[libc++] ABI compat breakage using reference type for std::set key_compare

[hubert-reinterpretcast]

The following compiled fine with libc++ version 19:

#include <set>

struct Cmp {
  __int128 x;
  bool operator()(__int128, __int128) const;
};

struct MySet {
  unsigned char b;
  std::set<__int128, const Cmp &> s;
};

extern char q[sizeof(MySet)];
extern char q[40];

It no longer compiles with libc++ trunk (https://godbolt.org/z/cTbEovcer). It seems the size of MySet has changed because of a change in std::set.
The draft release notes (https://libcxx.llvm.org/ReleaseNotes.html#abi-affecting-changes) mention issues with overaligned empty types, but nothing about reference types.

[rnk]

I am personally surprised by how alignof works on references, but it is consistent with sizeof. alignof(__int128&) is 16, not 8 or pointer-aligned. That seems to be the root cause here.

__compressed_pair_padding has some accommodation to avoid adding extra padding around references, but there isn't analogous logic to fix up the alignment attribute.

[jwakely]

std::set<__int128, const Cmp &> is somewhere in the intersection of underspecified, undefined, and just a very bad idea.

[hubert-reinterpretcast]

std::set<__int128, const Cmp &> is somewhere in the intersection of underspecified, undefined, and just a very bad idea.

I tend to agree; however, from that point of view, I think a deliberate compile-time error is better than the ABI break (documented or not).

[ldionne]

I do agree that reference comparators are a terrible idea. But this was an unintended ABI break and (unless I missed something) this is easy to address, at least until we officially decide to make such comparators ill-formed: #118685