Lines Matching refs:m0
88 mova m0, [aq]
91 mova [dstq+strideq*0], m0
92 mova [dstq+strideq*1], m0
93 mova [dstq+strideq*2], m0
94 mova [dstq+stride3q ], m0
100 mova m0, [aq]
103 mova [dstq+strideq*0], m0
104 mova [dstq+strideq*1], m0
105 mova [dstq+strideq*2], m0
106 mova [dstq+stride3q ], m0
108 mova [dstq+strideq*0], m0
109 mova [dstq+strideq*1], m0
110 mova [dstq+strideq*2], m0
111 mova [dstq+stride3q ], m0
117 mova m0, [aq]
123 mova [dstq+strideq*0+ 0], m0
125 mova [dstq+strideq*1+ 0], m0
127 mova [dstq+strideq*2+ 0], m0
129 mova [dstq+stride3q + 0], m0
139 mova m0, [aq+mmsize*0]
146 mova [dstq+strideq*0+ 0], m0
150 mova [dstq+strideq*1+ 0], m0
164 pshufw m0, m3, q3333
168 mova [dstq+strideq*0], m0
180 pshufd m0, m3, q3333
182 mova [dstq+strideq*0], m0
184 pshufd m0, m3, q1111
186 mova [dstq+strideq*2], m0
190 pshufd m0, m2, q3333
192 mova [dstq+strideq*0], m0
194 pshufd m0, m2, q1111
196 mova [dstq+strideq*2], m0
207 pshufd m0, m3, q3333
211 mova [dstq+strideq*0+ 0], m0
212 mova [dstq+strideq*0+16], m0
231 pshufd m0, m3, q3333
235 mova [dstq+strideq*0+ 0], m0
236 mova [dstq+strideq*0+16], m0
237 mova [dstq+strideq*0+32], m0
238 mova [dstq+strideq*0+48], m0
258 mova m0, [lq]
259 paddw m0, [aq]
262 pmaddwd m0, [pw_1]
263 pshufw m1, m0, q3232
264 paddd m0, [pd_4]
265 paddd m0, m1
266 psrad m0, 3
267 pshufw m0, m0, q0000
268 mova [dstq+strideq*0], m0
269 mova [dstq+strideq*1], m0
270 mova [dstq+strideq*2], m0
271 mova [dstq+stride3q ], m0
276 mova m0, [lq]
277 paddw m0, [aq]
280 pmaddwd m0, [pw_1]
281 pshufd m1, m0, q3232
282 paddd m0, m1
283 pshufd m1, m0, q1111
284 paddd m0, [pd_8]
285 paddd m0, m1
286 psrad m0, 4
287 pshuflw m0, m0, q0000
288 punpcklqdq m0, m0
289 mova [dstq+strideq*0], m0
290 mova [dstq+strideq*1], m0
291 mova [dstq+strideq*2], m0
292 mova [dstq+stride3q ], m0
294 mova [dstq+strideq*0], m0
295 mova [dstq+strideq*1], m0
296 mova [dstq+strideq*2], m0
297 mova [dstq+stride3q ], m0
302 mova m0, [lq]
303 paddw m0, [lq+mmsize]
304 paddw m0, [aq]
305 paddw m0, [aq+mmsize]
309 pmaddwd m0, [pw_1]
310 pshufd m1, m0, q3232
311 paddd m0, m1
312 pshufd m1, m0, q1111
313 paddd m0, [pd_16]
314 paddd m0, m1
315 psrad m0, 5
316 pshuflw m0, m0, q0000
317 punpcklqdq m0, m0
319 mova [dstq+strideq*0+ 0], m0
320 mova [dstq+strideq*0+16], m0
321 mova [dstq+strideq*1+ 0], m0
322 mova [dstq+strideq*1+16], m0
323 mova [dstq+strideq*2+ 0], m0
324 mova [dstq+strideq*2+16], m0
325 mova [dstq+stride3q + 0], m0
326 mova [dstq+stride3q +16], m0
334 mova m0, [lq+mmsize*0]
335 paddw m0, [lq+mmsize*1]
336 paddw m0, [lq+mmsize*2]
337 paddw m0, [lq+mmsize*3]
338 paddw m0, [aq+mmsize*0]
339 paddw m0, [aq+mmsize*1]
340 paddw m0, [aq+mmsize*2]
341 paddw m0, [aq+mmsize*3]
345 pmaddwd m0, [pw_1]
346 pshufd m1, m0, q3232
347 paddd m0, m1
348 pshufd m1, m0, q1111
349 paddd m0, [pd_32]
350 paddd m0, m1
351 psrad m0, 6
352 pshuflw m0, m0, q0000
353 punpcklqdq m0, m0
355 mova [dstq+strideq*0+ 0], m0
