On Wed, 23 Nov 2005 08:34:26 -0200, Peter Dimov wrote:

Bruno Martínez wrote:

...

Hi.

I had the same question that is answered there, but the answer didn't fit my problem. I'm using scoped_ptr<node> inside node itself, so node is incomplete at the point of use. That forbids using auto_ptr, IIUC.

You can use a scoped_ptr member with an incomplete type, as long as you include out of line default constructor (if you need one) and destructor.

I know. The problem is that I want incomplete types support *and* release member function. I don't seem to require a destructor in this case, for vc71 or gcc3.4: #include struct node { boost::scoped_ptr<node> next; }; int main() { node n; } Bruno

On Thu, 24 Nov 2005 07:10:41 -0200, Daniel Krügler wrote:

Bruno Martínez wrote:

...

I know. The problem is that I want incomplete types support *and* release member function.

Use std::auto_ptr and ensure that you obey Peter Dimov's recommendations. I know, that the standard says, in 5.3.5

"If the object being deleted has incomplete class type at the point of deletion and the complete class has a non-trivial destructor or a deallocation function, the behavior is undefined."

and in 17.4.3.6

"if an incomplete type (3.9) is used as a template argument when instantiating a template component."

There exists no special restriction on std::auto_ptr for incomplete classes

I don't understand. There is a restriction. It's 17.4.3.6!

and the main difference between std::auto_ptr and boost::scoped_ptr in this point is that the standard says, that "no diagnose is required", while boost will diagnose that due a safety belt in boost::checked_delete.

Yes, scoped_ptr is very well the behaved.

...

I don't seem to require a destructor in this case, for vc71 or gcc3.4:

#include

struct node { boost::scoped_ptr<node> next; };

int main() { node n; }

This snippet does not proof, that no d'tor is needed, because node::next is empty for the complete program run-time. Just assign a node to next, and you will see...

I did, but it still works: #include struct node { boost::scoped_ptr<node> next; node(node* next = 0) : next(next) {} }; int main() { node n; n.next.reset(new node); } Bruno

On Mon, 28 Nov 2005 06:57:07 -0200, Daniel Krügler wrote:

Bruno Martínez wrote:

...

On Thu, 24 Nov 2005 07:10:41 -0200, Daniel Krügler wrote:

...

Bruno Martínez wrote:

...

I know. The problem is that I want incomplete types support *and* release member function.

Use std::auto_ptr and ensure that you obey Peter Dimov's recommendations. I know, that the standard says, in 5.3.5

"If the object being deleted has incomplete class type at the point of deletion and the complete class has a non-trivial destructor or a deallocation function, the behavior is undefined."

and in 17.4.3.6

"if an incomplete type (3.9) is used as a template argument when instantiating a template component."

There exists no special restriction on std::auto_ptr for incomplete classes

I don't understand. There is a restriction. It's 17.4.3.6!

OK, so you have to write your own auto_ptr or move_ptr.

From a legal point of view, you can't use auto_ptr for incomplete classes, from a practical point of view you can do that provided that some rules concerning d'tor definition are obeyed.

Yes, that's it. This conclusion is the one I think should be added to the FAQ.

...

#include

struct node { boost::scoped_ptr<node> next; node(node* next = 0) : next(next) {} };

int main() { node n; n.next.reset(new node); }

I wasn't carefully reading your code, obviously ;-) It works, because at the point of the destructor of next the hosting class node is completely defined. You have a special case here, not the classical pimpl problem:

struct A { struct Impl; boost::scoped_ptr<Impl> pimpl; };

because of your next member referencing the hosting class itself. In the A case pimpl's d'tor does not have the complete definition of Impl available, in contrast to your node class.

Yes, it's not exactly the same. Bruno