66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
84 	vpshufb t4, x5, x5; \
99 	filter_8bit(x5, t5, t6, t7, t4); \
107 	vextracti128 $1, x5, t5##_x; \
127 	vaesenclast t4##_x, x5##_x, x5##_x; \
129 	vinserti128 $1, t5##_x, x5, x5; \
149 	filter_8bit(x5, t2, t3, t7, t6); \
164 	vpxor x5, x0, x0; \
176 	vpxor x3, x5, x5; \
186 	vpxor x5, x2, x2; \
190 	vpxor x0, x5, x5; \
215 	vpxor t2, x5, x5; \
216 	vpxor 1 * 32(mem_cd), x5, x5; \
250 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
260 	vmovdqu x5, 1 * 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
274 	vmovdqu x5, 5 * 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
535 	vpxor 10 * 32(rio), x0, x5; \
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, \
553 	vmovdqu x5, 5 * 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, \
586 	vpxor x0, x5, x5; \
593 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
600 	vmovdqu x5, 5 * 32(rio); \