+ } else {
+ // i am the ROOT: wait for all buffers by creating one request by sender
+ int src;
+ requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t));
+
+ void **tmpbufs = xbt_malloc((size-1) * sizeof(void *));
+ for (i = 0; i < size-1; i++) {
+ // we need 1 buffer per request to store intermediate receptions
+ tmpbufs[i] = xbt_malloc(count * datatype->size);
+ }
+ // root: initiliaze recv buf with my own snd buf
+ memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char));
+
+ // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
+ // since we should op values as soon as one receiving request matches.
+ for (i = 0; i < size-1; i++) {
+ // reminder: for smpi_create_request() the src is always the process sending.
+ src = i < root ? i : i + 1;
+ retval = smpi_create_request(tmpbufs[i], count, datatype,
+ src, root, tag, comm, &(requests[i]));
+ if (NULL != requests[i] && MPI_SUCCESS == retval) {
+ if (MPI_SUCCESS == retval) {
+ smpi_mpi_irecv(requests[i]);
+ }
+ }
+ }
+ // now, wait for completion of all irecv's.
+ for (i = 0; i < size-1; i++) {
+ int index = MPI_UNDEFINED;
+ smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE);
+#ifdef DEBUG_REDUCE
+ printf ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index);
+ print_buffer_int(tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)", index),
+ rank);
+#endif
+
+ // arg 2 is modified
+ op->func(tmpbufs[index], recvbuf, &count, &datatype);
+#ifdef DEBUG_REDUCE
+ print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank);
+#endif
+ xbt_free(tmpbufs[index]);
+ /* FIXME: with the following line, it generates an
+ * [xbt_ex/CRITICAL] Conditional list not empty 162518800.
+ */
+ // xbt_mallocator_release(smpi_global->request_mallocator, requests[index]);
+ }
+ xbt_free(requests);
+ xbt_free(tmpbufs);
+ }
+ smpi_bench_begin();
+ return retval;
+}