PyQt6 on Python 3.14: pyqtSignal(..., object, ...).emit(..., None, ...) corrupts the immortal None refcount (not on 3.13)

Darren Powierski darren.powierski at tmeic.com.au
Sun Nov 15 20:17:39 GMT 2026


Thank you Phil.

The old PyQt6 6.11.0 reproduces the bug on 3.14.5 and the new 6.12.1 does not. Everything else was identical between the two runs, and the results were the same on all three repeats.
Setup	          Python	PyQt6	Qt	sip	None refcount delta	Result
Old	         3.14.5	6.11.0	6.11.0	6.15.1	+200,000	               BUG
New	         3.14.5	6.12.1	6.12.0	6.17.3	0	                             OK
Mitigation     3.13.16	6.11.0	6.11.0	6.15.2	0	                             OK
Each test does 100,000 emits of pyqtSignal(str, object, object) with two None arguments (the same repro_none_refcount.py as before). Old and new each ran 3 times.

-----Original Message-----
From: Darren Powierski <darren.powierski at tmeic.com.au> 
Sent: Monday, 9 November 2026 8:49 AM
To: Phil Thompson <phil at riverbankcomputing.com>
Cc: pyqt at riverbankcomputing.com
Subject: RE: PyQt6 on Python 3.14: pyqtSignal(..., object, ...).emit(..., None, ...) corrupts the immortal None refcount (not on 3.13)

Thank you Phil for your help with this.

-----Original Message-----
From: Phil Thompson <phil at riverbankcomputing.com>
Sent: Saturday, 7 November 2026 10:18 PM
To: Darren Powierski <darren.powierski at tmeic.com.au>
Cc: pyqt at riverbankcomputing.com
Subject: Re: PyQt6 on Python 3.14: pyqtSignal(..., object, ...).emit(..., None, ...) corrupts the immortal None refcount (not on 3.13)

Just to tie this off, it is a bug in Python - see...

https://www.riverbankcomputing.com/pipermail/pyqt/2026-November/046562.html
https://github.com/python/cpython/issues/158401

I will experiment with building using the v3.12 headers as a workaround.

