Lines Matching refs:movaps
246 movaps %8, [mask_mpmppmpm]
247 movaps %7, [s16_perm]
252 movaps %7, [s16_perm]
343 movaps %13, [mask_pmpmpmpm] ; "subaddps? pfft, who needs that!"
347 movaps %13, [mask_pmpmpmpm] ; "subaddps? pfft, who needs that!"
400 movaps %10, [mask_pmpmpmpm]
405 movaps %10, [mask_pmpmpmpm]
475 movaps [outq + 0*mmsize], m0
476 movaps [outq + 4*mmsize], m1
477 movaps [outq + 8*mmsize], tx1_e0
478 movaps [outq + 12*mmsize], tx2_e0
482 movaps [outq + 2*mmsize], m2
483 movaps [outq + 6*mmsize], m3
484 movaps [outq + 10*mmsize], tx1_o0
485 movaps [outq + 14*mmsize], tx2_o0
487 movaps tw_e, [tab_64_float + mmsize]
490 movaps m0, [outq + 1*mmsize]
491 movaps m1, [outq + 3*mmsize]
492 movaps m2, [outq + 5*mmsize]
493 movaps m3, [outq + 7*mmsize]
498 movaps [outq + 1*mmsize], m0
499 movaps [outq + 3*mmsize], m1
500 movaps [outq + 5*mmsize], m2
501 movaps [outq + 7*mmsize], m3
503 movaps [outq + 9*mmsize], tx1_e1
504 movaps [outq + 11*mmsize], tx1_o1
505 movaps [outq + 13*mmsize], tx2_e1
506 movaps [outq + 15*mmsize], tx2_o1
514 movaps m8, [rtabq + (%5)*mmsize + %7]
517 movaps m0, [outq + (0 + %4)*mmsize + %6]
518 movaps m2, [outq + (2 + %4)*mmsize + %6]
519 movaps m1, [outq + %1 + (0 + %4)*mmsize + %6]
520 movaps m3, [outq + %1 + (2 + %4)*mmsize + %6]
522 movaps m4, [outq + %2 + (0 + %4)*mmsize + %6]
523 movaps m6, [outq + %2 + (2 + %4)*mmsize + %6]
524 movaps m5, [outq + %3 + (0 + %4)*mmsize + %6]
525 movaps m7, [outq + %3 + (2 + %4)*mmsize + %6]
532 movaps [outq + (0 + %4)*mmsize + %6], m0
533 movaps [outq + (2 + %4)*mmsize + %6], m2
534 movaps [outq + %1 + (0 + %4)*mmsize + %6], m1
535 movaps [outq + %1 + (2 + %4)*mmsize + %6], m3
537 movaps [outq + %2 + (0 + %4)*mmsize + %6], m4
538 movaps [outq + %2 + (2 + %4)*mmsize + %6], m6
539 movaps [outq + %3 + (0 + %4)*mmsize + %6], m5
540 movaps [outq + %3 + (2 + %4)*mmsize + %6], m7
571 movaps m8, [rtabq + (0 + %2)*mmsize]
574 movaps m0, [outq + (0 + 0 + %1)*mmsize + %6]
575 movaps m2, [outq + (2 + 0 + %1)*mmsize + %6]
576 movaps m1, [outq + %3 + (0 + 0 + %1)*mmsize + %6]
577 movaps m3, [outq + %3 + (2 + 0 + %1)*mmsize + %6]
579 movaps m4, [outq + %4 + (0 + 0 + %1)*mmsize + %6]
580 movaps m6, [outq + %4 + (2 + 0 + %1)*mmsize + %6]
581 movaps m5, [outq + %5 + (0 + 0 + %1)*mmsize + %6]
582 movaps m7, [outq + %5 + (2 + 0 + %1)*mmsize + %6]
613 movaps m8, [rtabq + (1 + %2)*mmsize]
616 movaps m0, [outq + (0 + 1 + %1)*mmsize + %6]
617 movaps m2, [outq + (2 + 1 + %1)*mmsize + %6]
