crypto/aes/asm/aes-parisc.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2009-2020 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the Apache License 2.0 (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9
  10 # ====================================================================
  11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
  12 # project. The module is, however, dual licensed under OpenSSL and
  13 # CRYPTOGAMS licenses depending on where you obtain it. For further
  14 # details see http://www.openssl.org/~appro/cryptogams/.
  15 # ====================================================================
  16
  17 # AES for PA-RISC.
  18 #
  19 # June 2009.
  20 #
  21 # The module is mechanical transliteration of aes-sparcv9.pl, but with
  22 # a twist: S-boxes are compressed even further down to 1K+256B. On
  23 # PA-7100LC performance is ~40% better than gcc 3.2 generated code and
  24 # is about 33 cycles per byte processed with 128-bit key. Newer CPUs
  25 # perform at 16 cycles per byte. It's not faster than code generated
  26 # by vendor compiler, but recall that it has compressed S-boxes, which
  27 # requires extra processing.
  28 #
  29 # Special thanks to polarhome.com for providing HP-UX account.
  30
  31 # $output is the last argument if it looks like a file (it has an extension)
  32 # $flavour is the first argument if it doesn't look like a file
  33 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
  34 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
  35
  36 $output and open STDOUT,">$output";
  37
  38 if ($flavour =~ /64/) {
  39         $LEVEL          ="2.0W";
  40         $SIZE_T         =8;
  41         $FRAME_MARKER   =80;
  42         $SAVED_RP       =16;
  43         $PUSH           ="std";
  44         $PUSHMA         ="std,ma";
  45         $POP            ="ldd";
  46         $POPMB          ="ldd,mb";
  47 } else {
  48         $LEVEL          ="1.0";
  49         $SIZE_T         =4;
  50         $FRAME_MARKER   =48;
  51         $SAVED_RP       =20;
  52         $PUSH           ="stw";
  53         $PUSHMA         ="stwm";
  54         $POP            ="ldw";
  55         $POPMB          ="ldwm";
  56 }
  57
  58 $FRAME=16*$SIZE_T+$FRAME_MARKER;# 16 saved regs + frame marker
  59                                 #                 [+ argument transfer]
  60 $inp="%r26";    # arg0
  61 $out="%r25";    # arg1
  62 $key="%r24";    # arg2
  63
  64 ($s0,$s1,$s2,$s3) = ("%r1","%r2","%r3","%r4");
  65 ($t0,$t1,$t2,$t3) = ("%r5","%r6","%r7","%r8");
  66
  67 ($acc0, $acc1, $acc2, $acc3, $acc4, $acc5, $acc6, $acc7,
  68  $acc8, $acc9,$acc10,$acc11,$acc12,$acc13,$acc14,$acc15) =
  69 ("%r9","%r10","%r11","%r12","%r13","%r14","%r15","%r16",
  70 "%r17","%r18","%r19","%r20","%r21","%r22","%r23","%r26");
  71
  72 $tbl="%r28";
  73 $rounds="%r29";
  74
  75 $code=<<___;
  76         .LEVEL  $LEVEL
  77         .SPACE  \$TEXT\$
  78         .SUBSPA \$CODE\$,QUAD=0,ALIGN=8,ACCESS=0x2C,CODE_ONLY
  79
  80         .EXPORT AES_encrypt,ENTRY,ARGW0=GR,ARGW1=GR,ARGW2=GR
  81         .ALIGN  64
  82 AES_encrypt
  83         .PROC
  84         .CALLINFO       FRAME=`$FRAME-16*$SIZE_T`,NO_CALLS,SAVE_RP,ENTRY_GR=18
  85         .ENTRY
  86         $PUSH   %r2,-$SAVED_RP(%sp)     ; standard prologue
  87         $PUSHMA %r3,$FRAME(%sp)
  88         $PUSH   %r4,`-$FRAME+1*$SIZE_T`(%sp)
  89         $PUSH   %r5,`-$FRAME+2*$SIZE_T`(%sp)
  90         $PUSH   %r6,`-$FRAME+3*$SIZE_T`(%sp)
  91         $PUSH   %r7,`-$FRAME+4*$SIZE_T`(%sp)
  92         $PUSH   %r8,`-$FRAME+5*$SIZE_T`(%sp)
  93         $PUSH   %r9,`-$FRAME+6*$SIZE_T`(%sp)
  94         $PUSH   %r10,`-$FRAME+7*$SIZE_T`(%sp)
  95         $PUSH   %r11,`-$FRAME+8*$SIZE_T`(%sp)
  96         $PUSH   %r12,`-$FRAME+9*$SIZE_T`(%sp)
  97         $PUSH   %r13,`-$FRAME+10*$SIZE_T`(%sp)
  98         $PUSH   %r14,`-$FRAME+11*$SIZE_T`(%sp)
  99         $PUSH   %r15,`-$FRAME+12*$SIZE_T`(%sp)
 100         $PUSH   %r16,`-$FRAME+13*$SIZE_T`(%sp)
 101         $PUSH   %r17,`-$FRAME+14*$SIZE_T`(%sp)
 102         $PUSH   %r18,`-$FRAME+15*$SIZE_T`(%sp)
 103
 104         blr     %r0,$tbl
 105         ldi     3,$t0
 106 L\$enc_pic
 107         andcm   $tbl,$t0,$tbl
 108         ldo     L\$AES_Te-L\$enc_pic($tbl),$tbl
 109
 110         and     $inp,$t0,$t0
 111         sub     $inp,$t0,$inp
 112         ldw     0($inp),$s0
 113         ldw     4($inp),$s1
 114         ldw     8($inp),$s2
 115         comib,= 0,$t0,L\$enc_inp_aligned
 116         ldw     12($inp),$s3
 117
 118         sh3addl $t0,%r0,$t0
 119         subi    32,$t0,$t0
 120         mtctl   $t0,%cr11
 121         ldw     16($inp),$t1
 122         vshd    $s0,$s1,$s0
 123         vshd    $s1,$s2,$s1
 124         vshd    $s2,$s3,$s2
 125         vshd    $s3,$t1,$s3
 126
 127 L\$enc_inp_aligned
 128         bl      _parisc_AES_encrypt,%r31
 129         nop
 130
 131         extru,<> $out,31,2,%r0
 132         b       L\$enc_out_aligned
 133         nop
 134
 135         _srm    $s0,24,$acc0
 136         _srm    $s0,16,$acc1
 137         stb     $acc0,0($out)
 138         _srm    $s0,8,$acc2
 139         stb     $acc1,1($out)
 140         _srm    $s1,24,$acc4
 141         stb     $acc2,2($out)
 142         _srm    $s1,16,$acc5
 143         stb     $s0,3($out)
 144         _srm    $s1,8,$acc6
 145         stb     $acc4,4($out)
 146         _srm    $s2,24,$acc0
 147         stb     $acc5,5($out)
 148         _srm    $s2,16,$acc1
 149         stb     $acc6,6($out)
 150         _srm    $s2,8,$acc2
 151         stb     $s1,7($out)
 152         _srm    $s3,24,$acc4
 153         stb     $acc0,8($out)
 154         _srm    $s3,16,$acc5
