crypto/des/asm/des_enc.m4

   1 ! Copyright 2000-2016 The OpenSSL Project Authors. All Rights Reserved.
   2 !
   3 ! Licensed under the OpenSSL license (the "License").  You may not use
   4 ! this file except in compliance with the License.  You can obtain a copy
   5 ! in the file LICENSE in the source distribution or at
   6 ! https://www.openssl.org/source/license.html
   7 !
   8 !  To expand the m4 macros: m4 -B 8192 des_enc.m4 > des_enc.S
   9 !
  10 !  Global registers 1 to 5 are used. This is the same as done by the
  11 !  cc compiler. The UltraSPARC load/store little endian feature is used.
  12 !
  13 !  Instruction grouping often refers to one CPU cycle.
  14 !
  15 !  Assemble through gcc: gcc -c -mcpu=ultrasparc -o des_enc.o des_enc.S
  16 !
  17 !  Assemble through cc:  cc -c -xarch=v8plusa -o des_enc.o des_enc.S
  18 !
  19 !  Performance improvement according to './apps/openssl speed des'
  20 !
  21 !       32-bit build:
  22 !               23%  faster than cc-5.2 -xarch=v8plus -xO5
  23 !               115% faster than gcc-3.2.1 -m32 -mcpu=ultrasparc -O5
  24 !       64-bit build:
  25 !               50%  faster than cc-5.2 -xarch=v9 -xO5
  26 !               100% faster than gcc-3.2.1 -m64 -mcpu=ultrasparc -O5
  27 !
  28
  29 .ident "des_enc.m4 2.1"
  30 .file  "des_enc-sparc.S"
  31
  32 #include <openssl/opensslconf.h>
  33
  34 #if defined(__SUNPRO_C) && defined(__sparcv9)
  35 # define ABI64  /* They've said -xarch=v9 at command line */
  36 #elif defined(__GNUC__) && defined(__arch64__)
  37 # define ABI64  /* They've said -m64 at command line */
  38 #endif
  39
  40 #ifdef ABI64
  41   .register     %g2,#scratch
  42   .register     %g3,#scratch
  43 # define        FRAME   -192
  44 # define        BIAS    2047
  45 # define        LDPTR   ldx
  46 # define        STPTR   stx
  47 # define        ARG0    128
  48 # define        ARGSZ   8
  49 #else
  50 # define        FRAME   -96
  51 # define        BIAS    0
  52 # define        LDPTR   ld
  53 # define        STPTR   st
  54 # define        ARG0    68
  55 # define        ARGSZ   4
  56 #endif
  57
  58 #define LOOPS 7
  59
  60 #define global0 %g0
  61 #define global1 %g1
  62 #define global2 %g2
  63 #define global3 %g3
  64 #define global4 %g4
  65 #define global5 %g5
  66
  67 #define local0 %l0
  68 #define local1 %l1
  69 #define local2 %l2
  70 #define local3 %l3
  71 #define local4 %l4
  72 #define local5 %l5
  73 #define local7 %l6
  74 #define local6 %l7
  75
  76 #define in0 %i0
  77 #define in1 %i1
  78 #define in2 %i2
  79 #define in3 %i3
  80 #define in4 %i4
  81 #define in5 %i5
  82 #define in6 %i6
  83 #define in7 %i7
  84
  85 #define out0 %o0
  86 #define out1 %o1
  87 #define out2 %o2
  88 #define out3 %o3
  89 #define out4 %o4
  90 #define out5 %o5
  91 #define out6 %o6
  92 #define out7 %o7
  93
  94 #define stub stb
  95
  96 changequote({,})
  97
  98
  99 ! Macro definitions:
 100
 101
 102 ! {ip_macro}
 103 !
 104 ! The logic used in initial and final permutations is the same as in
 105 ! the C code. The permutations are done with a clever shift, xor, and
 106 ! technique.
 107 !
 108 ! The macro also loads address sbox 1 to 5 to global 1 to 5, address
 109 ! sbox 6 to local6, and addres sbox 8 to out3.
 110 !
 111 ! Rotates the halfs 3 left to bring the sbox bits in convenient positions.
 112 !
 113 ! Loads key first round from address in parameter 5 to out0, out1.
 114 !
 115 ! After the the original LibDES initial permutation, the resulting left
 116 ! is in the variable initially used for right and vice versa. The macro
 117 ! implements the possibility to keep the halfs in the original registers.
 118 !
 119 ! parameter 1  left
 120 ! parameter 2  right
 121 ! parameter 3  result left (modify in first round)
 122 ! parameter 4  result right (use in first round)
 123 ! parameter 5  key address
 124 ! parameter 6  1/2 for include encryption/decryption
 125 ! parameter 7  1 for move in1 to in3
 126 ! parameter 8  1 for move in3 to in4, 2 for move in4 to in3
 127 ! parameter 9  1 for load ks3 and ks2 to in4 and in3
 128
 129 define(ip_macro, {
 130
 131 ! {ip_macro}
 132 ! $1 $2 $4 $3 $5 $6 $7 $8 $9
 133
 134         ld      [out2+256], local1
 135         srl     $2, 4, local4
 136
 137         xor     local4, $1, local4
 138         ifelse($7,1,{mov in1, in3},{nop})
 139
 140         ld      [out2+260], local2
 141         and     local4, local1, local4
 142         ifelse($8,1,{mov in3, in4},{})
 143         ifelse($8,2,{mov in4, in3},{})
 144
 145         ld      [out2+280], out4          ! loop counter
 146         sll     local4, 4, local1
 147         xor     $1, local4, $1
 148
 149         ld      [out2+264], local3
 150         srl     $1, 16, local4
 151         xor     $2, local1, $2
 152
 153         ifelse($9,1,{LDPTR      KS3, in4},{})
 154         xor     local4, $2, local4
 155         nop     !sethi  %hi(DES_SPtrans), global1 ! sbox addr
 156
 157         ifelse($9,1,{LDPTR      KS2, in3},{})
 158         and     local4, local2, local4
 159         nop     !or     global1, %lo(DES_SPtrans), global1   ! sbox addr
 160
 161         sll     local4, 16, local1
 162         xor     $2, local4, $2
 163
 164         srl     $2, 2, local4
 165         xor     $1, local1, $1
 166
 167         sethi   %hi(16711680), local5
 168         xor     local4, $1, local4
 169
 170         and     local4, local3, local4
 171         or      local5, 255, local5
 172
 173         sll     local4, 2, local2
 174         xor     $1, local4, $1
 175
 176         srl     $1, 8, local4
 177         xor     $2, local2, $2
 178
 179         xor     local4, $2, local4
 180         add     global1, 768, global4
 181
 182         and     local4, local5, local4
 183         add     global1, 1024, global5
 184
 185         ld      [out2+272], local7
 186         sll     local4, 8, local1
 187         xor     $2, local4, $2
 188
 189         srl     $2, 1, local4
 190         xor     $1, local1, $1
 191
 192         ld      [$5], out0                ! key 7531
 193         xor     local4, $1, local4
 194         add     global1, 256, global2
 195
 196         ld      [$5+4], out1              ! key 8642
 197         and     local4, local7, local4
 198         add     global1, 512, global3
 199
 200         sll     local4, 1, local1
 201         xor     $1, local4, $1
 202
 203         sll     $1, 3, local3
 204         xor     $2, local1, $2
 205
 206         sll     $2, 3, local2
 207         add     global1, 1280, local6     ! address sbox 8
 208
 209         srl     $1, 29, local4
 210         add     global1, 1792, out3       ! address sbox 8
 211
 212         srl     $2, 29, local1
 213         or      local4, local3, $4
 214
 215         or      local2, local1, $3
 216
 217         ifelse($6, 1, {
 218
 219                 ld      [out2+284], local5     ! 0x0000FC00 used in the rounds
 220                 or      local2, local1, $3
 221                 xor     $4, out0, local1
 222
 223                 call .des_enc.1
 224                 and     local1, 252, local1
 225
 226         },{})
 227
 228         ifelse($6, 2, {
 229
 230                 ld      [out2+284], local5     ! 0x0000FC00 used in the rounds
 231                 or      local2, local1, $3
 232                 xor     $4, out0, local1
 233
 234                 call .des_dec.1
 235                 and     local1, 252, local1
 236
 237         },{})
 238 })
 239
 240
 241 ! {rounds_macro}
 242 !
 243 ! The logic used in the DES rounds is the same as in the C code,
 244 ! except that calculations for sbox 1 and sbox 5 begin before
 245 ! the previous round is finished.
 246 !
 247 ! In each round one half (work) is modified based on key and the
 248 ! other half (use).
 249 !
 250 ! In this version we do two rounds in a loop repeated 7 times
 251 ! and two rounds separately.
 252 !
 253 ! One half has the bits for the sboxes in the following positions:
 254 !
 255 !       777777xx555555xx333333xx111111xx
 256 !
 257 !       88xx666666xx444444xx222222xx8888
 258 !
 259 ! The bits for each sbox are xor-ed with the key bits for that box.
 260 ! The above xx bits are cleared, and the result used for lookup in
 261 ! the sbox table. Each sbox entry contains the 4 output bits permuted
 262 ! into 32 bits according to the P permutation.
 263 !
