ICE due to new bitfield algorithm (PR65742)

[urnathan]

Sadly, #65742 (commit 49839f9) introduced a bug for a corner case of vbase placement:

struct __attribute__((aligned(4))) Empty {} empty;

struct Char { char a; } cbase;

struct D : virtual Char {
  [[no_unique_address]] Empty e0;
  [[no_unique_address]] Empty e1;
  unsigned a : 24; // keep as 24bits
} d;

in the complete D, the Char vbase is placed 3 bytes after the vptr (i.e. at offset 7 or 11), so the bitfield must remain 24-bits. We need dsize to limit the bitfield expansion, not nvsize. They're nearly always the same, except for cases like this involving multiple zero-sized members (or bases).

I will shortly be posting a patch to address this. (FWIW, I don't expect the triggering cases to be at all common).

[llvmbot]

@llvm/issue-subscribers-clang-codegen

Author: Nathan Sidwell (urnathan)

Sadly, #65742 (commit 49839f9) introduced a bug for a corner case of vbase placement: ``` struct __attribute__((aligned(4))) Empty {} empty;

struct Char { char a; } cbase;

struct D : virtual Char {
[[no_unique_address]] Empty e0;
[[no_unique_address]] Empty e1;
unsigned a : 24; // keep as 24bits
} d;

in the complete D, the Char vbase is placed 3 bytes after the vptr (i.e. at offset 7 or 11), so the bitfield must remain 24-bits.  We need dsize to limit the bitfield expansion, not nvsize.  They're nearly always the same, except for cases like this involving multiple zero-sized members (or bases).

I will shortly be posting a patch to address this. (FWIW, I don't expect the triggering cases to be at all common).
</details>