 155         stb     $acc1,9($out)
 156         _srm    $s3,8,$acc6
 157         stb     $acc2,10($out)
 158         stb     $s2,11($out)
 159         stb     $acc4,12($out)
 160         stb     $acc5,13($out)
 161         stb     $acc6,14($out)
 162         b       L\$enc_done
 163         stb     $s3,15($out)
 164
 165 L\$enc_out_aligned
 166         stw     $s0,0($out)
 167         stw     $s1,4($out)
 168         stw     $s2,8($out)
 169         stw     $s3,12($out)
 170
 171 L\$enc_done
 172         $POP    `-$FRAME-$SAVED_RP`(%sp),%r2    ; standard epilogue
 173         $POP    `-$FRAME+1*$SIZE_T`(%sp),%r4
 174         $POP    `-$FRAME+2*$SIZE_T`(%sp),%r5
 175         $POP    `-$FRAME+3*$SIZE_T`(%sp),%r6
 176         $POP    `-$FRAME+4*$SIZE_T`(%sp),%r7
 177         $POP    `-$FRAME+5*$SIZE_T`(%sp),%r8
 178         $POP    `-$FRAME+6*$SIZE_T`(%sp),%r9
 179         $POP    `-$FRAME+7*$SIZE_T`(%sp),%r10
 180         $POP    `-$FRAME+8*$SIZE_T`(%sp),%r11
 181         $POP    `-$FRAME+9*$SIZE_T`(%sp),%r12
 182         $POP    `-$FRAME+10*$SIZE_T`(%sp),%r13
 183         $POP    `-$FRAME+11*$SIZE_T`(%sp),%r14
 184         $POP    `-$FRAME+12*$SIZE_T`(%sp),%r15
 185         $POP    `-$FRAME+13*$SIZE_T`(%sp),%r16
 186         $POP    `-$FRAME+14*$SIZE_T`(%sp),%r17
 187         $POP    `-$FRAME+15*$SIZE_T`(%sp),%r18
 188         bv      (%r2)
 189         .EXIT
 190         $POPMB  -$FRAME(%sp),%r3
 191         .PROCEND
 192
 193         .ALIGN  16
 194 _parisc_AES_encrypt
 195         .PROC
 196         .CALLINFO       MILLICODE
 197         .ENTRY
 198         ldw     240($key),$rounds
 199         ldw     0($key),$t0
 200         ldw     4($key),$t1
 201         ldw     8($key),$t2
 202         _srm    $rounds,1,$rounds
 203         xor     $t0,$s0,$s0
 204         ldw     12($key),$t3
 205         _srm    $s0,24,$acc0
 206         xor     $t1,$s1,$s1
 207         ldw     16($key),$t0
 208         _srm    $s1,16,$acc1
 209         xor     $t2,$s2,$s2
 210         ldw     20($key),$t1
 211         xor     $t3,$s3,$s3
 212         ldw     24($key),$t2
 213         ldw     28($key),$t3
 214 L\$enc_loop
 215         _srm    $s2,8,$acc2
 216         ldwx,s  $acc0($tbl),$acc0
 217         _srm    $s3,0,$acc3
 218         ldwx,s  $acc1($tbl),$acc1
 219         _srm    $s1,24,$acc4
 220         ldwx,s  $acc2($tbl),$acc2
 221         _srm    $s2,16,$acc5
 222         ldwx,s  $acc3($tbl),$acc3
 223         _srm    $s3,8,$acc6
 224         ldwx,s  $acc4($tbl),$acc4
 225         _srm    $s0,0,$acc7
 226         ldwx,s  $acc5($tbl),$acc5
 227         _srm    $s2,24,$acc8
 228         ldwx,s  $acc6($tbl),$acc6
 229         _srm    $s3,16,$acc9
 230         ldwx,s  $acc7($tbl),$acc7
 231         _srm    $s0,8,$acc10
 232         ldwx,s  $acc8($tbl),$acc8
 233         _srm    $s1,0,$acc11
 234         ldwx,s  $acc9($tbl),$acc9
 235         _srm    $s3,24,$acc12
 236         ldwx,s  $acc10($tbl),$acc10
 237         _srm    $s0,16,$acc13
 238         ldwx,s  $acc11($tbl),$acc11
 239         _srm    $s1,8,$acc14
 240         ldwx,s  $acc12($tbl),$acc12
 241         _srm    $s2,0,$acc15
 242         ldwx,s  $acc13($tbl),$acc13
 243         ldwx,s  $acc14($tbl),$acc14
 244         ldwx,s  $acc15($tbl),$acc15
 245         addib,= -1,$rounds,L\$enc_last
 246         ldo     32($key),$key
 247
 248                 _ror    $acc1,8,$acc1
 249                 xor     $acc0,$t0,$t0
 250         ldw     0($key),$s0
 251                 _ror    $acc2,16,$acc2
 252                 xor     $acc1,$t0,$t0
 253         ldw     4($key),$s1
 254                 _ror    $acc3,24,$acc3
 255                 xor     $acc2,$t0,$t0
 256         ldw     8($key),$s2
 257                 _ror    $acc5,8,$acc5
 258                 xor     $acc3,$t0,$t0
 259         ldw     12($key),$s3
 260                 _ror    $acc6,16,$acc6
 261                 xor     $acc4,$t1,$t1
 262                 _ror    $acc7,24,$acc7
 263                 xor     $acc5,$t1,$t1
 264                 _ror    $acc9,8,$acc9
 265                 xor     $acc6,$t1,$t1
 266                 _ror    $acc10,16,$acc10
 267                 xor     $acc7,$t1,$t1
 268                 _ror    $acc11,24,$acc11
 269                 xor     $acc8,$t2,$t2
 270                 _ror    $acc13,8,$acc13
 271                 xor     $acc9,$t2,$t2
 272                 _ror    $acc14,16,$acc14
 273                 xor     $acc10,$t2,$t2
 274                 _ror    $acc15,24,$acc15
 275                 xor     $acc11,$t2,$t2
 276                 xor     $acc12,$acc14,$acc14
 277                 xor     $acc13,$t3,$t3
 278         _srm    $t0,24,$acc0
 279                 xor     $acc14,$t3,$t3
 280         _srm    $t1,16,$acc1
 281                 xor     $acc15,$t3,$t3
 282
 283         _srm    $t2,8,$acc2
 284         ldwx,s  $acc0($tbl),$acc0
 285         _srm    $t3,0,$acc3
 286         ldwx,s  $acc1($tbl),$acc1
 287         _srm    $t1,24,$acc4
 288         ldwx,s  $acc2($tbl),$acc2
 289         _srm    $t2,16,$acc5
 290         ldwx,s  $acc3($tbl),$acc3
 291         _srm    $t3,8,$acc6
 292         ldwx,s  $acc4($tbl),$acc4
 293         _srm    $t0,0,$acc7
 294         ldwx,s  $acc5($tbl),$acc5
 295         _srm    $t2,24,$acc8
 296         ldwx,s  $acc6($tbl),$acc6
 297         _srm    $t3,16,$acc9
 298         ldwx,s  $acc7($tbl),$acc7
 299         _srm    $t0,8,$acc10
 300         ldwx,s  $acc8($tbl),$acc8
 301         _srm    $t1,0,$acc11
 302         ldwx,s  $acc9($tbl),$acc9
 303         _srm    $t3,24,$acc12
 304         ldwx,s  $acc10($tbl),$acc10
 305         _srm    $t0,16,$acc13
 306         ldwx,s  $acc11($tbl),$acc11
 307         _srm    $t1,8,$acc14