356 mova [dstq+strideq*0+16], m0
357 mova [dstq+strideq*0+32], m0
358 mova [dstq+strideq*0+48], m0
359 mova [dstq+strideq*1+ 0], m0
360 mova [dstq+strideq*1+16], m0
361 mova [dstq+strideq*1+32], m0
362 mova [dstq+strideq*1+48], m0
371 mova m0, [%2]
374 pmaddwd m0, [pw_1]
375 pshufw m1, m0, q3232
376 paddd m0, [pd_2]
377 paddd m0, m1
378 psrad m0, 2
379 pshufw m0, m0, q0000
380 mova [dstq+strideq*0], m0
381 mova [dstq+strideq*1], m0
382 mova [dstq+strideq*2], m0
383 mova [dstq+stride3q ], m0
388 mova m0, [%2]
391 pmaddwd m0, [pw_1]
392 pshufd m1, m0, q3232
393 paddd m0, m1
394 pshufd m1, m0, q1111
395 paddd m0, [pd_4]
396 paddd m0, m1
397 psrad m0, 3
398 pshuflw m0, m0, q0000
399 punpcklqdq m0, m0
400 mova [dstq+strideq*0], m0
401 mova [dstq+strideq*1], m0
402 mova [dstq+strideq*2], m0
403 mova [dstq+stride3q ], m0
405 mova [dstq+strideq*0], m0
406 mova [dstq+strideq*1], m0
407 mova [dstq+strideq*2], m0
408 mova [dstq+stride3q ], m0
413 mova m0, [%2]
414 paddw m0, [%2+mmsize]
418 pmaddwd m0, [pw_1]
419 pshufd m1, m0, q3232
420 paddd m0, m1
421 pshufd m1, m0, q1111
422 paddd m0, [pd_8]
423 paddd m0, m1
424 psrad m0, 4
425 pshuflw m0, m0, q0000
426 punpcklqdq m0, m0
428 mova [dstq+strideq*0+ 0], m0
429 mova [dstq+strideq*0+16], m0
430 mova [dstq+strideq*1+ 0], m0
431 mova [dstq+strideq*1+16], m0
432 mova [dstq+strideq*2+ 0], m0
433 mova [dstq+strideq*2+16], m0
434 mova [dstq+stride3q + 0], m0
435 mova [dstq+stride3q +16], m0
443 mova m0, [%2+mmsize*0]
444 paddw m0, [%2+mmsize*1]
445 paddw m0, [%2+mmsize*2]
446 paddw m0, [%2+mmsize*3]
449 pmaddwd m0, [pw_1]
450 pshufd m1, m0, q3232
451 paddd m0, m1
452 pshufd m1, m0, q1111
453 paddd m0, [pd_16]
454 paddd m0, m1
455 psrad m0, 5
456 pshuflw m0, m0, q0000
457 punpcklqdq m0, m0
459 mova [dstq+strideq*0+ 0], m0
460 mova [dstq+strideq*0+16], m0
461 mova [dstq+strideq*0+32], m0
462 mova [dstq+strideq*0+48], m0
463 mova [dstq+strideq*1+ 0], m0
464 mova [dstq+strideq*1+16], m0
465 mova [dstq+strideq*1+32], m0
466 mova [dstq+strideq*1+48], m0
482 movd m0, [aq-4]
483 pshufw m0, m0, q1111
484 psubw m4, m0
487 pshufw m0, m3, q3333
491 paddw m0, m4
496 pmaxsw m0, m4
500 pminsw m0, m5
504 mova [dstq+strideq*0], m0
520 movd m0, [aq-4]
521 pshuflw m0, m0, q1111
522 punpcklqdq m0, m0
523 psubw m5, m0
530 pshufd m0, m3, q3333
534 paddw m0, m5
538 pmaxsw m0, m6
542 pminsw m0, m4
546 mova [dstq+strideq*0], m0
566 movd m0, [aq-4]
567 pshuflw m0, m0, q1111
568 punpcklqdq m0, m0
569 psubw m4, m0
570 psubw m5, m0
578 paddw m0, m2, m4
582 pmaxsw m0, m6
586 pminsw m0, m7
590 mova [dstq+strideq*0+ 0], m0
605 mova m0, [pw_1023]
614 mova [rsp+ 0], m0
624 movd m0, [aq-4]
625 pshuflw m0, m0, q1111
626 punpcklqdq m0, m0
627 psubw m4, m0
628 psubw m5, m0
629 psubw m6, m0
630 psubw m7, m0
637 paddw m0, m3, m4
641 pmaxsw m0, reg_min
645 pminsw m0, reg_max
649 mova [dstq+strideq*0+ 0], m0
659 mova m0, [pw_4095]
705 pshufhw m0, m1, q3310 ; abcdefhh
709 pshufd m1, m0, q3321 ; DEFGh...
