 .byte   0xd5,0xe7,0xe7,0x32,    0x8b,0xc8,0xc8,0x43
1323 .byte   0x6e,0x37,0x37,0x59,    0xda,0x6d,0x6d,0xb7
1324 .byte   0x01,0x8d,0x8d,0x8c,    0xb1,0xd5,0xd5,0x64
1325 .byte   0x9c,0x4e,0x4e,0xd2,    0x49,0xa9,0xa9,0xe0
1326 .byte   0xd8,0x6c,0x6c,0xb4,    0xac,0x56,0x56,0xfa
1327 .byte   0xf3,0xf4,0xf4,0x07,    0xcf,0xea,0xea,0x25
1328 .byte   0xca,0x65,0x65,0xaf,    0xf4,0x7a,0x7a,0x8e
1329 .byte   0x47,0xae,0xae,0xe9,    0x10,0x08,0x08,0x18
1330 .byte   0x6f,0xba,0xba,0xd5,    0xf0,0x78,0x78,0x88
1331 .byte   0x4a,0x25,0x25,0x6f,    0x5c,0x2e,0x2e,0x72
1332 .byte   0x38,0x1c,0x1c,0x24,    0x57,0xa6,0xa6,0xf1
1333 .byte   0x73,0xb4,0xb4,0xc7,    0x97,0xc6,0xc6,0x51
1334 .byte   0xcb,0xe8,0xe8,0x23,    0xa1,0xdd,0xdd,0x7c
1335 .byte   0xe8,0x74,0x74,0x9c,    0x3e,0x1f,0x1f,0x21
1336 .byte   0x96,0x4b,0x4b,0xdd,    0x61,0xbd,0xbd,0xdc
1337 .byte   0x0d,0x8b,0x8b,0x86,    0x0f,0x8a,0x8a,0x85
1338 .byte   0xe0,0x70,0x70,0x90,    0x7c,0x3e,0x3e,0x42
1339 .byte   0x71,0xb5,0xb5,0xc4,    0xcc,0x66,0x66,0xaa
1340 .byte   0x90,0x48,0x48,0xd8,    0x06,0x03,0x03,0x05
1341 .byte   0xf7,0xf6,0xf6,0x01,    0x1c,0x0e,0x0e,0x12
1342 .byte   0xc2,0x61,0x61,0xa3,    0x6a,0x35,0x35,0x5f
1343 .byte   0xae,0x57,0x57,0xf9,    0x69,0xb9,0xb9,0xd0
1344 .byte   0x17,0x86,0x86,0x91,    0x99,0xc1,0xc1,0x58
1345 .byte   0x3a,0x1d,0x1d,0x27,    0x27,0x9e,0x9e,0xb9
1346 .byte   0xd9,0xe1,0xe1,0x38,    0xeb,0xf8,0xf8,0x13
1347 .byte   0x2b,0x98,0x98,0xb3,    0x22,0x11,0x11,0x33
1348 .byte   0xd2,0x69,0x69,0xbb,    0xa9,0xd9,0xd9,0x70
1349 .byte   0x07,0x8e,0x8e,0x89,    0x33,0x94,0x94,0xa7
1350 .byte   0x2d,0x9b,0x9b,0xb6,    0x3c,0x1e,0x1e,0x22
1351 .byte   0x15,0x87,0x87,0x92,    0xc9,0xe9,0xe9,0x20
1352 .byte   0x87,0xce,0xce,0x49,    0xaa,0x55,0x55,0xff
1353 .byte   0x50,0x28,0x28,0x78,    0xa5,0xdf,0xdf,0x7a
1354 .byte   0x03,0x8c,0x8c,0x8f,    0x59,0xa1,0xa1,0xf8
1355 .byte   0x09,0x89,0x89,0x80,    0x1a,0x0d,0x0d,0x17
1356 .byte   0x65,0xbf,0xbf,0xda,    0xd7,0xe6,0xe6,0x31
1357 .byte   0x84,0x42,0x42,0xc6,    0xd0,0x68,0x68,0xb8
1358 .byte   0x82,0x41,0x41,0xc3,    0x29,0x99,0x99,0xb0
1359 .byte   0x5a,0x2d,0x2d,0x77,    0x1e,0x0f,0x0f,0x11
1360 .byte   0x7b,0xb0,0xb0,0xcb,    0xa8,0x54,0x54,0xfc
1361 .byte   0x6d,0xbb,0xbb,0xd6,    0x2c,0x16,0x16,0x3a
1362
1363 .byte   0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5  # Te4
1364 .byte   0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
1365 .byte   0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
1366 .byte   0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
1367 .byte   0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
1368 .byte   0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
