std::forward acting differently - c++

I'm currently learning perfect forwarding in c++ and I came across something that confused me. Pretty sure it's something stupid. When I used std::forward on a lvalue, it used the rvalue function. This is a pretty bad explanation so I'm just gonna show the code.
#include <iostream>
void check(int&& other) {
std::cout << "Rvalue" << std::endl;
}
void check(int& other) {
std::cout << "Lvalue" << std::endl;
}
void call(int& other) {
check(std::forward<int>(other));
}
int main()
{
int i = 4;
call(i);
}
This outputs "Rvalue". Please help me understand why.

"Perfect forwarding" is used in templates, i.e.:
template<typename T>
void func(T && arg)
{
func2(std::forward<T>(arg));
}
Outside of template context the end result drastically changes. All that std::forward does is return static_cast<T &&>, so your call function becomes nothing more than:
void call(int& other) {
check(static_cast<int &&>(other));
}
Hence you get an rvalue. The reason this works differently in templates is because a && template parameter is a forwarding reference (a fancy term for either an lvalue or an rvalue-deduced reference, depending on what gets dropped in that parameter), and because of reference collapsing rules. Briefly, when used in a template context, the end result is:
T gets deduced as either an lvalue or an rvalue reference, depending on what the parameter is.
The result of the static_cast<T &&> is an lvalue reference, if T is an lvalue reference, or an rvalue reference if T is an rvalue reference, due to reference collapsing rules.
The end result is that the same kind of a reference gets forwarded. But this only works in template context, since it requires both forwarding reference semantics and reference collapsing rules to work just right.

std::forward() isn't exactly magic. Which is one of the reasons one has to give it the appropriate type with the appropriate reference-category (rvalue-reference, lvalue-reference, no reference).
Normally, you get the type from the template-argument. If it's an auto&&-argument (C++14 lambda, C++20 abbreviated template or use of concept), one uses decltype() to get it from the function-argument.
Manually specifying it works, but goes against the spirit of using the function. std::move() or the argument itself is much easier to use in that case, depending on the template-argument you specify.
And due to how reference-collapsing and std::forward() are defined, Type and Type&& both result in an rvalue-reference, while Type& results in an lvalue-reference.

Related

Why is T const&& not a forwarding reference?

In the context of a template, the following "reference collapsing" rules are applied:
template <typename T>
void foo(T && t)
{
//T& & -> T&
//T& && -> T&
//T&& & -> T&
//T&& && -> T&&
}
Why does the language prohibit "universal references" from having const qualifiers?
template <typename T>
void foo(T const && t)
It would seem to make sense if the type had resolved to a reference (3 out of the 4 cases).
I'm sure this idea is incompatible with some other design aspect of the language, but I can't quite see the full picture.
Originally the rvalue reference proposal said that the transformation happens if P is "an rvalue reference type". However, a defect report later noticed
Additionally, consider this case:
template <class T> void f(const T&&);
...
int i;
f(i);
If we deduce T as int& in this case then f(i) calls f<int&>(int&), which seems counterintuitive. We prefer that f<int>(const int&&) be called. Therefore, we would like the wording clarified that the A& deduction rule in 14.8.2.1 [temp.deduct.call] paragraph 3 applies only to the form T&& and not to cv T&& as the note currently implies.
There appears to have been a time period where const T &&, with T being U&, was transformed to const U&. That was changed to be consistent with another rule that says that const T, where T is U& would stay U& (cv-qualifiers on references are ignored). So, when you would deduce T in above example to int&, the function parameter would stay int&, not const int&.
In the defect report, the reporter states "We prefer that f<int>(const int&&) be called", however provides no reason in the defect report. I can imagine that the reason was that it seemed too intricate to fix this without introducing inconsistency with other rules, however.
We should also keep in mind that the defect report was made at a time where rvalue references could still bind to lvalues - i.e const int&& could bind to an int lvalue. This was prohibited only later on, when a paper by Dave & Doug, "A Safety Problem with RValue References", appeared. So, it seems to me that a deduction that works (at that time) was worth more than a deduction that simply was counter intuitive and dropped qualifiers.
This does already happen for references; if your T is a U const &, then T && will collapse to U const &. The term "universal reference" really does mean universal reference: you don't need to specify const in there to get a constant reference.
If you want to have a truly universal reference mechanism, you need your T && to be able to become all kinds of references, will all kinds of constness. And, T && does exactly that. It collapses to all four cases: both l- and r-value references, and both const and non-const.
Explained another way, the constness is an attribute of the type, not the reference, i.e. when you say T const &, you are actually talking about a U &, where U is T const. The same is true for && (although an r-value reference to a const is less useful).
This means that if you want your universal reference to collapse to a U const &, just pass it something that is of the type you want: a U const &, and it will collapse to exactly that.
To answer you question more directly: the language does not "prohibit" the use of const in the declaration of a universal reference, per sé. It is saying that if you change the mechanism for declaring a universal reference even a little bit - even by inserting a lowly const between the T and the && - then you won't have a (literally) "universal" reference anymore, because it just won't accept anything and everything.
Why do you think the language does not allow const r-value references?
In the following code, what will be printed?
#include <iostream>
struct Foo
{
void bar() const &
{
std::cout << "&\n";
}
void bar() const &&
{
std::cout << "&&\n";
}
};
const Foo make() {
return Foo{};
}
int main()
{
make().bar();
}
answer:
&&
why? Because make() returns a const object and in this context it's a temporary. Therefore r-value reference to const.
Template argument deduction has a special case for "rvalue reference to cv-unqualified template parameters". It is this very special case that forwarding/universal references rely on. See section "Deduction from a function call" in the linked article for details.
Note that before template argument deduction, all top-level cv-qualifiers are removed; however, references never have top-level cv-qualifiers and above rule does not apply, so the special rule also does not apply. (In contrast to pointers, there is no "const reference", only "reference to const")

