When writing single-threaded programs in C++, there is no risk of data races occurring. We can write our programs happily without being aware of instruction reorderings. However, when it comes to shared variables in multi-threaded programs, it is a completely different story. The compiler (and hardware) does all its optimizations based on what is true and observable for one thread only. The compiler cannot know what other threads are able to observe by shared variables, so it is our job as programmers to inform the compiler of what reorderings are allowed. In fact, that is exactly what we are doing when we are using an atomic variable or a mutex to protect us from data races.
When protecting a critical section with ...
