bn-586.S revision 8ff035535f7cf2903f02bbe94d2fa10b7ab855f1
1#if defined(__i386__) 2.text 3.globl _bn_mul_add_words 4.private_extern _bn_mul_add_words 5.align 4 6_bn_mul_add_words: 7L_bn_mul_add_words_begin: 8 call L000PIC_me_up 9L000PIC_me_up: 10 popl %eax 11 movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L000PIC_me_up(%eax),%eax 12 btl $26,(%eax) 13 jnc L001maw_non_sse2 14 movl 4(%esp),%eax 15 movl 8(%esp),%edx 16 movl 12(%esp),%ecx 17 movd 16(%esp),%mm0 18 pxor %mm1,%mm1 19 jmp L002maw_sse2_entry 20.align 4,0x90 21L003maw_sse2_unrolled: 22 movd (%eax),%mm3 23 paddq %mm3,%mm1 24 movd (%edx),%mm2 25 pmuludq %mm0,%mm2 26 movd 4(%edx),%mm4 27 pmuludq %mm0,%mm4 28 movd 8(%edx),%mm6 29 pmuludq %mm0,%mm6 30 movd 12(%edx),%mm7 31 pmuludq %mm0,%mm7 32 paddq %mm2,%mm1 33 movd 4(%eax),%mm3 34 paddq %mm4,%mm3 35 movd 8(%eax),%mm5 36 paddq %mm6,%mm5 37 movd 12(%eax),%mm4 38 paddq %mm4,%mm7 39 movd %mm1,(%eax) 40 movd 16(%edx),%mm2 41 pmuludq %mm0,%mm2 42 psrlq $32,%mm1 43 movd 20(%edx),%mm4 44 pmuludq %mm0,%mm4 45 paddq %mm3,%mm1 46 movd 24(%edx),%mm6 47 pmuludq %mm0,%mm6 48 movd %mm1,4(%eax) 49 psrlq $32,%mm1 50 movd 28(%edx),%mm3 51 addl $32,%edx 52 pmuludq %mm0,%mm3 53 paddq %mm5,%mm1 54 movd 16(%eax),%mm5 55 paddq %mm5,%mm2 56 movd %mm1,8(%eax) 57 psrlq $32,%mm1 58 paddq %mm7,%mm1 59 movd 20(%eax),%mm5 60 paddq %mm5,%mm4 61 movd %mm1,12(%eax) 62 psrlq $32,%mm1 63 paddq %mm2,%mm1 64 movd 24(%eax),%mm5 65 paddq %mm5,%mm6 66 movd %mm1,16(%eax) 67 psrlq $32,%mm1 68 paddq %mm4,%mm1 69 movd 28(%eax),%mm5 70 paddq %mm5,%mm3 71 movd %mm1,20(%eax) 72 psrlq $32,%mm1 73 paddq %mm6,%mm1 74 movd %mm1,24(%eax) 75 psrlq $32,%mm1 76 paddq %mm3,%mm1 77 movd %mm1,28(%eax) 78 leal 32(%eax),%eax 79 psrlq $32,%mm1 80 subl $8,%ecx 81 jz L004maw_sse2_exit 82L002maw_sse2_entry: 83 testl $4294967288,%ecx 84 jnz L003maw_sse2_unrolled 85.align 2,0x90 86L005maw_sse2_loop: 87 movd (%edx),%mm2 88 movd (%eax),%mm3 89 pmuludq %mm0,%mm2 90 leal 4(%edx),%edx 91 paddq %mm3,%mm1 92 paddq %mm2,%mm1 93 movd %mm1,(%eax) 94 subl $1,%ecx 95 psrlq $32,%mm1 96 leal 4(%eax),%eax 97 jnz L005maw_sse2_loop 98L004maw_sse2_exit: 99 movd %mm1,%eax 100 emms 101 ret 102.align 4,0x90 103L001maw_non_sse2: 104 pushl %ebp 105 pushl %ebx 106 pushl %esi 107 pushl %edi 108 109 xorl %esi,%esi 110 movl 20(%esp),%edi 111 movl 28(%esp),%ecx 112 movl 24(%esp),%ebx 113 andl $4294967288,%ecx 114 movl 32(%esp),%ebp 115 pushl %ecx 116 jz L006maw_finish 117.align 4,0x90 118L007maw_loop: 119 # Round 0 120 movl (%ebx),%eax 121 mull %ebp 122 addl %esi,%eax 123 adcl $0,%edx 124 addl (%edi),%eax 125 adcl $0,%edx 126 movl %eax,(%edi) 127 movl %edx,%esi 128 # Round 4 129 movl 4(%ebx),%eax 130 mull %ebp 131 addl %esi,%eax 132 adcl $0,%edx 133 addl 4(%edi),%eax 134 adcl $0,%edx 135 movl %eax,4(%edi) 136 movl %edx,%esi 137 # Round 8 138 movl 8(%ebx),%eax 139 mull %ebp 140 addl %esi,%eax 141 adcl $0,%edx 142 addl 8(%edi),%eax 143 adcl $0,%edx 144 movl %eax,8(%edi) 145 movl %edx,%esi 146 # Round 12 147 movl 12(%ebx),%eax 148 mull %ebp 149 addl %esi,%eax 150 adcl $0,%edx 151 addl 12(%edi),%eax 152 adcl $0,%edx 153 movl %eax,12(%edi) 154 movl %edx,%esi 155 # Round 16 156 movl 16(%ebx),%eax 157 mull %ebp 158 addl %esi,%eax 159 adcl $0,%edx 160 addl 16(%edi),%eax 161 adcl $0,%edx 162 movl %eax,16(%edi) 163 movl %edx,%esi 164 # Round 20 165 movl 20(%ebx),%eax 166 mull %ebp 167 addl %esi,%eax 168 adcl $0,%edx 169 addl 20(%edi),%eax 170 adcl $0,%edx 171 movl %eax,20(%edi) 172 movl %edx,%esi 173 # Round 24 174 movl 24(%ebx),%eax 175 mull %ebp 176 addl %esi,%eax 177 adcl $0,%edx 178 addl 24(%edi),%eax 179 adcl $0,%edx 180 movl %eax,24(%edi) 181 movl %edx,%esi 182 # Round 28 183 movl 28(%ebx),%eax 184 mull %ebp 185 addl %esi,%eax 186 adcl $0,%edx 