diff --git a/amalgamate.py b/amalgamate.py
index 9775e061..2f8c6e41 100644
--- a/amalgamate.py
+++ b/amalgamate.py
@@ -74,4 +74,7 @@ if os.path.exists("plugins/unity/PocketPyUnityPlugin"):
 if os.path.exists("plugins/godot/godot-cpp/pocketpy"):
 	shutil.copy("amalgamated/pocketpy.h", "plugins/godot/godot-cpp/pocketpy/src/pocketpy.h")
 
+if os.path.exists("/mnt/e/PainterEngine/project/pocketpy.h"):
+	shutil.copy("amalgamated/pocketpy.h", "/mnt/e/PainterEngine/project/pocketpy.h")
+
 print("amalgamated/pocketpy.h")
\ No newline at end of file
diff --git a/src/obj.h b/src/obj.h
index 6e16fe5f..519ccd76 100644
--- a/src/obj.h
+++ b/src/obj.h
@@ -105,6 +105,10 @@ struct PyObject{
         _attr->~NameDict();
         pool64.dealloc(_attr);
     }
+
+    void enable_instance_dict(float lf=kInstAttrLoadFactor) noexcept {
+        _attr = new(pool64.alloc<NameDict>()) NameDict(lf);
+    }
 };
 
 template<typename T>
diff --git a/src/tuplelist.h b/src/tuplelist.h
index 80779d7f..0c8e47f7 100644
--- a/src/tuplelist.h
+++ b/src/tuplelist.h
@@ -38,6 +38,9 @@ public:
         for(PyObject* p : list) _args[i++] = p;
     }
 
+    // TODO: poor performance
+    // List is allocated by pool128 while tuple is by pool64
+    // ...
     Tuple(List&& other) noexcept : Tuple(other.size()){
         for(int i=0; i<_size; i++) _args[i] = other[i];
         other.clear();
diff --git a/src/vm.h b/src/vm.h
index 38b3cc5a..d51040a0 100644
--- a/src/vm.h
+++ b/src/vm.h
@@ -689,7 +689,13 @@ inline PyObject* VM::_py_call(PyObject* callable, ArgsView args, ArgsView kwargs
 
     int i = 0;
     if(args.size() < fn.decl->args.size()){
-        vm->TypeError(fmt("expected ", fn.decl->args.size(), " positional arguments, but got ", args.size()));
+        vm->TypeError(fmt(
+            "expected ",
+            fn.decl->args.size(),
+            " positional arguments, but got ",
+            args.size(),
+            " (", fn.decl->code->name, ')'
+        ));
     }
 
     // prepare args
@@ -711,7 +717,7 @@ inline PyObject* VM::_py_call(PyObject* callable, ArgsView args, ArgsView kwargs
                 break;
             }
         }
-        if(i < args.size()) TypeError("too many arguments");
+        if(i < args.size()) TypeError(fmt("too many arguments", " (", fn.decl->code->name, ')'));
     }
     
     for(int i=0; i<kwargs.size(); i+=2){
@@ -756,13 +762,17 @@ inline PyObject* VM::call(PyObject* callable, Args args, const Args& kwargs, boo
         PyObject* new_f = callable->attr().try_get(__new__);
         PyObject* obj;
         if(new_f != nullptr){
-            obj = call(new_f, std::move(args), kwargs, false);
+            // should not use std::move here, since we will reuse args in possible __init__
+            obj = call(new_f, args, kwargs, false);
+            if(!isinstance(obj, OBJ_GET(Type, callable))) return obj;
         }else{
             obj = heap.gcnew<DummyInstance>(OBJ_GET(Type, callable), {});
-            PyObject* self;
-            PyObject* init_f = get_unbound_method(obj, __init__, &self, false);
+        }
+        PyObject* self;
+        PyObject* init_f = get_unbound_method(obj, __init__, &self, false);
+        if (self != _py_null) {
             args.extend_self(self);
-            if (self != _py_null) call(init_f, std::move(args), kwargs, false);
+            call(init_f, std::move(args), kwargs, false);
         }
         return obj;
     }