1 .machine "any" 2 .text 3 4 .globl bn_mul_mont_int 5 .type bn_mul_mont_int,@function 6 .align 5 7 bn_mul_mont_int: 8 mr 9,3 9 li 3,0 10 cmpwi 8,32 11 bgelr 12 slwi 8,8,2 13 li 12,-4096 14 addi 3,8,256 15 subf 3,3,1 16 and 3,3,12 17 subf 3,1,3 18 mr 12,1 19 srwi 8,8,2 20 stwux 1,1,3 21 22 stw 20,-48(12) 23 stw 21,-44(12) 24 stw 22,-40(12) 25 stw 23,-36(12) 26 stw 24,-32(12) 27 stw 25,-28(12) 28 stw 26,-24(12) 29 stw 27,-20(12) 30 stw 28,-16(12) 31 stw 29,-12(12) 32 stw 30,-8(12) 33 stw 31,-4(12) 34 35 lwz 7,0(7) 36 addi 8,8,-2 37 38 lwz 23,0(5) 39 lwz 10,0(4) 40 addi 22,1,32 41 mullw 25,10,23 42 mulhwu 26,10,23 43 44 lwz 10,4(4) 45 lwz 11,0(6) 46 47 mullw 24,25,7 48 49 mullw 29,10,23 50 mulhwu 30,10,23 51 52 mullw 27,11,24 53 mulhwu 28,11,24 54 lwz 11,4(6) 55 addc 27,27,25 56 addze 28,28 57 58 mullw 31,11,24 59 mulhwu 0,11,24 60 61 mtctr 8 62 li 21,8 63 .align 4 64 .L1st: 65 lwzx 10,4,21 66 addc 25,29,26 67 lwzx 11,6,21 68 addze 26,30 69 mullw 29,10,23 70 addc 27,31,28 71 mulhwu 30,10,23 72 addze 28,0 73 mullw 31,11,24 74 addc 27,27,25 75 mulhwu 0,11,24 76 addze 28,28 77 stw 27,0(22) 78 79 addi 21,21,4 80 addi 22,22,4 81 bc 16,0,.L1st 82 83 addc 25,29,26 84 addze 26,30 85 86 addc 27,31,28 87 addze 28,0 88 addc 27,27,25 89 addze 28,28 90 stw 27,0(22) 91 92 li 3,0 93 addc 28,28,26 94 addze 3,3 95 stw 28,4(22) 96 97 li 20,4 98 .align 4 99 .Louter: 100 lwzx 23,5,20 101 lwz 10,0(4) 102 addi 22,1,32 103 lwz 12,32(1) 104 mullw 25,10,23 105 mulhwu 26,10,23 106 lwz 10,4(4) 107 lwz 11,0(6) 108 addc 25,25,12 109 mullw 29,10,23 110 addze 26,26 111 mullw 24,25,7 112 mulhwu 30,10,23 113 mullw 27,11,24 114 mulhwu 28,11,24 115 lwz 11,4(6) 116 addc 27,27,25 117 mullw 31,11,24 118 addze 28,28 119 mulhwu 0,11,24 120 121 mtctr 8 122 li 21,8 123 .align 4 124 .Linner: 125 lwzx 10,4,21 126 addc 25,29,26 127 lwz 12,4(22) 128 addze 26,30 129 lwzx 11,6,21 130 addc 27,31,28 131 mullw 29,10,23 132 addze 28,0 133 mulhwu 30,10,23 134 addc 25,25,12 135 mullw 31,11,24 136 addze 26,26 137 mulhwu 0,11,24 138 addc 27,27,25 139 addi 21,21,4 140 addze 28,28 141 stw 27,0(22) 142 addi 22,22,4 143 bc 16,0,.Linner 144 145 lwz 12,4(22) 146 addc 25,29,26 147 addze 26,30 148 addc 25,25,12 149 addze 26,26 150 151 addc 27,31,28 152 addze 28,0 153 addc 27,27,25 154 addze 28,28 155 stw 27,0(22) 156 157 addic 3,3,-1 158 li 3,0 159 adde 28,28,26 160 addze 3,3 161 stw 28,4(22) 162 163 slwi 12,8,2 164 .long 0x7c146040 165 addi 20,20,4 166 ble .Louter 167 168 addi 8,8,2 169 subfc 21,21,21 170 addi 22,1,32 171 mtctr 8 172 173 .align 4 174 .Lsub: lwzx 12,22,21 175 lwzx 11,6,21 176 subfe 10,11,12 177 stwx 10,9,21 178 addi 21,21,4 179 bc 16,0,.Lsub 180 181 li 21,0 182 mtctr 8 183 subfe 3,21,3 184 185 .align 4 186 .Lcopy: 187 lwzx 12,22,21 188 lwzx 10,9,21 189 and 12,12,3 190 andc 10,10,3 191 stwx 21,22,21 192 or 10,10,12 193 stwx 10,9,21 194 addi 21,21,4 195 bc 16,0,.Lcopy 196 197 lwz 12,0(1) 198 li 3,1 199 lwz 20,-48(12) 200 lwz 21,-44(12) 201 lwz 22,-40(12) 202 lwz 23,-36(12) 203 lwz 24,-32(12) 204 lwz 25,-28(12) 205 lwz 26,-24(12) 206 lwz 27,-20(12) 207 lwz 28,-16(12) 208 lwz 29,-12(12) 209 lwz 30,-8(12) 210 lwz 31,-4(12) 211 mr 1,12 212 blr 213 .long 0 214 .byte 0,12,4,0,0x80,12,6,0 215 .long 0 216 217 .globl bn_mul4x_mont_int 218 .type bn_mul4x_mont_int,@function 219 .align 5 220 bn_mul4x_mont_int: 221 andi. 0,8,7 222 bne .Lmul4x_do 223 .long 0x7c042840 224 bne .Lmul4x_do 225 b .Lsqr8x_do 226 .Lmul4x_do: 227 slwi 8,8,2 228 mr 9,1 229 li 10,-32*4 230 sub 10,10,8 231 stwux 1,1,10 232 233 stw 14,-4*18(9) 234 stw 15,-4*17(9) 235 stw 16,-4*16(9) 236 stw 17,-4*15(9) 237 stw 18,-4*14(9) 238 stw 19,-4*13(9) 239 stw 20,-4*12(9) 240 stw 21,-4*11(9) 241 stw 22,-4*10(9) 242 stw 23,-4*9(9) 243 stw 24,-4*8(9) 244 stw 25,-4*7(9) 245 stw 26,-4*6(9) 