1 #if defined(__x86_64__) 2 .text 3 4 .extern OPENSSL_ia32cap_P 5 .hidden OPENSSL_ia32cap_P 6 7 .globl bn_mul_mont 8 .hidden bn_mul_mont 9 .type bn_mul_mont,@function 10 .align 16 11 bn_mul_mont: 12 testl $3,%r9d 13 jnz .Lmul_enter 14 cmpl $8,%r9d 15 jb .Lmul_enter 16 cmpq %rsi,%rdx 17 jne .Lmul4x_enter 18 testl $7,%r9d 19 jz .Lsqr8x_enter 20 jmp .Lmul4x_enter 21 22 .align 16 23 .Lmul_enter: 24 pushq %rbx 25 pushq %rbp 26 pushq %r12 27 pushq %r13 28 pushq %r14 29 pushq %r15 30 31 movl %r9d,%r9d 32 leaq 2(%r9),%r10 33 movq %rsp,%r11 34 negq %r10 35 leaq (%rsp,%r10,8),%rsp 36 andq $-1024,%rsp 37 38 movq %r11,8(%rsp,%r9,8) 39 .Lmul_body: 40 movq %rdx,%r12 41 movq (%r8),%r8 42 movq (%r12),%rbx 43 movq (%rsi),%rax 44 45 xorq %r14,%r14 46 xorq %r15,%r15 47 48 movq %r8,%rbp 49 mulq %rbx 50 movq %rax,%r10 51 movq (%rcx),%rax 52 53 imulq %r10,%rbp 54 movq %rdx,%r11 55 56 mulq %rbp 57 addq %rax,%r10 58 movq 8(%rsi),%rax 59 adcq $0,%rdx 60 movq %rdx,%r13 61 62 leaq 1(%r15),%r15 63 jmp .L1st_enter 64 65 .align 16 66 .L1st: 67 addq %rax,%r13 68 movq (%rsi,%r15,8),%rax 69 adcq $0,%rdx 70 addq %r11,%r13 71 movq %r10,%r11 72 adcq $0,%rdx 73 movq %r13,-16(%rsp,%r15,8) 74 movq %rdx,%r13 75 76 .L1st_enter: 77 mulq %rbx 78 addq %rax,%r11 79 movq (%rcx,%r15,8),%rax 80 adcq $0,%rdx 81 leaq 1(%r15),%r15 82 movq %rdx,%r10 83 84 mulq %rbp 85 cmpq %r9,%r15 86 jne .L1st 87 88 addq %rax,%r13 89 movq (%rsi),%rax 90 adcq $0,%rdx 91 addq %r11,%r13 92 adcq $0,%rdx 93 movq %r13,-16(%rsp,%r15,8) 94 movq %rdx,%r13 95 movq %r10,%r11 96 97 xorq %rdx,%rdx 98 addq %r11,%r13 99 adcq $0,%rdx 100 movq %r13,-8(%rsp,%r9,8) 101 movq %rdx,(%rsp,%r9,8) 102 103 leaq 1(%r14),%r14 104 jmp .Louter 105 .align 16 106 .Louter: 107 movq (%r12,%r14,8),%rbx 108 xorq %r15,%r15 109 movq %r8,%rbp 110 movq (%rsp),%r10 111 mulq %rbx 112 addq %rax,%r10 113 movq (%rcx),%rax 114 adcq $0,%rdx 115 116 imulq %r10,%rbp 117 movq %rdx,%r11 118 119 mulq %rbp 120 addq %rax,%r10 121 movq 8(%rsi),%rax 122 adcq $0,%rdx 123 movq 8(%rsp),%r10 124 movq %rdx,%r13 125 126 leaq 1(%r15),%r15 127 jmp .Linner_enter 128 129 .align 16 130 .Linner: 131 addq %rax,%r13 132 movq (%rsi,%r15,8),%rax 133 adcq $0,%rdx 134 addq %r10,%r13 135 movq (%rsp,%r15,8),%r10 136 adcq $0,%rdx 137 movq %r13,-16(%rsp,%r15,8) 