[dune-functions] More build failures

[Carsten Gräser]

Am 18.09.2015 um 21:30 schrieb Steffen Müthing:
> 
>> Am 18.09.2015 um 21:12 schrieb Carsten Gräser <graeser at mi.fu-Berlin.de>:
>>
>> Am 18.09.2015 um 15:19 schrieb Christian Engwer:
>>> On Fri, Sep 18, 2015 at 01:59:58PM +0200, Carsten Gräser wrote:
>>>> Hi,
>>>>
>>>> [...]
>>>>> Ups, you are right... but the error is still the same
>>>>>      in-class initializer for static data member is not a constant expression
>>>>> 	      static const int value = ((const typename std::decay<T>::type*)(nullptr))->size();
>>>>> 		                           ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
>>>> it's really hard to call a constexpr method if you don't
>>>> have an object. After digging some more time it seems that
>>>> the only possible way is to call a constructor inside of
>>>> decltype. Unfortunately the constructor arguments are not
>>>> known. Maybe it's OK to restrict this to default-constructible
>>>> types. I'll have look.
>>>
>>> After discussing with Jö, we agreed that it is necessary to have
>>> default constructible types. casting from a null ptr and then calling
>>> a member is undefined behaviour. On the other hand we can safely
>>> assume that all types we want to pass are either std::array or
>>> somethingin our scopte. Thus we can always make sure that they are
>>> default constructible.
>>>
>>> With this assumption the code got a lot simpler and I pushed a fix
>>> which works with clang and g++.
>>
>> I also discussed this with Steffen. The main problem is, that
>> calling a constexpr function is only a constexpr if 'this' is
>> also a constexpr which is never true for the casted nullptr.
>> Default constructability is only needed because one has to settle
>> on one constructor to call and the default constructor is the
>> only viable solution.
>>
>> But this will also only solve the problem if the default
>> constructor is constexpr. Hence your commit only partly
>> fixes the problem because the HierarchicVectorWrapper
>> is not used in the posisson example (which is not necessary
>> but should also work). I just push a follow up fix allowing
>> to also use FlatMultiIndex in through the wrapper.
> 
> Another point Carsten and I discussed was the annoying fact that
> the size method of std::array<> isn’t constexpr, so you can’t use this
> for arrays…
No Steffen, this is not true. The current solution is working.
The method is constexpr but not static. So we need a constexpr
object which introduces the requirement to have a constexpr default
constructor.

Carsten




> 
> Steffen


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 473 bytes
Desc: OpenPGP digital signature
URL: <https://lists.dune-project.org/pipermail/dune-functions/attachments/20150918/cce08916/attachment.sig>