1 #if defined(__x86_64__) 2 .text 3 4 5 6 .globl _bn_mul_mont 7 .private_extern _bn_mul_mont 8 9 .p2align 4 10 _bn_mul_mont: 11 testl $3,%r9d 12 jnz L$mul_enter 13 cmpl $8,%r9d 14 jb L$mul_enter 15 cmpq %rsi,%rdx 16 jne L$mul4x_enter 17 testl $7,%r9d 18 jz L$sqr8x_enter 19 jmp L$mul4x_enter 20 21 .p2align 4 22 L$mul_enter: 23 pushq %rbx 24 pushq %rbp 25 pushq %r12 26 pushq %r13 27 pushq %r14 28 pushq %r15 29 30 movl %r9d,%r9d 31 leaq 2(%r9),%r10 32 movq %rsp,%r11 33 negq %r10 34 leaq (%rsp,%r10,8),%rsp 35 andq $-1024,%rsp 36 37 movq %r11,8(%rsp,%r9,8) 38 L$mul_body: 39 movq %rdx,%r12 40 movq (%r8),%r8 41 movq (%r12),%rbx 42 movq (%rsi),%rax 43 44 xorq %r14,%r14 45 xorq %r15,%r15 46 47 movq %r8,%rbp 48 mulq %rbx 49 movq %rax,%r10 50 movq (%rcx),%rax 51 52 imulq %r10,%rbp 53 movq %rdx,%r11 54 55 mulq %rbp 56 addq %rax,%r10 57 movq 8(%rsi),%rax 58 adcq $0,%rdx 59 movq %rdx,%r13 60 61 leaq 1(%r15),%r15 62 jmp L$1st_enter 63 64 .p2align 4 65 L$1st: 66 addq %rax,%r13 67 movq (%rsi,%r15,8),%rax 68 adcq $0,%rdx 69 addq %r11,%r13 70 movq %r10,%r11 71 adcq $0,%rdx 72 movq %r13,-16(%rsp,%r15,8) 73 movq %rdx,%r13 74 75 L$1st_enter: 76 mulq %rbx 77 addq %rax,%r11 78 movq (%rcx,%r15,8),%rax 79 adcq $0,%rdx 80 leaq 1(%r15),%r15 81 movq %rdx,%r10 82 83 mulq %rbp 84 cmpq %r9,%r15 85 jne L$1st 86 87 addq %rax,%r13 88 movq (%rsi),%rax 89 adcq $0,%rdx 90 addq %r11,%r13 91 adcq $0,%rdx 92 movq %r13,-16(%rsp,%r15,8) 93 movq %rdx,%r13 94 movq %r10,%r11 95 96 xorq %rdx,%rdx 97 addq %r11,%r13 98 adcq $0,%rdx 99 movq %r13,-8(%rsp,%r9,8) 100 movq %rdx,(%rsp,%r9,8) 101 102 leaq 1(%r14),%r14 103 jmp L$outer 104 .p2align 4 105 L$outer: 106 movq (%r12,%r14,8),%rbx 107 xorq %r15,%r15 108 movq %r8,%rbp 109 movq (%rsp),%r10 110 mulq %rbx 111 addq %rax,%r10 112 movq (%rcx),%rax 113 adcq $0,%rdx 114 115 imulq %r10,%rbp 116 movq %rdx,%r11 117 118 mulq %rbp 119 addq %rax,%r10 120 movq 8(%rsi),%rax 121 adcq $0,%rdx 122 movq 8(%rsp),%r10 123 movq %rdx,%r13 124 125 leaq 1(%r15),%r15 126 jmp L$inner_enter 127 128 .p2align 4 129 L$inner: 130 addq %rax,%r13 131 movq (%rsi,%r15,8),%rax 132 adcq $0,%rdx 133 addq %r10,%r13 134 movq (%rsp,%r15,8),%r10 135 adcq $0,%rdx 136 movq %r13,-16(%rsp,%r15,8) 137 movq %rdx,%r13 138 