release/src/router/nettle/x86_64/serpent-decrypt.asm

   1 C nettle, low-level cryptographics library
   2 C
   3 C Copyright (C) 2011 Niels Möller
   4 C
   5 C The nettle library is free software; you can redistribute it and/or modify
   6 C it under the terms of the GNU Lesser General Public License as published by
   7 C the Free Software Foundation; either version 2.1 of the License, or (at your
   8 C option) any later version.
   9 C
  10 C The nettle library is distributed in the hope that it will be useful, but
  11 C WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  12 C or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
  13 C License for more details.
  14 C
  15 C You should have received a copy of the GNU Lesser General Public License
  16 C along with the nettle library; see the file COPYING.LIB.  If not, write to
  17 C the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston,
  18 C MA 02111-1301, USA.
  19
  20 include_src(<x86_64/serpent.m4>)
  21
  22 C Register usage:
  23
  24 C Single block serpent state, two copies
  25 define(<x0>, <%eax>)
  26 define(<x1>, <%ebx>)
  27 define(<x2>, <%ebp>)
  28 define(<x3>, <%r8d>)
  29
  30 define(<y0>, <%r9d>)
  31 define(<y1>, <%r10d>)
  32 define(<y2>, <%r11d>)
  33 define(<y3>, <%r12d>)
  34
  35 C Quadruple block serpent state, two copies
  36 define(<X0>, <%xmm0>)
  37 define(<X1>, <%xmm1>)
  38 define(<X2>, <%xmm2>)
  39 define(<X3>, <%xmm3>)
  40
  41 define(<Y0>, <%xmm4>)
  42 define(<Y1>, <%xmm5>)
  43 define(<Y2>, <%xmm6>)
  44 define(<Y3>, <%xmm7>)
  45
  46 define(<MINUS1>, <%xmm8>)
  47 define(<T0>, <%xmm9>)
  48 define(<T1>, <%xmm10>)
  49 define(<T2>, <%xmm11>)
  50 define(<T3>, <%xmm12>)
  51
  52 C Arguments
  53 define(<CTX>, <%rdi>)
  54 define(<N>, <%rsi>)
  55 define(<DST>, <%rdx>)
  56 define(<SRC>, <%rcx>)
  57
  58 define(<CNT>, <%r13>)
  59 define(<TMP32>, <%r14d>)
  60
  61 C SBOX macros. Inputs $1 - $4 (destroyed), outputs $5 - $8
  62
  63 define(<SBOX0I>, <
  64         mov     $1, $5
  65         xor     $3, $5
  66         mov     $1, $7
  67         or      $2, $7
  68         mov     $3, $6
  69         xor     $4, $6
  70         xor     $6, $7
  71         and     $3, $6
  72         or      $2, $3
  73         xor     $4, $2
  74         or      $1, $6
  75         and     $3, $2
  76         xor     $2, $6
  77         or      $7, $1
  78         xor     $6, $1
  79         mov     $7, $2
  80         and     $1, $2
  81         not     $7
  82         or      $7, $4
  83         xor     $3, $4
  84         mov     $1, $8
  85         xor     $4, $8
  86         or      $4, $2
  87         xor     $2, $5
  88 >)
  89
  90 define(<SBOX1I>, <
  91         mov     $2, $6
  92         or      $4, $6
  93         xor     $3, $6
  94         mov     $1, $8
  95         xor     $2, $8
  96         mov     $1, $5
  97         or      $6, $5
  98         and     $8, $5
  99         xor     $5, $2
 100         xor     $6, $8
 101         and     $4, $2
 102         mov     $1, $7
 103         and     $3, $7
 104         or      $7, $6
 105         or      $4, $7
 106         xor     $5, $7
 107         not     $7
 108         xor     $2, $6
 109         xor     $6, $5
 110         xor     $3, $5