 264 ! In the description of DES, left and right are switched after
 265 ! each round, except after last round. In this code the original
 266 ! left and right are kept in the same register in all rounds, meaning
 267 ! that after the 16 rounds the result for right is in the register
 268 ! originally used for left.
 269 !
 270 ! parameter 1  first work (left in first round)
 271 ! parameter 2  first use (right in first round)
 272 ! parameter 3  enc/dec  1/-1
 273 ! parameter 4  loop label
 274 ! parameter 5  key address register
 275 ! parameter 6  optional address for key next encryption/decryption
 276 ! parameter 7  not empty for include retl
 277 !
 278 ! also compares in2 to 8
 279
 280 define(rounds_macro, {
 281
 282 ! {rounds_macro}
 283 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 284
 285         xor     $2, out0, local1
 286
 287         ld      [out2+284], local5        ! 0x0000FC00
 288         ba      $4
 289         and     local1, 252, local1
 290
 291         .align 32
 292
 293 $4:
 294         ! local6 is address sbox 6
 295         ! out3   is address sbox 8
 296         ! out4   is loop counter
 297
 298         ld      [global1+local1], local1
 299         xor     $2, out1, out1            ! 8642
 300         xor     $2, out0, out0            ! 7531
 301         ! fmovs %f0, %f0                  ! fxor used for alignment
 302
 303         srl     out1, 4, local0           ! rotate 4 right
 304         and     out0, local5, local3      ! 3
 305         ! fmovs %f0, %f0
 306
 307         ld      [$5+$3*8], local7         ! key 7531 next round
 308         srl     local3, 8, local3         ! 3
 309         and     local0, 252, local2       ! 2
 310         ! fmovs %f0, %f0
 311
 312         ld      [global3+local3],local3   ! 3
 313         sll     out1, 28, out1            ! rotate
 314         xor     $1, local1, $1            ! 1 finished, local1 now sbox 7
 315
 316         ld      [global2+local2], local2  ! 2
 317         srl     out0, 24, local1          ! 7
 318         or      out1, local0, out1        ! rotate
 319
 320         ldub    [out2+local1], local1     ! 7 (and 0xFC)
 321         srl     out1, 24, local0          ! 8
 322         and     out1, local5, local4      ! 4
 323
 324         ldub    [out2+local0], local0     ! 8 (and 0xFC)
 325         srl     local4, 8, local4         ! 4
 326         xor     $1, local2, $1            ! 2 finished local2 now sbox 6
 327
 328         ld      [global4+local4],local4   ! 4
 329         srl     out1, 16, local2          ! 6
 330         xor     $1, local3, $1            ! 3 finished local3 now sbox 5
 331
 332         ld      [out3+local0],local0      ! 8
 333         and     local2, 252, local2       ! 6
 334         add     global1, 1536, local5     ! address sbox 7
 335
 336         ld      [local6+local2], local2   ! 6
 337         srl     out0, 16, local3          ! 5
 338         xor     $1, local4, $1            ! 4 finished
 339
 340         ld      [local5+local1],local1    ! 7
 341         and     local3, 252, local3       ! 5
 342         xor     $1, local0, $1            ! 8 finished
 343
 344         ld      [global5+local3],local3   ! 5
 345         xor     $1, local2, $1            ! 6 finished
 346         subcc   out4, 1, out4
 347
 348         ld      [$5+$3*8+4], out0         ! key 8642 next round
 349         xor     $1, local7, local2        ! sbox 5 next round
 350         xor     $1, local1, $1            ! 7 finished
 351
 352         srl     local2, 16, local2        ! sbox 5 next round
 353         xor     $1, local3, $1            ! 5 finished
 354
 355         ld      [$5+$3*16+4], out1        ! key 8642 next round again
 356         and     local2, 252, local2       ! sbox5 next round
 357 ! next round
 358         xor     $1, local7, local7        ! 7531
 359
 360         ld      [global5+local2], local2  ! 5
 361         srl     local7, 24, local3        ! 7
 362         xor     $1, out0, out0            ! 8642
 363
 364         ldub    [out2+local3], local3     ! 7 (and 0xFC)
 365         srl     out0, 4, local0           ! rotate 4 right
 366         and     local7, 252, local1       ! 1
 367
 368         sll     out0, 28, out0            ! rotate
 369         xor     $2, local2, $2            ! 5 finished local2 used
 370
 371         srl     local0, 8, local4         ! 4
 372         and     local0, 252, local2       ! 2
 373         ld      [local5+local3], local3   ! 7
 374
 375         srl     local0, 16, local5        ! 6
 376         or      out0, local0, out0        ! rotate
 377         ld      [global2+local2], local2  ! 2
 378
 379         srl     out0, 24, local0
 380         ld      [$5+$3*16], out0          ! key 7531 next round
 381         and     local4, 252, local4       ! 4
 382
 383         and     local5, 252, local5       ! 6
 384         ld      [global4+local4], local4  ! 4
 385         xor     $2, local3, $2            ! 7 finished local3 used
 386
 387         and     local0, 252, local0       ! 8
 388         ld      [local6+local5], local5   ! 6
 389         xor     $2, local2, $2            ! 2 finished local2 now sbox 3
 390
 391         srl     local7, 8, local2         ! 3 start
 392         ld      [out3+local0], local0     ! 8
 393         xor     $2, local4, $2            ! 4 finished
 394
 395         and     local2, 252, local2       ! 3
 396         ld      [global1+local1], local1  ! 1
 397         xor     $2, local5, $2            ! 6 finished local5 used
 398
 399         ld      [global3+local2], local2  ! 3
 400         xor     $2, local0, $2            ! 8 finished
 401         add     $5, $3*16, $5             ! enc add 8, dec add -8 to key pointer
 402
 403         ld      [out2+284], local5        ! 0x0000FC00
 404         xor     $2, out0, local4          ! sbox 1 next round
 405         xor     $2, local1, $2            ! 1 finished
 406
 407         xor     $2, local2, $2            ! 3 finished
 408         bne     $4
 409         and     local4, 252, local1       ! sbox 1 next round
 410
 411 ! two rounds more:
 412
 413         ld      [global1+local1], local1
 414         xor     $2, out1, out1
 415         xor     $2, out0, out0
 416
 417         srl     out1, 4, local0           ! rotate
 418         and     out0, local5, local3
 419
 420         ld      [$5+$3*8], local7         ! key 7531
 421         srl     local3, 8, local3
 422         and     local0, 252, local2
 423
 424         ld      [global3+local3],local3
 425         sll     out1, 28, out1            ! rotate
 426         xor     $1, local1, $1            ! 1 finished, local1 now sbox 7
 427
 428         ld      [global2+local2], local2
 429         srl     out0, 24, local1
 430         or      out1, local0, out1        ! rotate
 431
 432         ldub    [out2+local1], local1
 433         srl     out1, 24, local0
 434         and     out1, local5, local4
 435
 436         ldub    [out2+local0], local0
 437         srl     local4, 8, local4
 438         xor     $1, local2, $1            ! 2 finished local2 now sbox 6
 439
 440         ld      [global4+local4],local4
 441         srl     out1, 16, local2
 442         xor     $1, local3, $1            ! 3 finished local3 now sbox 5
 443
 444         ld      [out3+local0],local0
 445         and     local2, 252, local2
 446         add     global1, 1536, local5     ! address sbox 7
 447
 448         ld      [local6+local2], local2
 449         srl     out0, 16, local3
 450         xor     $1, local4, $1            ! 4 finished
 451
 452         ld      [local5+local1],local1
 453         and     local3, 252, local3
 454         xor     $1, local0, $1
 455
 456         ld      [global5+local3],local3
 457         xor     $1, local2, $1            ! 6 finished
 458         cmp     in2, 8
 459
 460         ifelse($6,{}, {}, {ld   [out2+280], out4})  ! loop counter
 461         xor     $1, local7, local2        ! sbox 5 next round
 462         xor     $1, local1, $1            ! 7 finished
 463
 464         ld      [$5+$3*8+4], out0
 465         srl     local2, 16, local2        ! sbox 5 next round
 466         xor     $1, local3, $1            ! 5 finished
 467
 468         and     local2, 252, local2
 469 ! next round (two rounds more)
 470         xor     $1, local7, local7        ! 7531
 471
 472         ld      [global5+local2], local2
 473         srl     local7, 24, local3
 474         xor     $1, out0, out0            ! 8642
 475
 476         ldub    [out2+local3], local3
 477         srl     out0, 4, local0           ! rotate
 478         and     local7, 252, local1
 479
 480         sll     out0, 28, out0            ! rotate
 481         xor     $2, local2, $2            ! 5 finished local2 used
 482
 483         srl     local0, 8, local4
 484         and     local0, 252, local2
 485         ld      [local5+local3], local3
 486
 487         srl     local0, 16, local5
 488         or      out0, local0, out0        ! rotate
 489         ld      [global2+local2], local2
 490
 491         srl     out0, 24, local0
 492         ifelse($6,{}, {}, {ld   [$6], out0})   ! key next encryption/decryption
 493         and     local4, 252, local4
 494
 495         and     local5, 252, local5
 496         ld      [global4+local4], local4
 497         xor     $2, local3, $2            ! 7 finished local3 used
 498
 499         and     local0, 252, local0
 500         ld      [local6+local5], local5
 501         xor     $2, local2, $2            ! 2 finished local2 now sbox 3
 502
 503         srl     local7, 8, local2         ! 3 start
 504         ld      [out3+local0], local0
 505         xor     $2, local4, $2
 506
 507         and     local2, 252, local2
 508         ld      [global1+local1], local1
 509         xor     $2, local5, $2            ! 6 finished local5 used
 510
 511         ld      [global3+local2], local2
 512         srl     $1, 3, local3
 513         xor     $2, local0, $2
 514
 515         ifelse($6,{}, {}, {ld   [$6+4], out1}) ! key next encryption/decryption
 516         sll     $1, 29, local4
 517         xor     $2, local1, $2
 518
 519         ifelse($7,{}, {}, {retl})
 520         xor     $2, local2, $2
 521 })
 522
 523
 524 ! {fp_macro}
 525 !
