diff --git a/src/if_python.c b/src/if_python.c
index 8cc50c3..08d67d6 100644
--- a/src/if_python.c
+++ b/src/if_python.c
@@ -21,6 +21,15 @@
 
 #include <limits.h>
 
+/* uncomment this if used with the debug version of python.
+ * Checked on 2.7.4. */
+/* #define Py_DEBUG */
+/* Note: most of time you can add -DPy_DEBUG to CFLAGS in place of uncommenting 
+ */
+/* uncomment this if used with the debug version of python, but without its 
+ * allocator */
+/* #define Py_DEBUG_NO_PYMALLOC */
+
 /* Python.h defines _POSIX_THREADS itself (if needed) */
 #ifdef _POSIX_THREADS
 # undef _POSIX_THREADS
@@ -240,8 +249,18 @@
 #  define PyType_IsSubtype dll_PyType_IsSubtype
 # endif
 # if defined(PY_VERSION_HEX) && PY_VERSION_HEX >= 0x02030000
-#  define PyObject_Malloc dll_PyObject_Malloc
-#  define PyObject_Free dll_PyObject_Free
+#  ifdef Py_DEBUG
+#   define _Py_NegativeRefcount dll__Py_NegativeRefcount
+#   define _Py_RefTotal (*dll__Py_RefTotal)
+#   define _Py_Dealloc dll__Py_Dealloc
+#  endif
+#  if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
+#   define _PyObject_DebugMalloc dll__PyObject_DebugMalloc
+#   define _PyObject_DebugFree dll__PyObject_DebugFree
+#  else
+#   define PyObject_Malloc dll_PyObject_Malloc
+#   define PyObject_Free dll_PyObject_Free
+#  endif
 # endif
 # ifdef PY_USE_CAPSULE
 #  define PyCapsule_New dll_PyCapsule_New
@@ -350,8 +369,18 @@
 static int (*dll_PyType_IsSubtype)(PyTypeObject *, PyTypeObject *);
 # endif
 # if defined(PY_VERSION_HEX) && PY_VERSION_HEX >= 0x02030000
+#  ifdef Py_DEBUG
+static void (*dll__Py_NegativeRefcount)(const char *fname, int lineno, PyObject *op);
+static Py_ssize_t* dll__Py_RefTotal;
+static void (*dll__Py_Dealloc)(PyObject *obj);
+#  endif
+#  if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
+static void (*dll__PyObject_DebugFree)(void*);
+static void* (*dll__PyObject_DebugMalloc)(size_t);
+#  else
 static void* (*dll_PyObject_Malloc)(size_t);
 static void (*dll_PyObject_Free)(void*);
+#  endif
 # endif
 # ifdef PY_USE_CAPSULE
 static PyObject* (*dll_PyCapsule_New)(void *, char *, PyCapsule_Destructor);
@@ -469,12 +498,6 @@
     {"PyType_Type", (PYTHON_PROC*)&dll_PyType_Type},
     {"PyType_Ready", (PYTHON_PROC*)&dll_PyType_Ready},
     {"Py_FindMethod", (PYTHON_PROC*)&dll_Py_FindMethod},
-# if defined(PY_VERSION_HEX) && PY_VERSION_HEX >= 0x02050000 \
-	&& SIZEOF_SIZE_T != SIZEOF_INT
-    {"Py_InitModule4_64", (PYTHON_PROC*)&dll_Py_InitModule4},
-# else
-    {"Py_InitModule4", (PYTHON_PROC*)&dll_Py_InitModule4},
-# endif
     {"Py_SetPythonHome", (PYTHON_PROC*)&dll_Py_SetPythonHome},
     {"Py_Initialize", (PYTHON_PROC*)&dll_Py_Initialize},
     {"Py_Finalize", (PYTHON_PROC*)&dll_Py_Finalize},
@@ -496,8 +519,32 @@
     {"PyType_IsSubtype", (PYTHON_PROC*)&dll_PyType_IsSubtype},
 # endif
 # if defined(PY_VERSION_HEX) && PY_VERSION_HEX >= 0x02030000
+#  ifdef Py_DEBUG
+    {"_Py_NegativeRefcount", (PYTHON_PROC*)&dll__Py_NegativeRefcount},
+    {"_Py_RefTotal", (PYTHON_PROC*)&dll__Py_RefTotal},
+    {"_Py_Dealloc", (PYTHON_PROC*)&dll__Py_Dealloc},
+#  endif
+#  if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
+    {"_PyObject_DebugFree", (PYTHON_PROC*)&dll__PyObject_DebugFree},
+    {"_PyObject_DebugMalloc", (PYTHON_PROC*)&dll__PyObject_DebugMalloc},
+#  else
     {"PyObject_Malloc", (PYTHON_PROC*)&dll_PyObject_Malloc},
     {"PyObject_Free", (PYTHON_PROC*)&dll_PyObject_Free},
+#  endif
+# endif
+# if defined(PY_VERSION_HEX) && PY_VERSION_HEX >= 0x02050000 \
+	&& SIZEOF_SIZE_T != SIZEOF_INT
+#  ifdef Py_DEBUG
+    {"Py_InitModule4TraceRefs_64", (PYTHON_PROC*)&dll_Py_InitModule4},
+#  else
+    {"Py_InitModule4_64", (PYTHON_PROC*)&dll_Py_InitModule4},
+#  endif
+# else
+#  ifdef Py_DEBUG
+    {"Py_InitModule4TraceRefs", (PYTHON_PROC*)&dll_Py_InitModule4},
+#  else
+    {"Py_InitModule4", (PYTHON_PROC*)&dll_Py_InitModule4},
+#  endif
 # endif
 # ifdef PY_USE_CAPSULE
     {"PyCapsule_New", (PYTHON_PROC*)&dll_PyCapsule_New},
diff --git a/src/if_python3.c b/src/if_python3.c
index 2ae876e..dfc558a 100644
--- a/src/if_python3.c
+++ b/src/if_python3.c
@@ -24,6 +24,11 @@
 