139 L$inner_enter: 140 mulq %rbx 141 addq %rax,%r11 142 movq (%rcx,%r15,8),%rax 143 adcq $0,%rdx 144 addq %r11,%r10 145 movq %rdx,%r11 146 adcq $0,%r11 147 leaq 1(%r15),%r15 148 149 mulq %rbp 150 cmpq %r9,%r15 151 jne L$inner 152 153 addq %rax,%r13 154 movq (%rsi),%rax 155 adcq $0,%rdx 156 addq %r10,%r13 157 movq (%rsp,%r15,8),%r10 158 adcq $0,%rdx 159 movq %r13,-16(%rsp,%r15,8) 160 movq %rdx,%r13 161 162 xorq %rdx,%rdx 163 addq %r11,%r13 164 adcq $0,%rdx 165 addq %r10,%r13 166 adcq $0,%rdx 167 movq %r13,-8(%rsp,%r9,8) 168 movq %rdx,(%rsp,%r9,8) 169 170 leaq 1(%r14),%r14 171 cmpq %r9,%r14 172 jb L$outer 173 174 xorq %r14,%r14 175 movq (%rsp),%rax 176 leaq (%rsp),%rsi 177 movq %r9,%r15 178 jmp L$sub 179 .p2align 4 180 L$sub: sbbq (%rcx,%r14,8),%rax 181 movq %rax,(%rdi,%r14,8) 182 movq 8(%rsi,%r14,8),%rax 183 leaq 1(%r14),%r14 184 decq %r15 185 jnz L$sub 186 187 sbbq $0,%rax 188 xorq %r14,%r14 189 movq %r9,%r15 190 .p2align 4 191 L$copy: 192 movq (%rsp,%r14,8),%rsi 193 movq (%rdi,%r14,8),%rcx 194 xorq %rcx,%rsi 195 andq %rax,%rsi 196 xorq %rcx,%rsi 197 movq %r14,(%rsp,%r14,8) 198 movq %rsi,(%rdi,%r14,8) 199 leaq 1(%r14),%r14 200 subq $1,%r15 201 jnz L$copy 202 203 movq 8(%rsp,%r9,8),%rsi 204 movq $1,%rax 205 movq (%rsi),%r15 206 movq 8(%rsi),%r14 207 movq 16(%rsi),%r13 208 movq 24(%rsi),%r12 209 movq 32(%rsi),%rbp 210 movq 40(%rsi),%rbx 211 leaq 48(%rsi),%rsp 212 L$mul_epilogue: 213 .byte 0xf3,0xc3 214 215 216 .p2align 4 217 bn_mul4x_mont: 218 L$mul4x_enter: 219 pushq %rbx 220 pushq %rbp 221 pushq %r12 222 pushq %r13 223 pushq %r14 224 pushq %r15 225 226 movl %r9d,%r9d 227 leaq 4(%r9),%r10 228 movq %rsp,%r11 229 negq %r10 230 leaq (%rsp,%r10,8),%rsp 231 andq $-1024,%rsp 232 233 movq %r11,8(%rsp,%r9,8) 234 L$mul4x_body: 235 movq %rdi,16(%rsp,%r9,8) 236 movq %rdx,%r12 237 movq (%r8),%r8 238 movq (%r12),%rbx 239 movq (%rsi),%rax 240 241 xorq %r14,%r14 242 xorq %r15,%r15 243 244 movq %r8,%rbp 245 mulq %rbx 246 movq %rax,%r10 247 movq (%rcx),%rax 248 249 imulq %r10,%rbp 250 movq %rdx,%r11 251 252 mulq %rbp 253 addq %rax,%r10 254 movq 8(%rsi),%rax 255 adcq $0,%rdx 256 movq %rdx,%rdi 257 258 mulq %rbx 259 addq %rax,%r11 260 movq 8(%rcx),%rax 261 adcq $0,%rdx 262 movq %rdx,%r10 263 