 111         or      $7, $1
 112         xor     $1, $5
 113 >)
 114
 115 define(<SBOX2I>, <
 116         mov     $1, $5
 117         xor     $4, $5
 118         mov     $3, $7
 119         xor     $4, $7
 120         mov     $2, $6
 121         or      $7, $6
 122         xor     $6, $5
 123         mov     $4, $6
 124         or      $5, $6
 125         and     $2, $6
 126         not     $4
 127         mov     $1, $8
 128         or      $3, $8
 129         and     $8, $7
 130         xor     $7, $6
 131         and     $2, $8
 132         and     $3, $1
 133         or      $4, $1
 134         xor     $1, $8
 135         and     $8, $3
 136         xor     $1, $3
 137         mov     $5, $7
 138         xor     $6, $7
 139         xor     $3, $7
 140 >)
 141
 142 define(<SBOX3I>, <
 143         mov     $3, $8
 144         or      $4, $8
 145         mov     $2, $5
 146         and     $8, $5
 147         mov     $1, $7
 148         or      $4, $7
 149         mov     $3, $6
 150         xor     $7, $6
 151         xor     $6, $5
 152         xor     $1, $4
 153         xor     $4, $8
 154         xor     $2, $7
 155         and     $6, $7
 156         xor     $4, $7
 157         xor     $1, $6
 158         or      $5, $4
 159         and     $4, $6
 160         xor     $2, $6
 161         and     $7, $1
 162         or      $2, $1
 163         xor     $1, $8
 164 >)
 165
 166 define(<SBOX4I>, <
 167         mov     $3, $6
 168         xor     $4, $6
 169         mov     $3, $7
 170         or      $4, $7
 171         xor     $2, $7
 172         or      $4, $2
 173         mov     $1, $5
 174         xor     $7, $5
 175         xor     $7, $4
 176         and     $1, $7
 177         xor     $7, $6
 178         xor     $1, $7
 179         or      $3, $7
 180         and     $2, $1
 181         mov     $1, $8
 182         xor     $4, $8
 183         not     $1
 184         or      $6, $1
 185         xor     $1, $5
 186         xor     $2, $1
 187         xor     $1, $7
 188 >)
 189
 190 define(<SBOX5I>, <
 191         mov     $1, $6
 192         and     $4, $6
 193         mov     $3, $8
 194         xor     $6, $8
 195         mov     $2, $5
 196         and     $8, $5
 197         mov     $1, $7
 198         xor     $4, $7
 199         xor     $2, $4
 200         xor     $7, $5
 201         and     $1, $3
 202         and     $5, $1
 203         or      $2, $3
 204         xor     $5, $6
 205         xor     $3, $6
 206         mov     $5, $7
 207         or      $6, $7
 208         xor     $8, $7
 209         xor     $4, $7
 210         not     $2
 211         or      $1, $2
 212         xor     $2, $8
 213 >)
 214
 215 define(<SBOX6I>, <
 216         mov     $1, $7
 217         xor     $3, $7
 218         not     $3
 219         mov     $2, $5
 220         xor     $4, $5
 221         mov     $1, $6
 222         or      $3, $6
 223         xor     $5, $6
 224         mov     $2, $8
 225         and     $7, $8
 226         or      $4, $8
 227         or      $3, $4
 228         or      $2, $3
 229         and     $1, $3
 230         mov     $3, $5
 231         xor     $8, $5
 232         not     $5
 233         and     $7, $8
 234         xor     $3, $8
 235         xor     $6, $1
