Make real functions in processes use parent stack for return value.

tgt-vvp/draw_ufunc.c

@@ -205,11 +205,8 @@ void draw_ufunc_real(ivl_expr_t expr)
 	/* Take in arguments to function and call the function code. */
       draw_ufunc_preamble(expr);
 
-	/* Return value signal cannot be an array. */
-      assert(ivl_signal_dimensions(retval) == 0);
-
-	/* Load the result into a word. */
-      fprintf(vvp_out, "  %%load/real v%p_0;\n", retval);
+	/* The %callf/real function emitted by the preamble leaves
+	   the result in the stack for us. */
 
       draw_ufunc_epilogue(expr);
 }

tgt-vvp/stmt_assign.c

@@ -574,6 +574,18 @@ static int show_stmt_assign_sig_real(ivl_statement_t net)
       var = ivl_lval_sig(lval);
       assert(var != 0);
 
+	/* 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. */
+      ivl_scope_t sig_scope = ivl_signal_scope(var);
+      if ((ivl_scope_type(sig_scope) == IVL_SCT_FUNCTION)
+	  && (strcmp(ivl_signal_basename(var), ivl_scope_basename(sig_scope)) == 0)) {
+	    assert(ivl_signal_dimensions(var) == 0);
+	    fprintf(vvp_out, "    %%ret/real 0; Assign to %s\n",
+		    ivl_signal_basename(var));
+	    return 0;
+      }
+
       if (ivl_signal_dimensions(var) == 0) {
 	    fprintf(vvp_out, "    %%store/real v%p_0;\n", var);
 	    return 0;

tgt-vvp/vvp_scope.c

@@ -449,6 +449,16 @@ static void draw_reg_in_scope(ivl_signal_t sig)
 	    break;
       }
 
+	/* Special Case: If this variable is the return value of a function,
+	   then it need to exist as an actual variable. */
+      if ((ivl_signal_data_type(sig)==IVL_VT_REAL)
+	  && (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 REAL 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" : "";
       const char *local_flag = local_flag_str(sig);

vvp/codes.h

@@ -198,6 +198,7 @@ extern bool of_RELEASE_NET(vthread_t thr, vvp_code_t code);
 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_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);

vvp/compile.cc

@@ -249,6 +249,7 @@ static const struct opcode_table_s opcode_table[] = {
       { "%release/reg",of_RELEASE_REG,3,{OA_FUNC_PTR,OA_BIT1,OA_BIT2} },
       { "%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} },
       { "%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} },

vvp/opcodes.txt

@@ -225,7 +225,7 @@ out, then this function is a no-op.
 * %callf/void <code-label>, <scope-label>
 
 More directly implement function calling. This subsumes the %fork and
-%join of the mroe general task and block calling, but is optimized for
+%join of the more general task and block calling, but is optimized for
 functions, which are threads of a special, constrained sort.
 
 The different variants reflect the different return types for the
@@ -1004,6 +1004,18 @@ Pop the vec4 value, replicate it <count> times, then push the
 result. In other words, push the concatenation of <count> copies.
 See also the %concat instruction.
 
+* %ret/obj <index>
+* %ret/real <index>
+* %ret/str <index>
+* %ret/vec4 <index>
+
+Write a value to the indexed function argument. The value is popped
+from the appropriate stack and written into the argument. The return
+value of a function is the first argument of the appropriate type, so
+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.
+
 * %set/dar/obj/real <index>
 * %set/dar/obj/str <index>
 * %set/dar/obj/vec4 <index>

vvp/vthread.cc

@@ -111,6 +111,11 @@ struct vthread_s {
 	    uint64_t w_uint;
       } words[WORDS_COUNT];
 
