Add layout rule for [[no_unique_address]] (C++20 P0840R2)

[zygoloid]

Fixes #49

[zygoloid]

For exposition: the idea here is to use max(nvsize(D),dsize(D)) as the allocated size for potentially-overlapping data members of type D in the same way we use nvsize(D) as the allocated size for base classes. We previously only used dsize(D) as an intermediate value when laying out D, and now becomes part of the data propagated to other classes' layouts.

We need to take the max of these two values because:

• we want [[no_unique_address]] to give the same layout as a base class whenever possible, which is whenever the field has no virtual bases of its own
• therefore in the "no virtual bases" case, we want to reserve nvsize(D) bytes (in this case, nvsize(D) is necessarily >= dsize(D))
• in the remaining cases, we need to reserve dsize(D) to make room for D's virtual bases

Note that there are corner cases where nvsize(C) > dsize(C)! This happens when we (strangely) increase nvsize(C) to include the complete size of an empty base class. (We do need to increase sizeof(C) for this case, but not nvsize(C).) Perhaps that's something we could consider fixing for ABI v2.

The rules could be simplified slightly by setting dsize(C) = max(dsize(C), nvsize(C)) at the end of class layout, and then only using dsize(C) in recursive layouts, but it seems unreasonable to change the dsize for existing classes -- who knows what exciting uses people have found for it :)

[rjmccall]

@zygoloid, would you mind checking my merge and verifying that this change still reflects the right rule according to the standard?

[jicama]

Shouldn't the "empty data members" be a qualification on "no non-static data members" rather than "no unnamed bit-fields"?

[rjmccall]

Yes, absolutely; that was a mistake introduced by my merge. Fixed.

[zygoloid]

Merged rule now looks good, thanks.

[rjmccall]

Okay. I'll leave this open for a week for further comment.

[rjmccall]

1 week, 4 weeks, eh.