 236         xor     $1, $8
 237         and     $5, $2
 238         xor     $2, $7
 239         xor     $4, $7
 240 >)
 241
 242 define(<SBOX7I>, <
 243         mov     $1, $8
 244         and     $2, $8
 245         mov     $2, $7
 246         xor     $4, $7
 247         or      $8, $7
 248         mov     $1, $6
 249         or      $4, $6
 250         and     $3, $6
 251         xor     $6, $7
 252         or      $3, $8
 253         mov     $1, $5
 254         or      $2, $5
 255         and     $4, $5
 256         xor     $5, $8
 257         xor     $2, $5
 258         mov     $4, $6
 259         xor     $8, $6
 260         not     $6
 261         or      $5, $6
 262         xor     $3, $5
 263         xor     $1, $6
 264         or      $6, $4
 265         xor     $4, $5
 266 >)
 267
 268 define(<LTI>, <
 269         rol     <$>10, $3
 270         rol     <$>27, $1
 271         mov     $2, TMP32
 272         shl     <$>7, TMP32
 273         xor     $4, $3
 274         xor     TMP32, $3
 275         xor     $2, $1
 276         xor     $4, $1
 277         rol     <$>25, $4
 278         rol     <$>31, $2
 279         mov     $1, TMP32
 280         shl     <$>3, TMP32
 281         xor     $3, $4
 282         xor     TMP32, $4
 283         xor     $1, $2
 284         xor     $3, $2
 285         rol     <$>29, $3
 286         rol     <$>19, $1
 287 >)
 288
 289 define(<PNOT>, <
 290         pxor    MINUS1, $1
 291 >)
 292
 293 define(<WSBOX0I>, <
 294         movdqa  $1, $5
 295         pxor    $3, $5
 296         movdqa  $1, $7
 297         por     $2, $7
 298         movdqa  $3, $6
 299         pxor    $4, $6
 300         pxor    $6, $7
 301         pand    $3, $6
 302         por     $2, $3
 303         pxor    $4, $2
 304         por     $1, $6
 305         pand    $3, $2
 306         pxor    $2, $6
 307         por     $7, $1
 308         pxor    $6, $1
 309         movdqa  $7, $2
 310         pand    $1, $2
 311         PNOT($7)
 312         por     $7, $4
 313         pxor    $3, $4
 314         movdqa  $1, $8
 315         pxor    $4, $8
 316         por     $4, $2
 317         pxor    $2, $5
 318 >)
 319
 320 define(<WSBOX1I>, <
 321         movdqa  $2, $6
 322         por     $4, $6
 323         pxor    $3, $6
 324         movdqa  $1, $8
 325         pxor    $2, $8
 326         movdqa  $1, $5
 327         por     $6, $5
 328         pand    $8, $5
 329         pxor    $5, $2
 330         pxor    $6, $8
 331         pand    $4, $2
 332         movdqa  $1, $7
 333         pand    $3, $7
 334         por     $7, $6
 335         por     $4, $7
 336         pxor    $5, $7
 337         PNOT($7)
 338         pxor    $2, $6
 339         pxor    $6, $5
 340         pxor    $3, $5
 341         por     $7, $1
 342         pxor    $1, $5
 343 >)
 344
 345 define(<WSBOX2I>, <
 346         movdqa  $1, $5
 347         pxor    $4, $5
 348         movdqa  $3, $7
 349         pxor    $4, $7
 350         movdqa  $2, $6
 351         por     $7, $6
 352         pxor    $6, $5
 353         movdqa  $4, $6
 354         por     $5, $6
 355         pand    $2, $6
 356         PNOT($4)
 357         movdqa  $1, $8
 358         por     $3, $8
 359         pand    $8, $7
 360         pxor    $7, $6
