crypt/sha512.s

   1         .text
   2         .p2align 4,,15
   3         .globl  __sha512_init_ctx
   4         .type   __sha512_init_ctx, @function
   5 __sha512_init_ctx:
   6         movabsq $7640891576956012808, %rax
   7         movq    $0, 72(%rdi)
   8         movq    $0, 64(%rdi)
   9         movq    %rax, (%rdi)
  10         movabsq $-4942790177534073029, %rax
  11         movq    $0, 80(%rdi)
  12         movq    %rax, 8(%rdi)
  13         movabsq $4354685564936845355, %rax
  14         movq    %rax, 16(%rdi)
  15         movabsq $-6534734903238641935, %rax
  16         movq    %rax, 24(%rdi)
  17         movabsq $5840696475078001361, %rax
  18         movq    %rax, 32(%rdi)
  19         movabsq $-7276294671716946913, %rax
  20         movq    %rax, 40(%rdi)
  21         movabsq $2270897969802886507, %rax
  22         movq    %rax, 48(%rdi)
  23         movabsq $6620516959819538809, %rax
  24         movq    %rax, 56(%rdi)
  25         ret
  26         .size   __sha512_init_ctx, .-__sha512_init_ctx
  27         .p2align 4,,15
  28         .globl  __sha512_process_block
  29         .type   __sha512_process_block, @function
  30 __sha512_process_block:
  31         pushq   %r15
  32         pushq   %r14
  33         pushq   %r13
  34         pushq   %r12
  35         pushq   %rbp
  36         pushq   %rbx
  37         movq    %rdx, %rbx
  38         subq    $624, %rsp
  39         movq    (%rdx), %rax
  40         movq    %rdi, -56(%rsp)
  41         movq    %rsi, %rdi
  42         movq    %rdx, -32(%rsp)
  43         shrq    $3, %rdi
  44         movq    %rax, -96(%rsp)
  45         movq    8(%rdx), %rax
  46         movq    %rdi, -48(%rsp)
  47         movq    %rax, -88(%rsp)
  48         movq    16(%rdx), %rax
  49         movq    %rax, -80(%rsp)
  50         movq    24(%rdx), %rax
  51         movq    %rax, -104(%rsp)
  52         movq    32(%rdx), %rax
  53         movq    %rax, -72(%rsp)
  54         movq    40(%rdx), %rax
  55         movq    %rax, -112(%rsp)
  56         movq    48(%rdx), %rax
  57         movq    %rax, -64(%rsp)
  58         movq    56(%rdx), %rax
  59         xorl    %edx, %edx
  60         addq    %rsi, 64(%rbx)
  61         adcq    %rdx, 72(%rbx)
  62         testq   %rdi, %rdi
  63         movq    %rax, -120(%rsp)
  64         je      .L4
  65         leaq    -24(%rsp), %r14
  66         leaq    K(%rip), %r15
  67         leaq    512(%r14), %rax
  68         movq    %rax, -40(%rsp)
  69         .p2align 4,,10
  70         .p2align 3
  71 .L9:
  72         movq    -56(%rsp), %rcx
  73         xorl    %eax, %eax
  74         .p2align 4,,10
  75         .p2align 3
  76 .L5:
  77         movq    (%rcx,%rax), %rdx
  78         bswap   %rdx
  79         movq    %rdx, (%r14,%rax)
  80         addq    $8, %rax
  81         cmpq    $128, %rax
  82         jne     .L5
  83         subq    $-128, -56(%rsp)
  84         movq    -40(%rsp), %rdi
  85         movq    %r14, %rsi
  86         .p2align 4,,10
  87         .p2align 3
  88 .L6:
  89         movq    112(%rsi), %rax
  90         movq    8(%rsi), %rcx
  91         addq    $8, %rsi
  92         movq    %rax, %rdx
  93         movq    %rax, %r8
  94         shrq    $6, %rax
  95         rolq    $3, %r8
  96         rorq    $19, %rdx
  97         xorq    %r8, %rdx
  98         movq    %rcx, %r8
  99         xorq    %rax, %rdx
 100         movq    -8(%rsi), %rax
 101         addq    64(%rsi), %rax
 102         rorq    $8, %r8
 103         addq    %rax, %rdx
 104         movq    %rcx, %rax
 105         shrq    $7, %rcx
 106         rorq    %rax
 107         xorq    %r8, %rax
 108         xorq    %rcx, %rax
 109         addq    %rdx, %rax
