Re: [boost] [shared_ptr] dangerous implementation of moveconstructor

From: Joe Gottman (jgottman_at_[hidden])
Date: 2007-04-16 21:19:58

Peter Dimov wrote:

> 
> To get back to shared_ptr, given p and q of type shared_ptr, would you be 
> extremely disappointed if after:
> 
>     p = move( q );
> 
> q holds the old value of p instead of being empty?
> 
> (Note that p may have had a value that you could consider "unsafe" in the 
> context of the valid values of q.) 
> 

    Not extremely disappointed.  I do think that move semantics would be 
very useful for shared_ptr's, and I know that there are classes where it 
would be prohibitively expensive to clear the source after a move 
assignment. But note that it is possible to efficiently make q hold a 
null pointer after a move assignment:

  shared_ptr & operator=( shared_ptr && r ) // never throws
{
     pn.swap(r.pn);
     px = r.px;
     r.px = 0;
     return *this;
}

template <class Y>
shared_ptr &operator=(shared_ptr<Y> &&r) //never throws
{
    pn.swap(r.pn);
    px = r.px;
    r.px = 0;
    return *this;
}

This is actually fractionally more efficient than calling
     swap(pn, r.pn); swap(px, r.px);

since swap(px, r.px) involves one copy constructor and two assignments 
and the implementation I gave just involves two assignments.  Also, in 
the case of the template assignment operator it is not required that px 
be convertible to Y*.

Joe Gottman