Exception on destruction of communicator mpi_comm_null

[aelaguiz@…]

Using boost trunk mpi the communicator destructor does not check for the possibility of an MPI_COMM_NULL underlying communicator, code which produces a null underlying can exception on ref counted destruction.

For instance, this code:

void pool(bmpi::communicator &world, bmpi::communicator &local) {

LDEBUG(PSTR("Pool - local: "), pantheios::integer(local.rank()), PSTR(" of "), pantheios::integer(local.size()));

bmpi::group localGroup = local.group();

dump_group(localGroup);

std::vector<int> inc(1, 0); bmpi::group tinyGroup = localGroup.include(inc.begin(), inc.end()); dump_group(tinyGroup);

bmpi::communicator tinyCom(local, tinyGroup);

}

When tinyCom goes out of scope, this exception gets thrown:

"MPI_Comm_free: MPI_ERR_COMM: invalid communicator"

[aelaguiz@…]

Proposed patch to check for MPI_COMM_NULL on destruction

[aelaguiz@…]

better fix

[anonymous]

Attached a better fix, the real problem is that the subgroup constructor allows direct assignment to the underlying of MPI_COMM_NULL without a check, modified this to call the normal constructor with a take ownership flag

[Matthias Troyer]

(In [84739]) Fixed #6436 #5596 and added threaded initialization