- x0 = _mm_insert_epi32(ivx, bswap32(cc + 0), 3);
- x1 = _mm_insert_epi32(ivx, bswap32(cc + 1), 3);
- x2 = _mm_insert_epi32(ivx, bswap32(cc + 2), 3);
- x3 = _mm_insert_epi32(ivx, bswap32(cc + 3), 3);
+ x0 = _mm_insert_epi32(ivx, br_bswap32(cc + 0), 3);
+ x1 = _mm_insert_epi32(ivx, br_bswap32(cc + 1), 3);
+ x2 = _mm_insert_epi32(ivx, br_bswap32(cc + 2), 3);
+ x3 = _mm_insert_epi32(ivx, br_bswap32(cc + 3), 3);