66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
80 	vpshufb t4, x7, x7; \
91 	filter_8bit(x7, t5, t6, t7, t4); \
93 	vextracti128 $1, x7, t1##_x; \
111 	vaesenclast t4##_x, x7##_x, x7##_x; \
113 	vinserti128 $1, t1##_x, x7, x7; \
141 	filter_8bit(x7, t0, t1, t7, t6); \
166 	vpxor x7, x2, x2; \
178 	vpxor x1, x7, x7; \
184 	vpxor x7, x0, x0; \
192 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
221 	vpxor t0, x7, x7; \
222 	vpxor 3 * 32(mem_cd), x7, x7;
250 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
262 	vmovdqu x7, 3 * 32(mem_cd); \
267 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
269 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
271 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
276 	vmovdqu x7, 7 * 32(mem_ab); \
282 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
284 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
286 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
288 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
291 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
293 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
295 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
297 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
520 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
533 	vpxor 8 * 32(rio), x0, x7; \
543 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
545 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
555 	vmovdqu x7, 7 * 32(mem_ab); \
566 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
569 			      y3, y7, x3, x7, stack_tmp0, stack_tmp1); \
584 	vpxor x0, x7, x7; \
593 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
602 	vmovdqu x7, 7 * 32(rio); \