The promise behind the limited ABI is exactly that if your extension works on 3.x, it will also work on 3.y, for y >= x.

One thing to watch out for: normally extension modules on Linux and MacOS don't try to link to libpython. Instead they trust that someone else will make sure they only get loaded into a compatible python interpreter, and that all the symbols they need from python will be injected by the host process.

On Windows, the way the dynamic loader works, you can't do this: every extension module has to explicitly say "I'm getting PyNumber_Add from the dll named: somethingsomething.dll"

But traditionally, version X.Y of python includes a pythonXY.dll, so there's no consistent name across releases. So even if your library uses only the limited ABI and all of your imports could just as well come from python36.dll or python37.dll... you need some way to express that, and only Windows has this problem.

I feel bad sending this without doing my own research, but I don't have access to a Windows box right now. Does anyone know if this problem has been solved? Is it still true that Windows python dlls always include the python version in their name?


On Tue, Jul 17, 2018, 09:38 Paul Moore <> wrote:
On 17 July 2018 at 16:59, Cosimo Lupo <> wrote:
> I would like to revive this 5 year old thread and see if we can stir things
> up a bit.
> Basically the problem is that, in the current state of the PEPs and build
> tools, it is still not possible to build and distribute a Windows wheel that
> includes an extension module compiled with Py_LIMITED_API.
> Setuptools can successfully build such extension module on Windows (with
> `.pyd` file extension and no extra specifiers in the module filename), and
> these seems to work at least on CPython 3.5 and above. However the
> `--py-limited-api cpXX` option of bdist_wheel command fails on Windows
> because it attempts to use the `abi3` tag but the latter is not in the list
> of compatible tags for that platform.
> One can work around this by creating a wheel with `none` as the abi tag, and
> `cp35.cp36.cp37` as the python implementation tag but this feels a bit
> hackish.
> Here are some unresolved questions from the old distutils-sig thread,
> quoting Paul Moore:
>> 2. How should tools determine which ABIs a given Python supports?
>> (This is the get_supported function in wheel and distlib). The "Use"
>> section of the PEP (
>> gives a Linux-based example, but nothing normative and nothing that is
>> understandable to a Windows user.
> And from Vinay Sajip
>> For Windows, we (eventually) need some low-level sysconfig-supported way
>> to get the ABI information in an analogous way to how it happens on POSIX:
>> and
>> because that's not currently there, distlib doesn't provide any ABI
>> information
>> on Windows other than "none".
> Other related links:
> So.. what needs to be done here to allow distributing/installing Windows
> wheels with Py_LIMITED_API support?

IMO, the question I posed back then remains key. Vinay's response is
fair, but I don't think that waiting for core Python to provide
something via sysconfig is practical (it's not happened yet, so why
would we expect things to change?). So I think the next step is
probably for someone to propose an algorithm that can be used.
Actually, what I'd like to see is a full end to end proposal of the
process someone would use to build and install a limited-ABI
extension. That would probably tease out a number of issues.

I imagine the steps would be something like this:

1. Create an extension. Presumably you'd need to #define
PY_LIMITED_ABI in the source.
2. Build a wheel from it - how would you do that? I gather it's
possible to do this with plain setuptools - would it be necessary to
do this with setuptools/bdist_wheel, or should there be a way to
request a limited ABI build via pip? If we do want to be able to
request this from a build tools like pip, do we need something in PEP
517? Are we only looking at the prebuilt wheel case, or do we need to
support building from source?
3. What tags would the built wheel have?
4. Install that wheel - `pip install xxx`. Pip needs to be able to
enumerate the full list of valid tags here (cp37-abi3, cp3-abi3, ...)
There are also questions like - if there's a limited ABI wheel and a
full ABI (version specific) wheel, which takes precedence?

I don't honestly know how well the limited ABI actually achieves its
goals - is "cp3-abi3-win_x86_64" a realistic tag to apply? Can limited
ABI wheels really be used on any version of Python 3? That's a
question for python-dev, rather than distutils-sig, but if we take the
position that this is what's promised, and it later turns out not to
be true, we've got a lot of wheel renaming to do when Python 3.10
comes out and it doesn't support the same limited ABI as 3.x for x <

Also, does the limited ABI work on other platforms? If it does, we
should ensure that the Windows support works the same. If it doesn't,
do we want a Windows-only solution (why is that OK?) or should we
extend to (say) manylinux or OSX (at the risk of making the job too
big for anyone to actually get anywhere with it).

So to move this forward, I think someone needs to write up the process
of building and using a limited ABI wheel, as described above, and
document suggested answers to the various questions that will come out
in the process of going through the details. Is that something you'd
be willing to take on?

From that, we'd have something concrete to debate. I'm not sure how
many people have an interest in this topic, so getting people with the
necessary knowledge to chime in might take some work (I'm interested,
but I don't have detail understanding of build options and linking
conventions). The ultimate goal would be some sort of PEP covering
handling limited ABI extensions within the packaging infrastructure.

Does that seem reasonable? Is that the sort of guidance you were
looking for? I doubt anything is going to happen on this subject until
someone with the interest in moving it forward steps up to do the work
of making a proposal and collecting community views.

Distutils-SIG mailing list --
To unsubscribe send an email to
Message archived at