Add support for MPI_IN_PLACE to MPI_Exscan. Required for MPI 2.2 compliance.

Reviewed by Jeff Squyres.  This fixes trac:2221.

This commit was SVN r25165.

The following Trac tickets were found above:
  Ticket 2221 --> https://svn.open-mpi.org/trac/ompi/ticket/2221

ompi/mca/coll/basic/coll_basic_exscan.c

@@ -9,6 +9,7 @@
  *                         University of Stuttgart.  All rights reserved.
  * Copyright (c) 2004-2005 The Regents of the University of California.
  *                         All rights reserved.
+ * Copyright (c) 2011      NVIDIA Corporation.  All rights reserved.
  * $COPYRIGHT$
  * 
  * Additional copyrights may follow
@@ -49,16 +50,18 @@ mca_coll_basic_exscan_intra(void *sbuf, void *rbuf, int count,
     ptrdiff_t true_lb, true_extent, lb, extent;
     char *free_buffer = NULL;
     char *reduce_buffer = NULL;
-    char *source;
-    MPI_Request req = MPI_REQUEST_NULL;
-
-    /* Initialize. */
 
     rank = ompi_comm_rank(comm);
     size = ompi_comm_size(comm);
 
-    /* If we're rank 0, then we send our sbuf to the next rank */
+    /* For MPI_IN_PLACE, just adjust send buffer to point to
+     * receive buffer. */
+    if (MPI_IN_PLACE == sbuf) {
+        sbuf = rbuf;
+    }
 
+    /* If we're rank 0, then just send our sbuf to the next rank, and
+     * we are done. */
     if (0 == rank) {
         return MCA_PML_CALL(send(sbuf, count, dtype, rank + 1,
                                  MCA_COLL_BASE_TAG_EXSCAN,
@@ -66,8 +69,7 @@ mca_coll_basic_exscan_intra(void *sbuf, void *rbuf, int count,
     }
 
     /* If we're the last rank, then just receive the result from the
-     * prior rank */
-
+     * prior rank, and we are done. */
     else if ((size - 1) == rank) {
         return MCA_PML_CALL(recv(rbuf, count, dtype, rank - 1,
                                  MCA_COLL_BASE_TAG_EXSCAN, comm,
@@ -77,17 +79,8 @@ mca_coll_basic_exscan_intra(void *sbuf, void *rbuf, int count,
     /* Otherwise, get the result from the prior rank, combine it with my
      * data, and send it to the next rank */
 
-    /* Start the receive for the prior rank's answer */
-
-    err = MCA_PML_CALL(irecv(rbuf, count, dtype, rank - 1,
-                             MCA_COLL_BASE_TAG_EXSCAN, comm, &req));
-    if (MPI_SUCCESS != err) {
-        goto error;
-    }
-
     /* Get a temporary buffer to perform the reduction into.  Rationale
      * for malloc'ing this size is provided in coll_basic_reduce.c. */
-
     ompi_datatype_get_extent(dtype, &lb, &extent);
     ompi_datatype_get_true_extent(dtype, &true_lb, &true_extent);
 
@@ -96,72 +89,29 @@ mca_coll_basic_exscan_intra(void *sbuf, void *rbuf, int count,
         return OMPI_ERR_OUT_OF_RESOURCE;
     }
     reduce_buffer = free_buffer - lb;
+    err = ompi_datatype_copy_content_same_ddt(dtype, count, 
+                                              reduce_buffer, (char*)sbuf);
 
-    if (ompi_op_is_commute(op)) {
-
-        /* If we're commutative, we can copy my sbuf into the reduction
-         * buffer before the receive completes */
-
-        err = ompi_datatype_copy_content_same_ddt(dtype, count, 
-                                             reduce_buffer, (char*)sbuf);
-        if (MPI_SUCCESS != err) {
-            goto error;
-        }
-
-        /* Now setup the reduction */
-
-        source = (char*)rbuf;
-
-        /* Finally, wait for the receive to complete (so that we can do
-         * the reduction).  */
-
-        err = ompi_request_wait(&req, MPI_STATUS_IGNORE);
-        if (MPI_SUCCESS != err) {
-            goto error;
-        }
-    } else {
-
-        /* Setup the reduction */
-
-        source = (char*)sbuf;
-
-        /* If we're not commutative, we have to wait for the receive to
-         * complete and then copy it into the reduce buffer */
-
-        err = ompi_request_wait(&req, MPI_STATUS_IGNORE);
-        if (MPI_SUCCESS != err) {
-            goto error;
-        }
-
-        err = ompi_datatype_copy_content_same_ddt(dtype, count, 
-                                             reduce_buffer, (char*)rbuf);
-        if (MPI_SUCCESS != err) {
-            goto error;
-        }
+    /* Receive the reduced value from the prior rank */
+    err = MCA_PML_CALL(recv(rbuf, count, dtype, rank - 1,
+                            MCA_COLL_BASE_TAG_EXSCAN, comm, MPI_STATUS_IGNORE));
+    if (MPI_SUCCESS != err) {
+        goto error;
     }
 
-    /* Now reduce the received answer with my source into the answer
-     * that we send off to the next rank */
-
-    ompi_op_reduce(op, source, reduce_buffer, count, dtype);
+    /* Now reduce the prior rank's result with my source buffer.  The source
+     * buffer had been previously copied into the temporary reduce_buffer. */
+    ompi_op_reduce(op, rbuf, reduce_buffer, count, dtype);
 
     /* Send my result off to the next rank */
-
     err = MCA_PML_CALL(send(reduce_buffer, count, dtype, rank + 1,
                             MCA_COLL_BASE_TAG_EXSCAN,
                             MCA_PML_BASE_SEND_STANDARD, comm));
-
     /* Error */
-
   error:
     free(free_buffer);
-    if (MPI_REQUEST_NULL != req) {
-        ompi_request_cancel(req);
-        ompi_request_wait(&req, MPI_STATUS_IGNORE);
-    }
 
     /* All done */
-
     return err;
 }