 361         pand    $2, $8
 362         pand    $3, $1
 363         por     $4, $1
 364         pxor    $1, $8
 365         pand    $8, $3
 366         pxor    $1, $3
 367         movdqa  $5, $7
 368         pxor    $6, $7
 369         pxor    $3, $7
 370 >)
 371
 372 define(<WSBOX3I>, <
 373         movdqa  $3, $8
 374         por     $4, $8
 375         movdqa  $2, $5
 376         pand    $8, $5
 377         movdqa  $1, $7
 378         por     $4, $7
 379         movdqa  $3, $6
 380         pxor    $7, $6
 381         pxor    $6, $5
 382         pxor    $1, $4
 383         pxor    $4, $8
 384         pxor    $2, $7
 385         pand    $6, $7
 386         pxor    $4, $7
 387         pxor    $1, $6
 388         por     $5, $4
 389         pand    $4, $6
 390         pxor    $2, $6
 391         pand    $7, $1
 392         por     $2, $1
 393         pxor    $1, $8
 394 >)
 395
 396 define(<WSBOX4I>, <
 397         movdqa  $3, $6
 398         pxor    $4, $6
 399         movdqa  $3, $7
 400         por     $4, $7
 401         pxor    $2, $7
 402         por     $4, $2
 403         movdqa  $1, $5
 404         pxor    $7, $5
 405         pxor    $7, $4
 406         pand    $1, $7
 407         pxor    $7, $6
 408         pxor    $1, $7
 409         por     $3, $7
 410         pand    $2, $1
 411         movdqa  $1, $8
 412         pxor    $4, $8
 413         PNOT($1)
 414         por     $6, $1
 415         pxor    $1, $5
 416         pxor    $2, $1
 417         pxor    $1, $7
 418 >)
 419
 420 define(<WSBOX5I>, <
 421         movdqa  $1, $6
 422         pand    $4, $6
 423         movdqa  $3, $8
 424         pxor    $6, $8
 425         movdqa  $2, $5
 426         pand    $8, $5
 427         movdqa  $1, $7
 428         pxor    $4, $7
 429         pxor    $2, $4
 430         pxor    $7, $5
 431         pand    $1, $3
 432         pand    $5, $1
 433         por     $2, $3
 434         pxor    $5, $6
 435         pxor    $3, $6
 436         movdqa  $5, $7
 437         por     $6, $7
 438         pxor    $8, $7
 439         pxor    $4, $7
 440         PNOT($2)
 441         por     $1, $2
 442         pxor    $2, $8
 443 >)
 444
 445 define(<WSBOX6I>, <
 446         movdqa  $1, $7
 447         pxor    $3, $7
 448         PNOT($3)
 449         movdqa  $2, $5
 450         pxor    $4, $5
 451         movdqa  $1, $6
 452         por     $3, $6
 453         pxor    $5, $6
 454         movdqa  $2, $8
 455         pand    $7, $8
 456         por     $4, $8
 457         por     $3, $4
 458         por     $2, $3
 459         pand    $1, $3
 460         movdqa  $3, $5
 461         pxor    $8, $5
 462         PNOT($5)
 463         pand    $7, $8
 464         pxor    $3, $8
 465         pxor    $6, $1
 466         pxor    $1, $8
 467         pand    $5, $2
 468         pxor    $2, $7
 469         pxor    $4, $7
 470 >)
 471
 472 define(<WSBOX7I>, <
 473         movdqa  $1, $8
 474         pand    $2, $8
 475         movdqa  $2, $7
 476         pxor    $4, $7
 477         por     $8, $7
 478         movdqa  $1, $6
 479         por     $4, $6
 480         pand    $3, $6
 481         pxor    $6, $7
 482         por     $3, $8
 483         movdqa  $1, $5
 484         por     $2, $5
 485         pand    $4, $5
