crypto/des/asm/dest4-sparcv9.pl

   1 #! /usr/bin/env perl
   2 # Copyright 2013-2016 The OpenSSL Project Authors. All Rights Reserved.
   3 #
   4 # Licensed under the OpenSSL license (the "License").  You may not use
   5 # this file except in compliance with the License.  You can obtain a copy
   6 # in the file LICENSE in the source distribution or at
   7 # https://www.openssl.org/source/license.html
   8
   9
  10 # ====================================================================
  11 # Written by David S. Miller and Andy Polyakov.
  12 # The module is licensed under 2-clause BSD
  13 # license. March 2013. All rights reserved.
  14 # ====================================================================
  15
  16 ######################################################################
  17 # DES for SPARC T4.
  18 #
  19 # As with other hardware-assisted ciphers CBC encrypt results [for
  20 # aligned data] are virtually identical to critical path lengths:
  21 #
  22 #               DES             Triple-DES
  23 # CBC encrypt   4.14/4.15(*)    11.7/11.7
  24 # CBC decrypt   1.77/4.11(**)   6.42/7.47
  25 #
  26 #                        (*)    numbers after slash are for
  27 #                               misaligned data;
  28 #                        (**)   this is result for largest
  29 #                               block size, unlike all other
  30 #                               cases smaller blocks results
  31 #                               are better[?];
  32
  33 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
  34 push(@INC,"${dir}","${dir}../../perlasm");
  35 require "sparcv9_modes.pl";
  36
  37 $output=pop;
  38 open STDOUT,">$output";
  39
  40 $code.=<<___;
  41 #include "sparc_arch.h"
  42
  43 #ifdef  __arch64__
  44 .register       %g2,#scratch
  45 .register       %g3,#scratch
  46 #endif
  47
  48 .text
  49 ___
  50
  51 { my ($inp,$out)=("%o0","%o1");
  52
  53 $code.=<<___;
  54 .align  32
  55 .globl  des_t4_key_expand
  56 .type   des_t4_key_expand,#function
  57 des_t4_key_expand:
  58         andcc           $inp, 0x7, %g0
  59         alignaddr       $inp, %g0, $inp
  60         bz,pt           %icc, 1f
  61         ldd             [$inp + 0x00], %f0
  62         ldd             [$inp + 0x08], %f2
  63         faligndata      %f0, %f2, %f0
  64 1:      des_kexpand     %f0, 0, %f0
  65         des_kexpand     %f0, 1, %f2
  66         std             %f0, [$out + 0x00]
  67         des_kexpand     %f2, 3, %f6
  68         std             %f2, [$out + 0x08]
  69         des_kexpand     %f2, 2, %f4
  70         des_kexpand     %f6, 3, %f10
  71         std             %f6, [$out + 0x18]
  72         des_kexpand     %f6, 2, %f8
  73         std             %f4, [$out + 0x10]
  74         des_kexpand     %f10, 3, %f14
  75         std             %f10, [$out + 0x28]
  76         des_kexpand     %f10, 2, %f12
  77         std             %f8, [$out + 0x20]
  78         des_kexpand     %f14, 1, %f16
  79         std             %f14, [$out + 0x38]
  80         des_kexpand     %f16, 3, %f20
  81         std             %f12, [$out + 0x30]
  82         des_kexpand     %f16, 2, %f18
  83         std             %f16, [$out + 0x40]
  84         des_kexpand     %f20, 3, %f24
  85         std             %f20, [$out + 0x50]
  86         des_kexpand     %f20, 2, %f22
  87         std             %f18, [$out + 0x48]
  88         des_kexpand     %f24, 3, %f28
  89         std             %f24, [$out + 0x60]
  90         des_kexpand     %f24, 2, %f26
  91         std             %f22, [$out + 0x58]
  92         des_kexpand     %f28, 1, %f30
  93         std             %f28, [$out + 0x70]
  94         std             %f26, [$out + 0x68]
  95         retl
  96         std             %f30, [$out + 0x78]
  97 .size   des_t4_key_expand,.-des_t4_key_expand
  98 ___
  99 }
 100 { my ($inp,$out,$len,$key,$ivec) = map("%o$_",(0..4));
 101   my ($ileft,$iright,$omask) = map("%g$_",(1..3));
 102
 103 $code.=<<___;
 104 .globl  des_t4_cbc_encrypt
 105 .align  32
 106 des_t4_cbc_encrypt:
 107         cmp             $len, 0
 108         be,pn           $::size_t_cc, .Lcbc_abort
