[Dune] Returning Geometries As Objects

[Martin Nolte]

Dear grid implementors,

there is a pitfall when starting to actually return geometries as objects (as 
this is, for example, done in IdentityGrid in the dune-grid-dev-howto: One 
easily forgets that, for the transition period, the Dune::Geometry facade can 
either store a reference or the object. To select the correct behavior, you 
have to specialize the "StoreGeometryReference" facade option (and set v = false).

To simplify the first part of the transition, which you will all have gone 
through by now, the default value for "StoreGeometryReference" is true, i.e., 
the Dune::Geometry facade stores a reference to the geometry it is passed.

Unfortunately, the compiler does not warn about returning a reference to a 
temporary, if you forget to specialize the facade class. This will ususally 
lead to unpredictable behavior. Fortunately, in my experience, the grid tests 
always fail in this situation with an unpredictable error, e.g., local( 
global( x ) ) != x.

I hope, this piece of information helps debugging these realy nasty memory errors.

Best,

Martin

On 02/15/2012 03:03 PM, Carsten Gräser wrote:
> Dear 3rd party grid implementors,
>
> this is a warning about a small pitfall with the
> new geometry behavior in dune-grid:
>
> Even if you want to use (for a transition) the possibility
> to still store references to geometries according to the
> temporary default behavior of the new Geometry interface
> class you have to adapt your grid.
>
> The reason is that until now you had to store a Geometry<GeometryImpl>
> (or maybe a Makeable...<...>) yourself in the Entity. Now you have
> to store only the plain GeometryImpl. Using your old implementation
> you may end up storing a Geometry<GeometryImpl>  which itself stores
> a reference to a no longer existing temporary GeometryImpl
>
> You may want to look at the merge commit r7869 as reference for the
> neccessary changes.
>
> The reason for warning about this seemingly obvious fact is,
> that I just realized that an unmodified grid implementation
> might surprisingly still compile and run with the modified grid
> interface.
>
> Best,
> Carsten
>
> _______________________________________________
> 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