138 movq %rdx,%r13 139 140 .Linner_enter: 141 mulq %rbx 142 addq %rax,%r11 143 movq (%rcx,%r15,8),%rax 144 adcq $0,%rdx 145 addq %r11,%r10 146 movq %rdx,%r11 147 adcq $0,%r11 148 leaq 1(%r15),%r15 149 150 mulq %rbp 151 cmpq %r9,%r15 152 jne .Linner 153 154 addq %rax,%r13 155 movq (%rsi),%rax 156 adcq $0,%rdx 157 addq %r10,%r13 158 movq (%rsp,%r15,8),%r10 159 adcq $0,%rdx 160 movq %r13,-16(%rsp,%r15,8) 161 movq %rdx,%r13 162 163 xorq %rdx,%rdx 164 addq %r11,%r13 165 adcq $0,%rdx 166 addq %r10,%r13 167 adcq $0,%rdx 168 movq %r13,-8(%rsp,%r9,8) 169 movq %rdx,(%rsp,%r9,8) 170 171 leaq 1(%r14),%r14 172 cmpq %r9,%r14 173 jb .Louter 174 175 xorq %r14,%r14 176 movq (%rsp),%rax 177 leaq (%rsp),%rsi 178 movq %r9,%r15 179 jmp .Lsub 180 .align 16 181 .Lsub: sbbq (%rcx,%r14,8),%rax 182 movq %rax,(%rdi,%r14,8) 183 movq 8(%rsi,%r14,8),%rax 184 leaq 1(%r14),%r14 185 decq %r15 186 jnz .Lsub 187 188 sbbq $0,%rax 189 xorq %r14,%r14 190 movq %r9,%r15 191 .align 16 192 .Lcopy: 193 movq (%rsp,%r14,8),%rsi 194 movq (%rdi,%r14,8),%rcx 195 xorq %rcx,%rsi 196 andq %rax,%rsi 197 xorq %rcx,%rsi 198 movq %r14,(%rsp,%r14,8) 199 movq %rsi,(%rdi,%r14,8) 200 leaq 1(%r14),%r14 201 subq $1,%r15 202 jnz .Lcopy 203 204 movq 8(%rsp,%r9,8),%rsi 205 movq $1,%rax 206 movq (%rsi),%r15 207 movq 8(%rsi),%r14 208 movq 16(%rsi),%r13 209 movq 24(%rsi),%r12 210 movq 32(%rsi),%rbp 211 movq 40(%rsi),%rbx 212 leaq 48(%rsi),%rsp 213 .Lmul_epilogue: 214 .byte 0xf3,0xc3 215 .size bn_mul_mont,.-bn_mul_mont 216 .type bn_mul4x_mont,@function 217 .align 16 218 bn_mul4x_mont: 219 .Lmul4x_enter: 220 pushq %rbx 221 pushq %rbp 222 pushq %r12 223 pushq %r13 224 pushq %r14 225 pushq %r15 226 227 movl %r9d,%r9d 228 leaq 4(%r9),%r10 229 movq %rsp,%r11 230 negq %r10 231 leaq (%rsp,%r10,8),%rsp 232 andq $-1024,%rsp 233 234 movq %r11,8(%rsp,%r9,8) 235 .Lmul4x_body: 236 movq %rdi,16(%rsp,%r9,8) 237 movq %rdx,%r12 238 movq (%r8),%r8 239 movq (%r12),%rbx 240 movq (%rsi),%rax 241 242 xorq %r14,%r14 243 xorq %r15,%r15 244 245 movq %r8,%rbp 246 mulq %rbx 247 movq %rax,%r10 248 movq (%rcx),%rax 249 250 imulq %r10,%rbp 251 movq %rdx,%r11 252 253 mulq %rbp 254 addq %rax,%r10 255 movq 8(%rsi),%rax 256 adcq $0,%rdx 257 movq %rdx,%rdi 258 259 mulq %rbx 260 addq %rax,%r11 261 movq 8(%rcx),%rax 262 