 109         srln            $len, 0, $len           ! needed on v8+, "nop" on v9
 110         ld              [$ivec + 0], %f0        ! load ivec
 111         ld              [$ivec + 4], %f1
 112
 113         and             $inp, 7, $ileft
 114         andn            $inp, 7, $inp
 115         sll             $ileft, 3, $ileft
 116         mov             0xff, $omask
 117         prefetch        [$inp], 20
 118         prefetch        [$inp + 63], 20
 119         sub             %g0, $ileft, $iright
 120         and             $out, 7, %g4
 121         alignaddrl      $out, %g0, $out
 122         srl             $omask, %g4, $omask
 123         srlx            $len, 3, $len
 124         movrz           %g4, 0, $omask
 125         prefetch        [$out], 22
 126
 127         ldd             [$key + 0x00], %f4      ! load key schedule
 128         ldd             [$key + 0x08], %f6
 129         ldd             [$key + 0x10], %f8
 130         ldd             [$key + 0x18], %f10
 131         ldd             [$key + 0x20], %f12
 132         ldd             [$key + 0x28], %f14
 133         ldd             [$key + 0x30], %f16
 134         ldd             [$key + 0x38], %f18
 135         ldd             [$key + 0x40], %f20
 136         ldd             [$key + 0x48], %f22
 137         ldd             [$key + 0x50], %f24
 138         ldd             [$key + 0x58], %f26
 139         ldd             [$key + 0x60], %f28
 140         ldd             [$key + 0x68], %f30
 141         ldd             [$key + 0x70], %f32
 142         ldd             [$key + 0x78], %f34
 143
 144 .Ldes_cbc_enc_loop:
 145         ldx             [$inp + 0], %g4
 146         brz,pt          $ileft, 4f
 147         nop
 148
 149         ldx             [$inp + 8], %g5
 150         sllx            %g4, $ileft, %g4
 151         srlx            %g5, $iright, %g5
 152         or              %g5, %g4, %g4
 153 4:
 154         movxtod         %g4, %f2
 155         prefetch        [$inp + 8+63], 20
 156         add             $inp, 8, $inp
 157         fxor            %f2, %f0, %f0           ! ^= ivec
 158         prefetch        [$out + 63], 22
 159
 160         des_ip          %f0, %f0
 161         des_round       %f4, %f6, %f0, %f0
 162         des_round       %f8, %f10, %f0, %f0
 163         des_round       %f12, %f14, %f0, %f0
 164         des_round       %f16, %f18, %f0, %f0
 165         des_round       %f20, %f22, %f0, %f0
 166         des_round       %f24, %f26, %f0, %f0
 167         des_round       %f28, %f30, %f0, %f0
 168         des_round       %f32, %f34, %f0, %f0
 169         des_iip         %f0, %f0
 170
 171         brnz,pn         $omask, 2f
 172         sub             $len, 1, $len
 173
 174         std             %f0, [$out + 0]
 175         brnz,pt         $len, .Ldes_cbc_enc_loop
 176         add             $out, 8, $out
 177
 178         st              %f0, [$ivec + 0]        ! write out ivec
 179         retl
 180         st              %f1, [$ivec + 4]
 181 .Lcbc_abort:
 182         retl
 183         nop
 184
 185 .align  16
 186 2:      ldxa            [$inp]0x82, %g4         ! avoid read-after-write hazard
 187                                                 ! and ~4x deterioration
 188                                                 ! in inp==out case
 189         faligndata      %f0, %f0, %f2           ! handle unaligned output
 190
 191         stda            %f2, [$out + $omask]0xc0        ! partial store
 192         add             $out, 8, $out
 193         orn             %g0, $omask, $omask
 194         stda            %f2, [$out + $omask]0xc0        ! partial store
 195
 196         brnz,pt         $len, .Ldes_cbc_enc_loop+4
 197         orn             %g0, $omask, $omask
 198
 199         st              %f0, [$ivec + 0]        ! write out ivec
 200         retl
 201         st              %f1, [$ivec + 4]
 202 .type   des_t4_cbc_encrypt,#function
 203 .size   des_t4_cbc_encrypt,.-des_t4_cbc_encrypt
 204
 205 .globl  des_t4_cbc_decrypt
 206 .align  32
 207 des_t4_cbc_decrypt:
 208         cmp             $len, 0
 209         be,pn           $::size_t_cc, .Lcbc_abort
