[Libmesh-devel] Virtual destructor ~Node()

[Roy S. <roy...@ic...>]

We have virtual destructors for Node and its parent classes
(DofObject, TypeVector, Point, ReferenceCountedObject<Node>, and
ReferenceCounter) now, and the resulting vptr tacks on 8 bytes (4 on
32-bit systems) to each Node in the mesh.  Since there's no longer any
other virtual functions in those classes, the vptr ought to go away if
we make the destructor non-virtual, saving us a few megabytes of
memory on large problems.  So I'm thinking of getting rid of the
virtual destructors.

The problem is then that if anyone tries to delete a Node through a
pointer-to-parent-class (or delete any other ReferenceCountedObject
subclass through a pointer-to-RCO, or delete an Elem through a
pointer-to-DofObject...), it would be undefined behavior.  Of course,
it would still be okay to delete a subclass of Elem, which will still
have its vptr regardless, through a pointer-to-Elem.

I'm happy with the alternative solution of "don't ever delete a Node
through a pointer-to-parent-class, etc.", but if anyone else isn't
happy with that then please speak up and veto me.

(Cc: to libmesh-users in case anyone's depending on that level of OOP
in their application code)


Sadly, although the above discussion is about efficiency, this issue
came up in the context of correctness: when generating MPI datatypes
for C++ classes with vptrs, you need to take the changed offsets into
account.  Expect to see libMesh 0.7.2.1 out soon, with the new
exciting feature of "Parallel::*(Point&) actually works properly in
3D"...
---
Roy