adcq $0,%rdx 263 movq %rdx,%r10 264 265 mulq %rbp 266 addq %rax,%rdi 267 movq 16(%rsi),%rax 268 adcq $0,%rdx 269 addq %r11,%rdi 270 leaq 4(%r15),%r15 271 adcq $0,%rdx 272 movq %rdi,(%rsp) 273 movq %rdx,%r13 274 jmp .L1st4x 275 .align 16 276 .L1st4x: 277 mulq %rbx 278 addq %rax,%r10 279 movq -16(%rcx,%r15,8),%rax 280 adcq $0,%rdx 281 movq %rdx,%r11 282 283 mulq %rbp 284 addq %rax,%r13 285 movq -8(%rsi,%r15,8),%rax 286 adcq $0,%rdx 287 addq %r10,%r13 288 adcq $0,%rdx 289 movq %r13,-24(%rsp,%r15,8) 290 movq %rdx,%rdi 291 292 mulq %rbx 293 addq %rax,%r11 294 movq -8(%rcx,%r15,8),%rax 295 adcq $0,%rdx 296 movq %rdx,%r10 297 298 mulq %rbp 299 addq %rax,%rdi 300 movq (%rsi,%r15,8),%rax 301 adcq $0,%rdx 302 addq %r11,%rdi 303 adcq $0,%rdx 304 movq %rdi,-16(%rsp,%r15,8) 305 movq %rdx,%r13 306 307 mulq %rbx 308 addq %rax,%r10 309 movq (%rcx,%r15,8),%rax 310 adcq $0,%rdx 311 movq %rdx,%r11 312 313 mulq %rbp 314 addq %rax,%r13 315 movq 8(%rsi,%r15,8),%rax 316 adcq $0,%rdx 317 addq %r10,%r13 318 adcq $0,%rdx 319 movq %r13,-8(%rsp,%r15,8) 320 movq %rdx,%rdi 321 322 mulq %rbx 323 addq %rax,%r11 324 movq 8(%rcx,%r15,8),%rax 325 adcq $0,%rdx 326 leaq 4(%r15),%r15 327 movq %rdx,%r10 328 329 mulq %rbp 330 addq %rax,%rdi 331 movq -16(%rsi,%r15,8),%rax 332 adcq $0,%rdx 333 addq %r11,%rdi 334 adcq $0,%rdx 335 movq %rdi,-32(%rsp,%r15,8) 336 movq %rdx,%r13 337 cmpq %r9,%r15 338 jb .L1st4x 339 340 mulq %rbx 341 addq %rax,%r10 342 movq -16(%rcx,%r15,8),%rax 343 adcq $0,%rdx 344 movq %rdx,%r11 345 346 mulq %rbp 347 addq %rax,%r13 348 movq -8(%rsi,%r15,8),%rax 349 adcq $0,%rdx 350 addq %r10,%r13 351 adcq $0,%rdx 352 movq %r13,-24(%rsp,%r15,8) 353 movq %rdx,%rdi 354 355 mulq %rbx 356 addq %rax,%r11 357 movq -8(%rcx,%r15,8),%rax 358 adcq $0,%rdx 359 movq %rdx,%r10 360 361 mulq %rbp 362 addq %rax,%rdi 363 movq (%rsi),%rax 364 adcq $0,%rdx 365 addq %r11,%rdi 366 adcq $0,%rdx 367 movq %rdi,-16(%rsp,%r15,8) 368 movq %rdx,%r13 369 370 xorq %rdi,%rdi 371 addq %r10,%r13 372 adcq $0,%rdi 373 movq %r13,-8(%rsp,%r15,8) 374 movq %rdi,(%rsp,%r15,8) 375 376 leaq 1(%r14),%r14 377 .align 4 378 .Louter4x: 379 movq (%r12,%r14,8),%rbx 380 xorq %r15,%r15 381 movq (%rsp),%r10 382 movq %r8,%rbp 383 mulq %rbx 384 addq %rax,%r10 385 movq (%rcx),%rax 386 adcq $0,%rdx 387 