A strategy I've seen in similar scenarios is to apply a kind of inverse-ordered read to the last chunk. Instead of reading bytes x through x+n (your actual data

A strategy I've seen in similar scenarios is to apply a kind of inverse-ordered read to the last chunk. Instead of reading bytes x through x+n (your actual data) and x+n+1 through x+15 (outside the edge of the buffer) you instead read bytes x+n-15 through x+n and then apply whatever logic you're doing to mask or exclude the unimportant bytes to the beginning of that chunk rather than the end. Note: you still have to be careful if your entire input is less than 16 bytes, there can be other alignment issues, etc..., but it _might_ be better in your situation than a similar fix that other comments have mentioned which applies the fast read to all but the last partial chunk and then finishes the tail with some slow cleanup code.