 210         srln            $len, 0, $len           ! needed on v8+, "nop" on v9
 211         ld              [$ivec + 0], %f2        ! load ivec
 212         ld              [$ivec + 4], %f3
 213
 214         and             $inp, 7, $ileft
 215         andn            $inp, 7, $inp
 216         sll             $ileft, 3, $ileft
 217         mov             0xff, $omask
 218         prefetch        [$inp], 20
 219         prefetch        [$inp + 63], 20
 220         sub             %g0, $ileft, $iright
 221         and             $out, 7, %g4
 222         alignaddrl      $out, %g0, $out
 223         srl             $omask, %g4, $omask
 224         srlx            $len, 3, $len
 225         movrz           %g4, 0, $omask
 226         prefetch        [$out], 22
 227
 228         ldd             [$key + 0x78], %f4      ! load key schedule
 229         ldd             [$key + 0x70], %f6
 230         ldd             [$key + 0x68], %f8
 231         ldd             [$key + 0x60], %f10
 232         ldd             [$key + 0x58], %f12
 233         ldd             [$key + 0x50], %f14
 234         ldd             [$key + 0x48], %f16
 235         ldd             [$key + 0x40], %f18
 236         ldd             [$key + 0x38], %f20
 237         ldd             [$key + 0x30], %f22
 238         ldd             [$key + 0x28], %f24
 239         ldd             [$key + 0x20], %f26
 240         ldd             [$key + 0x18], %f28
 241         ldd             [$key + 0x10], %f30
 242         ldd             [$key + 0x08], %f32
 243         ldd             [$key + 0x00], %f34
 244
 245 .Ldes_cbc_dec_loop:
 246         ldx             [$inp + 0], %g4
 247         brz,pt          $ileft, 4f
 248         nop
 249
 250         ldx             [$inp + 8], %g5
 251         sllx            %g4, $ileft, %g4
 252         srlx            %g5, $iright, %g5
 253         or              %g5, %g4, %g4
 254 4:
 255         movxtod         %g4, %f0
 256         prefetch        [$inp + 8+63], 20
 257         add             $inp, 8, $inp
 258         prefetch        [$out + 63], 22
 259
 260         des_ip          %f0, %f0
 261         des_round       %f4, %f6, %f0, %f0
 262         des_round       %f8, %f10, %f0, %f0
 263         des_round       %f12, %f14, %f0, %f0
 264         des_round       %f16, %f18, %f0, %f0
 265         des_round       %f20, %f22, %f0, %f0
 266         des_round       %f24, %f26, %f0, %f0
 267         des_round       %f28, %f30, %f0, %f0
 268         des_round       %f32, %f34, %f0, %f0
 269         des_iip         %f0, %f0
 270
 271         fxor            %f2, %f0, %f0           ! ^= ivec
 272         movxtod         %g4, %f2
 273
 274         brnz,pn         $omask, 2f
 275         sub             $len, 1, $len
 276
 277         std             %f0, [$out + 0]
 278         brnz,pt         $len, .Ldes_cbc_dec_loop
 279         add             $out, 8, $out
 280
 281         st              %f2, [$ivec + 0]        ! write out ivec
 282         retl
 283         st              %f3, [$ivec + 4]
 284
 285 .align  16
 286 2:      ldxa            [$inp]0x82, %g4         ! avoid read-after-write hazard
 287                                                 ! and ~4x deterioration
 288                                                 ! in inp==out case
 289         faligndata      %f0, %f0, %f0           ! handle unaligned output
 290
 291         stda            %f0, [$out + $omask]0xc0        ! partial store
 292         add             $out, 8, $out
 293         orn             %g0, $omask, $omask
 294         stda            %f0, [$out + $omask]0xc0        ! partial store
 295
 296         brnz,pt         $len, .Ldes_cbc_dec_loop+4
 297         orn             %g0, $omask, $omask
 298
 299         st              %f2, [$ivec + 0]        ! write out ivec
 300         retl
 301         st              %f3, [$ivec + 4]
 302 .type   des_t4_cbc_decrypt,#function
 303 .size   des_t4_cbc_decrypt,.-des_t4_cbc_decrypt
 304 ___
 305
 306 # One might wonder why does one have back-to-back des_iip/des_ip
 307 # pairs between EDE passes. Indeed, aren't they inverse of each other?
