[Bug c++/111284] [11/12/13/14 Regression] Some passing-by-value parameters are mishandled since GCC 9

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111284

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Actually, the binding of parms seems to work fine.
But am out of ideas where this should be fixed.  When evaluating the
comparison, we have lhs
(const struct S *) (struct S *) ((const struct S *) this)->s
and rhs
(const struct S *) this
and we constant evaluate both.  In both cases we find this (foo's artificial
argument in #c2) is &x where x is bar's PARM_DECL.  But in the lhs case, -> is
applied to it,
so x is evaluated as prvalue as {.s=&D.2836} where D.2836 is the local temp
built for AGGR_INIT_EXPR - but nothing during the constexpr evaluation actually
knows that &D.2836 is actually the address of the x argument.  So, lhs folds
into &D.2836 with some cast.
While then constexpr evaluating rhs sees just accessing this which is &x and in
that case x is used as an lvalue, so the folds into &x.  And those two are not
equal.
So, I think we need to treat somewhere the arguments (and return values?) with
TREE_ADDRESSABLE types during constexpr evaluations actually as references
rather than what they are in the source (so what we do after genericization).