“unpacking” a tuple to call a matching function pointer

Question

I'm trying to store in a std::tuple a varying number of values, which will later be used as arguments for a call to a function pointer which matches the stored types.

I've created a simplified example showing the problem I'm struggling to solve:

#include <iostream>
#include <tuple>

void f(int a, double b, void* c) {
  std::cout << a << ":" << b << ":" << c << std::endl;
}

template <typename ...Args>
struct save_it_for_later {
  std::tuple<Args...> params;
  void (*func)(Args...);

  void delayed_dispatch() {
     // How can I "unpack" params to call func?
     func(std::get<0>(params), std::get<1>(params), std::get<2>(params));
     // But I *really* don't want to write 20 versions of dispatch so I'd rather 
     // write something like:
     func(params...); // Not legal
  }
};

int main() {
  int a=666;
  double b = -1.234;
  void *c = NULL;

  save_it_for_later<int,double,void*> saved = {
                                 std::tuple<int,double,void*>(a,b,c), f};
  saved.delayed_dispatch();
}

Normally for problems involving std::tuple or variadic templates I'd write another template like template <typename Head, typename ...Tail> to recursively evaluate all of the types one by one, but I can't see a way of doing that for dispatching a function call.

The real motivation for this is somewhat more complex and it's mostly just a learning exercise anyway. You can assume that I'm handed the tuple by contract from another interface, so can't be changed but that the desire to unpack it into a function call is mine. This rules out using std::bind as a cheap way to sidestep the underlying problem.

What's a clean way of dispatching the call using the std::tuple, or an alternative better way of achieving the same net result of storing/forwarding some values and a function pointer until an arbitrary future point?

Answer 1

The C++17 solution is simply to use std::apply:

auto f = [](int a, double b, std::string c) { std::cout<<a<<" "<<b<<" "<<c<< std::endl; };
auto params = std::make_tuple(1,2.0,"Hello");
std::apply(f, params);

Just felt that should be stated once in an answer in this thread (after it already appeared in one of the comments).

---

The basic C++14 solution is still missing in this thread. EDIT: No, it's actually there in the answer of Walter.

This function is given:

void f(int a, double b, void* c)
{
      std::cout << a << ":" << b << ":" << c << std::endl;
}

Call it with the following snippet:

template<typename Function, typename Tuple, size_t ... I>
auto call(Function f, Tuple t, std::index_sequence<I ...>)
{
     return f(std::get<I>(t) ...);
}

template<typename Function, typename Tuple>
auto call(Function f, Tuple t)
{
    static constexpr auto size = std::tuple_size<Tuple>::value;
    return call(f, t, std::make_index_sequence<size>{});
}

Example:

int main()
{
    std::tuple<int, double, int*> t;
    //or std::array<int, 3> t;
    //or std::pair<int, double> t;
    call(f, t);    
}

DEMO

Answer 2

You need to build a parameter pack of numbers and unpack them

template<int ...>
struct seq { };

template<int N, int ...S>
struct gens : gens<N-1, N-1, S...> { };

template<int ...S>
struct gens<0, S...> {
  typedef seq<S...> type;
};


// ...
  void delayed_dispatch() {
     callFunc(typename gens<sizeof...(Args)>::type());
  }

  template<int ...S>
  void callFunc(seq<S...>) {
     func(std::get<S>(params) ...);
  }
// ...

Answer 3

This is a complete compilable version of Johannes' solution to awoodland's question, in the hope it may be useful to somebody. This was tested with a snapshot of g++ 4.7 on Debian squeeze.

