diff --git a/ompi/mca/coll/sm/coll_sm.h b/ompi/mca/coll/sm/coll_sm.h
index dcf730df03..9ce7dfc084 100644
--- a/ompi/mca/coll/sm/coll_sm.h
+++ b/ompi/mca/coll/sm/coll_sm.h
@@ -62,7 +62,7 @@ extern "C" {
 
         /** Number of processes in this communicator who have seen
             this value already. */
-        volatile  int smbcs_count;
+        volatile int smbcs_count;
 
         /** Mechanism for letting multiple processes know whether this
             allocation succeeded or not */
@@ -242,7 +242,7 @@ extern "C" {
      */
     struct mca_coll_base_mpool_index_t {
         /** Pointer to beginning of control data */
-        uint32_t *mcbmi_control;
+        uint32_t volatile *mcbmi_control;
         /** Pointer to beginning of message fragment data */
         char *mcbmi_data;
     };
@@ -514,7 +514,7 @@ extern int32_t mca_coll_sm_bogus_free_after;
  */
 #define CHILD_WAIT_FOR_NOTIFY(rank, index, value) \
     do { \
-        volatile uint32_t *ptr = ((uint32_t*) \
+        uint32_t volatile *ptr = ((uint32_t*) \
                                   (((char*) index->mcbmi_control) + \
                                    ((rank) * mca_coll_sm_component.sm_control_size))); \
         while (0 == *ptr) SPIN; \
@@ -527,7 +527,7 @@ extern int32_t mca_coll_sm_bogus_free_after;
  * segment.  Used for fan in operations.
  */
 #define CHILD_NOTIFY_PARENT(child_rank, parent_rank, index, value) \
-    ((size_t*) \
+    ((size_t volatile *) \
      (((char*) (index)->mcbmi_control) + \
       (mca_coll_sm_component.sm_control_size * \
        (parent_rank))))[(child_rank)] = (value)
@@ -539,7 +539,7 @@ extern int32_t mca_coll_sm_bogus_free_after;
  */
 #define PARENT_WAIT_FOR_NOTIFY_SPECIFIC(child_rank, parent_rank, index, value) \
     do { \
-        volatile size_t *ptr = ((size_t *) \
+        size_t volatile *ptr = ((size_t volatile *) \
                                 (((char*) index->mcbmi_control) + \
                                  (mca_coll_sm_component.sm_control_size * \
                                   (parent_rank)))) + child_rank; \
diff --git a/ompi/mca/coll/sm/coll_sm_module.c b/ompi/mca/coll/sm/coll_sm_module.c
index c1d2640fc5..e1e2e67464 100644
--- a/ompi/mca/coll/sm/coll_sm_module.c
+++ b/ompi/mca/coll/sm/coll_sm_module.c
@@ -405,9 +405,9 @@ sm_module_init(struct ompi_communicator_t *comm)
         /* Memory affinity: control */
 
         maffinity[j].mbs_len = c->sm_control_size;
-        maffinity[j].mbs_start_addr = 
-            data->mcb_mpool_index[i].mcbmi_control +
-            (rank * c->sm_control_size);
+        maffinity[j].mbs_start_addr = (void *)
+            (data->mcb_mpool_index[i].mcbmi_control +
+             (rank * c->sm_control_size));
         ++j;
 
         /* Memory affinity: data */
@@ -430,7 +430,7 @@ sm_module_init(struct ompi_communicator_t *comm)
     memset(data->mcb_barrier_control_me, 0, 
            num_barrier_buffers * 2 * c->sm_control_size);
     for (i = 0; i < c->sm_comm_num_segments; ++i) {
-        memset(data->mcb_mpool_index[i].mcbmi_control, 0,
+        memset((void *) data->mcb_mpool_index[i].mcbmi_control, 0,
                c->sm_control_size);
     }
 
diff --git a/ompi/mca/coll/sm/coll_sm_reduce.c b/ompi/mca/coll/sm/coll_sm_reduce.c
index 9df0f85a0a..cfebe05654 100644
--- a/ompi/mca/coll/sm/coll_sm_reduce.c
+++ b/ompi/mca/coll/sm/coll_sm_reduce.c
@@ -147,7 +147,6 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
     mca_coll_sm_in_use_flag_t *flag;
     ompi_convertor_t convertor;
     mca_coll_base_mpool_index_t *index;
-    mca_coll_sm_tree_node_t *me;
     int32_t ddt_size;
     size_t segment_ddt_count, segment_ddt_bytes, zero = 0;
 
@@ -156,8 +155,6 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
     rank = ompi_comm_rank(comm);
     size = ompi_comm_size(comm);
 
-    me = &data->mcb_tree[(rank + size - root) % size];
-
     /* Figure out how much we should have the convertor copy.  We need
        to have it be in units of a datatype -- i.e., we only want to
        copy a whole datatype worth of data or none at all (we've
@@ -255,10 +252,8 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
         /* Main loop over receiving / reducing fragments */
 
         do {
-
             flag_num = (data->mcb_operation_count % 
                         mca_coll_sm_component.sm_comm_num_in_use_flags);
-
             FLAG_SETUP(flag_num, flag, data);
             FLAG_WAIT_FOR_IDLE(flag);
             FLAG_RETAIN(flag, size, data->mcb_operation_count);
@@ -417,7 +412,6 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
      *********************************************************************/
 
     else {
-        int parent_rank = (me->mcstn_parent->mcstn_id + root) % size;
 
         /* Here we get a convertor for the full count that the user
            provided (as opposed to the convertor that the root got) */
@@ -454,7 +448,7 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
             do {
                 index = &(data->mcb_mpool_index[segment_num]);
 
-                /* Copy from the user's buffer to the shared mem
+                /* Copy from the user's buffer to my shared mem
                    segment */
                 COPY_FRAGMENT_IN(convertor, index, rank, iov, max_data);
                 bytes += max_data;
@@ -462,8 +456,10 @@ static int reduce_inorder(void *sbuf, void* rbuf, int count,
                 /* Wait for the write to absolutely complete */
                 opal_atomic_wmb();
                 
-                /* Tell my parent that this fragment is ready */
-                CHILD_NOTIFY_PARENT(rank, parent_rank, index, max_data);
+                /* Tell my parent (always the reduction root -- we're
+                   ignoring the mcb_tree parent/child relationships
+                   here) that this fragment is ready */
+                CHILD_NOTIFY_PARENT(rank, root, index, max_data);
 
                 ++segment_num;
             } while (bytes < total_size && segment_num < max_segment_num);