66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
85 	vpshufb t4, x1, x1; \
100 	filter_8bit(x1, t5, t6, t7, t4); \
120 	vextracti128 $1, x1, t3##_x; \
130 	vaesenclast t4##_x, x1##_x, x1##_x; \
132 	vinserti128 $1, t3##_x, x1, x1; \
154 	filter_8bit(x1, t4, t5, t7, t2); \
165 	vpxor x6, x1, x1; \
178 	vpxor x1, x7, x7; \
185 	vpxor x4, x1, x1; \
191 	vpxor x1, x6, x6; \
196 	vpxor t6, x1, x1; \
197 	vpxor 5 * 32(mem_cd), x1, x1; \
250 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
256 	vmovdqu x1, 5 * 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
278 	vmovdqu x1, 1 * 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, \
457 #define transpose_4x4(x0, x1, x2, x3, t1, t2) \  argument
458 	vpunpckhdq x1, x0, t2; \
459 	vpunpckldq x1, x0, x0; \
464 	vpunpckhqdq t1, x0, x1; \
520 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
539 	vpxor 14 * 32(rio), x0, x1; \
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, \
549 	vmovdqu x1, 1 * 32(mem_ab); \
566 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
568 	byteslice_16x16b_fast(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, \
590 	vpxor x0, x1, x1; \
593 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
596 	vmovdqu x1, 1 * 32(rio); \