248 		      y6, y7, mem_ab, mem_cd, i, dir, store_ab) \  argument
264 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
266 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
268 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
270 	vmovdqu x4, 4 * 32(mem_ab); \
271 	vmovdqu x5, 5 * 32(mem_ab); \
272 	vmovdqu x6, 6 * 32(mem_ab); \
273 	vmovdqu x7, 7 * 32(mem_ab); \
274 	vmovdqu x0, 0 * 32(mem_ab); \
275 	vmovdqu x1, 1 * 32(mem_ab); \
276 	vmovdqu x2, 2 * 32(mem_ab); \
277 	vmovdqu x3, 3 * 32(mem_ab);
280 		      y6, y7, mem_ab, mem_cd, i) \  argument
282 		      y6, y7, mem_ab, mem_cd, (i) + 2, 1, store_ab_state); \
284 		      y6, y7, mem_ab, mem_cd, (i) + 4, 1, store_ab_state); \
286 		      y6, y7, mem_ab, mem_cd, (i) + 6, 1, dummy_store);
289 		      y6, y7, mem_ab, mem_cd, i) \  argument
291 		      y6, y7, mem_ab, mem_cd, (i) + 7, -1, store_ab_state); \
293 		      y6, y7, mem_ab, mem_cd, (i) + 5, -1, store_ab_state); \
295 		      y6, y7, mem_ab, mem_cd, (i) + 3, -1, dummy_store);
541 		      y6, y7, mem_ab, mem_cd) \  argument
543 			      y4, y5, y6, y7, (mem_ab), (mem_cd)); \
545 	vmovdqu x0, 0 * 32(mem_ab); \
546 	vmovdqu x1, 1 * 32(mem_ab); \
547 	vmovdqu x2, 2 * 32(mem_ab); \
548 	vmovdqu x3, 3 * 32(mem_ab); \
549 	vmovdqu x4, 4 * 32(mem_ab); \
550 	vmovdqu x5, 5 * 32(mem_ab); \
551 	vmovdqu x6, 6 * 32(mem_ab); \
552 	vmovdqu x7, 7 * 32(mem_ab); \