187 addl 28(%edi),%eax 188 adcl $0,%edx 189 movl %eax,28(%edi) 190 movl %edx,%esi 191 192 subl $8,%ecx 193 leal 32(%ebx),%ebx 194 leal 32(%edi),%edi 195 jnz L007maw_loop 196L006maw_finish: 197 movl 32(%esp),%ecx 198 andl $7,%ecx 199 jnz L008maw_finish2 200 jmp L009maw_end 201L008maw_finish2: 202 # Tail Round 0 203 movl (%ebx),%eax 204 mull %ebp 205 addl %esi,%eax 206 adcl $0,%edx 207 addl (%edi),%eax 208 adcl $0,%edx 209 decl %ecx 210 movl %eax,(%edi) 211 movl %edx,%esi 212 jz L009maw_end 213 # Tail Round 1 214 movl 4(%ebx),%eax 215 mull %ebp 216 addl %esi,%eax 217 adcl $0,%edx 218 addl 4(%edi),%eax 219 adcl $0,%edx 220 decl %ecx 221 movl %eax,4(%edi) 222 movl %edx,%esi 223 jz L009maw_end 224 # Tail Round 2 225 movl 8(%ebx),%eax 226 mull %ebp 227 addl %esi,%eax 228 adcl $0,%edx 229 addl 8(%edi),%eax 230 adcl $0,%edx 231 decl %ecx 232 movl %eax,8(%edi) 233 movl %edx,%esi 234 jz L009maw_end 235 # Tail Round 3 236 movl 12(%ebx),%eax 237 mull %ebp 238 addl %esi,%eax 239 adcl $0,%edx 240 addl 12(%edi),%eax 241 adcl $0,%edx 242 decl %ecx 243 movl %eax,12(%edi) 244 movl %edx,%esi 245 jz L009maw_end 246 # Tail Round 4 247 movl 16(%ebx),%eax 248 mull %ebp 249 addl %esi,%eax 250 adcl $0,%edx 251 addl 16(%edi),%eax 252 adcl $0,%edx 253 decl %ecx 254 movl %eax,16(%edi) 255 movl %edx,%esi 256 jz L009maw_end 257 # Tail Round 5 258 movl 20(%ebx),%eax 259 mull %ebp 260 addl %esi,%eax 261 adcl $0,%edx 262 addl 20(%edi),%eax 263 adcl $0,%edx 264 decl %ecx 265 movl %eax,20(%edi) 266 movl %edx,%esi 267 jz L009maw_end 268 # Tail Round 6 269 movl 24(%ebx),%eax 270 mull %ebp 271 addl %esi,%eax 272 adcl $0,%edx 273 addl 24(%edi),%eax 274 adcl $0,%edx 275 movl %eax,24(%edi) 276 movl %edx,%esi 277L009maw_end: 278 movl %esi,%eax 279 popl %ecx 280 popl %edi 281 popl %esi 282 popl %ebx 283 popl %ebp 284 ret 285.globl _bn_mul_words 286.private_extern _bn_mul_words 287.align 4 288_bn_mul_words: 289L_bn_mul_words_begin: 290 call L010PIC_me_up 291L010PIC_me_up: 292 popl %eax 293 movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L010PIC_me_up(%eax),%eax 294 btl $26,(%eax) 295 jnc L011mw_non_sse2 296 movl 4(%esp),%eax 297 movl 8(%esp),%edx 298 movl 12(%esp),%ecx 299 movd 16(%esp),%mm0 300 pxor %mm1,%mm1 301.align 4,0x90 302L012mw_sse2_loop: 303 movd (%edx),%mm2 304 pmuludq %mm0,%mm2 305 leal 4(%edx),%edx 306 paddq %mm2,%mm1 307 movd %mm1,(%eax) 308 subl $1,%ecx 309 psrlq $32,%mm1 310 leal 4(%eax),%eax 311 jnz L012mw_sse2_loop 312 movd %mm1,%eax 313 emms 314 ret 315.align 4,0x90 316L011mw_non_sse2: 317 pushl %ebp 318 pushl %ebx 319 pushl %esi 320 pushl %edi 321 322 xorl %esi,%esi 323 movl 20(%esp),%edi 324 movl 24(%esp),%ebx 325 movl 28(%esp),%ebp 326 movl 32(%esp),%ecx 327 andl $4294967288,%ebp 328 jz L013mw_finish 329L014mw_loop: 330 # Round 0 331 movl (%ebx),%eax 332 mull %ecx 333 addl %esi,%eax 334 adcl $0,%edx 335 movl %eax,(%edi) 336 movl %edx,%esi 337 # Round 4 338 movl 4(%ebx),%eax 339 mull %ecx 340 addl %esi,%eax 341 adcl $0,%edx 342 movl %eax,4(%edi) 343 movl %edx,%esi 344 # Round 8 345 movl 8(%ebx),%eax 346 mull %ecx 347 addl %esi,%eax 348 adcl $0,%edx 349 movl %eax,8(%edi) 350 movl %edx,%esi 351 # Round 12 352 movl 12(%ebx),%eax 353 mull %ecx 354 addl %esi,%eax 355 adcl $0,%edx 356 movl %eax,12(%edi) 357 movl %edx,%esi 358 # Round 16 359 movl 16(%ebx),%eax 360 mull %ecx 361 addl %esi,%eax 362 adcl $0,%edx 363 movl %eax,16(%edi) 364 movl %edx,%esi 365 # Round 20 366 movl 20(%ebx),%eax 367 mull %ecx 368 addl %esi,%eax 369 adcl $0,%edx 370 movl %eax,20(%edi) 371 movl %edx,%esi 372 # Round 24 373 movl 24(%ebx),%eax 374 mull %ecx 375 addl %esi,%eax 376 adcl $0,%edx 377 movl %eax,24(%edi) 378 movl %edx,%esi 379 # Round 28 380 movl 28(%ebx),%eax 381 mull %ecx 382 addl %esi,%eax 383 