66 #define roundsm32(x0, x1, x2, x3, x4, x5, x6, x7, t0, t1, t2, t3, t4, t5, t6, \  argument
92 	vextracti128 $1, x0, t0##_x; \
109 	vaesenclast t4##_x, t0##_x, t0##_x; \
110 	vinserti128 $1, t0##_x, x0, x0; \
122 	vbroadcasti128 .Lpost_tf_lo_s1, t0; \
140 	filter_8bit(x0, t0, t1, t7, t6); \
141 	filter_8bit(x7, t0, t1, t7, t6); \
142 	filter_8bit(x3, t0, t1, t7, t6); \
143 	filter_8bit(x6, t0, t1, t7, t6); \
151 	vpbroadcastq key, t0; /* higher 64-bit duplicate ignored */ \
158 	vpsrldq $1, t0, t1; \
159 	vpsrldq $2, t0, t2; \
161 	vpsrldq $3, t0, t3; \
170 	vpsrldq $4, t0, t4; \
172 	vpsrldq $5, t0, t5; \
180 	vpsrldq $6, t0, t6; \
199 	vpsrldq $7, t0, t6; \
200 	vpshufb t7, t0, t0; \
221 	vpxor t0, x7, x7; \
306 #define rol32_1_32(v0, v1, v2, v3, t0, t1, t2, zero) \  argument
307 	vpcmpgtb v0, zero, t0; \
309 	vpabsb t0, t0; \
319 	vpor t0, v1, v1; \
321 	vpcmpgtb v3, zero, t0; \
323 	vpabsb t0, t0; \
327 	vpor t0, v0, v0;
336 #define fls32(l, l0, l1, l2, l3, l4, l5, l6, l7, r, t0, t1, t2, t3, tt0, \  argument
343 	vpbroadcastd kll, t0; /* only lowest 32-bit used */ \
345 	vpshufb tt0, t0, t3; \
346 	vpsrldq $1, t0, t0; \
347 	vpshufb tt0, t0, t2; \
348 	vpsrldq $1, t0, t0; \
349 	vpshufb tt0, t0, t1; \
350 	vpsrldq $1, t0, t0; \
351 	vpshufb tt0, t0, t0; \
353 	vpand l0, t0, t0; \
358 	rol32_1_32(t3, t2, t1, t0, tt1, tt2, tt3, tt0); \
360 	vpxor l4, t0, l4; \
361 	vpbroadcastd krr, t0; /* only lowest 32-bit used */ \
376 	vpshufb tt0, t0, t3; \
377 	vpsrldq $1, t0, t0; \
378 	vpshufb tt0, t0, t2; \
379 	vpsrldq $1, t0, t0; \
380 	vpshufb tt0, t0, t1; \
381 	vpsrldq $1, t0, t0; \
382 	vpshufb tt0, t0, t0; \
384 	vpor 4 * 32(r), t0, t0; \
389 	vpxor 0 * 32(r), t0, t0; \
393 	vmovdqu t0, 0 * 32(r); \
394 	vpbroadcastd krl, t0; /* only lowest 32-bit used */ \
404 	vpshufb tt0, t0, t3; \
405 	vpsrldq $1, t0, t0; \
406 	vpshufb tt0, t0, t2; \
407 	vpsrldq $1, t0, t0; \
408 	vpshufb tt0, t0, t1; \
409 	vpsrldq $1, t0, t0; \
410 	vpshufb tt0, t0, t0; \
412 	vpand 0 * 32(r), t0, t0; \
417 	rol32_1_32(t3, t2, t1, t0, tt1, tt2, tt3, tt0); \
419 	vpxor 4 * 32(r), t0, t0; \
423 	vmovdqu t0, 4 * 32(r); \
424 	vpbroadcastd klr, t0; /* only lowest 32-bit used */ \
435 	vpshufb tt0, t0, t3; \
436 	vpsrldq $1, t0, t0; \
437 	vpshufb tt0, t0, t2; \
438 	vpsrldq $1, t0, t0; \
439 	vpshufb tt0, t0, t1; \
440 	vpsrldq $1, t0, t0; \
441 	vpshufb tt0, t0, t0; \
443 	vpor l4, t0, t0; \
448 	vpxor l0, t0, l0; \