 526 !  parameter 1   right (original left)
 527 !  parameter 2   left (original right)
 528 !  parameter 3   1 for optional store to [in0]
 529 !  parameter 4   1 for load input/output address to local5/7
 530 !
 531 !  The final permutation logic switches the halves, meaning that
 532 !  left and right ends up the the registers originally used.
 533
 534 define(fp_macro, {
 535
 536 ! {fp_macro}
 537 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 538
 539         ! initially undo the rotate 3 left done after initial permutation
 540         ! original left is received shifted 3 right and 29 left in local3/4
 541
 542         sll     $2, 29, local1
 543         or      local3, local4, $1
 544
 545         srl     $2, 3, $2
 546         sethi   %hi(0x55555555), local2
 547
 548         or      $2, local1, $2
 549         or      local2, %lo(0x55555555), local2
 550
 551         srl     $2, 1, local3
 552         sethi   %hi(0x00ff00ff), local1
 553         xor     local3, $1, local3
 554         or      local1, %lo(0x00ff00ff), local1
 555         and     local3, local2, local3
 556         sethi   %hi(0x33333333), local4
 557         sll     local3, 1, local2
 558
 559         xor     $1, local3, $1
 560
 561         srl     $1, 8, local3
 562         xor     $2, local2, $2
 563         xor     local3, $2, local3
 564         or      local4, %lo(0x33333333), local4
 565         and     local3, local1, local3
 566         sethi   %hi(0x0000ffff), local1
 567         sll     local3, 8, local2
 568
 569         xor     $2, local3, $2
 570
 571         srl     $2, 2, local3
 572         xor     $1, local2, $1
 573         xor     local3, $1, local3
 574         or      local1, %lo(0x0000ffff), local1
 575         and     local3, local4, local3
 576         sethi   %hi(0x0f0f0f0f), local4
 577         sll     local3, 2, local2
 578
 579         ifelse($4,1, {LDPTR INPUT, local5})
 580         xor     $1, local3, $1
 581
 582         ifelse($4,1, {LDPTR OUTPUT, local7})
 583         srl     $1, 16, local3
 584         xor     $2, local2, $2
 585         xor     local3, $2, local3
 586         or      local4, %lo(0x0f0f0f0f), local4
 587         and     local3, local1, local3
 588         sll     local3, 16, local2
 589
 590         xor     $2, local3, local1
 591
 592         srl     local1, 4, local3
 593         xor     $1, local2, $1
 594         xor     local3, $1, local3
 595         and     local3, local4, local3
 596         sll     local3, 4, local2
 597
 598         xor     $1, local3, $1
 599
 600         ! optional store:
 601
 602         ifelse($3,1, {st $1, [in0]})
 603
 604         xor     local1, local2, $2
 605
 606         ifelse($3,1, {st $2, [in0+4]})
 607
 608 })
 609
 610
 611 ! {fp_ip_macro}
 612 !
 613 ! Does initial permutation for next block mixed with
 614 ! final permutation for current block.
 615 !
 616 ! parameter 1   original left
 617 ! parameter 2   original right
 618 ! parameter 3   left ip
 619 ! parameter 4   right ip
 620 ! parameter 5   1: load ks1/ks2 to in3/in4, add 120 to in4
 621 !                2: mov in4 to in3
 622 !
 623 ! also adds -8 to length in2 and loads loop counter to out4
 624
 625 define(fp_ip_macro, {
 626
 627 ! {fp_ip_macro}
 628 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 629
 630         define({temp1},{out4})
 631         define({temp2},{local3})
 632
 633         define({ip1},{local1})
 634         define({ip2},{local2})
 635         define({ip4},{local4})
 636         define({ip5},{local5})
 637
 638         ! $1 in local3, local4
 639
 640         ld      [out2+256], ip1
 641         sll     out5, 29, temp1
 642         or      local3, local4, $1
 643
 644         srl     out5, 3, $2
 645         ifelse($5,2,{mov in4, in3})
 646
 647         ld      [out2+272], ip5
 648         srl     $4, 4, local0
 649         or      $2, temp1, $2
 650
 651         srl     $2, 1, temp1
 652         xor     temp1, $1, temp1
 653
 654         and     temp1, ip5, temp1
 655         xor     local0, $3, local0
 656
 657         sll     temp1, 1, temp2
 658         xor     $1, temp1, $1
 659
 660         and     local0, ip1, local0
 661         add     in2, -8, in2
 662
 663         sll     local0, 4, local7
 664         xor     $3, local0, $3
 665
 666         ld      [out2+268], ip4
 667         srl     $1, 8, temp1
 668         xor     $2, temp2, $2
 669         ld      [out2+260], ip2
 670         srl     $3, 16, local0
 671         xor     $4, local7, $4
 672         xor     temp1, $2, temp1
 673         xor     local0, $4, local0
 674         and     temp1, ip4, temp1
 675         and     local0, ip2, local0
 676         sll     temp1, 8, temp2
 677         xor     $2, temp1, $2
 678         sll     local0, 16, local7
 679         xor     $4, local0, $4
 680
 681         srl     $2, 2, temp1
 682         xor     $1, temp2, $1
 683
 684         ld      [out2+264], temp2         ! ip3
 685         srl     $4, 2, local0
 686         xor     $3, local7, $3
 687         xor     temp1, $1, temp1
 688         xor     local0, $3, local0
 689         and     temp1, temp2, temp1
 690         and     local0, temp2, local0
 691         sll     temp1, 2, temp2
 692         xor     $1, temp1, $1
 693         sll     local0, 2, local7
 694         xor     $3, local0, $3
 695
 696         srl     $1, 16, temp1
 697         xor     $2, temp2, $2
 698         srl     $3, 8, local0
 699         xor     $4, local7, $4
 700         xor     temp1, $2, temp1
 701         xor     local0, $4, local0
 702         and     temp1, ip2, temp1
 703         and     local0, ip4, local0
 704         sll     temp1, 16, temp2
 705         xor     $2, temp1, local4
 706         sll     local0, 8, local7
 707         xor     $4, local0, $4
 708
 709         srl     $4, 1, local0
 710         xor     $3, local7, $3
 711
 712         srl     local4, 4, temp1
 713         xor     local0, $3, local0
 714
 715         xor     $1, temp2, $1
 716         and     local0, ip5, local0
 717
 718         sll     local0, 1, local7
 719         xor     temp1, $1, temp1
 720
 721         xor     $3, local0, $3
 722         xor     $4, local7, $4
 723
 724         sll     $3, 3, local5
 725         and     temp1, ip1, temp1
 726
 727         sll     temp1, 4, temp2
 728         xor     $1, temp1, $1
 729
 730         ifelse($5,1,{LDPTR      KS2, in4})
 731         sll     $4, 3, local2
 732         xor     local4, temp2, $2
 733
 734         ! reload since used as temporary:
 735
 736         ld      [out2+280], out4          ! loop counter
 737
 738         srl     $3, 29, local0
 739         ifelse($5,1,{add in4, 120, in4})
 740
 741         ifelse($5,1,{LDPTR      KS1, in3})
 742         srl     $4, 29, local7
 743
 744         or      local0, local5, $4
 745         or      local2, local7, $3
 746
 747 })
 748
 749
 750
 751 ! {load_little_endian}
 752 !
 753 ! parameter 1  address
 754 ! parameter 2  destination left
 755 ! parameter 3  destination right
 756 ! parameter 4  temporary
 757 ! parameter 5  label
 758
 759 define(load_little_endian, {
 760
 761 ! {load_little_endian}
 762 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 763
 764         ! first in memory to rightmost in register
 765
 766 $5:
 767         ldub    [$1+3], $2
 768
 769         ldub    [$1+2], $4
 770         sll     $2, 8, $2
 771         or      $2, $4, $2
 772
 773         ldub    [$1+1], $4
 774         sll     $2, 8, $2
 775         or      $2, $4, $2
 776
 777         ldub    [$1+0], $4
 778         sll     $2, 8, $2
 779         or      $2, $4, $2
 780
 781
 782         ldub    [$1+3+4], $3
 783
 784         ldub    [$1+2+4], $4
 785         sll     $3, 8, $3
 786         or      $3, $4, $3
 787
 788         ldub    [$1+1+4], $4
 789         sll     $3, 8, $3
 790         or      $3, $4, $3
 791
 792         ldub    [$1+0+4], $4
 793         sll     $3, 8, $3
 794         or      $3, $4, $3
 795 $5a:
 796
 797 })
 798
 799
 800 ! {load_little_endian_inc}
 801 !
