212 		      y6, y7, mem_ab, mem_cd, i, dir, store_ab) \  argument
228 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
230 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
232 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
234 	vmovdqu x0, 0 * 16(mem_ab); \
235 	vmovdqu x1, 1 * 16(mem_ab); \
236 	vmovdqu x2, 2 * 16(mem_ab); \
237 	vmovdqu x3, 3 * 16(mem_ab); \
238 	vmovdqu x4, 4 * 16(mem_ab); \
239 	vmovdqu x5, 5 * 16(mem_ab); \
240 	vmovdqu x6, 6 * 16(mem_ab); \
241 	vmovdqu x7, 7 * 16(mem_ab);
244 		      y6, y7, mem_ab, mem_cd, i) \  argument
246 		      y6, y7, mem_ab, mem_cd, (i) + 2, 1, store_ab_state); \
248 		      y6, y7, mem_ab, mem_cd, (i) + 4, 1, store_ab_state); \
250 		      y6, y7, mem_ab, mem_cd, (i) + 6, 1, dummy_store);
253 		      y6, y7, mem_ab, mem_cd, i) \  argument
255 		      y6, y7, mem_ab, mem_cd, (i) + 7, -1, store_ab_state); \
257 		      y6, y7, mem_ab, mem_cd, (i) + 5, -1, store_ab_state); \
259 		      y6, y7, mem_ab, mem_cd, (i) + 3, -1, dummy_store);
505 		      y6, y7, mem_ab, mem_cd) \  argument
507 			 y5, y6, y7, (mem_ab), (mem_cd)); \
509 	vmovdqu x0, 0 * 16(mem_ab); \
510 	vmovdqu x1, 1 * 16(mem_ab); \
511 	vmovdqu x2, 2 * 16(mem_ab); \
512 	vmovdqu x3, 3 * 16(mem_ab); \
513 	vmovdqu x4, 4 * 16(mem_ab); \
514 	vmovdqu x5, 5 * 16(mem_ab); \
515 	vmovdqu x6, 6 * 16(mem_ab); \
516 	vmovdqu x7, 7 * 16(mem_ab); \