On Wed, Oct 10, 2012 at 1:51 AM, Andrzej Krzemienski <akrzemi1@gmail.com> wrote:
Since you asked what I think,
"pseudo-signatures > usage patterns"
but that's probably not the kind of information you were looking for :-)
Yes, that's more of a critique to N3351 than my design to implement it but it's a fare comment. If N3351 approach is to be abandoned in favor of pseudo-signatures then no point in implementing N3351 but I don't think there's consensus on this either way yet. I personally don't have an opinion yet... (and maybe I should have an opinion before I go off an impl the entire N3351).
Even if you decide to go with pseudo-signatures, one novel thing from N3351 is worth adapting: treating every concept as a predicate, so that the programmer can check himself whether a certain set of types models the concept:
static_assert( ForwardIterator<int*>, "ouch" );
and so that you can combine concept requirements with any boolean meta-function:
template <class T> requires Copyable<T> && sizeof(T) < 8 ...
I agree that concepts should be boolean meta-functions as they are in N3351, they will be in Boost.Contract, and I think they are in Boost.Generic. Also I think that's the only (?) way for concepts to remove the function/class templates from the overload/specialization set instead of causing a hard compiler error. However, there might be some requirements that cannot be checked as a boolean meta-function but that can be checked as to generate an hard compiler error. For example, in C++03 we could not program an has_copy_constructor boolean meta-function but Boost.ConceptCheck can check the CopyConstructible concept to generate a hard compiler error. If there are similar situations in C++11 still, that's a reason to keep Boost.ConceptCheck around in C++11 together with Boost.Contract and Boost.Generic concepts. Thanks. --Lorenzo