173 		     y0, y1, y2, y3,			\  argument
187 	vmovdqu (11 * 16)(rio), y3;			\
196 		      y0, y1, y2, y3,			\  argument
201 			 y0, y1, y2, y3,		\
216 	vmovdqu y3, 3 * 16(mem_cd);			\
224 		     y0, y1, y2, y3,			\  argument
238 	vmovdqu y3, 11 * 16(mem);			\
381 		       y0, y1, y2, y3,			\  argument
387 	vpxor y3, x7, x7;				\
393 	vpxor y7, y3, y3;				\
411 	vpxor x3, y3, y3;				\
417 	vpxor y3, x7, x7;
421 		y0, y1, y2, y3,				\  argument
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);	\
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
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);	\
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
539 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
555 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
560 	aria_load_state_8way(y0, y1, y2, y3,		\
567 		     y0, y1, y2, y3,			\  argument
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);	\
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
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);	\
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
694 			    y0, y1, y2, y3, 		\
712 			    y0, y1, y2, y3, 		\
718 	aria_load_state_8way(y0, y1, y2, y3,		\