[smart_ptr] extra colon on make_shared variadic template instantiation

newer
[multi_index][intel-linux-14.0.0]...

older
Comparing floating point Alberto...

Vicente J. Botet Escriba

17 Sep 2013 17 Sep '13

5:45 p.m.

Hi, while instantiating this template, the Intel compiler finds that there is an extra colon when Args is empty. ::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... ); ^ Is this a bug on the compiler or on Boost.SmartPtr? Best, Vicente template< class T, class Arg1, class... Args > typename boost::detail::sp_if_not_array< T >::type make_shared( Arg1 && arg1, Args && ... args ) { boost::shared_ptr< T > pt( static_cast< T* >( 0 ), BOOST_SP_MSD( T ) ); boost::detail::sp_ms_deleter< T > * pd = static_cast *>( pt._internal_get_untyped_deleter() ); void * pv = pd->address(); ::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... ); pd->set_initialized(); T * pt2 = static_cast< T* >( pv ); boost::detail::sp_enable_shared_from_this( &pt, pt2, pt2 ); return boost::shared_ptr< T >( pt, pt2 ); }

Show replies by date

Vicente J. Botet Escriba

19 Sep 19 Sep

5:36 a.m.

New subject: [smart_ptr] extra colon on make_shared variadic template instantiation

Le 17/09/13 19:45, Vicente J. Botet Escriba a écrit :

...

Hi,

while instantiating this template, the Intel compiler finds that there is an extra colon when Args is empty.

::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... ); ^ Is this a bug on the compiler or on Boost.SmartPtr?

Best, Vicente

template< class T, class Arg1, class... Args > typename boost::detail::sp_if_not_array< T >::type make_shared( Arg1 && arg1, Args && ... args ) { boost::shared_ptr< T > pt( static_cast< T* >( 0 ), BOOST_SP_MSD( T ) );

boost::detail::sp_ms_deleter< T > * pd = static_cast *>( pt._internal_get_untyped_deleter() );

void * pv = pd->address();

::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... );

pd->set_initialized();

T * pt2 = static_cast< T* >( pv );

boost::detail::sp_enable_shared_from_this( &pt, pt2, pt2 ); return boost::shared_ptr< T >( pt, pt2 ); }

Hi, could I commit this patch pc3:smart_ptr viboes$ svn diff Index: make_shared_object.hpp =================================================================== --- make_shared_object.hpp (revision 85776) +++ make_shared_object.hpp (working copy) @@ -207,6 +207,23 @@ // Variadic templates, rvalue reference +template< class T, class... Args > typename boost::detail::sp_if_not_array< T >::type make_shared( Args && ... args ) +{ + boost::shared_ptr< T > pt( static_cast< T* >( 0 ), BOOST_SP_MSD( T ) ); + + boost::detail::sp_ms_deleter< T > * pd = static_cast *>( pt._internal_get_untyped_deleter() ); + + void * pv = pd->address(); + + ::new( pv ) T( boost::detail::sp_forward<Args>( args )... ); + pd->set_initialized(); + + T * pt2 = static_cast< T* >( pv ); + + boost::detail::sp_enable_shared_from_this( &pt, pt2, pt2 ); + return boost::shared_ptr< T >( pt, pt2 ); +} + Vicente

Peter Dimov

7:53 a.m.

New subject: [smart_ptr] extra colon on make_shared variadic template instantiation

Vicente J. Botet Escriba wrote:

...

Hi,

while instantiating this template, the Intel compiler finds that there is an extra colon when Args is empty.

::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... ); ^ Is this a bug on the compiler or on Boost.SmartPtr?

I have no idea where an extra colon could come from, but maybe you mean an extra comma. This should be a compiler bug.

...

Hi,

could I commit this patch ... +template< class T, class... Args > typename boost::detail::sp_if_not_array< T >::type make_shared( Args && ... args )

I'm not sure how this could help. This overload is either ambiguous with the Arg1&& one, or less specialized. Do the make_shared tests pass on all compilers after the patch? The reason that there are two overloads, one without arguments, the other with Arg1&& and Args&&..., is that earlier versions of g++ treated new T(args...) as new T when args... is empty, instead of the correct new T(). There is a test that checks for this bug. If the current crop of supported compilers no longer suffers from it, the two overloads can be replaced by the simple variadic Args... version. But adding the Args... version in addition to the other two makes no sense to me.

Jonathan Wakely

2:22 p.m.

New subject: [smart_ptr] extra colon on make_shared variadic template instantiation

On 19 September 2013 08:53, Peter Dimov wrote:

...

The reason that there are two overloads, one without arguments, the other with Arg1&& and Args&&..., is that earlier versions of g++ treated new T(args...) as new T when args... is empty, instead of the correct new T(). There is a test that checks for this bug. If the current crop of supported compilers no longer suffers from it, the two overloads can be replaced by the simple variadic Args... version.

FWIW that should be fixed in G++ 4.3.0 onwards: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32597

Vicente J. Botet Escriba

5:44 p.m.

New subject: [smart_ptr] extra colon on make_shared variadic template instantiation

Le 19/09/13 09:53, Peter Dimov a écrit :

...

Vicente J. Botet Escriba wrote:

...
Hi,

while instantiating this template, the Intel compiler finds that there is an extra colon when Args is empty.

::new( pv ) T( boost::detail::sp_forward<Arg1>( arg1 ), boost::detail::sp_forward<Args>( args )... ); ^ Is this a bug on the compiler or on Boost.SmartPtr?

I have no idea where an extra colon could come from, but maybe you mean an extra comma. This should be a compiler bug. Yes, an extra comma. Maybe it is an extra bug, but it is worth providing a workaround.

...
Hi,

could I commit this patch ... +template< class T, class... Args > typename boost::detail::sp_if_not_array< T >::type make_shared( Args && ... args )

I'm not sure how this could help. This overload is either ambiguous with the Arg1&& one, or less specialized. Do the make_shared tests pass on all compilers after the patch?

The reason that there are two overloads, one without arguments, the other with Arg1&& and Args&&..., is that earlier versions of g++ treated new T(args...) as new T when args... is empty, instead of the correct new T(). There is a test that checks for this bug. If the current crop of supported compilers no longer suffers from it, the two overloads can be replaced by the simple variadic Args... version. But adding the Args... version in addition to the other two makes no sense to me.

I recognize that it should be a better solution. I have compiled with gcc-4.5 and newer, clang-3.1+3.2 and Intel 12.1.3 Not tested with MSVC :( Best, Vicente

4021

Age (days ago)

4023

Last active (days ago)

List overview

Download

4 comments

3 participants

participants (3)

Jonathan Wakely
Peter Dimov
Vicente J. Botet Escriba