All three all-to-all scatter/gather routines (MPI_Alltoall,
MPI_Alltoallv, and MPI_Alltoallw) now accept MPI_IN_PLACE as a valid
sendbuf argument.

This is an implementation of MPI Forum ticket #31 for MPI-2.2
compliance.  See the following URL for more information:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/31

No reviewer.

No reviewer.

This change should help us keep our usage of context IDs straight.  It
adds a check that runs at MPI_Finalize time that ensures that all
allocated context IDs have been freed.

Reviewed by buntinas@.

Under multi-threaded contention the old context ID allocation functions
would leak context IDs as part of the error checking.  This commit
refactors the allocation interfaces to minimize programmer confusion and
eliminates the leak in the previous implementation.  Context ID
allocation error checking is also somewhat improved and a regression
test for this issue is included.

Reviewed by buntinas@.

Some of the wiki docs erroneously recommended using --with-pm=none when
building for SLURM.  This commit adds it as a synonym for "no" to make
this more user friendly.  Thanks to Ashley Pittman for reporting this
discrepancy.

No reviewer.

This is a similar change to r5258 but for intracommunicators instead.

No reviewer.

untested and a maintenance headache.

Thanks to Kim McMahon at Cray for spotting and contributing a fix for
this bug.

It made some potentially unsafe assumptions about the behavior of
MTestGetIntercomm and it had several unused variables left over from a
previous incarnation of this test.

No reviewer.

Otherwise the tests could end up calling communicator functions on
MPI_COMM_NULL, which is erroneous.

No reviewer.

This corrects a bug that was introduced in r5249.  When the is_low_group
field isn't set deadlocks will typically occur for "full-duplex"
collectives over intercommunicators (such as MPI_Allreduce).

This commit also runs each of the intercomm collective tests with a
higher process count to ensure that more intercommunicator code paths
are exercised.  In this particular case running at higher process counts
is an effective regression test.

No reviewer.

The logic added in r5229 to avoid deadlocks in the intercommunicator
tests was flawed.  It assumed that the intercomm index was updated on
every loop iteration, which wasn't actually true.  So for involving 5 or
less processes that required communicators of size 4 or greater, we
would spin forever.

No reviewer.

releases. Otherwise, a bunch of unnecessary files are being added in
the release tarballs (e.g., autom4te.cache). Thanks to Lucas Nussbaum
(Debian mpich2 maintainer) for catching this.

Running maint/updatefiles on a release tarball would clobber the README
and fedora.spec files because the .vin versions weren't present.

No reviewer.

This commit implements the new MPI_Reduce_local and MPI_Op_commutative
functions added in MPI-2.2.  It also includes C-binding tests for the
new functions.

This commit is still WIP because the following things are needed:
- at least one Fortran test for the functions

No reviewer.

Per MPI-2.2, MPI_Exscan now supports MPI_IN_PLACE when passed in the
sendbuf.  For more history and information see the following URL:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/94

No reviewer.

No reviewer.

Implementation of MPI-2.2 ticket #71.  Delete callbacks are now invoked
in LIFO order.  Once the MPI version number macros are updated to 2.2
the test added in r4346 will check for this behavior.  This change also
includes another test for LIFO ordering because I wrote it before I saw
Bill's original test.

For more information and history on this issue, please see this URL:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/71

The non-testing portions of this change were originally written by Bill Gropp.

This commit does 3 things.

- It adds two new types of intercommunicators to the MTestGetIntercomm
  routine, more thoroughly exercising the MPI implementation's handling
  of intercommunicators.

- It reworks MPI_Comm_create such that intercommunicators that do not
  include both rank 0s from the input communicator do not crash,
  corrupt, or hang the MPI implementation.

- It adds tests to ensure that the implementation is MPI-2.2 compliant
  with respect to MPI_Comm_create's new ability to create multiple
  disjoint intracommunicators. (MPI Forum ticket #66)

No reviewer.

MPI_REQUEST_NULL is now a valid argument to MPI_Request_get_status.
This makes MPI_Request_get_status more consistent with MPI_Wait and
MPI_Test, which always supported this argument.

This is an implementation of MPI Forum ticket #137 for MPI-2.2
compliance.  See the following URL for more information:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/137

No reviewer.

A trivial change for MPI-2.2 compliance.  See the following URL for more
information:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/6

Reviewed by buntinas@.

This is backwards compatible with MPI-2.0 and MPI-2.1.  The _function
names have been added and the _fn names have been deprecated but are
still present.  See the following ticket for more info:

https://svn.mpi-forum.org/trac/mpi-forum-web/ticket/7

Reviewed by buntinas@.

We now understand --do= as well as -do=.  Also we only run autoreconf in
the openpa and hydra dirs when the 'build_configure' or 'makefiles'
steps are enabled.  This way you can iterate on the bindings without
having to wait for both sub-projects to fully autoreconf.

No reviewer.

might use the same signal.

No reviewer.

The 'done' variable that is used to control exit from the communicator
creation loop was not guaranteed to be consistently set.  Since most
iterations of the loop require participation from all processes in
MPI_COMM_WORLD a deadlock was possible if some of the procs escape the
loop before the other procs.  This fix uses an MPI_Allreduce to ensure
that all processes are actually done, not just the local proc.

No reviewer.

No reviewer.

They were very broken prior to this commit and could have potentially
interfered with matching logic in extremely difficult to debug ways.

Unfortunately the fix required shaving yet another bit from the context
ID space, dropping the limit to 2048 regular communicators.

No reviewer.

The various intercommunicator creation functions didn't all agree
whether the (send)context_id or the recvcontext_id was allocated from
the local group's context pool.  This led to double-frees and leaks of
context IDs in some cases, particularly for intercommunicators created
with MPI_Comm_split or MPI_Comm_create.

No reviewer.