In C++, copy-swap idiom is typically implemented like this:
C& operator=(C rhs)
{
swap(*this, rhs);
return *this;
}
Now, if I want to add a move-assignment operator, it is supposed to look like this:
C& operator=(C&& rhs)
{
swap(*this, rhs);
return *this;
}
However, this creates ambiguity about which assignment operator should be called and compilers rightfully complain about it. So my question is the following: If I want to support copy-swap idiom together with move-assignment semantics what am I supposed to do?
Or is this a non-issue as having a move-copy constructor and copy-swap idiom, one doesn't really benefit from having a move-assignment operator?
After asking this question, I've written a code that demonstrates that move-assignment may result in fewer function calls than copy-swap idiom. First let me present my copy-swap version. Please bear with me; it appears like a long but a simple example:
#include <algorithm>
#include <iostream>
#include <new>
using namespace std;
bool printOutput = false;
void* operator new(std::size_t sz)
{
if (printOutput)
{
cout << "sz = " << sz << endl;
}
return std::malloc(sz);
}
class C
{
int* data;
public:
C() : data(nullptr)
{
if (printOutput)
{
cout << "C() called" << endl;
}
}
C(int data) : data(new int)
{
if (printOutput)
{
cout << "C(data) called" << endl;
}
*(this->data) = data;
}
C(const C& rhs) : data(new int)
{
if (printOutput)
{
cout << "C(&rhs) called" << endl;
}
*data = *(rhs.data);
}
C(C&& rhs) : C()
{
if (printOutput)
{
cout << "C(&&rhs) called" << endl;
}
swap(*this, rhs);
}
C& operator=(C rhs)
{
if (printOutput)
{
cout << "operator= called" << endl;
}
swap(*this, rhs);
return *this;
}
C operator+(const C& rhs)
{
C result(*data + *(rhs.data));
return result;
}
friend void swap(C& lhs, C& rhs);
~C()
{
delete data;
}
};
void swap(C& lhs, C& rhs)
{
std::swap(lhs.data, rhs.data);
}
int main()
{
C c1(7);
C c2;
printOutput = true;
c2 = c1 + c1;
return 0;
}
I have compiled this with g++ using the -fno-elide-constructors option as I want to see the no optimization behavior. The result is the following:
sz = 4
C(data) called // (due to the declaration of result)
C() called // (called from the rvalue copy-constructor)
C(&&rhs) called // (called due to copy to return temporary)
C() called // (called from the rvalue copy-constructor)
C(&&rhs) called // (called due to pass-by-value in the assignment operator)
operator= called
Now, if I choose not to make copy-swap idiom in the assignment operator, I will have something like this:
C& operator=(const C& rhs)
{
if (printOutput)
{
cout << "operator=(const C&) called" << endl;
}
if (this != &rhs)
{
delete data;
data = new int;
*data = *(rhs.data);
}
return *this;
}
This allows me to have the move-assignment operator as follows:
C& operator=(C&& rhs)
{
if (printOutput)
{
cout << "operator=(C&&) called" << endl;
}
swap(*this, rhs);
return *this;
}
Now, with everything else being the same, I get the following output:
sz = 4
C(data) called // (due to the declaration of result)
C() called // (called from the rvalue copy-constructor)
C(&&rhs) called // (called due to copy to return temporary)
operator=(C&&) called // (move-assignment)
As you can see this results in fewer function calls. Actually the last three function calls in the copySwapIdiom has now dropped down to a single function call. This is expected as we no longer pass the assignment operator parameter by value, hence no construction happens there.
However, I do not benefit from the beauty of copy-swap idiom in the assignment operator. Any insight is much appreciated.
There actually isn't a need to implement the move assignment operator if you provide a valid move constructor.
class Foo
{
public:
explicit Foo(Bar bar)
: bar(bar)
{ }
Foo(const Foo& other)
: bar(other.bar)
{ }
Foo(Foo&& other)
: bar(other.bar)
{ }
// other will be initialized using the move constructor if the actual
// argument in the assignment statement is an rvalue
Foo& operator=(Foo other)
{
std::swap(bar, other.bar);
return *this;
}
The motivation behind the copy-swap idiom here is to forward the copy/move work to constructors, so that you don't duplicate work for both constructors and assignment operators. That said,
C& operator=(C rhs) noexcept;
means to replace the pair
C& operator=(const C& rhs);
C& operator=(C&& rhs) noexcept;
Whether C& operator=(C rhs) noexcept; performs copy or move assignment depends on how rhs is constructed. For example,
a = std::move(b); // rhs is move-constructed from r-value std::move(b), and thus move-assignment
c = d; // rhs is copy-constructed from l-value d, and thus copy-assignment
Related
I'm struggling to understand the exact reason the program fails.
