<atomic>: Add member difference_type to atomic<void*> and its friends

[frederick-vs-ja]

Fixes #4688.

It seems intentional that atomic(_ref)<T*> are SFINAE-friendly on atomic pointer arithmetic. So this PR adds another mediate base class to avoid breaking the SFINAE-friendliness.

Also updates the comments in Dev11_0863628_atomic_compare_exchange/test.cpp to cite WG21-N4981.

[cpplearner]

It seems intentional that atomic(_ref)<T*> are SFINAE-friendly on atomic pointer arithmetic. So this PR adds another mediate base class to avoid breaking the SFINAE-friendliness.

I don't think that matches the standard, which uses "Mandates" ([atomics.types.pointer]/5, [atomics.ref.pointer]/3).

[frederick-vs-ja]

It seems intentional that atomic(_ref)<T*> are SFINAE-friendly on atomic pointer arithmetic. So this PR adds another mediate base class to avoid breaking the SFINAE-friendliness.

I don't think that matches the standard, which uses "Mandates" ([atomics.types.pointer]/5, [atomics.ref.pointer]/3).

Mandates doesn't mean the ill-formedness shall be SFINAE-unfriendly, although we ususally use static_asserts or straightforward hard errors for it.

Previously, #4014 made Mandates for forward_like SFINAE-friendly.

Edit: see also #3013 (comment).

[StephanTLavavej]

Thanks for the quick fix and comprehensive test coverage, this is great! 😻

[StephanTLavavej]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[StephanTLavavej]

Thanks for making a difference! 😹 🤪 💚

[frederick-vs-ja]

😹Well, we should partially revert this for atomic_ref<NonObject*> due to P3323R1. Although atomic<NonObject*> should still have difference_type.