LRC bug

[TobiasWrigstad]

The following code gets a crazy high LRC:

class Obj: pass

r = Region()
r # Region(lrc=1, osc=0, name=None, is_open=no) --- as expected

r.f = Obj()
r # Region(lrc=6, osc=0, name=None, is_open=yes)

I cannot fathom why lrc=6.

But things get weirder!

r.f = Obj()
r # Region(lrc=5, osc=0, name=None, is_open=yes) --- why LRC==5???
r.f = Obj()
r.f = Obj()
r.f = Obj()
r.f = Obj()
r.f = Obj()
r.f = Obj()
r # Region(lrc=-1, osc=0, name=None, is_open=yes) --- why LRF==-1???

(On phase3 branch.)

[TobiasWrigstad]

Also weird:

[TobiasWrigstad]

OK the above seems explainable:

r = Region()
r.try_close()

returns True but

r = Region()
r
r.try_close()

returns False because _ == r. Also:

r = Region()
r
42 # so that _ == 42
r.try_close()

returns True.

[xFrednet]

I'll investigate this next week, but I have a theory where the two effects could come from:

1. The high LRC count might come from objects pulled into the region from the Obj() class. Mainly the prototype, and meta type which should have a relatively high RCs leading to a high LRC. You could check for this by creating a second region, moving an Obj() into that one and taking notes of which objects can't be shared.

2. The negative LRC is potentially an effect of the missing write barrier on objects. When the Obj prototype is inside the region and a new Obj() is created it should increase the LRC. The write barrier will then decrease the LRC when the Obj() is slurped into the region. My guess is that only the decrease happens, you can that this with this code:

   class Obj: pass
   
   r = Region()
   r.f = Obj()
   
   # Take note of LRC
   str(r)
   
   # Create a new local object, which should increase the LRC
   o = Obj()
   
   # Check if the LRC actually increased?
   str(r)

Guess was correct, the object init function only sets the type without a write barrier:

cpython/Include/internal/pycore_object.h

Lines 175 to 184 in 6945ae8

	static inline void
	_PyObject_Init(PyObject *op, PyTypeObject *typeobj)
	{
	assert(op != NULL);
	Py_SET_TYPE(op, typeobj);
	if (_PyType_HasFeature(typeobj, Py_TPFLAGS_HEAPTYPE)) {
	Py_INCREF(typeobj);
	}
	_Py_NewReference(op);
	}

[xFrednet]

The high LRC at the beginning doesn't seem to be a bug but a result of the unicode type being pulled into the region:

[xFrednet]

OK the above seems explainable:

r = Region()
r.try_close()

returns True but

r = Region()
r
r.try_close()

returns False because _ == r. Also:

r = Region()
r
42 # so that _ == 42
r.try_close()

returns True.

Correct, there is still a bug that the binding of the region to the _ didn't increase the LRC.

The relevant code snippet is here:

cpython/Python/sysmodule.c

Lines 750 to 751 in 78cedf2

	if (PyObject_SetAttr(builtins, &_Py_ID(_), o) != 0)
	return NULL;