1369 .byte   0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
1370 .byte   0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
1371 .byte   0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
1372 .byte   0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
1373 .byte   0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
1374 .byte   0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
1375 .byte   0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
1376 .byte   0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
1377 .byte   0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
1378 .byte   0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
1379 .byte   0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
1380 .byte   0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
1381 .byte   0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
1382 .byte   0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
1383 .byte   0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
1384 .byte   0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
1385 .byte   0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
1386 .byte   0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
1387 .byte   0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
1388 .byte   0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
1389 .byte   0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
1390 .byte   0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
1391 .byte   0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
1392 .byte   0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
1393 .byte   0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
1394 .byte   0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
1395
1396 .byte   0x01,0x00,0x00,0x00,    0x02,0x00,0x00,0x00     # rcon
1397 .byte   0x04,0x00,0x00,0x00,    0x08,0x00,0x00,0x00
1398 .byte   0x10,0x00,0x00,0x00,    0x20,0x00,0x00,0x00
1399 .byte   0x40,0x00,0x00,0x00,    0x80,0x00,0x00,0x00
1400 .byte   0x1B,0x00,0x00,0x00,    0x36,0x00,0x00,0x00
1401
1402 .align  6
1403 AES_Td:
1404 .byte   0x51,0xf4,0xa7,0x50,    0x7e,0x41,0x65,0x53     # Td0
1405 .byte   0x1a,0x17,0xa4,0xc3,    0x3a,0x27,0x5e,0x96
1406 .byte   0x3b,0xab,0x6b,0xcb,    0x1f,0x9d,0x45,0xf1
1407 .byte   0xac,0xfa,0x58,0xab,    0x4b,0xe3,0x03,0x93
1408 .byte   0x20,0x30,0xfa,0x55,    0xad,0x76,0x6d,0xf6
1409 .byte   0x88,0xcc,0x76,0x91,    0xf5,0x02,0x4c,0x25
1410 .byte   0x4f,0xe5,0xd7,0xfc,    0xc5,0x2a,0xcb,0xd7
1411 .byte   0x26,0x35,0x44,0x80,    0xb5,0x62,0xa3,0x8f
1412 .byte   0xde,0xb1,0x5a,0x49,    0x25,0xba,0x1b,0x67
1413 .byte   0x45,0xea,0x0e,0x98,    0x5d,0xfe,0xc0,0xe1
1414 .byte   0xc3,0x2f,0x75,0x02,    0x81,0x4c,0xf0,0x12
