Re: [boost] [smart_ptr/static_pointer_cast] Trying to make ADL useful in my boost libraries

From: Peter Dimov (pdimov_at_[hidden])
Date: 2006-06-13 12:24:41

Ion Gaztañaga wrote:

[...]

> template<class Pointer>
> class container
> {
>    Pointer ptr;
>    public:
>    void func()
>    {  static_pointer_cast<int>(Pointer());   }
> };
>
> }  //namespace dummy {
> }  //namespace boost {

[...]

> This code compiles in VC 7.1 and VC 8 but fails with gcc, complaining
> the "there is no declaration for static_pointer_cast". If we change
> the header order to:
>
> #include "smart_ptr.h"
> #include "container.h"
> #include "container_factory.h"
>
> This compiles fine because the static_pointer_cast definition is found
> before the call. My first question is:
>
> * Is this gcc error correct? Revising a bit ADL + template issues (a
> fairly complicated logic that should be removed from C++, in my
> opinion) I think that gcc is right.

Questionable. The problem is that for static_pointer_cast<int> to be valid, 
static_pointer_cast must be known to be a template. It is understandable why 
g++ rejects it without a declaration, but I'm not sure whether this is 
conforming now, much less whether it'd be conforming tomorrow if there are 
open core issues for this case.

Even if static_pointer_cast<int> is parsed correctly, I recall a core issue 
about whether the call should be subject to ADL.

> * If gcc is correct, what is the implementor supposed to do? Every
> templatized class header shouldn't include smart_ptr.h because
> smart_ptr is unknown and it can even be a user-defined class.

To make the above work for raw pointers (these don't have an associated 
namespace), the implementor would need to #include boost/pointer_cast.hpp 
and add an using declaration for boost::static_pointer_cast. This may or may 
not make g++ happy.

> I must have missed something, because otherwise, I would find ADL
> really useless as a robust customization point.

The typical ADL expression doesn't have an explicit <int> template 
parameter.