 308         ldwx,s  $acc12($tbl),$acc12
 309         _srm    $t2,0,$acc15
 310         ldwx,s  $acc13($tbl),$acc13
 311                 _ror    $acc1,8,$acc1
 312         ldwx,s  $acc14($tbl),$acc14
 313
 314                 _ror    $acc2,16,$acc2
 315                 xor     $acc0,$s0,$s0
 316         ldwx,s  $acc15($tbl),$acc15
 317                 _ror    $acc3,24,$acc3
 318                 xor     $acc1,$s0,$s0
 319         ldw     16($key),$t0
 320                 _ror    $acc5,8,$acc5
 321                 xor     $acc2,$s0,$s0
 322         ldw     20($key),$t1
 323                 _ror    $acc6,16,$acc6
 324                 xor     $acc3,$s0,$s0
 325         ldw     24($key),$t2
 326                 _ror    $acc7,24,$acc7
 327                 xor     $acc4,$s1,$s1
 328         ldw     28($key),$t3
 329                 _ror    $acc9,8,$acc9
 330                 xor     $acc5,$s1,$s1
 331         ldw     1024+0($tbl),%r0                ; prefetch te4
 332                 _ror    $acc10,16,$acc10
 333                 xor     $acc6,$s1,$s1
 334         ldw     1024+32($tbl),%r0               ; prefetch te4
 335                 _ror    $acc11,24,$acc11
 336                 xor     $acc7,$s1,$s1
 337         ldw     1024+64($tbl),%r0               ; prefetch te4
 338                 _ror    $acc13,8,$acc13
 339                 xor     $acc8,$s2,$s2
 340         ldw     1024+96($tbl),%r0               ; prefetch te4
 341                 _ror    $acc14,16,$acc14
 342                 xor     $acc9,$s2,$s2
 343         ldw     1024+128($tbl),%r0              ; prefetch te4
 344                 _ror    $acc15,24,$acc15
 345                 xor     $acc10,$s2,$s2
 346         ldw     1024+160($tbl),%r0              ; prefetch te4
 347         _srm    $s0,24,$acc0
 348                 xor     $acc11,$s2,$s2
 349         ldw     1024+192($tbl),%r0              ; prefetch te4
 350                 xor     $acc12,$acc14,$acc14
 351                 xor     $acc13,$s3,$s3
 352         ldw     1024+224($tbl),%r0              ; prefetch te4
 353         _srm    $s1,16,$acc1
 354                 xor     $acc14,$s3,$s3
 355         b       L\$enc_loop
 356                 xor     $acc15,$s3,$s3
 357
 358         .ALIGN  16
 359 L\$enc_last
 360         ldo     1024($tbl),$rounds
 361                 _ror    $acc1,8,$acc1
 362                 xor     $acc0,$t0,$t0
 363         ldw     0($key),$s0
 364                 _ror    $acc2,16,$acc2
 365                 xor     $acc1,$t0,$t0
 366         ldw     4($key),$s1
 367                 _ror    $acc3,24,$acc3
 368                 xor     $acc2,$t0,$t0
 369         ldw     8($key),$s2
 370                 _ror    $acc5,8,$acc5
 371                 xor     $acc3,$t0,$t0
 372         ldw     12($key),$s3
 373                 _ror    $acc6,16,$acc6
 374                 xor     $acc4,$t1,$t1
 375                 _ror    $acc7,24,$acc7
 376                 xor     $acc5,$t1,$t1
 377                 _ror    $acc9,8,$acc9
 378                 xor     $acc6,$t1,$t1
 379                 _ror    $acc10,16,$acc10
 380                 xor     $acc7,$t1,$t1
 381                 _ror    $acc11,24,$acc11
 382                 xor     $acc8,$t2,$t2
 383                 _ror    $acc13,8,$acc13
 384                 xor     $acc9,$t2,$t2
 385                 _ror    $acc14,16,$acc14
 386                 xor     $acc10,$t2,$t2
 387                 _ror    $acc15,24,$acc15
 388                 xor     $acc11,$t2,$t2
 389                 xor     $acc12,$acc14,$acc14
 390                 xor     $acc13,$t3,$t3
 391         _srm    $t0,24,$acc0
 392                 xor     $acc14,$t3,$t3
 393         _srm    $t1,16,$acc1
 394                 xor     $acc15,$t3,$t3
 395
 396         _srm    $t2,8,$acc2
 397         ldbx    $acc0($rounds),$acc0
 398         _srm    $t1,24,$acc4
 399         ldbx    $acc1($rounds),$acc1
 400         _srm    $t2,16,$acc5
 401         _srm    $t3,0,$acc3
 402         ldbx    $acc2($rounds),$acc2
 403         ldbx    $acc3($rounds),$acc3
 404         _srm    $t3,8,$acc6
 405         ldbx    $acc4($rounds),$acc4
 406         _srm    $t2,24,$acc8
 407         ldbx    $acc5($rounds),$acc5
 408         _srm    $t3,16,$acc9
 409         _srm    $t0,0,$acc7
 410         ldbx    $acc6($rounds),$acc6
 411         ldbx    $acc7($rounds),$acc7
 412         _srm    $t0,8,$acc10
 413         ldbx    $acc8($rounds),$acc8
 414         _srm    $t3,24,$acc12
 415         ldbx    $acc9($rounds),$acc9
 416         _srm    $t0,16,$acc13
 417         _srm    $t1,0,$acc11
 418         ldbx    $acc10($rounds),$acc10
 419         _srm    $t1,8,$acc14
 420         ldbx    $acc11($rounds),$acc11
 421         ldbx    $acc12($rounds),$acc12
 422         ldbx    $acc13($rounds),$acc13
 423         _srm    $t2,0,$acc15
 424         ldbx    $acc14($rounds),$acc14
 425
 426                 dep     $acc0,7,8,$acc3
 427         ldbx    $acc15($rounds),$acc15
 428                 dep     $acc4,7,8,$acc7
 429                 dep     $acc1,15,8,$acc3
 430                 dep     $acc5,15,8,$acc7
 431                 dep     $acc2,23,8,$acc3
 432                 dep     $acc6,23,8,$acc7
 433                 xor     $acc3,$s0,$s0
 434                 xor     $acc7,$s1,$s1
 435                 dep     $acc8,7,8,$acc11
 436                 dep     $acc12,7,8,$acc15
 437                 dep     $acc9,15,8,$acc11
 438                 dep     $acc13,15,8,$acc15
 439                 dep     $acc10,23,8,$acc11
 440                 dep     $acc14,23,8,$acc15
 441                 xor     $acc11,$s2,$s2
 442
 443         bv      (%r31)
 444         .EXIT
