In the previous article I have shown you how variadic templates can be a type-safe alternative to va_arg macros in designing the variadic functions. Today I want to show you a few techniques of variadic templates, that can be found in many codebases.

Recursive variadic templates

This technique has been already covered in the previous post, but let’s keep it anyway for completeness.

#include <iostream>

void showRecursive() {}

template <typename T, typename... Args>
void showRecursive(T first, Args... args)
{
    std::cout << "showRecursive: " << first << '\n';
    showRecursive(args...);
}

int main()
{
    showRecursive(1, 2, 3, 4);
    return 0;
}

Here we have a slightly different example with only one template. showRecursive is a function that prints its arguments, each in a separate line. As we already know, we have to unpack the first argument from the parameter pack by explicitly declaring it in the function parameter list. Thus each instantiation of the template will assign its leading argument to the first variable and leave the rest as the remaining pack.

In order to enable the recursion, we call the same function with the expansion of the remaining parameter pack. In every iteration the argument pack will be shorter by one element. Finally, in the last round showRecursive will be called with an empty pack. To handle this base case we can use a simple overload that takes no arguments (this is why it doesn’t need to be a template).

The expected result is of course:

showRecursive: 1
showRecursive: 2
showRecursive: 3
showRecursive: 4

Non-recursive argument evaluation with std::initializer_list

Variadic templates don’t have to be recursive. In fact this approach has several drawbacks:

• it makes the code less readable,
• it potentially generates more code,
• it requires a correct base case (stop condition) implementation.

Our preferred way is to iterate over all arguments and pass them one by one to some single argument function (like in the for loop) and ultimately get something like this:

func(arg1), func(arg2), ..., func(argN);

C++ doesn’t have an explicit way of achieving that. Fortunately, there is an implicit method, that involves usage of the infamous std::initializer_list. Let’s see:

#include <initializer_list>
#include <iostream>

template <typename T>
void showImpl(T arg)
{
    std::cout << "showImpl: " << arg << '\n';
}

template <typename... Args>
void showNonRecursive(Args... args)
{
    std::initializer_list<int>{(showImpl(args), 0)...};
}

int main()
{
    showNonRecursive(1, 2, 3, 4);
    return 0;
}

The first notable difference in this example is lack of the first unpacked argument. We are not basing out implementation on the recursion, so it is not needed. Instead the showNonRecursive is creating an object of std::initializer_list<int>. The argument of its constructor is particularly interesting:

(showImpl(args), 0)...

This is a pack expansion, but used on the comma operator expression. Arguments of the comma operator are always evaluated left-to-right, but the result is equal to the last argument. So in our case the first argument (which is a call to showImpl) is evaluated with the first unpacked value from the parameter pack and 0 is returned as a result of the comma operator. Then the second unpacked argument is passed to the constructor of the std::initializer_list in the same manner. And again, the result of the operator is 0. In pseudo-code this operation may be executed like this:

std::initializer_list<int>{(showImpl(arg1), 0), (showImpl(arg2), 0), ..., (showImpl(argN), 0)};

If we leave only the side effects (results of the expressions), then it may look like this:

std::initializer_list<int>{0, 0, ..., 0};

This explains, why we used int as a template argument in the std::initializer_list. In the end, the created object is not assigned to any variable, so compiler is free to optimize it away (however I didn’t confirmed that it actually does that).

To sum up, we are creating a fake object only to use the fact, that initializer lists are guaranteed to be evaluated in order. Since we can’t create an initializer list out of the parameter pack (which could potentially have multiple types), we use a list of 0s preceded with a demanded function. The glue that makes it work is the comma operator, which evaluates both showImpl and 0, leaving 0 as the side effect.

Expression expansion

Parameter expansion can be used in many ways and in different code contexts. In the previous example we have seen that applying a function on the parameter pack in fact calls this function for every element of the pack. The same can be done with the single argument operators like incrementing, casting or extracting an address. Below you can find a list of most interesting use cases for the expression expansion:

Context	Expansion
Function arguments	f(sizeof(args)...) -> f(sizeof(arg1), sizeof(arg2), ..., sizeof(argN)
Template arguments	print<Param1, Args..., Param2> -> print<Param1, Arg1, Arg2, ..., ArgN, Param2>
Inheritance list	template class Derived : public Bases... {} -> template <typename Base1, typename Base2, ..., typename BaseN> class Derived : public Base1, public Base2, ..., public BaseN {}
Lambda capture list	[args...]() {} -> [arg1, arg2, ..., argN]() {}

C++17 fold expressions

C++17 added even more operations, that can be used with the parameter pack. The main idea is to allow using two-argument operators. C++17 standard specifies 4 acceptable syntax variations for this:

1. unary right fold (pack op ...),
2. unary left fold (... op pack),
3. binary right fold (pack op ... op init),
4. binary left fold (init op ... op pack).

So our summing example could be implemented using the unary right fold expression like this:

#include <iostream>

template <typename... Args>
auto sum(Args... args)
{
    return (args + ...);
}

int main()
{
   std::cout << sum(1, 2, 3, 4) << '\n';
   return 0;
}

On the other hand, the “no separate line” variation of the printing example could look like this using the binary left fold expression:

#include <iostream>

template <typename... Args>
void foldPrint(Args... args)
{
    (std::cout << "foldPrint: " << ... << args) << '\n';
}

int main()
{
    foldPrint(1, 2, 3, 4);
    return 0;
}

The result of the above code is:

foldPrint: 1234

Below you can find the list of all acceptable operators in the fold expressions:

+   -   *   /   %   ^   &   |   <<   >> 
+=  -=  *=  /=  %=  ^=  &=  |=  <<=  >>=  =
==  !=  <   >   <=  >=  &&  ||  ,    .*   ->*

Usage of the sizeof… operator

When using variadic templates you may end up in a situation, where you would like to know how many arguments are actually passed. Let’s say that you want to store them in a table. How big should it be? sizeof...() will tell you:

template <typename... Args>
void func(Args... args)
{
    int argsTable[sizeof...(args)] = {args...}; 
    // Some other code.
}

This technique is pretty straightforward. sizeof... operator called on the parameter pack will tell you the number of its elements at compile time. It can be used in any context, that requires the compile-time evaluation.

Perfect forwarding (FYI)

There is one last very important technique called perfect forwarding. It is not limited only to variadic templates, but in my opinion it brings the most benefits in this context. Its only purpose is to pass (forward) function template arguments into another function preserving all their properties like qualifiers or value semantics (lvalue, rvalue etc…). However, it requires more knowledge about universal/forward references (&&) and move semantics. It is also very rarely found in a typical business code, thus I will not cover its details here. Usually, it is exploited in the highly customizable libraries, where client can specify its own callback functions. STL would be an excellent example of this. Some of the most popular standard functions and classes that use perfect forwarding are:

• std::make_shared() / std::make_unique(),
• std::thread,
• std::function,
• std::vector::emplace_back().

I hope that now you know what can be done with variadic templates and how to start using it. cppreference can be a useful resource when in doubt.

---