246 stw 27,-4*5(9) 247 stw 28,-4*4(9) 248 stw 29,-4*3(9) 249 stw 30,-4*2(9) 250 stw 31,-4*1(9) 251 252 subi 4,4,4 253 subi 6,6,4 254 subi 3,3,4 255 lwz 7,0(7) 256 257 add 14,5,8 258 add 30,4,8 259 subi 14,14,4*4 260 261 lwz 27,4*0(5) 262 li 22,0 263 lwz 9,4*1(4) 264 li 23,0 265 lwz 10,4*2(4) 266 li 24,0 267 lwz 11,4*3(4) 268 li 25,0 269 lwzu 12,4*4(4) 270 lwz 18,4*1(6) 271 lwz 19,4*2(6) 272 lwz 20,4*3(6) 273 lwzu 21,4*4(6) 274 275 stw 3,4*6(1) 276 stw 14,4*7(1) 277 li 3,0 278 addic 29,1,4*7 279 li 31,0 280 li 0,0 281 b .Loop_mul4x_1st_reduction 282 283 .align 5 284 .Loop_mul4x_1st_reduction: 285 mullw 14,9,27 286 addze 3,3 287 mullw 15,10,27 288 addi 31,31,4 289 mullw 16,11,27 290 andi. 31,31,4*4-1 291 mullw 17,12,27 292 addc 22,22,14 293 mulhwu 14,9,27 294 adde 23,23,15 295 mulhwu 15,10,27 296 adde 24,24,16 297 mullw 28,22,7 298 adde 25,25,17 299 mulhwu 16,11,27 300 addze 26,0 301 mulhwu 17,12,27 302 lwzx 27,5,31 303 addc 23,23,14 304 305 stwu 28,4(29) 306 adde 24,24,15 307 mullw 15,19,28 308 adde 25,25,16 309 mullw 16,20,28 310 adde 26,26,17 311 mullw 17,21,28 312 313 314 315 316 317 318 319 320 321 322 addic 22,22,-1 323 mulhwu 14,18,28 324 adde 22,23,15 325 mulhwu 15,19,28 326 adde 23,24,16 327 mulhwu 16,20,28 328 adde 24,25,17 329 mulhwu 17,21,28 330 adde 25,26,3 331 addze 3,0 332 addc 22,22,14 333 adde 23,23,15 334 adde 24,24,16 335 adde 25,25,17 336 337 bne .Loop_mul4x_1st_reduction 338 339 .long 0x7c1e2040 340 beq .Lmul4x4_post_condition 341 342 lwz 9,4*1(4) 343 lwz 10,4*2(4) 344 lwz 11,4*3(4) 345 lwzu 12,4*4(4) 346 lwz 28,4*8(1) 347 lwz 18,4*1(6) 348 lwz 19,4*2(6) 349 lwz 20,4*3(6) 350 lwzu 21,4*4(6) 351 b .Loop_mul4x_1st_tail 352 353 .align 5 354 .Loop_mul4x_1st_tail: 355 mullw 14,9,27 356 addze 3,3 357 mullw 15,10,27 358 addi 31,31,4 359 mullw 16,11,27 360 andi. 31,31,4*4-1 361 mullw 17,12,27 362 addc 22,22,14 363 mulhwu 14,9,27 364 adde 23,23,15 365 mulhwu 15,10,27 366 adde 24,24,16 367 mulhwu 16,11,27 368 adde 25,25,17 369 mulhwu 17,12,27 370 addze 26,0 371 lwzx 27,5,31 372 addc 23,23,14 373 mullw 14,18,28 374 adde 24,24,15 375 mullw 15,19,28 376 adde 25,25,16 377 mullw 16,20,28 378 adde 26,26,17 379 mullw 17,21,28 380 addc 22,22,14 381 mulhwu 14,18,28 382 adde 23,23,15 383 mulhwu 15,19,28 384 adde 24,24,16 385 mulhwu 16,20,28 386 adde 25,25,17 387 adde 26,26,3 388 mulhwu 17,21,28 389 addze 3,0 390 addi 28,1,4*8 391 lwzx 28,28,31 392 stwu 22,4(29) 393 addc 22,23,14 394 adde 23,24,15 395 adde 24,25,16 396 adde 25,26,17 397 398 bne .Loop_mul4x_1st_tail 399 400 sub 15,30,8 401 .long 0x7c1e2040 402 beq .Lmul4x_proceed 403 404 lwz 9,4*1(4) 405 lwz 10,4*2(4) 406 lwz 11,4*3(4) 407 lwzu 12,4*4(4) 408 lwz 18,4*1(6) 409 lwz 19,4*2(6) 410 lwz 20,4*3(6) 411 lwzu 21,4*4(6) 412 b .Loop_mul4x_1st_tail 413 414 .align 5 415 .Lmul4x_proceed: 416 lwzu 27,4*4(5) 417 addze 3,3 418 lwz 9,4*1(15) 419 lwz 10,4*2(15) 420 lwz 11,4*3(15) 421 lwz 12,4*4(15) 422 addi 4,15,4*4 423 sub 6,6,8 424 425 stw 22,4*1(29) 426 stw 23,4*2(29) 427 stw 24,4*3(29) 428 stw 25,4*4(29) 429 stw 3,4*5(29) 430 lwz 22,4*12(1) 431 lwz 23,4*13(1) 432 lwz 24,4*14(1) 433 lwz 25,4*15(1) 434 435 lwz 18,4*1(6) 436 lwz 19,4*2(6) 437 lwz 20,4*3(6) 438 lwzu 21,4*4(6) 439 addic 29,1,4*7 440 li 3,0 441 b .Loop_mul4x_reduction 442 443 .align 5 444 .Loop_mul4x_reduction: 445 mullw 14,9,27 446 addze 3,3 447 mullw 15,10,27 448 addi 31,31,4 449 mullw 16,11,27 450 andi. 31,31,4*4-1 451 mullw 17,12,27 452 addc 22,22,14 453 mulhwu 14,9,27 454 adde 23,23,15 455 mulhwu 15,10,27 456 adde 24,24,16 457 mullw 28,22,7 458 adde 25,25,17 459 mulhwu 16,11,27 460 addze 26,0 461 mulhwu 17,12,27 462 lwzx 27,5,31 463 addc 23,23,14 464 465 stwu 28,4(29) 466 adde 24,24,15 467 mullw 15,19,28 468 adde 25,25,16 469 mullw 16,20,28 470 adde 26,26,17 471 mullw 17,21,28 472 473 addic 22,22,-1 474 mulhwu 14,18,28 475 adde 22,23,15 476 mulhwu 15,19,28 477 adde 23,24,16 478 mulhwu 16,20,28 479 adde 24,25,17 480 mulhwu 17,21,28 481 adde 25,26,3 482 addze 3,0 483 addc 22,22,14 484 adde 23,23,15 485 adde 24,24,16 486 adde 25,25,17 487 488 bne .Loop_mul4x_reduction 489 490 lwz 14,4*5(29) 491 addze 3,3 492 lwz 15,4*6(29) 493 lwz 16,4*7(29) 494 lwz 17,4*8(29) 495 lwz 9,4*1(4) 496 lwz 10,4*2(4) 497 lwz 11,4*3(4) 498 lwzu 12,4*4(4) 499 addc 22,22,14 500 adde 23,23,15 501 adde 24,24,16 502 adde 25,25,17 503 504 505 lwz 28,4*8(1) 506 lwz 18,4*1(6) 507 lwz 19,4*2(6) 508 lwz 20,4*3(6) 509 lwzu 21,4*4(6) 510 b .Loop_mul4x_tail 511 512 .align 5 513 .Loop_mul4x_tail: 514 mullw 14,9,27 515 addze 3,3 516 mullw 15,10,27 517 addi 31,31,4 518 mullw 16,11,27 519 andi. 31,31,4*4-1 520 mullw 17,12,27 521 addc 22,22,14 522 mulhwu 14,9,27 523 adde 23,23,15 524 mulhwu 15,10,27 525 adde 24,24,16 526 mulhwu 16,11,27 527 adde 25,25,17 528 mulhwu 17,12,27 529 addze 26,0 530 lwzx 27,5,31 531 addc 23,23,14 532 mullw 14,18,28 533 adde 24,24,15 534 mullw 15,19,28 535 adde 25,25,16 536 mullw 16,20,28 537 adde 26,26,17 538 mullw 17,21,28 539 addc 22,22,14 540 mulhwu 14,18,28 541 adde 23,23,15 542 mulhwu 15,19,28 543 adde 24,24,16 544 mulhwu 16,20,28 545 adde 25,25,17 546 mulhwu 17,21,28 547 adde 26,26,3 548 addi 28,1,4*8 549 lwzx 28,28,31 550 addze 3,0 551 stwu 22,4(29) 552 addc 22,23,14 553 adde 23,24,15 554 adde 24,25,16 555 adde 25,26,17 556 557 bne .Loop_mul4x_tail 558 559 lwz 14,4*5(29) 560 sub 15,6,8 561 addze 3,3 562 .long 0x7c1e2040 563 beq .Loop_mul4x_break 564 565 lwz 15,4*6(29) 566 lwz 16,4*7(29) 567 lwz 17,4*8(29) 568 lwz 9,4*1(4) 569 lwz 10,4*2(4) 570 lwz 11,4*3(4) 571 lwzu 12,4*4(4) 572 addc 22,22,14 573 adde 23,23,15 574 adde 24,24,16 575 adde 25,25,17 576 577 578 lwz 18,4*1(6) 579 lwz 19,4*2(6) 580 lwz 20,4*3(6) 581 lwzu 21,4*4(6) 582 b .Loop_mul4x_tail 583 584 .align 5 585 .Loop_mul4x_break: 586 lwz 16,4*6(1) 587 lwz 17,4*7(1) 588 addc 9,22,14 589 lwz 22,4*12(1) 590 addze 10,23 591 lwz 23,4*13(1) 592 addze 11,24 593 lwz 24,4*14(1) 594 addze 12,25 595 lwz 25,4*15(1) 596 addze 3,3 597 stw 9,4*1(29) 598 sub 4,30,8 599 stw 10,4*2(29) 600 stw 11,4*3(29) 601 stw 12,4*4(29) 602 stw 3,4*5(29) 603 604 lwz 18,4*1(15) 605 lwz 19,4*2(15) 606 lwz 20,4*3(15) 607 lwz 21,4*4(15) 608 addi 6,15,4*4 609 .long 0x7c058840 610 beq .Lmul4x_post 611 612 lwzu 27,4*4(5) 613 lwz 9,4*1(4) 614 lwz 10,4*2(4) 615 lwz 11,4*3(4) 616 lwzu 12,4*4(4) 617 li 3,0 618 addic 29,1,4*7 619 b .Loop_mul4x_reduction 620 621 .align 5 622 .Lmul4x_post: 623 624 625 626 627 srwi 31,8,4 628 mr 5,16 629 subi 31,31,1 630 mr 30,16 631 subfc 14,18,22 632 addi 29,1,4*15 633 subfe 15,19,23 634 635 mtctr 31 636 .Lmul4x_sub: 637 lwz 18,4*1(6) 638 lwz 22,4*1(29) 639 subfe 16,20,24 640 lwz 19,4*2(6) 641 lwz 23,4*2(29) 642 subfe 17,21,25 643 lwz 20,4*3(6) 644 lwz 24,4*3(29) 645 lwzu 21,4*4(6) 646 lwzu 25,4*4(29) 647 stw 14,4*1(5) 648 stw 15,4*2(5) 649 subfe 14,18,22 650 stw 16,4*3(5) 651 stwu 17,4*4(5) 652 subfe 15,19,23 653 bc 16,0,.Lmul4x_sub 654 655 lwz 9,4*1(30) 656 stw 14,4*1(5) 657 lwz 14,4*12(1) 658 subfe 16,20,24 659 lwz 10,4*2(30) 660 stw 15,4*2(5) 661 lwz 15,4*13(1) 662 subfe 17,21,25 663 subfe 3,0,3 664 addi 29,1,4*12 665 lwz 11,4*3(30) 666 stw 16,4*3(5) 667 lwz 16,4*14(1) 668 lwz 12,4*4(30) 669 stw 17,4*4(5) 670 lwz 17,4*15(1) 671 672 mtctr 31 673 .Lmul4x_cond_copy: 674 and 14,14,3 675 andc 9,9,3 676 stw 0,4*0(29) 677 and 15,15,3 678 andc 10,10,3 679 stw 0,4*1(29) 680 and 16,16,3 681 andc 11,11,3 682 stw 0,4*2(29) 683 and 17,17,3 684 andc 12,12,3 685 stw 0,4*3(29) 686 or 22,14,9 687 lwz 9,4*5(30) 688 lwz 14,4*4(29) 689 or 23,15,10 690 lwz 10,4*6(30) 691 lwz 15,4*5(29) 692 or 24,16,11 693 lwz 11,4*7(30) 694 lwz 16,4*6(29) 695 or 25,17,12 696 lwz 12,4*8(30) 697 lwz 17,4*7(29) 698 addi 29,29,4*4 699 stw 22,4*1(30) 700 stw 23,4*2(30) 701 stw 24,4*3(30) 702 stwu 25,4*4(30) 703 bc 16,0,.Lmul4x_cond_copy 704 705 lwz 5,0(1) 706 and 14,14,3 707 andc 9,9,3 708 stw 0,4*0(29) 709 and 15,15,3 710 andc 10,10,3 711 stw 0,4*1(29) 712 and 16,16,3 713 andc 11,11,3 714 stw 0,4*2(29) 715 and 17,17,3 716 andc 12,12,3 717 stw 0,4*3(29) 718 or 22,14,9 719 or 23,15,10 720 stw 0,4*4(29) 721 or 24,16,11 722 or 25,17,12 723 stw 22,4*1(30) 724 stw 23,4*2(30) 725 stw 24,4*3(30) 726 stw 25,4*4(30) 727 728 b .Lmul4x_done 729 730 .align 4 731 .Lmul4x4_post_condition: 732 lwz 4,4*6(1) 733 lwz 5,0(1) 734 addze 3,3 735 736 subfc 9,18,22 737 subfe 10,19,23 738 subfe 11,20,24 739 subfe 12,21,25 740 subfe 3,0,3 741 742 and 18,18,3 743 and 19,19,3 744 addc 9,9,18 745 and 20,20,3 746 adde 10,10,19 747 and 21,21,3 748 adde 11,11,20 749 adde 12,12,21 750 751 stw 9,4*1(4) 752 stw 10,4*2(4) 753 stw 11,4*3(4) 754 stw 12,4*4(4) 755 756 .Lmul4x_done: 757 stw 0,4*8(1) 758 stw 0,4*9(1) 759 stw 0,4*10(1) 760 stw 0,4*11(1) 761 li 3,1 762 lwz 14,-4*18(5) 763 lwz 15,-4*17(5) 764 lwz 16,-4*16(5) 765 lwz 17,-4*15(5) 766 lwz 18,-4*14(5) 767 lwz 19,-4*13(5) 768 lwz 20,-4*12(5) 769 lwz 21,-4*11(5) 770 lwz 22,-4*10(5) 771 lwz 23,-4*9(5) 772 lwz 24,-4*8(5) 773 lwz 25,-4*7(5) 774 lwz 26,-4*6(5) 775 lwz 27,-4*5(5) 776 lwz 28,-4*4(5) 777 lwz 29,-4*3(5) 778 lwz 30,-4*2(5) 779 lwz 31,-4*1(5) 780 mr 1,5 781 blr 782 .long 0 783 .byte 0,12,4,0x20,0x80,18,6,0 784 .long 0 785 786 .align 5 787 __bn_sqr8x_mont: 788 .Lsqr8x_do: 789 mr 9,1 790 slwi 10,8,3 791 li 11,-32*4 792 sub 10,11,10 793 slwi 8,8,2 794 stwux 1,1,10 795 796 stw 14,-4*18(9) 797 stw 15,-4*17(9) 798 stw 16,-4*16(9) 799 stw 17,-4*15(9) 800 stw 18,-4*14(9) 801 stw 19,-4*13(9) 802 stw 20,-4*12(9) 803 stw 21,-4*11(9) 804 stw 22,-4*10(9) 805 stw 23,-4*9(9) 806 stw 24,-4*8(9) 807 stw 25,-4*7(9) 808 stw 26,-4*6(9) 809 stw 27,-4*5(9) 810 stw 28,-4*4(9) 811 stw 29,-4*3(9) 812 stw 30,-4*2(9) 813 stw 31,-4*1(9) 814 815 subi 4,4,4 816 subi 18,6,4 817 subi 3,3,4 818 lwz 7,0(7) 819 li 0,0 820 821 add 6,4,8 822 lwz 9,4*1(4) 823 824 lwz 10,4*2(4) 825 li 23,0 826 lwz 11,4*3(4) 827 li 24,0 828 lwz 12,4*4(4) 829 li 25,0 830 lwz 14,4*5(4) 831 li 26,0 832 lwz 15,4*6(4) 833 li 27,0 834 lwz 16,4*7(4) 835 li 28,0 836 lwzu 17,4*8(4) 837 li 29,0 838 839 addi 5,1,4*11 840 subic. 30,8,4*8 841 b .Lsqr8x_zero_start 842 843 .align 5 844 .Lsqr8x_zero: 845 subic. 30,30,4*8 846 stw 0,4*1(5) 847 stw 0,4*2(5) 848 stw 0,4*3(5) 849 stw 0,4*4(5) 850 stw 0,4*5(5) 851 stw 0,4*6(5) 852 stw 0,4*7(5) 853 stw 0,4*8(5) 854 .Lsqr8x_zero_start: 855 stw 0,4*9(5) 856 stw 0,4*10(5) 857 stw 0,4*11(5) 858 stw 0,4*12(5) 859 stw 0,4*13(5) 860 stw 0,4*14(5) 861 stw 0,4*15(5) 862 stwu 0,4*16(5) 863 bne .Lsqr8x_zero 864 865 stw 3,4*6(1) 866 stw 18,4*7(1) 867 stw 7,4*8(1) 868 stw 5,4*9(1) 869 stw 0,4*10(1) 870 addi 5,1,4*11 871 872 873 .align 5 874 .Lsqr8x_outer_loop: 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 mullw 18,10,9 905 mullw 19,11,9 906 mullw 20,12,9 907 mullw 21,14,9 908 addc 23,23,18 909 mullw 18,15,9 910 adde 24,24,19 911 mullw 19,16,9 912 adde 25,25,20 913 mullw 20,17,9 914 adde 26,26,21 915 mulhwu 21,10,9 916 adde 27,27,18 917 mulhwu 18,11,9 918 adde 28,28,19 919 mulhwu 19,12,9 920 adde 29,29,20 921 mulhwu 20,14,9 922 stw 22,4*1(5) 923 addze 22,0 924 stw 23,4*2(5) 925 addc 24,24,21 926 mulhwu 21,15,9 927 adde 25,25,18 928 mulhwu 18,16,9 929 adde 26,26,19 930 mulhwu 19,17,9 931 adde 27,27,20 932 mullw 20,11,10 933 adde 28,28,21 934 mullw 21,12,10 935 adde 29,29,18 936 mullw 18,14,10 937 adde 22,22,19 938 939 mullw 19,15,10 940 addc 25,25,20 941 mullw 20,16,10 942 adde 26,26,21 943 mullw 21,17,10 944 adde 27,27,18 945 mulhwu 18,11,10 946 adde 28,28,19 947 mulhwu 19,12,10 948 adde 29,29,20 949 mulhwu 20,14,10 950 adde 22,22,21 951 mulhwu 21,15,10 952 stw 24,4*3(5) 953 addze 23,0 954 stw 25,4*4(5) 955 addc 26,26,18 956 mulhwu 18,16,10 957 adde 27,27,19 958 mulhwu 19,17,10 959 adde 28,28,20 960 mullw 20,12,11 961 adde 29,29,21 962 mullw 21,14,11 963 adde 22,22,18 964 mullw 18,15,11 965 adde 23,23,19 966 967 mullw 19,16,11 968 addc 27,27,20 969 mullw 20,17,11 970 adde 28,28,21 971 mulhwu 21,12,11 972 adde 29,29,18 973 mulhwu 18,14,11 974 adde 22,22,19 975 mulhwu 19,15,11 976 adde 23,23,20 977 mulhwu 20,16,11 978 stw 26,4*5(5) 979 addze 24,0 980 stw 27,4*6(5) 981 addc 28,28,21 982 mulhwu 21,17,11 983 adde 29,29,18 984 mullw 18,14,12 985 adde 22,22,19 986 mullw 19,15,12 987 adde 23,23,20 988 mullw 20,16,12 989 adde 24,24,21 990 991 mullw 21,17,12 992 addc 29,29,18 993 mulhwu 18,14,12 994 adde 22,22,19 995 mulhwu 19,15,12 996 adde 23,23,20 997 mulhwu 20,16,12 998 adde 24,24,21 999 mulhwu 21,17,12 1000 stw 28,4*7(5) 1001 addze 25,0 1002 stwu 29,4*8(5) 1003 addc 22,22,18 1004 mullw 18,15,14 1005 adde 23,23,19 1006 mullw 19,16,14 1007 adde 24,24,20 1008 mullw 20,17,14 1009 adde 25,25,21 1010 1011 mulhwu 21,15,14 1012 addc 23,23,18 1013 mulhwu 18,16,14 1014 adde 24,24,19 1015 mulhwu 19,17,14 1016 adde 25,25,20 1017 mullw 20,16,15 1018 addze 26,0 1019 addc 24,24,21 1020 mullw 21,17,15 1021 adde 25,25,18 1022 mulhwu 18,16,15 1023 adde 26,26,19 1024 1025 mulhwu 19,17,15 1026 addc 25,25,20 1027 mullw 20,17,16 1028 adde 26,26,21 1029 mulhwu 21,17,16 1030 addze 27,0 1031 addc 26,26,18 1032 .long 0x7c062040 1033 adde 27,27,19 1034 1035 addc 27,27,20 1036 sub 18,6,8 1037 addze 28,0 1038 add 28,28,21 1039 1040 beq .Lsqr8x_outer_break 1041 1042 mr 7,9 1043 lwz 9,4*1(5) 1044 lwz 10,4*2(5) 1045 lwz 11,4*3(5) 1046 lwz 12,4*4(5) 1047 lwz 14,4*5(5) 1048 lwz 15,4*6(5) 1049 lwz 16,4*7(5) 1050 lwz 17,4*8(5) 1051 addc 22,22,9 1052 lwz 9,4*1(4) 1053 adde 23,23,10 1054 lwz 10,4*2(4) 1055 adde 24,24,11 1056 lwz 11,4*3(4) 1057 adde 25,25,12 1058 lwz 12,4*4(4) 1059 adde 26,26,14 1060 lwz 14,4*5(4) 1061 adde 27,27,15 1062 lwz 15,4*6(4) 1063 adde 28,28,16 1064 lwz 16,4*7(4) 1065 subi 3,4,4*7 1066 addze 29,17 1067 lwzu 17,4*8(4) 1068 1069 li 30,0 1070 b .Lsqr8x_mul 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 .align 5 1095 .Lsqr8x_mul: 1096 mullw 18,9,7 1097 addze 31,0 1098 mullw 19,10,7 1099 addi 30,30,4 1100 mullw 20,11,7 1101 andi. 30,30,4*8-1 1102 mullw 21,12,7 1103 addc 22,22,18 1104 mullw 18,14,7 1105 adde 23,23,19 1106 mullw 19,15,7 1107 adde 24,24,20 1108 mullw 20,16,7 1109 adde 25,25,21 1110 mullw 21,17,7 1111 adde 26,26,18 1112 mulhwu 18,9,7 1113 adde 27,27,19 1114 mulhwu 19,10,7 1115 adde 28,28,20 1116 mulhwu 20,11,7 1117 adde 29,29,21 1118 mulhwu 21,12,7 1119 addze 31,31 1120 stwu 22,4(5) 1121 addc 22,23,18 1122 mulhwu 18,14,7 1123 adde 23,24,19 1124 mulhwu 19,15,7 1125 adde 24,25,20 1126 mulhwu 20,16,7 1127 adde 25,26,21 1128 mulhwu 21,17,7 1129 lwzx 7,3,30 1130 adde 26,27,18 1131 adde 27,28,19 1132 adde 28,29,20 1133 adde 29,31,21 1134 1135 bne .Lsqr8x_mul 1136 1137 1138 .long 0x7c043040 1139 beq .Lsqr8x_break 1140 1141 lwz 9,4*1(5) 1142 lwz 10,4*2(5) 1143 lwz 11,4*3(5) 1144 lwz 12,4*4(5) 1145 lwz 14,4*5(5) 1146 lwz 15,4*6(5) 1147 lwz 16,4*7(5) 1148 lwz 17,4*8(5) 1149 addc 22,22,9 1150 lwz 9,4*1(4) 1151 adde 23,23,10 1152 lwz 10,4*2(4) 1153 adde 24,24,11 1154 lwz 11,4*3(4) 1155 adde 25,25,12 1156 lwz 12,4*4(4) 1157 adde 26,26,14 1158 lwz 14,4*5(4) 1159 adde 27,27,15 1160 lwz 15,4*6(4) 1161 adde 28,28,16 1162 lwz 16,4*7(4) 1163 adde 29,29,17 1164 lwzu 17,4*8(4) 1165 1166 b .Lsqr8x_mul 1167 1168 .align 5 1169 .Lsqr8x_break: 1170 lwz 9,4*8(3) 1171 addi 4,3,4*15 1172 lwz 10,4*9(3) 1173 sub. 18,6,4 1174 lwz 11,4*10(3) 1175 sub 19,5,18 1176 lwz 12,4*11(3) 1177 lwz 14,4*12(3) 1178 lwz 15,4*13(3) 1179 lwz 16,4*14(3) 1180 lwz 17,4*15(3) 1181 beq .Lsqr8x_outer_loop 1182 1183 stw 22,4*1(5) 1184 lwz 22,4*1(19) 1185 stw 23,4*2(5) 1186 lwz 23,4*2(19) 1187 stw 24,4*3(5) 1188 lwz 24,4*3(19) 1189 stw 25,4*4(5) 1190 lwz 25,4*4(19) 1191 stw 26,4*5(5) 1192 lwz 26,4*5(19) 1193 stw 27,4*6(5) 1194 lwz 27,4*6(19) 1195 stw 28,4*7(5) 1196 lwz 28,4*7(19) 1197 stw 29,4*8(5) 1198 lwz 29,4*8(19) 1199 mr 5,19 1200 b .Lsqr8x_outer_loop 1201 1202 .align 5 1203 .Lsqr8x_outer_break: 1204 1205 1206 lwz 10,4*1(18) 1207 lwz 12,4*2(18) 1208 lwz 15,4*3(18) 1209 lwz 17,4*4(18) 1210 addi 4,18,4*4 1211 1212 lwz 19,4*13(1) 1213 lwz 20,4*14(1) 1214 lwz 21,4*15(1) 1215 lwz 18,4*16(1) 1216 1217 stw 22,4*1(5) 1218 srwi 30,8,4 1219 stw 23,4*2(5) 1220 subi 30,30,1 1221 stw 24,4*3(5) 1222 stw 25,4*4(5) 1223 stw 26,4*5(5) 1224 stw 27,4*6(5) 1225 stw 28,4*7(5) 1226 1227 addi 5,1,4*11 1228 mullw 22,10,10 1229 mulhwu 10,10,10 1230 add 23,19,19 1231 srwi 19,19,32-1 1232 mullw 11,12,12 1233 mulhwu 12,12,12 1234 addc 23,23,10 1235 add 24,20,20 1236 srwi 20,20,32-1 1237 add 25,21,21 1238 srwi 21,21,32-1 1239 or 24,24,19 1240 1241 mtctr 30 1242 .Lsqr4x_shift_n_add: 1243 mullw 14,15,15 1244 mulhwu 15,15,15 1245 lwz 19,4*6(5) 1246 lwz 10,4*1(4) 1247 adde 24,24,11 1248 add 26,18,18 1249 srwi 18,18,32-1 1250 or 25,25,20 1251 lwz 20,4*7(5) 1252 adde 25,25,12 1253 lwz 12,4*2(4) 1254 add 27,19,19 1255 srwi 19,19,32-1 1256 or 26,26,21 1257 lwz 21,4*8(5) 1258 mullw 16,17,17 1259 mulhwu 17,17,17 1260 adde 26,26,14 1261 add 28,20,20 1262 srwi 20,20,32-1 1263 or 27,27,18 1264 lwz 18,4*9(5) 1265 adde 27,27,15 1266 lwz 15,4*3(4) 1267 add 29,21,21 1268 srwi 21,21,32-1 1269 or 28,28,19 1270 lwz 19,4*10(5) 1271 mullw 9,10,10 1272 mulhwu 10,10,10 1273 adde 28,28,16 1274 stw 22,4*1(5) 1275 add 22,18,18 1276 srwi 18,18,32-1 1277 or 29,29,20 1278 lwz 20,4*11(5) 1279 adde 29,29,17 1280 lwzu 17,4*4(4) 1281 stw 23,4*2(5) 1282 add 23,19,19 1283 srwi 19,19,32-1 1284 or 22,22,21 1285 lwz 21,4*12(5) 1286 mullw 11,12,12 1287 mulhwu 12,12,12 1288 adde 22,22,9 1289 stw 24,4*3(5) 1290 add 24,20,20 1291 srwi 20,20,32-1 1292 or 23,23,18 1293 lwz 18,4*13(5) 1294 adde 23,23,10 1295 stw 25,4*4(5) 1296 stw 26,4*5(5) 1297 stw 27,4*6(5) 1298 stw 28,4*7(5) 1299 stwu 29,4*8(5) 1300 add 25,21,21 1301 srwi 21,21,32-1 1302 or 24,24,19 1303 bc 16,0,.Lsqr4x_shift_n_add 1304 lwz 4,4*7(1) 1305 lwz 7,4*8(1) 1306 1307 mullw 14,15,15 1308 mulhwu 15,15,15 1309 stw 22,4*1(5) 1310 lwz 22,4*12(1) 1311 lwz 19,4*6(5) 1312 adde 24,24,11 1313 add 26,18,18 1314 srwi 18,18,32-1 1315 or 25,25,20 1316 lwz 20,4*7(5) 1317 adde 25,25,12 1318 add 27,19,19 1319 srwi 19,19,32-1 1320 or 26,26,21 1321 mullw 16,17,17 1322 mulhwu 17,17,17 1323 adde 26,26,14 1324 add 28,20,20 1325 srwi 20,20,32-1 1326 or 27,27,18 1327 stw 23,4*2(5) 1328 lwz 23,4*13(1) 1329 adde 27,27,15 1330 or 28,28,19 1331 lwz 9,4*1(4) 1332 lwz 10,4*2(4) 1333 adde 28,28,16 1334 lwz 11,4*3(4) 1335 lwz 12,4*4(4) 1336 adde 29,17,20 1337 lwz 14,4*5(4) 1338 lwz 15,4*6(4) 1339 1340 1341 1342 mullw 31,7,22 1343 li 30,8 1344 lwz 16,4*7(4) 1345 add 6,4,8 1346 lwzu 17,4*8(4) 1347 stw 24,4*3(5) 1348 lwz 24,4*14(1) 1349 stw 25,4*4(5) 1350 lwz 25,4*15(1) 1351 stw 26,4*5(5) 1352 lwz 26,4*16(1) 1353 stw 27,4*6(5) 1354 lwz 27,4*17(1) 1355 stw 28,4*7(5) 1356 lwz 28,4*18(1) 1357 stw 29,4*8(5) 1358 lwz 29,4*19(1) 1359 addi 5,1,4*11 1360 mtctr 30 1361 b .Lsqr8x_reduction 1362 1363 .align 5 1364 .Lsqr8x_reduction: 1365 1366 mullw 19,10,31 1367 mullw 20,11,31 1368 stwu 31,4(5) 1369 mullw 21,12,31 1370 1371 addic 22,22,-1 1372 mullw 18,14,31 1373 adde 22,23,19 1374 mullw 19,15,31 1375 adde 23,24,20 1376 mullw 20,16,31 1377 adde 24,25,21 1378 mullw 21,17,31 1379 adde 25,26,18 1380 mulhwu 18,9,31 1381 adde 26,27,19 1382 mulhwu 19,10,31 1383 adde 27,28,20 1384 mulhwu 20,11,31 1385 adde 28,29,21 1386 mulhwu 21,12,31 1387 addze 29,0 1388 addc 22,22,18 1389 mulhwu 18,14,31 1390 adde 23,23,19 1391 mulhwu 19,15,31 1392 adde 24,24,20 1393 mulhwu 20,16,31 1394 adde 25,25,21 1395 mulhwu 21,17,31 1396 mullw 31,7,22 1397 adde 26,26,18 1398 adde 27,27,19 1399 adde 28,28,20 1400 adde 29,29,21 1401 bc 16,0,.Lsqr8x_reduction 1402 1403 lwz 18,4*1(5) 1404 lwz 19,4*2(5) 1405 lwz 20,4*3(5) 1406 lwz 21,4*4(5) 1407 subi 3,5,4*7 1408 .long 0x7c062040 1409 addc 22,22,18 1410 lwz 18,4*5(5) 1411 adde 23,23,19 1412 lwz 19,4*6(5) 1413 adde 24,24,20 1414 lwz 20,4*7(5) 1415 adde 25,25,21 1416 lwz 21,4*8(5) 1417 adde 26,26,18 1418 adde 27,27,19 1419 adde 28,28,20 1420 adde 29,29,21 1421 1422 beq .Lsqr8x8_post_condition 1423 1424 lwz 7,4*0(3) 1425 lwz 9,4*1(4) 1426 lwz 10,4*2(4) 1427 lwz 11,4*3(4) 1428 lwz 12,4*4(4) 1429 lwz 14,4*5(4) 1430 lwz 15,4*6(4) 1431 lwz 16,4*7(4) 1432 lwzu 17,4*8(4) 1433 li 30,0 1434 1435 .align 5 1436 .Lsqr8x_tail: 1437 mullw 18,9,7 1438 addze 31,0 1439 mullw 19,10,7 1440 addi 30,30,4 1441 mullw 20,11,7 1442 andi. 30,30,4*8-1 1443 mullw 21,12,7 1444 addc 22,22,18 1445 mullw 18,14,7 1446 adde 23,23,19 1447 mullw 19,15,7 1448 adde 24,24,20 1449 mullw 20,16,7 1450 adde 25,25,21 1451 mullw 21,17,7 1452 adde 26,26,18 1453 mulhwu 18,9,7 1454 adde 27,27,19 1455 mulhwu 19,10,7 1456 adde 28,28,20 1457 mulhwu 20,11,7 1458 adde 29,29,21 1459 mulhwu 21,12,7 1460 addze 31,31 1461 stwu 22,4(5) 1462 addc 22,23,18 1463 mulhwu 18,14,7 1464 adde 23,24,19 1465 mulhwu 19,15,7 1466 adde 24,25,20 1467 mulhwu 20,16,7 1468 adde 25,26,21 1469 mulhwu 21,17,7 1470 lwzx 7,3,30 1471 adde 26,27,18 1472 adde 27,28,19 1473 adde 28,29,20 1474 adde 29,31,21 1475 1476 bne .Lsqr8x_tail 1477 1478 1479 lwz 9,4*1(5) 1480 lwz 31,4*10(1) 1481 .long 0x7c062040 1482 lwz 10,4*2(5) 1483 sub 20,6,8 1484 lwz 11,4*3(5) 1485 lwz 12,4*4(5) 1486 lwz 14,4*5(5) 1487 lwz 15,4*6(5) 1488 lwz 16,4*7(5) 1489 lwz 17,4*8(5) 1490 beq .Lsqr8x_tail_break 1491 1492 addc 22,22,9 1493 lwz 9,4*1(4) 1494 adde 23,23,10 1495 lwz 10,4*2(4) 1496 adde 24,24,11 1497 lwz 11,4*3(4) 1498 adde 25,25,12 1499 lwz 12,4*4(4) 1500 adde 26,26,14 1501 lwz 14,4*5(4) 1502 adde 27,27,15 1503 lwz 15,4*6(4) 1504 adde 28,28,16 1505 lwz 16,4*7(4) 1506 adde 29,29,17 1507 lwzu 17,4*8(4) 1508 1509 b .Lsqr8x_tail 1510 1511 .align 5 1512 .Lsqr8x_tail_break: 1513 lwz 7,4*8(1) 1514 lwz 21,4*9(1) 1515 addi 30,5,4*8 1516 1517 addic 31,31,-1 1518 adde 18,22,9 1519 lwz 22,4*8(3) 1520 lwz 9,4*1(20) 1521 adde 19,23,10 1522 lwz 23,4*9(3) 1523 lwz 10,4*2(20) 1524 adde 24,24,11 1525 lwz 11,4*3(20) 1526 adde 25,25,12 1527 lwz 12,4*4(20) 1528 adde 26,26,14 1529 lwz 14,4*5(20) 1530 adde 27,27,15 1531 lwz 15,4*6(20) 1532 adde 28,28,16 1533 lwz 16,4*7(20) 1534 adde 29,29,17 1535 lwz 17,4*8(20) 1536 addi 4,20,4*8 1537 addze 20,0 1538 mullw 31,7,22 1539 stw 18,4*1(5) 1540 .long 0x7c1ea840 1541 stw 19,4*2(5) 1542 li 30,8 1543 stw 24,4*3(5) 1544 lwz 24,4*10(3) 1545 stw 25,4*4(5) 1546 lwz 25,4*11(3) 1547 stw 26,4*5(5) 1548 lwz 26,4*12(3) 1549 stw 27,4*6(5) 1550 lwz 27,4*13(3) 1551 stw 28,4*7(5) 1552 lwz 28,4*14(3) 1553 stw 29,4*8(5) 1554 lwz 29,4*15(3) 1555 stw 20,4*10(1) 1556 addi 5,3,4*7 1557 mtctr 30 1558 bne .Lsqr8x_reduction 1559 1560 1561 1562 1563 1564 1565 lwz 3,4*6(1) 1566 srwi 30,8,5 1567 mr 7,5 1568 addi 5,5,4*8 1569 subi 30,30,1 1570 subfc 18,9,22 1571 subfe 19,10,23 1572 mr 31,20 1573 mr 6,3 1574 1575 mtctr 30 1576 b .Lsqr8x_sub 1577 1578 .align 5 1579 .Lsqr8x_sub: 1580 lwz 9,4*1(4) 1581 lwz 22,4*1(5) 1582 lwz 10,4*2(4) 1583 lwz 23,4*2(5) 1584 subfe 20,11,24 1585 lwz 11,4*3(4) 1586 lwz 24,4*3(5) 1587 subfe 21,12,25 1588 lwz 12,4*4(4) 