third_party/boringssl/win-x86_64/crypto/bn/rsaz-x86_64.asm

   1 default rel
   2 %define XMMWORD
   3 %define YMMWORD
   4 %define ZMMWORD
   5 section .text code align=64
   6
   7
   8 EXTERN  OPENSSL_ia32cap_P
   9
  10 global  rsaz_512_sqr
  11
  12 ALIGN   32
  13 rsaz_512_sqr:
  14         mov     QWORD[8+rsp],rdi        ;WIN64 prologue
  15         mov     QWORD[16+rsp],rsi
  16         mov     rax,rsp
  17 $L$SEH_begin_rsaz_512_sqr:
  18         mov     rdi,rcx
  19         mov     rsi,rdx
  20         mov     rdx,r8
  21         mov     rcx,r9
  22         mov     r8,QWORD[40+rsp]
  23
  24
  25         push    rbx
  26         push    rbp
  27         push    r12
  28         push    r13
  29         push    r14
  30         push    r15
  31
  32         sub     rsp,128+24
  33 $L$sqr_body:
  34         mov     rbp,rdx
  35         mov     rdx,QWORD[rsi]
  36         mov     rax,QWORD[8+rsi]
  37         mov     QWORD[128+rsp],rcx
  38         jmp     NEAR $L$oop_sqr
  39
  40 ALIGN   32
  41 $L$oop_sqr:
  42         mov     DWORD[((128+8))+rsp],r8d
  43
  44         mov     rbx,rdx
  45         mul     rdx
  46         mov     r8,rax
  47         mov     rax,QWORD[16+rsi]
  48         mov     r9,rdx
  49
  50         mul     rbx
  51         add     r9,rax
  52         mov     rax,QWORD[24+rsi]
  53         mov     r10,rdx
  54         adc     r10,0
  55
  56         mul     rbx
  57         add     r10,rax
  58         mov     rax,QWORD[32+rsi]
  59         mov     r11,rdx
  60         adc     r11,0
  61
  62         mul     rbx
  63         add     r11,rax
  64         mov     rax,QWORD[40+rsi]
  65         mov     r12,rdx
  66         adc     r12,0
  67
  68         mul     rbx
  69         add     r12,rax
  70         mov     rax,QWORD[48+rsi]
  71         mov     r13,rdx
  72         adc     r13,0
  73
  74         mul     rbx
  75         add     r13,rax
  76         mov     rax,QWORD[56+rsi]
  77         mov     r14,rdx
  78         adc     r14,0
  79
  80         mul     rbx
  81         add     r14,rax
  82         mov     rax,rbx
  83         mov     r15,rdx
  84         adc     r15,0
  85
  86         add     r8,r8
  87         mov     rcx,r9
  88         adc     r9,r9
  89
  90         mul     rax
  91         mov     QWORD[rsp],rax
  92         add     r8,rdx
  93         adc     r9,0
  94
  95         mov     QWORD[8+rsp],r8
  96         shr     rcx,63
  97
  98
  99         mov     r8,QWORD[8+rsi]
 100         mov     rax,QWORD[16+rsi]
 101         mul     r8
 102         add     r10,rax
 103         mov     rax,QWORD[24+rsi]
 104         mov     rbx,rdx
 105         adc     rbx,0
 106
 107         mul     r8
 108         add     r11,rax
 109         mov     rax,QWORD[32+rsi]
 110         adc     rdx,0
 111         add     r11,rbx
 112         mov     rbx,rdx
 113         adc     rbx,0
 114
 115         mul     r8
 116         add     r12,rax
 117         mov     rax,QWORD[40+rsi]
 118         adc     rdx,0
 119         add     r12,rbx
 120         mov     rbx,rdx
 121         adc     rbx,0
 122
 123         mul     r8
 124         add     r13,rax
 125         mov     rax,QWORD[48+rsi]
 126         adc     rdx,0
 127         add     r13,rbx
 128         mov     rbx,rdx
 129         adc     rbx,0
 130
 131         mul     r8
 132         add     r14,rax
 133         mov     rax,QWORD[56+rsi]
 134         adc     rdx,0
 135         add     r14,rbx
 136         mov     rbx,rdx
 137         adc     rbx,0
 138
 139         mul     r8
 140         add     r15,rax
 141         mov     rax,r8
 142         adc     rdx,0
 143         add     r15,rbx
 144         mov     r8,rdx
 145         mov     rdx,r10
 146         adc     r8,0
 147
 148         add     rdx,rdx
 149         lea     r10,[r10*2+rcx]
 150         mov     rbx,r11
 151         adc     r11,r11
 152
 153         mul     rax
 154         add     r9,rax
 155         adc     r10,rdx
 156         adc     r11,0
 157
 158         mov     QWORD[16+rsp],r9
 159         mov     QWORD[24+rsp],r10
 160         shr     rbx,63
 161
 162
 163         mov     r9,QWORD[16+rsi]
 164         mov     rax,QWORD[24+rsi]
 165         mul     r9
 166         add     r12,rax
 167         mov     rax,QWORD[32+rsi]
 168         mov     rcx,rdx
 169         adc     rcx,0
 170
 171         mul     r9
 172         add     r13,rax
 173         mov     rax,QWORD[40+rsi]
 174         adc     rdx,0
 175         add     r13,rcx
 176         mov     rcx,rdx
 177         adc     rcx,0
 178
 179         mul     r9
 180         add     r14,rax
 181         mov     rax,QWORD[48+rsi]
 182         adc     rdx,0
 183         add     r14,rcx
 184         mov     rcx,rdx
 185         adc     rcx,0
 186
 187         mul     r9
 188         mov     r10,r12
