We just ended our rather informal HPy meetup in Brno (Czech Republic) which started on Saturday (October 7th) and lasted until Tuesday (October 10th).
The main topics of the HPy meetup were:
- The CPython Core Dev Sprint
The CPython Core Dev Sprint
Petr Viktorin, one of the CPython core developers, invited us to partially attend the CPython core developer sprint. As a part of the sprint, he organized a C API Summit Monday where we were invited and happily attended this event.
Besides of lots of discussions, there were also very interesting presentations about the C API:
1. C API Work Group Summary
Irit presented the current status of the C API Work Group GitHub organisation. The idea of this GitHub org was to gather curent problems of the C API and also to collect ideas for the (re-)volution of the API. Irit thankfully went over all issues and discussions to structure and summarize them. She identified nine categories for problems:
- API Evolution and Maintenance
- API Specification and Abstraction
- Object Reference Management
- Type Definition and Object Creation
- Error Handling
- API Tiers and Stability Guarantees
- The C Language
- Implementation Flaws
- Missing Functionality
She also mentioned strengths of the C API which the people contributing to the repository agree on:
- Heap types
- APIs taking a C string to look up based on Python strings
- Limited API and Stable ABI
Now, the idea is to write an Informational PEP and then to create:
- A process for discussions about replacement APIs
- Guidelines for evolution of the C API.
The PEP draft is already available.
2. Petr's Vision for the C API
Petr presented his understanding of how the C API should look in future. The slides for his talk are public. He made seven suggestions:
Split API from ABI so you can compile the API that it would use different ABI variants.
Consistency: The API should have a consistent behavior. E.g. APIs should not return borrowed references and the exception behavior should be clear by the naming convention.
Versioning: Essentially, a C extension should include the ABI version it was compiled with.
Errors: All API functions must be able to signal an error that can be handled.
Concrete types: The API should provide dedicated C types for certain built-in objects to have type safety already at the C language level.
Blueprint structures: Creating types and objects should rather be done by using specifications than by allocating empty objects and filling them with content. Just think of
Context argument: Every API function should have a context argument.
Except for points 4 and 5, these goals are very close to HPy's. We, the HPy team, have to say that this sounds very good and we would be happy to see that in CPython.
3. HPy's Lessons Learned
We were also asked to give a talk and to concentrate in the talk on two things:
- How do handles affect the API, and
- What is the reason for the HPy context to contain the function table.
We didn't advertise HPy in our talk but we talked about why opaque handles
are not only useful but sometimes necessary and how the API would change
compared to having only
We further showed that having the function table in the context gives us a lot
of flexibility and finally enables us to switch on the debug mode essentially at
run time (when an extension is loaded).
We are not sure if we managed to convince everyone that those concepts are useful. The biggest problem is that introducing handles and a context argument will radically change the API and require changes in many libraries (see also C API Workgroup - API Revolution #2 and C API Workgroup - API Revolution #7).
4. Python C API: Past, Present, and Future
Victor did a thorough presentation of the past, present, and (possible) future of the Python C API. He pointed out some past and present problems and gave some impression of its complexity. Victor also suggests to make the stable ABI be the default. Overall, he sees a positive trend in how the C API evolves.
We had a lot of good discussions and probably the most important one was about
the PEP 703 (noGIL) implementation and the
new version of ABI (abi4) that is planned to be compatible with both noGIL and
GIL builds of CPython. In order to achieve that, Sam, the author of PEP
703, plans to make
PyObject * opaque (so it will narrow to the semantics of a
handle) and also, the
Py_INCREF will, in some cases, use an indirection over a
function pointer to actually call the incref logic. This is, again, a big step
forward for alternative implementations.
Besides the main focus on the C API Summit at the CPython Core Dev Sprint, we
were also hacking a bit on Cython/HPy and made the doctest system work.
The problem there was that the doctest module is binding functions to self in
order to get a method but HPy did not yet provide the necessary slot
HPy_tp_descr_get for that. So, we've implemented that and will include it in
the next (minor) release.
We hoped that we would do more hacking on HPy itself but the CPython core dev sprint was worth to spend our time too.