264 mulq %rbp 265 addq %rax,%rdi 266 movq 16(%rsi),%rax 267 adcq $0,%rdx 268 addq %r11,%rdi 269 leaq 4(%r15),%r15 270 adcq $0,%rdx 271 movq %rdi,(%rsp) 272 movq %rdx,%r13 273 jmp L$1st4x 274 .p2align 4 275 L$1st4x: 276 mulq %rbx 277 addq %rax,%r10 278 movq -16(%rcx,%r15,8),%rax 279 adcq $0,%rdx 280 movq %rdx,%r11 281 282 mulq %rbp 283 addq %rax,%r13 284 movq -8(%rsi,%r15,8),%rax 285 adcq $0,%rdx 286 addq %r10,%r13 287 adcq $0,%rdx 288 movq %r13,-24(%rsp,%r15,8) 289 movq %rdx,%rdi 290 291 mulq %rbx 292 addq %rax,%r11 293 movq -8(%rcx,%r15,8),%rax 294 adcq $0,%rdx 295 movq %rdx,%r10 296 297 mulq %rbp 298 addq %rax,%rdi 299 movq (%rsi,%r15,8),%rax 300 adcq $0,%rdx 301 addq %r11,%rdi 302 adcq $0,%rdx 303 movq %rdi,-16(%rsp,%r15,8) 304 movq %rdx,%r13 305 306 mulq %rbx 307 addq %rax,%r10 308 movq (%rcx,%r15,8),%rax 309 adcq $0,%rdx 310 movq %rdx,%r11 311 312 mulq %rbp 313 addq %rax,%r13 314 movq 8(%rsi,%r15,8),%rax 315 adcq $0,%rdx 316 addq %r10,%r13 317 adcq $0,%rdx 318 movq %r13,-8(%rsp,%r15,8) 319 movq %rdx,%rdi 320 321 mulq %rbx 322 addq %rax,%r11 323 movq 8(%rcx,%r15,8),%rax 324 adcq $0,%rdx 325 leaq 4(%r15),%r15 326 movq %rdx,%r10 327 328 mulq %rbp 329 addq %rax,%rdi 330 movq -16(%rsi,%r15,8),%rax 331 adcq $0,%rdx 332 addq %r11,%rdi 333 adcq $0,%rdx 334 movq %rdi,-32(%rsp,%r15,8) 335 movq %rdx,%r13 336 cmpq %r9,%r15 337 jb L$1st4x 338 339 mulq %rbx 340 addq %rax,%r10 341 movq -16(%rcx,%r15,8),%rax 342 adcq $0,%rdx 343 movq %rdx,%r11 344 345 mulq %rbp 346 addq %rax,%r13 347 movq -8(%rsi,%r15,8),%rax 348 adcq $0,%rdx 349 addq %r10,%r13 350 adcq $0,%rdx 351 movq %r13,-24(%rsp,%r15,8) 352 movq %rdx,%rdi 353 354 mulq %rbx 355 addq %rax,%r11 356 movq -8(%rcx,%r15,8),%rax 357 adcq $0,%rdx 358 movq %rdx,%r10 359 360 mulq %rbp 361 addq %rax,%rdi 362 movq (%rsi),%rax 363 adcq $0,%rdx 364 addq %r11,%rdi 365 adcq $0,%rdx 366 movq %rdi,-16(%rsp,%r15,8) 367 movq %rdx,%r13 368 369 xorq %rdi,%rdi 370 addq %r10,%r13 371 adcq $0,%rdi 372 movq %r13,-8(%rsp,%r15,8) 373 movq %rdi,(%rsp,%r15,8) 374 375 leaq 1(%r14),%r14 376 .p2align 2 377 L$outer4x: 378 movq (%r12,%r14,8),%rbx 379 xorq %r15,%r15 380 movq (%rsp),%r10 381 movq %r8,%rbp 382 mulq %rbx 383 addq %rax,%r10 384 movq (%rcx),%rax 385 adcq $0,%rdx 386 387 imulq %r10,%rbp 388 movq %rdx,%r11 389 390 mulq %rbp 391 addq %rax,%r10 392 movq 8(%rsi),%rax 393 adcq $0,%rdx 394 movq %rdx,%rdi 395 396 mulq %rbx 397 addq %rax,%r11 398 movq 8(%rcx),%rax 399 adcq $0,%rdx 400 addq 8(%rsp),%r11 401 adcq $0,%rdx 402 movq %rdx,%r10 403 404 mulq %rbp 405 addq %rax,%rdi 406 movq 16(%rsi),%rax 407 adcq $0,%rdx 408 addq %r11,%rdi 409 leaq 4(%r15),%r15 410 adcq $0,%rdx 411 movq %rdi,(%rsp) 412 movq %rdx,%r13 413 jmp L$inner4x 414 .p2align 4 415 L$inner4x: 416 mulq %rbx 417 addq %rax,%r10 418 movq -16(%rcx,%r15,8),%rax 419 adcq $0,%rdx 420 addq -16(%rsp,%r15,8),%r10 421 adcq $0,%rdx 422 movq %rdx,%r11 423 424 mulq %rbp 425 addq %rax,%r13 426 movq -8(%rsi,%r15,8),%rax 427 adcq $0,%rdx 428 addq %r10,%r13 429 adcq $0,%rdx 430 movq %r13,-24(%rsp,%r15,8) 431 movq %rdx,%rdi 432 433 mulq %rbx 434 addq %rax,%r11 435 movq -8(%rcx,%r15,8),%rax 436 adcq $0,%rdx 437 addq -8(%rsp,%r15,8),%r11 438 adcq $0,%rdx 439 movq %rdx,%r10 440 441 mulq %rbp 442 addq %rax,%rdi 443 movq (%rsi,%r15,8),%rax 444 adcq $0,%rdx 445 addq %r11,%rdi 446 adcq $0,%rdx 447 movq %rdi,-16(%rsp,%r15,8) 448 movq %rdx,%r13 449 450 mulq %rbx 451 addq %rax,%r10 452 movq (%rcx,%r15,8),%rax 453 adcq $0,%rdx 454 addq (%rsp,%r15,8),%r10 455 adcq $0,%rdx 456 movq %rdx,%r11 457 458 mulq %rbp 459 addq %rax,%r13 460 movq 8(%rsi,%r15,8),%rax 461 adcq $0,%rdx 462 addq %r10,%r13 463 adcq $0,%rdx 464 movq %r13,-8(%rsp,%r15,8) 465 movq %rdx,%rdi 466 467 mulq %rbx 468 addq %rax,%r11 469 movq 8(%rcx,%r15,8),%rax 470 adcq $0,%rdx 471 addq 8(%rsp,%r15,8),%r11 472 adcq $0,%rdx 473 leaq 4(%r15),%r15 474 movq %rdx,%r10 475 476 mulq %rbp 477 addq %rax,%rdi 478 movq -16(%rsi,%r15,8),%rax 479 adcq $0,%rdx 480 addq %r11,%rdi 481 adcq $0,%rdx 482 movq %rdi,-32(%rsp,%r15,8) 483 movq %rdx,%r13 484 cmpq %r9,%r15 485 jb L$inner4x 486 487 mulq %rbx 488 addq %rax,%r10 489 movq -16(%rcx,%r15,8),%rax 490 adcq $0,%rdx 491 addq -16(%rsp,%r15,8),%r10 492 adcq $0,%rdx 493 movq %rdx,%r11 494 495 mulq %rbp 496 addq %rax,%r13 497 movq -8(%rsi,%r15,8),%rax 498 adcq $0,%rdx 499 addq %r10,%r13 500 adcq $0,%rdx 501 movq %r13,-24(%rsp,%r15,8) 502 movq %rdx,%rdi 503 504 mulq %rbx 505 addq %rax,%r11 506 movq -8(%rcx,%r15,8),%rax 