 445                 xor     $acc15,$s3,$s3
 446         .PROCEND
 447
 448         .ALIGN  64
 449 L\$AES_Te
 450         .WORD   0xc66363a5, 0xf87c7c84, 0xee777799, 0xf67b7b8d
 451         .WORD   0xfff2f20d, 0xd66b6bbd, 0xde6f6fb1, 0x91c5c554
 452         .WORD   0x60303050, 0x02010103, 0xce6767a9, 0x562b2b7d
 453         .WORD   0xe7fefe19, 0xb5d7d762, 0x4dababe6, 0xec76769a
 454         .WORD   0x8fcaca45, 0x1f82829d, 0x89c9c940, 0xfa7d7d87
 455         .WORD   0xeffafa15, 0xb25959eb, 0x8e4747c9, 0xfbf0f00b
 456         .WORD   0x41adadec, 0xb3d4d467, 0x5fa2a2fd, 0x45afafea
 457         .WORD   0x239c9cbf, 0x53a4a4f7, 0xe4727296, 0x9bc0c05b
 458         .WORD   0x75b7b7c2, 0xe1fdfd1c, 0x3d9393ae, 0x4c26266a
 459         .WORD   0x6c36365a, 0x7e3f3f41, 0xf5f7f702, 0x83cccc4f
 460         .WORD   0x6834345c, 0x51a5a5f4, 0xd1e5e534, 0xf9f1f108
 461         .WORD   0xe2717193, 0xabd8d873, 0x62313153, 0x2a15153f
 462         .WORD   0x0804040c, 0x95c7c752, 0x46232365, 0x9dc3c35e
 463         .WORD   0x30181828, 0x379696a1, 0x0a05050f, 0x2f9a9ab5
 464         .WORD   0x0e070709, 0x24121236, 0x1b80809b, 0xdfe2e23d
 465         .WORD   0xcdebeb26, 0x4e272769, 0x7fb2b2cd, 0xea75759f
 466         .WORD   0x1209091b, 0x1d83839e, 0x582c2c74, 0x341a1a2e
 467         .WORD   0x361b1b2d, 0xdc6e6eb2, 0xb45a5aee, 0x5ba0a0fb
 468         .WORD   0xa45252f6, 0x763b3b4d, 0xb7d6d661, 0x7db3b3ce
 469         .WORD   0x5229297b, 0xdde3e33e, 0x5e2f2f71, 0x13848497
 470         .WORD   0xa65353f5, 0xb9d1d168, 0x00000000, 0xc1eded2c
 471         .WORD   0x40202060, 0xe3fcfc1f, 0x79b1b1c8, 0xb65b5bed
 472         .WORD   0xd46a6abe, 0x8dcbcb46, 0x67bebed9, 0x7239394b
 473         .WORD   0x944a4ade, 0x984c4cd4, 0xb05858e8, 0x85cfcf4a
 474         .WORD   0xbbd0d06b, 0xc5efef2a, 0x4faaaae5, 0xedfbfb16
 475         .WORD   0x864343c5, 0x9a4d4dd7, 0x66333355, 0x11858594
 476         .WORD   0x8a4545cf, 0xe9f9f910, 0x04020206, 0xfe7f7f81
 477         .WORD   0xa05050f0, 0x783c3c44, 0x259f9fba, 0x4ba8a8e3
 478         .WORD   0xa25151f3, 0x5da3a3fe, 0x804040c0, 0x058f8f8a
 479         .WORD   0x3f9292ad, 0x219d9dbc, 0x70383848, 0xf1f5f504
 480         .WORD   0x63bcbcdf, 0x77b6b6c1, 0xafdada75, 0x42212163
 481         .WORD   0x20101030, 0xe5ffff1a, 0xfdf3f30e, 0xbfd2d26d
 482         .WORD   0x81cdcd4c, 0x180c0c14, 0x26131335, 0xc3ecec2f
 483         .WORD   0xbe5f5fe1, 0x359797a2, 0x884444cc, 0x2e171739
 484         .WORD   0x93c4c457, 0x55a7a7f2, 0xfc7e7e82, 0x7a3d3d47
 485         .WORD   0xc86464ac, 0xba5d5de7, 0x3219192b, 0xe6737395
 486         .WORD   0xc06060a0, 0x19818198, 0x9e4f4fd1, 0xa3dcdc7f
 487         .WORD   0x44222266, 0x542a2a7e, 0x3b9090ab, 0x0b888883
 488         .WORD   0x8c4646ca, 0xc7eeee29, 0x6bb8b8d3, 0x2814143c
 489         .WORD   0xa7dede79, 0xbc5e5ee2, 0x160b0b1d, 0xaddbdb76
 490         .WORD   0xdbe0e03b, 0x64323256, 0x743a3a4e, 0x140a0a1e
 491         .WORD   0x924949db, 0x0c06060a, 0x4824246c, 0xb85c5ce4
 492         .WORD   0x9fc2c25d, 0xbdd3d36e, 0x43acacef, 0xc46262a6
 493         .WORD   0x399191a8, 0x319595a4, 0xd3e4e437, 0xf279798b
 494         .WORD   0xd5e7e732, 0x8bc8c843, 0x6e373759, 0xda6d6db7
 495         .WORD   0x018d8d8c, 0xb1d5d564, 0x9c4e4ed2, 0x49a9a9e0
 496         .WORD   0xd86c6cb4, 0xac5656fa, 0xf3f4f407, 0xcfeaea25
 497         .WORD   0xca6565af, 0xf47a7a8e, 0x47aeaee9, 0x10080818
 498         .WORD   0x6fbabad5, 0xf0787888, 0x4a25256f, 0x5c2e2e72
 499         .WORD   0x381c1c24, 0x57a6a6f1, 0x73b4b4c7, 0x97c6c651
 500         .WORD   0xcbe8e823, 0xa1dddd7c, 0xe874749c, 0x3e1f1f21
 501         .WORD   0x964b4bdd, 0x61bdbddc, 0x0d8b8b86, 0x0f8a8a85
 502         .WORD   0xe0707090, 0x7c3e3e42, 0x71b5b5c4, 0xcc6666aa
 503         .WORD   0x904848d8, 0x06030305, 0xf7f6f601, 0x1c0e0e12
 504         .WORD   0xc26161a3, 0x6a35355f, 0xae5757f9, 0x69b9b9d0
 505         .WORD   0x17868691, 0x99c1c158, 0x3a1d1d27, 0x279e9eb9
 506         .WORD   0xd9e1e138, 0xebf8f813, 0x2b9898b3, 0x22111133
 507         .WORD   0xd26969bb, 0xa9d9d970, 0x078e8e89, 0x339494a7
 508         .WORD   0x2d9b9bb6, 0x3c1e1e22, 0x15878792, 0xc9e9e920
 509         .WORD   0x87cece49, 0xaa5555ff, 0x50282878, 0xa5dfdf7a
 510         .WORD   0x038c8c8f, 0x59a1a1f8, 0x09898980, 0x1a0d0d17
 511         .WORD   0x65bfbfda, 0xd7e6e631, 0x844242c6, 0xd06868b8
 512         .WORD   0x824141c3, 0x299999b0, 0x5a2d2d77, 0x1e0f0f11
 513         .WORD   0x7bb0b0cb, 0xa85454fc, 0x6dbbbbd6, 0x2c16163a
 514         .BYTE   0x63, 0x7c, 0x77, 0x7b, 0xf2, 0x6b, 0x6f, 0xc5
 515         .BYTE   0x30, 0x01, 0x67, 0x2b, 0xfe, 0xd7, 0xab, 0x76
 516         .BYTE   0xca, 0x82, 0xc9, 0x7d, 0xfa, 0x59, 0x47, 0xf0
 517         .BYTE   0xad, 0xd4, 0xa2, 0xaf, 0x9c, 0xa4, 0x72, 0xc0
 518         .BYTE   0xb7, 0xfd, 0x93, 0x26, 0x36, 0x3f, 0xf7, 0xcc
 519         .BYTE   0x34, 0xa5, 0xe5, 0xf1, 0x71, 0xd8, 0x31, 0x15
 520         .BYTE   0x04, 0xc7, 0x23, 0xc3, 0x18, 0x96, 0x05, 0x9a
 521         .BYTE   0x07, 0x12, 0x80, 0xe2, 0xeb, 0x27, 0xb2, 0x75
 522         .BYTE   0x09, 0x83, 0x2c, 0x1a, 0x1b, 0x6e, 0x5a, 0xa0
 523         .BYTE   0x52, 0x3b, 0xd6, 0xb3, 0x29, 0xe3, 0x2f, 0x84
 524         .BYTE   0x53, 0xd1, 0x00, 0xed, 0x20, 0xfc, 0xb1, 0x5b
 525         .BYTE   0x6a, 0xcb, 0xbe, 0x39, 0x4a, 0x4c, 0x58, 0xcf
 526         .BYTE   0xd0, 0xef, 0xaa, 0xfb, 0x43, 0x4d, 0x33, 0x85
 527         .BYTE   0x45, 0xf9, 0x02, 0x7f, 0x50, 0x3c, 0x9f, 0xa8
 528         .BYTE   0x51, 0xa3, 0x40, 0x8f, 0x92, 0x9d, 0x38, 0xf5
 529         .BYTE   0xbc, 0xb6, 0xda, 0x21, 0x10, 0xff, 0xf3, 0xd2
 530         .BYTE   0xcd, 0x0c, 0x13, 0xec, 0x5f, 0x97, 0x44, 0x17
 531         .BYTE   0xc4, 0xa7, 0x7e, 0x3d, 0x64, 0x5d, 0x19, 0x73
