on Mon Dec 01 2008, Arno Schödl <aschoedl-AT-think-cell.com> wrote:
Specifically for my application, you (as maintainer of boost.Iterator?) would have to change boost::transform_iterator/filter_iterator and Joaquin multi_index_container. If we can agree on that, I will change the code to CRTP.
I am willing to consider such changes.
CRTP it is.
How is that clever? It follows the same pattern as the one for a custom swap:
using std::swap; CustomType tA, tB; swap( tA, tB );
What do you propose instead?
I think if you want to call lower_bound in the most efficient way, it should be possible to do so with qualification. If the algorithm in std:: isn't sufficiently "hookable," one should create a different one that is. The universe of names to be found by ADL should be limited to fundamental operations like swap, since ADL reserves names "globally."
Hmmm. I see it this way: std::lower_bound has well-defined semantics, and we don't change them. We merely provide an implementation for a particular iterator.
But we have no way to provide an implementation of std::lower_bound. We can only provide an implementation of lower_bound (unqualified).
If that would be the end of the story, you would probably agree that ADL is the way to go and
using std::lower_bound; lowerbound(customIt, customIt, ...); // unqualified
is the right way to call the custom lower_bound.
Like I said, whether or not that is the right way to go IMO depends largely on whether lower_bound should be considered a fundamental operation akin to swap.
Now it happens that for some class of iterators, lower_bound/... can be implemented on top of partition_point.
Sorry, what is partition_point? Oh, it appears to be essentially std::lower_bound.
I consider this an implementation detail, which you specify by deriving this iterator from a particular base class. I think the situation is really the same as in Boost.Operators that implements operator+ via operator+=.
Unless I'm missing something, I think, in that case, that the fundamental operation is partition_point and not lower_bound or upper_bound. Can't the latter two be implemented in terms of partition_point? -- Dave Abrahams BoostPro Computing http://www.boostpro.com