diff --git a/vvp/codes.h b/vvp/codes.h
index d02a2711a..e646f9f07 100644
--- a/vvp/codes.h
+++ b/vvp/codes.h
@@ -242,6 +242,7 @@ extern bool of_XORR(vthread_t thr, vvp_code_t code);
 extern bool of_ZOMBIE(vthread_t thr, vvp_code_t code);
 
 extern bool of_EXEC_UFUNC(vthread_t thr, vvp_code_t code);
+extern bool of_REAP_UFUNC(vthread_t thr, vvp_code_t code);
 
 extern bool of_CHUNK_LINK(vthread_t thr, vvp_code_t code);
 
diff --git a/vvp/ufunc.cc b/vvp/ufunc.cc
index b89a10780..7ac1f00ba 100644
--- a/vvp/ufunc.cc
+++ b/vvp/ufunc.cc
@@ -173,20 +173,25 @@ void compile_ufunc(char*label, char*code, unsigned wid,
 
 	/* Construct some phantom code that is the thread of the
 	   function call. The first instruction, at the start_address
-	   of the function, loads the points and calls the function.
-	   The last instruction is the usual %end. So the thread looks
+	   of the function, loads the ports and calls the function.
+	   The second instruction collects the function result. The
+	   last instruction is the usual %end. So the thread looks
 	   like this:
 
 	      %exec_ufunc <core>;
+	      %reap_ufunc <core>;
 	      %end;
 
-	   The %exec_ufunc copies the input values into local regs,
-           runs the function code, then copies the output values to
-           the destination net functors. */
+	   The %exec_ufunc copies the input values into local regs
+           and runs the function code. The %reap_ufunc then copies
+	   the output value to the destination net functor. */
 
-      vvp_code_t start_code = codespace_allocate();
-      start_code->opcode = of_EXEC_UFUNC;
-      code_label_lookup(start_code, code);
+      vvp_code_t exec_code = codespace_allocate();
+      exec_code->opcode = of_EXEC_UFUNC;
+      code_label_lookup(exec_code, code);
+
+      vvp_code_t reap_code = codespace_allocate();
+      reap_code->opcode = of_REAP_UFUNC;
 
       vvp_code_t end_code = codespace_allocate();
       end_code->opcode = &of_END;
@@ -204,13 +209,14 @@ void compile_ufunc(char*label, char*code, unsigned wid,
 	   that will contain the execution. */
       vvp_net_t*ptr = new vvp_net_t;
       ufunc_core*fcore = new ufunc_core(wid, ptr, portc, ports,
-					start_code, call_scope,
+					exec_code, call_scope,
 					retv.text, scope_label);
       ptr->fun = fcore;
       define_functor_symbol(label, ptr);
       free(label);
 
-      start_code->ufunc_core_ptr = fcore;
+      exec_code->ufunc_core_ptr = fcore;
+      reap_code->ufunc_core_ptr = fcore;
 
       wide_inputs_connect(fcore, argc, argv);
 
diff --git a/vvp/vthread.cc b/vvp/vthread.cc
index 6c304944f..f2b4c5397 100644
--- a/vvp/vthread.cc
+++ b/vvp/vthread.cc
@@ -6468,23 +6468,40 @@ bool of_EXEC_UFUNC(vthread_t thr, vvp_code_t cp)
 	   atomically. */
       cp->ufunc_core_ptr->assign_bits_to_ports(child_context);
 
-	/* Create a temporary thread and run it immediately. A function
-           may not contain any blocking statements, so vthread_run() can
-           only return when the %end opcode is reached. */
+	/* Create a temporary thread and run it immediately. */
       vthread_t child = vthread_new(cp->cptr, child_scope);
       child->wt_context = child_context;
       child->rd_context = child_context;
+
+      child->parent = thr;
       child->is_scheduled = 1;
       vthread_run(child);
       running_thread = thr;
 
-	/* Now copy the output from the result variable to the output
+      if (child->i_have_ended)
+            return true;
+
+      thr->children.insert(child);
+      thr->i_am_joining = 1;
+      return false;
+}
+
+/*
+ * This is a phantom opcode used to harvest the result of calling a user
+ * defined function. It is used in code generated by the .ufunc statement.
+ */
+bool of_REAP_UFUNC(vthread_t thr, vvp_code_t cp)
+{
+      struct __vpiScope*child_scope = cp->ufunc_core_ptr->func_scope();
+      assert(child_scope);
+
+	/* Copy the output from the result variable to the output
 	   ports of the .ufunc device. */
       cp->ufunc_core_ptr->finish_thread();
 
         /* If an automatic function, free the context for this call. */
       if (child_scope->is_automatic) {
-            vthread_free_context(child_context, child_scope);
+            vthread_free_context(thr->rd_context, child_scope);
             thr->wt_context = 0;
             thr->rd_context = 0;
       }