gh-103951: enable optimization for fast attribute access on module subclasses

This patch relax specialization requirements from
``PyModule_CheckExact(op)`` to
``Py_TYPE(op)->tp_getattro != PyModule_Type.tp_getattro``.

Benchmarks:

```py
import pyperf
import b
import c
runner = pyperf.Runner()
runner.timeit('b.x', 'b.x', globals=globals())
runner.timeit('c.x', 'c.x', globals=globals())
```
```py
x = 1
```
```py
import sys, types
x = 1
class _Foo(types.ModuleType): pass
sys.modules[__name__].__class__ = _Foo
```

On the main:
```
$ python a.py -q
b.x: Mean +- std dev: 50.2 ns +- 2.7 ns
c.x: Mean +- std dev: 132 ns +- 7 ns
```

With the patch:
```
$ python a.py -q
b.x: Mean +- std dev: 52.9 ns +- 3.6 ns
c.x: Mean +- std dev: 52.6 ns +- 2.7 ns
```

Co-authored-by: Nicolas Tessore <[email protected]>

Author: skirpichev

Include/internal/pycore_moduleobject.h

@@ -16,6 +16,12 @@ extern int _PyModule_IsPossiblyShadowing(PyObject *);
 
 extern int _PyModule_IsExtension(PyObject *obj);
 
+static inline int
+_PyModule_HasSpecializedGetAttr(PyObject* op)
+{
+    return Py_TYPE(op)->tp_getattro != PyModule_Type.tp_getattro;
+}
+
 typedef struct {
     PyObject_HEAD
     PyObject *md_dict;

Misc/NEWS.d/next/Core_and_Builtins/2024-11-01-09-58-06.gh-issue-103951.6qduwj.rst

@@ -0,0 +1,2 @@
+Relax optimization requirements to allow fast attribute access to module
+subclasses.

Python/bytecodes.c

@@ -2108,7 +2108,7 @@ dummy_func(
 
         op(_CHECK_ATTR_MODULE, (dict_version/2, owner -- owner)) {
             PyObject *owner_o = PyStackRef_AsPyObjectBorrow(owner);
-            DEOPT_IF(!PyModule_CheckExact(owner_o));
+            DEOPT_IF(_PyModule_HasSpecializedGetAttr(owner_o));
             PyDictObject *dict = (PyDictObject *)((PyModuleObject *)owner_o)->md_dict;
             assert(dict != NULL);
             DEOPT_IF(dict->ma_keys->dk_version != dict_version);

Python/generated_cases.c.h

@@ -1143,7 +1143,7 @@ _Py_Specialize_LoadAttr(_PyStackRef owner_st, _Py_CODEUNIT *instr, PyObject *nam
         SPECIALIZATION_FAIL(LOAD_ATTR, SPEC_FAIL_OTHER);
         fail = true;
     }
-    else if (PyModule_CheckExact(owner)) {
+    else if (!_PyModule_HasSpecializedGetAttr(owner)) {
         fail = specialize_module_load_attr(owner, instr, name);
     }
     else if (PyType_Check(owner)) {

Python/specialize.c

@@ -598,6 +598,7 @@ def has_error_without_pop(op: parser.InstDef) -> bool:
     "_PyLong_Multiply",
     "_PyLong_Subtract",
     "_PyManagedDictPointer_IsValues",
+    "_PyModule_HasSpecializedGetAttr",
     "_PyObject_GC_IS_TRACKED",
     "_PyObject_GC_MAY_BE_TRACKED",
     "_PyObject_GC_TRACK",