904   const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS - 1));  in VE4_SSE2()  local
905   const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);  in VE4_SSE2()
906   const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);  in VE4_SSE2()
907   const __m128i a = _mm_avg_epu8(ABCDEFGH, CDEFGH00);  in VE4_SSE2()
908   const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGH00), one);  in VE4_SSE2()
920   const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));  in LD4_SSE2()  local
921   const __m128i BCDEFGH0 = _mm_srli_si128(ABCDEFGH, 1);  in LD4_SSE2()
922   const __m128i CDEFGH00 = _mm_srli_si128(ABCDEFGH, 2);  in LD4_SSE2()
924   const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, CDEFGHH0);  in LD4_SSE2()
925   const __m128i lsb = _mm_and_si128(_mm_xor_si128(ABCDEFGH, CDEFGHH0), one);  in LD4_SSE2()
961   const __m128i ABCDEFGH = _mm_loadl_epi64((__m128i*)(dst - BPS));  in VL4_SSE2()  local
962   const __m128i BCDEFGH_ = _mm_srli_si128(ABCDEFGH, 1);  in VL4_SSE2()
963   const __m128i CDEFGH__ = _mm_srli_si128(ABCDEFGH, 2);  in VL4_SSE2()
964   const __m128i avg1 = _mm_avg_epu8(ABCDEFGH, BCDEFGH_);  in VL4_SSE2()
968   const __m128i ab = _mm_xor_si128(ABCDEFGH, BCDEFGH_);  in VL4_SSE2()