bpo-40137: Micro-optimize _PyType_GetModuleByDef()

Make _PyType_GetModuleByDef() 2.3 ns faster.

Benchmark: 43.2 ns +- 0.7 ns -> 40.9 ns +- 1.0 ns: 1.05x faster

./python -m pyperf timeit \
    --duplicate=4096
    -s "from functools import lru_cache; f = lru_cache(lambda: 42)" \
    "f()"

Changes:

* _PyType_GetModuleByDef(): add fast-path for tp_mro[0] and use
  _PyType_HasFeature().
* Add a new pycore_moduleobject.h internal C API header file.
* Add _PyModule_GetDef() and _PyModule_GetState() which can be
  inlined without LTO and don't check invalid argument at runtime.
* Replace PyModule_GetState() with _PyModule_GetState() in _abc,
  _array, _operator, _pickle, _queue, _random, _struct and os
  extension modules. The _array, _queue and _struct extensions are
  now built with Py_BUILD_CORE_MODULE macro defined.

Author: vstinner

Include/internal/pycore_moduleobject.h

@@ -0,0 +1,34 @@
+#ifndef Py_INTERNAL_MODULEOBJECT_H
+#define Py_INTERNAL_MODULEOBJECT_H
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef Py_BUILD_CORE
+#  error "this header requires Py_BUILD_CORE define"
+#endif
+
+typedef struct {
+    PyObject_HEAD
+    PyObject *md_dict;
+    struct PyModuleDef *md_def;
+    void *md_state;
+    PyObject *md_weaklist;
+    PyObject *md_name;  /* for logging purposes after md_dict is cleared */
+} PyModuleObject;
+
+static inline PyModuleDef* _PyModule_GetDef(PyObject *m) {
+    assert(PyModule_Check(m));
+    return ((PyModuleObject *)m)->md_def;
+}
+
+static inline void* _PyModule_GetState(PyObject* m) {
+    assert(PyModule_Check(m));
+    return ((PyModuleObject *)m)->md_state;
+}
+
+
+#ifdef __cplusplus
+}
+#endif
+#endif /* !Py_INTERNAL_MODULEOBJECT_H */

Makefile.pre.in

@@ -1160,6 +1160,7 @@ PYTHON_HEADERS= \
 		$(srcdir)/Include/internal/pycore_interp.h \
 		$(srcdir)/Include/internal/pycore_list.h \
 		$(srcdir)/Include/internal/pycore_long.h \
+		$(srcdir)/Include/internal/pycore_moduleobject.h \
 		$(srcdir)/Include/internal/pycore_object.h \
 		$(srcdir)/Include/internal/pycore_pathconfig.h \
 		$(srcdir)/Include/internal/pycore_pyarena.h \

Misc/NEWS.d/next/Core and Builtins/2021-04-13-22-48-56.bpo-40137.Z77R2P.rst

@@ -0,0 +1,2 @@
+Micro-optimize _PyType_GetModuleByDef() to make it 2.3 ns faster (43.2 ns +-
+0.7 ns -> 40.9 ns +- 1.0 ns). Patch by Victor Stinner.

Modules/_abc.c

@@ -1,6 +1,7 @@
 /* ABCMeta implementation */
 
 #include "Python.h"
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "clinic/_abc.c.h"
 
 /*[clinic input]
@@ -27,7 +28,7 @@ typedef struct {
 static inline _abcmodule_state*
 get_abc_state(PyObject *module)
 {
-    void *state = PyModule_GetState(module);
+    void *state = _PyModule_GetState(module);
     assert(state != NULL);
     return (_abcmodule_state *)state;
 }

Modules/_functoolsmodule.c

@@ -1,5 +1,6 @@
 #include "Python.h"
 #include "pycore_long.h"          // _PyLong_GetZero()
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "pycore_object.h"        // _PyObject_GC_TRACK
 #include "pycore_pystate.h"       // _PyThreadState_GET()
 #include "pycore_tuple.h"         // _PyTuple_ITEMS()
@@ -35,7 +36,7 @@ typedef struct _functools_state {
 static inline _functools_state *
 get_functools_state(PyObject *module)
 {
-    void *state = PyModule_GetState(module);
+    void *state = _PyModule_GetState(module);
     assert(state != NULL);
     return (_functools_state *)state;
 }
@@ -52,8 +53,7 @@ get_functools_state_by_type(PyTypeObject *type)
     if (module == NULL) {
         return NULL;
     }
-    _functools_state *state = get_functools_state(module);
-    return state;
+    return get_functools_state(module);
 }
 
 static PyObject *

Modules/_operator.c

@@ -1,6 +1,6 @@
 
 #include "Python.h"
-
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "clinic/_operator.c.h"
 
 typedef struct {
@@ -12,7 +12,7 @@ typedef struct {
 static inline _operator_state*
 get_operator_state(PyObject *module)
 {
-    void *state = PyModule_GetState(module);
+    void *state = _PyModule_GetState(module);
     assert(state != NULL);
     return (_operator_state *)state;
 }

Modules/_pickle.c

@@ -9,6 +9,7 @@
 #endif
 
 #include "Python.h"
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "structmember.h"         // PyMemberDef
 
 PyDoc_STRVAR(pickle_module_doc,
@@ -182,7 +183,7 @@ static struct PyModuleDef _picklemodule;
 static PickleState *
 _Pickle_GetState(PyObject *module)
 {
-    return (PickleState *)PyModule_GetState(module);
+    return (PickleState *)_PyModule_GetState(module);
 }
 
 /* Find the module instance imported in the currently running sub-interpreter

Modules/_queuemodule.c

@@ -1,4 +1,5 @@
 #include "Python.h"
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "structmember.h"         // PyMemberDef
 #include <stddef.h>               // offsetof()
 
@@ -10,7 +11,7 @@ typedef struct {
 static simplequeue_state *
 simplequeue_get_state(PyObject *module)
 {
-    simplequeue_state *state = PyModule_GetState(module);
+    simplequeue_state *state = _PyModule_GetState(module);
     assert(state);
     return state;
 }

Modules/_randommodule.c

@@ -67,6 +67,7 @@
 /* ---------------------------------------------------------------*/
 
 #include "Python.h"
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #ifdef HAVE_PROCESS_H
 #  include <process.h>            // getpid()
 #endif
@@ -86,7 +87,7 @@ typedef struct {
 static inline _randomstate*
 get_random_state(PyObject *module)
 {
-    void *state = PyModule_GetState(module);
+    void *state = _PyModule_GetState(module);
     assert(state != NULL);
     return (_randomstate *)state;
 }
@@ -538,7 +539,7 @@ random_new(PyTypeObject *type, PyObject *args, PyObject *kwds)
 
     if (PyTuple_GET_SIZE(args) == 1)
         arg = PyTuple_GET_ITEM(args, 0);
-    
+
     tmp = random_seed(self, arg);
     if (tmp == NULL) {
         Py_DECREF(self);

Modules/_struct.c

@@ -6,6 +6,7 @@
 #define PY_SSIZE_T_CLEAN
 
 #include "Python.h"
+#include "pycore_moduleobject.h"  // _PyModule_GetState()
 #include "structmember.h"         // PyMemberDef
 #include <ctype.h>
 
@@ -24,7 +25,7 @@ typedef struct {
 static inline _structmodulestate*
 get_struct_state(PyObject *module)
 {
-    void *state = PyModule_GetState(module);
+    void *state = _PyModule_GetState(module);
     assert(state != NULL);
     return (_structmodulestate *)state;
 }