adcl $0,%edx 384 movl %eax,28(%edi) 385 movl %edx,%esi 386 387 addl $32,%ebx 388 addl $32,%edi 389 subl $8,%ebp 390 jz L013mw_finish 391 jmp L014mw_loop 392L013mw_finish: 393 movl 28(%esp),%ebp 394 andl $7,%ebp 395 jnz L015mw_finish2 396 jmp L016mw_end 397L015mw_finish2: 398 # Tail Round 0 399 movl (%ebx),%eax 400 mull %ecx 401 addl %esi,%eax 402 adcl $0,%edx 403 movl %eax,(%edi) 404 movl %edx,%esi 405 decl %ebp 406 jz L016mw_end 407 # Tail Round 1 408 movl 4(%ebx),%eax 409 mull %ecx 410 addl %esi,%eax 411 adcl $0,%edx 412 movl %eax,4(%edi) 413 movl %edx,%esi 414 decl %ebp 415 jz L016mw_end 416 # Tail Round 2 417 movl 8(%ebx),%eax 418 mull %ecx 419 addl %esi,%eax 420 adcl $0,%edx 421 movl %eax,8(%edi) 422 movl %edx,%esi 423 decl %ebp 424 jz L016mw_end 425 # Tail Round 3 426 movl 12(%ebx),%eax 427 mull %ecx 428 addl %esi,%eax 429 adcl $0,%edx 430 movl %eax,12(%edi) 431 movl %edx,%esi 432 decl %ebp 433 jz L016mw_end 434 # Tail Round 4 435 movl 16(%ebx),%eax 436 mull %ecx 437 addl %esi,%eax 438 adcl $0,%edx 439 movl %eax,16(%edi) 440 movl %edx,%esi 441 decl %ebp 442 jz L016mw_end 443 # Tail Round 5 444 movl 20(%ebx),%eax 445 mull %ecx 446 addl %esi,%eax 447 adcl $0,%edx 448 movl %eax,20(%edi) 449 movl %edx,%esi 450 decl %ebp 451 jz L016mw_end 452 # Tail Round 6 453 movl 24(%ebx),%eax 454 mull %ecx 455 addl %esi,%eax 456 adcl $0,%edx 457 movl %eax,24(%edi) 458 movl %edx,%esi 459L016mw_end: 460 movl %esi,%eax 461 popl %edi 462 popl %esi 463 popl %ebx 464 popl %ebp 465 ret 466.globl _bn_sqr_words 467.private_extern _bn_sqr_words 468.align 4 469_bn_sqr_words: 470L_bn_sqr_words_begin: 471 call L017PIC_me_up 472L017PIC_me_up: 473 popl %eax 474 movl L_OPENSSL_ia32cap_P$non_lazy_ptr-L017PIC_me_up(%eax),%eax 475 btl $26,(%eax) 476 jnc L018sqr_non_sse2 477 movl 4(%esp),%eax 478 movl 8(%esp),%edx 479 movl 12(%esp),%ecx 480.align 4,0x90 481L019sqr_sse2_loop: 482 movd (%edx),%mm0 483 pmuludq %mm0,%mm0 484 leal 4(%edx),%edx 485 movq %mm0,(%eax) 486 subl $1,%ecx 487 leal 8(%eax),%eax 488 jnz L019sqr_sse2_loop 489 emms 490 ret 491.align 4,0x90 492L018sqr_non_sse2: 493 pushl %ebp 494 pushl %ebx 495 pushl %esi 496 pushl %edi 497 498 movl 20(%esp),%esi 499 movl 24(%esp),%edi 500 movl 28(%esp),%ebx 501 andl $4294967288,%ebx 502 jz L020sw_finish 503L021sw_loop: 504 # Round 0 505 movl (%edi),%eax 506 mull %eax 507 movl %eax,(%esi) 508 movl %edx,4(%esi) 509 # Round 4 510 movl 4(%edi),%eax 511 mull %eax 512 movl %eax,8(%esi) 513 movl %edx,12(%esi) 514 # Round 8 515 movl 8(%edi),%eax 516 mull %eax 517 movl %eax,16(%esi) 518 movl %edx,20(%esi) 519 # Round 12 520 movl 12(%edi),%eax 521 mull %eax 522 movl %eax,24(%esi) 523 movl %edx,28(%esi) 524 # Round 16 525 movl 16(%edi),%eax 526 mull %eax 527 movl %eax,32(%esi) 528 movl %edx,36(%esi) 529 # Round 20 530 movl 20(%edi),%eax 531 mull %eax 532 movl %eax,40(%esi) 533 movl %edx,44(%esi) 534 # Round 24 535 movl 24(%edi),%eax 536 mull %eax 537 movl %eax,48(%esi) 538 movl %edx,52(%esi) 539 # Round 28 540 movl 28(%edi),%eax 541 mull %eax 542 movl %eax,56(%esi) 543 movl %edx,60(%esi) 544 545 addl $32,%edi 546 addl $64,%esi 547 subl $8,%ebx 548 jnz L021sw_loop 549L020sw_finish: 550 movl 28(%esp),%ebx 551 andl $7,%ebx 552 jz L022sw_end 553 # Tail Round 0 554 movl (%edi),%eax 555 mull %eax 556 movl %eax,(%esi) 557 decl %ebx 558 movl %edx,4(%esi) 559 jz L022sw_end 560 # Tail Round 1 561 movl 4(%edi),%eax 562 mull %eax 563 movl %eax,8(%esi) 564 decl %ebx 565 movl %edx,12(%esi) 566 jz L022sw_end 567 # Tail Round 2 568 movl 8(%edi),%eax 569 mull %eax 570 movl %eax,16(%esi) 571 decl %ebx 572 movl %edx,20(%esi) 573 jz L022sw_end 574 # Tail Round 3 575 movl 12(%edi),%eax 576 mull %eax 577 movl %eax,24(%esi) 578 decl %ebx 579 movl %edx,28(%esi) 580 jz