04:36 <fijal> ltfish: yes, it is expected pypy interpreter is slower

07:22 <phlebas> can anyone point me to how pypy is dealt with in conda package build recipes (maybe antocuni?). that is, I'm trying to understand how e.g. the numpy recipe is built for pypy. it only declares python as a dependency, but I assume there should be variants for different python versions (since numpy doesn't compile only against the stable abi), and there should be a variant for each pypy version.

07:47 <antocuni> despite my affiliation, I don't know anything about conda 😂

07:47 <antocuni> phlebas: I think that mattip knows

07:51 <phlebas> antocuni: what?? I thought you are now the "alternative python implementation champion" on the inside! :D

07:53 <antocuni> I don't do anything related to conda 😂

08:03 <mattip> there is a python_abi feedstock that allows specifying python implementations

08:03 <mattip> https://github.com/conda-forge/python_abi-feedstock

08:04 <mattip> the "meat" is here

08:04 <mattip> https://github.com/conda-forge/python_abi-feedstock/blob/main/recipe/conda_build_config.yaml

08:04 <mattip> then there is a pypy feedstock unfortunately named pypy36.-feedstock

08:04 <mattip> https://github.com/conda-forge/pypy3.6-feedstock

08:05 <mattip> that provides the python package

08:05 <mattip> end users specify pypy oer cpython as documented here

08:06 <mattip> https://conda-forge.org/docs/user/tipsandtricks.html#using-pypy-as-an-interpreter

08:06 <phlebas> thanks. what's the pypy-meta-feedstock for?

08:08 <mattip> that specifies the connection between the python_abi_feedstock specifier (pypy38_pp73) and a pypy python version 3.8.12

08:08 <phlebas> hm. that's still a bit unclear to me. I read about the python_abi thing and that you can then specify pypy or cpython versions, but the numpy-feedstock seems to only require "python", not "python_abi". is that some magic somewhere that is not expressed via dependencies?

08:10 <phlebas> ah, so the chain is something ike python -> python_abi -> pypy-meta -> pypy ?

08:10 <mattip> in this case, you can see the various packages provided by numpy: all the variations in python_abi_feedstock are rendered into CI versions here

08:10 <mattip> https://github.com/conda-forge/numpy-feedstock/tree/main/.ci_support

08:11 <mattip> and each variant is a package on anaconda.org ready for installation

08:11 <mattip> https://anaconda.org/conda-forge/numpy/files

08:12 <mattip> once a new version is added to the python_abi, there is a bot that starts a migration which goes through all the recipes and starts to add the new CI variants

08:13 <mattip> you can track migrations via https://conda-forge.org/status/#pypy38

08:14 <phlebas> ok, thank you, that's a good chunk of pointers! I'll have to go through that stuff to understand how it all fits together in my head 🙂

08:14 <mattip> this might live in a document somewhere, but I don't know of one

08:16 <mattip> the last step in the chain is to get the feedstock maintainer to approve the migration PR

08:17 <mattip> Of the ~1000 packages that made it through the pypy38/39 migration, most were effortless,

08:17 <mattip> I needed to gently push the successful PR through in maybe 20 cases (remind the maintainer to hit the merge button)

08:17 <mattip> and needed to fix PyPy or project incompatibilities in another 20 or so

08:53 <phlebas> mattip: cffi is provided in pypy-meta-feedstock, I guess we need to add hpy there, too, right? to be able to provide hpy packages

09:02 <mattip> right, once hpy starts becoming widely supported, we will need to add it because the version is tied tightly to pypy

09:03 <mattip> installing pypy fixes the version of cffi/hpy you can use

11:10 <cfbolz> just found a JIT optimization bug :-(

11:11 <cfbolz> a pretty theoretical one for pypy, admittedly

11:11 <cfbolz> you need to use __pypy__.intop.int_add to get it from normal python code

11:30 <cfbolz> https://foss.heptapod.net/pypy/pypy/-/issues/3832