177 		     y0, y1, y2, y3,			\  argument
191 	vmovdqu (11 * 16)(rio), y3;			\
200 		      y0, y1, y2, y3,			\  argument
205 			 y0, y1, y2, y3,		\
220 	vmovdqu y3, 3 * 16(mem_cd);			\
228 		     y0, y1, y2, y3,			\  argument
242 	vmovdqu y3, 11 * 16(mem);			\
374 		       y0, y1, y2, y3,			\  argument
380 	vpxor y3, x7, x7;				\
386 	vpxor y7, y3, y3;				\
404 	vpxor x3, y3, y3;				\
410 	vpxor y3, x7, x7;
414 		y0, y1, y2, y3,				\  argument
421 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
423 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
424 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
436 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
438 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
439 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
443 	aria_load_state_8way(y0, y1, y2, y3,		\
448 		       y0, y1, y2, y3,			\
460 		       y0, y1, y2, y3,			\
468 		y0, y1, y2, y3,				\  argument
475 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
477 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
478 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
490 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
492 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
493 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
497 	aria_load_state_8way(y0, y1, y2, y3,		\
502 		       y0, y1, y2, y3,			\
514 		       y2, y3, y0, y1,			\
522 		y0, y1, y2, y3,				\  argument
529 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
545 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
550 	aria_load_state_8way(y0, y1, y2, y3,		\
556 		     y0, y1, y2, y3,			\  argument
564 			    y0, y1, y2, y3, 		\
567 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
568 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
581 			    y0, y1, y2, y3, 		\
584 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
585 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
589 	aria_load_state_8way(y0, y1, y2, y3,		\
594 		       y0, y1, y2, y3,			\
606 		       y0, y1, y2, y3,			\
614 		     y0, y1, y2, y3,			\  argument
622 			    y0, y1, y2, y3, 		\
625 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
626 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
639 			    y0, y1, y2, y3, 		\
642 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
643 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
647 	aria_load_state_8way(y0, y1, y2, y3,		\
652 		       y0, y1, y2, y3,			\
664 		       y2, y3, y0, y1,			\
672 		y0, y1, y2, y3,				\  argument
680 			    y0, y1, y2, y3, 		\
698 			    y0, y1, y2, y3, 		\
704 	aria_load_state_8way(y0, y1, y2, y3,		\