 802 ! parameter 1  address
 803 ! parameter 2  destination left
 804 ! parameter 3  destination right
 805 ! parameter 4  temporary
 806 ! parameter 4  label
 807 !
 808 ! adds 8 to address
 809
 810 define(load_little_endian_inc, {
 811
 812 ! {load_little_endian_inc}
 813 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 814
 815         ! first in memory to rightmost in register
 816
 817 $5:
 818         ldub    [$1+3], $2
 819
 820         ldub    [$1+2], $4
 821         sll     $2, 8, $2
 822         or      $2, $4, $2
 823
 824         ldub    [$1+1], $4
 825         sll     $2, 8, $2
 826         or      $2, $4, $2
 827
 828         ldub    [$1+0], $4
 829         sll     $2, 8, $2
 830         or      $2, $4, $2
 831
 832         ldub    [$1+3+4], $3
 833         add     $1, 8, $1
 834
 835         ldub    [$1+2+4-8], $4
 836         sll     $3, 8, $3
 837         or      $3, $4, $3
 838
 839         ldub    [$1+1+4-8], $4
 840         sll     $3, 8, $3
 841         or      $3, $4, $3
 842
 843         ldub    [$1+0+4-8], $4
 844         sll     $3, 8, $3
 845         or      $3, $4, $3
 846 $5a:
 847
 848 })
 849
 850
 851 ! {load_n_bytes}
 852 !
 853 ! Loads 1 to 7 bytes little endian
 854 ! Remaining bytes are zeroed.
 855 !
 856 ! parameter 1  address
 857 ! parameter 2  length
 858 ! parameter 3  destination register left
 859 ! parameter 4  destination register right
 860 ! parameter 5  temp
 861 ! parameter 6  temp2
 862 ! parameter 7  label
 863 ! parameter 8  return label
 864
 865 define(load_n_bytes, {
 866
 867 ! {load_n_bytes}
 868 ! $1 $2 $5 $6 $7 $8 $7 $8 $9
 869
 870 $7.0:   call    .+8
 871         sll     $2, 2, $6
 872
 873         add     %o7,$7.jmp.table-$7.0,$5
 874
 875         add     $5, $6, $5
 876         mov     0, $4
 877
 878         ld      [$5], $5
 879
 880         jmp     %o7+$5
 881         mov     0, $3
 882
 883 $7.7:
 884         ldub    [$1+6], $5
 885         sll     $5, 16, $5
 886         or      $3, $5, $3
 887 $7.6:
 888         ldub    [$1+5], $5
 889         sll     $5, 8, $5
 890         or      $3, $5, $3
 891 $7.5:
 892         ldub    [$1+4], $5
 893         or      $3, $5, $3
 894 $7.4:
 895         ldub    [$1+3], $5
 896         sll     $5, 24, $5
 897         or      $4, $5, $4
 898 $7.3:
 899         ldub    [$1+2], $5
 900         sll     $5, 16, $5
 901         or      $4, $5, $4
 902 $7.2:
 903         ldub    [$1+1], $5
 904         sll     $5, 8, $5
 905         or      $4, $5, $4
 906 $7.1:
 907         ldub    [$1+0], $5
 908         ba      $8
 909         or      $4, $5, $4
 910
 911         .align 4
 912
 913 $7.jmp.table:
 914         .word   0
 915         .word   $7.1-$7.0
 916         .word   $7.2-$7.0
 917         .word   $7.3-$7.0
 918         .word   $7.4-$7.0
 919         .word   $7.5-$7.0
 920         .word   $7.6-$7.0
 921         .word   $7.7-$7.0
 922 })
 923
 924
 925 ! {store_little_endian}
 926 !
 927 ! parameter 1  address
 928 ! parameter 2  source left
 929 ! parameter 3  source right
 930 ! parameter 4  temporary
 931
 932 define(store_little_endian, {
 933
 934 ! {store_little_endian}
 935 ! $1 $2 $3 $4 $5 $6 $7 $8 $9
 936
 937         ! rightmost in register to first in memory
 938
 939 $5:
 940         and     $2, 255, $4
 941         stub    $4, [$1+0]
 942
 943         srl     $2, 8, $4
 944         and     $4, 255, $4
 945         stub    $4, [$1+1]
 946
 947         srl     $2, 16, $4
 948         and     $4, 255, $4
 949         stub    $4, [$1+2]
 950
 951         srl     $2, 24, $4
 952         stub    $4, [$1+3]
 953
 954
 955         and     $3, 255, $4
 956         stub    $4, [$1+0+4]
 957
 958         srl     $3, 8, $4
 959         and     $4, 255, $4
 960         stub    $4, [$1+1+4]
 961
 962         srl     $3, 16, $4
 963         and     $4, 255, $4
 964         stub    $4, [$1+2+4]
 965
 966         srl     $3, 24, $4
 967         stub    $4, [$1+3+4]
 968
 969 $5a:
 970
 971 })
 972
 973
 974 ! {store_n_bytes}
 975 !
 976 ! Stores 1 to 7 bytes little endian
 977 !
 978 ! parameter 1  address
 979 ! parameter 2  length
 980 ! parameter 3  source register left
 981 ! parameter 4  source register right
 982 ! parameter 5  temp
 983 ! parameter 6  temp2
 984 ! parameter 7  label
 985 ! parameter 8  return label
 986
 987 define(store_n_bytes, {
 988
 989 ! {store_n_bytes}
 990 ! $1 $2 $5 $6 $7 $8 $7 $8 $9
 991
 992 $7.0:   call    .+8
 993         sll     $2, 2, $6
 994
 995         add     %o7,$7.jmp.table-$7.0,$5
 996
 997         add     $5, $6, $5
 998
 999         ld      [$5], $5
1000
1001         jmp     %o7+$5
1002         nop
1003
1004 $7.7:
1005         srl     $3, 16, $5
1006         and     $5, 0xff, $5
1007         stub    $5, [$1+6]
1008 $7.6:
1009         srl     $3, 8, $5
1010         and     $5, 0xff, $5
1011         stub    $5, [$1+5]
1012 $7.5:
1013         and     $3, 0xff, $5
1014         stub    $5, [$1+4]
1015 $7.4:
1016         srl     $4, 24, $5
1017         stub    $5, [$1+3]
1018 $7.3:
1019         srl     $4, 16, $5
1020         and     $5, 0xff, $5
1021         stub    $5, [$1+2]
1022 $7.2:
1023         srl     $4, 8, $5
1024         and     $5, 0xff, $5
1025         stub    $5, [$1+1]
1026 $7.1:
1027         and     $4, 0xff, $5
1028
1029
1030         ba      $8
1031         stub    $5, [$1]
1032
1033         .align 4
1034
1035 $7.jmp.table:
1036
1037         .word   0
1038         .word   $7.1-$7.0
1039         .word   $7.2-$7.0
1040         .word   $7.3-$7.0
1041         .word   $7.4-$7.0
1042         .word   $7.5-$7.0
1043         .word   $7.6-$7.0
1044         .word   $7.7-$7.0
1045 })
1046
1047
1048 define(testvalue,{1})
1049
1050 define(register_init, {
1051
1052 ! For test purposes:
1053
1054         sethi   %hi(testvalue), local0
1055         or      local0, %lo(testvalue), local0
1056
1057         ifelse($1,{},{}, {mov   local0, $1})
1058         ifelse($2,{},{}, {mov   local0, $2})
1059         ifelse($3,{},{}, {mov   local0, $3})
1060         ifelse($4,{},{}, {mov   local0, $4})
1061         ifelse($5,{},{}, {mov   local0, $5})
1062         ifelse($6,{},{}, {mov   local0, $6})
1063         ifelse($7,{},{}, {mov   local0, $7})
1064         ifelse($8,{},{}, {mov   local0, $8})
1065
1066         mov     local0, local1
1067         mov     local0, local2
1068         mov     local0, local3
1069         mov     local0, local4
1070         mov     local0, local5
1071         mov     local0, local7
1072         mov     local0, local6
1073         mov     local0, out0
1074         mov     local0, out1
1075         mov     local0, out2
1076         mov     local0, out3
1077         mov     local0, out4
1078         mov     local0, out5
1079         mov     local0, global1
1080         mov     local0, global2
1081         mov     local0, global3
1082         mov     local0, global4
1083         mov     local0, global5
1084
1085 })
1086
1087 .section        ".text"
1088
1089         .align 32
1090
1091 .des_enc:
1092
1093         ! key address in3
1094         ! loads key next encryption/decryption first round from [in4]
1095
1096         rounds_macro(in5, out5, 1, .des_enc.1, in3, in4, retl)
1097
1098
1099         .align 32
1100
1101 .des_dec:
1102
1103         ! implemented with out5 as first parameter to avoid
1104         ! register exchange in ede modes
1105
1106         ! key address in4
1107         ! loads key next encryption/decryption first round from [in3]
1108
1109         rounds_macro(out5, in5, -1, .des_dec.1, in4, in3, retl)
1110
1111
1112
1113 ! void DES_encrypt1(data, ks, enc)
1114 ! *******************************
1115
1116         .align 32
1117         .global DES_encrypt1
1118         .type    DES_encrypt1,#function
1119
1120 DES_encrypt1:
1121
1122         save    %sp, FRAME, %sp
1123
1124         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1125         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1126 1:      call    .+8
1127         add     %o7,global1,global1
1128         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1129
1130         ld      [in0], in5                ! left
1131         cmp     in2, 0                    ! enc
1132
1133         be      .encrypt.dec
1134         ld      [in0+4], out5             ! right
1135
1136         ! parameter 6  1/2 for include encryption/decryption
1137         ! parameter 7  1 for move in1 to in3
1138         ! parameter 8  1 for move in3 to in4, 2 for move in4 to in3
1139
1140         ip_macro(in5, out5, in5, out5, in3, 0, 1, 1)
1141
1142         rounds_macro(in5, out5, 1, .des_encrypt1.1, in3, in4) ! in4 not used
1143
1144         fp_macro(in5, out5, 1)            ! 1 for store to [in0]
1145
1146         ret
1147         restore
1148
1149 .encrypt.dec:
1150
1151         add     in1, 120, in3             ! use last subkey for first round
1152
1153         ! parameter 6  1/2 for include encryption/decryption
1154         ! parameter 7  1 for move in1 to in3
1155         ! parameter 8  1 for move in3 to in4, 2 for move in4 to in3
1156
1157         ip_macro(in5, out5, out5, in5, in4, 2, 0, 1) ! include dec,  ks in4
1158
1159         fp_macro(out5, in5, 1)            ! 1 for store to [in0]