 /* uncomment this if used with the debug version of python */
 /* #define Py_DEBUG */
+/* Note: most of time you can add -DPy_DEBUG to CFLAGS in place of uncommenting 
+ */
+/* uncomment this if used with the debug version of python, but without its 
+ * allocator */
+/* #define Py_DEBUG_NO_PYMALLOC */
 
 #include "vim.h"
 
@@ -207,6 +212,11 @@
 #  define _Py_NegativeRefcount py3__Py_NegativeRefcount
 #  define _Py_RefTotal (*py3__Py_RefTotal)
 #  define _Py_Dealloc py3__Py_Dealloc
+#  define PyModule_Create2TraceRefs py3_PyModule_Create2TraceRefs
+# else
+#  define PyModule_Create2 py3_PyModule_Create2
+# endif
+# if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
 #  define _PyObject_DebugMalloc py3__PyObject_DebugMalloc
 #  define _PyObject_DebugFree py3__PyObject_DebugFree
 # else
@@ -218,7 +228,6 @@
 # define PyObject_GC_UnTrack py3_PyObject_GC_UnTrack
 # define PyType_GenericAlloc py3_PyType_GenericAlloc
 # define PyType_GenericNew py3_PyType_GenericNew
-# define PyModule_Create2 py3_PyModule_Create2
 # undef PyUnicode_FromString
 # define PyUnicode_FromString py3_PyUnicode_FromString
 # undef PyUnicode_Decode
@@ -227,7 +236,7 @@
 # define PyCapsule_New py3_PyCapsule_New
 # define PyCapsule_GetPointer py3_PyCapsule_GetPointer
 