507 adcq $0,%rdx 508 addq -8(%rsp,%r15,8),%r11 509 adcq $0,%rdx 510 leaq 1(%r14),%r14 511 movq %rdx,%r10 512 513 mulq %rbp 514 addq %rax,%rdi 515 movq (%rsi),%rax 516 adcq $0,%rdx 517 addq %r11,%rdi 518 adcq $0,%rdx 519 movq %rdi,-16(%rsp,%r15,8) 520 movq %rdx,%r13 521 522 xorq %rdi,%rdi 523 addq %r10,%r13 524 adcq $0,%rdi 525 addq (%rsp,%r9,8),%r13 526 adcq $0,%rdi 527 movq %r13,-8(%rsp,%r15,8) 528 movq %rdi,(%rsp,%r15,8) 529 530 cmpq %r9,%r14 531 jb L$outer4x 532 movq 16(%rsp,%r9,8),%rdi 533 movq 0(%rsp),%rax 534 movq 8(%rsp),%rdx 535 shrq $2,%r9 536 leaq (%rsp),%rsi 537 xorq %r14,%r14 538 539 subq 0(%rcx),%rax 540 movq 16(%rsi),%rbx 541 movq 24(%rsi),%rbp 542 sbbq 8(%rcx),%rdx 543 leaq -1(%r9),%r15 544 jmp L$sub4x 545 .p2align 4 546 L$sub4x: 547 movq %rax,0(%rdi,%r14,8) 548 movq %rdx,8(%rdi,%r14,8) 549 sbbq 16(%rcx,%r14,8),%rbx 550 movq 32(%rsi,%r14,8),%rax 551 movq 40(%rsi,%r14,8),%rdx 552 sbbq 24(%rcx,%r14,8),%rbp 553 movq %rbx,16(%rdi,%r14,8) 554 movq %rbp,24(%rdi,%r14,8) 555 sbbq 32(%rcx,%r14,8),%rax 556 movq 48(%rsi,%r14,8),%rbx 557 movq 56(%rsi,%r14,8),%rbp 558 sbbq 40(%rcx,%r14,8),%rdx 559 leaq 4(%r14),%r14 560 decq %r15 561 jnz L$sub4x 562 563 movq %rax,0(%rdi,%r14,8) 564 movq 32(%rsi,%r14,8),%rax 565 sbbq 16(%rcx,%r14,8),%rbx 566 movq %rdx,8(%rdi,%r14,8) 567 sbbq 24(%rcx,%r14,8),%rbp 568 movq %rbx,16(%rdi,%r14,8) 569 570 sbbq $0,%rax 571 movq %rax,%xmm0 572 punpcklqdq %xmm0,%xmm0 573 movq %rbp,24(%rdi,%r14,8) 574 xorq %r14,%r14 575 576 movq %r9,%r15 577 pxor %xmm5,%xmm5 578 jmp L$copy4x 579 .p2align 4 580 L$copy4x: 581 movdqu (%rsp,%r14,1),%xmm2 582 movdqu 16(%rsp,%r14,1),%xmm4 583 movdqu (%rdi,%r14,1),%xmm1 584 movdqu 16(%rdi,%r14,1),%xmm3 585 pxor %xmm1,%xmm2 586 pxor %xmm3,%xmm4 587 pand %xmm0,%xmm2 588 pand %xmm0,%xmm4 589 pxor %xmm1,%xmm2 590 pxor %xmm3,%xmm4 591 movdqu %xmm2,(%rdi,%r14,1) 592 movdqu %xmm4,16(%rdi,%r14,1) 593 movdqa %xmm5,(%rsp,%r14,1) 594 movdqa %xmm5,16(%rsp,%r14,1) 595 596 leaq 32(%r14),%r14 597 decq %r15 598 jnz L$copy4x 599 600 shlq $2,%r9 601 movq 8(%rsp,%r9,8),%rsi 602 movq $1,%rax 603 movq (%rsi),%r15 604 movq 8(%rsi),%r14 605 movq 16(%rsi),%r13 606 movq 24(%rsi),%r12 607 movq 32(%rsi),%rbp 608 movq 40(%rsi),%rbx 609 leaq 