Rule. In x86 SIMD routines, use width-matched load helpers or intrinsics (such as xx_loadl_32 or _mm_cvtsi32_si128 for 4-byte/32-bit rows, and xx_loadl_64 or _mm_loadl_epi64 for 8-byte/64-bit rows) when reading lines, reference buffers, or mask rows narrower than full vector registers. Do not use wider or full-register loads (such as _mm_loadl_epi64 on 4-byte buffers or xx_loadu_128 / _mm_loadu_si128 on 8-byte rows) on buffers that are only allocated or valid for the narrower width.
Why. Loading 8 or 16 bytes from buffers or row strides that only allocate 4 or 8 valid bytes over-reads past allocated buffer boundaries or row strides. This causes memory over-read errors, AddressSanitizer violations, or potential segmentation faults when processing narrow blocks near image or buffer edges.
Bad:
// 4-byte reference buffer: over-reading 8 bytes with _mm_loadl_epi64 __m128i d = _mm_loadl_epi64((const __m128i *)above); // 8-byte mask row: over-reading 16 bytes with xx_loadu_128 const __m128i v_ra = xx_loadu_128(mask); const __m128i v_rb = xx_loadu_128(mask + mask_stride);
Good:
// 4-byte reference buffer: load exact 4 bytes (32 bits) __m128i d = _mm_cvtsi32_si128(*(const uint32_t *)above); // 8-byte mask row: load exact 8 bytes (64 bits) const __m128i v_ra = xx_loadl_64(mask); const __m128i v_rb = xx_loadl_64(mask + mask_stride);