diff --git a/tgt-vvp/draw_ufunc.c b/tgt-vvp/draw_ufunc.c
index f5903fff3..477ddfab8 100644
--- a/tgt-vvp/draw_ufunc.c
+++ b/tgt-vvp/draw_ufunc.c
@@ -191,9 +191,6 @@ void draw_ufunc_vec4(ivl_expr_t expr)
 	/* Take in arguments to function and call function code. */
       draw_ufunc_preamble(expr);
 
-      assert(ivl_signal_dimensions(retval) == 0);
-      fprintf(vvp_out, "    %%load/vec4  v%p_0;\n", retval);
-
       draw_ufunc_epilogue(expr);
 }
 
diff --git a/tgt-vvp/draw_vpi.c b/tgt-vvp/draw_vpi.c
index 3587815eb..14e18e53a 100644
--- a/tgt-vvp/draw_vpi.c
+++ b/tgt-vvp/draw_vpi.c
@@ -99,6 +99,12 @@ static int get_vpi_taskfunc_signal_arg(struct args_info *result,
 		    /* This should never happen since we have IVL_EX_SELECT. */
 		  return 0;
 
+	    } else if (signal_is_return_value(ivl_expr_signal(expr))) {
+		    /* If the signal is the return value of a function,
+		       then this can't be handled as a true signal, so
+		       fall back on general expression processing. */
+		  return 0;
+
 	    } else if (ivl_expr_signed(expr) !=
 	               ivl_signal_signed(ivl_expr_signal(expr))) {
 		  return 0;
diff --git a/tgt-vvp/eval_vec4.c b/tgt-vvp/eval_vec4.c
index 322b73573..2afd1fbbc 100644
--- a/tgt-vvp/eval_vec4.c
+++ b/tgt-vvp/eval_vec4.c
@@ -1005,6 +1005,15 @@ static void draw_signal_vec4(ivl_expr_t expr)
 {
       ivl_signal_t sig = ivl_expr_signal(expr);
 
+	/* Special Case: If the signal is the return value of the function,
+	   then use a different opcode to get the value. */
+      if (signal_is_return_value(sig)) {
+	    assert(ivl_signal_dimensions(sig) == 0);
+	    fprintf(vvp_out, "    %%retload/vec4 0; Load %s (draw_signal_vec4)\n",
+		    ivl_signal_basename(sig));
+	    return;
+      }
+
 	/* Handle the simple case, a signal expression that is a
 	   simple vector, no array dimensions. */
       if (ivl_signal_dimensions(sig) == 0) {
diff --git a/tgt-vvp/stmt_assign.c b/tgt-vvp/stmt_assign.c
index 8804c6ca2..a28f04ff5 100644
--- a/tgt-vvp/stmt_assign.c
+++ b/tgt-vvp/stmt_assign.c
@@ -209,6 +209,17 @@ static void put_vec_to_lval_slice(ivl_lval_t lval, struct vec_slice_info*slice,
       ivl_signal_t sig = ivl_lval_sig(lval);
       int part_off_idx;
 
+
+	/* Special Case: If the l-value signal is named after its scope,
+	   and the scope is a function, then this is an assign to a return
+	   value and should be handled differently. */
+      if (signal_is_return_value(sig)) {
+	    assert(ivl_signal_dimensions(sig) == 0);
+	    fprintf(vvp_out, "    %%ret/vec4 0; Assign to %s\n",
+		    ivl_signal_basename(sig));
+	    return;
+      }
+
 	/* If the slice of the l-value is a BOOL variable, then cast
 	   the data to a BOOL vector so that the stores can be valid. */
       if (ivl_signal_data_type(sig) == IVL_VT_BOOL) {
@@ -345,6 +356,16 @@ static void store_vec4_to_lval(ivl_statement_t net)
 	    if (lidx+1 < ivl_stmt_lvals(net))
 		  fprintf(vvp_out, "    %%split/vec4 %u;\n", lwid);
 
