Thanks, Nathan!

On 7/26/07, Nathan Sidwell <nathan@codesourcery.com> wrote:
> Index: cp/semantics.c
> +        case BIT_FIELD_REF:
> +          sorry ("bit-field accesses in a decltype expression");
> +          return error_mark_node;
>
> +        default:
> +          if (TYPE_P (expr) || DECL_P (expr))
> +            error ("argument to decltype must be an expression");
> +          else
> +            sorry ("unhandled decltype expression kind: %s",
> +                   tree_code_name[(int) TREE_CODE (expr)]);
>
> why sorry and not gcc_assert or gcc_unreachable?

The first sorry(), for BIT_FIELD_REF, should be a gcc_unreachable; it
just can't happen here. The second "sorry" I'd like to keep... if we
get here, it means that I somehow missed a particular member-access
expression node. So, apologize to the user and when they report the
bug, we'll see what the expression code is to fix the bug. Perhaps
internal_error would be better?

> +      if (TREE_CODE (expr) == CALL_EXPR
> +          && (fndecl = get_callee_fndecl (expr))
> +          && (fndecl != error_mark_node))
>
> +      else if ((type = is_bitfield_expr_with_lowered_type (expr)))
>
> ow :)  can we avoid assignments in conditionals here?

Heh, sure.

> Index: cp/parser.c
> +static tree
> +cp_parser_decltype (cp_parser *parser)
>
> here you parse tentatively, but don't appear to be committing to the tentative
> parse when it succeeds.

Ah, thanks. The fix is:

  if (id_expression_or_member_access_p)
    /* We have parsed the complete id-expression or member access.  */
    cp_parser_parse_definitely (parser);
  else
    {
      /* Abort our attempt to parse an id-expression or member access
         expression.  */
      cp_parser_abort_tentative_parse (parser);

      /* Parse a full expression.  */
      expr = cp_parser_expression (parser, /*cast_p=*/false);
    }

Essentially, by the time we reach this spot, we've either parsed an
id-expression or class member access expression, in which case there
is nothing left to tentatively parse (so we commit); or, we need to
abort the previous attempts and parse a full expression.

> Also, I can't see where decltype is gated on using c++0x extensions

In lex.c, if the keyword is only available when in C++0x mode, then
the feature is only available in C++0x mode:

  { "decltype",         RID_DECLTYPE,   D_CXX0X },

That said, decltype can save us (and the compiler) a LOT of work in
libstdc++, so I would suggest also:

  { "__decltype",       RID_DECLTYPE,   0 },

Updated patch attached, with those changes/fixes mention above.
Regtested powerpc-apple-darwin8.10.0, no regressions. Okay?

  - Doug