diff --git a/sql/sql_plugin.cc b/sql/sql_plugin.cc
index 994a09a8177..27d62387a8c 100644
--- a/sql/sql_plugin.cc
+++ b/sql/sql_plugin.cc
@@ -622,7 +622,10 @@ static plugin_ref intern_plugin_lock(LEX *lex, plugin_ref rc CALLER_INFO_PROTO)
   {
     plugin_ref plugin;
 #ifdef DBUG_OFF
-    /* built-in plugins don't need ref counting */
+    /*
+      In optimized builds we don't do reference counting for built-in
+      (plugin->plugin_dl == 0) plugins.
+    */
     if (!pi->plugin_dl)
       DBUG_RETURN(pi);
 
@@ -655,6 +658,26 @@ plugin_ref plugin_lock(THD *thd, plugin_ref ptr CALLER_INFO_PROTO)
   LEX *lex= thd ? thd->lex : 0;
   plugin_ref rc;
   DBUG_ENTER("plugin_lock");
+
+#ifdef DBUG_OFF
+  /*
+    In optimized builds we don't do reference counting for built-in
+    (plugin->plugin_dl == 0) plugins.
+
+    Note that we access plugin->plugin_dl outside of LOCK_plugin, and for
+    dynamic plugins a 'plugin' could correspond to plugin that was unloaded
+    meanwhile!  But because st_plugin_int is always allocated on
+    plugin_mem_root, the pointer can never be invalid - the memory is never
+    freed.
+    Of course, the memory that 'plugin' points to can be overwritten by
+    another plugin being loaded, but plugin->plugin_dl can never change
+    from zero to non-zero or vice versa.
+    That is, it's always safe to check for plugin->plugin_dl==0 even
+    without a mutex.
+  */
+  if (! plugin_dlib(ptr))
+    DBUG_RETURN(ptr);
+#endif
   pthread_mutex_lock(&LOCK_plugin);
   rc= my_intern_plugin_lock_ci(lex, ptr);
   pthread_mutex_unlock(&LOCK_plugin);