 532         .BYTE   0x60, 0x81, 0x4f, 0xdc, 0x22, 0x2a, 0x90, 0x88
 533         .BYTE   0x46, 0xee, 0xb8, 0x14, 0xde, 0x5e, 0x0b, 0xdb
 534         .BYTE   0xe0, 0x32, 0x3a, 0x0a, 0x49, 0x06, 0x24, 0x5c
 535         .BYTE   0xc2, 0xd3, 0xac, 0x62, 0x91, 0x95, 0xe4, 0x79
 536         .BYTE   0xe7, 0xc8, 0x37, 0x6d, 0x8d, 0xd5, 0x4e, 0xa9
 537         .BYTE   0x6c, 0x56, 0xf4, 0xea, 0x65, 0x7a, 0xae, 0x08
 538         .BYTE   0xba, 0x78, 0x25, 0x2e, 0x1c, 0xa6, 0xb4, 0xc6
 539         .BYTE   0xe8, 0xdd, 0x74, 0x1f, 0x4b, 0xbd, 0x8b, 0x8a
 540         .BYTE   0x70, 0x3e, 0xb5, 0x66, 0x48, 0x03, 0xf6, 0x0e
 541         .BYTE   0x61, 0x35, 0x57, 0xb9, 0x86, 0xc1, 0x1d, 0x9e
 542         .BYTE   0xe1, 0xf8, 0x98, 0x11, 0x69, 0xd9, 0x8e, 0x94
 543         .BYTE   0x9b, 0x1e, 0x87, 0xe9, 0xce, 0x55, 0x28, 0xdf
 544         .BYTE   0x8c, 0xa1, 0x89, 0x0d, 0xbf, 0xe6, 0x42, 0x68
 545         .BYTE   0x41, 0x99, 0x2d, 0x0f, 0xb0, 0x54, 0xbb, 0x16
 546 ___
 547
 548 $code.=<<___;
 549         .EXPORT AES_decrypt,ENTRY,ARGW0=GR,ARGW1=GR,ARGW2=GR
 550         .ALIGN  16
 551 AES_decrypt
 552         .PROC
 553         .CALLINFO       FRAME=`$FRAME-16*$SIZE_T`,NO_CALLS,SAVE_RP,ENTRY_GR=18
 554         .ENTRY
 555         $PUSH   %r2,-$SAVED_RP(%sp)     ; standard prologue
 556         $PUSHMA %r3,$FRAME(%sp)
 557         $PUSH   %r4,`-$FRAME+1*$SIZE_T`(%sp)
 558         $PUSH   %r5,`-$FRAME+2*$SIZE_T`(%sp)
 559         $PUSH   %r6,`-$FRAME+3*$SIZE_T`(%sp)
 560         $PUSH   %r7,`-$FRAME+4*$SIZE_T`(%sp)
 561         $PUSH   %r8,`-$FRAME+5*$SIZE_T`(%sp)
 562         $PUSH   %r9,`-$FRAME+6*$SIZE_T`(%sp)
 563         $PUSH   %r10,`-$FRAME+7*$SIZE_T`(%sp)
 564         $PUSH   %r11,`-$FRAME+8*$SIZE_T`(%sp)
 565         $PUSH   %r12,`-$FRAME+9*$SIZE_T`(%sp)
 566         $PUSH   %r13,`-$FRAME+10*$SIZE_T`(%sp)
 567         $PUSH   %r14,`-$FRAME+11*$SIZE_T`(%sp)
 568         $PUSH   %r15,`-$FRAME+12*$SIZE_T`(%sp)
 569         $PUSH   %r16,`-$FRAME+13*$SIZE_T`(%sp)
 570         $PUSH   %r17,`-$FRAME+14*$SIZE_T`(%sp)
 571         $PUSH   %r18,`-$FRAME+15*$SIZE_T`(%sp)
 572
 573         blr     %r0,$tbl
 574         ldi     3,$t0
 575 L\$dec_pic
 576         andcm   $tbl,$t0,$tbl
 577         ldo     L\$AES_Td-L\$dec_pic($tbl),$tbl
 578
 579         and     $inp,$t0,$t0
 580         sub     $inp,$t0,$inp
 581         ldw     0($inp),$s0
 582         ldw     4($inp),$s1
 583         ldw     8($inp),$s2
 584         comib,= 0,$t0,L\$dec_inp_aligned
 585         ldw     12($inp),$s3
 586
 587         sh3addl $t0,%r0,$t0
 588         subi    32,$t0,$t0
 589         mtctl   $t0,%cr11
 590         ldw     16($inp),$t1
 591         vshd    $s0,$s1,$s0
 592         vshd    $s1,$s2,$s1
 593         vshd    $s2,$s3,$s2
 594         vshd    $s3,$t1,$s3
 595
 596 L\$dec_inp_aligned
 597         bl      _parisc_AES_decrypt,%r31
 598         nop
 599
 600         extru,<> $out,31,2,%r0
 601         b       L\$dec_out_aligned
 602         nop
 603
 604         _srm    $s0,24,$acc0
 605         _srm    $s0,16,$acc1
 606         stb     $acc0,0($out)
 607         _srm    $s0,8,$acc2
 608         stb     $acc1,1($out)
 609         _srm    $s1,24,$acc4
 610         stb     $acc2,2($out)
 611         _srm    $s1,16,$acc5
 612         stb     $s0,3($out)
 613         _srm    $s1,8,$acc6
 614         stb     $acc4,4($out)
 615         _srm    $s2,24,$acc0
 616         stb     $acc5,5($out)
 617         _srm    $s2,16,$acc1
 618         stb     $acc6,6($out)
 619         _srm    $s2,8,$acc2
 620         stb     $s1,7($out)
 621         _srm    $s3,24,$acc4
 622         stb     $acc0,8($out)
 623         _srm    $s3,16,$acc5
 624         stb     $acc1,9($out)
 625         _srm    $s3,8,$acc6
 626         stb     $acc2,10($out)
 627         stb     $s2,11($out)
 628         stb     $acc4,12($out)
 629         stb     $acc5,13($out)
 630         stb     $acc6,14($out)
 631         b       L\$dec_done
 632         stb     $s3,15($out)
 633
 634 L\$dec_out_aligned
 635         stw     $s0,0($out)
 636         stw     $s1,4($out)
 637         stw     $s2,8($out)
 638         stw     $s3,12($out)
 639
 640 L\$dec_done
 641         $POP    `-$FRAME-$SAVED_RP`(%sp),%r2    ; standard epilogue
 642         $POP    `-$FRAME+1*$SIZE_T`(%sp),%r4
 643         $POP    `-$FRAME+2*$SIZE_T`(%sp),%r5
 644         $POP    `-$FRAME+3*$SIZE_T`(%sp),%r6
 645         $POP    `-$FRAME+4*$SIZE_T`(%sp),%r7
 646         $POP    `-$FRAME+5*$SIZE_T`(%sp),%r8
 647         $POP    `-$FRAME+6*$SIZE_T`(%sp),%r9
 648         $POP    `-$FRAME+7*$SIZE_T`(%sp),%r10
 649         $POP    `-$FRAME+8*$SIZE_T`(%sp),%r11
 650         $POP    `-$FRAME+9*$SIZE_T`(%sp),%r12
 651         $POP    `-$FRAME+10*$SIZE_T`(%sp),%r13
 652         $POP    `-$FRAME+11*$SIZE_T`(%sp),%r14
 653         $POP    `-$FRAME+12*$SIZE_T`(%sp),%r15
 654         $POP    `-$FRAME+13*$SIZE_T`(%sp),%r16
 655         $POP    `-$FRAME+14*$SIZE_T`(%sp),%r17
 656         $POP    `-$FRAME+15*$SIZE_T`(%sp),%r18
 657         bv      (%r2)
 658         .EXIT
 659         $POPMB  -$FRAME(%sp),%r3
 660         .PROCEND
 661
 662         .ALIGN  16
 663 _parisc_AES_decrypt
 664         .PROC
 665         .CALLINFO       MILLICODE
 666         .ENTRY
 667         ldw     240($key),$rounds
 668         ldw     0($key),$t0
 669         ldw     4($key),$t1
 670         ldw     8($key),$t2
 671         ldw     12($key),$t3
 672         _srm    $rounds,1,$rounds
 673         xor     $t0,$s0,$s0
 674         ldw     16($key),$t0
 675         xor     $t1,$s1,$s1
 676         ldw     20($key),$t1
 677         _srm    $s0,24,$acc0
 678         xor     $t2,$s2,$s2
 679         ldw     24($key),$t2
