Change sub-array offsets to save data move.

src/xt_xmap_intersection_ext.c

@@ -716,9 +716,9 @@ exchange_pos_ext_modifications(
   MPI_Request * requests
     = xmalloc((size_t)(num_src_intersections + 2 * num_dst_intersections) *
               sizeof(*requests));
-  MPI_Request *restrict recv_requests = requests,
-    *restrict send_header_requests = requests + num_src_intersections,
-    *restrict send_data_requests = send_header_requests + num_dst_intersections;
+  MPI_Request *restrict send_header_requests = requests,
+    *restrict recv_requests = requests + num_dst_intersections,
+    *restrict send_data_requests = recv_requests + num_src_intersections;
 
   // set up receives for indices that need to be removed from the send messages
   for (int i = 0; i < num_src_intersections; ++i)
@@ -758,7 +758,7 @@ exchange_pos_ext_modifications(
 
   // wait for the receiving of headers to complete
   xt_mpi_call(MPI_Waitall(num_src_intersections + num_dst_intersections,
-                          recv_requests, MPI_STATUSES_IGNORE), comm);
+                          requests, MPI_STATUSES_IGNORE), comm);
 
   size_t total_num_pos_ext_to_recv = 0;
 
@@ -777,26 +777,23 @@ exchange_pos_ext_modifications(
     size_t offset = 0;
     for (int i = 0; i < num_src_intersections; ++i)
       if (src_removals_per_intersection[i][1] > 0) {
+        ++num_src_changes;
         xt_mpi_call(MPI_Irecv(
                       src_updated_pos_ext + offset,
                       src_removals_per_intersection[i][1], MPI_2INT,
                       src_com[i].rank,
                       tag_offset + xt_mpi_tag_xmap_intersection_data_exchange,
-                      comm, recv_requests + num_src_changes), comm);
+                      comm, send_data_requests - num_src_changes), comm);
 
         offset += (size_t)src_removals_per_intersection[i][1];
-        ++num_src_changes;
       }
   } else
     src_updated_pos_ext = NULL;
 
-  /* move data request handles for compact wait */
-  memcpy(recv_requests + num_src_changes, send_data_requests,
-         num_dst_changes * sizeof (recv_requests[0]));
-
   // wait until all communication is completed
   xt_mpi_call(MPI_Waitall((int)num_src_changes + (int)num_dst_changes,
-                          recv_requests, MPI_STATUSES_IGNORE), comm);
+                          send_data_requests - num_src_changes,
+                          MPI_STATUSES_IGNORE), comm);
 
   free(requests);
   return src_updated_pos_ext;