388 imulq %r10,%rbp 389 movq %rdx,%r11 390 391 mulq %rbp 392 addq %rax,%r10 393 movq 8(%rsi),%rax 394 adcq $0,%rdx 395 movq %rdx,%rdi 396 397 mulq %rbx 398 addq %rax,%r11 399 movq 8(%rcx),%rax 400 adcq $0,%rdx 401 addq 8(%rsp),%r11 402 adcq $0,%rdx 403 movq %rdx,%r10 404 405 mulq %rbp 406 addq %rax,%rdi 407 movq 16(%rsi),%rax 408 adcq $0,%rdx 409 addq %r11,%rdi 410 leaq 4(%r15),%r15 411 adcq $0,%rdx 412 movq %rdi,(%rsp) 413 movq %rdx,%r13 414 jmp .Linner4x 415 .align 16 416 .Linner4x: 417 mulq %rbx 418 addq %rax,%r10 419 movq -16(%rcx,%r15,8),%rax 420 adcq $0,%rdx 421 addq -16(%rsp,%r15,8),%r10 422 adcq $0,%rdx 423 movq %rdx,%r11 424 425 mulq %rbp 426 addq %rax,%r13 427 movq -8(%rsi,%r15,8),%rax 428 adcq $0,%rdx 429 addq %r10,%r13 430 adcq $0,%rdx 431 movq %r13,-24(%rsp,%r15,8) 432 movq %rdx,%rdi 433 434 mulq %rbx 435 addq %rax,%r11 436 movq -8(%rcx,%r15,8),%rax 437 adcq $0,%rdx 438 addq -8(%rsp,%r15,8),%r11 439 adcq $0,%rdx 440 movq %rdx,%r10 441 442 mulq %rbp 443 addq %rax,%rdi 444 movq (%rsi,%r15,8),%rax 445 adcq $0,%rdx 446 addq %r11,%rdi 447 adcq $0,%rdx 448 movq %rdi,-16(%rsp,%r15,8) 449 movq %rdx,%r13 450 451 mulq %rbx 452 addq %rax,%r10 453 movq (%rcx,%r15,8),%rax 454 adcq $0,%rdx 455 addq (%rsp,%r15,8),%r10 456 adcq $0,%rdx 457 movq %rdx,%r11 458 459 mulq %rbp 460 addq %rax,%r13 461 movq 8(%rsi,%r15,8),%rax 462 adcq $0,%rdx 463 addq %r10,%r13 464 adcq $0,%rdx 465 movq %r13,-8(%rsp,%r15,8) 466 movq %rdx,%rdi 467 468 mulq %rbx 469 addq %rax,%r11 470 movq 8(%rcx,%r15,8),%rax 471 adcq $0,%rdx 472 addq 8(%rsp,%r15,8),%r11 473 adcq $0,%rdx 474 leaq 4(%r15),%r15 475 movq %rdx,%r10 476 477 mulq %rbp 478 addq %rax,%rdi 479 movq -16(%rsi,%r15,8),%rax 480 adcq $0,%rdx 481 addq %r11,%rdi 482 adcq $0,%rdx 483 movq %rdi,-32(%rsp,%r15,8) 484 movq %rdx,%r13 485 cmpq %r9,%r15 486 jb .Linner4x 487 488 mulq %rbx 489 addq %rax,%r10 490 movq -16(%rcx,%r15,8),%rax 491 adcq $0,%rdx 492 addq -16(%rsp,%r15,8),%r10 493 adcq $0,%rdx 494 movq %rdx,%r11 495 496 mulq %rbp 497 addq %rax,%r13 498 movq -8(%rsi,%r15,8),%rax 499 adcq $0,%rdx 500 addq %r10,%r13 501 adcq $0,%rdx 502 movq %r13,-24(%rsp,%r15,8) 503 movq %rdx,%rdi 504 505 mulq %rbx 506 addq %rax,%r11 507 movq -8(%rcx,%r15,8),%rax 508 adcq $0,%rdx 509 addq -8(%rsp,%r15,8),%r11 510 