[Bug c++/83204] [6/7/8 Regression] c++ -std=c++14 ICE in maybe_undo_parenthesized_ref, at cp/semantics.c:1694

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83204

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2017-11-30
                 CC|                            |jakub at gcc dot gnu.org,
                   |                            |jason at gcc dot gnu.org
     Ever confirmed|0                           |1

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Started with r236616.
I think this is related to maybe_undo_parenthesized_ref relying on
REF_PARENTHESIZED_P operands being in a certain form, which is what
force_paren_expr creates.  But in pt.c we have:
        tree r = RECUR (TREE_OPERAND (t, 0));

        if (REFERENCE_REF_P (t))
          {
            /* A type conversion to reference type will be enclosed in
               such an indirect ref, but the substitution of the cast
               will have also added such an indirect ref.  */
            r = convert_from_reference (r);
          }
        else
          r = build_x_indirect_ref (input_location, r, RO_UNARY_STAR,
                                    complain|decltype_flag);

        if (TREE_CODE (r) == INDIRECT_REF)
          REF_PARENTHESIZED_P (r) = REF_PARENTHESIZED_P (t);
where we copy over the bit, even when the tsubsted operand might be completely
different.  So should maybe_undo_parenthesized_ref be able to cope with any
other form (and how), or should we not set REF_PARENTHESIZED_P if the form of
r's operands no longer matches, something else?