Match iterable types (arrays and classes with begin()/end())

Question

I wrote type traits like classes that can be used test if a given type is "iterable". This is true for arrays (for T[N], not for T[]) and for classes that have a begin and an end method that return things that look like iterators. I wonder if it can be done more concise/simpler than I did it?

Especially the stuff in the impl namespace look a bit roundabout/hacky. It all looks a bit ugly to me. For an example that uses this and can be compiled with g++ and clang++ see: https://gist.github.com/panzi/869728c9879dcd4fffa8

template<typename T>
struct is_iterator {
private:
    template<typename I> static constexpr auto test(void*)
        -> decltype(
            *std::declval<const I>(),
            std::declval<const I>() == std::declval<const I>(),
            std::declval<const I>() != std::declval<const I>(),
            ++ (*std::declval<I*>()),
            (*std::declval<I*>()) ++,
            std::true_type()) { return std::true_type(); }

    template<typename I> static constexpr std::false_type test(...) { return std::false_type(); }

public:
    static constexpr const bool value = std::is_same<decltype(test<T>(0)), std::true_type>::value;
};

namespace impl {
    // implementation details

    template<typename T>
    struct has_iterable_methods {
    private:

        template<typename C> static constexpr auto test(void*)
            -> decltype(
                std::declval<C>().begin(),
                std::declval<C>().end(),
                std::true_type()) { return std::true_type(); }

        template<typename C> static constexpr std::false_type test(...) { return std::false_type(); }

    public:
        static constexpr const bool value = std::is_same<decltype(test<T>(0)), std::true_type>::value;
    };

    template<typename T, bool HasIterableMethods>
    struct returns_iterators : public std::false_type {};

    template<typename T>
    struct returns_iterators<T, true> {
        typedef decltype(std::declval<T>().begin()) begin_type;
        typedef decltype(std::declval<T>().end())   end_type;

        static constexpr const bool value =
            std::is_same<begin_type, end_type>::value &&
            is_iterator<begin_type>::value;
    };
}

template<typename T>
struct is_iterable : public std::integral_constant<
    bool,
    impl::returns_iterators<
        typename std::remove_const<T>::type,
        impl::has_iterable_methods<typename std::remove_const<T>::type>::value>::value> {};

template<typename T, std::size_t N>
struct is_iterable<T[N]> : public std::true_type {};

template<typename T>
struct is_iterable<T*> : public std::false_type {};

Answer 1

First, some boilerplate to do easy argument dependent lookup of begin in a context where std::begin is visible:

#include <utility>
#include <iterator>
namespace adl_details {
  using std::begin; using std::end;
  template<class R>
  decltype(begin(std::declval<R>())) adl_begin(R&&r){
    return begin(std::forward<R>(r));
  }
  template<class R>
  decltype(end(std::declval<R>())) adl_end(R&&r){
    return end(std::forward<R>(r));
  }
}
using adl_details::adl_begin;
using adl_details::adl_end;

This is required to reasonably emulate how range-based for(:) loops find their begin/end iterators. By packaging it up like this, we reduce boilerplate below.

Next, some C++1y style utility aliases:

template<class>struct sink {using type=void;};
template<class X>using sink_t=typename sink<X>::type;
template<bool b, class T=void>using enable_if_t=typename std::enable_if<b,T>::type;

sink_t takes any type, and throws it away replacing it with void.

enable_if_t removes annoying typename spam below.

In an industrial strength library, we'd put this in details, and have a 1-type-argument version that dispatches to it. But I don't care:

template<class I,class=void> struct is_iterator:std::false_type{};
template<> struct is_iterator<void*,void>:std::false_type{};
template<> struct is_iterator<void const*,void>:std::false_type{};
template<> struct is_iterator<void volatile*,void>:std::false_type{};
template<> struct is_iterator<void const volatile*,void>:std::false_type{};
template<class I>struct is_iterator<I,
  sink_t< typename std::iterator_traits<I>::value_type >
>:std::true_type{};

is_iterator doesn't do heavy auditing of the iterator_traits of I. But it is enough.

template<class R>
using begin_t=decltype(adl_begin(std::declval<R&>()));
template<class R>
using end_t=decltype(adl_end(std::declval<R&>()));

These two type aliases make the stuff below less annoying.

Again, in industrial strength libraries, put 2-arg-with-void into details:

template<class R,class=void> struct has_iterator:std::false_type{};
template<class R>
struct has_iterator<
  R,
  enable_if_t<
    is_iterator<begin_t<R>>::value
    && is_iterator<end_t<R>>::value
    // && std::is_same<begin_t<R>,end_t<R>>::value
  >
>:std::true_type{};

Note the commented out line in the enable_if_t above. I left that out to allow asymmetric iteration to work, where the end is a type that has a different operator== overload. Such is being considered for C++17: it allows really, really efficient algorithms on null-terminated strings (for example).

Finally, the final output:

template<class R>using iterator_t=enable_if_t<has_iterator<R>::type, begin_t<R>>;

which evaluates to the iterator of the iterable range R iff it has one.

There are cases where this won't work, but they are pathological.

live example