Isn't this the purpose of well-defined atomic primitives?

After all, not just the compiler, but also the processor can reorder operations. So you have to annotate synchronizing memory operations regardless of whether the compiler is optimizing. e.g., a lock-free algorithm implemented using only volatile (what ACCESS_ONCE does), even with -O0, is almost certainly wrong.

The alternative to explicit annotation is for the compiler to generate full memory barriers around every memory access. That would indeed preserve semantics in a multithreaded context, at a ridiculous performance cost.