 110         movq    %rax, 120(%rsi)
 111         cmpq    %rsi, %rdi
 112         jne     .L6
 113         movq    -120(%rsp), %rax
 114         movq    -64(%rsp), %rbx
 115         xorl    %r9d, %r9d
 116         movq    -112(%rsp), %rbp
 117         movq    -72(%rsp), %rdi
 118         movabsq $4794697086780616226, %r12
 119         movq    -104(%rsp), %r13
 120         movq    -80(%rsp), %r10
 121         movq    -88(%rsp), %r11
 122         movq    -96(%rsp), %r8
 123         movq    %rax, %rsi
 124         jmp     .L8
 125         .p2align 4,,10
 126         .p2align 3
 127 .L17:
 128         movq    (%r15,%r9), %r12
 129         movq    %r10, %r13
 130         movq    %rbx, %rsi
 131         movq    %r11, %r10
 132         movq    %rbp, %rbx
 133         movq    %r8, %r11
 134         movq    %rdi, %rbp
 135         movq    %rax, %r8
 136         movq    %rcx, %rdi
 137 .L8:
 138         movq    %rdi, %rax
 139         movq    %rdi, %rdx
 140         movq    %rdi, %rcx
 141         rorq    $18, %rdx
 142         rorq    $14, %rax
 143         andq    %rbp, %rcx
 144         xorq    %rdx, %rax
 145         movq    %rdi, %rdx
 146         rolq    $23, %rdx
 147         xorq    %rax, %rdx
 148         movq    %rdi, %rax
 149         notq    %rax
 150         andq    %rbx, %rax
 151         xorq    %rcx, %rax
 152         movq    %r8, %rcx
 153         addq    %rdx, %rax
 154         movq    %r8, %rdx
 155         addq    (%r14,%r9), %rax
 156         rolq    $30, %rcx
 157         rorq    $28, %rdx
 158         addq    $8, %r9
 159         xorq    %rcx, %rdx
 160         movq    %r8, %rcx
 161         rolq    $25, %rcx
 162         xorq    %rdx, %rcx
 163         movq    %r11, %rdx
 164         addq    %rsi, %rax
 165         xorq    %r10, %rdx
 166         movq    %r11, %rsi
 167         addq    %r12, %rax
 168         andq    %r8, %rdx
 169         andq    %r10, %rsi
 170         xorq    %rsi, %rdx
 171         addq    %rcx, %rdx
 172         leaq    (%rax,%r13), %rcx
 173         addq    %rdx, %rax
 174         cmpq    $640, %r9
 175         jne     .L17
 176         addq    %rax, -96(%rsp)
 177         addq    %r8, -88(%rsp)
 178         addq    %r11, -80(%rsp)
 179         addq    %r10, -104(%rsp)
 180         addq    %rcx, -72(%rsp)
 181         addq    %rdi, -112(%rsp)
 182         addq    %rbp, -64(%rsp)
 183         addq    %rbx, -120(%rsp)
 184         subq    $16, -48(%rsp)
 185         jne     .L9
 186 .L4:
 187         movq    -32(%rsp), %rax
 188         movq    -96(%rsp), %rbx
 189         movq    %rbx, (%rax)
 190         movq    -88(%rsp), %rbx
 191         movq    %rbx, 8(%rax)
 192         movq    -80(%rsp), %rbx
 193         movq    %rbx, 16(%rax)
 194         movq    -104(%rsp), %rbx
 195         movq    %rbx, 24(%rax)
 196         movq    -72(%rsp), %rbx
 197         movq    %rbx, 32(%rax)
 198         movq    -112(%rsp), %rbx
 199         movq    %rbx, 40(%rax)
 200         movq    -64(%rsp), %rbx
 201         movq    %rbx, 48(%rax)
 202         movq    -120(%rsp), %rbx
 203         movq    %rbx, 56(%rax)
 204         addq    $624, %rsp
 205         popq    %rbx
 206         popq    %rbp
 207         popq    %r12
 208         popq    %r13
 209         popq    %r14
 210         popq    %r15
 211         ret
 212         .size   __sha512_process_block, .-__sha512_process_block
 213         .p2align 4,,15
 214         .globl  __sha512_finish_ctx
 215         .type   __sha512_finish_ctx, @function
 216 __sha512_finish_ctx:
 217         pushq   %r14
 218         pushq   %r13
 219         pushq   %r12
 220         pushq   %rbp
 221         movq    %rsi, %rbp
 222         pushq   %rbx
 223         movq    80(%rdi), %rax
