Failing test when run as C++11

[Jonathan Wakely]

On 09/05/14 12:20 +0100, Jonathan Wakely wrote:
>On 09/05/14 13:15 +0200, Daniel Krügler wrote:
>>2014-05-09 12:53 GMT+02:00 Jonathan Wakely <jwakely@redhat.com>:
>>>20_util/specialized_algorithms/uninitialized_copy/808590.cc fails if
>>>run in C++11 mode because is_nothrow_move_constructible<c> is false,
>>>which means vector::resize() copies the elements using non-const
>>>lvalue arguments, which selects the template constructor.
>>>
>>>(In C++03 it copies from *const* lvalue arguments, selecting the copy
>>>constructor.)
>>>
>>>We can either fix it by adding throw() to the copy constructor, or by
>>>adding a move cosntructor (with a comment explaining why):
>>>
>>>#if __cplusplus >= 201103L
>>>  c(c&&) = default;
>>>#endif
>>>
>>>I can't decide which I prefer, does anyone else have a preference?
>>
>>Do you know whether the test case was added to validate a bug fix and
>>if so, which one? I would like to ensure that the change still
>>reflects the original intention for adding it.
>
>See https://bugzilla.redhat.com/show_bug.cgi?id=808590
>
>That was filed against Red Hat's GCC 4.4 and at the time the upstream
>GCC already passed the test. Benjamin added the test to catch any
>regressions. Arguably, the C++11 behaviour (copying from non-const
>lvalue instead of const lvalue) can be considered a regression, but I
>think it's the correct behaviour.

To explain why I think it's correct...

In C++03 the vector::resize() copied objects via (the equivalent of)
allocator<c>::construct(pointer, const c&) which always calls the
copy constructor, c::c(const c&).

In C++11 it uses the allocator_traits equivalent of
allocator<c>::construct(pointer, Args&&...) which can call any
constructor, depending on the value category, cv-qualification and
type of the argument. When is_nothrow_move_constructible<c> is false
(as in the current test code) the argument is an lvalue, so we call
c::c<c>(c&) instead of c::c(const c&).

If is_nothrow_move_constructible<c> was true then the argument would
be forwarded as an rvalue, so the template constructor isn't viable
and we would call c::c(const c&), or c::c(c&&) if we add that.