+	      /* Special Case: If the l-value signal is named after its scope,
+		 and the scope is a function, then this is an assign to a return
+		 value and should be handled differently. */
+	    if (signal_is_return_value(lsig)) {
+		  assert(ivl_signal_dimensions(lsig) == 0);
+		  fprintf(vvp_out, "    %%ret/vec4 0; Assign to %s (store_vec4_to_lval)\n",
+			  ivl_signal_basename(lsig));
+		  continue;
+	    }
+
 	    if (word_ex) {
 		    /* Handle index into an array */
 		  int word_index = allocate_word();
diff --git a/tgt-vvp/vvp_priv.h b/tgt-vvp/vvp_priv.h
index 5255058cc..3a4c17bfc 100644
--- a/tgt-vvp/vvp_priv.h
+++ b/tgt-vvp/vvp_priv.h
@@ -208,6 +208,10 @@ extern void draw_eval_expr_into_integer(ivl_expr_t expr, unsigned ix);
  */
 extern int draw_eval_condition(ivl_expr_t expr);
 
+/*
+ * Return true if the signal is the return value of a function.
+ */
+extern int signal_is_return_value(ivl_signal_t sig);
 
 extern int number_is_unknown(ivl_expr_t ex);
 extern int number_is_immediate(ivl_expr_t ex, unsigned lim_wid, int negative_is_ok);
diff --git a/tgt-vvp/vvp_scope.c b/tgt-vvp/vvp_scope.c
index 503682806..51a2e2615 100644
--- a/tgt-vvp/vvp_scope.c
+++ b/tgt-vvp/vvp_scope.c
@@ -290,6 +290,16 @@ static unsigned is_netlist_signal(ivl_net_logic_t net, ivl_nexus_t nex)
       return rtn;
 }
 
+int signal_is_return_value(ivl_signal_t sig)
+{
+      ivl_scope_t sig_scope = ivl_signal_scope(sig);
+      if (ivl_scope_type(sig_scope) != IVL_SCT_FUNCTION)
+	    return 0;
+      if (strcmp(ivl_signal_basename(sig), ivl_scope_basename(sig_scope)) == 0)
+	    return 1;
+      return 0;
+}
+
 /*
  * This tests a bufz device against an output receiver, and determines
  * if the device can be skipped. If this function returns false, then a
@@ -458,6 +468,13 @@ static void draw_reg_in_scope(ivl_signal_t sig)
 		    ivl_signal_basename(sig), ivl_signal_scope(sig));
 	    return;
       }
+      if ((ivl_signal_data_type(sig)==IVL_VT_LOGIC)
+	  && (ivl_scope_type(ivl_signal_scope(sig))==IVL_SCT_FUNCTION)
+	  && (strcmp(ivl_signal_basename(sig),ivl_scope_basename(ivl_signal_scope(sig))) == 0)) {
+	    fprintf(vvp_out, "; Variable %s is vec4 return value of scope S_%p\n",
+		    ivl_signal_basename(sig), ivl_signal_scope(sig));
+	    return;
+      }
 
       const char *datatype_flag = ivl_signal_integer(sig) ? "/i" :
 			       ivl_signal_signed(sig)? "/s" : "";
diff --git a/vvp/README.txt b/vvp/README.txt
index f74cf2fa6..16de088f0 100644
--- a/vvp/README.txt
+++ b/vvp/README.txt
@@ -702,11 +702,14 @@ STRUCTURAL FUNCTION CALLS:
 
 The .ufunc statements define a call to a user defined function.
 
-	<label> .ufunc   <flabel>, <wid>,
-            <isymbols> ( <psymbols> ) <rsymbol> <ssymbol>;
+	<label> .ufunc/real <flabel>, <wid>,
+            <isymbols> ( <psymbols> ) <ssymbol>;
+
+	<label> .ufunc/vec4 <flabel>, <wid>,
+            <isymbols> ( <psymbols> ) <ssymbol>;
 
 	<label> .ufunc/e <flabel>, <wid>, <trigger>,
-            <isymbols> ( <psymbols> ) <rsymbol> <ssymbol>;
+            <isymbols> ( <psymbols> ) <ssymbol>;
 
 The first variant is used for functions that only need to be called
 when one of their inputs changes value. The second variant is used
@@ -729,10 +732,6 @@ list. The <psymbols> are variables that represent the input ports for
 the function. The ufunc performs an assignment to these variables
 before calling the function.
 