###################
johannes.cc
###################
#include <tuple>
#include <iostream>
using std::cout;
using std::endl;

template<int ...> struct seq {};

template<int N, int ...S> struct gens : gens<N-1, N-1, S...> {};

template<int ...S> struct gens<0, S...>{ typedef seq<S...> type; };

double foo(int x, float y, double z)
{
  return x + y + z;
}

template <typename ...Args>
struct save_it_for_later
{
  std::tuple<Args...> params;
  double (*func)(Args...);

  double delayed_dispatch()
  {
    return callFunc(typename gens<sizeof...(Args)>::type());
  }

  template<int ...S>
  double callFunc(seq<S...>)
  {
    return func(std::get<S>(params) ...);
  }
};

#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-parameter"
#pragma GCC diagnostic ignored "-Wunused-variable"
#pragma GCC diagnostic ignored "-Wunused-but-set-variable"
int main(void)
{
  gens<10> g;
  gens<10>::type s;
  std::tuple<int, float, double> t = std::make_tuple(1, 1.2, 5);
  save_it_for_later<int,float, double> saved = {t, foo};
  cout << saved.delayed_dispatch() << endl;
}
#pragma GCC diagnostic pop

One can use the following SConstruct file

#####################
SConstruct
#####################
#!/usr/bin/python

env = Environment(CXX="g++-4.7", CXXFLAGS="-Wall -Werror -g -O3 -std=c++11")
env.Program(target="johannes", source=["johannes.cc"])

On my machine, this gives

g++-4.7 -o johannes.o -c -Wall -Werror -g -O3 -std=c++11 johannes.cc
g++-4.7 -o johannes johannes.o

Answer 4

Here is a C++14 solution.

template <typename ...Args>
struct save_it_for_later
{
  std::tuple<Args...> params;
  void (*func)(Args...);

  template<std::size_t ...I>
  void call_func(std::index_sequence<I...>)
  { func(std::get<I>(params)...); }
  void delayed_dispatch()
  { call_func(std::index_sequence_for<Args...>{}); }
};

This still needs one helper function (call_func). Since this is a common idiom, perhaps the standard should support it directly as std::call with possible implementation

// helper class
template<typename R, template<typename...> class Params, typename... Args, std::size_t... I>
R call_helper(std::function<R(Args...)> const&func, Params<Args...> const&params, std::index_sequence<I...>)
{ return func(std::get<I>(params)...); }

// "return func(params...)"
template<typename R, template<typename...> class Params, typename... Args>
R call(std::function<R(Args...)> const&func, Params<Args...> const&params)
{ return call_helper(func,params,std::index_sequence_for<Args...>{}); }

Then our delayed dispatch becomes

template <typename ...Args>
struct save_it_for_later
{
  std::tuple<Args...> params;
  std::function<void(Args...)> func;
  void delayed_dispatch()
  { std::call(func,params); }
};

Answer 5

This is a bit complicated to achieve (even though it is possible). I advise you to use a library where this is already implemented, namely Boost.Fusion (the invoke function). As a bonus, Boost Fusion works with C++03 compilers as well.

Answer 6

c++14 solution. First, some utility boilerplate:

template<std::size_t...Is>
auto index_over(std::index_sequence<Is...>){
  return [](auto&&f)->decltype(auto){
    return decltype(f)(f)( std::integral_constant<std::size_t, Is>{}... );
  };
}
template<std::size_t N>
auto index_upto(std::integral_constant<std::size_t, N> ={}){
  return index_over( std::make_index_sequence<N>{} );
}

These let you call a lambda with a series of compile-time integers.

void delayed_dispatch() {
  auto indexer = index_upto<sizeof...(Args)>();
  indexer([&](auto...Is){
    func(std::get<Is>(params)...);
  });
}

and we are done.

index_upto and index_over let you work with parameter packs without having to generate a new external overloads.

