Pass by value vs pass by rvalue reference

When should I declare my function as:

void foo(Widget w);

as opposed to

void foo(Widget&& w);?

Assume this is the only overload (as in, I pick one or the other, not both, and no other overloads). No templates involved. Assume that the function foo requires ownership of the Widget (e.g. const Widget& is not part of this discussion). I'm not interested in any answer outside the scope of these circumstances. See addendum at end of post for why these constraints are part of the question.

The primary difference that my colleagues and I can come up with is that the rvalue reference parameter forces you to be explicit about copies. The caller is responsible for making an explicit copy and then passing it in with std::move when you want a copy. In the pass by value case, the cost of the copy is hidden:

    //If foo is a pass by value function, calling + making a copy:
    Widget x{};
    foo(x); //Implicit copy
    //Not shown: continues to use x locally
    
    //If foo is a pass by rvalue reference function, calling + making a copy:
    Widget x{};
    //foo(x); //This would be a compiler error
    auto copy = x; //Explicit copy
    foo(std::move(copy));
    //Not shown: continues to use x locally

Other than that difference. Other than forcing people to be explicit about copying and changing how much syntactic sugar you get when calling the function, how else are these different? What do they say differently about the interface? Are they more or less efficient than one another?

Other things that my colleagues and I have already thought of:

• The rvalue reference parameter means that you may move the argument, but does not mandate it. It is possible that the argument you passed in at the call site will be in its original state afterwards. It's also possible the function would eat/change the argument without ever calling a move constructor but assume that because it was an rvalue reference, the caller relinquished control. Pass by value, if you move into it, you must assume that a move happened; there's no choice.
• Assuming no elisions, a single move constructor call is eliminated with pass by rvalue.
• The compiler has better opportunity to elide copies/moves with pass by value. Can anyone substantiate this claim? Preferably with a link to gcc.godbolt.org showing optimized generated code from gcc/clang rather than a line in the standard. My attempt at showing this was probably not able to successfully isolate the behavior: https://godbolt.org/g/4yomtt

Addendum: why am I constraining this problem so much?

• No overloads - if there were other overloads, this would devolve into a discussion of pass by value vs a set of overloads that include both const reference and rvalue reference, at which point the set of overloads is obviously more efficient and wins. This is well known, and therefore not interesting.
• No templates - I'm not interested in how forwarding references fit into the picture. If you have a forwarding reference, you call std::forward anyway. The goal with a forwarding reference is to pass things as you received them. Copies aren't relevant because you just pass an lvalue instead. It's well known, and not interesting.
• foo requires ownership of Widget (aka no const Widget&) - We're not talking about read-only functions. If the function was read-only or didn't need to own or extend the lifetime of the Widget, then the answer trivially becomes const Widget&, which again, is well known, and not interesting. I also refer you to why we don't want to talk about overloads.

What do rvalue usages say about an interface versus copying? rvalue suggests to the caller that the function both wants to own the value and has no intention of letting the caller know of any changes it has made. Consider the following (I know you said no lvalue references in your example, but bear with me):

//Hello. I want my own local copy of your Widget that I will manipulate,
//but I don't want my changes to affect the one you have. I may or may not
//hold onto it for later, but that's none of your business.
void foo(Widget w);

//Hello. I want to take your Widget and play with it. It may be in a
//different state than when you gave it to me, but it'll still be yours
//when I'm finished. Trust me!
void foo(Widget& w);

//Hello. Can I see that Widget of yours? I don't want to mess with it;
//I just want to check something out on it. Read that one value from it,
//or observe what state it's in. I won't touch it and I won't keep it.
void foo(const Widget& w);

//Hello. Ooh, I like that Widget you have. You're not going to use it
//anymore, are you? Please just give it to me. Thank you! It's my
//responsibility now, so don't worry about it anymore, m'kay?
void foo(Widget&& w);

For another way of looking at it:

