This is an first step in implementing MPI_Get_accumulate.  Currently, only
communication with self is supported.

Reviewer: buntinas

Moved RMA implementation constants from mpidimpl.h to the RMA implementation
header.  Also updated constants to use enumeration types and removed an old
fixme note, which indicated that this should be done.

Reviewer: buntinas

[svn-r10135] Partial fixes for broken attribute handling, including FIXMEs that must still be addressed

Updates from Jeff Hammond and a few more that I caught to enable more tests to
build cleanly without MPI-3 or on non-MPICH MPI implementations.  This includes
the patch Jeff submitted in #1510.

Reviewer: buntinas

AFAIK we have not deleted or otherwise incompatibly altered any of the
interfaces that we expose.  However, several new APIs and a few new
constants have been added, so we need to bump the "current" field while
still supporting the previous interface.  Hence, from "6:1:0" to
"7:0:1".

No reviewer.

[svn-r10122] Correct test for the case where Fortran DOUBLE is C long double (e.g., default Fortran integers are 8 bytes)

Libtool handles the ar flags selection for us, so putting the ar flags
inside of the $AR variable led to "ar cr cr BLAH".  The tflops configure
code is mostly harmless, but contained a similar AR assignment bug and
the platform is *long* gone.

Partial fix for tt#1639.

No reviewer.  (RobL verbally approved, though)

Also error out at configure time if no module outdir flag can be
determined.

Partial fix for tt#1639.

No reviewer.

It's not portable to Solaris (and probably elsewhere).  All of this should
be moot soon, since the PGCC depcomp fix has been accepted upstream
into automake and should be released soon-ish.

Partial fix for tt#1639.

No reviewer.

No reviewer.

Based on suggestions from Jim Hoekstra at Iowa State University.

No reviewer.

This patch was a quick fix, which moved shared memory segment support from
Nemesis up to ch3 in order to support shared memory windows.  This
functionality has been pushed down into Nemesis, so the r10057 fix is no longer
needed.

Reviewer: buntinas

This adds the win_fns table to ch3, which allows the channel to override the
default implementation of window creation routines provided by ch3.  This also
pushes the implementation of shared memory windows down into Nemesis, includes
window functions for sock, and contains multiple improvements to the window
creation functions code.

Reviewer: buntinas

Removed the old RMA virtual function infrastructure from CH3 -- this code was
all already dead.  Function overrides are already provided per-window in the
MPID_Win structure.  Overrides for non-window-specific (window creation)
operations will be added shortly.

Reviewer: buntinas

Also drop an unnecessary "$(top_builddir)" that seemed to be confusing
the FreeBSD make program for some reason.

No reviewer.

No reviewer.

I just moved 4 bits from the number of blocks bit-field to the number of
indices (within a block) bit-field.  This makes the new ranges [1,16384]
and [1,4096] respectively.  The default values have been set at 8192 and
1024, which offers 8 Mi objects but only a minimum of 1024 objects after
spilling to indirect storage.  This should keep communicator usage under
8 MiB on BG/Q.

No reviewer.

Also update the mprobe regression test to cover this case.

Thanks to Lisandro Dalcin for reporting this bug.

Reviewed by buntinas@.

Shortcomings of this test:

 * only tests 1-D Cartesian topology
 * does not directly test the nonblocking interface (although indirectly
   does so in the current MPICH2 implementation)

Nonetheless it was sufficient to find a number of small bugs in the
initial implementation.

No reviewer.

The default nonblocking implementations use the naive "as-if" code from
the draft MPI-3 standard.  No attempt is made to optimize the
communication patterns at this time.

The default blocking implementations simply call the corresponding
nonblocking versions.  NBC overheads are unlikely to be too large for
most neighborhood collective use cases.  If this assertion turns out to
be false at some point then we can revisit this implementation choice.

Reviewed by buntinas@.  (modulo a few last minute fixes)

Split up so that the later impl code is easier to find/read.

Reviewed by buntinas@.

These routines will assist with the neighborhood collectives implementation
effort.

Reviewed by buntinas@.

Specifically for:
* MPI_Cart_shift
* MPI_Dist_graph_neighbors_count
* MPI_Dist_graph_neighbors
* MPI_Graph_neighbors_count
* MPI_Graph_neighbors

This is a precursor to some of the neighborhood collectives
implementation work.

Reviewed by buntinas@.

The &ReplaceIfDifferent logic was causing some problems with
the old logic.

No reviewer.

Also permit it to be run with more than one process without generating
multiple " No errors" lines.

No reviewer.

These annotations allow the compiler to perform type checking between
the C language type of a buffer and the MPI_Datatype specified by the
user.

This new code relies on experimental features currently under
consideration for addition to Clang:

http://lists.cs.uiuc.edu/pipermail/cfe-commits/Week-of-Mon-20120521/058137.html

Code courtesy of Dmitri Gribenko <gribozavr@gmail.com>, with help from
Hal Finkel @ ALCF.

No reviewer.

[svn-r10093] Force use of at least -O in compiling performance tests. Eventually need something like AX_CC_MAXOPT

Reviewer: buntinas

MPID_Win structure and the name should reflect that it's part of the MPID
interface.

Reviewer: balaji