 189         lea     r12,[r12*2+rbx]
 190         add     r15,rax
 191         mov     rax,QWORD[56+rsi]
 192         adc     rdx,0
 193         add     r15,rcx
 194         mov     rcx,rdx
 195         adc     rcx,0
 196
 197         mul     r9
 198         shr     r10,63
 199         add     r8,rax
 200         mov     rax,r9
 201         adc     rdx,0
 202         add     r8,rcx
 203         mov     r9,rdx
 204         adc     r9,0
 205
 206         mov     rcx,r13
 207         lea     r13,[r13*2+r10]
 208
 209         mul     rax
 210         add     r11,rax
 211         adc     r12,rdx
 212         adc     r13,0
 213
 214         mov     QWORD[32+rsp],r11
 215         mov     QWORD[40+rsp],r12
 216         shr     rcx,63
 217
 218
 219         mov     r10,QWORD[24+rsi]
 220         mov     rax,QWORD[32+rsi]
 221         mul     r10
 222         add     r14,rax
 223         mov     rax,QWORD[40+rsi]
 224         mov     rbx,rdx
 225         adc     rbx,0
 226
 227         mul     r10
 228         add     r15,rax
 229         mov     rax,QWORD[48+rsi]
 230         adc     rdx,0
 231         add     r15,rbx
 232         mov     rbx,rdx
 233         adc     rbx,0
 234
 235         mul     r10
 236         mov     r12,r14
 237         lea     r14,[r14*2+rcx]
 238         add     r8,rax
 239         mov     rax,QWORD[56+rsi]
 240         adc     rdx,0
 241         add     r8,rbx
 242         mov     rbx,rdx
 243         adc     rbx,0
 244
 245         mul     r10
 246         shr     r12,63
 247         add     r9,rax
 248         mov     rax,r10
 249         adc     rdx,0
 250         add     r9,rbx
 251         mov     r10,rdx
 252         adc     r10,0
 253
 254         mov     rbx,r15
 255         lea     r15,[r15*2+r12]
 256
 257         mul     rax
 258         add     r13,rax
 259         adc     r14,rdx
 260         adc     r15,0
 261
 262         mov     QWORD[48+rsp],r13
 263         mov     QWORD[56+rsp],r14
 264         shr     rbx,63
 265
 266
 267         mov     r11,QWORD[32+rsi]
 268         mov     rax,QWORD[40+rsi]
 269         mul     r11
 270         add     r8,rax
 271         mov     rax,QWORD[48+rsi]
 272         mov     rcx,rdx
 273         adc     rcx,0
 274
 275         mul     r11
 276         add     r9,rax
 277         mov     rax,QWORD[56+rsi]
 278         adc     rdx,0
 279         mov     r12,r8
 280         lea     r8,[r8*2+rbx]
 281         add     r9,rcx
 282         mov     rcx,rdx
 283         adc     rcx,0
 284
 285         mul     r11
 286         shr     r12,63
 287         add     r10,rax
 288         mov     rax,r11
 289         adc     rdx,0
 290         add     r10,rcx
 291         mov     r11,rdx
 292         adc     r11,0
 293
 294         mov     rcx,r9
 295         lea     r9,[r9*2+r12]
 296
 297         mul     rax
 298         add     r15,rax
 299         adc     r8,rdx
 300         adc     r9,0
 301
 302         mov     QWORD[64+rsp],r15
 303         mov     QWORD[72+rsp],r8
 304         shr     rcx,63
 305
 306
 307         mov     r12,QWORD[40+rsi]
 308         mov     rax,QWORD[48+rsi]
 309         mul     r12
 310         add     r10,rax
 311         mov     rax,QWORD[56+rsi]
 312         mov     rbx,rdx
 313         adc     rbx,0
 314
 315         mul     r12
 316         add     r11,rax
 317         mov     rax,r12
 318         mov     r15,r10
 319         lea     r10,[r10*2+rcx]
 320         adc     rdx,0
 321         shr     r15,63
 322         add     r11,rbx
 323         mov     r12,rdx
 324         adc     r12,0
 325
 326         mov     rbx,r11
 327         lea     r11,[r11*2+r15]
 328
 329         mul     rax
 330         add     r9,rax
 331         adc     r10,rdx
 332         adc     r11,0
 333
 334         mov     QWORD[80+rsp],r9
 335         mov     QWORD[88+rsp],r10
 336
 337
 338         mov     r13,QWORD[48+rsi]
 339         mov     rax,QWORD[56+rsi]
 340         mul     r13
 341         add     r12,rax
 342         mov     rax,r13
 343         mov     r13,rdx
 344         adc     r13,0
 345
 346         xor     r14,r14
 347         shl     rbx,1
 348         adc     r12,r12
 349         adc     r13,r13
 350         adc     r14,r14
 351
 352         mul     rax
 353         add     r11,rax
 354         adc     r12,rdx
 355         adc     r13,0
 356
 357         mov     QWORD[96+rsp],r11
 358         mov     QWORD[104+rsp],r12
 359
 360
 361         mov     rax,QWORD[56+rsi]
 362         mul     rax
 363         add     r13,rax
 364         adc     rdx,0
 365
 366         add     r14,rdx
 367
 368         mov     QWORD[112+rsp],r13
 369         mov     QWORD[120+rsp],r14
 370
 371         mov     r8,QWORD[rsp]
 372         mov     r9,QWORD[8+rsp]
 373         mov     r10,QWORD[16+rsp]
