Re: [boost] Projection Iterator Adaptor

From: Beman Dawes (bdawes_at_[hidden])
Date: 2001-03-18 11:09:38

At 05:22 PM 3/17/2001, Jeremy Siek wrote:

 >On Sat, 17 Mar 2001, John E. Potter wrote:
 >jpotte>
 >jpotte> Sorry for being out of contact, but I only have time to
 >jpotte> delete without reading most of the boost stuff for now.
 >jpotte>
 >jpotte> The action of iter[n] is the same as *(iter + n) and in
 >jpotte> both cases should be a reference or reference to const.
 >jpotte> I know that the standard says convertable to T, but
 >jpotte> forward iterator has it right and there is an issue.
 >
 >There is a subtle problem with using reference return for
 >iterator_adaptor::operator[]. The problem shows up for iterators like
 >counting_iterator that return a reference to a value object that is part
 >of the iterator object itself, and has the same lifetime as the iterator.
 >The expressions (iter + n) creates a temporary iterator object, then
 >*(iter + n) returns a reference to part of this temporary iterator 
object.
 >Upon return from operator[], the temporary iterator goes away and we are
 >left with a dangling reference.
 >
 >One option is to say iterator_adaptor can't be used with this kind of
 >iterator. However, if a user were to go ahead and use such an iterator,
 >they would not find out about the problem until run time! Even so I'm
 >slighly in favor of this option, but Dave is strongly against it.

Note that the C++ committee's Library Working Group has an open issue:

198.  Validity of pointers and references unspecified after iterator 
destruction.

Matt Austern has been working on a new proposed resolution which includes 
my original general fix plus Andy Koenig's very perceptive fix to 
reverse_iterator.  Matt is very concerned about operator[]; you might want 
to take this up with him privately.

--Beman