adcq $0,%rdx 511 leaq 1(%r14),%r14 512 movq %rdx,%r10 513 514 mulq %rbp 515 addq %rax,%rdi 516 movq (%rsi),%rax 517 adcq $0,%rdx 518 addq %r11,%rdi 519 adcq $0,%rdx 520 movq %rdi,-16(%rsp,%r15,8) 521 movq %rdx,%r13 522 523 xorq %rdi,%rdi 524 addq %r10,%r13 525 adcq $0,%rdi 526 addq (%rsp,%r9,8),%r13 527 adcq $0,%rdi 528 movq %r13,-8(%rsp,%r15,8) 529 movq %rdi,(%rsp,%r15,8) 530 531 cmpq %r9,%r14 532 jb .Louter4x 533 movq 16(%rsp,%r9,8),%rdi 534 movq 0(%rsp),%rax 535 movq 8(%rsp),%rdx 536 shrq $2,%r9 537 leaq (%rsp),%rsi 538 xorq %r14,%r14 539 540 subq 0(%rcx),%rax 541 movq 16(%rsi),%rbx 542 movq 24(%rsi),%rbp 543 sbbq 8(%rcx),%rdx 544 leaq -1(%r9),%r15 545 jmp .Lsub4x 546 .align 16 547 .Lsub4x: 548 movq %rax,0(%rdi,%r14,8) 549 movq %rdx,8(%rdi,%r14,8) 550 sbbq 16(%rcx,%r14,8),%rbx 551 movq 32(%rsi,%r14,8),%rax 552 movq 40(%rsi,%r14,8),%rdx 553 sbbq 24(%rcx,%r14,8),%rbp 554 movq %rbx,16(%rdi,%r14,8) 555 movq %rbp,24(%rdi,%r14,8) 556 sbbq 32(%rcx,%r14,8),%rax 557 movq 48(%rsi,%r14,8),%rbx 558 movq 56(%rsi,%r14,8),%rbp 559 sbbq 40(%rcx,%r14,8),%rdx 560 leaq 4(%r14),%r14 561 decq %r15 562 jnz .Lsub4x 563 564 movq %rax,0(%rdi,%r14,8) 565 movq 32(%rsi,%r14,8),%rax 566 sbbq 16(%rcx,%r14,8),%rbx 567 movq %rdx,8(%rdi,%r14,8) 568 sbbq 24(%rcx,%r14,8),%rbp 569 movq %rbx,16(%rdi,%r14,8) 570 571 sbbq $0,%rax 572 movq %rax,%xmm0 573 punpcklqdq %xmm0,%xmm0 574 movq %rbp,24(%rdi,%r14,8) 575 xorq %r14,%r14 576 577 movq %r9,%r15 578 pxor %xmm5,%xmm5 579 jmp .Lcopy4x 580 .align 16 581 .Lcopy4x: 582 movdqu (%rsp,%r14,1),%xmm2 583 movdqu 16(%rsp,%r14,1),%xmm4 584 movdqu (%rdi,%r14,1),%xmm1 585 movdqu 16(%rdi,%r14,1),%xmm3 586 pxor %xmm1,%xmm2 587 pxor %xmm3,%xmm4 588 pand %xmm0,%xmm2 589 pand %xmm0,%xmm4 590 pxor %xmm1,%xmm2 591 pxor %xmm3,%xmm4 592 movdqu %xmm2,(%rdi,%r14,1) 593 movdqu %xmm4,16(%rdi,%r14,1) 594 movdqa %xmm5,(%rsp,%r14,1) 595 movdqa %xmm5,16(%rsp,%r14,1) 596 597 leaq 32(%r14),%r14 598 decq %r15 599 jnz .Lcopy4x 600 601 shlq $2,%r9 602 movq 8(%rsp,%r9,8),%rsi 603 movq $1,%rax 604 movq (%rsi),%r15 605 movq 8(%rsi),%r14 606 movq 16(%rsi),%r13 607 movq 24(%rsi),%r12 608 movq 32(%rsi),%rbp 609 movq 40(%rsi),%rbx 610 leaq 48(%rsi),%rsp 611 .Lmul4x_epilogue: 612 .byte 0xf3,0xc3 613 .size bn_mul4x_mont,.