GitLab

Added builtin op implementations for RMA-only ops.  MPI_REPLACE was not defined
in the op_table prior to this commit.  Also includes an alternate
implementation of the linked list test that uses fetch-and-op with these
operations to enable the use of shared locks.

Reviewer: goodell

Pulled bit manipulation arithmetic into macros for converting handles to
indices in MPIR_Op and MPIR_Op_check_dtype function pointer tables.

Reviewer: buntinas

disabled for the public interfaces, but all internal functions use
const.

No reviewer.

The old one was stale and yielded a 404.

No reviewer.

This changes allreduce and reduce to use MPIR_Reduce_local_impl instead
of manually managing operator invocation, which greatly de-clutters the
code and reduces the chances of making an error through copy-paste
coding.  It also strengthens the error checking when using built-in
reduction operators.

No reviewer.

[svn-r8153] COMML_FORGET was called after the communicator was freed, so the communicator's next pointer was invalid.  Fixes #1451.

[svn-r8135] Enhancements to the Fortran binding, including code to not set output character variables when an error occurs in the MPI routine (bug #1448) and partial support for the case where Fortran INTEGER and C int are not the same size (This requires a special option currently to buildiface, as the support is not complete.  However, it does work for many of the MPI routines, and can be used on an experimental basis by friendly users)

[svn-r7720] Fix collectives to not hang if the communicator contains a failed process.  The collectives will not return an error immediately upon detecting a failure, rather they'll return the error at the end of the function and continue the communication pattern so that other processes waiting to receive messages will not hang.  This means that, although the collective should complete at all processes, some processes will receive an error, and some processes may not get a valid result.  Since some processes may not receive an error and still receive an invalid result, a separate mechanism is needed to confirm that the collective has completed correctly, such as MPI_Comm_validate of the MPI3 FT proposal.

This should make it much easier to tune the collective algorithm
selection on a given platform.  It also provides some knobs to expose
to apps/tools via the upcoming MPIT interface.

Reviewed by buntinas@.

[svn-r5355] Added (or moved) MPIU_THREADPRIV_DECL so that the thread-private area is acquired only once within a routine, since the data (pointer to the thread-private area) is fixed during the execution.  This reduces the overhead of the nest increment/decrement, for example

valid, but the reduce and allreduce implementations aren't currently written
correctly to deal with non-commutative operations this way.

No reviewer (old code).

[svn-r3749] Use Rabenseifner's Algorithm for user defined reduction operators in MPI_Reduce and MPI_Allreduce.

Reviewed by thakur@.

Reviewed by thakur@.

[svn-r3456] Getting rid of SVN IDs that are present in 5% of the files and not in the other 95%. This is creating unnecessary conflicts for collaborators who merge from MPICH2.

[svn-r3177] Switch to the new macro for the global thread critical section (most of this is an automated change, tested against the MPICH2 test suite)