 308 # They almost are. Outcome of the pair is 32-bit words being swapped
 309 # in target register. Consider pair of des_iip/des_ip as a way to
 310 # perform the due swap, it's actually fastest way in this case.
 311
 312 $code.=<<___;
 313 .globl  des_t4_ede3_cbc_encrypt
 314 .align  32
 315 des_t4_ede3_cbc_encrypt:
 316         cmp             $len, 0
 317         be,pn           $::size_t_cc, .Lcbc_abort
 318         srln            $len, 0, $len           ! needed on v8+, "nop" on v9
 319         ld              [$ivec + 0], %f0        ! load ivec
 320         ld              [$ivec + 4], %f1
 321
 322         and             $inp, 7, $ileft
 323         andn            $inp, 7, $inp
 324         sll             $ileft, 3, $ileft
 325         mov             0xff, $omask
 326         prefetch        [$inp], 20
 327         prefetch        [$inp + 63], 20
 328         sub             %g0, $ileft, $iright
 329         and             $out, 7, %g4
 330         alignaddrl      $out, %g0, $out
 331         srl             $omask, %g4, $omask
 332         srlx            $len, 3, $len
 333         movrz           %g4, 0, $omask
 334         prefetch        [$out], 22
 335
 336         ldd             [$key + 0x00], %f4      ! load key schedule
 337         ldd             [$key + 0x08], %f6
 338         ldd             [$key + 0x10], %f8
 339         ldd             [$key + 0x18], %f10
 340         ldd             [$key + 0x20], %f12
 341         ldd             [$key + 0x28], %f14
 342         ldd             [$key + 0x30], %f16
 343         ldd             [$key + 0x38], %f18
 344         ldd             [$key + 0x40], %f20
 345         ldd             [$key + 0x48], %f22
 346         ldd             [$key + 0x50], %f24
 347         ldd             [$key + 0x58], %f26
 348         ldd             [$key + 0x60], %f28
 349         ldd             [$key + 0x68], %f30
 350         ldd             [$key + 0x70], %f32
 351         ldd             [$key + 0x78], %f34
 352
 353 .Ldes_ede3_cbc_enc_loop:
 354         ldx             [$inp + 0], %g4
 355         brz,pt          $ileft, 4f
 356         nop
 357
 358         ldx             [$inp + 8], %g5
 359         sllx            %g4, $ileft, %g4
 360         srlx            %g5, $iright, %g5
 361         or              %g5, %g4, %g4
 362 4:
 363         movxtod         %g4, %f2
 364         prefetch        [$inp + 8+63], 20
 365         add             $inp, 8, $inp
 366         fxor            %f2, %f0, %f0           ! ^= ivec
 367         prefetch        [$out + 63], 22
 368
 369         des_ip          %f0, %f0
 370         des_round       %f4, %f6, %f0, %f0
 371         des_round       %f8, %f10, %f0, %f0
 372         des_round       %f12, %f14, %f0, %f0
 373         des_round       %f16, %f18, %f0, %f0
 374         ldd             [$key + 0x100-0x08], %f36
 375         ldd             [$key + 0x100-0x10], %f38
 376         des_round       %f20, %f22, %f0, %f0
 377         ldd             [$key + 0x100-0x18], %f40
 378         ldd             [$key + 0x100-0x20], %f42
 379         des_round       %f24, %f26, %f0, %f0
 380         ldd             [$key + 0x100-0x28], %f44
 381         ldd             [$key + 0x100-0x30], %f46
 382         des_round       %f28, %f30, %f0, %f0
 383         ldd             [$key + 0x100-0x38], %f48
 384         ldd             [$key + 0x100-0x40], %f50
 385         des_round       %f32, %f34, %f0, %f0
 386         ldd             [$key + 0x100-0x48], %f52
 387         ldd             [$key + 0x100-0x50], %f54