 374         mov     r11,QWORD[24+rsp]
 375         mov     r12,QWORD[32+rsp]
 376         mov     r13,QWORD[40+rsp]
 377         mov     r14,QWORD[48+rsp]
 378         mov     r15,QWORD[56+rsp]
 379
 380         call    __rsaz_512_reduce
 381
 382         add     r8,QWORD[64+rsp]
 383         adc     r9,QWORD[72+rsp]
 384         adc     r10,QWORD[80+rsp]
 385         adc     r11,QWORD[88+rsp]
 386         adc     r12,QWORD[96+rsp]
 387         adc     r13,QWORD[104+rsp]
 388         adc     r14,QWORD[112+rsp]
 389         adc     r15,QWORD[120+rsp]
 390         sbb     rcx,rcx
 391
 392         call    __rsaz_512_subtract
 393
 394         mov     rdx,r8
 395         mov     rax,r9
 396         mov     r8d,DWORD[((128+8))+rsp]
 397         mov     rsi,rdi
 398
 399         dec     r8d
 400         jnz     NEAR $L$oop_sqr
 401
 402         lea     rax,[((128+24+48))+rsp]
 403         mov     r15,QWORD[((-48))+rax]
 404         mov     r14,QWORD[((-40))+rax]
 405         mov     r13,QWORD[((-32))+rax]
 406         mov     r12,QWORD[((-24))+rax]
 407         mov     rbp,QWORD[((-16))+rax]
 408         mov     rbx,QWORD[((-8))+rax]
 409         lea     rsp,[rax]
 410 $L$sqr_epilogue:
 411         mov     rdi,QWORD[8+rsp]        ;WIN64 epilogue
 412         mov     rsi,QWORD[16+rsp]
 413         DB      0F3h,0C3h               ;repret
 414 $L$SEH_end_rsaz_512_sqr:
 415 global  rsaz_512_mul
 416
 417 ALIGN   32
 418 rsaz_512_mul:
 419         mov     QWORD[8+rsp],rdi        ;WIN64 prologue
 420         mov     QWORD[16+rsp],rsi
 421         mov     rax,rsp
 422 $L$SEH_begin_rsaz_512_mul:
 423         mov     rdi,rcx
 424         mov     rsi,rdx
 425         mov     rdx,r8
 426         mov     rcx,r9
 427         mov     r8,QWORD[40+rsp]
 428
 429
 430         push    rbx
 431         push    rbp
 432         push    r12
 433         push    r13
 434         push    r14
 435         push    r15
 436
 437         sub     rsp,128+24
 438 $L$mul_body:
 439 DB      102,72,15,110,199
 440 DB      102,72,15,110,201
 441         mov     QWORD[128+rsp],r8
 442         mov     rbx,QWORD[rdx]
 443         mov     rbp,rdx
 444         call    __rsaz_512_mul
 445
 446 DB      102,72,15,126,199
 447 DB      102,72,15,126,205
 448
 449         mov     r8,QWORD[rsp]
 450         mov     r9,QWORD[8+rsp]
 451         mov     r10,QWORD[16+rsp]
 452         mov     r11,QWORD[24+rsp]
 453         mov     r12,QWORD[32+rsp]
 454         mov     r13,QWORD[40+rsp]
 455         mov     r14,QWORD[48+rsp]
 456         mov     r15,QWORD[56+rsp]
 457
 458         call    __rsaz_512_reduce
 459         add     r8,QWORD[64+rsp]
 460         adc     r9,QWORD[72+rsp]
 461         adc     r10,QWORD[80+rsp]
 462         adc     r11,QWORD[88+rsp]
 463         adc     r12,QWORD[96+rsp]
 464         adc     r13,QWORD[104+rsp]
 465         adc     r14,QWORD[112+rsp]
 466         adc     r15,QWORD[120+rsp]
 467         sbb     rcx,rcx
 468
 469         call    __rsaz_512_subtract
 470
 471         lea     rax,[((128+24+48))+rsp]
 472         mov     r15,QWORD[((-48))+rax]
 473         mov     r14,QWORD[((-40))+rax]
 474         mov     r13,QWORD[((-32))+rax]
 475         mov     r12,QWORD[((-24))+rax]
 476         mov     rbp,QWORD[((-16))+rax]
 477         mov     rbx,QWORD[((-8))+rax]
 478         lea     rsp,[rax]
 479 $L$mul_epilogue:
 480         mov     rdi,QWORD[8+rsp]        ;WIN64 epilogue
 481         mov     rsi,QWORD[16+rsp]
 482         DB      0F3h,0C3h               ;repret
 483 $L$SEH_end_rsaz_512_mul:
 484 global  rsaz_512_mul_gather4
 485
 486 ALIGN   32
 487 rsaz_512_mul_gather4:
 488         mov     QWORD[8+rsp],rdi        ;WIN64 prologue
 489         mov     QWORD[16+rsp],rsi
 490         mov     rax,rsp
 491 $L$SEH_begin_rsaz_512_mul_gather4:
 492         mov     rdi,rcx
 493         mov     rsi,rdx
 494         mov     rdx,r8
 495         mov     rcx,r9
 496         mov     r8,QWORD[40+rsp]
 497         mov     r9,QWORD[48+rsp]
 498
 499
 500         push    rbx
 501         push    rbp
 502         push    r12
 503         push    r13
 504         push    r14
 505         push    r15
 506
 507         mov     r9d,r9d
 508         sub     rsp,128+24
 509 $L$mul_gather4_body:
 510         mov     eax,DWORD[64+r9*4+rdx]
 511 DB      102,72,15,110,199
 512         mov     ebx,DWORD[r9*4+rdx]
 513 DB      102,72,15,110,201
 514         mov     QWORD[128+rsp],r8
 515
 516         shl     rax,32
 517         or      rbx,rax
 518         mov     rax,QWORD[rsi]
 519         mov     rcx,QWORD[8+rsi]
 520         lea     rbp,[128+r9*4+rdx]
 521         mul     rbx
 522         mov     QWORD[rsp],rax
 523         mov     rax,rcx
 524         mov     r8,rdx
 525
 526         mul     rbx
 527         movd    xmm4,DWORD[rbp]
