------- Comment #12 from jason at redhat dot com 2008-04-04 18:10 -------
Subject: Re: [4.3/4.4 Regression] vector_size attribute lost
in function arguments for templates
jakub at gcc dot gnu dot org wrote:
> Actually, to clarify #c10, attributes on parameter packs just make things
> harder on the compiler side, but even in C++98 the same issue is present:
> #define vector __attribute__((__vector_size__ (16)))
>
> template <typename T> void foo (int x, vector T y) { }
> void bar (vector long a, vector double b)
> {
> foo<long> (5, a);
> foo (5, b);
> }
This functionality seems desirable, but cannot be considered a regression.
> Are there any attributes other than vector_size which affect the decls
> similarly?
I don't think so.
> If not, I'd say that the C++ FE should hardcode some knowledge about this
> attribute, e.g. know that it applies to the type, so if
> processing_template_decl
> move them from DECL_ATTRIBUTES to corresponding type's TYPE_ATTRIBUTES (either
> the parameter type such that it would be in TYPE_ARG_TYPES too, or for
> FUNCTION_TYPE/METHOD_TYPE stick it into return type's TYPE_ATTRIBUTES).
> And in type_unification_real take it into account.
This makes sense to me. Though I think that if we just push the
attribute down to the type that it actually modifies, we don't have to
think about TYPE_ARG_TYPES. That should also avoid the need for
reconstruct_complex_type.
Jason
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35758
