How to extract the highest-indexed specialization

问题:

I'm trying to do some template metaprogramming and I'm finding the need to "extract" the highest index of a specialization of some structure in some type.

For example, if I have some types:

struct A
{
    template<unsigned int> struct D;
    template<> struct D<0> { };
};

struct B
{
    template<unsigned int> struct D;
    template<> struct D<0> { };
    template<> struct D<1> { };
};

struct C
{
    template<unsigned int> struct D;
    template<> struct D<0> { };
    template<> struct D<1> { };
    template<> struct D<2> { };
};

How can I then write a metafunction like this:

template<class T>
struct highest_index
{
    typedef ??? type;
    // could also be:   static size_t const index = ???;
};

to give me the highest-indexed D that has been specialized inside an arbitrary struct like the ones above, without requiring the struct to have declared the count explicitly?

回答1:

This is the first version which gets you the maximum index for which specialization is defined. From this, you will get the corresponding type!

Implementation:

template<class T>
struct highest_index
{
  private:
     template<int i>
     struct is_defined {};

     template<int i>
     static char f(is_defined<sizeof(typename T::template D<i>)> *);

     template<int i>
     static int f(...);

     template<int i>
     struct get_index;

     template<bool b, int j>
     struct next
     {
        static const int value = get_index<j>::value;
     };
     template<int j>
     struct next<false, j>
     {
        static const int value = j-2;
     };
     template<int i>
     struct get_index
     {
        static const bool exists = sizeof(f<i>(0)) == sizeof(char);
        static const int value = next<exists, i+1>::value;
     };

    public:
     static const int index = get_index<0>::value; 
};

Test code:

#include <iostream>

struct A
{
    template<unsigned int> struct D;
};
template<> struct A::D<0> { };
template<> struct A::D<1> { };

struct B
{
    template<unsigned int> struct D;
};
template<> struct B::D<0> { };
template<> struct B::D<1> { };
template<> struct B::D<2> { };


int main()
{
    std::cout << highest_index<A>::index << std::endl;
    std::cout << highest_index<B>::index << std::endl;
}

Output:

1
2

Live demo. :-)

回答2:

Figured it out with the help from the comments under the question!

struct A { template<unsigned int> struct D; };
template<> struct A::D<0> { };

struct B { template<unsigned int> struct D; };
template<> struct B::D<0> { };
template<> struct B::D<1> { };

struct C { template<unsigned int> struct D; };
template<> struct C::D<0> { };
template<> struct C::D<1> { };
template<> struct C::D<2> { };
template<> struct C::D<3> { };

template<unsigned int>
static unsigned char test(...);

template<unsigned int N, class T>
static typename enable_if<
    sizeof(typename T::template D<N>),
    unsigned char (&)[1 + sizeof(test<N + 1>(T()))]
>::type test(T, typename T::template D<N> = typename T::template D<N>());

int main()
{
    return sizeof(test<0>(C())) - 1;  // Evaluates to number of specializations
}

回答3:

Here is my little contribution.

We start off with the existence methods:

template <unsigned>
static unsigned char exists_impl(...);

template <unsigned N, typename T>
static auto exists_impl(T const&&) ->
    typename std::enable_if<sizeof(typename T::template D<N>),
                            unsigned char (&)[2]>::type;

template <typename T, unsigned N>
static constexpr bool exists() {
    return sizeof(exists_impl<N>(std::declval<T>())) != 1;
}

I believe here that constexpr and function usage do bring a lot to the table in terms of readability, so I don't use the typical types.

Then, we use a typical binary search (2nd attempt, see first attempt at bottom), at a loss of readability, but to benefit from lazy instantiation, we use partial template specialization and std::conditional:

template <typename T, unsigned low, unsigned high, typename = void>
struct highest_index_in;

template <typename T, unsigned low>
struct highest_index_in<T, low, low>: std::integral_constant<unsigned, low> {};

template <typename T, unsigned low, unsigned high>
struct highest_index_in<T, low, high, typename std::enable_if<(high == low + 1)>::type>:
  std::integral_constant<unsigned, low + exists<T, low+1>()> {};

template <typename T, unsigned low, unsigned high>
struct highest_index_in<T, low, high, typename std::enable_if<(high > low + 1)>::type>:
  std::conditional< exists<T, (low+high)/2>(),
                    highest_index_in<T, (low+high)/2, high>,
                    highest_index_in<T, low, (low+high)/2> >::type
{};

template <typename T>
static constexpr unsigned highest_index() {
   return highest_index_in<T, 0, ~(0u)>::value;
} // highest_index

Demo at liveworkspace, computing highest_index<C>() is near instantaneous.

---

First attempt at binary search, unfortunately the compiler need instantiate the function bodies recursively (to prove they can be instantiated) and thus the work it has to do is tremendous:

template <typename T, unsigned low, unsigned high>
static constexpr auto highest_index_in() ->
   typename std::enable_if<high >= low, unsigned>::type
{
   return low == high                 ? low :
          high == low + 1             ? (exists<T, high>() ? high : low) :
          exists<T, (high + low)/2>() ? highest_index_in<T, (high+low)/2, high>() :
                                        highest_index_in<T, low, (high+low)/2>();
} // highest_index_in

template <typename T>
static constexpr unsigned highest_index() {
   return highest_index_in<T, 0, ~(0u)>();
} // highest_index

So, unfortunately, highest_index is not usable and the clang is dang slow (not that gcc appears to be doing better).