 528         add     r8,rax
 529         mov     rax,QWORD[16+rsi]
 530         mov     r9,rdx
 531         adc     r9,0
 532
 533         mul     rbx
 534         movd    xmm5,DWORD[64+rbp]
 535         add     r9,rax
 536         mov     rax,QWORD[24+rsi]
 537         mov     r10,rdx
 538         adc     r10,0
 539
 540         mul     rbx
 541         pslldq  xmm5,4
 542         add     r10,rax
 543         mov     rax,QWORD[32+rsi]
 544         mov     r11,rdx
 545         adc     r11,0
 546
 547         mul     rbx
 548         por     xmm4,xmm5
 549         add     r11,rax
 550         mov     rax,QWORD[40+rsi]
 551         mov     r12,rdx
 552         adc     r12,0
 553
 554         mul     rbx
 555         add     r12,rax
 556         mov     rax,QWORD[48+rsi]
 557         mov     r13,rdx
 558         adc     r13,0
 559
 560         mul     rbx
 561         lea     rbp,[128+rbp]
 562         add     r13,rax
 563         mov     rax,QWORD[56+rsi]
 564         mov     r14,rdx
 565         adc     r14,0
 566
 567         mul     rbx
 568 DB      102,72,15,126,227
 569         add     r14,rax
 570         mov     rax,QWORD[rsi]
 571         mov     r15,rdx
 572         adc     r15,0
 573
 574         lea     rdi,[8+rsp]
 575         mov     ecx,7
 576         jmp     NEAR $L$oop_mul_gather
 577
 578 ALIGN   32
 579 $L$oop_mul_gather:
 580         mul     rbx
 581         add     r8,rax
 582         mov     rax,QWORD[8+rsi]
 583         mov     QWORD[rdi],r8
 584         mov     r8,rdx
 585         adc     r8,0
 586
 587         mul     rbx
 588         movd    xmm4,DWORD[rbp]
 589         add     r9,rax
 590         mov     rax,QWORD[16+rsi]
 591         adc     rdx,0
 592         add     r8,r9
 593         mov     r9,rdx
 594         adc     r9,0
 595
 596         mul     rbx
 597         movd    xmm5,DWORD[64+rbp]
 598         add     r10,rax
 599         mov     rax,QWORD[24+rsi]
 600         adc     rdx,0
 601         add     r9,r10
 602         mov     r10,rdx
 603         adc     r10,0
 604
 605         mul     rbx
 606         pslldq  xmm5,4
 607         add     r11,rax
 608         mov     rax,QWORD[32+rsi]
 609         adc     rdx,0
 610         add     r10,r11
 611         mov     r11,rdx
 612         adc     r11,0
 613
 614         mul     rbx
 615         por     xmm4,xmm5
 616         add     r12,rax
 617         mov     rax,QWORD[40+rsi]
 618         adc     rdx,0
 619         add     r11,r12
 620         mov     r12,rdx
 621         adc     r12,0
 622
 623         mul     rbx
 624         add     r13,rax
 625         mov     rax,QWORD[48+rsi]
 626         adc     rdx,0
 627         add     r12,r13
 628         mov     r13,rdx
 629         adc     r13,0
 630
 631         mul     rbx
 632         add     r14,rax
 633         mov     rax,QWORD[56+rsi]
 634         adc     rdx,0
 635         add     r13,r14
 636         mov     r14,rdx
 637         adc     r14,0
 638
 639         mul     rbx
 640 DB      102,72,15,126,227
 641         add     r15,rax
 642         mov     rax,QWORD[rsi]
 643         adc     rdx,0
 644         add     r14,r15
 645         mov     r15,rdx
 646         adc     r15,0
 647
 648         lea     rbp,[128+rbp]
 649         lea     rdi,[8+rdi]
 650
 651         dec     ecx
 652         jnz     NEAR $L$oop_mul_gather
 653
 654         mov     QWORD[rdi],r8
 655         mov     QWORD[8+rdi],r9
 656         mov     QWORD[16+rdi],r10
 657         mov     QWORD[24+rdi],r11
 658         mov     QWORD[32+rdi],r12
 659         mov     QWORD[40+rdi],r13
 660         mov     QWORD[48+rdi],r14
 661         mov     QWORD[56+rdi],r15
 662
 663 DB      102,72,15,126,199
 664 DB      102,72,15,126,205
 665
 666         mov     r8,QWORD[rsp]
 667         mov     r9,QWORD[8+rsp]
 668         mov     r10,QWORD[16+rsp]
 669         mov     r11,QWORD[24+rsp]
 670         mov     r12,QWORD[32+rsp]
 671         mov     r13,QWORD[40+rsp]
 672         mov     r14,QWORD[48+rsp]
 673         mov     r15,QWORD[56+rsp]
 674
 675         call    __rsaz_512_reduce
 676         add     r8,QWORD[64+rsp]
 677         adc     r9,QWORD[72+rsp]
 678         adc     r10,QWORD[80+rsp]
 679         adc     r11,QWORD[88+rsp]
 680         adc     r12,QWORD[96+rsp]
 681         adc     r13,QWORD[104+rsp]
 682         adc     r14,QWORD[112+rsp]
 683         adc     r15,QWORD[120+rsp]
 684         sbb     rcx,rcx
 685
 686         call    __rsaz_512_subtract
 687
 688         lea     rax,[((128+24+48))+rsp]
 689         mov     r15,QWORD[((-48))+rax]
 690         mov     r14,QWORD[((-40))+rax]
 691         mov     r13,QWORD[((-32))+rax]
 692         mov     r12,QWORD[((-24))+rax]
 693         mov     rbp,QWORD[((-16))+rax]
 694         mov     rbx,QWORD[((-8))+rax]
