diff --git a/openmp/runtime/src/kmp_barrier.cpp b/openmp/runtime/src/kmp_barrier.cpp
--- a/openmp/runtime/src/kmp_barrier.cpp
+++ b/openmp/runtime/src/kmp_barrier.cpp
@@ -538,6 +538,8 @@
         __itt_get_timestamp();
   }
 #endif
+  KMP_MB(); // Synchronize writes to child threads.
+
   /* Perform a hypercube-embedded tree gather to wait until all of the threads
      have arrived, and reduce any required data as we go.  */
   kmp_flag_64 p_flag(&thr_bar->b_arrived);
@@ -589,6 +591,7 @@
       // Wait for child to arrive
       kmp_flag_64 c_flag(&child_bar->b_arrived, new_state);
       c_flag.wait(this_thr, FALSE USE_ITT_BUILD_ARG(itt_sync_obj));
+      KMP_MB(); // Synchronize writes to child threads.
       ANNOTATE_BARRIER_END(child_thr);
 #if USE_ITT_BUILD && USE_ITT_NOTIFY
       // Barrier imbalance - write min of the thread time and a child time to