One can push_back
rvalues of a noncopyable-but-movable type into a vector of that type:
#include <vector>
struct S
{
S(int);
S(S&&);
};
int main()
{
std::vector<S> v;
v.push_back(S(1));
v.push_back(S(2));
v.push_back(S(3));
}
However, when I try to initializer-list-construct the vector with the same rvalues, I get errors about a copy constructor being required:
#include <vector>
struct S
{
S(int);
S(S&&);
};
int main()
{
std::vector<S> v = {S(1), S(2), S(3)};
}
I get the following errors with GCC 4.7:
In file included from include/c++/4.7.0/vector:63:0,
from test.cpp:1:
include/c++/4.7.0/bits/stl_construct.h: In instantiation of 'void std::_Construct(_T1*, _Args&& ...) [with _T1 = S, _Args = {const S&}]':
include/c++/4.7.0/bits/stl_uninitialized.h:77:3: required from 'static _ForwardIterator std::__uninitialized_copy<_TrivialValueTypes>::__uninit_copy(_InputIterator, _InputIterator, _ForwardIterator) [with _InputIterator = const S*, _ForwardIterator = S*, bool _TrivialValueTypes = false]'
include/c++/4.7.0/bits/stl_uninitialized.h:119:41: required from '_ForwardIterator std::uninitialized_copy(_InputIterator, _InputIterator, _ForwardIterator) [with _InputIterator = const S*, _ForwardIterator = S*]'
include/c++/4.7.0/bits/stl_uninitialized.h:260:63: required from '_ForwardIterator std::__uninitialized_copy_a(_InputIterator, _InputIterator, _ForwardIterator, std::allocator<_Tp>&) [with _InputIterator = const S*, _ForwardIterator = S*, _Tp = S]'
include/c++/4.7.0/bits/stl_vector.h:1185:4: required from 'void std::vector<_Tp, _Alloc>::_M_range_initialize(_ForwardIterator, _ForwardIterator, std::forward_iterator_tag) [with _ForwardIterator = const S*, _Tp = S, _Alloc = std::allocator<S>]'
include/c++/4.7.0/bits/stl_vector.h:362:2: required from 'std::vector<_Tp, _Alloc>::vector(std::initializer_list<_Tp>, const allocator_type&) [with _Tp = S, _Alloc = std::allocator<S>, std::vector<_Tp, _Alloc>::allocator_type = std::allocator<S>]'
test.cpp:11:41: required from here
include/c++/4.7.0/bits/stl_construct.h:77:7: error: no matching function for call to 'S::S(const S&)'
include/c++/4.7.0/bits/stl_construct.h:77:7: note: candidates are:
test.cpp:6:5: note: S::S(S&&)
test.cpp:6:5: note: no known conversion for argument 1 from 'const S' to 'S&&'
test.cpp:5:5: note: S::S(int)
test.cpp:5:5: note: no known conversion for argument 1 from 'const S' to 'int'
Should this be allowed? I see no technical obstacles to it being allowed, but I don't have the Standard handy at the moment...
The initializer_list only provides const references and const iterators. There is no way for the vector to move from that.
Maybe this clause from 8.5.4.5 explains it (my emphasis):
So you can only initialize from lists if the objects are copyable.
Update: As Johannes points out, copy-initialization can be realized by both copy and move constructors, so that alone isn't enough to answer the question. Here is, however, an excerpt of the specification of the
initializer_list
class as described in 18.9:Note how there are no non-constant typedefs!
I just tried making an IL constructor which would traverse the initializer list via
std::make_move_iterator
, which failed becauseconst T &
cannot be converted toT&&
.So the answer is: You cannot move from the IL, because the standard says so.
It seems the answer is No per Kerrek SB's answer. But you can achieve something similar by small helper functions using variadic templates:
It appears it might be a compiler issue. This works in g++ 4.5.1 (click for IdeOne online demo)
Conclusion: It was, in the sense that older g++ implementations did not correctly flag an error; initializer lists do not support moving their elements (elements are implicitely copied in the process). Thank to Kerrek SB for quoting the helpful phrase from the standard.
Old proceedings (for the sake of understanding the comments:)
Edit Found out that at least g++ 4.6.1+ seem to have your complaint about this code.
Edit Upon reading the source to
std::initializer_list<T>
I'm starting to get the impression that this is not supported by the library (it looks intentional). Whether the standard actually allows for an initializer list to forward the xvalue-ness of it's elements... I wouldn't be surprised if they stopped there (perfect forwarding is still not easily supported in C++0x I think, and not all initializer parameters would need to have the same (deductable) type.Anyone with more standardese under his belt care to help out? http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2640.pdf