66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
75 	vbroadcasti128 .Lpre_tf_lo_s4, t2; \
94 	filter_8bit(x3, t2, t3, t7, t4); \
95 	filter_8bit(x6, t2, t3, t7, t4); \
97 	vextracti128 $1, x6, t2##_x; \
118 	vaesenclast t4##_x, t2##_x, t2##_x; \
119 	vinserti128 $1, t2##_x, x6, x6; \
121 	vextracti128 $1, x4, t2##_x; \
134 	vaesenclast t4##_x, t2##_x, t2##_x; \
135 	vinserti128 $1, t2##_x, x4, x4; \
138 	vbroadcasti128 .Lpost_tf_lo_s3, t2; \
148 	filter_8bit(x2, t2, t3, t7, t6); \
149 	filter_8bit(x5, t2, t3, t7, t6); \
154 	filter_8bit(x1, t4, t5, t7, t2); \
155 	filter_8bit(x4, t4, t5, t7, t2); \
159 	vpsrldq $2, t0, t2; \
169 	vpshufb t7, t2, t2; \
215 	vpxor t2, x5, x5; \
306 #define rol32_1_32(v0, v1, v2, v3, t0, t1, t2, zero) \  argument
315 	vpcmpgtb v2, zero, t2; \
317 	vpabsb t2, t2; \
326 	vpor t2, v3, v3; \
336 #define fls32(l, l0, l1, l2, l3, l4, l5, l6, l7, r, t0, t1, t2, t3, tt0, \  argument
347 	vpshufb tt0, t0, t2; \
355 	vpand l2, t2, t2; \
358 	rol32_1_32(t3, t2, t1, t0, tt1, tt2, tt3, tt0); \
365 	vpxor l6, t2, l6; \
378 	vpshufb tt0, t0, t2; \
386 	vpor 6 * 32(r), t2, t2; \
391 	vpxor 2 * 32(r), t2, t2; \
396 	vmovdqu t2, 2 * 32(r); \
406 	vpshufb tt0, t0, t2; \
414 	vpand 2 * 32(r), t2, t2; \
417 	rol32_1_32(t3, t2, t1, t0, tt1, tt2, tt3, tt0); \
421 	vpxor 6 * 32(r), t2, t2; \
426 	vmovdqu t2, 6 * 32(r); \
437 	vpshufb tt0, t0, t2; \
445 	vpor l6, t2, t2; \
452 	vpxor l2, t2, l2; \
457 #define transpose_4x4(x0, x1, x2, x3, t1, t2) \  argument
458 	vpunpckhdq x1, x0, t2; \
467 	vpunpckhqdq x2, t2, x3; \
468 	vpunpcklqdq x2, t2, x2;