+	// These vectors are depths within the parent thread's
+	// corresponding stack.  This is how the %ret/* instructions
+	// get at parent thread arguments.
+      vector<unsigned> args_real;
+
     private:
       vector<vvp_vector4_t>stack_vec4_;
     public:
@@ -167,6 +172,12 @@ struct vthread_s {
 	    unsigned use_index = stack_real_.size()-1-depth;
 	    return stack_real_[use_index];
       }
+      inline void poke_real(unsigned depth, double val)
+      {
+	    assert(depth < stack_real_.size());
+	    unsigned use_index = stack_real_.size()-1-depth;
+	    stack_real_[use_index] = val;
+      }
       inline void pop_real(unsigned cnt)
       {
 	    while (cnt > 0) {
@@ -1270,11 +1281,10 @@ bool of_BREAKPOINT(vthread_t, vvp_code_t)
  * %callf/void <code-label>, <scope-label>
  * Combine the %fork and %join steps for invoking a function.
  */
-static bool do_callf_void(vthread_t thr, vvp_code_t cp)
+static bool do_callf_void(vthread_t thr, vthread_t child)
 {
-      vthread_t child = vthread_new(cp->cptr2, cp->scope);
 
-      if (cp->scope->is_automatic()) {
+      if (child->parent_scope->is_automatic()) {
 	      /* The context allocated for this child is the top entry
 		 on the write context stack */
 	    child->wt_context = thr->wt_context;
@@ -1289,7 +1299,7 @@ static bool do_callf_void(vthread_t thr, vvp_code_t cp)
 
         // Execute the function. This SHOULD run the function to completion,
         // but there are some exceptional situations where it won't.
-      assert(cp->scope->get_type_code() == vpiFunction);
+      assert(child->parent_scope->get_type_code() == vpiFunction);
       thr->task_func_children.insert(child);
       child->is_scheduled = 1;
       child->i_am_in_function = 1;
@@ -1309,36 +1319,44 @@ static bool do_callf_void(vthread_t thr, vvp_code_t cp)
 
 bool of_CALLF_OBJ(vthread_t thr, vvp_code_t cp)
 {
-      return do_callf_void(thr, cp);
+      vthread_t child = vthread_new(cp->cptr2, cp->scope);
+      return do_callf_void(thr, child);
 
       // XXXX NOT IMPLEMENTED
 }
 
 bool of_CALLF_REAL(vthread_t thr, vvp_code_t cp)
 {
-      // XXXX Here, I should arrange for a reference to the destination variable
-      // XXXX as a place in my stack. The function will write to that place in
-      // XXXX my stack for me.
-      return do_callf_void(thr, cp);
+      vthread_t child = vthread_new(cp->cptr2, cp->scope);
+
+	// 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_real(0.0);
+      child->args_real.push_back(0);
+
+      return do_callf_void(thr, child);
 }
 
 bool of_CALLF_STR(vthread_t thr, vvp_code_t cp)
 {
-      return do_callf_void(thr, cp);
+      vthread_t child = vthread_new(cp->cptr2, cp->scope);
+      return do_callf_void(thr, child);
 
       // XXXX NOT IMPLEMENTED
 }
 
 bool of_CALLF_VEC4(vthread_t thr, vvp_code_t cp)
 {
-      return do_callf_void(thr, cp);
+      vthread_t child = vthread_new(cp->cptr2, cp->scope);
+      return do_callf_void(thr, child);
 
       // XXXX NOT IMPLEMENTED
 }
 
 bool of_CALLF_VOID(vthread_t thr, vvp_code_t cp)
 {
-      return do_callf_void(thr, cp);
+      vthread_t child = vthread_new(cp->cptr2, cp->scope);
+      return do_callf_void(thr, child);
 }
 
 /*
@@ -4930,6 +4948,22 @@ bool of_REPLICATE(vthread_t thr, vvp_code_t cp)
       return true;
 }
 
+/*
+ * %ret/real <index>
+ */
+bool of_RET_REAL(vthread_t thr, vvp_code_t cp)
+{
+      size_t index = cp->number;
+      double val = thr->pop_real();
+
+      assert(index >= 0 && index < thr->args_real.size());
+      unsigned depth = thr->args_real[index];
+	// Use the depth to put the value into the stack of
+	// the parent thread.
+      thr->parent->poke_real(depth, val);
+      return true;
+}
+
 bool of_SCOPY(vthread_t thr, vvp_code_t)
 {
       vvp_object_t tmp;