-The <rsymbol> is the variable within the function where the result
-will be found when the function code ends. This value is picked up
-and propagated to the output of the functor.
-
 The <ssymbol> is the function scope name.
 
 THREAD STATEMENTS:
diff --git a/vvp/codes.h b/vvp/codes.h
index fed8a6737..8466386b4 100644
--- a/vvp/codes.h
+++ b/vvp/codes.h
@@ -199,6 +199,9 @@ extern bool of_RELEASE_REG(vthread_t thr, vvp_code_t code);
 extern bool of_RELEASE_WR(vthread_t thr, vvp_code_t code);
 extern bool of_REPLICATE(vthread_t thr, vvp_code_t code);
 extern bool of_RET_REAL(vthread_t thr, vvp_code_t code);
+extern bool of_RET_VEC4(vthread_t thr, vvp_code_t code);
+extern bool of_RETLOAD_REAL(vthread_t thr, vvp_code_t code);
+extern bool of_RETLOAD_VEC4(vthread_t thr, vvp_code_t code);
 extern bool of_SCOPY(vthread_t thr, vvp_code_t code);
 extern bool of_SET_DAR_OBJ_REAL(vthread_t thr, vvp_code_t code);
 extern bool of_SET_DAR_OBJ_STR(vthread_t thr, vvp_code_t code);
diff --git a/vvp/compile.cc b/vvp/compile.cc
index 9422c5619..2191cd2a6 100644
--- a/vvp/compile.cc
+++ b/vvp/compile.cc
@@ -250,6 +250,9 @@ static const struct opcode_table_s opcode_table[] = {
       { "%release/wr", of_RELEASE_WR, 2,{OA_FUNC_PTR,OA_BIT1,OA_NONE} },
       { "%replicate", of_REPLICATE,   1,{OA_NUMBER,  OA_NONE,OA_NONE} },
       { "%ret/real",  of_RET_REAL,    1,{OA_NUMBER,  OA_NONE,OA_NONE} },
+      { "%ret/vec4",  of_RET_VEC4,    1,{OA_NUMBER,  OA_NONE,OA_NONE} },
+      { "%retload/real",of_RETLOAD_REAL,1,{OA_NUMBER,  OA_NONE,OA_NONE} },
+      { "%retload/vec4",of_RETLOAD_VEC4,1,{OA_NUMBER,  OA_NONE,OA_NONE} },
       { "%scopy",  of_SCOPY,  0,  {OA_NONE,     OA_NONE,     OA_NONE} },
       { "%set/dar/obj/real",of_SET_DAR_OBJ_REAL,1,{OA_NUMBER,OA_NONE,OA_NONE} },
       { "%set/dar/obj/str", of_SET_DAR_OBJ_STR, 1,{OA_NUMBER,OA_NONE,OA_NONE} },
diff --git a/vvp/lexor.lex b/vvp/lexor.lex
index 6ce4af9fb..e161f0593 100644
--- a/vvp/lexor.lex
+++ b/vvp/lexor.lex
@@ -209,6 +209,7 @@ static char* strdupnew(char const *str)
 ".tranif1"  { return K_TRANIF1; }
 ".tranvp"   { return K_TRANVP; }
 ".ufunc/real" { return K_UFUNC_REAL; }
+".ufunc/vec4" { return K_UFUNC_VEC4; }
 ".ufunc/e"  { return K_UFUNC_E; }
 ".var"      { return K_VAR; }
 ".var/cobj" { return K_VAR_COBJECT; }
diff --git a/vvp/opcodes.txt b/vvp/opcodes.txt
index b82f88719..6403fc969 100644
--- a/vvp/opcodes.txt
+++ b/vvp/opcodes.txt
@@ -1016,6 +1016,11 @@ for example to store the return value for a real function, use
 "%ret/real 0;". It is up to the function caller to set up the argument
 references.
 
