63 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
79 	vpshufb t4, x6, x6; \
92 	filter_8bit(x6, t2, t3, t7, t4); \
94 	vextracti128 $1, x6, t2##_x; \
114 	vaesenclast t4##_x, x6##_x, x6##_x; \
116 	vinserti128 $1, t2##_x, x6, x6; \
140 	filter_8bit(x6, t0, t1, t7, t6); \
162 	vpxor x6, x1, x1; \
174 	vpxor x0, x6, x6; \
184 	vpxor x6, x3, x3; \
188 	vpxor x1, x6, x6; \
215 	vpxor t1, x6, x6; \
216 	vpxor 2 * 32(mem_cd), x6, x6; \
247 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
258 	vmovdqu x6, 2 * 32(mem_cd); \
264 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
266 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
268 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
272 	vmovdqu x6, 6 * 32(mem_ab); \
279 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
281 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
283 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
285 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
288 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
290 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
292 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
294 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
517 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
531 	vpxor 9 * 32(rio), x0, x6; \
540 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
542 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
551 	vmovdqu x6, 6 * 32(mem_ab); \
563 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
565 	byteslice_16x16b_fast(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, \
582 	vpxor x0, x6, x6; \
590 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
598 	vmovdqu x6, 6 * 32(rio); \