66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
81 	vpshufb t4, x3, x3; \
94 	filter_8bit(x3, t2, t3, t7, t4); \
96 	vextracti128 $1, x3, t3##_x; \
114 	vaesenclast t4##_x, x3##_x, x3##_x; \
116 	vinserti128 $1, t3##_x, x3, x3; \
142 	filter_8bit(x3, t0, t1, t7, t6); \
167 	vpxor x4, x3, x3; \
176 	vpxor x3, x5, x5; \
187 	vpxor x6, x3, x3; \
189 	vpxor x3, x4, x4; \
209 	vpxor t4, x3, x3; \
210 	vpxor 7 * 32(mem_cd), x3, x3; \
250 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
258 	vmovdqu x3, 7 * 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
280 	vmovdqu x3, 3 * 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
461 	vpunpckldq x3, x2, t1; \
462 	vpunpckhdq x3, x2, x2; \
467 	vpunpckhqdq x2, t2, x3; \
520 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
537 	vpxor 12 * 32(rio), x0, x3; \
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, \
551 	vmovdqu x3, 3 * 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); \
588 	vpxor x0, x3, x3; \
593 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
598 	vmovdqu x3, 3 * 32(rio); \