L022sw_end 581 # Tail Round 4 582 movl 16(%edi),%eax 583 mull %eax 584 movl %eax,32(%esi) 585 decl %ebx 586 movl %edx,36(%esi) 587 jz L022sw_end 588 # Tail Round 5 589 movl 20(%edi),%eax 590 mull %eax 591 movl %eax,40(%esi) 592 decl %ebx 593 movl %edx,44(%esi) 594 jz L022sw_end 595 # Tail Round 6 596 movl 24(%edi),%eax 597 mull %eax 598 movl %eax,48(%esi) 599 movl %edx,52(%esi) 600L022sw_end: 601 popl %edi 602 popl %esi 603 popl %ebx 604 popl %ebp 605 ret 606.globl _bn_div_words 607.private_extern _bn_div_words 608.align 4 609_bn_div_words: 610L_bn_div_words_begin: 611 movl 4(%esp),%edx 612 movl 8(%esp),%eax 613 movl 12(%esp),%ecx 614 divl %ecx 615 ret 616.globl _bn_add_words 617.private_extern _bn_add_words 618.align 4 619_bn_add_words: 620L_bn_add_words_begin: 621 pushl %ebp 622 pushl %ebx 623 pushl %esi 624 pushl %edi 625 626 movl 20(%esp),%ebx 627 movl 24(%esp),%esi 628 movl 28(%esp),%edi 629 movl 32(%esp),%ebp 630 xorl %eax,%eax 631 andl $4294967288,%ebp 632 jz L023aw_finish 633L024aw_loop: 634 # Round 0 635 movl (%esi),%ecx 636 movl (%edi),%edx 637 addl %eax,%ecx 638 movl $0,%eax 639 adcl %eax,%eax 640 addl %edx,%ecx 641 adcl $0,%eax 642 movl %ecx,(%ebx) 643 # Round 1 644 movl 4(%esi),%ecx 645 movl 4(%edi),%edx 646 addl %eax,%ecx 647 movl $0,%eax 648 adcl %eax,%eax 649 addl %edx,%ecx 650 adcl $0,%eax 651 movl %ecx,4(%ebx) 652 # Round 2 653 movl 8(%esi),%ecx 654 movl 8(%edi),%edx 655 addl %eax,%ecx 656 movl $0,%eax 657 adcl %eax,%eax 658 addl %edx,%ecx 659 adcl $0,%eax 660 movl %ecx,8(%ebx) 661 # Round 3 662 movl 12(%esi),%ecx 663 movl 12(%edi),%edx 664 addl %eax,%ecx 665 movl $0,%eax 666 adcl %eax,%eax 667 addl %edx,%ecx 668 adcl $0,%eax 669 movl %ecx,12(%ebx) 670 # Round 4 671 movl 16(%esi),%ecx 672 movl 16(%edi),%edx 673 addl %eax,%ecx 674 movl $0,%eax 675 adcl %eax,%eax 676 addl %edx,%ecx 677 adcl $0,%eax 678 movl %ecx,16(%ebx) 679 # Round 5 680 movl 20(%esi),%ecx 681 movl 20(%edi),%edx 682 addl %eax,%ecx 683 movl $0,%eax 684 adcl %eax,%eax 685 addl %edx,%ecx 686 adcl $0,%eax 687 movl %ecx,20(%ebx) 688 # Round 6 689 movl 24(%esi),%ecx 690 movl 24(%edi),%edx 691 addl %eax,%ecx 692 movl $0,%eax 693 adcl %eax,%eax 694 addl %edx,%ecx 695 adcl $0,%eax 696 movl %ecx,24(%ebx) 697 # Round 7 698 movl 28(%esi),%ecx 699 movl 28(%edi),%edx 700 addl %eax,%ecx 701 movl $0,%eax 702 adcl %eax,%eax 703 addl %edx,%ecx 704 adcl $0,%eax 705 movl %ecx,28(%ebx) 706 707 addl $32,%esi 708 addl $32,%edi 709 addl $32,%ebx 710 subl $8,%ebp 711 jnz L024aw_loop 712L023aw_finish: 713 movl 32(%esp),%ebp 714 andl $7,%ebp 715 jz L025aw_end 716 # Tail Round 0 717 movl (%esi),%ecx 718 movl (%edi),%edx 719 addl %eax,%ecx 720 movl $0,%eax 721 adcl %eax,%eax 722 addl %edx,%ecx 723 adcl $0,%eax 724 decl %ebp 725 movl %ecx,(%ebx) 726 jz L025aw_end 727 # Tail Round 1 728 movl 4(%esi),%ecx 729 movl 4(%edi),%edx 730 addl %eax,%ecx 731 movl $0,%eax 732 adcl %eax,%eax 733 addl %edx,%ecx 734 adcl $0,%eax 735 decl %ebp 736 movl %ecx,4(%ebx) 737 jz L025aw_end 738 # Tail Round 2 739 movl 8(%esi),%ecx 740 movl 8(%edi),%edx 741 addl %eax,%ecx 742 movl $0,%eax 743 adcl %eax,%eax 744 addl %edx,%ecx 745 adcl $0,%eax 746 decl %ebp 747 movl %ecx,8(%ebx) 748 jz L025aw_end 749 # Tail Round 3 750 movl 12(%esi),%ecx 751 movl 12(%edi),%edx 752 addl %eax,%ecx 753 movl $0,%eax 754 adcl %eax,%eax 755 addl %edx,%ecx 756 adcl $0,%eax 757 decl %ebp 758 movl %ecx,12(%ebx) 759 jz L025aw_end 760 # Tail Round 4 761 movl 16(%esi),%ecx 762 movl 16(%edi),%edx 763 addl %eax,%ecx 764 movl $0,%eax 765 adcl %eax,%eax 766 addl %edx,%ecx 767 adcl $0,%eax 768 decl %ebp 769 movl %ecx,16(%ebx) 770 jz L025aw_end 771 # Tail Round 5 772 movl 20(%esi),%ecx 773 movl 20(%edi),%edx 774 addl %eax,%ecx 775 movl $0,%eax 776 adcl %eax,%eax 777 