51 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
69 	vpshufb t4, x6, x6; \
84 	filter_8bit(x6, t2, t3, t7, t6); \
96 	vaesenclast t4, x6, x6; \
104 	filter_8bit(x6, t0, t1, t7, t6); \
136 	vpxor x6, x1, x1; \
142 	vpxor x0, x6, x6; \
148 	vpxor x6, x3, x3; \
152 	vpxor x1, x6, x6; \
169 	vpxor t1, x6, x6; \
170 	vpxor 2 * 16(mem_cd), x6, x6; \
213 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
220 	vmovdqu x6, 2 * 16(mem_cd); \
230 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
232 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
234 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
242 	vmovdqu x6, 6 * 16(mem_ab); \
245 #define enc_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
247 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
249 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
251 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
254 #define dec_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
256 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
258 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
260 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
483 #define inpack16_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
497 	vpxor 9 * 16(rio), x0, x6; \
506 #define inpack16_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
508 	byteslice_16x16b(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \
517 	vmovdqu x6, 6 * 16(mem_ab); \
529 #define outunpack16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
531 	byteslice_16x16b(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, y3, \
548 	vpxor x0, x6, x6; \
556 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
564 	vmovdqu x6, 6 * 16(rio); \