+* %retload/vec4 <index>
+
+Read a value from the indexed function argument. The value is read
+from the argument and pushed to the appropriate stack.
+
 * %set/dar/obj/real <index>
 * %set/dar/obj/str <index>
 * %set/dar/obj/vec4 <index>
diff --git a/vvp/parse.y b/vvp/parse.y
index d91b7b3b8..de29c4750 100644
--- a/vvp/parse.y
+++ b/vvp/parse.y
@@ -95,7 +95,7 @@ static struct __vpiModPath*modpath_dst = 0;
 %token K_RESOLV K_SCOPE K_SFUNC K_SFUNC_E K_SHIFTL K_SHIFTR K_SHIFTRS
 %token K_SUBSTITUTE
 %token K_THREAD K_TIMESCALE K_TRAN K_TRANIF0 K_TRANIF1 K_TRANVP
-%token K_UFUNC_REAL K_UFUNC_E K_UDP K_UDP_C K_UDP_S
+%token K_UFUNC_REAL K_UFUNC_VEC4 K_UFUNC_E K_UDP K_UDP_C K_UDP_S
 %token K_VAR K_VAR_COBJECT K_VAR_DARRAY
 %token K_VAR_QUEUE
 %token K_VAR_S K_VAR_STR K_VAR_I K_VAR_R K_VAR_2S K_VAR_2U
@@ -257,6 +257,9 @@ statement
   | T_LABEL K_UFUNC_REAL T_SYMBOL ',' T_NUMBER ',' symbols '(' symbols ')' T_SYMBOL ';'
       { compile_ufunc_real($1, $3, $5, $7.cnt, $7.vect, $9.cnt, $9.vect, $11, 0); }
 
+  | T_LABEL K_UFUNC_VEC4 T_SYMBOL ',' T_NUMBER ',' symbols '(' symbols ')' T_SYMBOL ';'
+      { compile_ufunc_vec4($1, $3, $5, $7.cnt, $7.vect, $9.cnt, $9.vect, $11, 0); }
+
   | T_LABEL K_UFUNC_E T_SYMBOL ',' T_NUMBER ',' T_SYMBOL ',' symbols '(' symbols ')' T_SYMBOL ';'
       { compile_ufunc_vec4($1, $3, $5, $9.cnt, $9.vect, $11.cnt, $11.vect, $13, $7); }
 
diff --git a/vvp/ufunc.cc b/vvp/ufunc.cc
index d4e5f6779..a8f64eead 100644
--- a/vvp/ufunc.cc
+++ b/vvp/ufunc.cc
@@ -38,6 +38,30 @@
 #include <windows.h>
 #endif
 
+class ufunc_real : public ufunc_core {
+   public:
+      ufunc_real(unsigned ow, vvp_net_t*ptr,
+		 unsigned nports, vvp_net_t**ports,
+		 vvp_code_t start_address,
+		 __vpiScope*call_scope,
+		 char*scope_label);
+      ~ufunc_real();
+
+      void finish_thread();
+};
+
+class ufunc_vec4 : public ufunc_core {
+   public:
+      ufunc_vec4(unsigned ow, vvp_net_t*ptr,
+		 unsigned nports, vvp_net_t**ports,
+		 vvp_code_t start_address,
+		 __vpiScope*call_scope,
+		 char*scope_label);
+      ~ufunc_vec4();
+
+      void finish_thread();
+};
+
 ufunc_core::ufunc_core(unsigned owid, vvp_net_t*ptr,
 		       unsigned nports, vvp_net_t**ports,
 		       vvp_code_t sa, __vpiScope*call_scope__,
@@ -114,6 +138,18 @@ void ufunc_core::finish_thread_real_()
       thread_ = 0;
 }
 
+void ufunc_core::finish_thread_vec4_()
+{
+      assert(thread_);
+
+      vvp_vector4_t val = vthread_get_vec4_stack(thread_, 0);
+      vthread_pop_vec4(thread_, 1);
+
+      propagate_vec4(val);
+
+      thread_ = 0;
+}
+
 /*
  * This method is only called when a trigger event occurs. Just arrange for
  * the function to be called.
@@ -147,6 +183,24 @@ void ufunc_core::invoke_thread_()
       }
 }
 
