3

I have this piece of code:

#include <iostream>
#include <vector>

using namespace std;

class Foo{
public:
    Foo() noexcept {cout << "ctor" << endl;}
    Foo(const Foo&) noexcept {cout << "copy ctor" << endl;}
    Foo(Foo&&) noexcept {cout << "move ctor" << endl;}

    Foo& operator=(Foo&&) noexcept {cout << "move assn" << endl; return *this;}
    Foo& operator=(const Foo&) noexcept {cout << "copy assn" << endl; return *this;}

    ~Foo() noexcept {cout << "dtor" << endl;}
};


int main()
{   
    Foo foo;

    vector<Foo> v;
    v.push_back(std::move(foo)); 

    // comment the above 2 lines and replace by
    // vector<Foo> v{std::move(foo)}; 
}

The output is what I expect (compiled with g++ -std=c++11 --no-elide-constructors, same output without the flag)

ctor
move ctor
dtor
dtor

Now instead of using push_back initialize directly the vector v as

vector<Foo> v{std::move(foo)};

I do not understand why I get the outputs:

1) (without --no-elide-constructors)

ctor
move ctor
copy ctor
dtor
dtor
dtor

2) (with --no-elide-constructors)

ctor
move ctor
move ctor
copy ctor
dtor
dtor
dtor
dtor

In the first case, why is the copy ctor invoked? And in the second case, when the compiler does not perform elision, I have absolutely no idea why the move ctor is invoked twice. Any ideas?

vsoftco
  • 55,410
  • 12
  • 139
  • 252

2 Answers2

12
vector<Foo> v{std::move(foo)};

Here you're calling the vector constructor that takes an std::initializer_list. An initializer list only allows const access to its elements, so the vector is going to have to copy each element from the initializer_list to its own storage. That's what causes the call to the copy constructor.

From §8.5.4/5 [dcl.init.list]

An object of type std::initializer_list<E> is constructed from an initializer list as if the implementation allocated a temporary array of N elements of type const E, where N is the number of elements in the initializer list.

See also https://tristanbrindle.com/posts/beware-copies-initializer-list


As for the extra move constructor call with -fno-elide-constructors, this was discussed in another answer a couple of days ago. It seems as though g++ takes a very literal approach to the example implementation of an initializer_list shown in the standard in same section I've quoted above.

The same example, when compiled using clang, doesn't produce the extra move constructor call.

Colonel Panic
  • 132,665
  • 89
  • 401
  • 465
Praetorian
  • 106,671
  • 19
  • 240
  • 328
  • Ohhh, got it... The syntax `{...}` for uniform initialization is quite tricky, and I now realized that `vector` is not an aggregate. And yes, same on my clang, no extra move. Great answer, thanks! – vsoftco Jul 17 '14 at 00:49
  • I agree with T.C.: the fact that gcc produces two move constructor calls seems like a bug. But I suppose as long as this behaviour is only observable when you pass `-fno-elide-constructors`, there's no problem. – Brian Bi Jul 17 '14 at 00:52
  • T.C.'s answer nails it, I agree that clang's behaviour seems correct. The following code should be the same: `Foo arr[1] = { std::move(foo) }; vector v(arr, arr + 1);` . The `initializer_list` should only serve the purpose of holding `arr` and `arr + 1`, not moving/copying the elements. – M.M Jul 17 '14 at 00:55
  • @MattMcNabb I agree, even if gcc'c behavior is not incorrect, it is definitely strange, clang's seems more intuitive. Anyway, as Brian says, as long as it only happens with `-fno-elide-constructors`, it's unlikely to bother anyone. – Praetorian Jul 17 '14 at 00:58
3

The containers try very hard to make sure they remain usable should an exception occur. As part of this, they'll only use std::move internally if your class' move constructor is exception safe. If it is not, (or it can't tell), it will copy just to be safe.

The correct move operations are

Foo(Foo&&) noexcept {cout << "move ctor" << endl;}
Foo& operator=(Foo&&) noexcept {cout << "move assn" << endl;}
Mooing Duck
  • 64,318
  • 19
  • 100
  • 158
  • Thanks, I changed it, now all of the ctors (together with the added `operator=`) are marked as `noexcept`, however I still get exactly the same output (cases 1) and 2) above). I use `g++4.9` btw. – vsoftco Jul 17 '14 at 00:37
  • @vsoftco: Praetorian figured it out, click accept on his answer. – Mooing Duck Jul 17 '14 at 00:42