 680         xor     $t3,$s3,$s3
 681         ldw     28($key),$t3
 682         _srm    $s3,16,$acc1
 683 L\$dec_loop
 684         _srm    $s2,8,$acc2
 685         ldwx,s  $acc0($tbl),$acc0
 686         _srm    $s1,0,$acc3
 687         ldwx,s  $acc1($tbl),$acc1
 688         _srm    $s1,24,$acc4
 689         ldwx,s  $acc2($tbl),$acc2
 690         _srm    $s0,16,$acc5
 691         ldwx,s  $acc3($tbl),$acc3
 692         _srm    $s3,8,$acc6
 693         ldwx,s  $acc4($tbl),$acc4
 694         _srm    $s2,0,$acc7
 695         ldwx,s  $acc5($tbl),$acc5
 696         _srm    $s2,24,$acc8
 697         ldwx,s  $acc6($tbl),$acc6
 698         _srm    $s1,16,$acc9
 699         ldwx,s  $acc7($tbl),$acc7
 700         _srm    $s0,8,$acc10
 701         ldwx,s  $acc8($tbl),$acc8
 702         _srm    $s3,0,$acc11
 703         ldwx,s  $acc9($tbl),$acc9
 704         _srm    $s3,24,$acc12
 705         ldwx,s  $acc10($tbl),$acc10
 706         _srm    $s2,16,$acc13
 707         ldwx,s  $acc11($tbl),$acc11
 708         _srm    $s1,8,$acc14
 709         ldwx,s  $acc12($tbl),$acc12
 710         _srm    $s0,0,$acc15
 711         ldwx,s  $acc13($tbl),$acc13
 712         ldwx,s  $acc14($tbl),$acc14
 713         ldwx,s  $acc15($tbl),$acc15
 714         addib,= -1,$rounds,L\$dec_last
 715         ldo     32($key),$key
 716
 717                 _ror    $acc1,8,$acc1
 718                 xor     $acc0,$t0,$t0
 719         ldw     0($key),$s0
 720                 _ror    $acc2,16,$acc2
 721                 xor     $acc1,$t0,$t0
 722         ldw     4($key),$s1
 723                 _ror    $acc3,24,$acc3
 724                 xor     $acc2,$t0,$t0
 725         ldw     8($key),$s2
 726                 _ror    $acc5,8,$acc5
 727                 xor     $acc3,$t0,$t0
 728         ldw     12($key),$s3
 729                 _ror    $acc6,16,$acc6
 730                 xor     $acc4,$t1,$t1
 731                 _ror    $acc7,24,$acc7
 732                 xor     $acc5,$t1,$t1
 733                 _ror    $acc9,8,$acc9
 734                 xor     $acc6,$t1,$t1
 735                 _ror    $acc10,16,$acc10
 736                 xor     $acc7,$t1,$t1
 737                 _ror    $acc11,24,$acc11
 738                 xor     $acc8,$t2,$t2
 739                 _ror    $acc13,8,$acc13
 740                 xor     $acc9,$t2,$t2
 741                 _ror    $acc14,16,$acc14
 742                 xor     $acc10,$t2,$t2
 743                 _ror    $acc15,24,$acc15
 744                 xor     $acc11,$t2,$t2
 745                 xor     $acc12,$acc14,$acc14
 746                 xor     $acc13,$t3,$t3
 747         _srm    $t0,24,$acc0
 748                 xor     $acc14,$t3,$t3
 749                 xor     $acc15,$t3,$t3
 750         _srm    $t3,16,$acc1
 751
 752         _srm    $t2,8,$acc2
 753         ldwx,s  $acc0($tbl),$acc0
 754         _srm    $t1,0,$acc3
 755         ldwx,s  $acc1($tbl),$acc1
 756         _srm    $t1,24,$acc4
 757         ldwx,s  $acc2($tbl),$acc2
 758         _srm    $t0,16,$acc5
 759         ldwx,s  $acc3($tbl),$acc3
 760         _srm    $t3,8,$acc6
 761         ldwx,s  $acc4($tbl),$acc4
 762         _srm    $t2,0,$acc7
 763         ldwx,s  $acc5($tbl),$acc5
 764         _srm    $t2,24,$acc8
 765         ldwx,s  $acc6($tbl),$acc6
 766         _srm    $t1,16,$acc9
 767         ldwx,s  $acc7($tbl),$acc7
 768         _srm    $t0,8,$acc10
 769         ldwx,s  $acc8($tbl),$acc8
 770         _srm    $t3,0,$acc11
 771         ldwx,s  $acc9($tbl),$acc9
 772         _srm    $t3,24,$acc12
 773         ldwx,s  $acc10($tbl),$acc10
 774         _srm    $t2,16,$acc13
 775         ldwx,s  $acc11($tbl),$acc11
 776         _srm    $t1,8,$acc14
 777         ldwx,s  $acc12($tbl),$acc12
 778         _srm    $t0,0,$acc15
 779         ldwx,s  $acc13($tbl),$acc13
 780                 _ror    $acc1,8,$acc1
 781         ldwx,s  $acc14($tbl),$acc14
 782
 783                 _ror    $acc2,16,$acc2
 784                 xor     $acc0,$s0,$s0
 785         ldwx,s  $acc15($tbl),$acc15
 786                 _ror    $acc3,24,$acc3
 787                 xor     $acc1,$s0,$s0
 788         ldw     16($key),$t0
 789                 _ror    $acc5,8,$acc5
 790                 xor     $acc2,$s0,$s0
 791         ldw     20($key),$t1
 792                 _ror    $acc6,16,$acc6
 793                 xor     $acc3,$s0,$s0
 794         ldw     24($key),$t2
 795                 _ror    $acc7,24,$acc7
 796                 xor     $acc4,$s1,$s1
 797         ldw     28($key),$t3
 798                 _ror    $acc9,8,$acc9
 799                 xor     $acc5,$s1,$s1
 800         ldw     1024+0($tbl),%r0                ; prefetch td4
 801                 _ror    $acc10,16,$acc10
 802                 xor     $acc6,$s1,$s1
 803         ldw     1024+32($tbl),%r0               ; prefetch td4
 804                 _ror    $acc11,24,$acc11
 805                 xor     $acc7,$s1,$s1
 806         ldw     1024+64($tbl),%r0               ; prefetch td4
 807                 _ror    $acc13,8,$acc13
 808                 xor     $acc8,$s2,$s2
 809         ldw     1024+96($tbl),%r0               ; prefetch td4
 810                 _ror    $acc14,16,$acc14
 811                 xor     $acc9,$s2,$s2
 812         ldw     1024+128($tbl),%r0              ; prefetch td4
 813                 _ror    $acc15,24,$acc15
 814                 xor     $acc10,$s2,$s2
 815         ldw     1024+160($tbl),%r0              ; prefetch td4
 816         _srm    $s0,24,$acc0
 817                 xor     $acc11,$s2,$s2
 818         ldw     1024+192($tbl),%r0              ; prefetch td4
 819                 xor     $acc12,$acc14,$acc14
 820                 xor     $acc13,$s3,$s3
