New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[C API] Remove _Py_NewReference() and _Py_ForgetReference() from the public C API #85161
Comments
The _Py_NewReference() and _Py_ForgetReference() functions are tightly coupled to CPython internals. _Py_NewReference() is only exposed because it used by the PyObject_INIT() macro which is the fast inlined flavor of PyObject_Init(). If we make PyObject_INIT() as alias to PyObject_Init(), as already done for the limited C API, _Py_NewReference() can be removed from the public C API (moved to the internal C API). _Py_ForgetReference() function is only defined if Py_TRACE_REFS macro is defined. I propose to also removed it from the public C API (move it to the internal C API). In the CPython code base, _Py_NewReference() is used:
These are corner cases which can be avoided in third party C extension modules. |
Work on this issue started in Python 3.9 with the following change of bpo-39542: commit f58bd7c
|
See "Removal of _Py_ForgetReference from public header in 3.9 issue" discussion on capi-sig: |
I found a similar code pattern outside CPython code base. Like Nuitka generic macros for free lists: Another example in the old unmaintained gmpy project (last release in 2013): These functions can opt-in for the internal C API to continue to get access to _Py_NewReference(). Or maybe free lists should be implemented differently? (is it possible?)
Third-party code can use the public PyUnicode_Resize() function and the private _PyBytes_Resize() function. Later, if needed, we can consider to expose _PyBytes_Resize() in the limited C API.
I found a few projects which basically reimplement the PEP-442 logic in their tp_dealloc function. Example with pyuv resurrect_object() function: For these, using PEP-442 tp_finalizer here would avoid relying on private functions like _Py_NewReference(), make the code simpler and more reliable. |
To add to the list, Cython also calls _Py_NewReference() in its async generator code, which uses a free-list. That code was mostly copied from the CPython-internal implementation. Other freelist implementations in Cython call PyObject_INIT() instead, so I guess calling _Py_NewReference() directly here isn't actually required and could be avoided. |
I looked into the freelists a bit more and (as always) it's not quite that simple. Calling _Py_NewReference() allows keeping the "ob_type" pointer in place, whereas PyObject_INIT() requires a blank object in order to set the type and properly ref-count heap types. I think what that means is that there are at least two different cases for freelists: those that only keep the bare object memory alive (and can also support subtypes of the same size), and those that keep the (cleared) object alive, including its type. For the first, PyObject_INIT() works. For the latter, _Py_NewReference() seems the right helper function. The advantage of keeping the object as it is is the much simpler freelist code in tp_dealloc(). All it needs to do is 1) clear the ref-counted object fields (if any) and 2) put it in the freelist. No other C-API interaction is needed. If we only want to keep the object memory, then we need C-API support in both tp_new() and tp_dealloc(). If _Py_NewReference() is removed/hidden, then it would be nice if there was a replacement for the use case of initialising an object from a freelist that already knows its type. |
See also "Removal of _Py_ForgetReference from public header in 3.9 issue" thread on python-dev list: And "Re: [Python-Dev] Removal of _Py_ForgetReference from public header in 3.9 issue" thread on capi-sig list: |
I still consider that these functions must be moved to the internal C API. But I failed to find time to design a *public* C API for that, or help projects using these functions to avoid it. I prefer to close the isseu for now. |
Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.
Show more details
GitHub fields:
bugs.python.org fields:
The text was updated successfully, but these errors were encountered: