The send and recv buffers aliased each other in an "in place" fashion,
yet MPI_IN_PLACE was not used.  Thanks to Joe Ratterman @ IBM for the
bug report.

No reviewer.

Thanks to Joe Ratterman @ IBM for the bug report.

No reviewer.

This required moving CHKMEMCPYANDSTMT above the "/* --BEGIN ERROR MACROS-- */"
line.  This lets maint/extracterrmsgs know that **memcpyalias is
actually used and avoids an "(unknown)(): Internal MPI error!" message
when memcpy arguments are aliased in MPIR_Localcopy.

Also fix a format string mismatch.

No reviewer.

Allows specifying the errnames.txt file on the command line without
specifying the whole src/mpi/errhan directory.

No reviewer.

[svn-r6949] Added Mac OSX Darwin to known architectures for ROMIO.  This eliminates warnings about unrecognized architecture when building ROMIO

[svn-r6948] Thread id type is defined only if thread level is >= THREAD_SERIALIZED, so define cs_holders -added in r6911- only in that case

No reviewer.

[svn-r6940] Replaced enable_fc=default setting by enable_fc=yes and set enable_fc=no if any Fortran90 test fails.  Guarded all AC_PROG_[CC|F77|FC|CXX] so none of the *FLAGS would be set -g -O2.  Eliminated duplicated PAC_PROG_FC/PAC_PROG_FC_WORKS caused by inconsistent use of enable_fc=default.  Moved AC_PROG_[CC|F77|FC|CXX] right before the corresponding language tests to improve readability of the code.

[svn-r6939] bug fix:  the communicator wasn't passed to the err_return function, so we weren't checking whether to return errors or abort

This commit may be slightly conservative in some of the pt2pt code, but
it is not obvious that persistent requests handles don't need to be
published safely.  Until it is proven otherwise, we will take the
write memory barrier hit (on some platforms).

No reviewer.

[svn-r6937] add MPIU_OBJ_PUBLISH_HANDLE macro to enable lockfree access to read-only MPI object fields

No reviewer.

No reviewer.

No reviewer.

No reviewer.

No reviewer.

Fine-grained multithreading still breaks dynamic processes in a pretty big way,
but at least the test suite passes a bit more reliably with these in place.

No reviewer.

No reviewer.

This suppresses warnings about defined but not used.

No reviewer.

No reviewer.

This seems like it is a real bug, but there's no obvious fix right now.

No reviewer.

No reviewer.

No reviewer.

Racy accesses don't happen any longer, although I think this code now runs
slower in some cases.  Also, unrelated to this change, this algorithm probably
scales poorly with the number of contending threads.

No reviewer.

This commit begins to parallelize the progress engine slightly by making
it possible to signal request completion without acquiring the (bloated
and largely serial) MPIDCOMM critical section.

No reviewer.

No reviewer.

No reviewer.

When compiled for fine-grained threading, the completion counter serves
as a form of lockfree signalling.  As such, atomic access and memory
barriers must be used to ensure correctness.

In per-object mode, this code also contains valgrind client request annotations
to inform Helgrind/DRD/TSan about the lockfree signalling pattern.

No reviewer.

This code is still really, really nasty and shouldn't be doing what it
is doing.  The progress engine is indirectly called recursively, even
with a single thread.

No reviewer.

Reviewed by buntinas@.

No reviewer.

No reviewer.