Repository navigation
[BUG] 3.3.0 regression: dict[str, str].get(missing_key) is typed as str, so the returned None takes str fast paths (bool/len/in give wrong results) #7991
Description
Since 3.3.0, the declared item type of a dict[str, str] is used to type the result of .get(key) as str. When the key is missing, .get() returns None, but the generated code applies the str fast paths to it without a None check. Truthiness silently returns the wrong answer, and len() reads out of bounds.
Reproducer
# g.py
def truth(d: dict[str, str]):
v = d.get('missing')
return bool(v)
def length(d: dict[str, str]):
v = d.get('missing')
return len(v)
def contains(d: dict[str, str]):
v = d.get('missing')
return 'a' in vImportant
The bug affects not only missing keys but also None values:
def length(d: dict[str, str]):
v = d.get('none')
return len(v)
length({'none': None})# setup.py
from setuptools import setup, Extension
from Cython.Build import cythonize
setup(name='g', ext_modules=cythonize([Extension("g", ["g.py"])],
compiler_directives={"language_level": 3}))Called as f({'a': 'b'}), so the key is missing and .get() returns None:
| CPython | Cython 3.2.5 | Cython 3.3.0 | |
|---|---|---|---|
bool(v) |
False |
False |
True |
len(v) |
TypeError |
TypeError |
4294967295 |
'a' in v |
TypeError |
TypeError |
True |
v == 'a' and v + 'a' behave correctly.
Generated C
3.3.0, for bool(v) — no None check:
Py_ssize_t __pyx_temp = __Pyx_PyUnicode_IS_TRUE(__pyx_v_v);3.2.5, same line:
if (__pyx_v_v == Py_None) __pyx_t_1 = 0;
else {
Py_ssize_t __pyx_temp = __Pyx_PyUnicode_IS_TRUE(__pyx_v_v);
...
}The assignment itself explicitly permits None:
if (!(likely(PyUnicode_CheckExact(__pyx_t_1))||((__pyx_t_1) == Py_None) || __Pyx_RaiseUnexpectedTypeError("str", __pyx_t_1))) ...so nothing raises there — only the later fast paths assume a real str.
Notes
- Independent of
infer_types; reproduces with and without it. - Annotating
v: str | None = d.get('missing')does not help, and neither does passing the default explicitly asd.get('missing', None). - Dropping the item type (
d: dict) avoids it, as does an explicitv is not None and v. - Presumably comes from "Declared container item types are now used by the type system" (Allow Cython to understand subscripted types #7288, Introduce support of
tuple[TYP, ...]#7798).
Environment
Cython 3.3.0 (vs 3.2.5), CPython 3.12.3, Linux x86_64 (also observed on aarch64), gcc 15.
This hit us in production: a value read with .get() from a dict[str, str] | None passed an if v guard while being None, so None flowed into code that could not receive it.
Activity
scoder commentedon Sep 13, 2026
matusvalo commentedon Sep 13, 2026
i confirm the bug. But I am puzzled what is purpose of the test
__pyx_t_1 == Py_Nonein line:It should not guard againstUpdate: The condition just allows None and str values... Now it is clear for me the purpouse of the test...Nonevalue? And if it is not guarding what is the purpose of that expression?Also it is interesting that annotating variable instead it adds correct
Noneguard:Produces the same code just adds extra None guard:
Hence two points are unclear for me:
What is the purpose of the firstUpdate, now I understand that code it basically allowsPy_Noneequality teststrorNonevalues ...strvariable annotation adds second (correct) guard but inference to the same type (str) does not.matusvalo commentedon Sep 13, 2026
OK the fix was more simpler than I thought. Fixing PR is linked to this issue.
matusvalo commentedon Sep 15, 2026
I updated description with one additional case.
Fix handling default values in typed dictionaries (GH-8003)