--enable-sparse-groups causes assert failures in intel and ibm tests

[jsquyres]

v2.x results https://mtt.open-mpi.org/index.php?do_redir=2306 shows that --enable-sparse-groups is returning in assertion failures in:

• Intel MPI_Comm_split3_c
• Intel MPI_Comm_split3_f
• Intel MPI_Comm_split4_c
• Intel MPI_Comm_split4_f

And v2.x results https://mtt.open-mpi.org/index.php?do_redir=2309 shows an --enable-sparse-groups fail in

• IBM group/compare
• Similar failures occurred in IBM communicator/create_group_* (Fortran) in different MTT results

There's a bunch of similar failures in master results: https://mtt.open-mpi.org/index.php?do_redir=2310

@edgargabriel Can you have a look?

[edgargabriel]

Honestly, I suspect that these failures have to do something with the addproc changes, since the sparse group code has not changes since ~2008

[rhc54]

Is this on the master, or 2.x? I thought we had fixed such things on the master, and so this might just be something that hasn't come across.

[jsquyres]

This is on v2.x. But I see the same failure signature on master.

[jsquyres]

Looks like there's more failures of the same type on master (i.e., more fails on master than there are on v2.x). I added an MTT link for master fails to the initial description.

[edgargabriel]

I can have a look, but it will take me a couple of days to get to it

[jsquyres]

@edgargabriel Many thanks. There's no immediate rush, but given all the progress we've made in terms of scalability for the 2.x series, having sparse groups continue to work is probably a Good Thing. Doesn't need to be v2.0.0, though.

[rhc54]

FWIW: I believe @ggouaillardet was looking at this and (IIRC) had a patch for it that is pending for master

[jsquyres]

Also FWIW: I don't see an obvious PR pending about sparse groups. Shrug. @hjelmn thinks it may be a missing OBJ_RETAIN or OBJ_RELEASE.

[hjelmn]

I couldn't find an existing PR so I opened #1595 to fix this. Tested and the failure is gone and I am not seeing any proc leaks.