1415 .byte   0x8d,0x46,0x97,0xa3,    0x6b,0xd3,0xf9,0xc6
1416 .byte   0x03,0x8f,0x5f,0xe7,    0x15,0x92,0x9c,0x95
1417 .byte   0xbf,0x6d,0x7a,0xeb,    0x95,0x52,0x59,0xda
1418 .byte   0xd4,0xbe,0x83,0x2d,    0x58,0x74,0x21,0xd3
1419 .byte   0x49,0xe0,0x69,0x29,    0x8e,0xc9,0xc8,0x44
1420 .byte   0x75,0xc2,0x89,0x6a,    0xf4,0x8e,0x79,0x78
1421 .byte   0x99,0x58,0x3e,0x6b,    0x27,0xb9,0x71,0xdd
1422 .byte   0xbe,0xe1,0x4f,0xb6,    0xf0,0x88,0xad,0x17
1423 .byte   0xc9,0x20,0xac,0x66,    0x7d,0xce,0x3a,0xb4
1424 .byte   0x63,0xdf,0x4a,0x18,    0xe5,0x1a,0x31,0x82
1425 .byte   0x97,0x51,0x33,0x60,    0x62,0x53,0x7f,0x45
1426 .byte   0xb1,0x64,0x77,0xe0,    0xbb,0x6b,0xae,0x84
1427 .byte   0xfe,0x81,0xa0,0x1c,    0xf9,0x08,0x2b,0x94
1428 .byte   0x70,0x48,0x68,0x58,    0x8f,0x45,0xfd,0x19
1429 .byte   0x94,0xde,0x6c,0x87,    0x52,0x7b,0xf8,0xb7
1430 .byte   0xab,0x73,0xd3,0x23,    0x72,0x4b,0x02,0xe2
1431 .byte   0xe3,0x1f,0x8f,0x57,    0x66,0x55,0xab,0x2a
1432 .byte   0xb2,0xeb,0x28,0x07,    0x2f,0xb5,0xc2,0x03
1433 .byte   0x86,0xc5,0x7b,0x9a,    0xd3,0x37,0x08,0xa5
1434 .byte   0x30,0x28,0x87,0xf2,    0x23,0xbf,0xa5,0xb2
1435 .byte   0x02,0x03,0x6a,0xba,    0xed,0x16,0x82,0x5c
1436 .byte   0x8a,0xcf,0x1c,0x2b,    0xa7,0x79,0xb4,0x92
1437 .byte   0xf3,0x07,0xf2,0xf0,    0x4e,0x69,0xe2,0xa1
1438 .byte   0x65,0xda,0xf4,0xcd,    0x06,0x05,0xbe,0xd5
1439 .byte   0xd1,0x34,0x62,0x1f,    0xc4,0xa6,0xfe,0x8a
1440 .byte   0x34,0x2e,0x53,0x9d,    0xa2,0xf3,0x55,0xa0
1441 .byte   0x05,0x8a,0xe1,0x32,    0xa4,0xf6,0xeb,0x75
1442 .byte   0x0b,0x83,0xec,0x39,    0x40,0x60,0xef,0xaa
1443 .byte   0x5e,0x71,0x9f,0x06,    0xbd,0x6e,0x10,0x51
1444 .byte   0x3e,0x21,0x8a,0xf9,    0x96,0xdd,0x06,0x3d
1445 .byte   0xdd,0x3e,0x05,0xae,    0x4d,0xe6,0xbd,0x46
1446 .byte   0x91,0x54,0x8d,0xb5,    0x71,0xc4,0x5d,0x05
1447 .byte   0x04,0x06,0xd4,0x6f,    0x60,0x50,0x15,0xff
1448 .byte   0x19,0x98,0xfb,0x24,    0xd6,0xbd,0xe9,0x97
1449 .byte   0x89,0x40,0x43,0xcc,    0x67,0xd9,0x9e,0x77
1450 .byte   0xb0,0xe8,0x42,0xbd,    0x07,0x89,0x8b,0x88
1451 .byte   0xe7,0x19,0x5b,0x38,    0x79,0xc8,0xee,0xdb
1452 .byte   0xa1,0x7c,0x0a,0x47,    0x7c,0x42,0x0f,0xe9
1453 .byte   0xf8,0x84,0x1e,0xc9,    0x00,0x00,0x00,0x00
1454 .byte   0x09,0x80,0x86,0x83,    0x32,0x2b,0xed,0x48
1455 .byte   0x1e,0x11,0x70,0xac,    0x6c,0x5a,0x72,0x4e
1456 .byte   0xfd,0x0e,0xff,0xfb,    0x0f,0x85,0x38,0x56
1457 .byte   0x3d,0xae,0xd5,0x1e,    0x36,0x2d,0x39,0x27
1458 .byte   0x0a,0x0f,0xd9,0x64,    0x68,0x5c,0xa6,0x21
1459 .byte   0x9b,0x5b,0x54,0xd1,    0x24,0x36,0x2e,0x3a
1460 .byte   0x0c,0x0a,0x67,0xb1,    0x93,0x57,0xe7,0x0f
1461 .byte   0xb4,0xee,0x96,0xd2,    0x1b,0x9b,0x91,0x9e
