[cfe-dev] swap, noexcept and non-movable types

[Sean Hunt]

On 05/31/11 10:53, Howard Hinnant wrote:
> Consider this C++0x code:
>
> template<class T>
> T&&
> declval() noexcept;
>
> template<class T>
> struct some_trait
> {
>      static const bool value = false;
> };
>
> template<class T>
> void swap(T&  x, T&  y) noexcept(some_trait<T>::value)
> {
>      T tmp(static_cast<T&&>(x));
>      x = static_cast<T&&>(y);
>      y = static_cast<T&&>(tmp);
> }
>
> template<class T, unsigned N>
> struct array
> {
>      T data[N];
>
>      void swap(array&  a) noexcept(noexcept(swap(declval<T&>(), declval<T&>())));
> };
>
> struct DefaultOnly
> {
>      DefaultOnly() = default;
>      DefaultOnly(const DefaultOnly&) = delete;
>      DefaultOnly&  operator=(const DefaultOnly&) = delete;
>      ~DefaultOnly() = default;
> };
>
> int main()
> {
>      array<DefaultOnly, 1>  a, b;
> }
>
> We currently do not compile this code.  The errors are:
>
> test.cpp:14:7: error: call to deleted constructor of 'DefaultOnly'
>      T tmp(static_cast<T&&>(x));
>        ^   ~~~~~~~~~~~~~~~~~~~
> test.cpp:24:43: note: in instantiation of function template specialization 'swap<DefaultOnly>' requested here
>      void swap(array&  a) noexcept(noexcept(swap(declval<T&>(), declval<T&>())));
>                                            ^
> test.cpp:30:5: note: function has been explicitly marked deleted here
>      DefaultOnly(const DefaultOnly&) = delete;
>      ^
> test.cpp:15:7: error: overload resolution selected deleted operator '='
>      x = static_cast<T&&>(y);
>      ~ ^ ~~~~~~~~~~~~~~~~~~~
> test.cpp:31:18: note: candidate function has been explicitly deleted
>      DefaultOnly&  operator=(const DefaultOnly&) = delete;
>                   ^
> 2 errors generated.
>
> I recently learned that a recent gcc 4.7 snapshot /does/ compile this code.
>
> So I'm asking the language lawyers here:  should we compile this code or not?
>
> The underlying question is whether or not swap<DefaultOnly>() should be instantiated.  It is not called except as an unevaluated operand.
>
> Thanks,
> Howard

We definitely shouldn't be instantiating swap here. Looks like a bug. 
I'll poke at it a bit.

Sean