C++ won't implicitly cast pointers to/from (void *) in the same way that
C will.  Since MPI_UNWEIGHTED is only ever passed to functions expecting
an int pointer, (int *) is a more appropriate type for mixed C/C++
usage.

Thanks to Lisandro Dalcín for reporting this bug.

No reviewer.

This isn't a proper fix, per se, but it does make the specific problem
type reported in ticket #972 work.  A sufficiently deep type will still
make the implementation blow up.

No reviewer.

This makes the crash when we overflow the dataloop stack depth
(as described in ticket #972) prettier.  We don't segfault anymore, we
just get an assertion failure.

No reviewer.

MPID_Dataloop_create still has void return type, so error handling is a
bit broken in general.

No reviewer.

A deeply nested type that has structures and vectors in alternation with
a deep dataloop depth.

No reviewer.

No reviewer.

No reviewer.

MAC/OS. We work around this by checking if stdin is valid and only
forward it on valid platforms.

for a close event on it. Also, while shutting down, the bootstrap
servers now don't wait for the stdin socket to be closed.

still not automated, but is useful for debugging.

the spawned processes, though.

[svn-r5985] Use standard environment variables for LIBTOOLIZE and AUTORECONF as well as the MPICH2_xxx versions

least one nasty bug remaining, but committing this part in so the
nightly tests can make sure no errors got introduced in the
non-dynamic process case in the meanwhile.

[svn-r5983] Fixed the debugger message queue interface.  There is more to do, but this brings the interface up-to-date with the organization of the MPIDI_Request in ch3, including the message match, and includes thread-cs (type handle) around the code to remember send requests, as well as the installation of a finalize handler to free the pool of structures use to remember send requests.  Still needed: a way to request that that device maintain a queue of send requests, rather than the after-the-fact method currently in use.

[svn-r5981] Moved the release of the request after the sendq_forget for the case where the application in multithreaded - otherwise, the sendq could forget the wrong request

[svn-r5980] Added fixme about the placement ot he MPIR_WaitForDebugger call in the init routine.  This does not matter for the classic process acquistion approach (where the debugger examines the mpiexec process for the proctable) but is relevant when examining the MPI processes

[svn-r5979] Fixed minor typo in message and added a check for the weak,alias attribute (a gcc features)

[svn-r5977] Prepare for fix to the debugging interface by providing a name for the message_match_parts structure

[svn-r5976] Update pm/util and pm/gforker to support Totalview process acquisition interface (run totalview directly on the mpiexec process).  This will work with an pm that uses the pm/util routines and invokes the MPIE_InitForDebugger routine.  The changes to util eliminate the need to link with libmpich.a at the cost of making copies of a few files and eliminating use of the trmem routines

slurm) for Esteban to hack on. It's not configured by default.

user-specified timeout with 1000 to convert to milliseconds. For
infinite timeout (-1), this will give a value of -1000, which is not
portable. Thanks to Esteban Meneses for reporting the problem on AIX.

PMI_KVS_Put(). This case can occur, for example, during the business
card publish phase in nemesis when no netmods are configured. Thanks
to Guillaume for reporting this error.

information to the UI and move generic functionality to utility
functions. Fixed a bug in the MPICH_PORT_RANGE code along the way.

rebuilds. Fixes ticket #930.

Reviewed by goodell.

- an api change: zoidfs routines take hints now
- bug fixes with listio parameters

thanks!

[svn-r5953] Merging r5570, r5577, r5579, r5580, r5581, r5616 from wintcp_async_progress branch - Initial cut of integrating EX with nemesis