618 movaps m1, [outq + %3 + (0 + 1 + %1)*mmsize + %6]
619 movaps m3, [outq + %3 + (2 + 1 + %1)*mmsize + %6]
621 movaps [outq + (0 + 1 + %1)*mmsize + %6], m10 ; m0 conflict
622 movaps [outq + %3 + (0 + 1 + %1)*mmsize + %6], m11 ; m1 conflict
624 movaps m4, [outq + %4 + (0 + 1 + %1)*mmsize + %6]
625 movaps m6, [outq + %4 + (2 + 1 + %1)*mmsize + %6]
626 movaps m5, [outq + %5 + (0 + 1 + %1)*mmsize + %6]
627 movaps m7, [outq + %5 + (2 + 1 + %1)*mmsize + %6]
629 movaps [outq + %4 + (0 + 1 + %1)*mmsize + %6], m12 ; m4 conflict
630 movaps [outq + %5 + (0 + 1 + %1)*mmsize + %6], m13 ; m5 conflict
679 movaps m0, [inq]
681 movaps [outq], m0
687 movaps m0, [inq + 0*mmsize]
688 movaps m1, [inq + 1*mmsize]
693 movaps m1, m2
701 movaps [outq + 0*mmsize], m2
702 movaps [outq + 1*mmsize], m0
720 movaps m0, [inq + 0*mmsize]
721 movaps m1, [inq + 1*mmsize]
722 movaps m2, [inq + 2*mmsize]
723 movaps m3, [inq + 3*mmsize]
752 movaps m0, [inq + 0*mmsize]
753 movaps m1, [inq + 1*mmsize]
777 movaps m0, [inq + 0*mmsize]
778 movaps m1, [inq + 1*mmsize]
779 movaps m2, [inq + 2*mmsize]
780 movaps m3, [inq + 3*mmsize]
817 movaps m4, [inq + 4*mmsize]
818 movaps m5, [inq + 5*mmsize]
819 movaps m6, [inq + 6*mmsize]
820 movaps m7, [inq + 7*mmsize]
832 movaps m0, [inq + 0*mmsize]
833 movaps m1, [inq + 1*mmsize]
834 movaps m2, [inq + 2*mmsize]
835 movaps m3, [inq + 3*mmsize]
843 movaps m8, [tab_32_float]
937 movaps m4, [inq + 4*mmsize]
938 movaps m5, [inq + 5*mmsize]
939 movaps m6, [inq + 6*mmsize]
940 movaps m7, [inq + 7*mmsize]
951 movaps m0, [inq + 0*mmsize]
952 movaps m1, [inq + 1*mmsize]
953 movaps m2, [inq + 2*mmsize]
954 movaps m3, [inq + 3*mmsize]
962 movaps m8, [tab_32_float]
970 movaps [outq + 1*mmsize], m1
971 movaps [outq + 3*mmsize], m3
972 movaps [outq + 5*mmsize], m5
973 movaps [outq + 7*mmsize], m7
983 movaps [outq + 0*mmsize], m0
984 movaps [outq + 2*mmsize], m2
985 movaps [outq + 4*mmsize], m4
986 movaps [outq + 6*mmsize], m6
1011 movaps tx1_e0, [inq + 0*mmsize]
1012 movaps tx1_e1, [inq + 1*mmsize]
1013 movaps tx1_o0, [inq + 2*mmsize]
1014 movaps tx1_o1, [inq + 3*mmsize]
1025 movaps tx2_e0, [inq + 4*mmsize]
1026 movaps tx2_e1, [inq + 5*mmsize]
1027 movaps tx2_o0, [inq + 6*mmsize]
1028 movaps tx2_o1, [inq + 7*mmsize]
1038 movaps tw_e, [tab_64_float]
1241 movaps tw_e, [tab_64_float + mmsize]
1244 movaps m0, [outq + 1*mmsize]
1245 movaps m1, [outq + 3*mmsize]
1246 movaps m2, [outq + 5*mmsize]
1247 movaps m3, [outq + 7*mmsize]
1249 movaps [outq + 1*mmsize], tmp1
1250 movaps [outq + 5*mmsize], tmp2