1160
1161         ret
1162         restore
1163
1164 .DES_encrypt1.end:
1165         .size    DES_encrypt1,.DES_encrypt1.end-DES_encrypt1
1166
1167
1168 ! void DES_encrypt2(data, ks, enc)
1169 !*********************************
1170
1171         ! encrypts/decrypts without initial/final permutation
1172
1173         .align 32
1174         .global DES_encrypt2
1175         .type    DES_encrypt2,#function
1176
1177 DES_encrypt2:
1178
1179         save    %sp, FRAME, %sp
1180
1181         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1182         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1183 1:      call    .+8
1184         add     %o7,global1,global1
1185         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1186
1187         ! Set sbox address 1 to 6 and rotate halfs 3 left
1188         ! Errors caught by destest? Yes. Still? *NO*
1189
1190         !sethi  %hi(DES_SPtrans), global1 ! address sbox 1
1191
1192         !or     global1, %lo(DES_SPtrans), global1  ! sbox 1
1193
1194         add     global1, 256, global2     ! sbox 2
1195         add     global1, 512, global3     ! sbox 3
1196
1197         ld      [in0], out5               ! right
1198         add     global1, 768, global4     ! sbox 4
1199         add     global1, 1024, global5    ! sbox 5
1200
1201         ld      [in0+4], in5              ! left
1202         add     global1, 1280, local6     ! sbox 6
1203         add     global1, 1792, out3       ! sbox 8
1204
1205         ! rotate
1206
1207         sll     in5, 3, local5
1208         mov     in1, in3                  ! key address to in3
1209
1210         sll     out5, 3, local7
1211         srl     in5, 29, in5
1212
1213         srl     out5, 29, out5
1214         add     in5, local5, in5
1215
1216         add     out5, local7, out5
1217         cmp     in2, 0
1218
1219         ! we use our own stackframe
1220
1221         be      .encrypt2.dec
1222         STPTR   in0, [%sp+BIAS+ARG0+0*ARGSZ]
1223
1224         ld      [in3], out0               ! key 7531 first round
1225         mov     LOOPS, out4               ! loop counter
1226
1227         ld      [in3+4], out1             ! key 8642 first round
1228         sethi   %hi(0x0000FC00), local5
1229
1230         call .des_enc
1231         mov     in3, in4
1232
1233         ! rotate
1234         sll     in5, 29, in0
1235         srl     in5, 3, in5
1236         sll     out5, 29, in1
1237         add     in5, in0, in5
1238         srl     out5, 3, out5
1239         LDPTR   [%sp+BIAS+ARG0+0*ARGSZ], in0
1240         add     out5, in1, out5
1241         st      in5, [in0]
1242         st      out5, [in0+4]
1243
1244         ret
1245         restore
1246
1247
1248 .encrypt2.dec:
1249
1250         add in3, 120, in4
1251
1252         ld      [in4], out0               ! key 7531 first round
1253         mov     LOOPS, out4               ! loop counter
1254
1255         ld      [in4+4], out1             ! key 8642 first round
1256         sethi   %hi(0x0000FC00), local5
1257
1258         mov     in5, local1               ! left expected in out5
1259         mov     out5, in5
1260
1261         call .des_dec
1262         mov     local1, out5
1263
1264 .encrypt2.finish:
1265
1266         ! rotate
1267         sll     in5, 29, in0
1268         srl     in5, 3, in5
1269         sll     out5, 29, in1
1270         add     in5, in0, in5
1271         srl     out5, 3, out5
1272         LDPTR   [%sp+BIAS+ARG0+0*ARGSZ], in0
1273         add     out5, in1, out5
1274         st      out5, [in0]
1275         st      in5, [in0+4]
1276
1277         ret
1278         restore
1279
1280 .DES_encrypt2.end:
1281         .size    DES_encrypt2, .DES_encrypt2.end-DES_encrypt2
1282
1283
1284 ! void DES_encrypt3(data, ks1, ks2, ks3)
1285 ! **************************************
1286
1287         .align 32
1288         .global DES_encrypt3
1289         .type    DES_encrypt3,#function
1290
1291 DES_encrypt3:
1292
1293         save    %sp, FRAME, %sp
1294
1295         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1296         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1297 1:      call    .+8
1298         add     %o7,global1,global1
1299         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1300
1301         ld      [in0], in5                ! left
1302         add     in2, 120, in4             ! ks2
1303
1304         ld      [in0+4], out5             ! right
1305         mov     in3, in2                  ! save ks3
1306
1307         ! parameter 6  1/2 for include encryption/decryption
1308         ! parameter 7  1 for mov in1 to in3
1309         ! parameter 8  1 for mov in3 to in4
1310         ! parameter 9  1 for load ks3 and ks2 to in4 and in3
1311
1312         ip_macro(in5, out5, in5, out5, in3, 1, 1, 0, 0)
1313
1314         call    .des_dec
1315         mov     in2, in3                  ! preload ks3
1316
1317         call    .des_enc
1318         nop
1319
1320         fp_macro(in5, out5, 1)
1321
1322         ret
1323         restore
1324
1325 .DES_encrypt3.end:
1326         .size    DES_encrypt3,.DES_encrypt3.end-DES_encrypt3
1327
1328
1329 ! void DES_decrypt3(data, ks1, ks2, ks3)
1330 ! **************************************
1331
1332         .align 32
1333         .global DES_decrypt3
1334         .type    DES_decrypt3,#function
1335
1336 DES_decrypt3:
1337
1338         save    %sp, FRAME, %sp
1339
1340         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1341         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1342 1:      call    .+8
1343         add     %o7,global1,global1
1344         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1345
1346         ld      [in0], in5                ! left
1347         add     in3, 120, in4             ! ks3
1348
1349         ld      [in0+4], out5             ! right
1350         mov     in2, in3                  ! ks2
1351
1352         ! parameter 6  1/2 for include encryption/decryption
1353         ! parameter 7  1 for mov in1 to in3
1354         ! parameter 8  1 for mov in3 to in4
1355         ! parameter 9  1 for load ks3 and ks2 to in4 and in3
1356
1357         ip_macro(in5, out5, out5, in5, in4, 2, 0, 0, 0)
1358
1359         call    .des_enc
1360         add     in1, 120, in4             ! preload ks1
1361
1362         call    .des_dec
1363         nop
1364
1365         fp_macro(out5, in5, 1)
1366
1367         ret
1368         restore
1369
1370 .DES_decrypt3.end:
1371         .size    DES_decrypt3,.DES_decrypt3.end-DES_decrypt3
1372
1373 ! void DES_ncbc_encrypt(input, output, length, schedule, ivec, enc)
1374 ! *****************************************************************
1375
1376
1377         .align 32
1378         .global DES_ncbc_encrypt
1379         .type    DES_ncbc_encrypt,#function
1380
1381 DES_ncbc_encrypt:
1382
1383         save    %sp, FRAME, %sp
1384
1385         define({INPUT},  { [%sp+BIAS+ARG0+0*ARGSZ] })
1386         define({OUTPUT}, { [%sp+BIAS+ARG0+1*ARGSZ] })
1387         define({IVEC},   { [%sp+BIAS+ARG0+4*ARGSZ] })
1388
1389         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1390         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1391 1:      call    .+8
1392         add     %o7,global1,global1
1393         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1394
1395         cmp     in5, 0                    ! enc
1396
1397         be      .ncbc.dec
1398         STPTR   in4, IVEC
1399
1400         ! addr  left  right  temp  label
1401         load_little_endian(in4, in5, out5, local3, .LLE1)  ! iv
1402
1403         addcc   in2, -8, in2              ! bytes missing when first block done
1404
1405         bl      .ncbc.enc.seven.or.less
1406         mov     in3, in4                  ! schedule
1407
1408 .ncbc.enc.next.block:
1409
1410         load_little_endian(in0, out4, global4, local3, .LLE2)  ! block
1411
1412 .ncbc.enc.next.block_1:
1413
1414         xor     in5, out4, in5            ! iv xor
1415         xor     out5, global4, out5       ! iv xor
1416
1417         ! parameter 8  1 for move in3 to in4, 2 for move in4 to in3
1418         ip_macro(in5, out5, in5, out5, in3, 0, 0, 2)
1419
1420 .ncbc.enc.next.block_2:
1421
1422 !//     call .des_enc                     ! compares in2 to 8
1423 !       rounds inlined for alignment purposes
1424
1425         add     global1, 768, global4     ! address sbox 4 since register used below
1426
1427         rounds_macro(in5, out5, 1, .ncbc.enc.1, in3, in4) ! include encryption  ks in3
1428
1429         bl      .ncbc.enc.next.block_fp
1430         add     in0, 8, in0               ! input address
1431
1432         ! If 8 or more bytes are to be encrypted after this block,
1433         ! we combine final permutation for this block with initial
1434         ! permutation for next block. Load next block:
1435
1436         load_little_endian(in0, global3, global4, local5, .LLE12)
1437
1438         !  parameter 1   original left
1439         !  parameter 2   original right
1440         !  parameter 3   left ip
1441         !  parameter 4   right ip
1442         !  parameter 5   1: load ks1/ks2 to in3/in4, add 120 to in4
1443         !                2: mov in4 to in3
1444         !
