176 		     x4, x5, x6, x7,			\  argument
184 	vmovdqu (4 * 16)(rio), x4;			\
199 		      x4, x5, x6, x7,			\  argument
204 			 x4, x5, x6, x7,		\
213 	vmovdqu x4, 4 * 16(mem_ab);			\
227 		     x4, x5, x6, x7,			\  argument
235 	vmovdqu x4, 4 * 16(mem);			\
249 			      x4, x5, x6, x7,		\  argument
255 	vmovdqu x4, ((idx + 4) * 16)(mem_tmp);		\
261 			     x4, x5, x6, x7,		\  argument
267 	vmovdqu ((idx + 4) * 16)(mem_tmp), x4;		\
273 		      x4, x5, x6, x7,			\  argument
285 	vpxor t0, x4, x4;				\
294 			    x4, x5, x6, x7,		\  argument
309 	vgf2p8affineinvqb $(tf_aff_const), t3, x4, x4;	\
316 		       x4, x5, x6, x7,			\  argument
329 	vaesenclast t7, x4, x4;				\
337 	vpshufb t0, x4, x4;				\
373 		       x4, x5, x6, x7,			\  argument
377 	vpxor y0, x4, x4;				\
389 	vpxor x4, x0, x0;				\
395 	vpxor x4, y4, y4;				\
407 	vpxor y0, x4, x4;				\
413 		x4, x5, x6, x7,				\  argument
417 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
420 	aria_sbox_8way(x2, x3, x0, x1, x6, x7, x4, x5,	\
424 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
426 			      x4, x5, x6, x7,		\
430 			     x4, x5, x6, x7,		\
432 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
435 	aria_sbox_8way(x2, x3, x0, x1, x6, x7, x4, x5,	\
439 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
441 			      x4, x5, x6, x7,		\
447 		       x4, x5, x6, x7,			\
459 		       x7, x6, x5, x4,			\
463 			      x6, x7, x4, x5,		\
467 		x4, x5, x6, x7,				\  argument
471 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
474 	aria_sbox_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
478 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
480 			      x4, x5, x6, x7,		\
484 			     x4, x5, x6, x7,		\
486 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
489 	aria_sbox_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
493 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
495 			      x4, x5, x6, x7,		\
501 		       x4, x5, x6, x7,			\
513 		       x5, x4, x7, x6,			\
517 			      x6, x7, x4, x5,		\
521 		x4, x5, x6, x7,				\  argument
525 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
528 	aria_sbox_8way(x2, x3, x0, x1, x6, x7, x4, x5,	\
531 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
535 			      x4, x5, x6, x7,		\
539 			     x4, x5, x6, x7,		\
541 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
544 	aria_sbox_8way(x2, x3, x0, x1, x6, x7, x4, x5,	\
547 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
555 		     x4, x5, x6, x7,			\  argument
559 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
563 			    x6, x7, x4, x5,		\
568 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
570 			      x4, x5, x6, x7,		\
574 			     x4, x5, x6, x7,		\
576 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
580 			    x6, x7, x4, x5,		\
585 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
587 			      x4, x5, x6, x7,		\
593 		       x4, x5, x6, x7,			\
605 		       x7, x6, x5, x4,			\
609 			      x6, x7, x4, x5,		\
613 		     x4, x5, x6, x7,			\  argument
617 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
621 			    x4, x5, x6, x7,		\
626 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
628 			      x4, x5, x6, x7,		\
632 			     x4, x5, x6, x7,		\
634 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
638 			    x4, x5, x6, x7,		\
643 	aria_diff_m(x4, x5, x6, x7, y0, y1, y2, y3);	\
645 			      x4, x5, x6, x7,		\
651 		       x4, x5, x6, x7,			\
663 		       x5, x4, x7, x6,			\
667 			      x6, x7, x4, x5,		\
671 		x4, x5, x6, x7,				\  argument
675 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
679 			    x6, x7, x4, x5,		\
683 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
687 			      x4, x5, x6, x7,		\
691 			     x4, x5, x6, x7,		\
693 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\
697 			    x6, x7, x4, x5,		\
701 	aria_ark_8way(x0, x1, x2, x3, x4, x5, x6, x7,	\