 388         des_iip         %f0, %f0
 389
 390         ldd             [$key + 0x100-0x58], %f56
 391         ldd             [$key + 0x100-0x60], %f58
 392         des_ip          %f0, %f0
 393         ldd             [$key + 0x100-0x68], %f60
 394         ldd             [$key + 0x100-0x70], %f62
 395         des_round       %f36, %f38, %f0, %f0
 396         ldd             [$key + 0x100-0x78], %f36
 397         ldd             [$key + 0x100-0x80], %f38
 398         des_round       %f40, %f42, %f0, %f0
 399         des_round       %f44, %f46, %f0, %f0
 400         des_round       %f48, %f50, %f0, %f0
 401         ldd             [$key + 0x100+0x00], %f40
 402         ldd             [$key + 0x100+0x08], %f42
 403         des_round       %f52, %f54, %f0, %f0
 404         ldd             [$key + 0x100+0x10], %f44
 405         ldd             [$key + 0x100+0x18], %f46
 406         des_round       %f56, %f58, %f0, %f0
 407         ldd             [$key + 0x100+0x20], %f48
 408         ldd             [$key + 0x100+0x28], %f50
 409         des_round       %f60, %f62, %f0, %f0
 410         ldd             [$key + 0x100+0x30], %f52
 411         ldd             [$key + 0x100+0x38], %f54
 412         des_round       %f36, %f38, %f0, %f0
 413         ldd             [$key + 0x100+0x40], %f56
 414         ldd             [$key + 0x100+0x48], %f58
 415         des_iip         %f0, %f0
 416
 417         ldd             [$key + 0x100+0x50], %f60
 418         ldd             [$key + 0x100+0x58], %f62
 419         des_ip          %f0, %f0
 420         ldd             [$key + 0x100+0x60], %f36
 421         ldd             [$key + 0x100+0x68], %f38
 422         des_round       %f40, %f42, %f0, %f0
 423         ldd             [$key + 0x100+0x70], %f40
 424         ldd             [$key + 0x100+0x78], %f42
 425         des_round       %f44, %f46, %f0, %f0
 426         des_round       %f48, %f50, %f0, %f0
 427         des_round       %f52, %f54, %f0, %f0
 428         des_round       %f56, %f58, %f0, %f0
 429         des_round       %f60, %f62, %f0, %f0
 430         des_round       %f36, %f38, %f0, %f0
 431         des_round       %f40, %f42, %f0, %f0
 432         des_iip         %f0, %f0
 433
 434         brnz,pn         $omask, 2f
 435         sub             $len, 1, $len
 436
 437         std             %f0, [$out + 0]
 438         brnz,pt         $len, .Ldes_ede3_cbc_enc_loop
 439         add             $out, 8, $out
 440
 441         st              %f0, [$ivec + 0]        ! write out ivec
 442         retl
 443         st              %f1, [$ivec + 4]
 444
 445 .align  16
 446 2:      ldxa            [$inp]0x82, %g4         ! avoid read-after-write hazard
 447                                                 ! and ~2x deterioration
 448                                                 ! in inp==out case
 449         faligndata      %f0, %f0, %f2           ! handle unaligned output
 450
 451         stda            %f2, [$out + $omask]0xc0        ! partial store
 452         add             $out, 8, $out
 453         orn             %g0, $omask, $omask
 454         stda            %f2, [$out + $omask]0xc0        ! partial store
 455
 456         brnz,pt         $len, .Ldes_ede3_cbc_enc_loop+4
 457         orn             %g0, $omask, $omask
 458
 459         st              %f0, [$ivec + 0]        ! write out ivec
 460         retl
 461         st              %f1, [$ivec + 4]
 462 .type   des_t4_ede3_cbc_encrypt,#function
 463 .size   des_t4_ede3_cbc_encrypt,.-des_t4_ede3_cbc_encrypt
 464
 465 .globl  des_t4_ede3_cbc_decrypt
