dolfin team mailing list archive

["Garth N. Wells" <gnw20@xxxxxxxxx>]

On 10/02/11 09:19, Anders Logg wrote:
> On Thu, Feb 10, 2011 at 09:11:26AM +0000, Garth N. Wells wrote:
>>
>>
>> On 10/02/11 09:09, Anders Logg wrote:
>>> On Thu, Feb 10, 2011 at 08:48:55AM +0000, Garth N. Wells wrote:
>>>>
>>>>
>>>> On 10/02/11 07:38, Johan Hake wrote:
>>>>> On Wednesday February 9 2011 23:32:55 Anders Logg wrote:
>>>>>> On Wed, Feb 09, 2011 at 04:40:19PM -0800, Johan Hake wrote:
>>>>>>> Nice fix!
>>>>>>>
>>>>>>> I thought we needed to introduce shared_ptr versions of refinements call.
>>>>>>> But I realise this fix only work for SWIG 2.0. As all shared_ptr_foo are
>>>>>>> renamed to foo.
>>>>>>>
>>>>>>> We need to add an extra layer of %rename/%ignore for the SWIG 2.0 case.
>>>>>>>
>>>>>>> I can do this. But again it introduces another layer of complexity you
>>>>>>> mention in your other post.
>>>>>>
>>>>>> That didn't seem to work. The buildbot now says:
>>>>>>
>>>>>> File
>>>>>> "/home/fenicsslave/jhbuildbot/fenics/lib/python2.6/site-packages/dolfin/mes
>>>>>> h/refine.py", line 30, in refine
>>>>>>   return mesh.child_shared_ptr()
>>>>>> TypeError: in method 'HierarchicalMesh_child_shared_ptr', argument 1
>>>>>> of type 'dolfin::Hierarchical< dolfin::Mesh > *'
>>>>>>
>>>>>> What does that mean? mesh.child_shared_ptr() should return a
>>>>>> shared_ptr to an object of class T (in this case Mesh), which it does
>>>>>> in the C++ interface. It should not return a pointer to an object of
>>>>>> class Hierarchical<T>.
>>>>>
>>>>> It is related to the "fix" Garth introduced. He used foo_shared_ptr, but that
>>>>> one does not work for SWIG < 2.0. These methods are now ignored or renamed for
>>>>> all SWIG versions, and everything should be well and fine for both versions of
>>>>> SWIG.
>>>>>
>>>>
>>>> I don't think that this is really a fix. The Python refine should return
>>>> a shared_ptr, otherwise we can have problems with things going out of
>>>> scope. The underlying problem is SWIG not wrapping the Mesh shared_ptrs
>>>> properly.
>>>
>>> But isn't that exactly what's happening now? The C++ shared_ptr_foo()
>>> functions are renamed to foo() so when you call foo() in Python you
>>> get the shared_ptr.
>>>
>>
>> That was my fix, which was reverted. Looks to me like a reference is
>> returned.
> 
> From site-packages/dolfin/mesh/refine.py:
> 
>   return mesh.child()
> 
> From dolfin/swig/shared_ptr_classes.i:
> 
>    boost::shared_ptr<DERIVED_TYPE> child()
>    { return self->child_shared_ptr(); }
> 
> So when mesh.child() is called in refine.py, it really returns
> mesh.child_shared_ptr().
> 

OK.

It' pretty confusing having stuff dotted around in different files.
We'll be able to remove it when we moving to SWIG 2.

Garth

> 
> The point is that whenever we return something in a Python function
> named foo(), we should always return a shared_ptr.
> 
> The Hierarchical/adaptive classes are probably (now) in better shape
> than other parts of DOLFIN wrt correct use of shared_ptr.
> 
> --
> Anders