51 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
63 	vpshufb t4, x7, x7; \
75 	filter_8bit(x7, t0, t1, t7, t6); \
90 	vaesenclast t4, x7, x7; \
102 	filter_8bit(x7, t0, t1, t7, t6); \
137 	vpxor x7, x2, x2; \
143 	vpxor x1, x7, x7; \
145 	vpxor x7, x0, x0; \
153 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
172 	vpxor t0, x7, x7; \
173 	vpxor 3 * 16(mem_cd), x7, x7; \
213 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
221 	vmovdqu x7, 3 * 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
243 	vmovdqu x7, 7 * 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
496 	vpxor 8 * 16(rio), x0, x7; \
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, \
518 	vmovdqu x7, 7 * 16(mem_ab); \
529 #define outunpack16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
532 			 y7, x3, x7, stack_tmp0, stack_tmp1); \
547 	vpxor x0, x7, x7; \
556 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
565 	vmovdqu x7, 7 * 16(rio); \