 695         lea     rsp,[rax]
 696 $L$mul_gather4_epilogue:
 697         mov     rdi,QWORD[8+rsp]        ;WIN64 epilogue
 698         mov     rsi,QWORD[16+rsp]
 699         DB      0F3h,0C3h               ;repret
 700 $L$SEH_end_rsaz_512_mul_gather4:
 701 global  rsaz_512_mul_scatter4
 702
 703 ALIGN   32
 704 rsaz_512_mul_scatter4:
 705         mov     QWORD[8+rsp],rdi        ;WIN64 prologue
 706         mov     QWORD[16+rsp],rsi
 707         mov     rax,rsp
 708 $L$SEH_begin_rsaz_512_mul_scatter4:
 709         mov     rdi,rcx
 710         mov     rsi,rdx
 711         mov     rdx,r8
 712         mov     rcx,r9
 713         mov     r8,QWORD[40+rsp]
 714         mov     r9,QWORD[48+rsp]
 715
 716
 717         push    rbx
 718         push    rbp
 719         push    r12
 720         push    r13
 721         push    r14
 722         push    r15
 723
 724         mov     r9d,r9d
 725         sub     rsp,128+24
 726 $L$mul_scatter4_body:
 727         lea     r8,[r9*4+r8]
 728 DB      102,72,15,110,199
 729 DB      102,72,15,110,202
 730 DB      102,73,15,110,208
 731         mov     QWORD[128+rsp],rcx
 732
 733         mov     rbp,rdi
 734         mov     rbx,QWORD[rdi]
 735         call    __rsaz_512_mul
 736
 737 DB      102,72,15,126,199
 738 DB      102,72,15,126,205
 739
 740         mov     r8,QWORD[rsp]
 741         mov     r9,QWORD[8+rsp]
 742         mov     r10,QWORD[16+rsp]
 743         mov     r11,QWORD[24+rsp]
 744         mov     r12,QWORD[32+rsp]
 745         mov     r13,QWORD[40+rsp]
 746         mov     r14,QWORD[48+rsp]
 747         mov     r15,QWORD[56+rsp]
 748
 749         call    __rsaz_512_reduce
 750         add     r8,QWORD[64+rsp]
 751         adc     r9,QWORD[72+rsp]
 752         adc     r10,QWORD[80+rsp]
 753         adc     r11,QWORD[88+rsp]
 754         adc     r12,QWORD[96+rsp]
 755         adc     r13,QWORD[104+rsp]
 756         adc     r14,QWORD[112+rsp]
 757         adc     r15,QWORD[120+rsp]
 758 DB      102,72,15,126,214
 759         sbb     rcx,rcx
 760
 761         call    __rsaz_512_subtract
 762
 763         mov     DWORD[rsi],r8d
 764         shr     r8,32
 765         mov     DWORD[128+rsi],r9d
 766         shr     r9,32
 767         mov     DWORD[256+rsi],r10d
 768         shr     r10,32
 769         mov     DWORD[384+rsi],r11d
 770         shr     r11,32
 771         mov     DWORD[512+rsi],r12d
 772         shr     r12,32
 773         mov     DWORD[640+rsi],r13d
 774         shr     r13,32
 775         mov     DWORD[768+rsi],r14d
 776         shr     r14,32
 777         mov     DWORD[896+rsi],r15d
 778         shr     r15,32
 779         mov     DWORD[64+rsi],r8d
 780         mov     DWORD[192+rsi],r9d
 781         mov     DWORD[320+rsi],r10d
 782         mov     DWORD[448+rsi],r11d
 783         mov     DWORD[576+rsi],r12d
 784         mov     DWORD[704+rsi],r13d
 785         mov     DWORD[832+rsi],r14d
 786         mov     DWORD[960+rsi],r15d
 787
 788         lea     rax,[((128+24+48))+rsp]
 789         mov     r15,QWORD[((-48))+rax]
 790         mov     r14,QWORD[((-40))+rax]
 791         mov     r13,QWORD[((-32))+rax]
 792         mov     r12,QWORD[((-24))+rax]
 793         mov     rbp,QWORD[((-16))+rax]
 794         mov     rbx,QWORD[((-8))+rax]
 795         lea     rsp,[rax]
 796 $L$mul_scatter4_epilogue:
 797         mov     rdi,QWORD[8+rsp]        ;WIN64 epilogue
 798         mov     rsi,QWORD[16+rsp]
 799         DB      0F3h,0C3h               ;repret
 800 $L$SEH_end_rsaz_512_mul_scatter4:
 801 global  rsaz_512_mul_by_one
 802
 803 ALIGN   32
 804 rsaz_512_mul_by_one:
 805         mov     QWORD[8+rsp],rdi        ;WIN64 prologue
 806         mov     QWORD[16+rsp],rsi
 807         mov     rax,rsp
 808 $L$SEH_begin_rsaz_512_mul_by_one:
 809         mov     rdi,rcx
 810         mov     rsi,rdx
 811         mov     rdx,r8
 812         mov     rcx,r9
 813
 814
 815         push    rbx
 816         push    rbp
 817         push    r12
 818         push    r13
 819         push    r14
 820         push    r15
 821
 822         sub     rsp,128+24
 823 $L$mul_by_one_body:
 824         mov     rbp,rdx
 825         mov     QWORD[128+rsp],rcx
 826
 827         mov     r8,QWORD[rsi]
 828         pxor    xmm0,xmm0
 829         mov     r9,QWORD[8+rsi]
 830         mov     r10,QWORD[16+rsi]
 831         mov     r11,QWORD[24+rsi]
 832         mov     r12,QWORD[32+rsi]
 833         mov     r13,QWORD[40+rsi]
 834         mov     r14,QWORD[48+rsi]
 835         mov     r15,QWORD[56+rsi]
 836
 837         movdqa  XMMWORD[rsp],xmm0
 838         movdqa  XMMWORD[16+rsp],xmm0
 839         movdqa  XMMWORD[32+rsp],xmm0
 840         movdqa  XMMWORD[48+rsp],xmm0
 841         movdqa  XMMWORD[64+rsp],xmm0
 842         movdqa  XMMWORD[80+rsp],xmm0
 843         movdqa  XMMWORD[96+rsp],xmm0
