Otherwise they will get freed again by the MPIR_Comm_free that is called
by CH3's MPID_Finalize.

No reviewer.

No reviewer.

This optimization can greatly improve multi-threaded message rates for
MPI_COMM_WORLD, particularly on platforms where a fetch-and-inc is
potentially expensive (like BG/P).

No reviewer.

There were several layers of brokenness that prevented lock-free
reference counting from occurring. Some of this code has been brought
over from the threads development branch, but most of it is new.  This
code still contains many questionable uses of the reference count
variables that need to be investigated and fixed (search for FIXME-MT).

No reviewer.

No reviewer.

At some point we need to define MPIDU_Busy_wait and MPIDU_Yield
operations that are similar to what's available in the process locks
code but without actually using that horrible mess itself.

No reviewer.

No reviewer.

I'm deleting them because they are too hard to get working early in the
include sequence.  Ideally we would be able to re-enable them at some
point after some rationalization of our header file ordering.

No reviewer.

This is necessary to permit non-ch3 devices (such as dcmfd) to override
the implementation of the critical section macro in mpidpre.h.  For
example, dcmfd may want to use a lockbox lock instead of a pthread
mutex when running on a BG/P.

No reviewer.

It provides little to no performance gains over GLOBAL while requiring
plenty of additional testing and maintenance.  Deleting it should
simplify working with the fine-grained threading code and reduce the
overall maintenance burden.

No reviewer.

Prior to this commit the constants used to indicate the thread
granularity level were defined after mpidpre.h was included in mpiimpl.h
This prevented the device from making conditional compilation decisions
in mpidpre.h based on the thread granularity level selected by
configure.

No reviewer.

No reviewer.

tarball. Update the Hydra configure.in to call hwloc's configure as a
sub-configure, rather than directly calling the configure executable.

[svn-r5488] Move the include of mpif.h so that the Fortran 90 versions that use the mpi module will have the module in the correct place in the file

which caused the RMK to be left behind in the previous cleanup.

[svn-r5484] Added check for nbytes_ > 0 to the memory macros to allow the use of 0 length memory requests; these macros will then signal an error only if both the pointer returned by malloc is null and the length is greater than zero

not selected.

Build hwloc when its selected, instead of just configuring it.

(aclocal.m4 should only contain includes to other m4 files). The
problem is basically a small behavioral difference between automake
and simplemake.

Simplemake requires the special file "aclocal.m4" to be present in the
included path (confdb/aclocal.m4). The confdb/aclocal.m4 file needs to
explicitly include all other m4 files as needed.

Automake creates its own aclocal.m4 if its not present, and
automatically includes all m4 files present in the included path
(confdb/*.m4). However, if the included path has another aclocal.m4,
it ignores it.

This is mostly fine, except that confdb/aclocal.m4 contains some
macros which are also being ignored by automake (since it ignores the
entire file), but not by simplemake. This patch just moves them to a
different file. This way, simplemake can continue to use
confdb/aclocal.m4 and automake can continue to ignore it, but things
will still work.

Reviewed by goodell.

alternative to plpa. The following parts still need to be done:

1. Check when hwloc is usable and when it is not. Ideally the hwloc
configure should deal with this, and we should be able to always
enable it.

2. Initialize appropriate structures with hwloc. We might need to
modify the underlying hydra binding infrastructure a bit, since hwloc
is more powerful than plpa (assuming we want to use the extra
features).

3. This commit directly drops in the released version of hwloc, which
contains a bunch of autogenerated files. The problem with using an svn
version is that hwloc uses libtool, while we don't. We need to either
edit the released version of hwloc to manually remove parts of it, or
start using libtool.

[svn-r5478] Correct the datatype name tests for both C (make sure that all required types are before the optional types) and Fortran (test was broken because of implicit typing in Fortran)

[svn-r5476] Printing out the context name and event retval instead of a context id/number for sock events

maintenance cleanup. Tools functionality is all HYDT_, utilities are
all HYDU_ and the rest are HYD_.

[svn-r5471] remove r5470. Added better support to avoid installing all *.in that includes mpecc.in,mpefc.in,traceTOslog2.in and Makefile.in.

it clearer with respect to what structures are shared by what parts of
the stack.

segments on the same node is still possible, but we use multiple
proxies in that case. This simplifies the code, but might cause minor
performance loss in cases where the user provides weird non-contiguous
segments of ranks on each node.

[svn-r5463] Added a test program for use of MPI_Comm_disconnect, provided by the Parallel R group that thought they had a problem with MPI_Comm_disconnect

[svn-r5462] Enabling user selection of the PM port during installation on windows. Fixes ticket #897