[Dune] dune-grid does not compile when implementations of EntityPointer and Iterator are not interoperable

[Martin Nolte]

Hi Carsten,

I have no objections to removing it (it is deprecated and I never got the 
warning even before the 2.1 release).

Best,

Martin

On 04/27/2012 02:33 PM, Carsten Gräser wrote:
> Hi,
> Am 27.04.2012 13:00, schrieb Martin Nolte:
>> Hi Steffen,
>>
>> you are right, the cast operator is indeed deprecated in the 2.1 release
>> branch. This would mean it should be gone by now. Unfortunately, we are
>> close to a release and I would ask the release managers whether such a
>> change would still be in order. Is anybody still using the deprecated
>> cast operator?
> in my opinion we should really remove it. And I don't expect any
> problems in the core modules.
>
>> For the casting problem, you might want to have a look at GeometryGrid's
>> entity pointer (see dune/grid/geometrygrid/entitypointer.hh). Here, the
>> entity pointer stores the host iterator (which could, e.g., be a leaf
>> iterator). You can then derive your iterator from the entity pointer
>> (with the correct iterator passed as template argument) and, thus,
>> satisfy the interoperability desired by Dune::EntityPointer. Of course,
>> this trick has some consequences:
>> (a) There is no guarantee that the code will work
>>      (but that holds for the cast in Dune::EntityPointer as well).
>> (b) The last member of your entity pointer must be the host iterator
>>      (which is the one actually casted by the reinterpret_cast).
>> Apart from that, the code seems to work wonderfully. I have to admit,
>> though, that I never use the deprecated cast operator in my code.
>>
>> Best,
>>
>> Martin
>>
>> On 04/27/2012 12:08 PM, Steffen Müthing wrote:
>>> Hi Dune,
>>>
>>> I just noticed that it is not possible to have non-interoperable
>>> implementations of EntityPointer and any
>>> of the iterators - doing so will make the compiler choke on the
>>> deprecated EntityPointer reference cast.
>>>
>>> I stumbled over the problem when I forcibly disabled the
>>> interoperability (by specializing EnableIfInteroperable)
>>> to debug an issue in MultiDomainGrid. I was able to make fix the issue
>>> (see attached patch), but from the wording
>>> in entitypointer.hh, I really don't know whether or not the
>>> compilation is expected to fail for non-interoperable
>>> implementations (as some kind of compile-time safeguard).
> I suppose it was intended to be a compile-time safeguard.
> However, it does not really provide any safety because
> the interoperability check works fine if you can convert
> in any direction. IMHO one would need to use enable_if with
>
>    IsBaseOf<EntityPointerImp,IteratorImp>
>
> instead of IsInteroperable to check if the reinterpret_cast
> could do something meaningful (at least if you disregard the
> standard). Even then one would need to apply your trick to
> make the code compile if the latter is not true.
>
>>>
>>> That said, being able to disable the possibility to do the reference
>>> cast would really simplify the MultiDomainGrid
>>> code quite a bit in a number of places - right now I have the problem
>>> that the cast matches rather greedily and I have to
>>> provide pretty elaborate template specializations of constructors and
>>> other stuff to work around that.
>>>
>>> So, is this intended behavior or should we fix it? Or do we throw out
>>> the reference cast for the new release (it was deprecated
>>> in 2.1)?
> It's not intended but I suggest to remove the cast anyway. If there are
> any objections I'd fix the problem as you suggested and be more strict
> for the cast to be enabled (see attached patch).
>
> Best,
> Carsten
>
>>>
>>> Best,
>>>
>>> Steffen
>>>
>>>
>>> PS: In order to fix the problem, I had to do some trickery with the
>>> cast operator - it can't be disabled using SFINAE, because the return
>>> type *is* the function name. So I did a less than pretty hack and just
>>> made it convert to void if the cast is disabled (a marker type like
>>> InvalidCast or something might be nicer).
>>>
>>>
>>> _______________________________________________
>>> Dune mailing list
>>> Dune at dune-project.org
>>> http://lists.dune-project.org/mailman/listinfo/dune

-- 
Dr. Martin Nolte <nolte at mathematik.uni-freiburg.de>

Universität Freiburg                                   phone: +49-761-203-5630
Abteilung für angewandte Mathematik                    fax:   +49-761-203-5632
Hermann-Herder-Straße 10
79104 Freiburg, Germany