//Here, let me buy you a new car just like mine. I don't care if you wreck
//it or give it a new paint job; you have yours and I have mine.
void foo(Car c);

//Here are the keys to my car. I understand that it may come back...
//not quite the same... as I lent it to you, but I'm okay with that.
void foo(Car& c);

//Here are the keys to my car as long as you promise to not give it a
//paint job or anything like that
void foo(const Car& c);

//I don't need my car anymore, so I'm signing the title over to you now.
//Happy birthday!
void foo(Car&& c);

Now, if Widgets have to remain unique (as actual widgets in, say, GTK do) then the first option cannot work. The second, third and fourth options make sense, because there's still only one real representation of the data. Anyway, that's what those semantics say to me when I see them in code.

Now, as for efficiency: it depends. rvalue references can save a lot of time if Widget has a pointer to a data member whose pointed-to contents can be rather large (think an array). Since the caller used an rvalue, they're saying they don't care about what they're giving you anymore. So, if you want to move the caller's Widget's contents into your Widget, just take their pointer. No need to meticulously copy each element in the data structure their pointer points to. This can lead to pretty good improvements in speed (again, think arrays). But if the Widget class doesn't have any such thing, this benefit is nowhere to be seen.

Hopefully that gets at what you were asking; if not, I can perhaps expand/clarify things.

> The rvalue reference parameter forces you to be explicit about copies.

Yes, pass-by-rvalue-reference got a point.

> The rvalue reference parameter means that you may move the argument, but does not mandate it.

Yes, pass-by-value got a point.

But that also gives to pass-by-rvalue the opportunity to handle exception guarantee: if foo throws, widget value is not necessary consumed.

For move-only types (as std::unique_ptr), pass-by-value seems to be the norm (mostly for your second point, and first point is not applicable anyway).

EDIT: standard library contradicts my previous sentence, one of shared_ptr's constructor takes std::unique_ptr<T, D>&&.

For types which have both copy/move (as std::shared_ptr), we have the choice of the coherency with previous types or force to be explicit on copy.

Unless you want to guarantee there is no unwanted copy, I would use pass-by-value for coherency.

Unless you want guaranteed and/or immediate sink, I would use pass-by-rvalue.

For existing code base, I would keep consistency.

Unless the type is a move-only type you normally have an option to pass by reference-to-const and it seems arbitrary to make it "not part of the discussion" but I will try.

I think the choice partly depends on what foo is going to do with the parameter.

The function needs a local copy

Let's say Widget is an iterator and you want to implement your own std::next function. next needs its own copy to advance and then return. In this case your choice is something like:

Widget next(Widget it, int n = 1){
    std::advance(it, n);
    return it;
}

vs

Widget next(Widget&& it, int n = 1){
    std::advance(it, n);
    return std::move(it);
}

I think by-value is better here. From the signature you can see it is taking a copy. If the caller wants to avoid a copy they can do a std::move and guarantee the variable is moved from but they can still pass lvalues if they want to. With pass-by-rvalue-reference the caller cannot guarantee that the variable has been moved from.

Move-assignment to a copy

Let's say you have a class WidgetHolder:

class WidgetHolder {
    Widget widget;
   //...
};

and you need to implement a setWidget member function. I'm going to assume you already have an overload that takes a reference-to-const:

WidgetHolder::setWidget(const Widget& w) {
    widget = w;
}

but after measuring performance you decide you need to optimize for r-values. You have a choice between replacing it with:

WidgetHolder::setWidget(Widget w) {
    widget = std::move(w);
}

Or overloading with:

WidgetHolder::setWidget(Widget&& widget) {
    widget = std::move(w);
}

This one is a little bit more tricky. It is tempting choose pass-by-value because it accepts both rvalues and lvalues so you don't need two overloads. However it is unconditionally taking a copy so you can't take advantage of any existing capacity in the member variable. The pass by reference-to-const and pass by r-value reference overloads use assignment without taking a copy which might be faster

