gh-112529: Implement GC for free-threaded builds (#114262)

* gh-112529: Implement GC for free-threaded builds

This implements a mark and sweep GC for the free-threaded builds of
CPython. The implementation relies on mimalloc to find GC tracked
objects (i.e., "containers").

Python/gc.c

@@ -15,6 +15,8 @@
 #include "pycore_weakref.h"       // _PyWeakref_ClearRef()
 #include "pydtrace.h"
 
+#ifndef Py_GIL_DISABLED
+
 typedef struct _gc_runtime_state GCState;
 
 #ifdef Py_DEBUG
@@ -964,10 +966,10 @@ finalize_garbage(PyThreadState *tstate, PyGC_Head *collectable)
         PyGC_Head *gc = GC_NEXT(collectable);
         PyObject *op = FROM_GC(gc);
         gc_list_move(gc, &seen);
-        if (!_PyGCHead_FINALIZED(gc) &&
+        if (!_PyGC_FINALIZED(op) &&
             (finalize = Py_TYPE(op)->tp_finalize) != NULL)
         {
-            _PyGCHead_SET_FINALIZED(gc);
+            _PyGC_SET_FINALIZED(op);
             Py_INCREF(op);
             finalize(op);
             assert(!_PyErr_Occurred(tstate));
@@ -1942,3 +1944,5 @@ PyUnstable_GC_VisitObjects(gcvisitobjects_t callback, void *arg)
 done:
     gcstate->enabled = origenstate;
 }
+
+#endif  // Py_GIL_DISABLED

Python/object_stack.c

@@ -0,0 +1,87 @@
+// Stack of Python objects
+
+#include "Python.h"
+#include "pycore_freelist.h"
+#include "pycore_pystate.h"
+#include "pycore_object_stack.h"
+
+extern _PyObjectStackChunk *_PyObjectStackChunk_New(void);
+extern void _PyObjectStackChunk_Free(_PyObjectStackChunk *);
+
+static struct _Py_object_stack_state *
+get_state(void)
+{
+    _PyFreeListState *state = _PyFreeListState_GET();
+    return &state->object_stack_state;
+}
+
+_PyObjectStackChunk *
+_PyObjectStackChunk_New(void)
+{
+    _PyObjectStackChunk *buf;
+    struct _Py_object_stack_state *state = get_state();
+    if (state->numfree > 0) {
+        buf = state->free_list;
+        state->free_list = buf->prev;
+        state->numfree--;
+    }
+    else {
+        // NOTE: we use PyMem_RawMalloc() here because this is used by the GC
+        // during mimalloc heap traversal. In that context, it is not safe to
+        // allocate mimalloc memory, such as via PyMem_Malloc().
+        buf = PyMem_RawMalloc(sizeof(_PyObjectStackChunk));
+        if (buf == NULL) {
+            return NULL;
+        }
+    }
+    buf->prev = NULL;
+    buf->n = 0;
+    return buf;
+}
+
+void
+_PyObjectStackChunk_Free(_PyObjectStackChunk *buf)
+{
+    assert(buf->n == 0);
+    struct _Py_object_stack_state *state = get_state();
+    if (state->numfree >= 0 &&
+        state->numfree < _PyObjectStackChunk_MAXFREELIST)
+    {
+        buf->prev = state->free_list;
+        state->free_list = buf;
+        state->numfree++;
+    }
+    else {
+        PyMem_RawFree(buf);
+    }
+}
+
+void
+_PyObjectStack_Clear(_PyObjectStack *queue)
+{
+    while (queue->head != NULL) {
+        _PyObjectStackChunk *buf = queue->head;
+        buf->n = 0;
+        queue->head = buf->prev;
+        _PyObjectStackChunk_Free(buf);
+    }
+}
+
+void
+_PyObjectStackChunk_ClearFreeList(_PyFreeListState *free_lists, int is_finalization)
+{
+    if (!is_finalization) {
+        // Ignore requests to clear the free list during GC. We use object
+        // stacks during GC, so emptying the free-list is counterproductive.
+        return;
+    }
+
+    struct _Py_object_stack_state *state = &free_lists->object_stack_state;
+    while (state->numfree > 0) {
+        _PyObjectStackChunk *buf = state->free_list;
+        state->free_list = buf->prev;
+        state->numfree--;
+        PyMem_RawFree(buf);
+    }
+    state->numfree = -1;
+}

Python/pystate.c

@@ -10,6 +10,7 @@
 #include "pycore_frame.h"
 #include "pycore_initconfig.h"    // _PyStatus_OK()
 #include "pycore_object.h"        // _PyType_InitCache()
+#include "pycore_object_stack.h"  // _PyObjectStackChunk_ClearFreeList()
 #include "pycore_parking_lot.h"   // _PyParkingLot_AfterFork()
 #include "pycore_pyerrors.h"      // _PyErr_Clear()
 #include "pycore_pylifecycle.h"   // _PyAST_Fini()
@@ -1468,6 +1469,7 @@ _Py_ClearFreeLists(_PyFreeListState *state, int is_finalization)
     _PyList_ClearFreeList(state, is_finalization);
     _PyContext_ClearFreeList(state, is_finalization);
     _PyAsyncGen_ClearFreeLists(state, is_finalization);
+    _PyObjectStackChunk_ClearFreeList(state, is_finalization);
 }
 
 void
@@ -2055,7 +2057,6 @@ start_the_world(struct _stoptheworld_state *stw)
     HEAD_LOCK(runtime);
     stw->requested = 0;
     stw->world_stopped = 0;
-    stw->requester = NULL;
     // Switch threads back to the detached state.
     PyInterpreterState *i;
     PyThreadState *t;
@@ -2066,6 +2067,7 @@ start_the_world(struct _stoptheworld_state *stw)
             _PyParkingLot_UnparkAll(&t->state);
         }
     }
+    stw->requester = NULL;
     HEAD_UNLOCK(runtime);
     if (stw->is_global) {
         _PyRWMutex_Unlock(&runtime->stoptheworld_mutex);