What is Perfect Forwarding equal to

I know that this can be used to perform perfect forwarding:
template <typename A>
void foo(A&&) { /* */ }
This can be used to perform perfect forwarding on a certain type:
template <typename A, std::enable_if_t<std::is_same<std::decay_t<A>, int>::value, int> = 0>
void foo(A&&) { /* */ }
But these are just templates for functions, which means, that these get expanded to some functions, which are then used for every special case in which it might be used. However do these get expanded to:
void foo(A&) and void foo(A&&)
OR
void foo(A&) and void foo(A)
I always thought, it would be the first one, but then I noticed, that in that case, you wouldn't be able to use A const as an argument to the function, which certainly works.
However the second would be ambiguous, if you used a normal non-const lvalue. Does it call foo(A&) or foo(A)?
It's the first one. The second wouldn't make very much sense: there is no A such that A&& is a non-reference type.
If the argument is an lvalue of type cv T, then A is deduced as cv T&. If the argument is an rvalue of type cv T, then A is deduced as cv T and A&& is cv T&&. So when you pass in a const lvalue, the specialization generated is one that can accept a const argument.
They were called originally "Univeral References" by Scott Meyers, and now "Forwarding References".
As you can see, the references part has not changed. You pass in any kind of rvalue, you get a rvalue reference. You pass in any kind of lvalue, and you get a lvalue reference. Life is that simple.

How is the move constructor of member variable invoked without using std::forward?