1462 .byte   0x80,0xc0,0xc5,0x4f,    0x61,0xdc,0x20,0xa2
1463 .byte   0x5a,0x77,0x4b,0x69,    0x1c,0x12,0x1a,0x16
1464 .byte   0xe2,0x93,0xba,0x0a,    0xc0,0xa0,0x2a,0xe5
1465 .byte   0x3c,0x22,0xe0,0x43,    0x12,0x1b,0x17,0x1d
1466 .byte   0x0e,0x09,0x0d,0x0b,    0xf2,0x8b,0xc7,0xad
1467 .byte   0x2d,0xb6,0xa8,0xb9,    0x14,0x1e,0xa9,0xc8
1468 .byte   0x57,0xf1,0x19,0x85,    0xaf,0x75,0x07,0x4c
1469 .byte   0xee,0x99,0xdd,0xbb,    0xa3,0x7f,0x60,0xfd
1470 .byte   0xf7,0x01,0x26,0x9f,    0x5c,0x72,0xf5,0xbc
1471 .byte   0x44,0x66,0x3b,0xc5,    0x5b,0xfb,0x7e,0x34
1472 .byte   0x8b,0x43,0x29,0x76,    0xcb,0x23,0xc6,0xdc
1473 .byte   0xb6,0xed,0xfc,0x68,    0xb8,0xe4,0xf1,0x63
1474 .byte   0xd7,0x31,0xdc,0xca,    0x42,0x63,0x85,0x10
1475 .byte   0x13,0x97,0x22,0x40,    0x84,0xc6,0x11,0x20
1476 .byte   0x85,0x4a,0x24,0x7d,    0xd2,0xbb,0x3d,0xf8
1477 .byte   0xae,0xf9,0x32,0x11,    0xc7,0x29,0xa1,0x6d
1478 .byte   0x1d,0x9e,0x2f,0x4b,    0xdc,0xb2,0x30,0xf3
1479 .byte   0x0d,0x86,0x52,0xec,    0x77,0xc1,0xe3,0xd0
1480 .byte   0x2b,0xb3,0x16,0x6c,    0xa9,0x70,0xb9,0x99
1481 .byte   0x11,0x94,0x48,0xfa,    0x47,0xe9,0x64,0x22
1482 .byte   0xa8,0xfc,0x8c,0xc4,    0xa0,0xf0,0x3f,0x1a
1483 .byte   0x56,0x7d,0x2c,0xd8,    0x22,0x33,0x90,0xef
1484 .byte   0x87,0x49,0x4e,0xc7,    0xd9,0x38,0xd1,0xc1
1485 .byte   0x8c,0xca,0xa2,0xfe,    0x98,0xd4,0x0b,0x36
1486 .byte   0xa6,0xf5,0x81,0xcf,    0xa5,0x7a,0xde,0x28
1487 .byte   0xda,0xb7,0x8e,0x26,    0x3f,0xad,0xbf,0xa4
1488 .byte   0x2c,0x3a,0x9d,0xe4,    0x50,0x78,0x92,0x0d
1489 .byte   0x6a,0x5f,0xcc,0x9b,    0x54,0x7e,0x46,0x62
1490 .byte   0xf6,0x8d,0x13,0xc2,    0x90,0xd8,0xb8,0xe8
1491 .byte   0x2e,0x39,0xf7,0x5e,    0x82,0xc3,0xaf,0xf5
1492 .byte   0x9f,0x5d,0x80,0xbe,    0x69,0xd0,0x93,0x7c
1493 .byte   0x6f,0xd5,0x2d,0xa9,    0xcf,0x25,0x12,0xb3
1494 .byte   0xc8,0xac,0x99,0x3b,    0x10,0x18,0x7d,0xa7
1495 .byte   0xe8,0x9c,0x63,0x6e,    0xdb,0x3b,0xbb,0x7b
1496 .byte   0xcd,0x26,0x78,0x09,    0x6e,0x59,0x18,0xf4
1497 .byte   0xec,0x9a,0xb7,0x01,    0x83,0x4f,0x9a,0xa8
1498 .byte   0xe6,0x95,0x6e,0x65,    0xaa,0xff,0xe6,0x7e
1499 .byte   0x21,0xbc,0xcf,0x08,    0xef,0x15,0xe8,0xe6
1500 .byte   0xba,0xe7,0x9b,0xd9,    0x4a,0x6f,0x36,0xce
1501 .byte   0xea,0x9f,0x09,0xd4,    0x29,0xb0,0x7c,0xd6
1502 .byte   0x31,0xa4,0xb2,0xaf,    0x2a,0x3f,0x23,0x31
1503 .byte   0xc6,0xa5,0x94,0x30,    0x35,0xa2,0x66,0xc0
1504 .byte   0x74,0x4e,0xbc,0x37,    0xfc,0x82,0xca,0xa6
1505 .byte   0xe0,0x90,0xd0,0xb0,    0x33,0xa7,0xd8,0x15
1506 .byte   0xf1,0x04,0x98,0x4a,    0x41,0xec,0xda,0xf7
1507 .byte   0x7f,0xcd,0x50,0x0e,    0x17,0x91,0xf6,0x2f
1508 .byte   0x76,0x4d,0xd6,0x8d,    0x43,0xef,0xb0,0x4d
1509 .byte   0xcc,0xaa,0x4d,0x54,    0xe4,0x96,0x04,0xdf