Move-construct a copy

Now lets say you are writing the constructor for WidgetHolder and as before you have already implemented a constructor that takes an reference-to-const:

WidgetHolder::WidgetHolder(const Widget& w) : widget(w) {
}

and as before you have measured peformance and decided you need to optimize for rvalues. You have a choice between replacing it with:

WidgetHolder::WidgetHolder(Widget w) : widget(std::move(w)) {
}

Or overloading with:

WidgetHolder::WidgetHolder(Widget&& w) : widget(std:move(w)) {
}

In this case, the member variable cannot have any existing capacity since this is the constructor. You are move-constucting a copy. Also, constructors often take many parameters so it can be quite a pain to write all the different permutations of overloads to optimize for r-value references. So in this case it is a good idea to use pass-by-value, especially if the constructor takes many such parameters.

Passing unique_ptr

With unique_ptr the efficiency concerns are less important given that a move is so cheap and it doesn't have any capacity. More important is expressiveness and correctness. There is a good discussion of how to pass unique_ptr here.

One issue not mentioned in the other answers is the idea of exception-safety.

In general, if the function throws an exception, we would ideally like to have the strong exception guarantee, meaning that the call has no effect other than raising the exception. If pass-by-value uses the move constructor, then such an effect is essentially unavoidable. So an rvalue-reference argument may be superior in some cases. (Of course, there are various cases where the strong exception guarantee isn't achievable either way, as well as various cases where the no-throw guarantee is available either way. So this is not relevant in 100% of cases. But it's relevant sometimes.)

When you pass by rvalue reference object lifetimes get complicated. If the callee does not move out of the argument, the destruction of the argument is delayed. I think this is interesting in two cases.

First, you have an RAII class

void fn(RAII &&);

RAII x{underlying_resource};
fn(std::move(x));
// later in the code
RAII y{underlying_resource};

When initializing y, the resource could still be held by x if fn doesn't move out of the rvalue reference. In the pass by value code, we know that x gets moved out of, and fn releases x. This is probably a case where you would want to pass by value, and the copy constructor would likely be deleted, so you wouldn't have to worry about accidental copies.

Second, if the argument is a large object and the function doesn't move out, the lifetime of the vectors data is larger than in the case of pass by value.

vector<B> fn1(vector<A> &&x);
vector<C> fn2(vector<B> &&x);

vector<A> va;  // large vector
vector<B> vb = fn1(std::move(va));
vector<C> vc = fn2(std::move(vb));

In the example above, if fn1 and fn2 don't move out of x, then you will end up with all of the data in all of the vectors still alive. If you instead pass by value, only the last vector's data will still be alive (assuming vectors move constructor clears the sources vector).

Choosing between by-value and by-rvalue-ref, with no other overloads, is not meaningful.

With pass by value the actual argument can be an lvalue expression.

With pass by rvalue-ref the actual argument must be an rvalue.

---

If the function is storing a copy of the argument, then a sensible choice is between pass-by-value, and a set of overloads with pass-by-ref-to-const and pass-by-rvalue-ref. For an rvalue expression as actual argument the set of overloads can avoid one move. It's an engineering gut-feeling decision whether the micro-optimization is worth the added complexity and typing.

One notable difference is that if you move to an pass-by-value function:

void foo(Widget w);
foo(std::move(copy));

compiler must generate a move-constructor call Widget(Widget&&) to create the value object. In case of pass-by-rvalue-reference no such call is needed as the rvalue-reference is passed directly to the method. Usually this does not matter, as move constructors are trivial (or default) and are inlined most of the time. (you can check it on gcc.godbolt.org -- in your example declare move constructor Widget(Widget&&); and it will show up in assembly)

So my rule of thumb is this:

• if the object represents a unique resource (without copy semantics) I prefer to use pass-by-rvalue-reference,
• otherwise if it logically makes sense to either move or copy the object, I use pass-by-value.