173 		     y0, y1, y2, y3,			\  argument
184 	vmovdqu (8 * 16)(rio), y0;			\
196 		      y0, y1, y2, y3,			\  argument
201 			 y0, y1, y2, y3,		\
213 	vmovdqu y0, 0 * 16(mem_cd);			\
224 		     y0, y1, y2, y3,			\  argument
235 	vmovdqu y0, 8 * 16(mem);			\
381 		       y0, y1, y2, y3,			\  argument
384 	vpxor y0, x4, x4;				\
390 	vpxor y4, y0, y0;				\
408 	vpxor x0, y0, y0;				\
414 	vpxor y0, x4, x4;				\
421 		y0, y1, y2, y3,				\  argument
426 		      y0, y7, y2, rk, 8, round);	\
429 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
431 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
432 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
441 		      y0, y7, y2, rk, 0, round);	\
444 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
446 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
447 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
451 	aria_load_state_8way(y0, y1, y2, y3,		\
456 		       y0, y1, y2, y3,			\
468 		       y0, y1, y2, y3,			\
476 		y0, y1, y2, y3,				\  argument
481 		      y0, y7, y2, rk, 8, round);	\
484 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
486 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
487 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
496 		      y0, y7, y2, rk, 0, round);	\
499 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
501 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
502 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
506 	aria_load_state_8way(y0, y1, y2, y3,		\
511 		       y0, y1, y2, y3,			\
523 		       y2, y3, y0, y1,			\
531 		y0, y1, y2, y3,				\  argument
536 		      y0, y7, y2, rk, 8, round);	\
539 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
542 		      y0, y7, y2, rk, 8, last_round);	\
552 		      y0, y7, y2, rk, 0, round);	\
555 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
558 		      y0, y7, y2, rk, 0, last_round);	\
560 	aria_load_state_8way(y0, y1, y2, y3,		\
567 		     y0, y1, y2, y3,			\  argument
572 		      y0, y7, y2, rk, 8, round);	\
576 			    y0, y1, y2, y3, 		\
579 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
580 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
589 		      y0, y7, y2, rk, 0, round);	\
593 			    y0, y1, y2, y3, 		\
596 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
597 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
601 	aria_load_state_8way(y0, y1, y2, y3,		\
606 		       y0, y1, y2, y3,			\
618 		       y0, y1, y2, y3,			\
626 		     y0, y1, y2, y3,			\  argument
631 		      y0, y7, y2, rk, 8, round);	\
635 			    y0, y1, y2, y3, 		\
638 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
639 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
648 		      y0, y7, y2, rk, 0, round);	\
652 			    y0, y1, y2, y3, 		\
655 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
656 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
660 	aria_load_state_8way(y0, y1, y2, y3,		\
665 		       y0, y1, y2, y3,			\
677 		       y2, y3, y0, y1,			\
685 		y0, y1, y2, y3,				\  argument
690 		      y0, y7, y2, rk, 8, round);	\
694 			    y0, y1, y2, y3, 		\
698 		      y0, y7, y2, rk, 8, last_round);	\
708 		      y0, y7, y2, rk, 0, round);	\
712 			    y0, y1, y2, y3, 		\
716 		      y0, y7, y2, rk, 0, last_round);	\
718 	aria_load_state_8way(y0, y1, y2, y3,		\