Fix free threading.

Author: ZeroIntensity

Include/internal/pycore_pystate.h

@@ -113,11 +113,12 @@ PyAPI_FUNC(PyThreadState *) _PyThreadState_GetCurrent(void);
 
    See also PyThreadState_Get() and PyThreadState_GetUnchecked(). */
 
-#ifdef Py_DEBUG
+static PyThreadState*
 // Forcing this to be non-inlined is nice for finding where this is called often.
-static Py_NO_INLINE PyThreadState*
+#ifdef Py_DEBUG
+_Py_MAYBE_UNUSED Py_NO_INLINE
 #else
-static inline PyThreadState*
+inline
 #endif
 _PyThreadState_GET(void)
 {
@@ -211,11 +212,11 @@ _Py_EnsureFuncTstateNotNULL(const char *func, PyThreadState *tstate)
 
    See also PyInterpreterState_Get()
    and _PyGILState_GetInterpreterStateUnsafe(). */
+static _Py_MAYBE_UNUSED PyInterpreterState*
 #ifdef Py_DEBUG
-// Forcing this to be non-inlined is nice for finding where this is called often.
-static Py_NO_INLINE PyInterpreterState*
+Py_NO_INLINE
 #else
-static inline PyInterpreterState*
+inline
 #endif
 _PyInterpreterState_GET(void) {
 #ifdef Py_DEBUG

Include/pyport.h

@@ -667,5 +667,13 @@ extern "C" {
 #  define _Py_STACK_GROWS_DOWN 1
 #endif
 
+#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 202311L
+// C23 standard attribute
+#  define _Py_MAYBE_UNUSED [[maybe_unused]]
+#elif _Py__has_attribute(__unused__)
+#  define _Py_MAYBE_UNUSED __attribute__((__unused__))
+#else
+#  define _Py_MAYBE_UNUSED
+#endif
 
 #endif /* Py_PYPORT_H */

Include/refcount.h

@@ -314,14 +314,17 @@ static inline Py_ALWAYS_INLINE void Py_INCREF(PyObject *op)
 #if !defined(Py_LIMITED_API)
 #if defined(Py_GIL_DISABLED)
 // Implements Py_DECREF on objects not owned by the current thread.
-PyAPI_FUNC(void) _Py_DecRefShared(PyThreadState *, PyObject *);
-PyAPI_FUNC(void) _Py_DecRefSharedDebug(PyThreadState *, PyObject *, const char *, int);
+PyAPI_FUNC(void) _Py_DecRefShared(PyObject *);
+PyAPI_FUNC(void) _Py_DecRefSharedDebug(PyObject *, const char *, int);
+PyAPI_FUNC(void) _Py_DecRefSharedTstate(PyThreadState *, PyObject *);
+PyAPI_FUNC(void) _Py_DecRefSharedDebugTstate(PyThreadState *, PyObject *, const char *, int);
 
 // Called from Py_DECREF by the owning thread when the local refcount reaches
 // zero. The call will deallocate the object if the shared refcount is also
 // zero. Otherwise, the thread gives up ownership and merges the reference
 // count fields.
-PyAPI_FUNC(void) _Py_MergeZeroLocalRefcount(PyThreadState *, PyObject *);
+PyAPI_FUNC(void) _Py_MergeZeroLocalRefcount(PyObject *);
+PyAPI_FUNC(void) _Py_MergeZeroLocalRefcountTstate(PyThreadState *, PyObject *);
 #endif  // Py_GIL_DISABLED
 #endif  // Py_LIMITED_API
 
@@ -385,7 +388,7 @@ static inline void Py_DECREF(PyObject *op)
         }
     }
     else {
-        _Py_DecRefShared(_PyThreadState_GET(), op);
+        _Py_DecRefShared(op);
     }
 }
 #define Py_DECREF(op) Py_DECREF(_PyObject_CAST(op))
@@ -402,11 +405,11 @@ static inline void _Py_DECREF(PyThreadState *tstate, PyObject *op)
         local--;
         _Py_atomic_store_uint32_relaxed(&op->ob_ref_local, local);
         if (local == 0) {
-            _Py_MergeZeroLocalRefcount(tstate, op);
+            _Py_MergeZeroLocalRefcountTstate(tstate, op);
         }
     }
     else {
-        _Py_DecRefShared(tstate, op);
+        _Py_DecRefSharedTstate(tstate, op);
     }
 }
 

Objects/object.c

@@ -420,21 +420,33 @@ _Py_DecRefSharedIsDead(PyObject *o, const char *filename, int lineno)
 }
 
 void
-_Py_DecRefSharedDebug(PyThreadState *tstate, PyObject *o, const char *filename, int lineno)
+_Py_DecRefSharedDebugTstate(PyThreadState *tstate, PyObject *o, const char *filename, int lineno)
 {
     if (_Py_DecRefSharedIsDead(o, filename, lineno)) {
-        _Py_Dealloc(tstate, o);
+        _Py_DeallocTstate(tstate, o);
     }
 }
 
 void
-_Py_DecRefShared(PyThreadState *tstate, PyObject *o)
+_Py_DecRefSharedDebug(PyObject *o, const char *filename, int lineno)
 {
-    _Py_DecRefSharedDebug(tstate, o, NULL, 0);
+    _Py_DecRefSharedDebugTstate(_PyThreadState_GET(), o, filename, lineno);
 }
 
 void
-_Py_MergeZeroLocalRefcount(PyThreadState *tstate, PyObject *op)
+_Py_DecRefSharedTstate(PyThreadState *tstate, PyObject *o)
+{
+    _Py_DecRefSharedDebugTstate(tstate, o, NULL, 0);
+}
+
+void
+_Py_DecRefShared(PyObject *o)
+{
+    _Py_DecRefSharedTstate(_PyThreadState_GET(), o);
+}
+
+void
+_Py_MergeZeroLocalRefcountTstate(PyThreadState *tstate, PyObject *op)
 {
     assert(op->ob_ref_local == 0);
 
@@ -464,6 +476,12 @@ _Py_MergeZeroLocalRefcount(PyThreadState *tstate, PyObject *op)
     }
 }
 
+void
+_Py_MergeZeroLocalRefcount(PyObject *op)
+{
+    _Py_MergeZeroLocalRefcountTstate(_PyThreadState_GET(), op);
+}
+
 Py_ssize_t
 _Py_ExplicitMergeRefcount(PyObject *op, Py_ssize_t extra)
 {

Python/gc_free_threading.c

@@ -1,6 +1,7 @@
 // Cyclic garbage collector implementation for free-threaded build.
 #include "Python.h"
 #include "pycore_brc.h"           // struct _brc_thread_state
+#include "pycore_call.h"          // _PyObject_VectorcallTstate()
 #include "pycore_ceval.h"         // _Py_set_eval_breaker_bit()
 #include "pycore_dict.h"          // _PyInlineValuesSize()
 #include "pycore_frame.h"         // FRAME_CLEARED