on Tue Dec 11 2007, Thorsten Ottosen <thorsten.ottosen-AT-dezide.com> wrote:
David Abrahams skrev:
on Fri Dec 07 2007, Ion Gaztañaga <igaztanaga-AT-gmail.com> wrote:
Thorsten Ottosen wrote:
One thing puzzled me though: (see http://igaztanaga.drivehq.com/unordered/unordered/comparison.html)
"The containers hash or predicate function can throw exceptions from erase"
Certainly, Daniel has tried to achieve strong exception guarantees and the implementation becomes quite complicated if comparison/hash throws. Double buffering and other tricks are needed. I think this is a very good question both for boosters and people from the LWG.
When specifying requirements in the standard, we (the LWG) don't like to constrain implementations or users if possible. I don't see any reason to forbid a throwing comparison or hash.
Right, but this is often a double-edged sword. Think of problems with allowing std::list<T>::size() to be O(n).
Or allowing ptr_container<T> to not be assignable and copy-constructible?
What is flexibility for the implementer is often a synonym for non-portability for the user. As a library implementer I have to be conservative w.r.t. the exceotion-safety guarantees given which means the additional guarantees given by some implementations cannot be taken advantage of. Of course, I might take advantage of them in non-generic code, but then at the expense of being non-portable.
I'm aware of all the tradeoffs. It's a balancing act, as ever. -- Dave Abrahams Boost Consulting http://www.boost-consulting.com