diff --git a/ompi/mca/coll/sm2/coll_sm2_allreduce.c b/ompi/mca/coll/sm2/coll_sm2_allreduce.c
index 0f1c590642..974fbac2de 100644
--- a/ompi/mca/coll/sm2/coll_sm2_allreduce.c
+++ b/ompi/mca/coll/sm2/coll_sm2_allreduce.c
@@ -52,11 +52,6 @@ int mca_coll_sm2_allreduce_intra_fanin_fanout(void *sbuf, void *rbuf, int count,
 
     sm_module=(mca_coll_sm2_module_t *) module;
 
-    /* get unique tag for this collective - assume only one collective
-     *  per communicator at a given time, so no locking needed
-     *  for atomic update of the tag */
-    tag=sm_module->collective_tag;
-    sm_module->collective_tag++;
 
     /* get size of data needed - same layout as user data, so that
      *   we can apply the reudction routines directly on these buffers
@@ -95,6 +90,12 @@ int mca_coll_sm2_allreduce_intra_fanin_fanout(void *sbuf, void *rbuf, int count,
     /* NOTE: starting with a rather synchronous approach */
     for( stripe_number=0 ; stripe_number < n_data_segments ; stripe_number++ ) {
     
+        /* get unique tag for this stripe - assume only one collective
+         *  per communicator at a given time, so no locking needed
+         *  for atomic update of the tag */
+        tag=sm_module->collective_tag;
+        sm_module->collective_tag++;
+
         sm_buffer_desc=alloc_sm2_shared_buffer(sm_module);
 
         /* get number of elements to process in this stripe */
diff --git a/ompi/op/op_predefined.c b/ompi/op/op_predefined.c
index 3b935cacac..98201bb6df 100644
--- a/ompi/op/op_predefined.c
+++ b/ompi/op/op_predefined.c
@@ -680,8 +680,8 @@ LOC_FUNC(minloc, long_double_int, <)
  *    routines, needed for some optimizations.
  */
 #define OP_FUNC_3BUF(name, type_name, type, op) \
-  void ompi_mpi_op_three_buff_##name##_##type_name(restrict void *in1, \
-          restrict void *in2, restrict void *out, int *count,            \
+  void ompi_mpi_op_three_buff_##name##_##type_name(void * restrict in1, \
+          void * restrict in2, void * restrict out, int *count,            \
                                         MPI_Datatype *dtype)             \
   {                                                                      \
     int i;                                                               \
@@ -694,8 +694,8 @@ LOC_FUNC(minloc, long_double_int, <)
   }
 
 #define COMPLEX_OP_FUNC_SUM_3BUF(type_name, type) \
-  void ompi_mpi_op_sum_three_buff_##type_name(restrict void *in1,             \
-          restrict void * in2, restrict void *out, int *count,           \
+  void ompi_mpi_op_sum_three_buff_##type_name(void * restrict in1,             \
+          void * restrict in2, void * restrict out, int *count,           \
                                    MPI_Datatype *dtype)                  \
   {                                                                      \
     int i;                                                               \
@@ -709,8 +709,8 @@ LOC_FUNC(minloc, long_double_int, <)
   }
 
 #define COMPLEX_OP_FUNC_PROD_3BUF(type_name, type) \
-  void ompi_mpi_op_prod_three_buff_##type_name(restrict void *in1,             \
-          restrict void *in2, restrict void *out, int *count,            \
+  void ompi_mpi_op_prod_three_buff_##type_name(void * restrict in1,      \
+          void * restrict in2, void * restrict out, int *count,          \
                                    MPI_Datatype *dtype)                  \
   {                                                                      \
     int i;                                                               \
@@ -734,8 +734,8 @@ LOC_FUNC(minloc, long_double_int, <)
  * This macro is for (out = op(in1, in2))
  */
 #define FUNC_FUNC_3BUF(name, type_name, type) \
-  void ompi_mpi_op_three_buff_##name##_##type_name(restrict void *in1, \
-          restrict void *in2, restrict void *out, int *count,            \
+  void ompi_mpi_op_three_buff_##name##_##type_name(void * restrict in1, \
+          void * restrict in2, void * restrict out, int *count,          \
                                         MPI_Datatype *dtype)             \
   {                                                                      \
     int i;                                                               \
@@ -766,9 +766,9 @@ LOC_FUNC(minloc, long_double_int, <)
 */
 
 #define LOC_FUNC_3BUF(name, type_name, op) \
-  void ompi_mpi_op_three_buff_##name##_##type_name(restrict void *in1,  \
-          restrict void *in2, restrict void *out, int *count,         \
-                                        MPI_Datatype *dtype)          \
+  void ompi_mpi_op_three_buff_##name##_##type_name(void * restrict in1,  \
+          void * restrict in2, void * restrict out, int *count,          \
+                                        MPI_Datatype *dtype)             \
   { \
     int i; \
     ompi_op_predefined_##type_name##_t *a1 = (ompi_op_predefined_##type_name##_t*) in1; \