 466 .align  32
 467 des_t4_ede3_cbc_decrypt:
 468         cmp             $len, 0
 469         be,pn           $::size_t_cc, .Lcbc_abort
 470         srln            $len, 0, $len           ! needed on v8+, "nop" on v9
 471         ld              [$ivec + 0], %f2        ! load ivec
 472         ld              [$ivec + 4], %f3
 473
 474         and             $inp, 7, $ileft
 475         andn            $inp, 7, $inp
 476         sll             $ileft, 3, $ileft
 477         mov             0xff, $omask
 478         prefetch        [$inp], 20
 479         prefetch        [$inp + 63], 20
 480         sub             %g0, $ileft, $iright
 481         and             $out, 7, %g4
 482         alignaddrl      $out, %g0, $out
 483         srl             $omask, %g4, $omask
 484         srlx            $len, 3, $len
 485         movrz           %g4, 0, $omask
 486         prefetch        [$out], 22
 487
 488         ldd             [$key + 0x100+0x78], %f4        ! load key schedule
 489         ldd             [$key + 0x100+0x70], %f6
 490         ldd             [$key + 0x100+0x68], %f8
 491         ldd             [$key + 0x100+0x60], %f10
 492         ldd             [$key + 0x100+0x58], %f12
 493         ldd             [$key + 0x100+0x50], %f14
 494         ldd             [$key + 0x100+0x48], %f16
 495         ldd             [$key + 0x100+0x40], %f18
 496         ldd             [$key + 0x100+0x38], %f20
 497         ldd             [$key + 0x100+0x30], %f22
 498         ldd             [$key + 0x100+0x28], %f24
 499         ldd             [$key + 0x100+0x20], %f26
 500         ldd             [$key + 0x100+0x18], %f28
 501         ldd             [$key + 0x100+0x10], %f30
 502         ldd             [$key + 0x100+0x08], %f32
 503         ldd             [$key + 0x100+0x00], %f34
 504
 505 .Ldes_ede3_cbc_dec_loop:
 506         ldx             [$inp + 0], %g4
 507         brz,pt          $ileft, 4f
 508         nop
 509
 510         ldx             [$inp + 8], %g5
 511         sllx            %g4, $ileft, %g4
 512         srlx            %g5, $iright, %g5
 513         or              %g5, %g4, %g4
 514 4:
 515         movxtod         %g4, %f0
 516         prefetch        [$inp + 8+63], 20
 517         add             $inp, 8, $inp
 518         prefetch        [$out + 63], 22
 519
 520         des_ip          %f0, %f0
 521         des_round       %f4, %f6, %f0, %f0
 522         des_round       %f8, %f10, %f0, %f0
 523         des_round       %f12, %f14, %f0, %f0
 524         des_round       %f16, %f18, %f0, %f0
 525         ldd             [$key + 0x80+0x00], %f36
 526         ldd             [$key + 0x80+0x08], %f38
 527         des_round       %f20, %f22, %f0, %f0
 528         ldd             [$key + 0x80+0x10], %f40
 529         ldd             [$key + 0x80+0x18], %f42
 530         des_round       %f24, %f26, %f0, %f0
 531         ldd             [$key + 0x80+0x20], %f44
 532         ldd             [$key + 0x80+0x28], %f46
 533         des_round       %f28, %f30, %f0, %f0
 534         ldd             [$key + 0x80+0x30], %f48
 535         ldd             [$key + 0x80+0x38], %f50
 536         des_round       %f32, %f34, %f0, %f0
 537         ldd             [$key + 0x80+0x40], %f52
 538         ldd             [$key + 0x80+0x48], %f54
 539         des_iip         %f0, %f0
 540
 541         ldd             [$key + 0x80+0x50], %f56
 542         ldd             [$key + 0x80+0x58], %f58
 543         des_ip          %f0, %f0
 544         ldd             [$key + 0x80+0x60], %f60
