Maybe replace the < 32 bytes implementation to avoid the double scan + two functions call, if it happens to be an issue with production data.