-# ifdef Py_DEBUG
+# if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
 #  undef PyObject_NEW
 #  define PyObject_NEW(type, typeobj) \
 ( (type *) PyObject_Init( \
@@ -317,7 +326,6 @@
 static PyObject* (*py3_PyFloat_FromDouble)(double num);
 static double (*py3_PyFloat_AsDouble)(PyObject *);
 static PyObject* (*py3_PyObject_GenericGetAttr)(PyObject *obj, PyObject *name);
-static PyObject* (*py3_PyModule_Create2)(struct PyModuleDef* module, int module_api_version);
 static PyObject* (*py3_PyType_GenericAlloc)(PyTypeObject *type, Py_ssize_t nitems);
 static PyObject* (*py3_PyType_GenericNew)(PyTypeObject *type, PyObject *args, PyObject *kwds);
 static PyTypeObject* py3_PyType_Type;
@@ -328,14 +336,19 @@
 static PyObject* (*py3_PyCapsule_New)(void *, char *, PyCapsule_Destructor);
 static void* (*py3_PyCapsule_GetPointer)(PyObject *, char *);
 # ifdef Py_DEBUG
-    static void (*py3__Py_NegativeRefcount)(const char *fname, int lineno, PyObject *op);
-    static Py_ssize_t* py3__Py_RefTotal;
-    static void (*py3__Py_Dealloc)(PyObject *obj);
-    static void (*py3__PyObject_DebugFree)(void*);
-    static void* (*py3__PyObject_DebugMalloc)(size_t);
+static void (*py3__Py_NegativeRefcount)(const char *fname, int lineno, PyObject *op);
+static Py_ssize_t* py3__Py_RefTotal;
+static void (*py3__Py_Dealloc)(PyObject *obj);
+static PyObject* (*py3_PyModule_Create2TraceRefs)(struct PyModuleDef* module, int module_api_version);
 # else
-    static void (*py3_PyObject_Free)(void*);
-    static void* (*py3_PyObject_Malloc)(size_t);
+static PyObject* (*py3_PyModule_Create2)(struct PyModuleDef* module, int module_api_version);
+# endif
+# if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
+static void (*py3__PyObject_DebugFree)(void*);
+static void* (*py3__PyObject_DebugMalloc)(size_t);
+# else
+static void (*py3_PyObject_Free)(void*);
+static void* (*py3_PyObject_Malloc)(size_t);
 # endif
 static PyObject*(*py3__PyObject_GC_New)(PyTypeObject *);
 static void(*py3_PyObject_GC_Del)(void *);
@@ -451,7 +464,6 @@
     {"PyFloat_FromDouble", (PYTHON_PROC*)&py3_PyFloat_FromDouble},
     {"PyFloat_AsDouble", (PYTHON_PROC*)&py3_PyFloat_AsDouble},
     {"PyObject_GenericGetAttr", (PYTHON_PROC*)&py3_PyObject_GenericGetAttr},
-    {"PyModule_Create2", (PYTHON_PROC*)&py3_PyModule_Create2},
     {"PyType_GenericAlloc", (PYTHON_PROC*)&py3_PyType_GenericAlloc},
     {"PyType_GenericNew", (PYTHON_PROC*)&py3_PyType_GenericNew},
     {"PyType_Type", (PYTHON_PROC*)&py3_PyType_Type},
@@ -463,6 +475,11 @@
     {"_Py_NegativeRefcount", (PYTHON_PROC*)&py3__Py_NegativeRefcount},
     {"_Py_RefTotal", (PYTHON_PROC*)&py3__Py_RefTotal},
     {"_Py_Dealloc", (PYTHON_PROC*)&py3__Py_Dealloc},
+    {"PyModule_Create2TraceRefs", (PYTHON_PROC*)&py3_PyModule_Create2TraceRefs},
+# else
+    {"PyModule_Create2", (PYTHON_PROC*)&py3_PyModule_Create2},
+# endif
+# if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
     {"_PyObject_DebugFree", (PYTHON_PROC*)&py3__PyObject_DebugFree},
     {"_PyObject_DebugMalloc", (PYTHON_PROC*)&py3__PyObject_DebugMalloc},
 # else
@@ -656,7 +673,7 @@
     static void
 call_PyObject_Free(void *p)
 {
-#ifdef Py_DEBUG
+#if defined(Py_DEBUG) && !defined(Py_DEBUG_NO_PYMALLOC)
     _PyObject_DebugFree(p);
 #else
     PyObject_Free(p);
diff --git a/src/version.c b/src/version.c
index b9db33d..62d0bab 100644
--- a/src/version.c
+++ b/src/version.c
@@ -729,6 +729,8 @@
 static int included_patches[] =
 {   /* Add new patch number below this line */
 /**/
+    1040,
+/**/
     1039,
 /**/
     1038,