Expand two parameter packs

Question

Consider following piece of code:

static constexpr size_t Num {2};
struct S {
    std::array<size_t, Num> get () { return {1, 2}; }
};

struct S1 : S {};
struct S2 : S {};

struct M {
    template <typename T>
    typename std::enable_if<std::is_same<T, S1>::value, S1>::type get () const { 
        return S1 {}; 
    }

    template <typename T>
    typename std::enable_if<std::is_same<T, S2>::value, S2>::type get () const { 
        return S2 {}; 
    }
};

I want to have a function which merges two or more std::arrays making one std::array.

So far I ended with something like this:

template <typename Mode, typename... Rs, size_t... Ns>
std::array<size_t, sizeof... (Rs)*Num> get_array (const Mode& mode, Sequence::Sequence<Ns...>) {
    return {std::get<Ns> (mode.template get<Rs...> ().get ())...};
}

I want to have that the following code

M m;
auto x = get_array<M, S1, S2> (m, Sequence::Make<2> {});

produces std::array<size_t, 4> filled with {1, 2, 1, 2}.

Where Sequence::Sequence and Sequence::Make are described here.

I know that placing ... of Rs is incorrect in this context (If sizeof... (Rs) is 1 then it is fine, std::array<size_t, 2> with {1, 2} is returned) but I have no idea where to put it to make expansion which looks like this:

std::get<0> (mode.template get<Rs[0]> ().get ()), 
std::get<1> (mode.template get<Rs[0]> ().get ()),
std::get<0> (mode.template get<Rs[1]> ().get ()), 
std::get<1> (mode.template get<Rs[1]> ().get ());

Of course Rs[0] I mean first type from parameter pack.

Is it even possible?

Answer 1

So here's for an arbitrary number of same-type arrays. We are basically implementing a highly restrictive version of tuple_cat, made substantially easier because the number of elements in the arrays is the same. I make use of a couple C++14 and 17 library features that are all readily implementable in C++11.

template<class, size_t> struct div_sequence;
template<size_t...Is, size_t Divisor>
struct div_sequence<std::index_sequence<Is...>, Divisor>
{
   using quot = std::index_sequence<Is / Divisor...>;
   using rem = std::index_sequence<Is % Divisor...>;
};


template<class T, size_t...Ns, size_t...Is, class ToA>
std::array<T, sizeof...(Ns)> array_cat_impl(std::index_sequence<Ns...>,
                                            std::index_sequence<Is...>,
                                            ToA&& t) 
{
    // NB: get gives you perfect forwarding; [] doesn't.
    return {std::get<Is>(std::get<Ns>(std::forward<ToA>(t)))... }; 
}

template<class Array, class... Arrays,
         class VT = typename std::decay_t<Array>::value_type,
         size_t S = std::tuple_size<std::decay_t<Array>>::value,
         size_t N = S * (1 + sizeof...(Arrays))>
std::array<VT, N> array_cat(Array&& a1, Arrays&&... as) 
{
     static_assert(std::conjunction_v<std::is_same<std::decay_t<Array>,
                                                   std::decay_t<Arrays>>...
                                      >, "Array type mismatch");

     using ind_seq = typename div_sequence<std::make_index_sequence<N>, S>::rem;
     using arr_seq = typename div_sequence<std::make_index_sequence<N>, S>::quot;
     return array_cat_impl<VT>(arr_seq(), ind_seq(), 
                               std::forward_as_tuple(std::forward<Array>(a1),
                                                     std::forward<Arrays>(as)...)
                               );
}

---

We can also reuse the tuple_cat machinery, as in @Barry's answer. To sidestep potential QoI issues, avoid depending on extensions and also extra moves, we don't want to tuple_cat std::arrays directly. Instead, we transform the array into a tuple of references first.

template<class TupleLike, size_t... Is>
auto as_tuple_ref(TupleLike&& t, std::index_sequence<Is...>)
    -> decltype(std::forward_as_tuple(std::get<Is>(std::forward<TupleLike>(t))...))
{
    return std::forward_as_tuple(std::get<Is>(std::forward<TupleLike>(t))...);
}

template<class TupleLike,
         size_t S = std::tuple_size<std::decay_t<TupleLike>>::value >
auto as_tuple_ref(TupleLike&& t)
    -> decltype(as_tuple_ref(std::forward<TupleLike>(t), std::make_index_sequence<S>()))
{
    return as_tuple_ref(std::forward<TupleLike>(t), std::make_index_sequence<S>());
}

We can then transform the tuple_cat'd references back into an array:

template <class R1, class...Rs, size_t... Is>
std::array<std::decay_t<R1>, sizeof...(Is)> 
   to_array(std::tuple<R1, Rs...> t, std::index_sequence<Is...>) 
{
   return { std::get<Is>(std::move(t))... };
}

template <class R1, class...Rs>
std::array<std::decay_t<R1>, sizeof...(Rs) + 1> to_array(std::tuple<R1, Rs...> t) 
{
   static_assert(std::conjunction_v<std::is_same<std::decay_t<R1>, std::decay_t<Rs>>...>,
                 "Array element type mismatch");
   return to_array(t, std::make_index_sequence<sizeof...(Rs) + 1>());
}

Finally, array_cat itself is just

template <class... Arrays>
auto array_cat(Arrays&&... arrays) 
    -> decltype(to_array(std::tuple_cat(as_tuple_ref(std::forward<Arrays>(arrays))...))) 
{
    return to_array(std::tuple_cat(as_tuple_ref(std::forward<Arrays>(arrays))...));
}

Any decent optimizer should have little difficulty optimizing the intermediate tuples of references away.

