52 		  t7, mem_cd, key) \  argument
160 	vpxor 0 * 16(mem_cd), x4, x4; \
163 	vpxor 1 * 16(mem_cd), x5, x5; \
170 	vpxor 2 * 16(mem_cd), x6, x6; \
173 	vpxor 3 * 16(mem_cd), x7, x7; \
176 	vpxor 4 * 16(mem_cd), x0, x0; \
179 	vpxor 5 * 16(mem_cd), x1, x1; \
182 	vpxor 6 * 16(mem_cd), x2, x2; \
185 	vpxor 7 * 16(mem_cd), x3, x3;
214 		      y6, y7, mem_ab, mem_cd, i, dir, store_ab) \  argument
218 	vmovdqu x4, 0 * 16(mem_cd); \
219 	vmovdqu x5, 1 * 16(mem_cd); \
220 	vmovdqu x6, 2 * 16(mem_cd); \
221 	vmovdqu x7, 3 * 16(mem_cd); \
222 	vmovdqu x0, 4 * 16(mem_cd); \
223 	vmovdqu x1, 5 * 16(mem_cd); \
224 	vmovdqu x2, 6 * 16(mem_cd); \
225 	vmovdqu x3, 7 * 16(mem_cd); \
246 		      y6, y7, mem_ab, mem_cd, i) \  argument
248 		      y6, y7, mem_ab, mem_cd, (i) + 2, 1, store_ab_state); \
250 		      y6, y7, mem_ab, mem_cd, (i) + 4, 1, store_ab_state); \
252 		      y6, y7, mem_ab, mem_cd, (i) + 6, 1, dummy_store);
255 		      y6, y7, mem_ab, mem_cd, i) \  argument
257 		      y6, y7, mem_ab, mem_cd, (i) + 7, -1, store_ab_state); \
259 		      y6, y7, mem_ab, mem_cd, (i) + 5, -1, store_ab_state); \
261 		      y6, y7, mem_ab, mem_cd, (i) + 3, -1, dummy_store);
507 		      y6, y7, mem_ab, mem_cd) \  argument
509 			 y5, y6, y7, (mem_ab), (mem_cd)); \
519 	vmovdqu y0, 0 * 16(mem_cd); \
520 	vmovdqu y1, 1 * 16(mem_cd); \
521 	vmovdqu y2, 2 * 16(mem_cd); \
522 	vmovdqu y3, 3 * 16(mem_cd); \
523 	vmovdqu y4, 4 * 16(mem_cd); \
524 	vmovdqu y5, 5 * 16(mem_cd); \
525 	vmovdqu y6, 6 * 16(mem_cd); \
526 	vmovdqu y7, 7 * 16(mem_cd);