176 		     y0, y1, y2, y3,			\  argument
189 	vmovdqu (10 * 16)(rio), y2;			\
199 		      y0, y1, y2, y3,			\  argument
204 			 y0, y1, y2, y3,		\
218 	vmovdqu y2, 2 * 16(mem_cd);			\
227 		     y0, y1, y2, y3,			\  argument
240 	vmovdqu y2, 10 * 16(mem);			\
373 		       y0, y1, y2, y3,			\  argument
378 	vpxor y2, x6, x6;				\
384 	vpxor y6, y2, y2;				\
402 	vpxor x2, y2, y2;				\
408 	vpxor y2, x6, x6;				\
413 		y0, y1, y2, y3,				\  argument
420 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
422 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
423 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
435 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
437 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
438 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
442 	aria_load_state_8way(y0, y1, y2, y3,		\
447 		       y0, y1, y2, y3,			\
459 		       y0, y1, y2, y3,			\
467 		y0, y1, y2, y3,				\  argument
474 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
476 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
477 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
489 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
491 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
492 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
496 	aria_load_state_8way(y0, y1, y2, y3,		\
501 		       y0, y1, y2, y3,			\
507 	 * T2 = y0, y1, y2, y3, -> y2, y3, y0, y1 	\
513 		       y2, y3, y0, y1,			\
521 		y0, y1, y2, y3,				\  argument
528 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
544 		       y0, y1, y2, y3, y4, y5, y6, y7);	\
549 	aria_load_state_8way(y0, y1, y2, y3,		\
555 		     y0, y1, y2, y3,			\  argument
563 			    y0, y1, y2, y3, 		\
566 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
567 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
580 			    y0, y1, y2, y3, 		\
583 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
584 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
588 	aria_load_state_8way(y0, y1, y2, y3,		\
593 		       y0, y1, y2, y3,			\
605 		       y0, y1, y2, y3,			\
613 		     y0, y1, y2, y3,			\  argument
621 			    y0, y1, y2, y3, 		\
624 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
625 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
638 			    y0, y1, y2, y3, 		\
641 	aria_diff_m(x0, x1, x2, x3, y0, y1, y2, y3);	\
642 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
646 	aria_load_state_8way(y0, y1, y2, y3,		\
651 		       y0, y1, y2, y3,			\
657 	 * T2 = y0, y1, y2, y3, -> y2, y3, y0, y1 	\
663 		       y2, y3, y0, y1,			\
671 		y0, y1, y2, y3,				\  argument
679 			    y0, y1, y2, y3, 		\
697 			    y0, y1, y2, y3, 		\
703 	aria_load_state_8way(y0, y1, y2, y3,		\