Of course, in c++17 you just

void delayed_dispatch() {
  std::apply( func, params );
}

Now, if we like that, in c++14 we can write:

namespace notstd {
  template<class T>
  constexpr auto tuple_size_v = std::tuple_size<T>::value;
  template<class F, class Tuple>
  decltype(auto) apply( F&& f, Tuple&& tup ) {
    auto indexer = index_upto<
      tuple_size_v<std::remove_reference_t<Tuple>>
    >();
    return indexer(
      [&](auto...Is)->decltype(auto) {
        return std::forward<F>(f)(
          std::get<Is>(std::forward<Tuple>(tup))...
        );
      }
    );
  }
}

relatively easily and get the cleaner c++17 syntax ready to ship.

void delayed_dispatch() {
  notstd::apply( func, params );
}

just replace notstd with std when your compiler upgrades and bob is your uncle.

Answer 7

Thinking about the problem some more based on the answer given I've found another way of solving the same problem:

template <int N, int M, typename D>
struct call_or_recurse;

template <typename ...Types>
struct dispatcher {
  template <typename F, typename ...Args>
  static void impl(F f, const std::tuple<Types...>& params, Args... args) {
     call_or_recurse<sizeof...(Args), sizeof...(Types), dispatcher<Types...> >::call(f, params, args...);
  }
};

template <int N, int M, typename D>
struct call_or_recurse {
  // recurse again
  template <typename F, typename T, typename ...Args>
  static void call(F f, const T& t, Args... args) {
     D::template impl(f, t, std::get<M-(N+1)>(t), args...);
  }
};

template <int N, typename D>
struct call_or_recurse<N,N,D> {
  // do the call
  template <typename F, typename T, typename ...Args>
  static void call(F f, const T&, Args... args) {
     f(args...);
  }
};

Which requires changing the implementation of delayed_dispatch() to:

  void delayed_dispatch() {
     dispatcher<Args...>::impl(func, params);
  }

This works by recursively converting the std::tuple into a parameter pack in its own right. call_or_recurse is needed as a specialization to terminate the recursion with the real call, which just unpacks the completed parameter pack.

I'm not sure this is in anyway a "better" solution, but it's another way of thinking about and solving it.

---

As another alternative solution you can use enable_if, to form something arguably simpler than my previous solution:

#include <iostream>
#include <functional>
#include <tuple>

void f(int a, double b, void* c) {
  std::cout << a << ":" << b << ":" << c << std::endl;
}

template <typename ...Args>
struct save_it_for_later {
  std::tuple<Args...> params;
  void (*func)(Args...);

  template <typename ...Actual>
  typename std::enable_if<sizeof...(Actual) != sizeof...(Args)>::type
  delayed_dispatch(Actual&& ...a) {
    delayed_dispatch(std::forward<Actual>(a)..., std::get<sizeof...(Actual)>(params));
  }

  void delayed_dispatch(Args ...args) {
    func(args...);
  }
};

int main() {
  int a=666;
  double b = -1.234;
  void *c = NULL;

  save_it_for_later<int,double,void*> saved = {
                                 std::tuple<int,double,void*>(a,b,c), f};
  saved.delayed_dispatch();
}

The first overload just takes one more argument from the tuple and puts it into a parameter pack. The second overload takes a matching parameter pack and then makes the real call, with the first overload being disabled in the one and only case where the second would be viable.

Answer 8

My variation of the solution from Johannes using the C++14 std::index_sequence (and function return type as template parameter RetT):

template <typename RetT, typename ...Args>
struct save_it_for_later
{
    RetT (*func)(Args...);
    std::tuple<Args...> params;

    save_it_for_later(RetT (*f)(Args...), std::tuple<Args...> par) : func { f }, params { par } {}

    RetT delayed_dispatch()
    {
        return callFunc(std::index_sequence_for<Args...>{});
    }

    template<std::size_t... Is>
    RetT callFunc(std::index_sequence<Is...>)
    {
        return func(std::get<Is>(params) ...);
    }
};

double foo(int x, float y, double z)
{
  return x + y + z;
}

int testTuple(void)
{
  std::tuple<int, float, double> t = std::make_tuple(1, 1.2, 5);
  save_it_for_later<double, int, float, double> saved (&foo, t);
  cout << saved.delayed_dispatch() << endl;
  return 0;
}