66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
83 	vpshufb t4, x2, x2; \
98 	filter_8bit(x2, t5, t6, t7, t4); \
106 	vextracti128 $1, x2, t6##_x; \
124 	vaesenclast t4##_x, x2##_x, x2##_x; \
126 	vinserti128 $1, t6##_x, x2, x2; \
148 	filter_8bit(x2, t2, t3, t7, t6); \
166 	vpxor x7, x2, x2; \
175 	vpxor x2, x4, x4; \
186 	vpxor x5, x2, x2; \
192 	vpxor x2, x7, x7; /* note: high and low parts swapped */ \
206 	vpxor t5, x2, x2; \
207 	vpxor 6 * 32(mem_cd), x2, x2; \
250 #define two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
257 	vmovdqu x2, 6 * 32(mem_cd); \
267 	store_ab(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab);
269 #define dummy_store(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) /* do nothing */  argument
271 #define store_ab_state(x0, x1, x2, x3, x4, x5, x6, x7, mem_ab) \  argument
279 	vmovdqu x2, 2 * 32(mem_ab); \
282 #define enc_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
284 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
286 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
288 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
291 #define dec_rounds32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
293 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
295 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
297 	two_roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \
457 #define transpose_4x4(x0, x1, x2, x3, t1, t2) \  argument
461 	vpunpckldq x3, x2, t1; \
462 	vpunpckhdq x3, x2, x2; \
467 	vpunpckhqdq x2, t2, x3; \
468 	vpunpcklqdq x2, t2, x2;
520 #define inpack32_pre(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
538 	vpxor 13 * 32(rio), x0, x2; \
543 #define inpack32_post(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
545 	byteslice_16x16b_fast(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, \
550 	vmovdqu x2, 2 * 32(mem_ab); \
566 #define outunpack32(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, \  argument
568 	byteslice_16x16b_fast(y0, y4, x0, x4, y1, y5, x1, x5, y2, y6, x2, x6, \
589 	vpxor x0, x2, x2; \
593 #define write_output(x0, x1, x2, x3, x4, x5, x6, x7, y0, y1, y2, y3, y4, y5, \  argument
597 	vmovdqu x2, 2 * 32(rio); \