51 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
68 	vpshufb t4, x3, x3; \
83 	filter_8bit(x3, t2, t3, t7, t6); \
95 	vaesenclast t4, x3, x3; \
103 	filter_8bit(x3, t0, t1, t7, t6); \
138 	vpxor x4, x3, x3; \
141 	vpxor x3, x5, x5; \
148 	vpxor x6, x3, x3; \
150 	vpxor x3, x4, x4; \
184 	vpxor t4, x3, x3; \
185 	vpxor 7 * 16(mem_cd), x3, x3;
213 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
225 	vmovdqu x3, 7 * 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
239 	vmovdqu x3, 3 * 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, \
420 #define transpose_4x4(x0, x1, x2, x3, t1, t2) \  argument
424 	vpunpckldq x3, x2, t1; \
425 	vpunpckhdq x3, x2, x2; \
430 	vpunpckhqdq x2, t2, x3; \
483 #define inpack16_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
500 	vpxor 12 * 16(rio), x0, x3; \
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, \
514 	vmovdqu x3, 3 * 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); \
551 	vpxor x0, x3, x3; \
556 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
561 	vmovdqu x3, 3 * 16(rio); \