710 movh [dstq+strideq*0], m0
713 psrldq m0, 2 ; CDEFGh..
715 movh [dstq+strideq*0], m0
721 mova m0, [aq] ; abcdefgh
725 SHIFT_RIGHTx2 m1, m2, m0, m4 ; bcdefghh/cdefghhh
727 shufps m1, m0, m2, q3332 ; FGHhhhhh
728 shufps m3, m0, m1, q2121 ; DEFGHhhh
732 mova [dstq+strideq*0], m0
734 SHIFT_RIGHT m0, m0, m4 ; CDEFGHhh
736 pshufd m2, m0, q3321 ; EFGHhhhh
737 mova [dstq+strideq*1], m0
750 mova m0, [aq] ; abcdefgh
752 PALIGNR m1, m3, m0, 2, m4 ; bcdefghi
753 PALIGNR m2, m3, m0, 4, m4 ; cdefghij
765 mova [dstq+strideq*0+ 0], m0
771 vpalignr m0, m1, m0, 2
773 PALIGNR m3, m1, m0, 2, m4
774 mova m0, m3
783 mova m0, [aq+mmsize*0] ; abcdefgh
787 PALIGNR m4, m1, m0, 2, m6
788 PALIGNR m5, m1, m0, 4, m6
808 mova [dstq+stride8q*0+ 0], m0
826 vpalignr m0, m1, m0, 2
830 PALIGNR m5, m1, m0, 2, m6
831 mova m0, m5
854 mova m0, [aq] ; abcdefghijklmnop
856 vperm2i128 m2, m0, m1, q0201 ; ijklmnoppppppppp
857 vpalignr m3, m2, m0, 2 ; bcdefghijklmnopp
858 vpalignr m4, m2, m0, 4 ; cdefghijklmnoppp
860 vperm2i128 m2, m0, m1, q0201 ; JKLMNOPppppppppp
866 mova [dstq+strideq*0], m0
867 vpalignr m3, m2, m0, 2
868 vpalignr m4, m2, m0, 4
871 vpalignr m3, m2, m0, 6
872 vpalignr m4, m2, m0, 8
876 vpalignr m3, m2, m0, 10
877 vpalignr m4, m2, m0, 12
880 vpalignr m3, m2, m0, 14
883 mova m0, m2
891 mova m0, [aq+mmsize*0+ 0] ; abcdefghijklmnop
894 vperm2i128 m5, m0, m1, q0201 ; ijklmnopqrstuvwx
895 vpalignr m2, m5, m0, 2 ; bcdefghijklmnopq
896 vpalignr m3, m5, m0, 4 ; cdefghijklmnopqr
903 vperm2i128 m5, m0, m1, q0201 ; JKLMNOPQRSTUVWXY
909 mova [dstq+strideq*0 + 0], m0
911 vpalignr m3, m5, m0, 2
915 vpalignr m3, m5, m0, 4
919 vpalignr m3, m5, m0, 6
924 vpalignr m3, m5, m0, 8
928 vpalignr m3, m5, m0, 10
932 vpalignr m3, m5, m0, 12
936 vpalignr m3, m5, m0, 14
940 vpalignr m3, m5, m0, 16
944 mova m0, m3
954 movh m0, [lq] ; wxyz....
955 movhps m0, [aq-2] ; wxyz*abc
957 PALIGNR m1, m0, 2, m2 ; xyz*abcd
963 movh [dstq+stride3q ], m0
964 psrldq m0, 2 ; YZ#ABC..