 821         ldw     1024+224($tbl),%r0              ; prefetch td4
 822                 xor     $acc14,$s3,$s3
 823                 xor     $acc15,$s3,$s3
 824         b       L\$dec_loop
 825         _srm    $s3,16,$acc1
 826
 827         .ALIGN  16
 828 L\$dec_last
 829         ldo     1024($tbl),$rounds
 830                 _ror    $acc1,8,$acc1
 831                 xor     $acc0,$t0,$t0
 832         ldw     0($key),$s0
 833                 _ror    $acc2,16,$acc2
 834                 xor     $acc1,$t0,$t0
 835         ldw     4($key),$s1
 836                 _ror    $acc3,24,$acc3
 837                 xor     $acc2,$t0,$t0
 838         ldw     8($key),$s2
 839                 _ror    $acc5,8,$acc5
 840                 xor     $acc3,$t0,$t0
 841         ldw     12($key),$s3
 842                 _ror    $acc6,16,$acc6
 843                 xor     $acc4,$t1,$t1
 844                 _ror    $acc7,24,$acc7
 845                 xor     $acc5,$t1,$t1
 846                 _ror    $acc9,8,$acc9
 847                 xor     $acc6,$t1,$t1
 848                 _ror    $acc10,16,$acc10
 849                 xor     $acc7,$t1,$t1
 850                 _ror    $acc11,24,$acc11
 851                 xor     $acc8,$t2,$t2
 852                 _ror    $acc13,8,$acc13
 853                 xor     $acc9,$t2,$t2
 854                 _ror    $acc14,16,$acc14
 855                 xor     $acc10,$t2,$t2
 856                 _ror    $acc15,24,$acc15
 857                 xor     $acc11,$t2,$t2
 858                 xor     $acc12,$acc14,$acc14
 859                 xor     $acc13,$t3,$t3
 860         _srm    $t0,24,$acc0
 861                 xor     $acc14,$t3,$t3
 862                 xor     $acc15,$t3,$t3
 863         _srm    $t3,16,$acc1
 864
 865         _srm    $t2,8,$acc2
 866         ldbx    $acc0($rounds),$acc0
 867         _srm    $t1,24,$acc4
 868         ldbx    $acc1($rounds),$acc1
 869         _srm    $t0,16,$acc5
 870         _srm    $t1,0,$acc3
 871         ldbx    $acc2($rounds),$acc2
 872         ldbx    $acc3($rounds),$acc3
 873         _srm    $t3,8,$acc6
 874         ldbx    $acc4($rounds),$acc4
 875         _srm    $t2,24,$acc8
 876         ldbx    $acc5($rounds),$acc5
 877         _srm    $t1,16,$acc9
 878         _srm    $t2,0,$acc7
 879         ldbx    $acc6($rounds),$acc6
 880         ldbx    $acc7($rounds),$acc7
 881         _srm    $t0,8,$acc10
 882         ldbx    $acc8($rounds),$acc8
 883         _srm    $t3,24,$acc12
 884         ldbx    $acc9($rounds),$acc9
 885         _srm    $t2,16,$acc13
 886         _srm    $t3,0,$acc11
 887         ldbx    $acc10($rounds),$acc10
 888         _srm    $t1,8,$acc14
 889         ldbx    $acc11($rounds),$acc11
 890         ldbx    $acc12($rounds),$acc12
 891         ldbx    $acc13($rounds),$acc13
 892         _srm    $t0,0,$acc15
 893         ldbx    $acc14($rounds),$acc14
 894
 895                 dep     $acc0,7,8,$acc3
 896         ldbx    $acc15($rounds),$acc15
 897                 dep     $acc4,7,8,$acc7
 898                 dep     $acc1,15,8,$acc3
 899                 dep     $acc5,15,8,$acc7
 900                 dep     $acc2,23,8,$acc3
 901                 dep     $acc6,23,8,$acc7
 902                 xor     $acc3,$s0,$s0
 903                 xor     $acc7,$s1,$s1
 904                 dep     $acc8,7,8,$acc11
 905                 dep     $acc12,7,8,$acc15
 906                 dep     $acc9,15,8,$acc11
 907                 dep     $acc13,15,8,$acc15
 908                 dep     $acc10,23,8,$acc11
 909                 dep     $acc14,23,8,$acc15
 910                 xor     $acc11,$s2,$s2
 911
 912         bv      (%r31)
 913         .EXIT
 914                 xor     $acc15,$s3,$s3
 915         .PROCEND
 916
 917         .ALIGN  64
 918 L\$AES_Td
 919         .WORD   0x51f4a750, 0x7e416553, 0x1a17a4c3, 0x3a275e96
 920         .WORD   0x3bab6bcb, 0x1f9d45f1, 0xacfa58ab, 0x4be30393
 921         .WORD   0x2030fa55, 0xad766df6, 0x88cc7691, 0xf5024c25
 922         .WORD   0x4fe5d7fc, 0xc52acbd7, 0x26354480, 0xb562a38f
 923         .WORD   0xdeb15a49, 0x25ba1b67, 0x45ea0e98, 0x5dfec0e1
 924         .WORD   0xc32f7502, 0x814cf012, 0x8d4697a3, 0x6bd3f9c6
 925         .WORD   0x038f5fe7, 0x15929c95, 0xbf6d7aeb, 0x955259da
 926         .WORD   0xd4be832d, 0x587421d3, 0x49e06929, 0x8ec9c844
 927         .WORD   0x75c2896a, 0xf48e7978, 0x99583e6b, 0x27b971dd
 928         .WORD   0xbee14fb6, 0xf088ad17, 0xc920ac66, 0x7dce3ab4
 929         .WORD   0x63df4a18, 0xe51a3182, 0x97513360, 0x62537f45
 930         .WORD   0xb16477e0, 0xbb6bae84, 0xfe81a01c, 0xf9082b94
 931         .WORD   0x70486858, 0x8f45fd19, 0x94de6c87, 0x527bf8b7
 932         .WORD   0xab73d323, 0x724b02e2, 0xe31f8f57, 0x6655ab2a
 933         .WORD   0xb2eb2807, 0x2fb5c203, 0x86c57b9a, 0xd33708a5
 934         .WORD   0x302887f2, 0x23bfa5b2, 0x02036aba, 0xed16825c
 935         .WORD   0x8acf1c2b, 0xa779b492, 0xf307f2f0, 0x4e69e2a1
 936         .WORD   0x65daf4cd, 0x0605bed5, 0xd134621f, 0xc4a6fe8a
 937         .WORD   0x342e539d, 0xa2f355a0, 0x058ae132, 0xa4f6eb75
 938         .WORD   0x0b83ec39, 0x4060efaa, 0x5e719f06, 0xbd6e1051
 939         .WORD   0x3e218af9, 0x96dd063d, 0xdd3e05ae, 0x4de6bd46
 940         .WORD   0x91548db5, 0x71c45d05, 0x0406d46f, 0x605015ff
 941         .WORD   0x1998fb24, 0xd6bde997, 0x894043cc, 0x67d99e77
 942         .WORD   0xb0e842bd, 0x07898b88, 0xe7195b38, 0x79c8eedb
 943         .WORD   0xa17c0a47, 0x7c420fe9, 0xf8841ec9, 0x00000000
 944         .WORD   0x09808683, 0x322bed48, 0x1e1170ac, 0x6c5a724e
 945         .WORD   0xfd0efffb, 0x0f853856, 0x3daed51e, 0x362d3927
 946         .WORD   0x0a0fd964, 0x685ca621, 0x9b5b54d1, 0x24362e3a