addl %edx,%ecx 778 adcl $0,%eax 779 decl %ebp 780 movl %ecx,20(%ebx) 781 jz L025aw_end 782 # Tail Round 6 783 movl 24(%esi),%ecx 784 movl 24(%edi),%edx 785 addl %eax,%ecx 786 movl $0,%eax 787 adcl %eax,%eax 788 addl %edx,%ecx 789 adcl $0,%eax 790 movl %ecx,24(%ebx) 791L025aw_end: 792 popl %edi 793 popl %esi 794 popl %ebx 795 popl %ebp 796 ret 797.globl _bn_sub_words 798.private_extern _bn_sub_words 799.align 4 800_bn_sub_words: 801L_bn_sub_words_begin: 802 pushl %ebp 803 pushl %ebx 804 pushl %esi 805 pushl %edi 806 807 movl 20(%esp),%ebx 808 movl 24(%esp),%esi 809 movl 28(%esp),%edi 810 movl 32(%esp),%ebp 811 xorl %eax,%eax 812 andl $4294967288,%ebp 813 jz L026aw_finish 814L027aw_loop: 815 # Round 0 816 movl (%esi),%ecx 817 movl (%edi),%edx 818 subl %eax,%ecx 819 movl $0,%eax 820 adcl %eax,%eax 821 subl %edx,%ecx 822 adcl $0,%eax 823 movl %ecx,(%ebx) 824 # Round 1 825 movl 4(%esi),%ecx 826 movl 4(%edi),%edx 827 subl %eax,%ecx 828 movl $0,%eax 829 adcl %eax,%eax 830 subl %edx,%ecx 831 adcl $0,%eax 832 movl %ecx,4(%ebx) 833 # Round 2 834 movl 8(%esi),%ecx 835 movl 8(%edi),%edx 836 subl %eax,%ecx 837 movl $0,%eax 838 adcl %eax,%eax 839 subl %edx,%ecx 840 adcl $0,%eax 841 movl %ecx,8(%ebx) 842 # Round 3 843 movl 12(%esi),%ecx 844 movl 12(%edi),%edx 845 subl %eax,%ecx 846 movl $0,%eax 847 adcl %eax,%eax 848 subl %edx,%ecx 849 adcl $0,%eax 850 movl %ecx,12(%ebx) 851 # Round 4 852 movl 16(%esi),%ecx 853 movl 16(%edi),%edx 854 subl %eax,%ecx 855 movl $0,%eax 856 adcl %eax,%eax 857 subl %edx,%ecx 858 adcl $0,%eax 859 movl %ecx,16(%ebx) 860 # Round 5 861 movl 20(%esi),%ecx 862 movl 20(%edi),%edx 863 subl %eax,%ecx 864 movl $0,%eax 865 adcl %eax,%eax 866 subl %edx,%ecx 867 adcl $0,%eax 868 movl %ecx,20(%ebx) 869 # Round 6 870 movl 24(%esi),%ecx 871 movl 24(%edi),%edx 872 subl %eax,%ecx 873 movl $0,%eax 874 adcl %eax,%eax 875 subl %edx,%ecx 876 adcl $0,%eax 877 movl %ecx,24(%ebx) 878 # Round 7 879 movl 28(%esi),%ecx 880 movl 28(%edi),%edx 881 subl %eax,%ecx 882 movl $0,%eax 883 adcl %eax,%eax 884 subl %edx,%ecx 885 adcl $0,%eax 886 movl %ecx,28(%ebx) 887 888 addl $32,%esi 889 addl $32,%edi 890 addl $32,%ebx 891 subl $8,%ebp 892 jnz L027aw_loop 893L026aw_finish: 894 movl 32(%esp),%ebp 895 andl $7,%ebp 896 jz L028aw_end 897 # Tail Round 0 898 movl (%esi),%ecx 899 movl (%edi),%edx 900 subl %eax,%ecx 901 movl $0,%eax 902 adcl %eax,%eax 903 subl %edx,%ecx 904 adcl $0,%eax 905 decl %ebp 906 movl %ecx,(%ebx) 907 jz L028aw_end 908 # Tail Round 1 909 movl 4(%esi),%ecx 910 movl 4(%edi),%edx 911 subl %eax,%ecx 912 movl $0,%eax 913 adcl %eax,%eax 914 subl %edx,%ecx 915 adcl $0,%eax 916 decl %ebp 917 movl %ecx,4(%ebx) 918 jz L028aw_end 919 # Tail Round 2 920 movl 8(%esi),%ecx 921 movl 8(%edi),%edx 922 subl %eax,%ecx 923 movl $0,%eax 924 adcl %eax,%eax 925 subl %edx,%ecx 926 adcl $0,%eax 927 decl %ebp 928 movl %ecx,8(%ebx) 929 jz L028aw_end 930 # Tail Round 3 931 movl 12(%esi),%ecx 932 movl 12(%edi),%edx 933 subl %eax,%ecx 934 movl $0,%eax 935 adcl %eax,%eax 936 subl %edx,%ecx 937 adcl $0,%eax 938 decl %ebp 939 movl %ecx,12(%ebx) 940 jz L028aw_end 941 # Tail Round 4 942 movl 16(%esi),%ecx 943 movl 16(%edi),%edx 944 subl %eax,%ecx 945 movl $0,%eax 946 adcl %eax,%eax 947 subl %edx,%ecx 948 adcl $0,%eax 949 decl %ebp 950 movl %ecx,16(%ebx) 951 jz L028aw_end 952 # Tail Round 5 953 movl 20(%esi),%ecx 954 movl 20(%edi),%edx 955 subl %eax,%ecx 956 movl $0,%eax 957 adcl %eax,%eax 958 subl %edx,%ecx 959 adcl $0,%eax 960 decl %ebp 961 movl %ecx,20(%ebx) 962 jz L028aw_end 963 # Tail Round 6 964 movl 24(%esi),%ecx 965 movl 24(%edi),%edx 966 subl %eax,%ecx 967 movl $0,%eax 968 adcl %eax,%eax 969 subl %edx,%ecx 970 adcl $0,%eax 971 movl %ecx,24(%ebx) 972L028aw_end: 973 popl %edi 974 popl %esi 975 popl %ebx 976 popl %ebp 977 ret 978.globl _bn_sub_part_words 979.private_extern _bn_sub_part_words 980.align 4 981_bn_sub_part_words: 982L_bn_sub_part_words_begin: 983 pushl %ebp 984 pushl %ebx 985 pushl %esi 986 pushl %edi 987 988 movl 20(%esp),%ebx 989 movl 24(%esp),%esi 990 movl 28(%esp),%edi 991 movl 32(%esp),%ebp 992 xorl %eax,%eax 993 andl $4294967288,%ebp 994 jz L029aw_finish 995L030aw_loop: 996 # Round 0 997 movl (%esi),%ecx 998 movl (%edi),%edx 999 subl %eax,%ecx 1000 movl $0,%eax 1001 adcl %eax,%eax 1002 subl %edx,%ecx 1003 adcl $0,%eax 1004 movl %ecx,(%ebx) 1005 # Round 1 1006 movl 4(%esi),%ecx 1007 movl 4(%edi),%edx 1008 subl %eax,%ecx 1009 movl $0,%eax 1010 adcl %eax,%eax 1011 subl %edx,%ecx 1012 adcl $0,%eax 1013 movl %ecx,4(%ebx) 1014 # Round 2 1015 movl 8(%esi),%ecx 1016 movl 8(%edi),%edx 1017 subl %eax,%ecx 1018 movl $0,%eax 1019 adcl %eax,%eax 1020 subl %edx,%ecx 1021 adcl $0,%eax 1022 movl %ecx,8(%ebx) 1023 # Round 3 1024 movl 12(%esi),%ecx 1025 movl 12(%edi),%edx 1026 subl %eax,%ecx 1027 movl $0,%eax 1028 adcl %eax,%eax 1029 subl %edx,%ecx 1030 adcl $0,%eax 1031 movl %ecx,12(%ebx) 1032 # Round 4 1033 movl 16(%esi),%ecx 1034 movl 16(%edi),%edx 1035 subl %eax,%ecx 1036 movl $0,%eax 1037 adcl %eax,%eax 1038 subl %edx,%ecx 1039 adcl $0,%eax 1040 movl %ecx,16(%ebx) 1041 # Round 5 1042 movl 20(%esi),%ecx 1043 movl 20(%edi),%edx 1044 subl %eax,%ecx 1045 movl $0,%eax 1046 adcl %eax,%eax 1047 subl %edx,%ecx 1048 adcl $0,%eax 1049 movl %ecx,20(%ebx) 1050 # Round 6 1051 movl 24(%esi),%ecx 1052 movl 24(%edi),%edx 1053 subl %eax,%ecx 1054 movl $0,%eax 1055 adcl %eax,%eax 1056 subl %edx,%ecx 1057 adcl $0,%eax 1058 movl %ecx,24(%ebx) 1059 # Round 7 1060 movl 28(%esi),%ecx 1061 movl 28(%edi),%edx 1062 subl %eax,%ecx 1063 movl $0,%eax 1064 adcl %eax,%eax 1065 subl %edx,%ecx 1066 adcl $0,%eax 1067 movl %ecx,28(%ebx) 1068 1069 addl $32,%esi 1070 addl $32,%edi 1071 addl $32,%ebx 1072 subl $8,%ebp 1073 jnz L030aw_loop 1074L029aw_finish: 1075 movl 32(%esp),%ebp 1076 andl $7,%ebp 1077 jz L031aw_end 1078 # Tail Round 0 1079 movl (%esi),%ecx 1080 movl (%edi),%edx 1081 subl %eax,%ecx 1082 movl $0,%eax 1083 adcl %eax,%eax 1084 subl %edx,%ecx 1085 adcl $0,%eax 1086 movl %ecx,(%ebx) 1087 addl $4,%esi 1088 addl $4,%edi 1089 addl $4,%ebx 1090 decl %ebp 1091 jz L031aw_end 1092 # Tail Round 1 1093 movl (%esi),%ecx 1094 movl (%edi),%edx 1095 subl %eax,%ecx 1096 movl $0,%eax 1097 adcl %eax,%eax 1098 subl %edx,%ecx 1099 adcl $0,%eax 1100 movl %ecx,(%ebx) 1101 addl $4,%esi 1102 addl $4,%edi 1103 addl $4,%ebx 1104 decl %ebp 1105 jz L031aw_end 1106 # Tail Round 2 1107 movl (%esi),%ecx 1108 movl (%edi),%edx 1109 subl %eax,%ecx 1110 movl $0,%eax 1111 adcl %eax,%eax 1112 subl %edx,%ecx 1113 adcl $0,%eax 1114 movl %ecx,(%ebx) 1115 addl $4,%esi 1116 addl $4,%edi 1117 addl $4,%ebx 1118 decl %ebp 1119 jz L031aw_end 1120 # Tail Round 3 1121 movl (%esi),%ecx 1122 movl (%edi),%edx 1123 subl %eax,%ecx 1124 movl $0,%eax 1125 adcl %eax,%eax 1126 subl %edx,%ecx 1127 adcl $0,%eax 1128 movl %ecx,(%ebx) 1129 addl $4,%esi 1130 addl $4,%edi 1131 addl $4,%ebx 1132 decl %ebp 1133 jz L031aw_end 1134 # Tail Round 4 1135 movl (%esi),%ecx 1136 movl (%edi),%edx 1137 subl %eax,%ecx 1138 movl $0,%eax 1139 adcl %eax,%eax 1140 subl %edx,%ecx 1141 adcl $0,%eax 1142 movl %ecx,(%ebx) 1143 addl $4,%esi 1144 addl $4,%edi 1145 addl $4,%ebx 1146 decl %ebp 1147 jz L031aw_end 1148 # Tail Round 5 1149 movl (%esi),%ecx 1150 movl (%edi),%edx 1151 subl %eax,%ecx 1152 movl $0,%eax 1153 adcl %eax,%eax 1154 subl %edx,%ecx 1155 adcl $0,%eax 1156 movl %ecx,(%ebx) 1157 addl $4,%esi 1158 addl $4,%edi 1159 addl $4,%ebx 1160 decl %ebp 1161 jz L031aw_end 1162 # Tail Round 6 1163 movl (%esi),%ecx 1164 movl (%edi),%edx 1165 subl %eax,%ecx 1166 movl $0,%eax 1167 adcl %eax,%eax 1168 subl %edx,%ecx 1169 adcl $0,%eax 1170 movl %ecx,(%ebx) 1171 addl $4,%esi 1172 addl $4,%edi 1173 addl $4,%ebx 1174L031aw_end: 1175 cmpl $0,36(%esp) 1176 je L032pw_end 1177 movl 36(%esp),%ebp 1178 cmpl $0,%ebp 1179 je L032pw_end 1180 jge L033pw_pos 1181 # pw_neg 1182 movl $0,%edx 1183 subl %ebp,%edx 1184 movl %edx,%ebp 1185 andl $4294967288,%ebp 1186 jz L034pw_neg_finish 1187L035pw_neg_loop: 1188 # dl<0 Round 0 1189 movl $0,%ecx 1190 movl (%edi),%edx 1191 subl %eax,%ecx 1192 movl $0,%eax 1193 adcl %eax,%eax 1194 subl %edx,%ecx 1195 adcl $0,%eax 1196 movl %ecx,(%ebx) 1197 # dl<0 Round 1 1198 movl $0,%ecx 1199 movl 4(%edi),%edx 1200 subl %eax,%ecx 1201 movl $0,%eax 1202 adcl %eax,%eax 1203 subl %edx,%ecx 1204 adcl $0,%eax 1205 movl %ecx,4(%ebx) 1206 # dl<0 Round 2 1207 movl $0,%ecx 1208 movl 8(%edi),%edx 1209 subl %eax,%ecx 1210 movl $0,%eax 1211 adcl %eax,%eax 1212 subl %edx,%ecx 1213 adcl $0,%eax 1214 movl %ecx,8(%ebx) 1215 # dl<0 Round 3 1216 movl $0,%ecx 1217 movl 12(%edi),%edx 1218 subl %eax,%ecx 1219 movl $0,%eax 1220 adcl %eax,%eax 1221 subl %edx,%ecx 1222 adcl $0,%eax 1223 movl %ecx,12(%ebx) 1224 # dl<0 Round 4 1225 movl $0,%ecx 1226 movl 16(%edi),%edx 1227 subl %eax,%ecx 1228 movl $0,%eax 1229 adcl %eax,%eax 1230 subl %edx,%ecx 1231 adcl $0,%eax 1232 movl %ecx,16(%ebx) 1233 # dl<0 Round 5 1234 movl $0,%ecx 1235 movl 20(%edi),%edx 1236 subl %eax,%ecx 1237 movl $0,%eax 1238 adcl %eax,%eax 1239 subl %edx,%ecx 1240 adcl $0,%eax 1241 movl %ecx,20(%ebx) 1242 # dl<0 Round 6 1243 movl $0,%ecx 1244 movl 24(%edi),%edx 1245 subl %eax,%ecx 1246 movl $0,%eax 1247 adcl %eax,%eax 1248 subl %edx,%ecx 1249 adcl $0,%eax 1250 movl %ecx,24(%ebx) 1251 # dl<0 Round 7 1252 movl $0,%ecx 1253 movl 28(%edi),%edx 1254 subl %eax,%ecx 1255 movl $0,%eax 1256 adcl %eax,%eax 1257 subl %edx,%ecx 1258 adcl $0,%eax 1259 movl %ecx,28(%ebx) 1260 1261 addl $32,%edi 1262 addl $32,%ebx 1263 subl $8,%ebp 1264 jnz L035pw_neg_loop 1265L034pw_neg_finish: 1266 movl 36(%esp),%edx 1267 movl $0,%ebp 1268 subl %edx,%ebp 1269 andl $7,%ebp 1270 jz L032pw_end 1271 # dl<0 Tail Round 0 1272 movl $0,%ecx 1273 movl (%edi),%edx 1274 subl %eax,%ecx 1275 movl $0,%eax 1276 adcl %eax,%eax 1277 subl %edx,%ecx 1278 adcl $0,%eax 1279 decl %ebp 1280 movl %ecx,(%ebx) 1281 jz L032pw_end 1282 # dl<0 Tail Round 1 1283 movl $0,%ecx 1284 movl 4(%edi),%edx 1285 subl %eax,%ecx 1286 movl $0,%eax 1287 adcl %eax,%eax 1288 subl %edx,%ecx 1289 adcl $0,%eax 1290 decl %ebp 1291 movl %ecx,4(%ebx) 1292 jz L032pw_end 1293 # dl<0 Tail Round 2 1294 movl $0,%ecx 1295 movl 8(%edi),%edx 1296 subl %eax,%ecx 1297 movl $0,%eax 1298 adcl %eax,%eax 1299 subl %edx,%ecx 1300 adcl $0,%eax 1301 decl %ebp 1302 movl %ecx,8(%ebx) 1303 jz L032pw_end 1304 # dl<0 Tail Round 3 1305 movl $0,%ecx 1306 movl 12(%edi),%edx 1307 subl %eax,%ecx 1308 movl $0,%eax 1309 adcl %eax,%eax 1310 subl %edx,%ecx 1311 adcl $0,%eax 1312 decl %ebp 1313 movl %ecx,12(%ebx) 1314 jz L032pw_end 1315 # dl<0 Tail Round 4 1316 movl $0,%ecx 1317 movl 16(%edi),%edx 1318 subl %eax,%ecx 1319 movl $0,%eax 1320 adcl %eax,%eax 1321 subl %edx,%ecx 1322 adcl $0,%eax 1323 decl %ebp 1324 movl %ecx,16(%ebx) 1325 jz L032pw_end 1326 # dl<0 Tail Round 5 1327 movl $0,%ecx 1328 movl 20(%edi),%edx 1329 subl %eax,%ecx 1330 movl $0,%eax 1331 adcl %eax,%eax 1332 subl %edx,%ecx 1333 adcl $0,%eax 1334 decl %ebp 1335 movl %ecx,20(%ebx) 1336 jz L032pw_end 1337 # dl<0 Tail Round 6 1338 movl $0,%ecx 1339 movl 24(%edi),%edx 1340 subl %eax,%ecx 1341 movl $0,%eax 1342 adcl %eax,%eax 1343 subl %edx,%ecx 1344 adcl $0,%eax 1345 movl %ecx,24(%ebx) 1346 jmp L032pw_end 1347L033pw_pos: 1348 andl $4294967288,%ebp 1349 jz L036pw_pos_finish 1350L037pw_pos_loop: 1351 # dl>0 