50 #define roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
62 	vpshufb t4, x7, x7; \
74 	filter_8bit(x7, t0, t1, t7, t6); \
89 	vaesenclast t4, x7, x7; \
101 	filter_8bit(x7, t0, t1, t7, t6); \
136 	vpxor x7, x2, x2; \
142 	vpxor x1, x7, x7; \
144 	vpxor x7, x0, x0; \
152 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
171 	vpxor t0, x7, x7; \
172 	vpxor 3 * 16(mem_cd), x7, x7; \
212 #define two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
220 	vmovdqu x7, 3 * 16(mem_cd); \
229 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
231 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
233 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
242 	vmovdqu x7, 7 * 16(mem_ab);
244 #define enc_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
246 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
248 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
250 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
253 #define dec_rounds16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
255 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
257 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
259 	two_roundsm16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
482 #define inpack16_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
495 	vpxor 8 * 16(rio), x0, x7; \
505 #define inpack16_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
507 	byteslice_16x16b(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \
517 	vmovdqu x7, 7 * 16(mem_ab); \
528 #define outunpack16(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
531 			 y7, x3, x7, stack_tmp0, stack_tmp1); \
546 	vpxor x0, x7, x7; \
555 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
564 	vmovdqu x7, 7 * 16(rio); \