1445         ! also adds -8 to length in2 and loads loop counter to out4
1446
1447         fp_ip_macro(out0, out1, global3, global4, 2)
1448
1449         store_little_endian(in1, out0, out1, local3, .SLE10)  ! block
1450
1451         ld      [in3], out0               ! key 7531 first round next block
1452         mov     in5, local1
1453         xor     global3, out5, in5        ! iv xor next block
1454
1455         ld      [in3+4], out1             ! key 8642
1456         add     global1, 512, global3     ! address sbox 3 since register used
1457         xor     global4, local1, out5     ! iv xor next block
1458
1459         ba      .ncbc.enc.next.block_2
1460         add     in1, 8, in1               ! output address
1461
1462 .ncbc.enc.next.block_fp:
1463
1464         fp_macro(in5, out5)
1465
1466         store_little_endian(in1, in5, out5, local3, .SLE1)  ! block
1467
1468         addcc   in2, -8, in2              ! bytes missing when next block done
1469
1470         bpos    .ncbc.enc.next.block
1471         add     in1, 8, in1
1472
1473 .ncbc.enc.seven.or.less:
1474
1475         cmp     in2, -8
1476
1477         ble     .ncbc.enc.finish
1478         nop
1479
1480         add     in2, 8, local1            ! bytes to load
1481
1482         ! addr, length, dest left, dest right, temp, temp2, label, ret label
1483         load_n_bytes(in0, local1, global4, out4, local2, local3, .LNB1, .ncbc.enc.next.block_1)
1484
1485         ! Loads 1 to 7 bytes little endian to global4, out4
1486
1487
1488 .ncbc.enc.finish:
1489
1490         LDPTR   IVEC, local4
1491         store_little_endian(local4, in5, out5, local5, .SLE2)  ! ivec
1492
1493         ret
1494         restore
1495
1496
1497 .ncbc.dec:
1498
1499         STPTR   in0, INPUT
1500         cmp     in2, 0                    ! length
1501         add     in3, 120, in3
1502
1503         LDPTR   IVEC, local7              ! ivec
1504         ble     .ncbc.dec.finish
1505         mov     in3, in4                  ! schedule
1506
1507         STPTR   in1, OUTPUT
1508         mov     in0, local5               ! input
1509
1510         load_little_endian(local7, in0, in1, local3, .LLE3)   ! ivec
1511
1512 .ncbc.dec.next.block:
1513
1514         load_little_endian(local5, in5, out5, local3, .LLE4)  ! block
1515
1516         ! parameter 6  1/2 for include encryption/decryption
1517         ! parameter 7  1 for mov in1 to in3
1518         ! parameter 8  1 for mov in3 to in4
1519
1520         ip_macro(in5, out5, out5, in5, in4, 2, 0, 1) ! include decryption  ks in4
1521
1522         fp_macro(out5, in5, 0, 1) ! 1 for input and output address to local5/7
1523
1524         ! in2 is bytes left to be stored
1525         ! in2 is compared to 8 in the rounds
1526
1527         xor     out5, in0, out4           ! iv xor
1528         bl      .ncbc.dec.seven.or.less
1529         xor     in5, in1, global4         ! iv xor
1530
1531         ! Load ivec next block now, since input and output address might be the same.
1532
1533         load_little_endian_inc(local5, in0, in1, local3, .LLE5)  ! iv
1534
1535         store_little_endian(local7, out4, global4, local3, .SLE3)
1536
1537         STPTR   local5, INPUT
1538         add     local7, 8, local7
1539         addcc   in2, -8, in2
1540
1541         bg      .ncbc.dec.next.block
1542         STPTR   local7, OUTPUT
1543
1544
1545 .ncbc.dec.store.iv:
1546
1547         LDPTR   IVEC, local4              ! ivec
1548         store_little_endian(local4, in0, in1, local5, .SLE4)
1549
1550 .ncbc.dec.finish:
1551
1552         ret
1553         restore
1554
1555 .ncbc.dec.seven.or.less:
1556
1557         load_little_endian_inc(local5, in0, in1, local3, .LLE13)     ! ivec
1558
1559         store_n_bytes(local7, in2, global4, out4, local3, local4, .SNB1, .ncbc.dec.store.iv)
1560
1561
1562 .DES_ncbc_encrypt.end:
1563         .size    DES_ncbc_encrypt, .DES_ncbc_encrypt.end-DES_ncbc_encrypt
1564
1565
1566 ! void DES_ede3_cbc_encrypt(input, output, length, ks1, ks2, ks3, ivec, enc)
1567 ! **************************************************************************
1568
1569
1570         .align 32
1571         .global DES_ede3_cbc_encrypt
1572         .type    DES_ede3_cbc_encrypt,#function
1573
1574 DES_ede3_cbc_encrypt:
1575
1576         save    %sp, FRAME, %sp
1577
1578         define({KS1}, { [%sp+BIAS+ARG0+3*ARGSZ] })
1579         define({KS2}, { [%sp+BIAS+ARG0+4*ARGSZ] })
1580         define({KS3}, { [%sp+BIAS+ARG0+5*ARGSZ] })
1581
1582         sethi   %hi(.PIC.DES_SPtrans-1f),global1
1583         or      global1,%lo(.PIC.DES_SPtrans-1f),global1
1584 1:      call    .+8
1585         add     %o7,global1,global1
1586         sub     global1,.PIC.DES_SPtrans-.des_and,out2
1587
1588         LDPTR   [%fp+BIAS+ARG0+7*ARGSZ], local3          ! enc
1589         LDPTR   [%fp+BIAS+ARG0+6*ARGSZ], local4          ! ivec
1590         cmp     local3, 0                 ! enc
1591
1592         be      .ede3.dec
1593         STPTR   in4, KS2
1594
1595         STPTR   in5, KS3
1596
1597         load_little_endian(local4, in5, out5, local3, .LLE6)  ! ivec
1598
1599         addcc   in2, -8, in2              ! bytes missing after next block
1600
1601         bl      .ede3.enc.seven.or.less
1602         STPTR   in3, KS1
1603
1604 .ede3.enc.next.block:
1605
1606         load_little_endian(in0, out4, global4, local3, .LLE7)
1607
1608 .ede3.enc.next.block_1:
1609
1610         LDPTR   KS2, in4
1611         xor     in5, out4, in5            ! iv xor
1612         xor     out5, global4, out5       ! iv xor
1613
1614         LDPTR   KS1, in3
1615         add     in4, 120, in4             ! for decryption we use last subkey first
1616         nop
1617
1618         ip_macro(in5, out5, in5, out5, in3)
1619
1620 .ede3.enc.next.block_2:
1621
1622         call .des_enc                     ! ks1 in3
1623         nop
1624
1625         call .des_dec                     ! ks2 in4
1626         LDPTR   KS3, in3
1627
1628         call .des_enc                     ! ks3 in3  compares in2 to 8
1629         nop
1630
1631         bl      .ede3.enc.next.block_fp
1632         add     in0, 8, in0
1633
1634         ! If 8 or more bytes are to be encrypted after this block,
1635         ! we combine final permutation for this block with initial
1636         ! permutation for next block. Load next block:
1637
1638         load_little_endian(in0, global3, global4, local5, .LLE11)
1639
1640         !  parameter 1   original left
1641         !  parameter 2   original right
1642         !  parameter 3   left ip
1643         !  parameter 4   right ip
1644         !  parameter 5   1: load ks1/ks2 to in3/in4, add 120 to in4
1645         !                2: mov in4 to in3
1646         !
