[Bug c++/68782] [6 regression] bad reference member formed with constexpr

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68782

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Jason Merrill from comment #3)
> Hmm, any element without TREE_CONSTANT should have caused us to return 
> the original CONSTRUCTOR.

Perhaps the TREE_SIDE_EFFECTS stuff is not needed, but for TREE_CONSTANT
perhaps the reason is that constexpr.c has different POV on what is a constant
compared to ../tree.c - at least it seems that cxx_eval_constant_expression
happily accepts &y->c as *non_constant_p = false, but if CONSTRUCTOR containing
that is marked TREE_CONSTANT (which is IMHO wrong, because in that case all
elements should be TREE_CONSTANT), then we e.g. trigger:
    case CONSTRUCTOR:
      if (TREE_CONSTANT (t))
        /* Don't re-process a constant CONSTRUCTOR, but do fold it to
           VECTOR_CST if applicable.  */
        return fold (t);
      r = cxx_eval_bare_aggregate (ctx, t, lval,
                                   non_constant_p, overflow_p);
      break;
and just fold it instead of calling cxx_eval_bare_aggregate on it.

> I thought there was already a function to recompute these flags, but I'm 
> not finding it.

I can't find it either, we have that only for ADDR_EXPR it seems -
recompute_tree_invariant_for_addr_expr.