64 		sk[u] = _mm_loadu_si128((void *)(ctx->skey.skni + (u << 4)));  in br_aes_x86ni_ctrcbc_ctr()
70 	erev = _mm_set_epi8(0, 1, 2, 3, 4, 5, 6, 7,  in br_aes_x86ni_ctrcbc_ctr()
74 	four = _mm_set_epi64x(0, 4);  in br_aes_x86ni_ctrcbc_ctr()
92 		__m128i x0, x1, x2, x3;  in br_aes_x86ni_ctrcbc_ctr()  local
100 		x2 = _mm_shuffle_epi8(ivx2, erev);  in br_aes_x86ni_ctrcbc_ctr()
105 		x2 = _mm_xor_si128(x2, sk[0]);  in br_aes_x86ni_ctrcbc_ctr()
109 		x2 = _mm_aesenc_si128(x2, sk[1]);  in br_aes_x86ni_ctrcbc_ctr()
113 		x2 = _mm_aesenc_si128(x2, sk[2]);  in br_aes_x86ni_ctrcbc_ctr()
117 		x2 = _mm_aesenc_si128(x2, sk[3]);  in br_aes_x86ni_ctrcbc_ctr()
119 		x0 = _mm_aesenc_si128(x0, sk[4]);  in br_aes_x86ni_ctrcbc_ctr()
120 		x1 = _mm_aesenc_si128(x1, sk[4]);  in br_aes_x86ni_ctrcbc_ctr()
121 		x2 = _mm_aesenc_si128(x2, sk[4]);  in br_aes_x86ni_ctrcbc_ctr()
122 		x3 = _mm_aesenc_si128(x3, sk[4]);  in br_aes_x86ni_ctrcbc_ctr()
125 		x2 = _mm_aesenc_si128(x2, sk[5]);  in br_aes_x86ni_ctrcbc_ctr()
129 		x2 = _mm_aesenc_si128(x2, sk[6]);  in br_aes_x86ni_ctrcbc_ctr()
133 		x2 = _mm_aesenc_si128(x2, sk[7]);  in br_aes_x86ni_ctrcbc_ctr()
137 		x2 = _mm_aesenc_si128(x2, sk[8]);  in br_aes_x86ni_ctrcbc_ctr()
141 		x2 = _mm_aesenc_si128(x2, sk[9]);  in br_aes_x86ni_ctrcbc_ctr()
146 			x2 = _mm_aesenclast_si128(x2, sk[10]);  in br_aes_x86ni_ctrcbc_ctr()
151 			x2 = _mm_aesenc_si128(x2, sk[10]);  in br_aes_x86ni_ctrcbc_ctr()
155 			x2 = _mm_aesenc_si128(x2, sk[11]);  in br_aes_x86ni_ctrcbc_ctr()
159 			x2 = _mm_aesenclast_si128(x2, sk[12]);  in br_aes_x86ni_ctrcbc_ctr()
164 			x2 = _mm_aesenc_si128(x2, sk[10]);  in br_aes_x86ni_ctrcbc_ctr()
168 			x2 = _mm_aesenc_si128(x2, sk[11]);  in br_aes_x86ni_ctrcbc_ctr()
172 			x2 = _mm_aesenc_si128(x2, sk[12]);  in br_aes_x86ni_ctrcbc_ctr()
176 			x2 = _mm_aesenc_si128(x2, sk[13]);  in br_aes_x86ni_ctrcbc_ctr()
180 			x2 = _mm_aesenclast_si128(x2, sk[14]);  in br_aes_x86ni_ctrcbc_ctr()
188 			x2 = _mm_xor_si128(x2,  in br_aes_x86ni_ctrcbc_ctr()
194 			_mm_storeu_si128((void *)(buf + 32), x2);  in br_aes_x86ni_ctrcbc_ctr()
203 			_mm_storeu_si128((void *)(tmp + 32), x2);  in br_aes_x86ni_ctrcbc_ctr()
223 		 * Add 4 to each counter value. For carry propagation  in br_aes_x86ni_ctrcbc_ctr()
225 		 * the results with 4, but SSE2+ has only _signed_  in br_aes_x86ni_ctrcbc_ctr()
269 		sk[u] = _mm_loadu_si128((void *)(ctx->skey.skni + (u << 4)));  in br_aes_x86ni_ctrcbc_mac()
279 		x = _mm_aesenc_si128(x, sk[4]);  in br_aes_x86ni_ctrcbc_mac()
321 		sk[u] = _mm_loadu_si128((void *)(ctx->skey.skni + (u << 4)));  in br_aes_x86ni_ctrcbc_encrypt()
327 	erev = _mm_set_epi8(0, 1, 2, 3, 4, 5, 6, 7,  in br_aes_x86ni_ctrcbc_encrypt()
361 		x0 = _mm_aesenc_si128(x0, sk[4]);  in br_aes_x86ni_ctrcbc_encrypt()
362 		x1 = _mm_aesenc_si128(x1, sk[4]);  in br_aes_x86ni_ctrcbc_encrypt()
424 			cmx = _mm_aesenc_si128(cmx, sk[4]);  in br_aes_x86ni_ctrcbc_encrypt()
469 		sk[u] = _mm_loadu_si128((void *)(ctx->skey.skni + (u << 4)));  in br_aes_x86ni_ctrcbc_decrypt()
475 	erev = _mm_set_epi8(0, 1, 2, 3, 4, 5, 6, 7,  in br_aes_x86ni_ctrcbc_decrypt()
508 		x0 = _mm_aesenc_si128(x0, sk[4]);  in br_aes_x86ni_ctrcbc_decrypt()
509 		x1 = _mm_aesenc_si128(x1, sk[4]);  in br_aes_x86ni_ctrcbc_decrypt()
570 	4,