 224         movq    %rdi, %rbx
 225         xorl    %edi, %edi
 226         addq    %rax, 64(%rbx)
 227         adcq    %rdi, 72(%rbx)
 228         cmpq    $111, %rax
 229         jbe     .L19
 230         movl    $240, %edx
 231         movl    $256, %r12d
 232         movl    $30, %r13d
 233         subq    %rax, %rdx
 234         movl    $31, %r14d
 235 .L20:
 236         leaq    88(%rbx,%rax), %rdi
 237         leaq    fillbuf(%rip), %rsi
 238         call    memcpy@PLT
 239         movq    64(%rbx), %rdx
 240         leaq    88(%rbx), %rdi
 241         movq    %r12, %rsi
 242         leaq    0(,%rdx,8), %rax
 243         shrq    $61, %rdx
 244         bswap   %rax
 245         movq    %rax, 88(%rbx,%r14,8)
 246         movq    72(%rbx), %rax
 247         salq    $3, %rax
 248         orq     %rdx, %rax
 249         movq    %rbx, %rdx
 250         bswap   %rax
 251         movq    %rax, 88(%rbx,%r13,8)
 252         call    __sha512_process_block
 253         xorl    %eax, %eax
 254         .p2align 4,,10
 255         .p2align 3
 256 .L21:
 257         movq    (%rbx,%rax), %rdx
 258         bswap   %rdx
 259         movq    %rdx, 0(%rbp,%rax)
 260         addq    $8, %rax
 261         cmpq    $64, %rax
 262         jne     .L21
 263         popq    %rbx
 264         movq    %rbp, %rax
 265         popq    %rbp
 266         popq    %r12
 267         popq    %r13
 268         popq    %r14
 269         ret
 270         .p2align 4,,10
 271         .p2align 3
 272 .L19:
 273         movl    $112, %edx
 274         movl    $128, %r12d
 275         movl    $14, %r13d
 276         subq    %rax, %rdx
 277         movl    $15, %r14d
 278         jmp     .L20
 279         .size   __sha512_finish_ctx, .-__sha512_finish_ctx
 280         .p2align 4,,15
 281         .globl  __sha512_process_bytes
 282         .type   __sha512_process_bytes, @function
 283 __sha512_process_bytes:
 284         pushq   %r15
 285         pushq   %r14
 286         pushq   %r13
 287         pushq   %r12
 288         movq    %rdi, %r12
 289         pushq   %rbp
 290         pushq   %rbx
 291         movq    %rdx, %rbp
 292         movq    %rsi, %rbx
 293         subq    $8, %rsp
 294         movq    80(%rdx), %r14
 295         testq   %r14, %r14
 296         jne     .L54
 297 .L25:
 298         cmpq    $127, %rbx
 299         ja      .L55
 300 .L31:
 301         testq   %rbx, %rbx
 302         jne     .L56
 303         addq    $8, %rsp
 304         popq    %rbx
 305         popq    %rbp
 306         popq    %r12
 307         popq    %r13
 308         popq    %r14
 309         popq    %r15
 310         ret
 311         .p2align 4,,10
 312         .p2align 3
 313 .L56:
 314         movq    80(%rbp), %r8
 315         cmpl    $8, %ebx
 316         leaq    88(%rbp,%r8), %rdx
 317         jnb     .L33
 318         testb   $4, %bl
 319         jne     .L57
 320         testl   %ebx, %ebx
 321         je      .L34
 322         movzbl  (%r12), %eax
 323         testb   $2, %bl
 324         movb    %al, (%rdx)
 325         jne     .L58
 326 .L34:
 327         addq    %r8, %rbx
 328         cmpq    $127, %rbx
 329         ja      .L59
 330 .L37:
 331         movq    %rbx, 80(%rbp)
 332         addq    $8, %rsp
 333         popq    %rbx
 334         popq    %rbp
 335         popq    %r12
 336         popq    %r13
 337         popq    %r14
 338         popq    %r15
 339         ret
 340         .p2align 4,,10
 341         .p2align 3
 342 .L55:
 343         movq    %rbx, %r13
 344         movq    %r12, %rdi
 345         movq    %rbp, %rdx
 346         andq    $-128, %r13
 347         andl    $127, %ebx
 348         movq    %r13, %rsi
 349         addq    %r13, %r12