1647         ! also adds -8 to length in2 and loads loop counter to out4
1648
1649         fp_ip_macro(out0, out1, global3, global4, 1)
1650
1651         store_little_endian(in1, out0, out1, local3, .SLE9)  ! block
1652
1653         mov     in5, local1
1654         xor     global3, out5, in5        ! iv xor next block
1655
1656         ld      [in3], out0               ! key 7531
1657         add     global1, 512, global3     ! address sbox 3
1658         xor     global4, local1, out5     ! iv xor next block
1659
1660         ld      [in3+4], out1             ! key 8642
1661         add     global1, 768, global4     ! address sbox 4
1662         ba      .ede3.enc.next.block_2
1663         add     in1, 8, in1
1664
1665 .ede3.enc.next.block_fp:
1666
1667         fp_macro(in5, out5)
1668
1669         store_little_endian(in1, in5, out5, local3, .SLE5)  ! block
1670
1671         addcc   in2, -8, in2              ! bytes missing when next block done
1672
1673         bpos    .ede3.enc.next.block
1674         add     in1, 8, in1
1675
1676 .ede3.enc.seven.or.less:
1677
1678         cmp     in2, -8
1679
1680         ble     .ede3.enc.finish
1681         nop
1682
1683         add     in2, 8, local1            ! bytes to load
1684
1685         ! addr, length, dest left, dest right, temp, temp2, label, ret label
1686         load_n_bytes(in0, local1, global4, out4, local2, local3, .LNB2, .ede3.enc.next.block_1)
1687
1688 .ede3.enc.finish:
1689
1690         LDPTR   [%fp+BIAS+ARG0+6*ARGSZ], local4          ! ivec
1691         store_little_endian(local4, in5, out5, local5, .SLE6)  ! ivec
1692
1693         ret
1694         restore
1695
1696 .ede3.dec:
1697
1698         STPTR   in0, INPUT
1699         add     in5, 120, in5
1700
1701         STPTR   in1, OUTPUT
1702         mov     in0, local5
1703         add     in3, 120, in3
1704
1705         STPTR   in3, KS1
1706         cmp     in2, 0
1707
1708         ble     .ede3.dec.finish
1709         STPTR   in5, KS3
1710
1711         LDPTR   [%fp+BIAS+ARG0+6*ARGSZ], local7          ! iv
1712         load_little_endian(local7, in0, in1, local3, .LLE8)
1713
1714 .ede3.dec.next.block:
1715
1716         load_little_endian(local5, in5, out5, local3, .LLE9)
1717
1718         ! parameter 6  1/2 for include encryption/decryption
1719         ! parameter 7  1 for mov in1 to in3
1720         ! parameter 8  1 for mov in3 to in4
1721         ! parameter 9  1 for load ks3 and ks2 to in4 and in3
1722
1723         ip_macro(in5, out5, out5, in5, in4, 2, 0, 0, 1) ! inc .des_dec ks3 in4
1724
1725         call .des_enc                     ! ks2 in3
1726         LDPTR   KS1, in4
1727
1728         call .des_dec                     ! ks1 in4
1729         nop
1730
1731         fp_macro(out5, in5, 0, 1)   ! 1 for input and output address local5/7
1732
1733         ! in2 is bytes left to be stored
1734         ! in2 is compared to 8 in the rounds
1735
1736         xor     out5, in0, out4
1737         bl      .ede3.dec.seven.or.less
1738         xor     in5, in1, global4
1739
1740         load_little_endian_inc(local5, in0, in1, local3, .LLE10)   ! iv next block
1741
1742         store_little_endian(local7, out4, global4, local3, .SLE7)  ! block
1743
1744         STPTR   local5, INPUT
1745         addcc   in2, -8, in2
1746         add     local7, 8, local7
1747
1748         bg      .ede3.dec.next.block
1749         STPTR   local7, OUTPUT
1750
1751 .ede3.dec.store.iv:
1752
1753         LDPTR   [%fp+BIAS+ARG0+6*ARGSZ], local4          ! ivec
1754         store_little_endian(local4, in0, in1, local5, .SLE8)  ! ivec
1755
1756 .ede3.dec.finish:
1757
1758         ret
1759         restore
1760
1761 .ede3.dec.seven.or.less:
1762
1763         load_little_endian_inc(local5, in0, in1, local3, .LLE14)     ! iv
1764
1765         store_n_bytes(local7, in2, global4, out4, local3, local4, .SNB2, .ede3.dec.store.iv)
1766
1767
1768 .DES_ede3_cbc_encrypt.end:
1769         .size    DES_ede3_cbc_encrypt,.DES_ede3_cbc_encrypt.end-DES_ede3_cbc_encrypt
1770
1771         .align  256
1772         .type    .des_and,#object
1773         .size    .des_and,284
1774
1775 .des_and:
1776
1777 ! This table is used for AND 0xFC when it is known that register
1778 ! bits 8-31 are zero. Makes it possible to do three arithmetic
1779 ! operations in one cycle.
1780
1781         .byte  0, 0, 0, 0, 4, 4, 4, 4
1782         .byte  8, 8, 8, 8, 12, 12, 12, 12
1783         .byte  16, 16, 16, 16, 20, 20, 20, 20
1784         .byte  24, 24, 24, 24, 28, 28, 28, 28
1785         .byte  32, 32, 32, 32, 36, 36, 36, 36
1786         .byte  40, 40, 40, 40, 44, 44, 44, 44
1787         .byte  48, 48, 48, 48, 52, 52, 52, 52
1788         .byte  56, 56, 56, 56, 60, 60, 60, 60
1789         .byte  64, 64, 64, 64, 68, 68, 68, 68
1790         .byte  72, 72, 72, 72, 76, 76, 76, 76
1791         .byte  80, 80, 80, 80, 84, 84, 84, 84
1792         .byte  88, 88, 88, 88, 92, 92, 92, 92
1793         .byte  96, 96, 96, 96, 100, 100, 100, 100
1794         .byte  104, 104, 104, 104, 108, 108, 108, 108
1795         .byte  112, 112, 112, 112, 116, 116, 116, 116
1796         .byte  120, 120, 120, 120, 124, 124, 124, 124
1797         .byte  128, 128, 128, 128, 132, 132, 132, 132
1798         .byte  136, 136, 136, 136, 140, 140, 140, 140
1799         .byte  144, 144, 144, 144, 148, 148, 148, 148
1800         .byte  152, 152, 152, 152, 156, 156, 156, 156
1801         .byte  160, 160, 160, 160, 164, 164, 164, 164
1802         .byte  168, 168, 168, 168, 172, 172, 172, 172
1803         .byte  176, 176, 176, 176, 180, 180, 180, 180
1804         .byte  184, 184, 184, 184, 188, 188, 188, 188
1805         .byte  192, 192, 192, 192, 196, 196, 196, 196
1806         .byte  200, 200, 200, 200, 204, 204, 204, 204
1807         .byte  208, 208, 208, 208, 212, 212, 212, 212
1808         .byte  216, 216, 216, 216, 220, 220, 220, 220
1809         .byte  224, 224, 224, 224, 228, 228, 228, 228
1810         .byte  232, 232, 232, 232, 236, 236, 236, 236
1811         .byte  240, 240, 240, 240, 244, 244, 244, 244
1812         .byte  248, 248, 248, 248, 252, 252, 252, 252
1813
1814         ! 5 numbers for initial/final permutation
1815
1816         .word   0x0f0f0f0f                ! offset 256
1817         .word   0x0000ffff                ! 260
1818         .word   0x33333333                ! 264
1819         .word   0x00ff00ff                ! 268
1820         .word   0x55555555                ! 272
1821
1822         .word   0                         ! 276
1823         .word   LOOPS                     ! 280
1824         .word   0x0000FC00                ! 284
1825
1826         .global DES_SPtrans
1827         .type   DES_SPtrans,#object
1828         .size   DES_SPtrans,2048
1829 .align  64
1830 DES_SPtrans:
1831 .PIC.DES_SPtrans:
1832         ! nibble 0
1833         .word   0x02080800, 0x00080000, 0x02000002, 0x02080802
1834         .word   0x02000000, 0x00080802, 0x00080002, 0x02000002
1835         .word   0x00080802, 0x02080800, 0x02080000, 0x00000802
1836         .word   0x02000802, 0x02000000, 0x00000000, 0x00080002
1837         .word   0x00080000, 0x00000002, 0x02000800, 0x00080800
1838         .word   0x02080802, 0x02080000, 0x00000802, 0x02000800
1839         .word   0x00000002, 0x00000800, 0x00080800, 0x02080002
1840         .word   0x00000800, 0x02000802, 0x02080002, 0x00000000
1841         .word   0x00000000, 0x02080802, 0x02000800, 0x00080002
1842         .word   0x02080800, 0x00080000, 0x00000802, 0x02000800
1843         .word   0x02080002, 0x00000800, 0x00080800, 0x02000002
1844         .word   0x00080802, 0x00000002, 0x02000002, 0x02080000
1845         .word   0x02080802, 0x00080800, 0x02080000, 0x02000802
1846         .word   0x02000000, 0x00000802, 0x00080002, 0x00000000
1847         .word   0x00080000, 0x02000000, 0x02000802, 0x02080800
