68 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
86 	vpshufb t4, x5, x5; \
101 	filter_8bit(x5, t5, t6, t7, t4); \
109 	vextracti128 $1, x5, t5##_x; \
129 	vaesenclast t4##_x, x5##_x, x5##_x; \
131 	vinserti128 $1, t5##_x, x5, x5; \
151 	filter_8bit(x5, t2, t3, t7, t6); \
166 	vpxor x5, x0, x0; \
178 	vpxor x3, x5, x5; \
188 	vpxor x5, x2, x2; \
192 	vpxor x0, x5, x5; \
217 	vpxor t2, x5, x5; \
218 	vpxor 1 * 32(mem_cd), x5, x5; \
252 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
262 	vmovdqu x5, 1 * 32(mem_cd); \
269 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
271 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
273 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
276 	vmovdqu x5, 5 * 32(mem_ab); \
284 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
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, \
290 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
293 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
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, \
299 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
522 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
537 	vpxor 10 * 32(rio), x0, x5; \
545 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
547 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
555 	vmovdqu x5, 5 * 32(mem_ab); \
568 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
570 	byteslice_16x16b_fast(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, \
588 	vpxor x0, x5, x5; \
595 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
602 	vmovdqu x5, 5 * 32(rio); \