48(%rsi),%rsp 610 L$mul4x_epilogue: 611 .byte 0xf3,0xc3 612 613 614 615 616 .p2align 5 617 bn_sqr8x_mont: 618 L$sqr8x_enter: 619 movq %rsp,%rax 620 pushq %rbx 621 pushq %rbp 622 pushq %r12 623 pushq %r13 624 pushq %r14 625 pushq %r15 626 627 movl %r9d,%r10d 628 shll $3,%r9d 629 shlq $3+2,%r10 630 negq %r9 631 632 633 634 635 636 637 leaq -64(%rsp,%r9,4),%r11 638 movq (%r8),%r8 639 subq %rsi,%r11 640 andq $4095,%r11 641 cmpq %r11,%r10 642 jb L$sqr8x_sp_alt 643 subq %r11,%rsp 644 leaq -64(%rsp,%r9,4),%rsp 645 jmp L$sqr8x_sp_done 646 647 .p2align 5 648 L$sqr8x_sp_alt: 649 leaq 4096-64(,%r9,4),%r10 650 leaq -64(%rsp,%r9,4),%rsp 651 subq %r10,%r11 652 movq $0,%r10 653 cmovcq %r10,%r11 654 subq %r11,%rsp 655 L$sqr8x_sp_done: 656 andq $-64,%rsp 657 movq %r9,%r10 658 negq %r9 659 660 leaq 64(%rsp,%r9,2),%r11 661 movq %r8,32(%rsp) 662 movq %rax,40(%rsp) 663 L$sqr8x_body: 664 665 movq %r9,%rbp 666 .byte 102,73,15,110,211 667 shrq $3+2,%rbp 668 movl _OPENSSL_ia32cap_P+8(%rip),%eax 669 jmp L$sqr8x_copy_n 670 671 .p2align 5 672 L$sqr8x_copy_n: 673 movq 0(%rcx),%xmm0 674 movq 8(%rcx),%xmm1 675 movq 16(%rcx),%xmm3 676 movq 24(%rcx),%xmm4 677 leaq 32(%rcx),%rcx 678 movdqa %xmm0,0(%r11) 679 movdqa %xmm1,16(%r11) 680 movdqa %xmm3,32(%r11) 681 movdqa %xmm4,48(%r11) 682 leaq 64(%r11),%r11 683 decq %rbp 684 jnz L$sqr8x_copy_n 685 686 pxor %xmm0,%xmm0 687 .byte 102,72,15,110,207 688 .byte 102,73,15,110,218 689 call _bn_sqr8x_internal 690 691 pxor %xmm0,%xmm0 692 leaq 48(%rsp),%rax 693 leaq 64(%rsp,%r9,2),%rdx 694 shrq $3+2,%r9 695 movq 40(%rsp),%rsi 696 jmp L$sqr8x_zero 697 698 .p2align 5 699 L$sqr8x_zero: 700 movdqa %xmm0,0(%rax) 701 movdqa %xmm0,16(%rax) 702 movdqa %xmm0,32(%rax) 703 movdqa %xmm0,48(%rax) 704 leaq 64(%rax),%rax 705 movdqa %xmm0,0(%rdx) 706 movdqa %xmm0,16(%rdx) 707 movdqa %xmm0,32(%rdx) 708 movdqa %xmm0,48(%rdx) 709 leaq 64(%rdx),%rdx 710 decq %r9 711 jnz L$sqr8x_zero 712 713 movq $1,%rax 714 movq -48(%rsi),%r15 715 movq -40(%rsi),%r14 716 movq -32(%rsi),%r13 717 movq -24(%rsi),%r12 718 movq -16(%rsi),%rbp 719 movq -8(%rsi),%rbx 720 leaq (%rsi),%rsp 721 L$sqr8x_epilogue: 722 .byte 0xf3,0xc3 723 724 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 725 .p2align 4 726 #endif 727
