I have a very basic question in C++. How to avoid copy when returning an object ?
Here is an example :
std::vector test(const uns
Referencing it would work.
Void(vector<> &x) {
}
Named Return Value Optimization will do the job for you since the compiler tries to eliminate redundant Copy constructor and Destructor calls while using it.
std::vector<unsigned int> test(const unsigned int n){
std::vector<unsigned int> x;
return x;
}
...
std::vector<unsigned int> y;
y = test(10);
with return value optimization:
(in case you want to try it yourself for deeper understanding, look at this example of mine)
or even better, just like Matthieu M. pointed out, if you call test
within the same line where y
is declared, you can also avoid construction of redundant object and redundant assignment as well (x
will be constructed within memory where y
will be stored):
std::vector<unsigned int> y = test(10);
check his answer for better understanding of that situation (you will also find out that this kind of optimization can not always be applied).
OR you could modify your code to pass the reference of vector to your function, which would be semantically more correct while avoiding copying:
void test(std::vector<unsigned int>& x){
// use x.size() instead of n
// do something with x...
}
...
std::vector<unsigned int> y;
test(y);
There seems to be some confusion as to how the RVO (Return Value Optimization) works.
A simple example:
#include <iostream>
struct A {
int a;
int b;
int c;
int d;
};
A create(int i) {
A a = {i, i+1, i+2, i+3 };
std::cout << &a << "\n";
return a;
}
int main(int argc, char*[]) {
A a = create(argc);
std::cout << &a << "\n";
}
And its output at ideone:
0xbf928684
0xbf928684
Surprising ?
Actually, that is the effect of RVO: the object to be returned is constructed directly in place in the caller.
How ?
Traditionally, the caller (main
here) will reserve some space on the stack for the return value: the return slot; the callee (create
here) is passed (somehow) the address of the return slot to copy its return value into. The callee then allocate its own space for the local variable in which it builds the result, like for any other local variable, and then copies it into the return slot upon the return
statement.
RVO is triggered when the compiler deduces from the code that the variable can be constructed directly into the return slot with equivalent semantics (the as-if rule).
Note that this is such a common optimization that it is explicitly white-listed by the Standard and the compiler does not have to worry about possible side-effects of the copy (or move) constructor.
When ?
The compiler is most likely to use simple rules, such as:
// 1. works
A unnamed() { return {1, 2, 3, 4}; }
// 2. works
A unique_named() {
A a = {1, 2, 3, 4};
return a;
}
// 3. works
A mixed_unnamed_named(bool b) {
if (b) { return {1, 2, 3, 4}; }
A a = {1, 2, 3, 4};
return a;
}
// 4. does not work
A mixed_named_unnamed(bool b) {
A a = {1, 2, 3, 4};
if (b) { return {4, 3, 2, 1}; }
return a;
}
In the latter case (4), the optimization cannot be applied when A
is returned because the compiler cannot build a
in the return slot, as it may need it for something else (depending on the boolean condition b
).
A simple rule of thumb is thus that:
RVO should be applied if no other candidate for the return slot has been declared prior to the return
statement.
This program can take advantage of named return value optimization (NRVO). See here: http://en.wikipedia.org/wiki/Copy_elision
In C++11 there are move constructors and assignment which are also cheap. You can read a tutorial here: http://thbecker.net/articles/rvalue_references/section_01.html
The move constructor is guaranteed to be used if NRVO does not happen
Therefore, if you return an object with move constructor (such as std::vector
) by value, it is guaranteed not to do a full vector copy, even if the compiler fails to do the optional NRVO optimization.
This is mentioned by two users who appear influential in the C++ specification itself:
Not satisfied by my Appeal to Celebrity?
OK. I can't fully understand the C++ standard, but I can understand the examples it has! ;-)
Quoting the C++17 n4659 standard draft 15.8.3 [class.copy.elision] "Copy/move elision"
3 In the following copy-initialization contexts, a move operation might be used instead of a copy operation:
- (3.1) — If the expression in a return statement (9.6.3) is a (possibly parenthesized) id-expression that names an object with automatic storage duration declared in the body or parameter-declaration-clause of the innermost enclosing function or lambda-expression, or
- (3.2) — if the operand of a throw-expression (8.17) is the name of a non-volatile automatic object (other than a function or catch-clause parameter) whose scope does not extend beyond the end of the innermost enclosing try-block (if there is one),
overload resolution to select the constructor for the copy is first performed as if the object were designated by an rvalue. If the first overload resolution fails or was not performed, or if the type of the first parameter of the selected constructor is not an rvalue reference to the object’s type (possibly cv-qualified), overload resolution is performed again, considering the object as an lvalue. [ Note: This two-stage overload resolution must be performed regardless of whether copy elision will occur. It determines the constructor to be called if elision is not performed, and the selected constructor must be accessible even if the call is elided. — end note ]
4 [ Example:
class Thing { public: Thing(); ~ Thing(); Thing(Thing&&); private: Thing(const Thing&); }; Thing f(bool b) { Thing t; if (b) throw t; // OK: Thing(Thing&&) used (or elided) to throw t return t; // OK: Thing(Thing&&) used (or elided) to return t } Thing t2 = f(false); // OK: no extra copy/move performed, t2 constructed by call to f struct Weird { Weird(); Weird(Weird&); }; Weird g() { Weird w; return w; // OK: first overload resolution fails, second overload resolution selects Weird(Weird&) }
— end example
I don't like the "might be used" wording, but I think the intent is to mean that if either "3.1" or "3.2" hold, then the rvalue return must happen.
This is pretty clear on the code comments for me.
Pass by reference + std::vector.resize(0)
for multiple calls
If you are making multiple calls to test
, I believe that this would be slightly more efficient as it saves a few malloc()
calls + relocation copies when the vector doubles in size:
void test(const unsigned int n, std::vector<int>& x) {
x.resize(0);
x.reserve(n);
for (unsigned int i = 0; i < n; ++i) {
x.push_back(i);
}
}
std::vector<int> x;
test(10, x);
test(20, x);
test(10, x);
given that https://en.cppreference.com/w/cpp/container/vector/resize says:
Vector capacity is never reduced when resizing to smaller size because that would invalidate all iterators, rather than only the ones that would be invalidated by the equivalent sequence of pop_back() calls.
and I don't think compilers are able to optimize the return by value version to prevent the extra mallocs.
On the other hand, this:
so there is a trade-off.
First of all, you could declare your return type to be std::vector & in which case a reference will be returned instead of a copy.
You could also define a pointer, build a pointer inside your method body and then return that pointer (or a copy of that pointer to be correct).
Finally, many C++ compilers may do return value optimization (http://en.wikipedia.org/wiki/Return_value_optimization) eliminating the temporary object in some cases.