Compilers are usually very reluctant to add "surprising" optimizations to atomics, even if they're fully justified by the standard, because (1) there are a lot

Compilers are usually very reluctant to add "surprising" optimizations to atomics, even if they're fully justified by the standard, because (1) there are a lot of compiler bugs and/or holes in the standard that are only exposed after combining multiple optimizations involving atomics, (2) most programs explicitly use atomics only rarely, making it not the most productive thing to optimize, and (3) the programs and libraries that do use them are often reliant on them for synchronization, "liveness," or performance behavior that's not mandated by the standard. Common pathological examples include checks that assume the compiler won't skip assigning values if it can prove there's some obscure possible execution where the intermediate value can never be read, assuming that checking the result of a relaxed load in an if statement will guarantee that accesses run in the if will be sequenced after the read, etc.