1848         .word   0x00000002, 0x02080002, 0x00000800, 0x00080802
1849         ! nibble 1
1850         .word   0x40108010, 0x00000000, 0x00108000, 0x40100000
1851         .word   0x40000010, 0x00008010, 0x40008000, 0x00108000
1852         .word   0x00008000, 0x40100010, 0x00000010, 0x40008000
1853         .word   0x00100010, 0x40108000, 0x40100000, 0x00000010
1854         .word   0x00100000, 0x40008010, 0x40100010, 0x00008000
1855         .word   0x00108010, 0x40000000, 0x00000000, 0x00100010
1856         .word   0x40008010, 0x00108010, 0x40108000, 0x40000010
1857         .word   0x40000000, 0x00100000, 0x00008010, 0x40108010
1858         .word   0x00100010, 0x40108000, 0x40008000, 0x00108010
1859         .word   0x40108010, 0x00100010, 0x40000010, 0x00000000
1860         .word   0x40000000, 0x00008010, 0x00100000, 0x40100010
1861         .word   0x00008000, 0x40000000, 0x00108010, 0x40008010
1862         .word   0x40108000, 0x00008000, 0x00000000, 0x40000010
1863         .word   0x00000010, 0x40108010, 0x00108000, 0x40100000
1864         .word   0x40100010, 0x00100000, 0x00008010, 0x40008000
1865         .word   0x40008010, 0x00000010, 0x40100000, 0x00108000
1866         ! nibble 2
1867         .word   0x04000001, 0x04040100, 0x00000100, 0x04000101
1868         .word   0x00040001, 0x04000000, 0x04000101, 0x00040100
1869         .word   0x04000100, 0x00040000, 0x04040000, 0x00000001
1870         .word   0x04040101, 0x00000101, 0x00000001, 0x04040001
1871         .word   0x00000000, 0x00040001, 0x04040100, 0x00000100
1872         .word   0x00000101, 0x04040101, 0x00040000, 0x04000001
1873         .word   0x04040001, 0x04000100, 0x00040101, 0x04040000
1874         .word   0x00040100, 0x00000000, 0x04000000, 0x00040101
1875         .word   0x04040100, 0x00000100, 0x00000001, 0x00040000
1876         .word   0x00000101, 0x00040001, 0x04040000, 0x04000101
1877         .word   0x00000000, 0x04040100, 0x00040100, 0x04040001
1878         .word   0x00040001, 0x04000000, 0x04040101, 0x00000001
1879         .word   0x00040101, 0x04000001, 0x04000000, 0x04040101
1880         .word   0x00040000, 0x04000100, 0x04000101, 0x00040100
1881         .word   0x04000100, 0x00000000, 0x04040001, 0x00000101
1882         .word   0x04000001, 0x00040101, 0x00000100, 0x04040000
1883         ! nibble 3
1884         .word   0x00401008, 0x10001000, 0x00000008, 0x10401008
1885         .word   0x00000000, 0x10400000, 0x10001008, 0x00400008
1886         .word   0x10401000, 0x10000008, 0x10000000, 0x00001008
1887         .word   0x10000008, 0x00401008, 0x00400000, 0x10000000
1888         .word   0x10400008, 0x00401000, 0x00001000, 0x00000008
1889         .word   0x00401000, 0x10001008, 0x10400000, 0x00001000
1890         .word   0x00001008, 0x00000000, 0x00400008, 0x10401000
1891         .word   0x10001000, 0x10400008, 0x10401008, 0x00400000
1892         .word   0x10400008, 0x00001008, 0x00400000, 0x10000008
1893         .word   0x00401000, 0x10001000, 0x00000008, 0x10400000
1894         .word   0x10001008, 0x00000000, 0x00001000, 0x00400008
1895         .word   0x00000000, 0x10400008, 0x10401000, 0x00001000
1896         .word   0x10000000, 0x10401008, 0x00401008, 0x00400000
1897         .word   0x10401008, 0x00000008, 0x10001000, 0x00401008
1898         .word   0x00400008, 0x00401000, 0x10400000, 0x10001008
1899         .word   0x00001008, 0x10000000, 0x10000008, 0x10401000
1900         ! nibble 4
1901         .word   0x08000000, 0x00010000, 0x00000400, 0x08010420
1902         .word   0x08010020, 0x08000400, 0x00010420, 0x08010000
1903         .word   0x00010000, 0x00000020, 0x08000020, 0x00010400
1904         .word   0x08000420, 0x08010020, 0x08010400, 0x00000000
1905         .word   0x00010400, 0x08000000, 0x00010020, 0x00000420
1906         .word   0x08000400, 0x00010420, 0x00000000, 0x08000020
1907         .word   0x00000020, 0x08000420, 0x08010420, 0x00010020
1908         .word   0x08010000, 0x00000400, 0x00000420, 0x08010400
1909         .word   0x08010400, 0x08000420, 0x00010020, 0x08010000
1910         .word   0x00010000, 0x00000020, 0x08000020, 0x08000400
1911         .word   0x08000000, 0x00010400, 0x08010420, 0x00000000
1912         .word   0x00010420, 0x08000000, 0x00000400, 0x00010020
1913         .word   0x08000420, 0x00000400, 0x00000000, 0x08010420
1914         .word   0x08010020, 0x08010400, 0x00000420, 0x00010000
1915         .word   0x00010400, 0x08010020, 0x08000400, 0x00000420
1916         .word   0x00000020, 0x00010420, 0x08010000, 0x08000020
1917         ! nibble 5
1918         .word   0x80000040, 0x00200040, 0x00000000, 0x80202000
1919         .word   0x00200040, 0x00002000, 0x80002040, 0x00200000
1920         .word   0x00002040, 0x80202040, 0x00202000, 0x80000000
1921         .word   0x80002000, 0x80000040, 0x80200000, 0x00202040
1922         .word   0x00200000, 0x80002040, 0x80200040, 0x00000000
1923         .word   0x00002000, 0x00000040, 0x80202000, 0x80200040
1924         .word   0x80202040, 0x80200000, 0x80000000, 0x00002040
1925         .word   0x00000040, 0x00202000, 0x00202040, 0x80002000
1926         .word   0x00002040, 0x80000000, 0x80002000, 0x00202040
1927         .word   0x80202000, 0x00200040, 0x00000000, 0x80002000
1928         .word   0x80000000, 0x00002000, 0x80200040, 0x00200000
1929         .word   0x00200040, 0x80202040, 0x00202000, 0x00000040
1930         .word   0x80202040, 0x00202000, 0x00200000, 0x80002040
1931         .word   0x80000040, 0x80200000, 0x00202040, 0x00000000
1932         .word   0x00002000, 0x80000040, 0x80002040, 0x80202000
1933         .word   0x80200000, 0x00002040, 0x00000040, 0x80200040
1934         ! nibble 6
1935         .word   0x00004000, 0x00000200, 0x01000200, 0x01000004
1936         .word   0x01004204, 0x00004004, 0x00004200, 0x00000000
1937         .word   0x01000000, 0x01000204, 0x00000204, 0x01004000
1938         .word   0x00000004, 0x01004200, 0x01004000, 0x00000204
1939         .word   0x01000204, 0x00004000, 0x00004004, 0x01004204
1940         .word   0x00000000, 0x01000200, 0x01000004, 0x00004200
1941         .word   0x01004004, 0x00004204, 0x01004200, 0x00000004
1942         .word   0x00004204, 0x01004004, 0x00000200, 0x01000000
1943         .word   0x00004204, 0x01004000, 0x01004004, 0x00000204
1944         .word   0x00004000, 0x00000200, 0x01000000, 0x01004004
1945         .word   0x01000204, 0x00004204, 0x00004200, 0x00000000
1946         .word   0x00000200, 0x01000004, 0x00000004, 0x01000200
1947         .word   0x00000000, 0x01000204, 0x01000200, 0x00004200
1948         .word   0x00000204, 0x00004000, 0x01004204, 0x01000000
1949         .word   0x01004200, 0x00000004, 0x00004004, 0x01004204
1950         .word   0x01000004, 0x01004200, 0x01004000, 0x00004004
1951         ! nibble 7
1952         .word   0x20800080, 0x20820000, 0x00020080, 0x00000000
1953         .word   0x20020000, 0x00800080, 0x20800000, 0x20820080
1954         .word   0x00000080, 0x20000000, 0x00820000, 0x00020080
1955         .word   0x00820080, 0x20020080, 0x20000080, 0x20800000
1956         .word   0x00020000, 0x00820080, 0x00800080, 0x20020000
1957         .word   0x20820080, 0x20000080, 0x00000000, 0x00820000
1958         .word   0x20000000, 0x00800000, 0x20020080, 0x20800080
1959         .word   0x00800000, 0x00020000, 0x20820000, 0x00000080
1960         .word   0x00800000, 0x00020000, 0x20000080, 0x20820080
1961         .word   0x00020080, 0x20000000, 0x00000000, 0x00820000
1962         .word   0x20800080, 0x20020080, 0x20020000, 0x00800080
1963         .word   0x20820000, 0x00000080, 0x00800080, 0x20020000
1964         .word   0x20820080, 0x00800000, 0x20800000, 0x20000080
1965         .word   0x00820000, 0x00020080, 0x20020080, 0x20800000
1966         .word   0x00000080, 0x20820000, 0x00820080, 0x00000000
1967         .word   0x20000000, 0x20800080, 0x00020000, 0x00820080
1968