An example here for std::forward,
// forward example
#include <utility> // std::forward
#include <iostream> // std::cout
// function with lvalue and rvalue reference overloads:
void overloaded (const int& x) {std::cout << "[lvalue]";}
void overloaded (int&& x) {std::cout << "[rvalue]";}
// function template taking rvalue reference to deduced type:
template <class T> void fn (T&& x) {
overloaded (x); // always an lvalue
overloaded (std::forward<T>(x)); // rvalue if argument is rvalue
}
int main () {
int a;
std::cout << "calling fn with lvalue: ";
fn (a);
std::cout << '\n';
std::cout << "calling fn with rvalue: ";
fn (0);
std::cout << '\n';
return 0;
}
Output:
calling fn with lvalue: [lvalue][lvalue]
calling fn with rvalue: [lvalue][rvalue]
mentions that
the fact that all named values (such as function parameters) always
evaluate as lvalues (even those declared as rvalue references)
Whereas, the typical move constructor looks like
ClassName(ClassName&& other)
: _data(other._data)
{
}
which looks like _data(other._data) should invoke the move constructor of _data's class. But, how is it possible without using std::forward? In other words, shouldn't it be
ClassName(ClassName&& other)
: _data(std::forward(other._data))
{
}
?
Because, as pointed out in std:forward case,
all then named values should evaluate as lvalue
I more and more like C++ because of the depth of issue like this and the fact that the language is bold enough to provide such features :) Thank you!
A typical move constructor looks like this (assuming it is explicitly implemented: you might want to prefer = default):
ClassName::ClassName(ClassName&& other)
: _data(std::move(other._data)) {
}
Without the std::move() the member is copied: since it has a name other is an lvalue. The object the reference is bound to is an rvalue or an object considered as such, however.
std::forward<T>(obj) is always used with an explicit template argument. In practice the type is that deduced for a forwarding reference. These look remarkably like rvalue references but are something entirely different! In particular, a forwarding reference may refer to an lvalue.
You may be interested in my Two Daemons article which describes the difference in detail.
std::forward should be used with a forwarding reference.
std::move should be used with an rvalue reference.
There is nothing particular about constructors. The rules apply the same to any function, member function or constructor.
The most important thing is to realize when you have a forwarding reference and when you have an rvalue reference. They look similar but are not.
A forwarding reference is always in the form:
T&& ref
for T some deduced type.
For instance, this is a forwarding reference:
template <class T>
auto foo(T&& ref) -> void;
All these are rvalue references:
auto foo(int&& ref) -> void; // int not deduced
template <class T>
auto foo(const T&& ref); // not in form `T&&` (note the const)
template <class T>
auto foo(std::vector<T>&& ref) -> void; // not in form `T&&`
template <class T>
struct X {
auto foo(T&& ref) -> T; // T not deduced. (It was deduced at class level)
};
For more please check this excellent in-depth article by Scott Meyers with the note that when the article was written the term "universal reference" was used (actually introduced by Scott himself). Now it is agreed that "forwarding reference" better describes it's purpose and usage.
So your example should be:
ClassName(ClassName&& other)
: _data(std::move(other._data))
{
}
as other is an rvalue reference because ClassName is not a deduced type.
This Ideone example should make things pretty clear for you. If not, keep reading.
The following constructor accepts Rvalues only. However, since the argument "other" got a name it lost its "rvalueness" and now is a Lvalue. To cast it back to Rvalue, you have to use std::move. There's no reason to use std::forward here because this constructor does not accept Lvalues. If you try to call it with a Lvalue, you will get compile error.
ClassName(ClassName&& other)
: _data(std::move(other._data))
{
// If you don't use move, you could have:
// cout << other._data;
// And you will notice "other" has not been moved.
}
The following constructor accepts both Lvalues and Rvalues. Scott Meyers called it "Universal Rerefences", but now it's called "Forwarding References". That's why, here, it's a must to use std::forward so that if other was an Rvalue, _data constructor will get called with an Rvalue. If other was an Lvalue, _data will be constructed with an Lvalue. That's why it's called perfect-forwarding.
template<typename T>
ClassName(T&& other)
: _data(std::forward<decltype(_data)>(other._data))
{
}
I've tried to use your constructors as an example so you could understand, but this is not specific to constructors. This applies to functions as well.
With the first example tho, since your first constructor only accepts Rvalues, you could perfectly use std::forward instead, and both would do the same thing. But it's best not to do it, because people may think that your constructor accepts a forwarding reference, when it actually doesn't.

Why does std::move take a forward reference?

The implementation of std::move basically looks like this:
template<typename T>
typename std::remove_reference<T>::type&&
move(T&& t)
{
return static_cast<typename std::remove_reference<T>::type&&>(t);
}
Note that the parameter of std::move is a universal reference (also known as a forwarding reference, but we're not forwarding here). That is, you can std::move both lvalues and rvalues:
std::string a, b, c;
// ...
foo(std::move(a)); // fine, a is an lvalue
foo(std::move(b + c)); // nonsense, b + c is already an rvalue
But since the whole point of std::move is to cast to an rvalue, why are we even allowed to std::move rvalues? Wouldn't it make more sense if std::move would only accept lvalues?
template<typename T>
T&&
move(T& t)
{
return static_cast<T&&>(t);
}
Then the nonsensical expression std::move(b + c) would cause a compile-time error.
The above implementation of std::move would also be much easier to understand for beginners, because the code does exactly what it appears to do: It takes an lvalue and returns an rvalue. You don't have to understand universal references, reference collapsing and meta functions.
So why was std::move designed to take both lvalues and rvalues?
Here is some example simplified to the extreme:
#include <iostream>
#include <vector>
template<typename T>
T&& my_move(T& t)
{
return static_cast<T&&>(t);
}
int main()
{
std::vector<bool> v{true};
std::move(v[0]); // std::move on rvalue, OK
my_move(v[0]); // my_move on rvalue, OOPS
}
Cases like the one above may appear in generic code, for example when using containers which have specializations that return proxy objects (rvalues), and you may not know whether the client will be using the specialization or not, so you want unconditional support for move semantics.
It doesn't hurt.
You're simply establishing a guarantee that code will treat the result as an rvalue. You certainly could write std::move in such way that it errors out when dealing with something that's already an rvalue, but what is the benefit?
In generic code, where you don't necessarily know what type(s) you're going to be working with, what gains in expressiveness would you extract out of a bunch of "if type is rvalue do nothing else std::move" plastered everywhere when you can simply say "I promise we can think of this as an rvalue".
You said it yourself, it is nothing more than a cast. Should *_cast operations also fail if the argument already matches the intended type?

