Dynamically call lambda based on stream input

Question

Some context: I have code that looks like this (minor issue noted here):

Statement  select("SELECT * FROM People WHERE ID > ? AND ID < ?");
select.execute(1462, 1477, [](int ID, std::string const& person, double item1, float item2){
     std::cout << "Got Row:" 
               << ID     << ", " 
               << person << ", " 
               << item1  << ", " 
               << item2  << "\n";
 });

Anyway this connects to the MySQL DB and starts pulling data from the server. So inside execute I loop over the results and call the lambda for each row:

    template<typename Action, typename ...Args>
    void execute(Args... param, Action action)
    {
        // STUFF TO SET up connection.
        // Start retrieving rows.


        while(row = results->getNextRow())
        {
            call(action, row);
        }
    }

So here row gets a single row from the socket connection with mysql (so it calls the lambda as it receives each row (no pulling the rows into memory first)). So the code I want to review is pulling the data and calling the lambda.

 // Statement::call

    template<typename Action>
    void call(Action action, std::unique_ptr<ResultSetRow>& row)
    {
        typedef CallerTraits<decltype(action)>   trait;
        typedef typename trait::AllArgs         AllArgs;
        Caller<trait::size, 0, AllArgs, Action>::call(action, row);
    }

This utilizes the helper class CallerTraits and Caller to pull the required rows from the stream and then call the lambda:

// CallerTraits
// Get information about the arguments in the lambda

template <typename T>
struct CallerTraits
    : public CallerTraits<decltype(&T::operator())>
{};

template<typename C, typename ...Args>
struct CallerTraits<void (C::*)(Args...) const>
{
    static const int                        size = sizeof...(Args);
    typedef std::tuple<Args...>             AllArgs;
};

Then the Caller:

// Caller::call()
//    Reads the next argument required by the lambda from the stream.
//    An exception will be generated if the next argument on the stream
//    does not match the type expected by the lambda.
template<int size, int index, typename ArgumentTupple, typename Action, typename ...Args>
struct Caller
{
    static void call(Action action, std::unique_ptr<ResultSetRow>& row, Args... args)
    {
        // Get the next argument type required by the lambda.
        // As defined by index. Then remove all ref and const
        // bindings.
        typedef typename std::tuple_element<index, ArgumentTupple>::type    NextArgBase;
        typedef typename std::remove_reference<NextArgBase>::type           NextArgCont;
        typedef typename std::remove_const<NextArgCont>::type               NextArg;

        // Read the next value from the stream.
        NextArg val;
        row->getValue(val);

        // Recursively call Caller::call() (via doCall())
        // To get the next argument we need. All the arguments
        // are accumulated in the var args parameter `args`
        doCall<size-1, index+1, ArgumentTupple>(action, row, args..., val);
    }
};

Specialization when no more args need to be retrieved:

// Specialization of Caller::call() when we have got all the arguments.
// This simply calls the lambda with the arguments we have accumulated.
template<int index, typename ArgumentTupple, typename Action, typename ...Args>
struct Caller<0, index, ArgumentTupple, Action, Args...>
{
    static void call(Action action, std::unique_ptr<ResultSetRow>&, Args... args)
    {
        action(args...);
    }
};

Function to deduce parameter types:

// Function template needed because we
// can not deduce the Args... parameter manually in the call.
// so we let the compiler deduce it for us.
template<int size, int index, typename ArgumentTupple, typename Action, typename ...Args>
void doCall(Action action, std::unique_ptr<ResultSetRow>& row, Args... args)
{
    Caller<size, index, ArgumentTupple, Action, Args...>::call(action, row, args...);
}

Answer 1

I find your implementation a bit more complex than necessary. What you want to do is

1. fetch arguments from your "result set" row by calling its getValue() in a particular order;

2. use them (as arguments) to call operator() on function object action.

This can be done without recursion in two lines:

Do{row->getValue(std::get<N>(args))...};
action(std::get<N>(args)...);

where args is a tuple.