Regarding the following program:
#include <iostream>
template < class T, size_t SIZE>
class Stack {
T arr[SIZE] = {};
int pos = 0;
public:
Stack & push(const T & t) {
arr[pos++] = t;
return *this;
}
Stack & push(T && t) {
arr[pos++] = std::move(t);
return *this;
}
T pop() {
return std::move(arr[--pos]);
}
};
class Foo {
std::string s;
public:
Foo(const char* s = "#") : s(s) {}
Foo(Foo && foo) : s(std::move(foo.s)) {}
Foo & operator=(Foo && foo) {
s = std::move(foo.s);
return *this;
}
void print() const { std::cout << s << std::endl; }
};
int main() {
Stack<std::string, 5> s1;
s1.push("hi").push(std::string{ "bye" });
std::cout << s1.pop() << std::endl;
Stack<Foo, 5> s3;
Foo f2;
s3.push(f2);
}
The program fails at s3.push(f2); is it because Foo doesn't have a copy ctor or is it because its assignment operator function only handles Foo&& types?
I'm suspecting it's the assignment operator function but I'm not sure it's not because of copy ctor as well.
Because you provided a custom move cosntructor and assignment operator, the compiler no longer generates default copy constructor and assignment operator.
You either need to write those too, or, even better, remove the custom move operations.
The compiler will then generate all 4 for you, and they might be better than your manually written ones (e.g. one problem is that you forgot noexcept on move operations, so standard containers will prefer making copies to moves in some scenarios).
Well, the compiler tells you:
error: use of deleted function ‘Foo& Foo::operator=(const Foo&)’
So you can solve it by adding that function:
Foo& operator=(const Foo& foo) {
s = foo.s;
return *this;
}
There are 2 solutions (I will not point out design issues assuming you are just experimenting).
As a internal data use vector<T> instead of the array of fixed size and in the push that accepts the const& move and use emplace_back (nasty and amoral, also it will leave the src in invalid state).
Provide copy constructor to Foo (this is much more easier)
I have the following code:
#include <bits/stdc++.h>
using namespace std;
class A {
public:
A(const A& a) noexcept { cout << "copy constructor" << endl; }
A& operator=(const A& a) noexcept { cout << "copy assignment operator" << endl; }
A(A&& a) noexcept { cout << "move constructor" << endl; }
A& operator=(A&& a) noexcept { cout << "move assignment operator" << endl; }
A() { cout << "default constructor" << endl; }
};
vector<A> aList;
void AddData(const A&& a)
{
aList.push_back(std::move(a));
}
int main()
{
AddData(A());
return 0;
}
The output is default constructor copy constructor. please tell me is the rvalue reference push_back(T&&)called? And when is copy constructor called?
The issue is with the a parameter in AddData():
void AddData(const A&& a) // <-- const reference!!!
{
aList.push_back(std::move(a)); // selects push_back(const A&)
}
The a parameter above is a const rvalue reference. You are marking with std::move() a const object.
Marking a const object with std::move() for moving has no effect when it comes to move semantics because you can't move from a const object (i.e., you need to alter the moved-from object, but it is const-qualified).
An rvalue reference doesn't bind to a const object, but a const lvalue reference does. As a result, the push_back(const A&) overload is selected instead of the push_back(A&&) one, and therefore the A object is copy constructed.
Solution
Use a non-const rvalue reference instead:
void AddData(A&& a) // <-- non-const reference
{
aList.push_back(std::move(a)); // selects push_back(A&&)
}
Let's say I have the following minimal example class:
#include <iostream>
class Foo {
public:
Foo() = default;
Foo(const Foo&) = default;
Foo(Foo&&) noexcept = default;
Foo& operator=(const Foo& rhs) {
std::cout << "copy\n";
return *this;
}
Foo& operator=(Foo&& rhs) noexcept {
std::cout << "move\n";
return *this;
}
Foo operator+(const Foo& rhs) const {
Foo x; // with some calculation
return x;
}
};
int main() {
Foo a, b, c;
a = b + c;
}
This prints move as expected. Now according to Effective C++ Item 3, I should return const Foo from operator+ to avoid construct like a + b = c, i.e.:
// To avoid a + b = c
const Foo operator+(const Foo& rhs) const {}
Unfortunately, this suddenly starts calling copy assignment instead of move assignment operator. [I'm using gcc 4.8.4 on Ubuntu, but it is probably nothing related to compiler]
How can I ensure that a + b = c fails to compile and in the same time move assignment is called for a = b + c? Or with the introduction of move semantics, is there no way to achieve both of them in the same time?
I have ended up using lvalue reference qualifier as pointed by Caninonos in comment and by max66 in now deleted answer (but 10k users can see it).
Foo& operator=(const Foo& rhs) & {}
Foo& operator=(Foo&& rhs) & noexcept {}
It is simple to implement and it provides a better interface design since assignment to anything other that lvalue doesn't sound meaningful and a possible source of bug.
