49 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
62 	vpshufb t4, x1, x1; \
74 	filter_8bit(x1, t0, t1, t7, t6); \
89 	vaesenclast t4, x1, x1; \
114 	filter_8bit(x1, t4, t5, t7, t2); \
134 	vpxor x6, x1, x1; \
141 	vpxor x1, x7, x7; \
144 	vpxor x4, x1, x1; \
150 	vpxor x1, x6, x6; \
176 	vpxor t6, x1, x1; \
177 	vpxor 5 * 16(mem_cd), x1, x1; \
211 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
221 	vmovdqu x1, 5 * 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
235 	vmovdqu x1, 1 * 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
419 	vpunpckhdq x1, x0, t2; \
420 	vpunpckldq x1, x0, x0; \
425 	vpunpckhqdq t1, x0, x1; \
481 #define inpack16_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
500 	vpxor 14 * 16(rio), x0, x1; \
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, \
510 	vmovdqu x1, 1 * 16(mem_ab); \
527 #define outunpack16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
529 	byteslice_16x16b(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, y3, \
551 	vpxor x0, x1, x1; \
554 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
557 	vmovdqu x1, 1 * 16(rio); \