 350         call    __sha512_process_block
 351         jmp     .L31
 352         .p2align 4,,10
 353         .p2align 3
 354 .L54:
 355         movl    $256, %r13d
 356         leaq    88(%rdx,%r14), %rdi
 357         subq    %r14, %r13
 358         cmpq    %rsi, %r13
 359         cmova   %rsi, %r13
 360         movq    %r12, %rsi
 361         movq    %r13, %rdx
 362         call    memcpy@PLT
 363         movq    80(%rbp), %rsi
 364         addq    %r13, %rsi
 365         cmpq    $128, %rsi
 366         movq    %rsi, 80(%rbp)
 367         ja      .L60
 368 .L26:
 369         addq    %r13, %r12
 370         subq    %r13, %rbx
 371         jmp     .L25
 372         .p2align 4,,10
 373         .p2align 3
 374 .L33:
 375         movq    (%r12), %rax
 376         leaq    8(%rdx), %rdi
 377         movq    %r12, %rsi
 378         andq    $-8, %rdi
 379         movq    %rax, (%rdx)
 380         movl    %ebx, %eax
 381         movq    -8(%r12,%rax), %rcx
 382         movq    %rcx, -8(%rdx,%rax)
 383         subq    %rdi, %rdx
 384         leal    (%rbx,%rdx), %ecx
 385         addq    %r8, %rbx
 386         subq    %rdx, %rsi
 387         shrl    $3, %ecx
 388         cmpq    $127, %rbx
 389         rep movsq
 390         jbe     .L37
 391 .L59:
 392         leaq    88(%rbp), %r12
 393         movq    %rbp, %rdx
 394         movl    $128, %esi
 395         addq    $-128, %rbx
 396         movq    %r12, %rdi
 397         call    __sha512_process_block
 398         leaq    216(%rbp), %rsi
 399         movq    %rbx, %rdx
 400         movq    %r12, %rdi
 401         call    memcpy@PLT
 402         jmp     .L37
 403         .p2align 4,,10
 404         .p2align 3
 405 .L60:
 406         leaq    88(%rbp), %r15
 407         andq    $-128, %rsi
 408         movq    %rbp, %rdx
 409         addq    %r13, %r14
 410         movq    %r15, %rdi
 411         andq    $-128, %r14
 412         call    __sha512_process_block
 413         movq    80(%rbp), %rax
 414         leaq    88(%rbp,%r14), %rsi
 415         andl    $127, %eax
 416         cmpl    $8, %eax
 417         movq    %rax, 80(%rbp)
 418         jnb     .L27
 419         testb   $4, %al
 420         jne     .L61
 421         testl   %eax, %eax
 422         je      .L26
 423         movzbl  (%rsi), %edx
 424         testb   $2, %al
 425         movb    %dl, 88(%rbp)
 426         je      .L26
 427         movl    %eax, %eax
 428         movzwl  -2(%rsi,%rax), %edx
 429         movw    %dx, -2(%r15,%rax)
 430         jmp     .L26
 431         .p2align 4,,10
 432         .p2align 3
 433 .L27:
 434         movq    (%rsi), %rdx
 435         leaq    96(%rbp), %rdi
 436         andq    $-8, %rdi
 437         movq    %rdx, 88(%rbp)
 438         movl    %eax, %edx
 439         movq    -8(%rsi,%rdx), %rcx
 440         movq    %rcx, -8(%r15,%rdx)
 441         subq    %rdi, %r15
 442         leal    (%rax,%r15), %ecx
 443         subq    %r15, %rsi
 444         movl    %ecx, %eax
 445         shrl    $3, %eax
 446         movl    %eax, %ecx
 447         rep movsq
 448         jmp     .L26
 449 .L61:
 450         movl    (%rsi), %edx
 451         movl    %eax, %eax
 452         movl    %edx, 88(%rbp)
 453         movl    -4(%rsi,%rax), %edx
 454         movl    %edx, -4(%r15,%rax)
 455         jmp     .L26
 456         .p2align 4,,10
 457         .p2align 3
 458 .L57:
 459         movl    (%r12), %eax
 460         movl    %eax, (%rdx)
 461         movl    %ebx, %eax
 462         movl    -4(%r12,%rax), %ecx
 463         movl    %ecx, -4(%rdx,%rax)
 464         jmp     .L34
 465 .L58:
 466         movl    %ebx, %eax
 467         movzwl  -2(%r12,%rax), %ecx
 468         movw    %cx, -2(%rdx,%rax)
 469         jmp     .L34
