Update threadmodule.c a bit

Will take some more work to actually free thread-locals since we don't expose that
I'm surprised we didn't actually free any of that before!

from_cpython/Include/pystate.h

@@ -144,7 +144,7 @@ PyAPI_FUNC(void) PyThreadState_DeleteCurrent(void) PYSTON_NOEXCEPT;
 
 PyAPI_FUNC(PyThreadState *) PyThreadState_Get(void) PYSTON_NOEXCEPT;
 PyAPI_FUNC(PyThreadState *) PyThreadState_Swap(PyThreadState *) PYSTON_NOEXCEPT;
-PyAPI_FUNC(PyObject *) PyThreadState_GetDict(void) PYSTON_NOEXCEPT;
+PyAPI_FUNC(BORROWED(PyObject *)) PyThreadState_GetDict(void) PYSTON_NOEXCEPT;
 PyAPI_FUNC(int) PyThreadState_SetAsyncExc(long, PyObject *) PYSTON_NOEXCEPT;
 
 

from_cpython/Modules/threadmodule.c

@@ -239,7 +239,7 @@ static PyTypeObject localdummytype = {
     /* tp_name           */ "_thread._localdummy",
     /* tp_basicsize      */ sizeof(localdummyobject),
     /* tp_itemsize       */ 0,
-    /* tp_dealloc        */ (destructor)/*localdummy_dealloc*/ NULL,
+    /* tp_dealloc        */ (destructor)localdummy_dealloc,
     /* tp_print          */ 0,
     /* tp_getattr        */ 0,
     /* tp_setattr        */ 0,
@@ -392,16 +392,13 @@ local_traverse(localobject *self, visitproc visit, void *arg)
 static int
 local_clear(localobject *self)
 {
-    // Pyston change:
-    Py_FatalError("unexpected call to local_clear()");
-    abort();
-#if 0
     PyThreadState *tstate;
     Py_CLEAR(self->args);
     Py_CLEAR(self->kw);
     Py_CLEAR(self->dummies);
     Py_CLEAR(self->wr_callback);
     /* Remove all strong references to dummies from the thread states */
+    /*
     if (self->key
         && (tstate = PyThreadState_Get())
         && tstate->interp) {
@@ -412,8 +409,9 @@ local_clear(localobject *self)
                 PyDict_GetItem(tstate->dict, self->key))
                 PyDict_DelItem(tstate->dict, self->key);
     }
+    */
+    printf("threadmodule.c:413: can't really free up the thread-specific dummy storage\n");
     return 0;
-#endif
 }
 
 static void
@@ -498,7 +496,7 @@ static PyTypeObject localtype = {
     /* tp_name           */ "thread._local",
     /* tp_basicsize      */ sizeof(localobject),
     /* tp_itemsize       */ 0,
-    /* tp_dealloc        */ (destructor)/*local_dealloc*/ NULL,
+    /* tp_dealloc        */ (destructor)local_dealloc,
     /* tp_print          */ 0,
     /* tp_getattr        */ 0,
     /* tp_setattr        */ 0,

src/runtime/capi.cpp

@@ -1296,7 +1296,7 @@ extern "C" char* Py_GetPythonHome(void) noexcept {
     return home;
 }
 
-extern "C" PyObject* PyThreadState_GetDict(void) noexcept {
+extern "C" BORROWED(PyObject*) PyThreadState_GetDict(void) noexcept {
     Box* dict = cur_thread_state.dict;
     if (!dict) {
         dict = cur_thread_state.dict = new BoxedDict();
@@ -1306,7 +1306,11 @@ extern "C" PyObject* PyThreadState_GetDict(void) noexcept {
 
 extern "C" void PyThreadState_Clear(PyThreadState* tstate) noexcept {
     assert(tstate == NULL);
+
     Py_CLEAR(cur_thread_state.dict);
+    Py_CLEAR(cur_thread_state.curexc_type);
+    Py_CLEAR(cur_thread_state.curexc_value);
+    Py_CLEAR(cur_thread_state.curexc_traceback);
 }
 
 extern "C" int _PyOS_URandom(void* buffer, Py_ssize_t size) noexcept {