63 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
77 	vpshufb t4, x7, x7; \
88 	filter_8bit(x7, t5, t6, t7, t4); \
90 	vextracti128 $1, x7, t1##_x; \
108 	vaesenclast t4##_x, x7##_x, x7##_x; \
110 	vinserti128 $1, t1##_x, x7, x7; \
138 	filter_8bit(x7, t0, t1, t7, t6); \
163 	vpxor x7, x2, x2; \
175 	vpxor x1, x7, x7; \
181 	vpxor x7, x0, x0; \
189 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
218 	vpxor t0, x7, x7; \
219 	vpxor 3 * 32(mem_cd), x7, x7;
247 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
259 	vmovdqu x7, 3 * 32(mem_cd); \
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
273 	vmovdqu x7, 7 * 32(mem_ab); \
279 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
281 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
283 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
285 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
288 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
290 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
292 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
294 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
517 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
530 	vpxor 8 * 32(rio), x0, x7; \
540 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
542 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
552 	vmovdqu x7, 7 * 32(mem_ab); \
563 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
566 			      y3, y7, x3, x7, stack_tmp0, stack_tmp1); \
581 	vpxor x0, x7, x7; \
590 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
599 	vmovdqu x7, 7 * 32(rio); \