 947         .WORD   0x0c0a67b1, 0x9357e70f, 0xb4ee96d2, 0x1b9b919e
 948         .WORD   0x80c0c54f, 0x61dc20a2, 0x5a774b69, 0x1c121a16
 949         .WORD   0xe293ba0a, 0xc0a02ae5, 0x3c22e043, 0x121b171d
 950         .WORD   0x0e090d0b, 0xf28bc7ad, 0x2db6a8b9, 0x141ea9c8
 951         .WORD   0x57f11985, 0xaf75074c, 0xee99ddbb, 0xa37f60fd
 952         .WORD   0xf701269f, 0x5c72f5bc, 0x44663bc5, 0x5bfb7e34
 953         .WORD   0x8b432976, 0xcb23c6dc, 0xb6edfc68, 0xb8e4f163
 954         .WORD   0xd731dcca, 0x42638510, 0x13972240, 0x84c61120
 955         .WORD   0x854a247d, 0xd2bb3df8, 0xaef93211, 0xc729a16d
 956         .WORD   0x1d9e2f4b, 0xdcb230f3, 0x0d8652ec, 0x77c1e3d0
 957         .WORD   0x2bb3166c, 0xa970b999, 0x119448fa, 0x47e96422
 958         .WORD   0xa8fc8cc4, 0xa0f03f1a, 0x567d2cd8, 0x223390ef
 959         .WORD   0x87494ec7, 0xd938d1c1, 0x8ccaa2fe, 0x98d40b36
 960         .WORD   0xa6f581cf, 0xa57ade28, 0xdab78e26, 0x3fadbfa4
 961         .WORD   0x2c3a9de4, 0x5078920d, 0x6a5fcc9b, 0x547e4662
 962         .WORD   0xf68d13c2, 0x90d8b8e8, 0x2e39f75e, 0x82c3aff5
 963         .WORD   0x9f5d80be, 0x69d0937c, 0x6fd52da9, 0xcf2512b3
 964         .WORD   0xc8ac993b, 0x10187da7, 0xe89c636e, 0xdb3bbb7b
 965         .WORD   0xcd267809, 0x6e5918f4, 0xec9ab701, 0x834f9aa8
 966         .WORD   0xe6956e65, 0xaaffe67e, 0x21bccf08, 0xef15e8e6
 967         .WORD   0xbae79bd9, 0x4a6f36ce, 0xea9f09d4, 0x29b07cd6
 968         .WORD   0x31a4b2af, 0x2a3f2331, 0xc6a59430, 0x35a266c0
 969         .WORD   0x744ebc37, 0xfc82caa6, 0xe090d0b0, 0x33a7d815
 970         .WORD   0xf104984a, 0x41ecdaf7, 0x7fcd500e, 0x1791f62f
 971         .WORD   0x764dd68d, 0x43efb04d, 0xccaa4d54, 0xe49604df
 972         .WORD   0x9ed1b5e3, 0x4c6a881b, 0xc12c1fb8, 0x4665517f
 973         .WORD   0x9d5eea04, 0x018c355d, 0xfa877473, 0xfb0b412e
 974         .WORD   0xb3671d5a, 0x92dbd252, 0xe9105633, 0x6dd64713
 975         .WORD   0x9ad7618c, 0x37a10c7a, 0x59f8148e, 0xeb133c89
 976         .WORD   0xcea927ee, 0xb761c935, 0xe11ce5ed, 0x7a47b13c
 977         .WORD   0x9cd2df59, 0x55f2733f, 0x1814ce79, 0x73c737bf
 978         .WORD   0x53f7cdea, 0x5ffdaa5b, 0xdf3d6f14, 0x7844db86
 979         .WORD   0xcaaff381, 0xb968c43e, 0x3824342c, 0xc2a3405f
 980         .WORD   0x161dc372, 0xbce2250c, 0x283c498b, 0xff0d9541
 981         .WORD   0x39a80171, 0x080cb3de, 0xd8b4e49c, 0x6456c190
 982         .WORD   0x7bcb8461, 0xd532b670, 0x486c5c74, 0xd0b85742
 983         .BYTE   0x52, 0x09, 0x6a, 0xd5, 0x30, 0x36, 0xa5, 0x38
 984         .BYTE   0xbf, 0x40, 0xa3, 0x9e, 0x81, 0xf3, 0xd7, 0xfb
 985         .BYTE   0x7c, 0xe3, 0x39, 0x82, 0x9b, 0x2f, 0xff, 0x87
 986         .BYTE   0x34, 0x8e, 0x43, 0x44, 0xc4, 0xde, 0xe9, 0xcb
 987         .BYTE   0x54, 0x7b, 0x94, 0x32, 0xa6, 0xc2, 0x23, 0x3d
 988         .BYTE   0xee, 0x4c, 0x95, 0x0b, 0x42, 0xfa, 0xc3, 0x4e
 989         .BYTE   0x08, 0x2e, 0xa1, 0x66, 0x28, 0xd9, 0x24, 0xb2
 990         .BYTE   0x76, 0x5b, 0xa2, 0x49, 0x6d, 0x8b, 0xd1, 0x25
 991         .BYTE   0x72, 0xf8, 0xf6, 0x64, 0x86, 0x68, 0x98, 0x16
 992         .BYTE   0xd4, 0xa4, 0x5c, 0xcc, 0x5d, 0x65, 0xb6, 0x92
 993         .BYTE   0x6c, 0x70, 0x48, 0x50, 0xfd, 0xed, 0xb9, 0xda
 994         .BYTE   0x5e, 0x15, 0x46, 0x57, 0xa7, 0x8d, 0x9d, 0x84
 995         .BYTE   0x90, 0xd8, 0xab, 0x00, 0x8c, 0xbc, 0xd3, 0x0a
 996         .BYTE   0xf7, 0xe4, 0x58, 0x05, 0xb8, 0xb3, 0x45, 0x06
 997         .BYTE   0xd0, 0x2c, 0x1e, 0x8f, 0xca, 0x3f, 0x0f, 0x02
 998         .BYTE   0xc1, 0xaf, 0xbd, 0x03, 0x01, 0x13, 0x8a, 0x6b
 999         .BYTE   0x3a, 0x91, 0x11, 0x41, 0x4f, 0x67, 0xdc, 0xea
1000         .BYTE   0x97, 0xf2, 0xcf, 0xce, 0xf0, 0xb4, 0xe6, 0x73
1001         .BYTE   0x96, 0xac, 0x74, 0x22, 0xe7, 0xad, 0x35, 0x85
1002         .BYTE   0xe2, 0xf9, 0x37, 0xe8, 0x1c, 0x75, 0xdf, 0x6e
1003         .BYTE   0x47, 0xf1, 0x1a, 0x71, 0x1d, 0x29, 0xc5, 0x89
1004         .BYTE   0x6f, 0xb7, 0x62, 0x0e, 0xaa, 0x18, 0xbe, 0x1b
1005         .BYTE   0xfc, 0x56, 0x3e, 0x4b, 0xc6, 0xd2, 0x79, 0x20
1006         .BYTE   0x9a, 0xdb, 0xc0, 0xfe, 0x78, 0xcd, 0x5a, 0xf4
1007         .BYTE   0x1f, 0xdd, 0xa8, 0x33, 0x88, 0x07, 0xc7, 0x31
1008         .BYTE   0xb1, 0x12, 0x10, 0x59, 0x27, 0x80, 0xec, 0x5f
1009         .BYTE   0x60, 0x51, 0x7f, 0xa9, 0x19, 0xb5, 0x4a, 0x0d
1010         .BYTE   0x2d, 0xe5, 0x7a, 0x9f, 0x93, 0xc9, 0x9c, 0xef
1011         .BYTE   0xa0, 0xe0, 0x3b, 0x4d, 0xae, 0x2a, 0xf5, 0xb0
1012         .BYTE   0xc8, 0xeb, 0xbb, 0x3c, 0x83, 0x53, 0x99, 0x61
1013         .BYTE   0x17, 0x2b, 0x04, 0x7e, 0xba, 0x77, 0xd6, 0x26
1014         .BYTE   0xe1, 0x69, 0x14, 0x63, 0x55, 0x21, 0x0c, 0x7d
1015         .STRINGZ "AES for PA-RISC, CRYPTOGAMS by <appro\@openssl.org>"
1016 ___
1017
1018 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
1019         =~ /GNU assembler/) {
1020     $gnuas = 1;
1021 }
1022
1023 foreach (split("\n",$code)) {
1024         s/\`([^\`]*)\`/eval $1/ge;
1025
1026         # translate made up instructions: _ror, _srm
1027         s/_ror(\s+)(%r[0-9]+),/shd$1$2,$2,/                             or
1028
1029         s/_srm(\s+%r[0-9]+),([0-9]+),/
1030                 $SIZE_T==4 ? sprintf("extru%s,%d,8,",$1,31-$2)
1031                 :            sprintf("extrd,u%s,%d,8,",$1,63-$2)/e;
1032
1033         s/(\.LEVEL\s+2\.0)W/$1w/        if ($gnuas && $SIZE_T==8);
1034         s/\.SPACE\s+\$TEXT\$/.text/     if ($gnuas && $SIZE_T==8);
1035         s/\.SUBSPA.*//                  if ($gnuas && $SIZE_T==8);
1036         s/,\*/,/                        if ($SIZE_T==4);
1037         s/\bbv\b(.*\(%r2\))/bve$1/      if ($SIZE_T==8);
1038
1039         print $_,"\n";
1040 }
1041 close STDOUT or die "error closing STDOUT: $!";