+ufunc_vec4::ufunc_vec4(unsigned ow, vvp_net_t*ptr,
+		       unsigned nports, vvp_net_t**ports,
+		       vvp_code_t start_address,
+		       __vpiScope*call_scope,
+		       char*scope_label)
+: ufunc_core(ow, ptr, nports, ports, start_address, call_scope, scope_label)
+{
+}
+
+ufunc_vec4::~ufunc_vec4()
+{
+}
+
+void ufunc_vec4::finish_thread()
+{
+      finish_thread_vec4_();
+}
+
 ufunc_real::ufunc_real(unsigned ow, vvp_net_t*ptr,
 		       unsigned nports, vvp_net_t**ports,
 		       vvp_code_t start_address,
@@ -195,8 +249,8 @@ void compile_ufunc_real(char*label, char*code, unsigned wid,
 	   last instruction is the usual %end. So the thread looks
 	   like this:
 
-	      %exec_ufunc <core>;
-	      %reap_ufunc <core>;
+	      %exec_ufunc/real <core>;
+	      %reap_ufunc;
 	      %end;
 
 	   The %exec_ufunc copies the input values into local regs
@@ -251,8 +305,70 @@ void compile_ufunc_vec4(char*label, char*code, unsigned wid,
 		   unsigned portc, struct symb_s*portv,
 		   char*scope_label, char*trigger_label)
 {
-	// XXXX NOT IMPLEMENTED
-      assert(0);
+	/* The input argument list and port list must have the same
+	   sizes, since internally we will be mapping the inputs list
+	   to the ports list. */
+      assert(argc == portc);
+
+      __vpiScope*call_scope = vpip_peek_current_scope();
+      assert(call_scope);
+
+	/* Construct some phantom code that is the thread of the
+	   function call. The first instruction, at the start_address
+	   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/vec4 <core>;
+	      %reap_ufunc;
+	      %end;
+
+	   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 exec_code = codespace_allocate();
+      exec_code->opcode = of_EXEC_UFUNC_VEC4;
+      code_label_lookup(exec_code, code, false);
+
+      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;
+
+	/* Run through the function ports (which are related to but
+	   not the same as the input ports) and arrange for their
+	   binding. */
+      vvp_net_t**ports = new vvp_net_t*[portc];
+      for (unsigned idx = 0 ;  idx < portc ;  idx += 1) {
+	    functor_ref_lookup(&ports[idx], portv[idx].text);
+      }
+
+	/* Create the output functor and attach it to the label. Tell
+	   it about the start address of the code stub, and the scope
+	   that will contain the execution. */
+      vvp_net_t*ptr = new vvp_net_t;
+      ufunc_core*fcore = new ufunc_vec4(wid, ptr, portc, ports,
+					exec_code, call_scope,
+					scope_label);
+      ptr->fun = fcore;
+      define_functor_symbol(label, ptr);
+      free(label);
+
+      exec_code->ufunc_core_ptr = fcore;
+      reap_code->ufunc_core_ptr = fcore;
+
+      wide_inputs_connect(fcore, argc, argv);
+
+        /* If this function has a trigger event, connect the functor to
+           that event. */
+      if (trigger_label)
+            input_connect(ptr, 0, trigger_label);
+
+      free(argv);
+      free(portv);
 }
 #ifdef CHECK_WITH_VALGRIND
 static map<ufunc_core*, bool> ufunc_map;
diff --git a/vvp/ufunc.h b/vvp/ufunc.h
index 437bf1d8b..4517bd47e 100644
--- a/vvp/ufunc.h
+++ b/vvp/ufunc.h
@@ -70,6 +70,7 @@ class ufunc_core : public vvp_wide_fun_core {
 
     protected:
       void finish_thread_real_();
+      void finish_thread_vec4_();
 
     private:
       void recv_vec4_from_inputs(unsigned port);
@@ -94,16 +95,4 @@ class ufunc_core : public vvp_wide_fun_core {
       vvp_code_t code_;
 };
 
