These kinds of analyses are all nice and funny and so on but the issue here is that on real computers it does not quite work like this. One thing is that an optimizer may change the order in which statements are executed and then all guarantees go out of the window. Another is when writing to main memory the hardware may also reorder writes. So the whole process is filled with gotchas on every level. What one should remember is that if multiple threads do things with the same memory location where one of the 'things' that are done is writing, it is always wrong. The way to fix it then is to use the appropriate protection. E.g., use a mutex or use an atomic variable.
Heh this reminds me of a really good book, The Art of Multiprocessor Programming [1]. It has two parts, the first one goes quite a hit into the theory. And the second part begins with a real example where the theory doesn’t hold and it’s aptly titled “Welcome to The Real World”.