1510 .byte   0x9e,0xd1,0xb5,0xe3,    0x4c,0x6a,0x88,0x1b
1511 .byte   0xc1,0x2c,0x1f,0xb8,    0x46,0x65,0x51,0x7f
1512 .byte   0x9d,0x5e,0xea,0x04,    0x01,0x8c,0x35,0x5d
1513 .byte   0xfa,0x87,0x74,0x73,    0xfb,0x0b,0x41,0x2e
1514 .byte   0xb3,0x67,0x1d,0x5a,    0x92,0xdb,0xd2,0x52
1515 .byte   0xe9,0x10,0x56,0x33,    0x6d,0xd6,0x47,0x13
1516 .byte   0x9a,0xd7,0x61,0x8c,    0x37,0xa1,0x0c,0x7a
1517 .byte   0x59,0xf8,0x14,0x8e,    0xeb,0x13,0x3c,0x89
1518 .byte   0xce,0xa9,0x27,0xee,    0xb7,0x61,0xc9,0x35
1519 .byte   0xe1,0x1c,0xe5,0xed,    0x7a,0x47,0xb1,0x3c
1520 .byte   0x9c,0xd2,0xdf,0x59,    0x55,0xf2,0x73,0x3f
1521 .byte   0x18,0x14,0xce,0x79,    0x73,0xc7,0x37,0xbf
1522 .byte   0x53,0xf7,0xcd,0xea,    0x5f,0xfd,0xaa,0x5b
1523 .byte   0xdf,0x3d,0x6f,0x14,    0x78,0x44,0xdb,0x86
1524 .byte   0xca,0xaf,0xf3,0x81,    0xb9,0x68,0xc4,0x3e
1525 .byte   0x38,0x24,0x34,0x2c,    0xc2,0xa3,0x40,0x5f
1526 .byte   0x16,0x1d,0xc3,0x72,    0xbc,0xe2,0x25,0x0c
1527 .byte   0x28,0x3c,0x49,0x8b,    0xff,0x0d,0x95,0x41
1528 .byte   0x39,0xa8,0x01,0x71,    0x08,0x0c,0xb3,0xde
1529 .byte   0xd8,0xb4,0xe4,0x9c,    0x64,0x56,0xc1,0x90
1530 .byte   0x7b,0xcb,0x84,0x61,    0xd5,0x32,0xb6,0x70
1531 .byte   0x48,0x6c,0x5c,0x74,    0xd0,0xb8,0x57,0x42
1532
1533 .byte   0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38  # Td4
1534 .byte   0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
1535 .byte   0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
1536 .byte   0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
1537 .byte   0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
1538 .byte   0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
1539 .byte   0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
1540 .byte   0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
1541 .byte   0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
1542 .byte   0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
1543 .byte   0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
1544 .byte   0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
1545 .byte   0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
1546 .byte   0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
1547 .byte   0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
1548 .byte   0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
1549 .byte   0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
1550 .byte   0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
1551 .byte   0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
1552 .byte   0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
1553 .byte   0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
