Re: [boost] Borland C++ 5.5.1 problem with the new shared_ptr tests

From: Peter Dimov (pdimov_at_[hidden])
Date: 2002-11-19 13:04:03

From: "Fernando Cacciola" <fernando_cacciola_at_[hidden]>
> > Borland C++ 5.5.1 fails the following test:
> >
> > struct X
> > {
> >     X(): next() {}
> >     boost::shared_ptr<X> next;
> > };
> >
> > void test()
> > {
> >     boost::shared_ptr<X> p(new X);
> >     p->next = boost::shared_ptr<X>(new X);
> >     BOOST_TEST(!p->next->next);
> >     p = p->next;
> >     BOOST_TEST(!p->next); // here
> > }
> >
> The problem is here;
>
>    p = p->next;
>
> This assignment is slicing. It is directly calling
>
>     shared_count & operator= (shared_count const & r) // nothrow
>
> on the lvalue (p). The pointee (p.px)  is destroyed during
"pi_->release();"
> but is not re-assigned.
>
> If the intention was that the assignment use:
>
>     template<typename Y>
>     shared_ptr & operator=(shared_ptr<Y> const & r) // never throws
>
> then Borland doesn't use it.

No. The intent is to use the automatically generated copy assignment
operator. By default, it should do memberwise assignment, and the end result
should be exactly

>     shared_ptr & operator=(shared_ptr const & r) // never throws
>     {
>         px = r.px;
>         pn = r.pn; // shared_count::op= doesn't throw
>         return *this;
>     }

that.