Usage of std::forward vs std::move

I always read that std::forward is only for use with template parameters. However, I was asking myself why. See the following example:
void ImageView::setImage(const Image& image){
_image = image;
}
void ImageView::setImage(Image&& image){
_image = std::move(image);
}
Those are two functions which basically do the same; one takes an l-value reference, the other an r-value reference. Now, I thought since std::forward is supposed to return an l-value reference if the argument is an l-value reference and an r-value reference if the argument is one, this code could be simplified to something like this:
void ImageView::setImage(Image&& image){
_image = std::forward(image);
}
Which is kind of similar to the example cplusplus.com mentions for std::forward (just without any template parameters). I'd just like to know, if this is correct or not, and if not why.
I was also asking myself what exactly would be the difference to
void ImageView::setImage(Image& image){
_image = std::forward(image);
}
You cannot use std::forward without explicitly specifying its template argument. It is intentionally used in a non-deduced context.
To understand this, you need to really understand how forwarding references (T&& for a deduced T) work internally, and not wave them away as "it's magic." So let's look at that.
template <class T>
void foo(T &&t)
{
bar(std::forward<T>(t));
}
Let's say we call foo like this:
foo(42);
42 is an rvalue of type int.
T is deduced to int.
The call to bar therefore uses int as the template argument for std::forward.
The return type of std::forward<U> is U && (in this case, that's int &&) so t is forwarded as an rvalue.
Now, let's call foo like this:
int i = 42;
foo(i);
i is an lvalue of type int.
Because of the special rule for perfect forwarding, when an lvalue of type V is used to deduce T in a parameter of type T &&, V & is used for deduction. Therefore, in our case, T is deduced to be int &.
Therefore, we specify int & as the template argument to std::forward. Its return type will therefore be "int & &&", which collapses to int &. That's an lvalue, so i is forwarded as an lvalue.
Summary
Why this works with templates is when you do std::forward<T>, T is sometimes a reference (when the original is an lvalue) and sometimes not (when the original is an rvalue). std::forward will therefore cast to an lvalue or rvalue reference as appropriate.
You cannot make this work in the non-template version precisely because you'll have only one type available. Not to mention the fact that setImage(Image&& image) would not accept lvalues at all—an lvalue cannot bind to rvalue references.
I recommend reading "Effective Modern C ++" by Scott Meyers, specifically:
Item 23: Understand std::move and std::forward.
Item 24: Distinguish universal references for rvalue references.
From a purely technical perspective, the answer is yes: std::forward
can do it all. std::move isn’t necessary. Of course, neither function
is really necessary, because we could write casts everywhere, but I
hope we agree that that would be, well, yucky. std::move’s attractions
are convenience, reduced likelihood of error, and greater clarity.
rvalue-reference
This function accepts rvalues and cannot accept lvalues.
void ImageView::setImage(Image&& image){
_image = std::forward(image); // error
_image = std::move(image); // conventional
_image = std::forward<Image>(image); // unconventional
}
Note first that std::move requires only a function argument, while std::forward requires both a function argument and a template type argument.
Universal references (forwarding references)
This function accepts all and does perfect forwarding.
template <typename T> void ImageView::setImage(T&& image){
_image = std::forward<T>(image);
}
You have to specify the template type in std::forward.
In this context Image&& image is always an r-value reference and std::forward<Image> will always move so you might as well use std::move.
Your function accepting an r-value reference cannot accept l-values so it is not equivalent to the first two functions.