RFA: PATCH to handle_transparent_union_attribute for c++/51228 (ICE on bogus use of transparent_union)

[Jakub Jelinek]

On Wed, Dec 14, 2011 at 05:14:18PM -0500, Jason Merrill wrote:
> --- a/gcc/c-family/c-common.c
> +++ b/gcc/c-family/c-common.c
> @@ -6286,13 +6286,21 @@ handle_transparent_union_attribute (tree *node, tree name,
>  
>    if (TREE_CODE (type) == UNION_TYPE)
>      {
> -      /* When IN_PLACE is set, leave the check for FIELDS and MODE to
> -	 the code in finish_struct.  */
> +      /* Make sure that the first field will work for a transparent union.
> +	 If the type isn't complete yet, leave the check to the code in
> +	 finish_struct.  */
> +      if (TYPE_SIZE (type))
> +	{
> +	  tree first = first_field (type);
> +	  if (first == NULL_TREE
> +	      || TYPE_MODE (type) != DECL_MODE (first))
> +	    goto ignored;
> +	}

I'd think it would be nicer to emit the same diagnostic from this
spot as we do if handle_transparent_union_attribute is called on
an incomplete type and finish_struct is called later.
Unfortunately that varies between C and C++ FEs, we'd need a langhook
for that, which is perhaps an overkill for such rarely used attribute.

So, I'm fine with this patch for the trunk too, though perhaps you
want also
	|| first == error_mark_node
	|| DECL_ARTIFICIAL (first)
for the condition to catch what finish_struct_1 does for C++?

	Jakub