 486         pxor    $5, $8
 487         pxor    $2, $5
 488         movdqa  $4, $6
 489         pxor    $8, $6
 490         PNOT($6)
 491         por     $5, $6
 492         pxor    $3, $5
 493         pxor    $1, $6
 494         por     $6, $4
 495         pxor    $4, $5
 496 >)
 497
 498 define(<WLTI>, <
 499         WROL(10, $3)
 500         WROL(27, $1)
 501         movdqa  $2, T0
 502         pslld   <$>7, T0
 503         pxor    $4, $3
 504         pxor    T0, $3
 505         pxor    $2, $1
 506         pxor    $4, $1
 507         WROL(25, $4)
 508         WROL(31, $2)
 509         movdqa  $1, T0
 510         pslld   <$>3, T0
 511         pxor    $3, $4
 512         pxor    T0, $4
 513         pxor    $1, $2
 514         pxor    $3, $2
 515         WROL(29, $3)
 516         WROL(19, $1)
 517 >)
 518
 519         .file "serpent-decrypt.asm"
 520
 521         C serpent_decrypt(struct serpent_context *ctx,
 522         C                 unsigned length, uint8_t *dst,
 523         C                 const uint8_t *src)
 524         .text
 525         ALIGN(16)
 526 PROLOGUE(nettle_serpent_decrypt)
 527         C save all registers that need to be saved
 528         W64_ENTRY(4, 13)
 529         push    %rbx
 530         push    %rbp
 531         push    %r12
 532         push    %r13
 533         push    %r14
 534
 535         lea     (SRC, N), SRC
 536         lea     (DST, N), DST
 537         neg     N
 538         jz      .Lend
 539
 540         cmp     $-64, N
 541         ja      .Lblock_loop
 542
 543         pcmpeqd MINUS1, MINUS1
 544
 545 .Lwblock_loop:
 546         movups  (SRC, N), X0
 547         movups  16(SRC, N), X1
 548         movups  32(SRC, N), X2
 549         movups  48(SRC, N), X3
 550
 551         WTRANSPOSE(X0,X1,X2,X3)
 552
 553         mov     $384, CNT
 554
 555         C FIXME: CNT known, no index register needed
 556         WKEYXOR(128, X0,X1,X2,X3)
 557
 558         jmp     .Lwround_start
 559
 560         ALIGN(16)
 561
 562 .Lwround_loop:
 563         WLTI(X0,X1,X2,X3)
 564 .Lwround_start:
 565         WSBOX7I(X0,X1,X2,X3, Y0,Y1,Y2,Y3)
 566         WKEYXOR(112, Y0,Y1,Y2,Y3)
 567
 568         WLTI(Y0,Y1,Y2,Y3)
 569         WSBOX6I(Y0,Y1,Y2,Y3, X0,X1,X2,X3)
 570         WKEYXOR(96, X0,X1,X2,X3)
 571
 572         WLTI(X0,X1,X2,X3)
 573         WSBOX5I(X0,X1,X2,X3, Y0,Y1,Y2,Y3)
 574         WKEYXOR(80, Y0,Y1,Y2,Y3)
 575
 576         WLTI(Y0,Y1,Y2,Y3)
 577         WSBOX4I(Y0,Y1,Y2,Y3, X0,X1,X2,X3)
 578         WKEYXOR(64, X0,X1,X2,X3)
 579
 580         WLTI(X0,X1,X2,X3)
 581         WSBOX3I(X0,X1,X2,X3, Y0,Y1,Y2,Y3)
 582         WKEYXOR(48, Y0,Y1,Y2,Y3)
 583
 584         WLTI(Y0,Y1,Y2,Y3)
 585         WSBOX2I(Y0,Y1,Y2,Y3, X0,X1,X2,X3)
 586         WKEYXOR(32, X0,X1,X2,X3)
 587
 588         WLTI(X0,X1,X2,X3)
 589         WSBOX1I(X0,X1,X2,X3, Y0,Y1,Y2,Y3)
 590         WKEYXOR(16, Y0,Y1,Y2,Y3)
 591
 592         WLTI(Y0,Y1,Y2,Y3)
 593         WSBOX0I(Y0,Y1,Y2,Y3, X0,X1,X2,X3)
 594         WKEYXOR(, X0,X1,X2,X3)
 595
 596         sub     $128, CNT
 597         jnc     .Lwround_loop
 598
 599         WTRANSPOSE(X0,X1,X2,X3)
 600
 601         movups  X0, (DST, N)
 602         movups  X1, 16(DST, N)
 603         movups  X2, 32(DST, N)
 604         movups  X3, 48(DST, N)