-class ufunc_real : public ufunc_core {
-   public:
-      ufunc_real(unsigned ow, vvp_net_t*ptr,
-		 unsigned nports, vvp_net_t**ports,
-		 vvp_code_t start_address,
-		 __vpiScope*call_scope,
-		 char*scope_label);
-      ~ufunc_real();
-
-      void finish_thread();
-};
-
 #endif /* IVL_ufunc_H */
diff --git a/vvp/vthread.cc b/vvp/vthread.cc
index 473e48ae8..4725d8b0b 100644
--- a/vvp/vthread.cc
+++ b/vvp/vthread.cc
@@ -115,6 +115,7 @@ struct vthread_s {
 	// corresponding stack.  This is how the %ret/* instructions
 	// get at parent thread arguments.
       vector<unsigned> args_real;
+      vector<unsigned> args_vec4;
 
     private:
       vector<vvp_vector4_t>stack_vec4_;
@@ -143,6 +144,12 @@ struct vthread_s {
 	    assert(use_index >= 1);
 	    return stack_vec4_[use_index-1];
       }
+      inline void poke_vec4(unsigned depth, const vvp_vector4_t&val)
+      {
+	    assert(depth < stack_vec4_.size());
+	    unsigned use_index = stack_vec4_.size()-1-depth;
+	    stack_vec4_[use_index] = val;
+      }
       inline void pop_vec4(unsigned cnt)
       {
 	    while (cnt > 0) {
@@ -1348,9 +1355,13 @@ bool of_CALLF_STR(vthread_t thr, vvp_code_t cp)
 bool of_CALLF_VEC4(vthread_t thr, vvp_code_t cp)
 {
       vthread_t child = vthread_new(cp->cptr2, cp->scope);
-      return do_callf_void(thr, child);
 
-      // XXXX NOT IMPLEMENTED
+	// This is the return value. Push a place-holder value. The function
+	// will replace this with the actual value using a %ret/real instruction.
+      thr->push_vec4(vvp_vector4_t());
+      child->args_vec4.push_back(0);
+
+      return do_callf_void(thr, child);
 }
 
 bool of_CALLF_VOID(vthread_t thr, vvp_code_t cp)
@@ -4964,6 +4975,47 @@ bool of_RET_REAL(vthread_t thr, vvp_code_t cp)
       return true;
 }
 
+/*
+ * %ret/vec4 <index>
+ */
+bool of_RET_VEC4(vthread_t thr, vvp_code_t cp)
+{
+      size_t index = cp->number;
+      vvp_vector4_t val = thr->pop_vec4();
+
+      assert(index >= 0 && index < thr->args_vec4.size());
+      unsigned depth = thr->args_vec4[index];
+	// Use the depth to put the value into the stack of
+	// the parent thread.
+      thr->parent->poke_vec4(depth, val);
+      return true;
+}
+
+/*
+ * %retload/real <index>
+ */
+bool of_RETLOAD_REAL(vthread_t thr, vvp_code_t cp)
+{
+	// NOT IMPLEMENTED
+      assert(0);
+      return true;
+}
+
+/*
+ * %retload/vec4 <index>
+ */
+bool of_RETLOAD_VEC4(vthread_t thr, vvp_code_t cp)
+{
+      size_t index = cp->number;
+
+      assert(index >= 0 && index < thr->args_vec4.size());
+      unsigned depth = thr->args_vec4[index];
+	// Use the depth to put the value into the stack of
+	// the parent thread.
+      thr->push_vec4(thr->parent->peek_vec4(depth));
+      return true;
+}
+
 bool of_SCOPY(vthread_t thr, vvp_code_t)
 {
       vvp_object_t tmp;
@@ -5971,7 +6023,20 @@ bool of_EXEC_UFUNC_REAL(vthread_t thr, vvp_code_t cp)
 	/* Create a temporary thread and run it immediately. */
       vthread_t child = vthread_new(cp->cptr, child_scope);
       thr->push_real(0.0);
-      child->args_real.push_back(0.0);
+      child->args_real.push_back(0);
+
+      return do_exec_ufunc(thr, cp, child);
+}
+
+bool of_EXEC_UFUNC_VEC4(vthread_t thr, vvp_code_t cp)
+{
+      __vpiScope*child_scope = cp->ufunc_core_ptr->func_scope();
+      assert(child_scope);
+
+	/* Create a temporary thread and run it immediately. */
+      vthread_t child = vthread_new(cp->cptr, child_scope);
+      thr->push_vec4(vvp_vector4_t());
+      child->args_vec4.push_back(0);
 
       return do_exec_ufunc(thr, cp, child);
 }