 844         call    __rsaz_512_reduce
 845         mov     QWORD[rdi],r8
 846         mov     QWORD[8+rdi],r9
 847         mov     QWORD[16+rdi],r10
 848         mov     QWORD[24+rdi],r11
 849         mov     QWORD[32+rdi],r12
 850         mov     QWORD[40+rdi],r13
 851         mov     QWORD[48+rdi],r14
 852         mov     QWORD[56+rdi],r15
 853
 854         lea     rax,[((128+24+48))+rsp]
 855         mov     r15,QWORD[((-48))+rax]
 856         mov     r14,QWORD[((-40))+rax]
 857         mov     r13,QWORD[((-32))+rax]
 858         mov     r12,QWORD[((-24))+rax]
 859         mov     rbp,QWORD[((-16))+rax]
 860         mov     rbx,QWORD[((-8))+rax]
 861         lea     rsp,[rax]
 862 $L$mul_by_one_epilogue:
 863         mov     rdi,QWORD[8+rsp]        ;WIN64 epilogue
 864         mov     rsi,QWORD[16+rsp]
 865         DB      0F3h,0C3h               ;repret
 866 $L$SEH_end_rsaz_512_mul_by_one:
 867
 868 ALIGN   32
 869 __rsaz_512_reduce:
 870         mov     rbx,r8
 871         imul    rbx,QWORD[((128+8))+rsp]
 872         mov     rax,QWORD[rbp]
 873         mov     ecx,8
 874         jmp     NEAR $L$reduction_loop
 875
 876 ALIGN   32
 877 $L$reduction_loop:
 878         mul     rbx
 879         mov     rax,QWORD[8+rbp]
 880         neg     r8
 881         mov     r8,rdx
 882         adc     r8,0
 883
 884         mul     rbx
 885         add     r9,rax
 886         mov     rax,QWORD[16+rbp]
 887         adc     rdx,0
 888         add     r8,r9
 889         mov     r9,rdx
 890         adc     r9,0
 891
 892         mul     rbx
 893         add     r10,rax
 894         mov     rax,QWORD[24+rbp]
 895         adc     rdx,0
 896         add     r9,r10
 897         mov     r10,rdx
 898         adc     r10,0
 899
 900         mul     rbx
 901         add     r11,rax
 902         mov     rax,QWORD[32+rbp]
 903         adc     rdx,0
 904         add     r10,r11
 905         mov     rsi,QWORD[((128+8))+rsp]
 906
 907
 908         adc     rdx,0
 909         mov     r11,rdx
 910
 911         mul     rbx
 912         add     r12,rax
 913         mov     rax,QWORD[40+rbp]
 914         adc     rdx,0
 915         imul    rsi,r8
 916         add     r11,r12
 917         mov     r12,rdx
 918         adc     r12,0
 919
 920         mul     rbx
 921         add     r13,rax
 922         mov     rax,QWORD[48+rbp]
 923         adc     rdx,0
 924         add     r12,r13
 925         mov     r13,rdx
 926         adc     r13,0
 927
 928         mul     rbx
 929         add     r14,rax
 930         mov     rax,QWORD[56+rbp]
 931         adc     rdx,0
 932         add     r13,r14
 933         mov     r14,rdx
 934         adc     r14,0
 935
 936         mul     rbx
 937         mov     rbx,rsi
 938         add     r15,rax
 939         mov     rax,QWORD[rbp]
 940         adc     rdx,0
 941         add     r14,r15
 942         mov     r15,rdx
 943         adc     r15,0
 944
 945         dec     ecx
 946         jne     NEAR $L$reduction_loop
 947
 948         DB      0F3h,0C3h               ;repret
 949
 950
 951 ALIGN   32
 952 __rsaz_512_subtract:
 953         mov     QWORD[rdi],r8
 954         mov     QWORD[8+rdi],r9
 955         mov     QWORD[16+rdi],r10
 956         mov     QWORD[24+rdi],r11
 957         mov     QWORD[32+rdi],r12
 958         mov     QWORD[40+rdi],r13
 959         mov     QWORD[48+rdi],r14
 960         mov     QWORD[56+rdi],r15
 961
 962         mov     r8,QWORD[rbp]
 963         mov     r9,QWORD[8+rbp]
 964         neg     r8
 965         not     r9
 966         and     r8,rcx
 967         mov     r10,QWORD[16+rbp]
 968         and     r9,rcx
 969         not     r10
 970         mov     r11,QWORD[24+rbp]
 971         and     r10,rcx
 972         not     r11
 973         mov     r12,QWORD[32+rbp]
 974         and     r11,rcx
 975         not     r12
 976         mov     r13,QWORD[40+rbp]
 977         and     r12,rcx
 978         not     r13
 979         mov     r14,QWORD[48+rbp]
 980         and     r13,rcx
 981         not     r14
 982         mov     r15,QWORD[56+rbp]
 983         and     r14,rcx
 984         not     r15
 985         and     r15,rcx
 986
 987         add     r8,QWORD[rdi]
 988         adc     r9,QWORD[8+rdi]
 989         adc     r10,QWORD[16+rdi]
 990         adc     r11,QWORD[24+rdi]
 991         adc     r12,QWORD[32+rdi]
 992         adc     r13,QWORD[40+rdi]
 993         adc     r14,QWORD[48+rdi]
 994         adc     r15,QWORD[56+rdi]
 995
 996         mov     QWORD[rdi],r8
 997         mov     QWORD[8+rdi],r9
 998         mov     QWORD[16+rdi],r10
 999         mov     QWORD[24+rdi],r11
1000         mov     QWORD[32+rdi],r12
1001         mov     QWORD[40+rdi],r13
1002         mov     QWORD[48+rdi],r14
1003         mov     QWORD[56+rdi],r15
1004
