Well, your memory model might fail to properly expose "atomic" access at the level you're actually relying on to prove that your 'data race' is safe. (Or using such fine-grained atomic accesses might have bad performance implications that can't be worked around using relaxed operations, etc.) But I don't think that's going to be a common case.

But any optimizing compiler is surely going to assume its memory model is adhered to, no? Therefore any "data races" will be undefined behaviour (even if they're atomic on physical hardware).

"Any" and "surely" are very strong words! Compilers implement different kinds of languages, guarantees, extensions and what not.

Even if your program is not conforming to whatever standard, it may be very well be useful and working in most compilers out there.

As an example, Linux is not conforming C, including triggering UB. But at least two major compilers support the intended semantics even if they are forbidden by the standard.

UB is UB. It might work. It might work now but not in the future. It might work in some situation but not others. It might appear to work but introduce subtle bugs elsewhere in the program.

Linux, as Linus has said, isn't written in ISO C it's written (or was) in GNU C. The two had something of a symbiotic relationship for a long time. Which is why porting it to LLVM was such a tremendous effort.

Which is to say, Linux conforms to the memory model of the compilers it compiles on (or vice versa). Not ISO C's memory model. So I'm not sure how your example is relevant.

There has been a significant push to get Linux to use the c11 conformant gcc builtins. There has been significant push back though, but at least some bits have been updated.

Linux arguably relies on compiler extensions. Compilers are allowed to define behavior that is undefined by the language standards, effectively extending the language.

No, it is not "dangerous". CPUs are designed to deal with that all the time (it is a whole field of research) and give you different guarantees when it does happen.

If your software is designed with that in mind, then a data race is not a bug.

I'm not sure I understand you. If it's not dangerous and there are guarantees then there isn't a data race to begin with.