P2985R0 A type trait for detecting virtual base classes

[Endilll]

This paper introduces std::is_virtual_base_of trait, and was voted into the working draft in St. Louis.
Like other traits, we're planning to support this with a __is_virtual_base_of builtin.
Tagging @stl, @jwakely, and @ldionne to make sure they are fine with this spelling, and hopefully to prevent implementation divergence on the name.

[llvmbot]

@llvm/issue-subscribers-clang-frontend

Author: Vlad Serebrennikov (Endilll)

This paper introduces `std::is_virtual_base_of` trait, and was voted into the working draft in St. Louis. Like other traits, we're planning to support this with a `__is_virtual_base_of` builtin. Tagging @STL, @jwakely, and @ldionne to make sure they are fine with this spelling, and hopefully to prevent implementation divergence on the name.

[Endilll]

@StephanTLavavej

[ldionne]

The name LGTM. For what it's worth, I would have preferred if all of the type trait builtins started with __builtin (eg __builtin_is_signed(...)), but at this point we should just be consistent with everything else.

[StephanTLavavej]

Name LGTM too.

[jwakely]

For what it's worth, I would have preferred if all of the type trait builtins started with __builtin (eg __builtin_is_signed(...)), but at this point we should just be consistent with everything else.

Yes, we had a similar conversation in GCC recently. That would have avoided all the problems with new built-in traits clashing with existing class templates already using the name (e.g. __is_pointer, __is_assignable). But it seems far too late to change this now.

The suggested name seems coherent with existing practice.

[ldionne]

@jwakely Would it be controversial to propose that we support both the __builtin_XXXXX spelling and the normal spelling in compilers for a while, and perhaps document that we want to move away from the non-__builtin_XXXXX spellings? If we started now, we could be in a better world just a few years from now.

What do you think the GCC folks would say?

[jwakely]

I asked, and the C++ maintainers would be happy with that direction.

[AaronBallman]

I prefer the __builtin_ spellings and think that's more clear, and I'd be okay with deprecating the existing spellings and guiding people towards the new ones, but I'm not certain how I feel about introducing new traits with both spellings. Is that the request, or am I misreading?

[ldionne]

My thinking was to support both spellings for existing traits only, and introduce new traits as __builtin_xxxxx going forward.

In that case, it looks like we can start right away and this trait should probably be spelled __builtin_is_virtual_base_of.

[AaronBallman]

Thanks! I think that's a reasonable approach.