 605
 606         C FIXME: Adjust N, so we can use just jnc without an extra cmp.
 607         add     $64, N
 608         jz      .Lend
 609
 610         cmp     $-64, N
 611         jbe     .Lwblock_loop
 612
 613 .Lblock_loop:
 614         movl    (SRC, N), x0
 615         movl    4(SRC, N), x1
 616         movl    8(SRC, N), x2
 617         movl    12(SRC, N), x3
 618
 619         xor     512(CTX), x0
 620         xor     516(CTX), x1
 621         xor     520(CTX), x2
 622         xor     524(CTX), x3
 623
 624         mov     $384, CNT
 625         jmp     .Lround_start
 626
 627         ALIGN(16)
 628 .Lround_loop:
 629         LTI(x0,x1,x2,x3)
 630 .Lround_start:
 631         SBOX7I(x0,x1,x2,x3, y0,y1,y2,y3)
 632         xor     112(CTX, CNT), y0
 633         xor     116(CTX, CNT), y1
 634         xor     120(CTX, CNT), y2
 635         xor     124(CTX, CNT), y3
 636
 637         LTI(y0,y1,y2,y3)
 638         SBOX6I(y0,y1,y2,y3, x0,x1,x2,x3)
 639         xor      96(CTX, CNT), x0
 640         xor     100(CTX, CNT), x1
 641         xor     104(CTX, CNT), x2
 642         xor     108(CTX, CNT), x3
 643
 644         LTI(x0,x1,x2,x3)
 645         SBOX5I(x0,x1,x2,x3, y0,y1,y2,y3)
 646         xor     80(CTX, CNT), y0
 647         xor     84(CTX, CNT), y1
 648         xor     88(CTX, CNT), y2
 649         xor     92(CTX, CNT), y3
 650
 651         LTI(y0,y1,y2,y3)
 652         SBOX4I(y0,y1,y2,y3, x0,x1,x2,x3)
 653         xor     64(CTX, CNT), x0
 654         xor     68(CTX, CNT), x1
 655         xor     72(CTX, CNT), x2
 656         xor     76(CTX, CNT), x3
 657
 658         LTI(x0,x1,x2,x3)
 659         SBOX3I(x0,x1,x2,x3, y0,y1,y2,y3)
 660         xor     48(CTX, CNT), y0
 661         xor     52(CTX, CNT), y1
 662         xor     56(CTX, CNT), y2
 663         xor     60(CTX, CNT), y3
 664
 665         LTI(y0,y1,y2,y3)
 666         SBOX2I(y0,y1,y2,y3, x0,x1,x2,x3)
 667         xor     32(CTX, CNT), x0
 668         xor     36(CTX, CNT), x1
 669         xor     40(CTX, CNT), x2
 670         xor     44(CTX, CNT), x3
 671
 672         LTI(x0,x1,x2,x3)
 673         SBOX1I(x0,x1,x2,x3, y0,y1,y2,y3)
 674         xor     16(CTX, CNT), y0
 675         xor     20(CTX, CNT), y1
 676         xor     24(CTX, CNT), y2
 677         xor     28(CTX, CNT), y3
 678
 679         LTI(y0,y1,y2,y3)
 680         SBOX0I(y0,y1,y2,y3, x0,x1,x2,x3)
 681         xor       (CTX, CNT), x0
 682         xor      4(CTX, CNT), x1
 683         xor      8(CTX, CNT), x2
 684         xor     12(CTX, CNT), x3
 685         sub     $128, CNT
 686         jnc     .Lround_loop
 687
 688         movl    x0, (DST, N)
 689         movl    x1, 4(DST, N)
 690         movl    x2, 8(DST, N)
 691         movl    x3, 12(DST, N)
 692         add     $16, N
 693         jnc     .Lblock_loop
 694
 695 .Lend:
 696         pop     %r14
 697         pop     %r13
 698         pop     %r12
 699         pop     %rbp
 700         pop     %rbx
 701         W64_EXIT(4, 13)
 702         ret