 545         ldd             [$key + 0x80+0x68], %f62
 546         des_round       %f36, %f38, %f0, %f0
 547         ldd             [$key + 0x80+0x70], %f36
 548         ldd             [$key + 0x80+0x78], %f38
 549         des_round       %f40, %f42, %f0, %f0
 550         des_round       %f44, %f46, %f0, %f0
 551         des_round       %f48, %f50, %f0, %f0
 552         ldd             [$key + 0x80-0x08], %f40
 553         ldd             [$key + 0x80-0x10], %f42
 554         des_round       %f52, %f54, %f0, %f0
 555         ldd             [$key + 0x80-0x18], %f44
 556         ldd             [$key + 0x80-0x20], %f46
 557         des_round       %f56, %f58, %f0, %f0
 558         ldd             [$key + 0x80-0x28], %f48
 559         ldd             [$key + 0x80-0x30], %f50
 560         des_round       %f60, %f62, %f0, %f0
 561         ldd             [$key + 0x80-0x38], %f52
 562         ldd             [$key + 0x80-0x40], %f54
 563         des_round       %f36, %f38, %f0, %f0
 564         ldd             [$key + 0x80-0x48], %f56
 565         ldd             [$key + 0x80-0x50], %f58
 566         des_iip         %f0, %f0
 567
 568         ldd             [$key + 0x80-0x58], %f60
 569         ldd             [$key + 0x80-0x60], %f62
 570         des_ip          %f0, %f0
 571         ldd             [$key + 0x80-0x68], %f36
 572         ldd             [$key + 0x80-0x70], %f38
 573         des_round       %f40, %f42, %f0, %f0
 574         ldd             [$key + 0x80-0x78], %f40
 575         ldd             [$key + 0x80-0x80], %f42
 576         des_round       %f44, %f46, %f0, %f0
 577         des_round       %f48, %f50, %f0, %f0
 578         des_round       %f52, %f54, %f0, %f0
 579         des_round       %f56, %f58, %f0, %f0
 580         des_round       %f60, %f62, %f0, %f0
 581         des_round       %f36, %f38, %f0, %f0
 582         des_round       %f40, %f42, %f0, %f0
 583         des_iip         %f0, %f0
 584
 585         fxor            %f2, %f0, %f0           ! ^= ivec
 586         movxtod         %g4, %f2
 587
 588         brnz,pn         $omask, 2f
 589         sub             $len, 1, $len
 590
 591         std             %f0, [$out + 0]
 592         brnz,pt         $len, .Ldes_ede3_cbc_dec_loop
 593         add             $out, 8, $out
 594
 595         st              %f2, [$ivec + 0]        ! write out ivec
 596         retl
 597         st              %f3, [$ivec + 4]
 598
 599 .align  16
 600 2:      ldxa            [$inp]0x82, %g4         ! avoid read-after-write hazard
 601                                                 ! and ~3x deterioration
 602                                                 ! in inp==out case
 603         faligndata      %f0, %f0, %f0           ! handle unaligned output
 604
 605         stda            %f0, [$out + $omask]0xc0        ! partial store
 606         add             $out, 8, $out
 607         orn             %g0, $omask, $omask
 608         stda            %f0, [$out + $omask]0xc0        ! partial store
 609
 610         brnz,pt         $len, .Ldes_ede3_cbc_dec_loop+4
 611         orn             %g0, $omask, $omask
 612
 613         st              %f2, [$ivec + 0]        ! write out ivec
 614         retl
 615         st              %f3, [$ivec + 4]
 616 .type   des_t4_ede3_cbc_decrypt,#function
 617 .size   des_t4_ede3_cbc_decrypt,.-des_t4_ede3_cbc_decrypt
 618 ___
 619 }
 620 $code.=<<___;
 621 .asciz  "DES for SPARC T4, David S. Miller, Andy Polyakov"
 622 .align  4
 623 ___
 624
 625 &emit_assembler();
 626
 627 close STDOUT;