965 movh [dstq+strideq*2], m0
966 psrldq m0, 2 ; Z#ABC...
967 movh [dstq+strideq*1], m0
968 psrldq m0, 2 ; #ABC....
969 movh [dstq+strideq*0], m0
973 mova m0, [lq] ; stuvwxyz
978 PALIGNR m1, m0, 2, m4 ; tuvwxyz*
988 PALIGNR m1, m3, m2, 2, m0
993 PALIGNR m2, m3, m1, 2, m0
998 PALIGNR m1, m3, m2, 2, m0
1006 mova m0, [lq] ; klmnopqr
1019 PALIGNR m1, m0, 2, m3 ; lmnopqrs
1037 PALIGNR m0, m4, m2, 2, m1
1038 mova m2, m0
1039 PALIGNR m0, m5, m4, 2, m1
1040 mova m4, m0
1041 PALIGNR m0, m6, m5, 2, m1
1042 mova m5, m0
1051 mova m0, [aq+mmsize*3] ; a[24-31]
1053 psrldq m2, m0, 2 ; a[25-31].
1057 PALIGNR m0, m1, 2, m4 ; a[17-24]
1073 PALIGNR m5, m6, 2, m0 ; l[25-31]*
1074 PALIGNR m4, m5, 2, m0 ; l[26-31]*a
1077 PALIGNR m6, m7, 2, m0 ; l[17-24]
1078 PALIGNR m5, m6, 2, m0 ; l[18-25]
1081 PALIGNR m7, m1, 2, m0 ; l[9-16]
1082 PALIGNR m6, m7, 2, m0 ; l[10-17]
1085 PALIGNR m1, m3, 2, m0 ; l[1-8]
1086 PALIGNR m7, m1, 2, m0 ; l[2-9]
1113 mova [dst8q+stride8q*0+48], m0
1132 vpalignr m1, m0, m1, 2
1133 vpalignr m0, m2, m0, 2
1139 PALIGNR m2, m6, m7, 2, m0
1141 PALIGNR m2, m5, m6, 2, m0
1143 PALIGNR m2, m4, m5, 2, m0
1145 PALIGNR m2, m3, m4, 2, m0
1147 PALIGNR m2, m1, m3, 2, m0
1153 PALIGNR m2, m0, m1, 2, m3
1157 PALIGNR m1, m2, m0, 2, m3
1158 mova m0, m1
1176 mova m0, [lq] ; klmnopqrstuvwxyz
1181 vperm2i128 m3, m0, m1, q0201 ; stuvwxyz*abcdefg
1183 vpalignr m4, m3, m0, 2 ; lmnopqrstuvwxyz*
1184 vpalignr m5, m3, m0, 4 ; mnopqrstuvwxyz*a
1186 vperm2i128 m5, m0, m1, q0201 ; TUVWXYZ#ABCDEFGH
1192 vpalignr m3, m5, m0, 2
1196 vpalignr m3, m5, m0, 4
1201 mova [dst3q+stride3q*4], m0 ; 15
1202 vpalignr m3, m5, m0, 6
1206 vpalignr m3, m5, m0, 8
1210 vpalignr m3, m5, m0, 10
1214 vpalignr m3, m5, m0, 12
1218 vpalignr m3, m5, m0, 14
1227 mova m0, [lq+mmsize*0+0] ; l[0-15]
1232 vperm2i128 m5, m0, m1, q0201 ; lmnopqrstuvwxyz0
1233 vpalignr m6, m5, m0, 2 ; mnopqrstuvwxyz01
1234 vpalignr m7, m5, m0, 4 ; nopqrstuvwxyz012
1248 vperm2i128 m5, m0, m1, q0201 ; L[7-15]L[16-23]
1259 mova [dst24q+stride7q+0 ], m0 ; 31 23 15 7
1264 vpalignr m7, m5, m0, 2
1271 vpalignr m7, m5, m0, 4
1278 vpalignr m7, m5, m0, 6
1285 vpalignr m7, m5, m0, 8
1292 vpalignr m7, m5, m0, 10
1299 vpalignr m7, m5, m0, 12
1306 vpalignr m7, m5, m0, 14
1312 mova m0, m5
1331 movu m0, [aq] ; abcdefgh
1332 psrldq m1, m0, 2 ; bcdefgh.
