Passing by rvalue reference is typically an optimization for the case when data can be “stolen” from the parameter instead of copying from it.

struct S
{
    void init(const SomeType& param); // copy from param
    void init(SomeType&& param);      // steal from param
};

People noticed that if SomeType is cheaply movable, then we don’t have to write two overloads and can get away with a single pass-by-value:

struct S
{
    void initByVal(SomeType param); // can now steal from param
};

SomeType t;
s.initByVal(std::move(t)); // move data to param, then can steal from it

Compared to the first version, this will result in one more invocation of SomeType’s move constructor to construct param, but as it is cheap to move it does not matter (much).

It is often mentioned that the first approach would result in 2^N overloads in case of N arguments, and passing by value magically solves this issue. However, passing by reference solves it just as well (see below). Also, passing by value can easily lead to unspecified behavior and, in my opinion, this alone makes it worth considering alternative approaches.

First, note that similarly to only passing by value we might equally well only pass by rvalue reference:

struct S
{
    void initByRef(SomeType&& param);
};

SomeType t;
s.initByRef(std::move(t)); // this works, and even better (less moves)
s.initByRef(SomeType()); // this also works

The only case when it “does not work” is when the argument passed to initByRef() is an lvalue. Note, however, that in this case the pass-by-value approach will do no magic neither - it will copy-construct (not move!) param and this copy cannot be elided. The question of the day is: do you actually want such copies to be created implicitly? If the code above was initByVal(t);, did the author actually mean to copy the object or did they simply forget the std::move()?

Anyway, how to deal with non-movable lvalues? Apart from std::move(), there is a second simple way of turning an lvalue into an rvalue: build a temporary object. Actually, we are already doing this implicitly when passing by value, and of course nothing stops us from doing this explicitly now:

template<typename T>
T copy_to_temp(const T& t) { return t; }

SomeType t; // I'll need t later, so cannot std::move(t)
s.initByRef(copy_to_temp(t)); // but can copy

s.initByVal(t); // note that here t is also copied, it just happens implicitly

While copy_to_temp may not be a great name, I picked it just to show how exactly the same idea can be applied to move-only types (such as unique_ptr<T>). Obviously, we cannot copy_to_temp something that is not copyable, but then we can move_to_temp just as simply:

template<typename T>
T move_to_temp(T& t) { return std::move(t); }

void foo(unique_ptr<SomeType>&& param);

unique_ptr<SomeType> p;
foo(move_to_temp(p)); // note: this does move, p will become empty

But wait a second - why moving to a temporary when we can simply turn it into rvalue foo(std::move(p));? Does not it do the same thing but just less efficiently? Well, it turns out there is a subtle difference: foo(move_to_temp(p)) guarantees that after the operation p will be empty (because it was “moved from”). Herb Sutter called this “the killer argument” for passing by value vs passing by reference in the comments to Scott Meyers’ Should move-only types ever be passed by value?. Well, here is a way to provide the same guarantee when passing by reference if this actually is a requirement: move_to_temp.

On a side note, this is a very weak guarantee. All you get is that data was “moved from” the object that was passed to the function. The data might be copied somewhere, but it might as well be just ignored and destroyed when the temporary goes out of scope. My expectation is that in most cases foo(std::move(p)) will be just as good as foo(move_to_temp(p)) because no one will notice the difference.

To sum up, here is what only passing by reference does compared to only passing by value:

  1. Saves one move construction (modulo move elisions)
  2. Avoids the mentioned unspecified behavior situation
  3. Makes copies as explicit as moves. When you pass an lvalue to the function accepting rvalue references only, you get a compilation error. Is it a bad thing? In my opinion it is actually good - it forces you to decide whether you mean to move or to copy and indicate it explicitly for better maintainability.
  4. Turns adding a const T& overload into an optimization for the case when you cannot move.


blog comments powered by Disqus