On Wed, Apr 17, 2024 at 04:34:07PM +0200, Jan Hubicka wrote:
> I think for most scenarios this is OK, but I am not sure about
> incremental linking (both LTO and non-LTO). What seems wrong is that we
> produce C5 comdat that is not exporting what it should and thus breaking
> the invariant that in valid code all comdats of same name are
> semantically equivalent.
Yeah, exactly. That is what I'm worried about too.
> Perhaps it makes no difference since this
> scenario is pretty special and we know that the functions are
> semantically equivalent and their addresses are never compared for
> equality (at least I failed to produce some useful testcase).
Yes, I think one can't take address of a constructor/destructor and compare
that for equality; I guess the destructor address can be stored in vtables,
but code manually reading stuff from vtables and assuming pointer equality
is almost certainly not valid.
Jakub
