Parallel and high-performance computing

Several threads increment one shared integer, and the platform provides a supported atomic fetch-and-add operation. Why can this operation safely replace an unprotected read, add, and write sequence?