-bn_mul4x_mont 614 .extern bn_sqr8x_internal 615 .hidden bn_sqr8x_internal 616 617 .type bn_sqr8x_mont,@function 618 .align 32 619 bn_sqr8x_mont: 620 .Lsqr8x_enter: 621 movq %rsp,%rax 622 pushq %rbx 623 pushq %rbp 624 pushq %r12 625 pushq %r13 626 pushq %r14 627 pushq %r15 628 629 movl %r9d,%r10d 630 shll $3,%r9d 631 shlq $3+2,%r10 632 negq %r9 633 634 635 636 637 638 639 leaq -64(%rsp,%r9,4),%r11 640 movq (%r8),%r8 641 subq %rsi,%r11 642 andq $4095,%r11 643 cmpq %r11,%r10 644 jb .Lsqr8x_sp_alt 645 subq %r11,%rsp 646 leaq -64(%rsp,%r9,4),%rsp 647 jmp .Lsqr8x_sp_done 648 649 .align 32 650 .Lsqr8x_sp_alt: 651 leaq 4096-64(,%r9,4),%r10 652 leaq -64(%rsp,%r9,4),%rsp 653 subq %r10,%r11 654 movq $0,%r10 655 cmovcq %r10,%r11 656 subq %r11,%rsp 657 .Lsqr8x_sp_done: 658 andq $-64,%rsp 659 movq %r9,%r10 660 negq %r9 661 662 leaq 64(%rsp,%r9,2),%r11 663 movq %r8,32(%rsp) 664 movq %rax,40(%rsp) 665 .Lsqr8x_body: 666 667 movq %r9,%rbp 668 .byte 102,73,15,110,211 669 shrq $3+2,%rbp 670 movl OPENSSL_ia32cap_P+8(%rip),%eax 671 jmp .Lsqr8x_copy_n 672 673 .align 32 674 .Lsqr8x_copy_n: 675 movq 0(%rcx),%xmm0 676 movq 8(%rcx),%xmm1 677 movq 16(%rcx),%xmm3 678 movq 24(%rcx),%xmm4 679 leaq 32(%rcx),%rcx 680 movdqa %xmm0,0(%r11) 681 movdqa %xmm1,16(%r11) 682 movdqa %xmm3,32(%r11) 683 movdqa %xmm4,48(%r11) 684 leaq 64(%r11),%r11 685 decq %rbp 686 jnz .Lsqr8x_copy_n 687 688 pxor %xmm0,%xmm0 689 .byte 102,72,15,110,207 690 .byte 102,73,15,110,218 691 call bn_sqr8x_internal 692 693 pxor %xmm0,%xmm0 694 leaq 48(%rsp),%rax 695 leaq 64(%rsp,%r9,2),%rdx 696 shrq $3+2,%r9 697 movq 40(%rsp),%rsi 698 jmp .Lsqr8x_zero 699 700 .align 32 701 .Lsqr8x_zero: 702 movdqa %xmm0,0(%rax) 703 movdqa %xmm0,16(%rax) 704 movdqa %xmm0,32(%rax) 705 movdqa %xmm0,48(%rax) 706 leaq 64(%rax),%rax 707 movdqa %xmm0,0(%rdx) 708 movdqa %xmm0,16(%rdx) 709 movdqa %xmm0,32(%rdx) 710 movdqa %xmm0,48(%rdx) 711 leaq 64(%rdx),%rdx 712 decq %r9 713 jnz .Lsqr8x_zero 714 715 movq $1,%rax 716 movq -48(%rsi),%r15 717 movq -40(%rsi),%r14 718 movq -32(%rsi),%r13 719 movq -24(%rsi),%r12 720 movq -16(%rsi),%rbp 721 movq -8(%rsi),%rbx 722 leaq (%rsi),%rsp 723 .Lsqr8x_epilogue: 724 .byte 0xf3,0xc3 725 .size bn_sqr8x_mont,.-bn_sqr8x_mont 726 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 727 .align 16 728 #endif 729
