Re: [boost] Re: Re: Re: Traits-based design and ODR

From: Beman Dawes (bdawes_at_[hidden])
Date: 2002-05-01 15:22:04

At 10:51 PM 4/30/2002, Gennadiy Rozental wrote:
 >"Peter Dimov" <pdimov_at_[hidden]> wrote in message
 >news:00f801c1f068$0c7fc650$1d00a8c0_at_pdimov2...
 >
 >>
 >> X is defined in X.hpp, along with the associated traits 
specializations.
 >The
 >> problem, with a smart_ptr that uses the type of its template parameter 
to
 >> somehow alter its type (via traits, sizeof(expression-involving-X), or
 >some
 >> other mechanism), is that a.cpp sees a different smart_ptr<X>. So 
there's
 >an
 >> ODR violation.
 >>
 >> Note that shared_ptr doesn't have this problem.
 >>
 >
 >And how shared_ptr will automatically switch between 
intrusive/nonintrusive
 >counting?

Peter has figured out something really clever.  Here was his description:

 > What I have in mind is:
 >
 > * define template<class T> class intrusive_ptr that depends on
 > T::add_ref
 > and T::release (or, better yet, as Dave Harris pointed out, on
 > add_ref(T *)
 > and release(T *).) It doesn't require a common base class (per itself.)
 >
 > * add a shared_ptr<T>(intrusive_ptr<U> const & p) constructor that
 > initializes px to p.get() and pn to p.get() implicitly converted to
 > counted_base*.
 >
 > Voila, instant interoperability when U is derived from counted_base.

So at virtually no cost (the only change was the addition of a single 
constructor, which isn't used if not needed) he has allowed shared_ptr's to 
be constructed from invasive_ptr's.  A marked improvement in 
interoperability, particularly as needed in library interfaces.

If that isn't cool, I don't know what is.

Hopefully a policy-based smart pointers could also deliver an
invasive->shared constructor.

--Beman