When trying to use an auto_ptr with a type that was declared with forward-declaration, like this:
class A;
...
std::auto_ptr<A> a;
the destructor of A is not called (apparently, because auto_ptr internally deletes the underlying pointer and the destructor for an incomplete type cannot be called).
However, the same code works fine and the destructor is called when using std::shared_ptr instead of std::auto_ptr.
How can that be explained?
A shared_ptr can be declared with an incomplete type, yes. The type does not need to be complete until you initialize or reset it.
When you initialize or reset a shared_ptr to point to a new object, it creates a "deleter" that can be used to destroy the object. For example, consider the following:
// somewhere where A is incomplete:
std::shared_ptr<class A> p;
// define A
class A { /* ... */ };
p.reset(new A());
When you call reset, A is complete because you're creating an instance of it using new. The reset function creates and stores internally a deleter that will be used to destroy the object using delete. Because A is complete here, that delete will do the right thing.
By doing this, shared_ptr does not require that A be complete when the shared_ptr<A> is declared; it only requires that A is complete when the shared_ptr constructor that takes a raw pointer is invoked or when you call reset with a raw pointer.
Note that if A is not complete when you do one of those two things, shared_ptr won't do the right thing and the behavior is undefined (this is explained in the documentation for boost::shared_ptr, which is probably the best resource for learning how to use shared_ptr correctly, regardless which version of shared_ptr you are using (Boost, TR1, C++0x, etc.)).
However, as long as you always follow the best usage practices for shared_ptr--notably, if you always initialize and reset a shared_ptr directly with a pointer resulting from a call to new--you won't have to worry about violating this rule.
This functionality isn't free: shared_ptr has to create and store a pointer to the deleter functor; typically this is done by storing the deleter as part of the block that stores the strong and weak reference counts or by having a pointer as part of that block that points to the deleter (since you can provide your own deleter).
auto_ptr (and unique_ptr too) is designed to be free of any overhead: operations on it are supposed to be just as efficient as using a dumb pointer. Thus, auto_ptr doesn't have this functionality.
