Jtdcftul

Suppose I want to write a function that calls a nullary function 100 times. Which of these implementations is best and why?

template<typename F>

void call100(F f) {

    for (int i = 0; i < 100; i++)

        f();

}



template<typename F>

void call100(F& f) {

    for (int i = 0; i < 100; i++)

        f();

}



template<typename F>

void call100(const F& f) {

    for (int i = 0; i < 100; i++)

        f();

}





template<typename F>

void call100(F&& f) {

    for (int i = 0; i < 100; i++)

        f();

}

Or is there a better implementation?

Update regarding 4

struct S {

    S() {}

    S(const S&) = delete;

    void operator()() const {}

};



template<typename F>

void call100(F&& f) {

    for (int i = 0; i < 100; i++)

        f();

}



int main() {

    const S s;

    call100(s);

}

I would use the first one (pass the callable by value).

If a caller is concerned about the cost of copying the callable, then they can use std::ref(f) or std::cref(f) to pass it using reference_wrapper.

By doing this, you provide the most flexibility to the caller.

The only runtime cost of

template<typename F>

void call100(F&& f) {

  for (int i = 0; i < 100; ++i)

    f();

}

is that it can have more versions (copies of code) if you pass f in multiple ways. With MSVC or the gold linker with ICF, those copies only cost compile time unless they differ, and if they differ you probably want to keep them.

template<typename F>

void call100(F f) {

  for (int i = 0; i < 100; ++i)

    f();

}

this one has the advantage of being value semantics; and following the rule of taking values unless you have good reason not to is reasonable. std::ref/std::cref let you call it with a persistant reference, and for prvalues c++17 guaranteed elision will prevent a spurious copy.

As a joke you could do:

template<typename F>

void call100(F&& f) {

  for (int i = 0; i < 99; ++i)

    f();

  std::forward<F>(f)();

}

but that relies on people having && overloads on their operator(), which nobody does.

I do not think there is a definitive answer:

1. 
   

   The first one copies everything you pass in which might be expensive
   for capturing lambdas but otherwise provides the most flexibility:

   
   
   

   Pros

   
   
   
   
   
   • Const objects allowed
   
   
   • Mutable objects allowed (copied)
   
   
   • Copy can be elided (?)
   
   
   
   

   Cons

   
   
   
   
   
   • Copies everything you give it
   
   
   • You cannot call it with an existing object such as mutable lambda without copying it in
   
   
   
   


2. 
   

   The second one cannot be used for const objects. On the other hand
   it does not copy anything and allows mutable objects:

   
   
   

   Pros

   
   
   
   
   
   • Mutable objects allowed
   
   
   • Copies nothing
   
   
   
   

   Cons

   
   
   
   
   
   • Does not allow const objects
   
   
   
   


3. 
   

   The third one cannot be used for mutable lambdas so is a slight
   modification of the second one.

   
   
   

   Pros

   
   
   
   
   
   • Const objects allowed
   
   
   • Copies nothing
   
   
   
   

   Cons

   
   
   
   
   
   • Cannot be called with mutable objects
   
   
   
   


4. 
   

   The fourth one cannot be called with const objects unless you copy
   them which becomes quite awkward with lambdas. You also cannot use
   it with pre-existing mutable lambda object without copying it or
   moving from it (losing it in the process) which is similar
   limitation to 1.

   
   
   

   Pros

   
   
   
   
   
   • Avoids copies explicitely by forcing (requiring) move semanthics if the copy is needed
   
   
   • Mutable objects allowed.
   
   
   • Const objects allowed (except for mutable lambdas)
   
   
   
   

   Cons

   
   
   
   
   
   • Does not allow const mutable lambdas without a copy
   
   
   • You cannot call it with an existing object such as mutable lambda
   
   
   
   

And there you have it. There is no silver bullet here and there are different pros & cons to each of these versions. I tend to lean towards the first one being the default but with certain types of capturing lambdas or bigger callables, it might become an issue. And you cannot call the 1) with the mutable object and get an expected result. As mentioned in the other answer some of these can be overcome with std::ref and other ways of manipulating the actual T type. In my experience, these tend to be the source of pretty nasty bugs though when T is then something different than one expects to achieve i.e. mutability of a copy or such.