However, it should be noted that the possibility of writing a + b = c by mistake is very low. Also compiler generated assignment operators are not lvalue reference qualified and we can write a + b = c with standard types, e.g. with std::string or with std::complex.
This is my code:
#include <iostream>
using namespace std;
class A
{
int i;
public:
A(int v) : i(v) { }
A(const A& r) : i(r.i) {
cout << "Copy constructor" << endl;
}
A operator=(const A& r) {
cout << "Assignment function" << endl;
return r;
}
void show() {
cout << i << endl;
}
};
int main()
{
A a(1);
A b(2);
a = b;
a.show();
return 0;
}
Value of b is 2 and value of a is 1. In 'main', b is copied into a and this the output I get:
Assignment function
Copy constructor
This is understandable, but the output for a.show() comes out be 1. I can't understand this. How? Because b is copied into a using the copy constructor so shouldn't a.i have the value of b.i?
b is copied into a using the assignment operator, not the copy constructor. And your assignment operator doesn't assign i, so a.i keeps its original value.
The copy constructor you are seeing is for the return value of operator=. It is more customary to return the left-hand-side by reference, not the right-hand-side by value:
A& operator=(const A& r) {
cout << "Assignment function" << endl;
i = r.i;
return *this;
}
When you define the assignment operator you have to do all of the work to copy the data.
A operator=(const A& r) {
cout << "Assignment function" << endl;
i = r.i;
return r;
}
so a = b is calling your assignment operator (obviously, hence your output). But right now it isn't copying anything.
The assignment operator doesn't do anything to the "assigned to" object. The copy constructor call you see reported is from the creation of the assignment return value.
Your copy assignment should probably look loke this:
A& A::operator= (A const& other) {
// output
this->i = other.i;
return *this;
}
I have a user-defined class (tree structure) with implemented move semantics, and a swap function. I would like to implement a move function the proper way, working as standard std::move implementation.
In the tree node class, each child node has a parent pointer, pointing to the parent node. This means that for move operations, all children have to reparented (and there may be many children)
This means that use of swap for moving is not optimal, as children of both lists have to be reparented after being swapped. So I would like to implement a move function which clears the moved-from tree.
The declaration of std::move implementations is somewhat complicated, they use a std::remove_reference<T>::type&& return type. Do I need this?
You don't need to write a specialisation of std::move.
If you write a correct move constructor and move assignment operator, std::move will work on your class.
example:
#include <iostream>
#include <cstring>
using namespace std;
struct Thing {
Thing()
: _data(new int[100])
{
cout << "default construct\n";
}
// Copy operator
Thing(const Thing& other)
: _data(new int[100])
{
cout << "copy constructor\n";
memcpy(_data, other._data, sizeof(int) * 100);
}
// Move constructor
Thing(Thing&& other) noexcept
: _data(other._data)
{
cout << "move constructor\n";
other._data = nullptr;
}
// assignment operator
Thing& operator=(const Thing& rhs) {
cout << "copy operator\n";
if (&rhs != this) {
Thing tmp(rhs);
std::swap(*this, tmp);
}
return *this;
}
// move assignment operator
Thing& operator=(Thing&& rhs) noexcept {
cout << "move operator\n";
std::swap(_data, rhs._data);
return *this;
}
// destructor necessary since we are working in dangerous new/delete territory
~Thing() noexcept {
cout << "destructor " << (_data ? "object has data" : "object is empty") << "\n";
delete[] _data;
}
private:
int* _data;
};
int main()
{
cout << "constructing a\n";
Thing a;
cout << "constructing b with copy of a\n";
Thing b(a);
cout << "moving a to newly constructed c\n";
Thing c(std::move(a));
cout << "moving c back to a\n";
a = std::move(c);
cout << "create a new d\n";
Thing d;
cout << "replace d with a copy of a\n";
d = a;
return 0;
}
Program's output:
constructing a
default construct
constructing b with copy of a
copy constructor
moving a to newly constructed c
move constructor
moving c back to a
move operator
create a new d
default construct
replace d with a copy of a
copy operator
copy constructor
move constructor
move operator
move operator
destructor object is empty
destructor object has data
destructor object has data
destructor object is empty
destructor object has data
destructor object has data
To write move semantics is to write move constructor/move assignment, but not to implement move function. As result, you maybe move the root of the another tree in move ctor/assignment.
class binary_tree_node {};
class binary_tree
{
public:
binary_tree() : root(nullptr) {}
binary_tree(binary_tree &&rhs)
: root(rhs.root)
{
rhs.root = nullptr;
}
binary_tree& operator=(binary_tree rhs)
{
swap(rhs);
return *this;
}
void swap(binary_tree &rhs)
{
std::swap(root, rhs.root);
}
private:
binary_tree_node *root;
};
int main()
{
binary_tree tree1;
binary_tree tree2 = std::move(tree1);
return 0;
}