Phil

On 26/10/2026 20:51, Darren Powierski wrote:
> Hi Phil,
> Thanks for the reply. I agree it needs to be shown that this is not a 
> CPython bug, so I ran more tests, all on one machine with embedded 
> python.org builds.
> 
> 1. PyQt6 6.11.0, Qt 6.11, 100,000 x pyqtSignal(str, object, 
> object).emit("k", None, None),
>    reading the None refcount with ctypes before and after:
> 	 Python 3.13.16 (sip 6.15.2): delta 0        (None stays immortal)
> 	 Python 3.14.5  (sip 6.15.2): delta +200,000
> 	 Python 3.15.2  (sip 6.17.0, PyQt6-sip 13.13.0): delta +200,000
>    So it is not specific to 3.14.x; 3.15 behaves the same. The delta 
> is exactly one per
>    None argument per emit.
> 
> 2. Same emit with PySide6 6.12.0 on 3.14.5 and 3.15.2: delta 0.
> 
> 3. A minimal C extension (no Qt, no sip) doing 1,000,000 balanced 
> Py_INCREF/Py_DECREF and
>    PyTuple_Pack on None, built both with Py_LIMITED_API 3.10 (as in 
> the cp310-abi3 PyQt6
>    wheel) and as a native build, MSVC: delta 0 on 3.14.5 and 3.15.2.
> 
> 4. A pure Python / ctypes control (calls, dicts, lists, 
> Py_IncRef/Py_DecRef on None):
>    delta 0 on 3.13, 3.14 and 3.15.
> 
> On the same interpreters, only the PyQt6/sip path changes the None 
> refcount. That suggests an unbalanced Py_DECREF of None (or a missed 
> immortality check) in the code that converts the object arguments of a 
> signal emit, rather than a CPython fault. I have not located the exact 
> line, so this is evidence, not proof. If it helps, I can build
> sip/PyQt6 from source
> with a trace in the object-argument conversion.
> 
> Why it matters: in a long-running application (hours) the drift 
> eventually produced process-wide TypeErrors in json.dumps and
> "__init__() should return None, not NoneType", which is why we run on 
> 3.13.
> 
> Reproducer and test sources (repro_none_refcount.py, 
> repro_none_refcount_pyside6.py, control_no_pyqt.py, nonechurn.c,
> runchurn.py) are available; I have added them below.
> 
> Environment: Windows, PyQt6 6.11.0, PyQt6-Qt6 6.11.x, Qt 6.11.0.
> 
> 
> 
> 
> repro_none_refcount.py
> 
> """Minimal repro: PyQt6 signal emit with None changes the refcount of 
> None on Python 3.14.
> 
> None is immortal on 3.12+, so its refcount should never change. On
> 3.14 it grows on
> pyqtSignal(..., object, ...).emit(..., None, ...). Same PyQt6 / Qt / 
> sip on 3.13: no change.
> """
> import ctypes
> import os
> import sys
> 
> os.environ.setdefault("QT_QPA_PLATFORM", "offscreen") from PyQt6 
> import QtCore, QtWidgets from PyQt6.sip import SIP_VERSION_STR
> 
> 
> def none_refcount():
>     return ctypes.c_uint64.from_address(id(None)).value
> 
> 
> class W(QtCore.QObject):
>     sig = QtCore.pyqtSignal(str, object, object)
> 
>     @QtCore.pyqtSlot(str, object, object)
>     def slot(self, a, b, c):
>         pass
> 
> 
> app = QtWidgets.QApplication([])
> w = W()
> w.sig.connect(w.slot)
> 
> N = 100000
> before = none_refcount()
> for _ in range(N):
>     w.sig.emit("k", None, None)
> after = none_refcount()
> 
> print("python", sys.version.split()[0], "| PyQt6", 
> QtCore.PYQT_VERSION_STR, "| Qt", QtCore.QT_VERSION_STR,
>       "| sip", SIP_VERSION_STR)
> print("None refcount before: %#x  after: %#x  delta: %d (2 None args x 
> %d emits)" % (before, after, after - before, N)) print("RESULT:", "OK 
> (unchanged)" if after == before else "BUG (None refcount changed)")
> 
> 
> repro_none_refcount_pyside6.py
> 
> """Cross-check: same None-through-signal emit using PySide6 instead of 
> PyQt6."""
> import ctypes
> import os
> import sys
> 
> os.environ.setdefault("QT_QPA_PLATFORM", "offscreen") from PySide6 
> import QtCore, QtWidgets import PySide6
> 
> 
> def none_refcount():
>     return ctypes.c_uint64.from_address(id(None)).value
> 
> 
> class W(QtCore.QObject):
>     sig = QtCore.Signal(str, object, object)
> 
>     @QtCore.Slot(str, object, object)
>     def slot(self, a, b, c):
>         pass
> 
> 
> app = QtWidgets.QApplication([])
> w = W()
> w.sig.connect(w.slot)
> N = 100000
> before = none_refcount()
> for _ in range(N):
>     w.sig.emit("k", None, None)
> after = none_refcount()
> print("python", sys.version.split()[0], "| PySide6",
> PySide6.__version__)
> print("None refcount before: %#x after: %#x delta: %d" % (before, 
> after, after - before)) print("RESULT:", "OK (unchanged)" if after == 
> before else "BUG (None refcount changed)")
> 
> 
> 
> 
> control_no_pyqt.py
> 
> """Control: pure Python / ctypes, no PyQt. Does None's refcount 
> change?"""
> import ctypes
> import sys
> 
> 
> def rc():
>     return ctypes.c_uint64.from_address(id(None)).value
> 
> 
> def f(a, b, c):
>     return (a, b, c)
> 
> 
> b = rc()
> for _ in range(200000):
>     f("k", None, None)
>     d = {"x": None}
>     l = [None, None]
>     ctypes.pythonapi.Py_IncRef(ctypes.py_object(None))
>     ctypes.pythonapi.Py_DecRef(ctypes.py_object(None))
> a = rc()
> print("python", sys.version.split()[0], "before %#x after %#x delta 
> %d" % (b, a, a - b))
> 
> 
> 
> 
> 
> nonechurn.c
> 
> /* Minimal extension: balanced INCREF/DECREF on None, no Qt. Built 
> twice from this
>    source: -DUSE_ABI3 (Py_LIMITED_API 3.10, like the PyQt6 cp310-abi3
> wheel) and native. */
> #ifdef USE_ABI3
> #define Py_LIMITED_API 0x030A0000
> #endif
> #include <Python.h>
> 
> static PyObject *churn(PyObject *self, PyObject *args) {
> 	PyObject *obj;
> 	long n;
> 	if (!PyArg_ParseTuple(args, "Ol", &obj, &n))
> 		return NULL;
> 	for (long i = 0; i < n; i++) {
> 		PyObject *t;
> 		Py_INCREF(obj);
> 		Py_DECREF(obj);
> 		Py_XINCREF(obj);
> 		Py_XDECREF(obj);
> 		t = PyTuple_Pack(2, obj, obj);
> 		Py_XDECREF(t);
> 	}
> 	Py_RETURN_NONE;
> }
> 
> static PyMethodDef M[] = {{"churn", churn, METH_VARARGS, ""}, {NULL, 
> NULL, 0, NULL}}; static struct PyModuleDef D = {PyModuleDef_HEAD_INIT, 
> "nonechurn", NULL, -1, M}; PyMODINIT_FUNC PyInit_nonechurn(void) { 
> return PyModule_Create(&D); }
> 
> 
> 
> 
> runchurn.py
> 
> import ctypes
> import sys
> sys.path.insert(0, sys.argv[1])
> import nonechurn
> 
> rc = lambda: ctypes.c_uint64.from_address(id(None)).value
> b = rc()
> nonechurn.churn(None, 1000000)
> a = rc()
> print("python", sys.version.split()[0], sys.argv[1].split("\\")[-1], 
> "delta", a - b)
> 
> 
> Thank you,
> Darren
> 
> 
> 
> 
> -----Original Message-----
> From: Phil Thompson <phil at riverbankcomputing.com>
> Sent: Monday, 26 October 2026 8:26 PM
> To: Darren Powierski <darren.powierski at tmeic.com.au>
> Cc: pyqt at riverbankcomputing.com
> Subject: Re: PyQt6 on Python 3.14: pyqtSignal(..., object, 
> ...).emit(..., None, ...) corrupts the immortal None refcount (not on
> 3.13)
> 
> On 25/10/2026 23:06, Darren Powierski wrote:
>> Hi Phil,
>> 
>> 
>> 
>> I have added a test description of the issue. I am wondering if you 
>> can help understand the issue.
> 
> First you have to convince me that this isn't a bug in Python. What 
> results do you get with v3.15?
> 
> Phil


More information about the PyQt mailing list