62 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
76 	vpshufb t4, x7, x7; \
87 	filter_8bit(x7, t5, t6, t7, t4); \
89 	vextracti128 $1, x7, t1##_x; \
107 	vaesenclast t4##_x, x7##_x, x7##_x; \
109 	vinserti128 $1, t1##_x, x7, x7; \
137 	filter_8bit(x7, t0, t1, t7, t6); \
162 	vpxor x7, x2, x2; \
174 	vpxor x1, x7, x7; \
180 	vpxor x7, x0, x0; \
188 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
217 	vpxor t0, x7, x7; \
218 	vpxor 3 * 32(mem_cd), x7, x7;
244 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
256 	vmovdqu x7, 3 * 32(mem_cd); \
261 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
263 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
265 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
270 	vmovdqu x7, 7 * 32(mem_ab); \
276 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
278 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
280 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
282 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
285 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
287 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
289 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
291 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
514 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
527 	vpxor 8 * 32(rio), x0, x7; \
537 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
539 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
549 	vmovdqu x7, 7 * 32(mem_ab); \
560 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
563 			      y3, y7, x3, x7, stack_tmp0, stack_tmp1); \
578 	vpxor x0, x7, x7; \
587 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
596 	vmovdqu x7, 7 * 32(rio); \