189 		     y0, y1, y2, y3,			\  argument
202 	vmovdqu (10 * 32)(rio), y2;			\
212 		      y0, y1, y2, y3,			\  argument
217 			 y0, y1, y2, y3,		\
231 	vmovdqu y2, 2 * 32(mem_cd);			\
240 		     y0, y1, y2, y3,			\  argument
253 	vmovdqu y2, 10 * 32(mem);			\
422 		       y0, y1, y2, y3,			\  argument
427 	vpxor y2, x6, x6;				\
433 	vpxor y6, y2, y2;				\
451 	vpxor x2, y2, y2;				\
457 	vpxor y2, x6, x6;				\
462 		y0, y1, y2, y3,				\  argument
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);	\
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
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);	\
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,			\
562 		       y2, y3, y0, y1,			\
570 		y0, y1, y2, y3,				\  argument
577 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
593 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
598 	aria_load_state_8way(y0, y1, y2, y3,		\
604 		     y0, y1, y2, y3,			\  argument
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);	\
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
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);	\
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,			\
712 		       y2, y3, y0, y1,			\
720 		y0, y1, y2, y3,				\  argument
728 			    y0, y1, y2, y3,		\
746 			    y0, y1, y2, y3,		\
752 	aria_load_state_8way(y0, y1, y2, y3,		\