Answer 2

To start with, this is basically asking to concatenate an arbitrary number of arrays. Which is very similar to concatenate an arbitrary number of tuples, for which there is a standard library function, even in C++11: std::tuple_cat(). That gets us almost there:

template <class... Ts, class M>
auto get_array(M m) -> decltype(std::tuple_cat(m.template get<Ts>()...)) {
    return std::tuple_cat(m.template get<Ts>()...);
}

Note that I flipped the template parameters, so this is just get_array<T1, T2>(m) instead of having to write get_array<M, T1, T2>(m).

Now the question is, how do we write array_cat? We'll just use tuple_cat and convert the resulting tuple to an array. Assume an implementation of index_sequence is available (which is something you'll want in your collection anyway):

template <class T, class... Ts, size_t... Is>
std::array<T, sizeof...(Ts)+1> to_array_impl(std::tuple<T, Ts...>&& tup,
                                             std::index_sequence<Is...> ) {
    return {{std::get<Is>(std::move(tup))...}};
}

template <class T, class... Ts>
std::array<T, sizeof...(Ts)+1> to_array(std::tuple<T, Ts...>&& tup) {
    return to_array_impl(std::move(tup), std::index_sequence_for<T, Ts...>());
}

template <class... Tuples>
auto array_cat(Tuples&&... tuples) -> decltype(to_array(std::tuple_cat(std::forward<Tuples>(tuples)...))) {
    return to_array(std::tuple_cat(std::forward<Tuples>(tuples)...));
}

And that gives you:

template <class... Ts, class M>
auto get_array(M m) -> decltype(array_cat(m.template get<Ts>()...)) {
    return array_cat(m.template get<Ts>()...);
}

which handles arbitrarily many types.

Answer 3

Yes, this can be done, with the standard index_sequence tricks:

template <class T, std::size_t N1, std::size_t N2, std::size_t ... Is, std::size_t ... Js>
std::array<T, N1 + N2> merge_impl(const std::array<T, N1>& a1,
                                  const std::array<T, N2>& a2, 
                                  std::index_sequence<Is...>,
                                  std::index_sequence<Js...>) {
    return {a1[Is]..., a2[Js]...};
}

template <class T, std::size_t N1, std::size_t N2>
std::array<T, N1 + N2> merge(const std::array<T, N1>& a1, const std::array<T, N2>& a2) {
    return merge_impl(a1, a2,
                      std::make_index_sequence<N1>{}, 
                      std::make_index_sequence<N2>{});
}

index_sequence is only in the 14 standard, but can be easily implemented in 11; there are many resources (including on SO) that describe how to do so (edit: it's basically equivalent to your Sequence stuff, may as well get used to the standard names for them). Live example: http://coliru.stacked-crooked.com/a/54dce4a695357359.

Answer 4

Assuming that we're using Xeo's index sequence implementation, we can do something like this:

First create a function for concatenating two arrays. It receives the arrays, plus an index sequence for each one (detail::seq is the index_sequence type)

template<class T, size_t N, size_t M, size_t... I, size_t... J>
std::array<T, N + M> concat(const std::array<T, N>& arr1, const std::array<T, M>& arr2, detail::seq<I...>, detail::seq<J...>)
{
     return {arr1[I]..., arr2[J]...};
}

Next, call this function from your get_array function, except we're going to double the seq that we received from the call in main:

template<class MODE, class... T, size_t... I>
auto get_array(MODE m, detail::seq<I...>) ->decltype(concat(m.template get<T>().get()..., detail::seq<I...>{}, detail::seq<I...>{})){
    return concat(m.template get<T>().get()..., detail::seq<I...>{}, detail::seq<I...>{});
}

The call in main looks just like it did in your code:

M m;
auto x = get_array<M, S1, S2>(m, detail::gen_seq<2>{});

Where detail::gen_seq is the implementation of make_index_sequence that Xeo had.

Live Demo

Note that I replaced unsigned with size_t in Xeo's index sequence impl.

In C++14 we don't need to implement seq or gen_seq, and we also wouldn't need a trailing -> decltype() after our function.

In C++17 it would be even easier to generalize our concatenation for an arbitrary number of arrays, using fold expressions.