Round 0 1352 movl (%esi),%ecx 1353 subl %eax,%ecx 1354 movl %ecx,(%ebx) 1355 jnc L038pw_nc0 1356 # dl>0 Round 1 1357 movl 4(%esi),%ecx 1358 subl %eax,%ecx 1359 movl %ecx,4(%ebx) 1360 jnc L039pw_nc1 1361 # dl>0 Round 2 1362 movl 8(%esi),%ecx 1363 subl %eax,%ecx 1364 movl %ecx,8(%ebx) 1365 jnc L040pw_nc2 1366 # dl>0 Round 3 1367 movl 12(%esi),%ecx 1368 subl %eax,%ecx 1369 movl %ecx,12(%ebx) 1370 jnc L041pw_nc3 1371 # dl>0 Round 4 1372 movl 16(%esi),%ecx 1373 subl %eax,%ecx 1374 movl %ecx,16(%ebx) 1375 jnc L042pw_nc4 1376 # dl>0 Round 5 1377 movl 20(%esi),%ecx 1378 subl %eax,%ecx 1379 movl %ecx,20(%ebx) 1380 jnc L043pw_nc5 1381 # dl>0 Round 6 1382 movl 24(%esi),%ecx 1383 subl %eax,%ecx 1384 movl %ecx,24(%ebx) 1385 jnc L044pw_nc6 1386 # dl>0 Round 7 1387 movl 28(%esi),%ecx 1388 subl %eax,%ecx 1389 movl %ecx,28(%ebx) 1390 jnc L045pw_nc7 1391 1392 addl $32,%esi 1393 addl $32,%ebx 1394 subl $8,%ebp 1395 jnz L037pw_pos_loop 1396L036pw_pos_finish: 1397 movl 36(%esp),%ebp 1398 andl $7,%ebp 1399 jz L032pw_end 1400 # dl>0 Tail Round 0 1401 movl (%esi),%ecx 1402 subl %eax,%ecx 1403 movl %ecx,(%ebx) 1404 jnc L046pw_tail_nc0 1405 decl %ebp 1406 jz L032pw_end 1407 # dl>0 Tail Round 1 1408 movl 4(%esi),%ecx 1409 subl %eax,%ecx 1410 movl %ecx,4(%ebx) 1411 jnc L047pw_tail_nc1 1412 decl %ebp 1413 jz L032pw_end 1414 # dl>0 Tail Round 2 1415 movl 8(%esi),%ecx 1416 subl %eax,%ecx 1417 movl %ecx,8(%ebx) 1418 jnc L048pw_tail_nc2 1419 decl %ebp 1420 jz L032pw_end 1421 # dl>0 Tail Round 3 1422 movl 12(%esi),%ecx 1423 subl %eax,%ecx 1424 movl %ecx,12(%ebx) 1425 jnc L049pw_tail_nc3 1426 decl %ebp 1427 jz L032pw_end 1428 # dl>0 Tail Round 4 1429 movl 16(%esi),%ecx 1430 subl %eax,%ecx 1431 movl %ecx,16(%ebx) 1432 jnc L050pw_tail_nc4 1433 decl %ebp 1434 jz L032pw_end 1435 # dl>0 Tail Round 5 1436 movl 20(%esi),%ecx 1437 subl %eax,%ecx 1438 movl %ecx,20(%ebx) 1439 jnc L051pw_tail_nc5 1440 decl %ebp 1441 jz L032pw_end 1442 # dl>0 Tail Round 6 1443 movl 24(%esi),%ecx 1444 subl %eax,%ecx 1445 movl %ecx,24(%ebx) 1446 jnc L052pw_tail_nc6 1447 movl $1,%eax 1448 jmp L032pw_end 1449L053pw_nc_loop: 1450 movl (%esi),%ecx 1451 movl %ecx,(%ebx) 1452L038pw_nc0: 1453 movl 4(%esi),%ecx 1454 movl %ecx,4(%ebx) 1455L039pw_nc1: 1456 movl 8(%esi),%ecx 1457 movl %ecx,8(%ebx) 1458L040pw_nc2: 1459 movl 12(%esi),%ecx 1460 movl %ecx,12(%ebx) 1461L041pw_nc3: 1462 movl 16(%esi),%ecx 1463 movl %ecx,16(%ebx) 1464L042pw_nc4: 1465 movl 20(%esi),%ecx 1466 movl %ecx,20(%ebx) 1467L043pw_nc5: 1468 movl 24(%esi),%ecx 1469 movl %ecx,24(%ebx) 1470L044pw_nc6: 1471 movl 28(%esi),%ecx 1472 movl %ecx,28(%ebx) 1473L045pw_nc7: 1474 1475 addl $32,%esi 1476 addl $32,%ebx 1477 subl $8,%ebp 1478 jnz L053pw_nc_loop 1479 movl 36(%esp),%ebp 1480 andl $7,%ebp 1481 jz L054pw_nc_end 1482 movl (%esi),%ecx 1483 movl %ecx,(%ebx) 1484L046pw_tail_nc0: 1485 decl %ebp 1486 jz L054pw_nc_end 1487 movl 4(%esi),%ecx 1488 movl %ecx,4(%ebx) 1489L047pw_tail_nc1: 1490 decl %ebp 1491 jz L054pw_nc_end 1492 movl 8(%esi),%ecx 1493 movl %ecx,8(%ebx) 1494L048pw_tail_nc2: 1495 decl %ebp 1496 jz L054pw_nc_end 1497 movl 12(%esi),%ecx 1498 movl %ecx,12(%ebx) 1499L049pw_tail_nc3: 1500 decl %ebp 1501 jz L054pw_nc_end 1502 movl 16(%esi),%ecx 1503 movl %ecx,16(%ebx) 1504L050pw_tail_nc4: 1505 decl %ebp 1506 jz L054pw_nc_end 1507 movl 20(%esi),%ecx 1508 movl %ecx,20(%ebx) 1509L051pw_tail_nc5: 1510 decl %ebp 1511 jz L054pw_nc_end 1512 movl 24(%esi),%ecx 1513 movl %ecx,24(%ebx) 1514L052pw_tail_nc6: 1515L054pw_nc_end: 1516 movl $0,%eax 1517L032pw_end: 1518 popl %edi 1519 popl %esi 1520 popl %ebx 1521 popl %ebp 1522 ret 1523.section __IMPORT,__pointers,non_lazy_symbol_pointers 1524L_OPENSSL_ia32cap_P$non_lazy_ptr: 1525.indirect_symbol _OPENSSL_ia32cap_P 1526.long 0 1527#endif 1528