1333 psrldq m2, m0, 4 ; cdefgh..
1335 pavgw m1, m0 ; ABCDEFG.
1349 mova m0, [aq] ; abcdefgh
1353 SHIFT_RIGHTx2 m1, m2, m0, m3 ; bcdefghh/cdefghhh
1355 pavgw m1, m0 ; ABCDEFGh
1378 mova m0, [aq]
1380 PALIGNR m2, m1, m0, 2, m3
1381 PALIGNR m3, m1, m0, 4, m4
1383 pavgw m2, m0
1387 SHIFT_RIGHTx2 m5, m0, m1, m4
1397 mova [dstq+strideq*1+16], m0
1401 vpalignr m3, m0, m3, 2
1405 PALIGNR m5, m0, m3, 2, m4
1409 SHIFT_RIGHT m0, m0, m4
1416 mova m0, [aq+mmsize*0]
1419 PALIGNR m6, m1, m0, 2, m5
1420 PALIGNR m7, m1, m0, 4, m5
1422 pavgw m6, m0
1424 PALIGNR m4, m2, m1, 2, m0
1425 PALIGNR m5, m2, m1, 4, m0
1428 mova m0, [aq+mmsize*3]
1429 PALIGNR m1, m0, m2, 2, m6
1430 PALIGNR m3, m0, m2, 4, m6
1436 SHIFT_RIGHTx2 m6, m1, m0, reg_shuf
1438 pavgw m0, m6
1459 mova [dstq+strideq*0+48], m0
1466 mova [dstq+stride16q+32], m0
1480 vpalignr m2, m0, m2, 2
1493 PALIGNR m3, m0, m2, 2, m1
1512 SHIFT_RIGHT m0, m0, reg_shuf
1521 mova [dstq+strideq*0+48], m0
1522 mova [dstq+strideq*1+48], m0
1523 mova [dstq+strideq*2+48], m0
1524 mova [dstq+stride3q +48], m0
1543 movu m0, [aq-2]
1545 PALIGNR m0, m1, 10, m2 ; xyz*abcd
1546 pslldq m1, m0, 2 ; .xyz*abc
1547 pslldq m2, m0, 4 ; ..xyz*ab
1549 pavgw m1, m0 ; ....#ABC
1555 shufps m0, m2, m1, q3210
1562 psrldq m0, 6
1563 movh [dstq+strideq*2], m0
1570 mova m0, [aq] ; abcdefgh
1573 pavgw m0, m1
1580 mova [dstq+strideq*0], m0
1582 PALIGNR m0, m4, 14, m1
1586 mova [dstq+strideq*2], m0
1589 PALIGNR m0, m4, 14, m1
1593 mova [dstq+strideq*0], m0
1595 PALIGNR m0, m4, 14, m1
1598 mova [dstq+strideq*2], m0
1608 PALIGNR m0, m1, m5, 14, m6 ; z*abcdef
1633 mova [dstq+strideq*1+ 0], m0
1639 PALIGNR m6, m0, 14, m4
1640 PALIGNR m0, m5, 14, m4
1647 movu m0, [aq+mmsize*0-2] ; *a[0-6]
1668 PALIGNR m6, m0, m5, 14, m7 ; l[31]*a[0-5]
1671 pavgw m4, m0
1672 PALIGNR m0, m5, 2, m7 ; l[25-31]*
1679 psrld m7, m0, 16
1681 pand m0, [pd_65535]
1684 packssdw m5, m0
1688 mova m0, [lq+mmsize*1+0] ; l[8-15]
1691 movu m0, [lq+mmsize*0+2] ; l[1-8]
1692 pslldq m5, m0, 2 ; .l[1-7]
1693 pslldq m7, m0, 4 ; ..l[1-6]
1696 psrld m7, m0, 16
1698 pand m0, [pd_65535]
1700 packssdw m0, m6
1743 PALIGNR m12, m0, 14, m5
1744 pslldq m0, 2
1772 PALIGNR m5, m0, 14, m6
1773 pslldq m0, 2
1789 movh m0, [lq] ; abcd
1791 pshufb m0, [pb_0to7_67x4] ; abcddddd
1793 punpcklqdq m0, m0
1794 pshufhw m0, m0, q3333 ; abcddddd
1796 psrldq m1, m0, 2 ; bcddddd.