1589 lwz 25,4*4(5) 1590 stw 18,4*1(3) 1591 subfe 18,14,26 1592 lwz 14,4*5(4) 1593 lwz 26,4*5(5) 1594 stw 19,4*2(3) 1595 subfe 19,15,27 1596 lwz 15,4*6(4) 1597 lwz 27,4*6(5) 1598 stw 20,4*3(3) 1599 subfe 20,16,28 1600 lwz 16,4*7(4) 1601 lwz 28,4*7(5) 1602 stw 21,4*4(3) 1603 subfe 21,17,29 1604 lwzu 17,4*8(4) 1605 lwzu 29,4*8(5) 1606 stw 18,4*5(3) 1607 subfe 18,9,22 1608 stw 19,4*6(3) 1609 subfe 19,10,23 1610 stw 20,4*7(3) 1611 stwu 21,4*8(3) 1612 bc 16,0,.Lsqr8x_sub 1613 1614 srwi 30,8,4 1615 lwz 9,4*1(6) 1616 lwz 22,4*1(7) 1617 subi 30,30,1 1618 lwz 10,4*2(6) 1619 lwz 23,4*2(7) 1620 subfe 20,11,24 1621 lwz 11,4*3(6) 1622 lwz 24,4*3(7) 1623 subfe 21,12,25 1624 lwz 12,4*4(6) 1625 lwzu 25,4*4(7) 1626 stw 18,4*1(3) 1627 subfe 18,14,26 1628 stw 19,4*2(3) 1629 subfe 19,15,27 1630 stw 20,4*3(3) 1631 subfe 20,16,28 1632 stw 21,4*4(3) 1633 subfe 21,17,29 1634 stw 18,4*5(3) 1635 subfe 31,0,31 1636 stw 19,4*6(3) 1637 stw 20,4*7(3) 1638 stw 21,4*8(3) 1639 1640 addi 5,1,4*11 1641 mtctr 30 1642 1643 .Lsqr4x_cond_copy: 1644 andc 9,9,31 1645 stw 0,-4*3(7) 1646 and 22,22,31 1647 stw 0,-4*2(7) 1648 andc 10,10,31 1649 stw 0,-4*1(7) 1650 and 23,23,31 1651 stw 0,-4*0(7) 1652 andc 11,11,31 1653 stw 0,4*1(5) 1654 and 24,24,31 1655 stw 0,4*2(5) 1656 andc 12,12,31 1657 stw 0,4*3(5) 1658 and 25,25,31 1659 stwu 0,4*4(5) 1660 or 18,9,22 1661 lwz 9,4*5(6) 1662 lwz 22,4*1(7) 1663 or 19,10,23 1664 lwz 10,4*6(6) 1665 lwz 23,4*2(7) 1666 or 20,11,24 1667 lwz 11,4*7(6) 1668 lwz 24,4*3(7) 1669 or 21,12,25 1670 lwz 12,4*8(6) 1671 lwzu 25,4*4(7) 1672 stw 18,4*1(6) 1673 stw 19,4*2(6) 1674 stw 20,4*3(6) 1675 stwu 21,4*4(6) 1676 bc 16,0,.Lsqr4x_cond_copy 1677 1678 lwz 4,0(1) 1679 andc 9,9,31 1680 and 22,22,31 1681 andc 10,10,31 1682 and 23,23,31 1683 andc 11,11,31 1684 and 24,24,31 1685 andc 12,12,31 1686 and 25,25,31 1687 or 18,9,22 1688 or 19,10,23 1689 or 20,11,24 1690 or 21,12,25 1691 stw 18,4*1(6) 1692 stw 19,4*2(6) 1693 stw 20,4*3(6) 1694 stw 21,4*4(6) 1695 1696 b .Lsqr8x_done 1697 1698 .align 5 1699 .Lsqr8x8_post_condition: 1700 lwz 3,4*6(1) 1701 lwz 4,0(1) 1702 addze 31,0 1703 1704 1705 subfc 22,9,22 1706 subfe 23,10,23 1707 stw 0,4*12(1) 1708 stw 0,4*13(1) 1709 subfe 24,11,24 1710 stw 0,4*14(1) 1711 stw 0,4*15(1) 1712 subfe 25,12,25 1713 stw 0,4*16(1) 1714 stw 0,4*17(1) 1715 subfe 26,14,26 1716 stw 0,4*18(1) 1717 stw 0,4*19(1) 1718 subfe 27,15,27 1719 stw 0,4*20(1) 1720 stw 0,4*21(1) 1721 subfe 28,16,28 1722 stw 0,4*22(1) 1723 stw 0,4*23(1) 1724 subfe 29,17,29 1725 stw 0,4*24(1) 1726 stw 0,4*25(1) 1727 subfe 31,0,31 1728 stw 0,4*26(1) 1729 stw 0,4*27(1) 1730 1731 and 9,9,31 1732 and 10,10,31 1733 addc 22,22,9 1734 and 11,11,31 1735 adde 23,23,10 1736 and 12,12,31 1737 adde 24,24,11 1738 and 14,14,31 1739 adde 25,25,12 1740 and 15,15,31 1741 adde 26,26,14 1742 and 16,16,31 1743 adde 27,27,15 1744 and 17,17,31 1745 adde 28,28,16 1746 adde 29,29,17 1747 stw 22,4*1(3) 1748 stw 23,4*2(3) 1749 stw 24,4*3(3) 1750 stw 25,4*4(3) 1751 stw 26,4*5(3) 1752 stw 27,4*6(3) 1753 stw 28,4*7(3) 1754 stw 29,4*8(3) 1755 1756 .Lsqr8x_done: 1757 stw 0,4*8(1) 1758 stw 0,4*10(1) 1759 1760 lwz 14,-4*18(4) 1761 li 3,1 1762 lwz 15,-4*17(4) 1763 lwz 16,-4*16(4) 1764 lwz 17,-4*15(4) 1765 lwz 18,-4*14(4) 1766 lwz 19,-4*13(4) 1767 lwz 20,-4*12(4) 1768 lwz 21,-4*11(4) 1769 lwz 22,-4*10(4) 1770 lwz 23,-4*9(4) 1771 lwz 24,-4*8(4) 1772 lwz 25,-4*7(4) 1773 lwz 26,-4*6(4) 1774 lwz 27,-4*5(4) 1775 lwz 28,-4*4(4) 1776 lwz 29,-4*3(4) 1777 lwz 30,-4*2(4) 1778 lwz 31,-4*1(4) 1779 mr 1,4 1780 blr 1781 .long 0 1782 .byte 0,12,4,0x20,0x80,18,6,0 1783 .long 0 1784 1785 .byte 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,80,80,67,44,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 1786 .align 2 1787
Indexes created Sat Feb 28 05:31:39 UTC 2026