We have a lot of stdlib code that looks like:
try:
nm_tpl.__module__ = sys._getframe(2).f_globals.get('__name__', '__main__')
except (AttributeError, ValueError):
pass
While technically it handles sys._getframe being missing, it would be nice to handle it better for this scenario.
I'm already using a sys._get_calling_module_name() as an internal patch to avoid exposing the _getframe() calls to Python code. (As I recall, it cleans up basically all of the uses apart from the traceback module.) This lets us treat sys._getframe() calls as suspicious, because now most code never uses it.
/*[clinic input]
sys._get_calling_module_name
Return the name of the calling module.
[clinic start generated code]*/
static PyObject *
sys__get_calling_module_name_impl(PyObject *module)
/*[clinic end generated code]*/
{
PyFrameObject *f = _PyThreadState_GET()->frame;
PyObject *r;
if (f == NULL) {
Py_RETURN_NONE;
}
f = f->f_back;
if (f == NULL) {
Py_RETURN_NONE;
}
r = _PyDict_GetItemIdWithError(f->f_globals, &PyId___name__);
if (!r) {
PyErr_Clear();
r = Py_None;
}
Py_INCREF(r);
return r;
}
For something that will live beyond a separate patch, it might make sense to add as much functionality as needed for the warning module, which currently skips some importlib frames and a caller-specified count. I wouldn't want to make it too much more complex though.
Separating this out would make it easier for other implementations to support enum, typing, collections, and any other modules that want the caller's name, even if they can't easily/efficiently support a full _getframe().
Thoughts?
|