1797 psrldq m2, m0, 4 ; cddddd..
1799 pavgw m1, m0 ; abcddddd
1801 PALIGNR m2, m1, 4, m0 ; bCcDdddd
1812 mova m0, [lq]
1816 SHIFT_RIGHTx2 m1, m2, m0, m3
1818 pavgw m1, m0
1820 shufps m0, m1, m2, q1032
1826 mova [dstq+strideq *2], m0
1833 PALIGNR m0, m2, m1, 4, m3
1834 mova m1, m0
1837 shufps m0, m1, m2, q1032
1840 mova [dstq+strideq *2], m0
1846 mova m0, [lq]
1851 pavgw m1, m0
1856 SHIFT_RIGHTx2 m0, m4, m3, m5
1858 pavgw m3, m0
1860 pshufd m0, m0, q3333
1873 mova [dstq+stride3q*4+16], m0
1879 vpalignr m4, m0, m4, 4
1887 PALIGNR m5, m0, m4, 4, m6
1898 movu m0, [lq+mmsize*0+4]
1918 mova m0, [pb_2to15_14_15]
1920 SHIFT_RIGHTx2 m1, m6, m7, m0
1942 mova [dstq+strideq *0+16], m0
1945 mova [dstq+stride4q*1+ 0], m0
1954 vpalignr m1, m0, m1, 4
1955 vpalignr m0, m3, m0, 4
1962 PALIGNR m6, m0, m1, 4, m7
1964 PALIGNR m6, m3, m0, 4, m7
1965 mova m0, m6
2008 PALIGNR m0, m5, m2, 4, m3
2009 mova m2, m0
2010 PALIGNR m0, m4, m5, 4, m3
2011 mova m5, m0
2012 PALIGNR m0, m7, m4, 4, m3
2013 mova m4, m0
2014 PALIGNR m0, m6, m7, 4, m3
2015 mova m7, m0
2016 PALIGNR m0, m1, m6, 4, m3
2017 mova m6, m0
2038 movh m0, [lq]
2039 movhps m0, [aq-2]
2040 psrldq m1, m0, 2
2041 psrldq m2, m0, 4
2043 pavgw m1, m0
2051 PALIGNR m2, m1, 4, m0
2057 mova m0, [lq]
2059 PALIGNR m2, m1, m0, 2, m3
2060 PALIGNR m3, m1, m0, 4, m4
2062 pavgw m2, m0
2064 psrldq m0, m1, 2
2080 PALIGNR m0, m3, m2, 4, m4
2081 mova m2, m0
2082 PALIGNR m0, m1, m3, 4, m4
2083 mova m3, m0
2093 movu m0, [lq+4]
2124 mova [dstq+mstrideq *4+ 0], m0
2127 mova [dstq+mstrideq *0+16], m0
2129 vpalignr m1, m0, m1, 4
2130 vpalignr m0, m3, m0, 4
2135 PALIGNR m6, m0, m1, 4, m7
2137 PALIGNR m6, m3, m0, 4, m7
2138 mova m0, m6
2155 movu m0, [lq+mmsize*0+4]
2175 mova m0, [lq+mmsize*3+0]
2177 PALIGNR m7, m1, m0, 2, m2
2178 PALIGNR m6, m1, m0, 4, m2
2180 pavgw m7, m0
2183 movu m0, [aq+mmsize*0+2]
2246 mova [dstq+strideq* 8+48], m0
2249 mova [dstq+strideq* 4+32], m0
2252 mova [dstq+strideq* 0+16], m0
2266 vpalignr m6, m0, m6, 4
2267 vpalignr m0, m1, m0, 4
2291 PALIGNR m3, m0, m6, 4, m2
2293 PALIGNR m3, m1, m0, 4, m2
2294 mova m0, m3
2299 PALIGNR m3, m2, m1, 4, m0
2301 PALIGNR m3, reg_sh, m2, 4, m0
2328 mova [dstq+stride3q*4+16], m0
2331 mova [dstq+strideq* 8+ 0], m0
2345 vpalignr m1, m0, m1, 4
2346 vpalignr m0, m3, m0, 4
2357 PALIGNR m6, m0, m1, 4, m7
2359 PALIGNR m6, m3, m0, 4, m7
2360 mova m0, m6