1005         DB      0F3h,0C3h               ;repret
1006
1007
1008 ALIGN   32
1009 __rsaz_512_mul:
1010         lea     rdi,[8+rsp]
1011
1012         mov     rax,QWORD[rsi]
1013         mul     rbx
1014         mov     QWORD[rdi],rax
1015         mov     rax,QWORD[8+rsi]
1016         mov     r8,rdx
1017
1018         mul     rbx
1019         add     r8,rax
1020         mov     rax,QWORD[16+rsi]
1021         mov     r9,rdx
1022         adc     r9,0
1023
1024         mul     rbx
1025         add     r9,rax
1026         mov     rax,QWORD[24+rsi]
1027         mov     r10,rdx
1028         adc     r10,0
1029
1030         mul     rbx
1031         add     r10,rax
1032         mov     rax,QWORD[32+rsi]
1033         mov     r11,rdx
1034         adc     r11,0
1035
1036         mul     rbx
1037         add     r11,rax
1038         mov     rax,QWORD[40+rsi]
1039         mov     r12,rdx
1040         adc     r12,0
1041
1042         mul     rbx
1043         add     r12,rax
1044         mov     rax,QWORD[48+rsi]
1045         mov     r13,rdx
1046         adc     r13,0
1047
1048         mul     rbx
1049         add     r13,rax
1050         mov     rax,QWORD[56+rsi]
1051         mov     r14,rdx
1052         adc     r14,0
1053
1054         mul     rbx
1055         add     r14,rax
1056         mov     rax,QWORD[rsi]
1057         mov     r15,rdx
1058         adc     r15,0
1059
1060         lea     rbp,[8+rbp]
1061         lea     rdi,[8+rdi]
1062
1063         mov     ecx,7
1064         jmp     NEAR $L$oop_mul
1065
1066 ALIGN   32
1067 $L$oop_mul:
1068         mov     rbx,QWORD[rbp]
1069         mul     rbx
1070         add     r8,rax
1071         mov     rax,QWORD[8+rsi]
1072         mov     QWORD[rdi],r8
1073         mov     r8,rdx
1074         adc     r8,0
1075
1076         mul     rbx
1077         add     r9,rax
1078         mov     rax,QWORD[16+rsi]
1079         adc     rdx,0
1080         add     r8,r9
1081         mov     r9,rdx
1082         adc     r9,0
1083
1084         mul     rbx
1085         add     r10,rax
1086         mov     rax,QWORD[24+rsi]
1087         adc     rdx,0
1088         add     r9,r10
1089         mov     r10,rdx
1090         adc     r10,0
1091
1092         mul     rbx
1093         add     r11,rax
1094         mov     rax,QWORD[32+rsi]
1095         adc     rdx,0
1096         add     r10,r11
1097         mov     r11,rdx
1098         adc     r11,0
1099
1100         mul     rbx
1101         add     r12,rax
1102         mov     rax,QWORD[40+rsi]
1103         adc     rdx,0
1104         add     r11,r12
1105         mov     r12,rdx
1106         adc     r12,0
1107
1108         mul     rbx
1109         add     r13,rax
1110         mov     rax,QWORD[48+rsi]
1111         adc     rdx,0
1112         add     r12,r13
1113         mov     r13,rdx
1114         adc     r13,0
1115
1116         mul     rbx
1117         add     r14,rax
1118         mov     rax,QWORD[56+rsi]
1119         adc     rdx,0
1120         add     r13,r14
1121         mov     r14,rdx
1122         lea     rbp,[8+rbp]
1123         adc     r14,0
1124
1125         mul     rbx
1126         add     r15,rax
1127         mov     rax,QWORD[rsi]
1128         adc     rdx,0
1129         add     r14,r15
1130         mov     r15,rdx
1131         adc     r15,0
1132
1133         lea     rdi,[8+rdi]
1134
1135         dec     ecx
1136         jnz     NEAR $L$oop_mul
1137
1138         mov     QWORD[rdi],r8
1139         mov     QWORD[8+rdi],r9
1140         mov     QWORD[16+rdi],r10
1141         mov     QWORD[24+rdi],r11
1142         mov     QWORD[32+rdi],r12
1143         mov     QWORD[40+rdi],r13
1144         mov     QWORD[48+rdi],r14
1145         mov     QWORD[56+rdi],r15
1146
1147         DB      0F3h,0C3h               ;repret
1148
1149 global  rsaz_512_scatter4
1150
1151 ALIGN   16
1152 rsaz_512_scatter4:
1153         lea     rcx,[r8*4+rcx]
1154         mov     r9d,8
1155         jmp     NEAR $L$oop_scatter
1156 ALIGN   16
1157 $L$oop_scatter:
1158         mov     rax,QWORD[rdx]
1159         lea     rdx,[8+rdx]
1160         mov     DWORD[rcx],eax
1161         shr     rax,32
1162         mov     DWORD[64+rcx],eax
1163         lea     rcx,[128+rcx]
1164         dec     r9d
1165         jnz     NEAR $L$oop_scatter
1166         DB      0F3h,0C3h               ;repret
1167
1168
1169 global  rsaz_512_gather4
1170
1171 ALIGN   16
1172 rsaz_512_gather4:
1173         lea     rdx,[r8*4+rdx]
1174         mov     r9d,8
1175         jmp     NEAR $L$oop_gather
1176 ALIGN   16
1177 $L$oop_gather:
1178         mov     eax,DWORD[rdx]
1179         mov     r8d,DWORD[64+rdx]
1180         lea     rdx,[128+rdx]
1181         shl     r8,32
1182         or      rax,r8
1183         mov     QWORD[rcx],rax
1184         lea     rcx,[8+rcx]
1185         dec     r9d
