[dune-functions] SegFault in TypeErasure wrapper

[Simon Praetorius]

See also the code snippet:

https://gitlab.dune-project.org/snippets/56

for an example that shows the segfault.

Am 25.07.20 um 22:02 schrieb Simon Praetorius:
> Dear all,
>
> I'm facing currently a difficult to analyze segfault when implementing 
> a type-erasure type based on the dune-functions wrappers in 
> typeerasure.hh.
>
> It is a simple design, a bit similar to a global-function - 
> local-function pair, so one class with some member functions and a 
> generator function for the other class. Internally some data might be 
> stored sometimes. So, depending on the concrete type, either the 
> polymorphic small object optimization kicks in or not. The design 
> works fine for several concrete types, but in some cases when the 
> class gets a bit too large, I get the segfault in the construction of 
> the second class, the instance of the local-function-like class.
>
> I was looking into this for some time now and found that in debug 
> mode, it works fine and with optimization enabled, I get the segfault. 
> It seems to be related to some copy-elision/return-value optimization 
> in the construction.
>
> Interestingly, a (simplified) re-implementation of the type-erasure 
> framework worked fine without any segfaults. So, it must be related to 
> a small difference in these implementations. Finally, I found, that 
> when removing the one virtual function `target_type()` from the 
> `TypeErasureBase` class it worked fine again. It even works fine, if I 
> just remove the `virtual` from that function, to make it a non-virtual 
> class. I think, the `virtual` is not needed for that function, so a 
> removal could be a quick fix of the problem.
>
> Is this a bug in the implementation? Is there a reason for this 
> behavior? A virtual function changes the structure of the class in 
> memory, an additional vtable-pointer is added, for example. Might this 
> be the underlying problem? Is the small-object optimization only 
> possible for non-polymorphic classes?
>
> If we agree, that the fix above is the right way to go, I could 
> provide a MR, but I'm not sure if this is the underlying source of the 
> error.
>
> Best regards,
> Simon
>
>
> _______________________________________________
> dune-functions mailing list
> dune-functions at lists.dune-project.org
> https://lists.dune-project.org/mailman/listinfo/dune-functions

-- 
Dr. Simon Praetorius
Technische Universität Dresden
Institute of Scientific Computing
phone: +49 351 463-34432
mail: simon.praetorius@​tu-dresden.de