Fixes trac:430. Fix a few places where optimization checking conflicted

with the use of MPI_IN_PLACE, and make some optimization checks more
correct.  Thanks to Lisandro Dalcin for reporting the problems.

This commit was SVN r11904.

The following Trac tickets were found above:
  Ticket 430 --> https://svn.open-mpi.org/trac/ompi/ticket/430

NEWS

@@ -68,8 +68,9 @@ version 1.0.
 
 - Fix for large-sized Fortran LOGICAL datatypes.
 - Fix various error checking in MPI_INFO_GET_NTHKEY and
-  MPI_GROUP_TRANSLATE_RANKS.  Thanks to Lisandro Dalcin for reporting
+  MPI_GROUP_TRANSLATE_RANKS, and some collective operations
-  the problem.
+  (particularly with regards to MPI_IN_PLACE).  Thanks to Lisandro
+  Dalcin for reporting the problems.
 - Fix receiving messages to buffers allocated by MPI_ALLOC_MEM.
 - Fix the "tuned" collective componenete where some cases where
   MPI_BCAST could hang.

ompi/mpi/c/allgather.c

@@ -60,12 +60,14 @@ int MPI_Allgather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
         OMPI_ERRHANDLER_CHECK(err, comm, err, FUNC_NAME);
     }
 
-    /* Can we optimize?  Everyone had to give the same send signature,
+    /* Do we need to do anything?  Everyone had to give the same send
-       which means that everyone must have given a sendcount > 0 if
+       signature, which means that everyone must have given a
-       there's anything to send. */
+       sendcount > 0 if there's anything to send.  If we're doing
+       IN_PLACE, however, check recvcount, not sendcount. */
 
-    if (sendcount == 0) {
+    if ((MPI_IN_PLACE != sendbuf && 0 == sendcount) ||
-      return MPI_SUCCESS;
+        (0 == recvcount)) {
+        return MPI_SUCCESS;
     }
 
     /* Invoke the coll component to perform the back-end operation */

ompi/mpi/c/alltoall.c

@@ -61,10 +61,9 @@ int MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype sendtype,
       OMPI_ERRHANDLER_CHECK(err, comm, err, FUNC_NAME);
     }
 
-    /* If the sendcount is 0, since everyone gave the same value, then
+    /* Do we need to do anything? */
-       we don't need to do anything */
 
-    if (0 == sendcount) {
+    if (0 == sendcount && 0 == recvcount) {
         return MPI_SUCCESS;
     }
 

ompi/mpi/c/bcast.c

@@ -71,17 +71,12 @@ int MPI_Bcast(void *buffer, int count, MPI_Datatype datatype,
       } 
     }
 
-    /* If there's only one node, we're done */
+    /* If there's only one node, or if the count is 0, we're done */
 
-    if (OMPI_COMM_IS_INTRA(comm) && ompi_comm_size(comm) <= 1) {
+    if ((OMPI_COMM_IS_INTRA(comm) && ompi_comm_size(comm) <= 1) ||
-      return MPI_SUCCESS;
+        0 == count) {
-    }
+        return MPI_SUCCESS;
-
+    } 
-    /* Can we optimize? */
-
-    if (count == 0) {
-      return MPI_SUCCESS;
-    }
 
     /* Invoke the coll component to perform the back-end operation */
 

ompi/mpi/c/exscan.c

@@ -62,6 +62,14 @@ int MPI_Exscan(void *sendbuf, void *recvbuf, int count,
         OMPI_ERRHANDLER_CHECK(err, comm, err, FUNC_NAME);
     }
 
+    /* Do we need to do anything? (MPI says that reductions have to
+       have a count of at least 1, but at least IMB calls reduce with
+       a count of 0 -- blah!) */
+
+    if (0 == count) {
+        return MPI_SUCCESS;
+    }
+
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);

ompi/mpi/c/gather.c

@@ -112,12 +112,14 @@ int MPI_Gather(void *sendbuf, int sendcount, MPI_Datatype sendtype,
         }
     }
 
-    /* Can we optimize?  Everyone had to give the same send signature,
+    /* Do we need to do anything? */
-       which means that everyone must have given a sendcount > 0 if
-       there's anything to send. */
 
-    if (sendcount == 0) {
+    if ((0 == sendcount &&
-      return MPI_SUCCESS;
+         (ompi_comm_rank(comm) != root ||
+          (ompi_comm_rank(comm) == root && MPI_IN_PLACE != sendbuf))) ||
+        (ompi_comm_rank(comm) == root && MPI_IN_PLACE == sendbuf && 
+         0 == recvbuf)) {
+        return MPI_SUCCESS;
     }
 
     /* Invoke the coll component to perform the back-end operation */

ompi/mpi/c/gatherv.c

@@ -128,12 +128,6 @@ int MPI_Gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
         }
     }
 
-    /* If we have nothing to do, just return */
-
-    if (0 == sendcount && ompi_comm_rank(comm) != root) {
-        return MPI_SUCCESS;
-    }
-
     /* Invoke the coll component to perform the back-end operation */
 	
     err = comm->c_coll.coll_gatherv(sendbuf, sendcount, sendtype, recvbuf,

ompi/mpi/c/reduce.c

@@ -81,10 +81,10 @@ int MPI_Reduce(void *sendbuf, void *recvbuf, int count,
         }
     }
 
-    /* MPI-1, p114, says that each process must supply at least
+    /* Do we need to do anything? (MPI says that reductions have to
-       one element.  But at least the Pallas benchmarks call
+       have a count of at least 1, but at least IMB calls reduce with
-       MPI_REDUCE with a count of 0.  So be sure to handle it. */
+       a count of 0 -- blah!) */
-    
+
     if (0 == count) {
         return MPI_SUCCESS;
     }

ompi/mpi/c/reduce_scatter.c

@@ -88,7 +88,7 @@ int MPI_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts,
     if (size == count) {
         return MPI_SUCCESS;
     }
-        
+
     /* Invoke the coll component to perform the back-end operation */
 
     OBJ_RETAIN(op);

ompi/mpi/c/scan.c

@@ -70,7 +70,9 @@ int MPI_Scan(void *sendbuf, void *recvbuf, int count,
         OMPI_ERRHANDLER_CHECK(err, comm, err, FUNC_NAME);
     }
 
-    /* If everyone supplied count == 0, we can just return */
+    /* Do we need to do anything? (MPI says that reductions have to
+       have a count of at least 1, but at least IMB calls reduce with
+       a count of 0 -- blah!) */
 
     if (0 == count) {
         return MPI_SUCCESS;

ompi/mpi/c/scatter.c

@@ -107,10 +107,13 @@ int MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
         }
     }
 
-    /* If we have nothing to receive, return success (everyone must
+    /* Do we need to do anything? */
-       have given the same recvcount) */
 
-    if (0 == recvcount) {
+    if ((0 == recvcount &&
+         (ompi_comm_rank(comm) != root ||
+          (ompi_comm_rank(comm) == root && MPI_IN_PLACE != recvbuf))) ||
+        (ompi_comm_rank(comm) == root && MPI_IN_PLACE == recvbuf &&
+         0 == sendcount)) {
         return MPI_SUCCESS;
     }
 

ompi/mpi/c/scatterv.c

@@ -136,12 +136,6 @@ int MPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
         }
     }
 
-    /* If we have nothing to do, just return */
-
-    if (0 == recvcount && ompi_comm_rank(comm) != root) {
-        return MPI_SUCCESS;
-    }
-
     /* Invoke the coll component to perform the back-end operation */
 	
     err = comm->c_coll.coll_scatterv(sendbuf, sendcounts, displs, sendtype,