1186         jnz     NEAR $L$oop_gather
1187         DB      0F3h,0C3h               ;repret
1188
1189 EXTERN  __imp_RtlVirtualUnwind
1190
1191 ALIGN   16
1192 se_handler:
1193         push    rsi
1194         push    rdi
1195         push    rbx
1196         push    rbp
1197         push    r12
1198         push    r13
1199         push    r14
1200         push    r15
1201         pushfq
1202         sub     rsp,64
1203
1204         mov     rax,QWORD[120+r8]
1205         mov     rbx,QWORD[248+r8]
1206
1207         mov     rsi,QWORD[8+r9]
1208         mov     r11,QWORD[56+r9]
1209
1210         mov     r10d,DWORD[r11]
1211         lea     r10,[r10*1+rsi]
1212         cmp     rbx,r10
1213         jb      NEAR $L$common_seh_tail
1214
1215         mov     rax,QWORD[152+r8]
1216
1217         mov     r10d,DWORD[4+r11]
1218         lea     r10,[r10*1+rsi]
1219         cmp     rbx,r10
1220         jae     NEAR $L$common_seh_tail
1221
1222         lea     rax,[((128+24+48))+rax]
1223
1224         mov     rbx,QWORD[((-8))+rax]
1225         mov     rbp,QWORD[((-16))+rax]
1226         mov     r12,QWORD[((-24))+rax]
1227         mov     r13,QWORD[((-32))+rax]
1228         mov     r14,QWORD[((-40))+rax]
1229         mov     r15,QWORD[((-48))+rax]
1230         mov     QWORD[144+r8],rbx
1231         mov     QWORD[160+r8],rbp
1232         mov     QWORD[216+r8],r12
1233         mov     QWORD[224+r8],r13
1234         mov     QWORD[232+r8],r14
1235         mov     QWORD[240+r8],r15
1236
1237 $L$common_seh_tail:
1238         mov     rdi,QWORD[8+rax]
1239         mov     rsi,QWORD[16+rax]
1240         mov     QWORD[152+r8],rax
1241         mov     QWORD[168+r8],rsi
1242         mov     QWORD[176+r8],rdi
1243
1244         mov     rdi,QWORD[40+r9]
1245         mov     rsi,r8
1246         mov     ecx,154
1247         DD      0xa548f3fc
1248
1249         mov     rsi,r9
1250         xor     rcx,rcx
1251         mov     rdx,QWORD[8+rsi]
1252         mov     r8,QWORD[rsi]
1253         mov     r9,QWORD[16+rsi]
1254         mov     r10,QWORD[40+rsi]
1255         lea     r11,[56+rsi]
1256         lea     r12,[24+rsi]
1257         mov     QWORD[32+rsp],r10
1258         mov     QWORD[40+rsp],r11
1259         mov     QWORD[48+rsp],r12
1260         mov     QWORD[56+rsp],rcx
1261         call    QWORD[__imp_RtlVirtualUnwind]
1262
1263         mov     eax,1
1264         add     rsp,64
1265         popfq
1266         pop     r15
1267         pop     r14
1268         pop     r13
1269         pop     r12
1270         pop     rbp
1271         pop     rbx
1272         pop     rdi
1273         pop     rsi
1274         DB      0F3h,0C3h               ;repret
1275
1276
1277 section .pdata rdata align=4
1278 ALIGN   4
1279         DD      $L$SEH_begin_rsaz_512_sqr wrt ..imagebase
1280         DD      $L$SEH_end_rsaz_512_sqr wrt ..imagebase
1281         DD      $L$SEH_info_rsaz_512_sqr wrt ..imagebase
1282
1283         DD      $L$SEH_begin_rsaz_512_mul wrt ..imagebase
1284         DD      $L$SEH_end_rsaz_512_mul wrt ..imagebase
1285         DD      $L$SEH_info_rsaz_512_mul wrt ..imagebase
1286
1287         DD      $L$SEH_begin_rsaz_512_mul_gather4 wrt ..imagebase
1288         DD      $L$SEH_end_rsaz_512_mul_gather4 wrt ..imagebase
1289         DD      $L$SEH_info_rsaz_512_mul_gather4 wrt ..imagebase
1290
1291         DD      $L$SEH_begin_rsaz_512_mul_scatter4 wrt ..imagebase
1292         DD      $L$SEH_end_rsaz_512_mul_scatter4 wrt ..imagebase
1293         DD      $L$SEH_info_rsaz_512_mul_scatter4 wrt ..imagebase
1294
1295         DD      $L$SEH_begin_rsaz_512_mul_by_one wrt ..imagebase
1296         DD      $L$SEH_end_rsaz_512_mul_by_one wrt ..imagebase
1297         DD      $L$SEH_info_rsaz_512_mul_by_one wrt ..imagebase
1298
1299 section .xdata rdata align=8
1300 ALIGN   8
1301 $L$SEH_info_rsaz_512_sqr:
1302 DB      9,0,0,0
1303         DD      se_handler wrt ..imagebase
1304         DD      $L$sqr_body wrt ..imagebase,$L$sqr_epilogue wrt ..imagebase
1305 $L$SEH_info_rsaz_512_mul:
1306 DB      9,0,0,0
1307         DD      se_handler wrt ..imagebase
1308         DD      $L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase
1309 $L$SEH_info_rsaz_512_mul_gather4:
1310 DB      9,0,0,0
1311         DD      se_handler wrt ..imagebase
1312         DD      $L$mul_gather4_body wrt ..imagebase,$L$mul_gather4_epilogue wrt ..imagebase
1313 $L$SEH_info_rsaz_512_mul_scatter4:
1314 DB      9,0,0,0
1315         DD      se_handler wrt ..imagebase
1316         DD      $L$mul_scatter4_body wrt ..imagebase,$L$mul_scatter4_epilogue wrt ..imagebase
1317 $L$SEH_info_rsaz_512_mul_by_one:
1318 DB      9,0,0,0
1319         DD      se_handler wrt ..imagebase
1320         DD      $L$mul_by_one_body wrt ..imagebase,$L$mul_by_one_epilogue wrt ..imagebase