Range

---

Ok, now let's step back to see how this is possible. First, we learn how to count from 0 to a given number L, in order to construct range 0, ..., L-1:

// holds any number of size_t parameters
template <size_t... N>
struct sizes { using type = sizes <N...>; };

// given L>=0, generate sequence <0, ..., L-1>
template <size_t L, size_t I = 0, typename S = sizes <> >
struct Range;

template <size_t L, size_t I, size_t... N>
struct Range <L, I, sizes <N...> > : Range <L, I+1, sizes <N..., I> > { };

template <size_t L, size_t... N>
struct Range <L, L, sizes <N...> > : sizes <N...> { };

This is a very common task, actually borrowed from here. There's a better implementation with logarithmic (rather than linear) template depth, but I want to keep it simple here.

"Do"?

---

Next, an extremely helpful struct lets us evaluate expressions in a given order:

// using a list-initializer constructor, evaluate arguments in order of appearance
struct Do { template <typename... T> Do(T&&...) { } };

But beware, due to a bug since at least version 4.7.0, GCC evaluates in the opposite order, right-to-left. A workaround is to provide a range in the opposite order, L-1, ..., 0, but I'm not doing this here.

Caller

---

Now, Caller has a generic definition with only two actual parameters, ArgumentTuple and Action. It also reads that tuple's size, say L, and constructs range 0, ..., L-1 in a third parameter:

// generic Caller
template<
    typename ArgumentTuple, typename Action,
    typename Indices = typename Range<std::tuple_size<ArgumentTuple>{}>::type
>
struct Caller;

Finally, a specialization deduces the generated range as variadic size_t parameters N.... A local tuple of type ArgumentTuple is used to store the arguments, and std::get<N> accesses its N-th element. That's it:

// Caller specialization, where indices N... have been deduced
template<typename ArgumentTuple, typename Action, size_t... N>
struct Caller<ArgumentTuple, Action, sizes<N...> >
{
    static void call(Action action, std::unique_ptr<ResultSetRow>& row)
    {
        ArgumentTuple args;
        Do{row->getValue(std::get<N>(args))...};
        action(std::get<N>(args)...);
    }
};

Please note that all the above code compiles but I have not seen it in action since I don't have the database infrastructure. I have just made a minimal definition

struct ResultSetRow { template<typename T> void getValue(T) { } };

So I can only hope it works for you.

I am sorry if this looks like a complete rewrite rather than a review, but I couldn't help it :-) At least I've kept the part of your code where you deduce ArgumentTuple from the lambda.

PS-1 If your ResultSetRow::getValue() is void, then you need to adjust its variadic call to

Do{(row->getValue(std::get<N>(args)), 0)...};

so that each sub-expression evaluates to int rather than void (you cannot have a list-initializer made of void arguments).

PS-2 I suspect you're not really managing resources here, so you don't need std::unique_ptr; a plain ResultSetRow& would suffice.

Answer 2

I would have probably applied the following changes:

• Make size a static constexpr variable in CallerTraits instead of simply static const.

• Wherever a function simply passes variadic arguments whose types have been deduced, I would have passed args by universal reference (now officially called forwarding reference) and used std::forward to forward the results to the following functions:

  template<int size, int index, typename ArgumentTupple, typename Action, typename ...Args>
  void doCall(Action action, std::unique_ptr<ResultSetRow>& row, Args&&... args)
  {
      Caller<size, index, ArgumentTupple, Action, Args...>::call(action, row, std::forward<Args>(args)...);
  }
  

  It's a bit hard and quite long to explain how it works exactly - you can find a great explanation in the answer linked above -, but the main point is that using this particular recipe implements perfect forwarding:

  template<typename X>
  void foo(X&& arg)
  {
      bar(std::forward<X>(arg));
  }
  

  The type of the parameters of X&&... in foo will have the same const and reference qualifications than the type of the corresponding parameters in bar. Anyway, the link is by far clearer than I am. Simply remember the recipe and that for this recipe to work, the type X has to be deduced by the function; it may not work if X is known from somewhere else.

• Instead of creating functions that take a std::unique_ptr<ResultSetRow>& parameters, I would have had Caller<...>::call and doCall them take a ResultSetRow& and dereferenced row right away. I don't know what is the exact return type of results->getNextRow() so I won't try to assume anything about it and the type that the main call should take as a parameter.