1554 .byte   0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
1555 .byte   0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
1556 .byte   0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
1557 .byte   0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
1558 .byte   0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
1559 .byte   0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
1560 .byte   0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
1561 .byte   0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
1562 .byte   0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
1563 .byte   0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
1564 .byte   0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
1565 ___
1566 \f
1567 foreach (split("\n",$code)) {
1568         s/\`([^\`]*)\`/eval $1/ge;
1569
1570         # made-up _instructions, _xtr, _ins, _ror and _bias, cope
1571         # with byte order dependencies...
1572         if (/^\s+_/) {
1573             s/(_[a-z]+\s+)(\$[0-9]+),([^,]+)(#.*)*$/$1$2,$2,$3/;
1574
1575             s/_xtr\s+(\$[0-9]+),(\$[0-9]+),([0-9]+(\-2)*)/
1576                 sprintf("srl\t$1,$2,%d",$big_endian ?   eval($3)
1577                                         :               eval("24-$3"))/e or
1578             s/_ins\s+(\$[0-9]+),(\$[0-9]+),([0-9]+)/
1579                 sprintf("sll\t$1,$2,%d",$big_endian ?   eval($3)
1580                                         :               eval("24-$3"))/e or
1581             s/_ror\s+(\$[0-9]+),(\$[0-9]+),(\-?[0-9]+)/
1582                 sprintf("srl\t$1,$2,%d",$big_endian ?   eval($3)
1583                                         :               eval("$3*-1"))/e or
1584             s/_bias\s+(\$[0-9]+),(\$[0-9]+),([0-9]+)/
1585                 sprintf("sll\t$1,$2,%d",$big_endian ?   eval($3)
1586                                         :               eval("($3-16)&31"))/e;
1587
1588             s/srl\s+(\$[0-9]+),(\$[0-9]+),\-([0-9]+)/
1589                 sprintf("sll\t$1,$2,$3")/e                              or
1590             s/srl\s+(\$[0-9]+),(\$[0-9]+),0/
1591                 sprintf("and\t$1,$2,0xff")/e                            or
1592             s/(sll\s+\$[0-9]+,\$[0-9]+,0)/#$1/;
1593         }
1594
1595         # convert lwl/lwr and swr/swl to little-endian order
1596         if (!$big_endian && /^\s+[sl]w[lr]\s+/) {
1597             s/([sl]wl.*)([0-9]+)\((\$[0-9]+)\)/
1598                 sprintf("$1%d($3)",eval("$2-$2%4+($2%4-1)&3"))/e        or
1599             s/([sl]wr.*)([0-9]+)\((\$[0-9]+)\)/
1600                 sprintf("$1%d($3)",eval("$2-$2%4+($2%4+1)&3"))/e;
1601         }
1602
1603         print $_,"\n";
1604 }
1605
1606 close STDOUT;