Can I typically/always use std::forward instead of std::move?

Joseph Mansfield picture Joseph Mansfield · Nov 4, 2012 · Viewed 7.6k times · Source

I've been watching Scott Meyers' talk on Universal References from the C++ and Beyond 2012 conference, and everything makes sense so far. However, an audience member asks a question at around 50 minutes in that I was also wondering about. Meyers says that he does not care about the answer because it is non-idiomatic and would silly his mind, but I'm still interested.

The code presented is as follows:

// Typical function bodies with overloading:
void doWork(const Widget& param)   // copy
{
  // ops and exprs using param
}
void doWork(Widget&& param)        // move
{
  // ops and exprs using std::move(param)
}

// Typical function implementations with universal reference:
template <typename T>
void doWork(T&& param)             // forward => copy and move
{
  // ops and exprs using std::forward<T>(param)
}

The point being that when we take an rvalue reference, we know we have an rvalue, so we should std::move it to preserve the fact that it's an rvalue. When we take a universal reference (T&&, where T is a deduced type), we want std::forward to preserve the fact that it may have been an lvalue or an rvalue.

So the question is: since std::forward preserves whether the value passed into the function was either an lvalue or an rvalue, and std::move simply casts its argument to an rvalue, could we just use std::forward everywhere? Would std::forward behave like std::move in all cases where we would use std::move, or are there some important differences in behaviour that are missed out by Meyers' generalisation?

I'm not suggesting that anybody should do it because, as Meyers correctly says, it's completely non-idiomatic, but is the following also a valid use of std::move:

void doWork(Widget&& param)         // move
{
  // ops and exprs using std::forward<Widget>(param)
}

Answer

Kerrek SB picture Kerrek SB · Nov 4, 2012

The two are very different and complementary tools.

  • std::move deduces the argument and unconditionally creates an rvalue expression. This makes sense to apply to an actual object or variable.

  • std::forward takes a mandatory template argument (you must specify this!) and magically creates an lvalue or an rvalue expression depending on what the type was (by virtue of adding && and the collapsing rules). This only makes sense to apply to a deduced, templated function argument.

Maybe the following examples illustrate this a bit better:

#include <utility>
#include <memory>
#include <vector>
#include "foo.hpp"

std::vector<std::unique_ptr<Foo>> v;

template <typename T, typename ...Args>
std::unique_ptr<T> make_unique(Args &&... args)
{
    return std::unique_ptr<T>(new T(std::forward<Args>(args)...));  // #1
}

int main()
{
    {
        std::unique_ptr<Foo> p(new Foo('a', true, Bar(1,2,3)));
        v.push_back(std::move(p));                                  // #2
    }

    {
        v.push_back(make_unique<Foo>('b', false, Bar(5,6,7)));      // #3
    }

    {
        Bar b(4,5,6);
        char c = 'x';
        v.push_back(make_unique<Foo>(c, b.ready(), b));             // #4
    }
}

In situation #2, we have an existing, concrete object p, and we want to move from it, unconditionally. Only std::move makes sense. There's nothing to "forward" here. We have a named variable and we want to move from it.

On the other hand, situation #1 accepts a list of any sort of arguments, and each argument needs to be forwarded as the same value category as it was in the original call. For example, in #3 the arguments are temporary expressions, and thus they will be forwarded as rvalues. But we could also have mixed in named objects in the constructor call, as in situation #4, and then we need forwarding as lvalues.