49 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
66 	vpshufb t4, x3, x3; \
81 	filter_8bit(x3, t2, t3, t7, t6); \
93 	vaesenclast t4, x3, x3; \
101 	filter_8bit(x3, t0, t1, t7, t6); \
136 	vpxor x4, x3, x3; \
139 	vpxor x3, x5, x5; \
146 	vpxor x6, x3, x3; \
148 	vpxor x3, x4, x4; \
182 	vpxor t4, x3, x3; \
183 	vpxor 7 * 16(mem_cd), x3, x3;
211 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
223 	vmovdqu x3, 7 * 16(mem_cd); \
228 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
230 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
232 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
237 	vmovdqu x3, 3 * 16(mem_ab); \
243 #define enc_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
245 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
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, \
252 #define dec_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
254 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
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, \
418 #define transpose_4x4(x0, x1, x2, x3, t1, t2) \  argument
422 	vpunpckldq x3, x2, t1; \
423 	vpunpckhdq x3, x2, x2; \
428 	vpunpckhqdq x2, t2, x3; \
481 #define inpack16_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
498 	vpxor 12 * 16(rio), x0, x3; \
504 #define inpack16_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
506 	byteslice_16x16b(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \
512 	vmovdqu x3, 3 * 16(mem_ab); \
527 #define outunpack16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
530 			 y7, x3, x7, stack_tmp0, stack_tmp1); \
549 	vpxor x0, x3, x3; \
554 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
559 	vmovdqu x3, 3 * 16(rio); \