For the deal.II wrappers of the N_Vector interface, we've had several lengthy discussions about how to deal with the ops.nvgetcommunicator callback. The basic problem is that that interface requires one to write a function that returns a pointer to a communicator -- and that requires that that communicator has a memory location where it can be accessed.

But that's not always the case. Imagine a case where we're trying to wrap PETSc vectors. Then one might be tempted to write a function of this sort:

void * GetCommunicator (NVector v)
{
   MPI_Comm comm = PetscObjectComm(reinterpret_cast<PetscObject>(unpack 'v' to get at PETSc vector));
   return &comm;
}

...
object.ops.nvgetcommunicator = &GetCommunicator;

Except, of course, this doesn't work: We're now returning a pointer to a local variable. In other words, short of finding out where exactly PETSc happens to store the communicator, we don't quite know what address to return.

There needs to be a better design than this. MPI_Comm objects are designed to be passed around by value, and asking for a void * to a specific memory location just doesn't work without an unnecessarily large amount of work.

@sebproell @stefanozampini FYI.
A pointer to one of the discussions we're having: dealii/dealii#15086 (comment)