 470         .size   __sha512_process_bytes, .-__sha512_process_bytes
 471         .section        .rodata
 472         .align 32
 473         .type   K, @object
 474         .size   K, 640
 475 K:
 476         .quad   4794697086780616226
 477         .quad   8158064640168781261
 478         .quad   -5349999486874862801
 479         .quad   -1606136188198331460
 480         .quad   4131703408338449720
 481         .quad   6480981068601479193
 482         .quad   -7908458776815382629
 483         .quad   -6116909921290321640
 484         .quad   -2880145864133508542
 485         .quad   1334009975649890238
 486         .quad   2608012711638119052
 487         .quad   6128411473006802146
 488         .quad   8268148722764581231
 489         .quad   -9160688886553864527
 490         .quad   -7215885187991268811
 491         .quad   -4495734319001033068
 492         .quad   -1973867731355612462
 493         .quad   -1171420211273849373
 494         .quad   1135362057144423861
 495         .quad   2597628984639134821
 496         .quad   3308224258029322869
 497         .quad   5365058923640841347
 498         .quad   6679025012923562964
 499         .quad   8573033837759648693
 500         .quad   -7476448914759557205
 501         .quad   -6327057829258317296
 502         .quad   -5763719355590565569
 503         .quad   -4658551843659510044
 504         .quad   -4116276920077217854
 505         .quad   -3051310485924567259
 506         .quad   489312712824947311
 507         .quad   1452737877330783856
 508         .quad   2861767655752347644
 509         .quad   3322285676063803686
 510         .quad   5560940570517711597
 511         .quad   5996557281743188959
 512         .quad   7280758554555802590
 513         .quad   8532644243296465576
 514         .quad   -9096487096722542874
 515         .quad   -7894198246740708037
 516         .quad   -6719396339535248540
 517         .quad   -6333637450476146687
 518         .quad   -4446306890439682159
 519         .quad   -4076793802049405392
 520         .quad   -3345356375505022440
 521         .quad   -2983346525034927856
 522         .quad   -860691631967231958
 523         .quad   1182934255886127544
 524         .quad   1847814050463011016
 525         .quad   2177327727835720531
 526         .quad   2830643537854262169
 527         .quad   3796741975233480872
 528         .quad   4115178125766777443
 529         .quad   5681478168544905931
 530         .quad   6601373596472566643
 531         .quad   7507060721942968483
 532         .quad   8399075790359081724
 533         .quad   8693463985226723168
 534         .quad   -8878714635349349518
 535         .quad   -8302665154208450068
 536         .quad   -8016688836872298968
 537         .quad   -6606660893046293015
 538         .quad   -4685533653050689259
 539         .quad   -4147400797238176981
 540         .quad   -3880063495543823972
 541         .quad   -3348786107499101689
 542         .quad   -1523767162380948706
 543         .quad   -757361751448694408
 544         .quad   500013540394364858
 545         .quad   748580250866718886
 546         .quad   1242879168328830382
 547         .quad   1977374033974150939
 548         .quad   2944078676154940804
 549         .quad   3659926193048069267
 550         .quad   4368137639120453308
 551         .quad   4836135668995329356
 552         .quad   5532061633213252278
 553         .quad   6448918945643986474
 554         .quad   6902733635092675308
 555         .quad   7801388544844847127
 556         .align 32
 557         .type   fillbuf, @object
 558         .size   fillbuf, 128
 559 fillbuf:
 560         .byte   -128
 561         .byte   0
 562         .zero   126