crypto/aesni/aesni_ghash.c

110 	__m128i tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8, tmp9;
145 	tmp2 = _mm_srli_epi32(tmp3, 1);
148 	tmp2 = _mm_xor_si128(tmp2, tmp4);
149 	tmp2 = _mm_xor_si128(tmp2, tmp5);
150 	tmp2 = _mm_xor_si128(tmp2, tmp8);
151 	tmp3 = _mm_xor_si128(tmp3, tmp2);
167 	__m128i tmp0, tmp1, tmp2, tmp3;
197 	tmp2 = _mm_shuffle_epi32(H3, 78);
199 	tmp2 = _mm_xor_si128(tmp2, H3);
208 	tmp2 = _mm_clmulepi64_si128(tmp2, tmp6, 0x00);
214 	tmp0 = _mm_xor_si128(tmp2, tmp0);
248 	tmp2 = _mm_srli_epi32(tmp3, 1);
251 	tmp2 = _mm_xor_si128(tmp2, tmp4);
252 	tmp2 = _mm_xor_si128(tmp2, tmp5);
253 	tmp2 = _mm_xor_si128(tmp2, tmp8);
254 	tmp3 = _mm_xor_si128(tmp3, tmp2);
275 	__m128i tmp1, tmp2, tmp3, tmp4;
295 		tmp2 = _mm_xor_si128(Y, KEY[0]);
298 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j]);
301 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j+1]);
304 		tmp2 = _mm_aesenc_si128(tmp2, KEY[nr-1]);
307 		T = _mm_aesenclast_si128(tmp2, KEY[nr]);
351 		tmp2 = _mm_loadu_si128(&((const __m128i *)addt)[i*4+1]);
356 		tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK);
361 		reduce4(H, H2, H3, H4, tmp4, tmp3, tmp2, tmp1, &X);
391 		tmp2 = _mm_shuffle_epi8(ctr2, BSWAP_EPI64);
409 		tmp2 =_mm_xor_si128(tmp2, KEY[0]);
419 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j]);
428 		tmp2 =_mm_aesenclast_si128(tmp2, KEY[nr]);
438 		tmp2 = _mm_xor_si128(tmp2,
454 		_mm_storeu_si128(&((__m128i*)out)[i*8+1], tmp2);
463 		tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK);
473 		reduce4(H, H2, H3, H4, tmp4, tmp3, tmp2, tmp1, &X);
536 	__m128i tmp1, tmp2, tmp3, tmp4;
556 		tmp2 = _mm_xor_si128(Y, KEY[0]);
559 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j]);
562 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j+1]);
565 		tmp2 = _mm_aesenc_si128(tmp2, KEY[nr-1]);
568 		T = _mm_aesenclast_si128(tmp2, KEY[nr]);
612 		tmp2 = _mm_loadu_si128(&((const __m128i *)addt)[i*4+1]);
617 		tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK);
623 		reduce4(H, H2, H3, H4, tmp4, tmp3, tmp2, tmp1, &X);
644 		tmp2 = _mm_loadu_si128(&((const __m128i *)in)[i*4+1]);
649 		tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK);
655 		reduce4(H, H2, H3, H4, tmp4, tmp3, tmp2, tmp1, &X);
696 		tmp2 = _mm_shuffle_epi8(ctr2, BSWAP_EPI64);
714 		tmp2 =_mm_xor_si128(tmp2, KEY[0]);
724 			tmp2 = _mm_aesenc_si128(tmp2, KEY[j]);
733 		tmp2 =_mm_aesenclast_si128(tmp2, KEY[nr]);
743 		tmp2 = _mm_xor_si128(tmp2,
759 		_mm_storeu_si128(&((__m128i*)out)[i*8+1], tmp2);
768 		tmp2 = _mm_shuffle_epi8(tmp2, BSWAP_MASK);