189 		     y0, y1, y2, y3,			\  argument
200 	vmovdqu (8 * 32)(rio), y0;			\
212 		      y0, y1, y2, y3,			\  argument
217 			 y0, y1, y2, y3,		\
229 	vmovdqu y0, 0 * 32(mem_cd);			\
240 		     y0, y1, y2, y3,			\  argument
251 	vmovdqu y0, 8 * 32(mem);			\
422 		       y0, y1, y2, y3,			\  argument
425 	vpxor y0, x4, x4;				\
431 	vpxor y4, y0, y0;				\
449 	vpxor x0, y0, y0;				\
455 	vpxor y0, x4, x4;				\
462 		y0, y1, y2, y3,				\  argument
466 		      y0, rk, 8, round);		\
469 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
471 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
472 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
481 		      y0, rk, 0, 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);	\
491 	aria_load_state_8way(y0, y1, y2, y3,		\
496 		       y0, y1, y2, y3,			\
508 		       y0, y1, y2, y3,			\
516 		y0, y1, y2, y3,				\  argument
520 		      y0, rk, 8, round);		\
523 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
525 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
526 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
535 		      y0, rk, 0, round);		\
538 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
540 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
541 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
545 	aria_load_state_8way(y0, y1, y2, y3,		\
550 		       y0, y1, y2, y3,			\
556 	 * T2 = y0, y1, y2, y3, -> y2, y3, y0, y1	\
562 		       y2, y3, y0, y1,			\
570 		y0, y1, y2, y3,				\  argument
574 		      y0, rk, 8, round);		\
577 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
580 		      y0, rk, 8, last_round);		\
590 		      y0, rk, 0, round);		\
593 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
596 		      y0, rk, 0, last_round);		\
598 	aria_load_state_8way(y0, y1, y2, y3,		\
604 		     y0, y1, y2, y3,			\  argument
608 		      y0, rk, 8, round);		\
612 			    y0, y1, y2, y3,		\
615 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
616 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
625 		      y0, rk, 0, round);		\
629 			    y0, y1, y2, y3,		\
632 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
633 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
637 	aria_load_state_8way(y0, y1, y2, y3,		\
642 		       y0, y1, y2, y3,			\
654 		       y0, y1, y2, y3,			\
662 		     y0, y1, y2, y3,			\  argument
666 		      y0, rk, 8, round);		\
670 			    y0, y1, y2, y3,		\
673 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
674 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
683 		      y0, rk, 0, round);		\
687 			    y0, y1, y2, y3,		\
690 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
691 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
695 	aria_load_state_8way(y0, y1, y2, y3,		\
700 		       y0, y1, y2, y3,			\
706 	 * T2 = y0, y1, y2, y3, -> y2, y3, y0, y1	\
712 		       y2, y3, y0, y1,			\
720 		y0, y1, y2, y3,				\  argument
724 		      y0, rk, 8, round);		\
728 			    y0, y1, y2, y3,		\
732 		      y0, rk, 8, last_round);		\
742 		      y0, rk, 0, round);		\
746 			    y0, y1, y2, y3,		\
750 		      y0, rk, 0, last_round);		\
752 	aria_load_state_8way(y0, y1, y2, y3,		\