05:51 <mattip> darn. Even this segfaults on the release since binary compatibility is broken: 'pypy -m pip install numpy; pypy -c "import numpy; numpy.test()"'

06:27 <mattip> I added a numpy build to https://github.com/pypy/binary-testing/actions that shows the problem :(

06:32 <mattip> mgorny, tumbleweed: is there a way to re-release 7.3.6 with a backward compatible API?

07:12 <mgorny> mattip: i'd make it 7.3.7 to avoid confusion

07:12 <mgorny> and add a big warning to 7.3.6

07:13 <mgorny> that said, are we talking about pypy3.8 too or just 2.7/3.7?

07:13 <mgorny> (i probably didn't notice because i went straight for 3.8, so total rebuild was expected anyway)

07:32 <arigato> mattip: not sure to see the point, just do a 7.4.0 and don't spend time reverting carefully some bits

08:04 <mattip> arigato: yeah, it appears we will have to

08:04 <mattip> but then this means projects like cryptography will have to release wheels for both pp37-pypy73 and pp37-pypy74

08:06 <mattip> mgorny: pypy3.8 would be OK to move to 7.4.0 since it is too early for wheels to have appeared with pp38-pypy73

08:13 <mgorny> mattip: yeah, i'd prefer not to have to rebuild everything there ;-)

08:13 <mgorny> i.e. leave the new ABI for pypy38 as-is

08:24 <arigato> is the ultimate goal of pypy3.7 to continue being updated for a while, or just dropped?

08:24 <arigato> if dropped, then I see the point of wanting to at least not change its C ABI right now

08:27 <mattip> arigato: correct. If progess on py3.9 moves quickly, and we do not see too many problems with py3.8, I will propose that the bugfix release be the last release of py3.7

08:30 <arigato> OK

08:31 <mattip> so it seems I should release pypy3.7 v7.3.7 and pypy3.8 v7.4.0. Complicated

08:33 <arigato> you could also say that older pypy3.8 < 7.3.6 were beta anyway, and not care about wheels built with them

08:34 <arigato> and only release pypy3.7 7.3.7

08:49 <mattip> well, NumPy for instance is releasing in 6 weeks.

08:49 <mattip> If we declare 7.3.6 broken (which it is) and do not release a pypy3.8 v7.4.0, then they will not release a pypy3.8 binary wheel at all

08:50 <arigato> no, I mean, pypy3.8 7.3.6 is not broken, if nobody used the older pypy3.8 < 7.3.6 because they were beta

08:50 <arigato> if I'm wrong about that, then fine :-)

08:56 <mattip> you are not wrong, we could continue calling the version "7.3.x" as long as we keep pypy3.7 with the older, 7.3.5, ABI

08:57 <mattip> we never released a pypy3.8 < 7.3.6

11:51 <cfbolz> mattip: hm, complicated

11:52 <cfbolz> mattip: are there any ABI changes we held back but like to do at some point?

11:57 <mattip> I would like to get our PyThreadState struct more inline with CPython,

11:58 <mattip> and design a stable API like CPython's abi3

11:58 <mattip> but there are none we held back from the release

12:06 <cfbolz> mattip: right

12:06 <cfbolz> Anyway, seems mgorny has pretty reasonable arguments

14:10 <mattip> ok, so we stick with 7.3.7 (for the c-api reversion on py3.7) and the next real release will be 7.3.8

14:10 <mattip> unless we decide it is worth the price to break ABI compatibility for PyThreadState or some kind of stable API

16:31 <cfbolz> mattip: thanks for dealing with this

16:32 <tumbleweed> doesn't matter to me (because we still don't build anything against pypy3.x in Debian)

19:16 <mattip> tumbleweed: thanks

19:21 <cfbolz> making progress with 3.9, very slowly

19:21 <cfbolz> trying to get the stdlib tests to run again

19:38 <Corbin> I'm trying to link against a single-file library (https://github.com/nothings/stb/blob/master/stb_image_write.h) and putting it in the includes seems not to work.

19:39 <Corbin> My rffi declaration is https://bpa.st/VL2Q and translation works, but the C compiler complains "undefined reference to `stbi_write_png'"

19:53 <Corbin> pre_include_bits=["#define STB_IMAGE_WRITE_IMPLEMENTATION"] helps a little? This gets the STB library to define its procedures, but it defines them multiple times, so now I get "multiple definition of `stbi_write_png'"

19:55 <mattip> cfbolz: we should try to get some of our changes to stdlib into upstream, so the merge is easier

19:56 <cfbolz> mattip: actually I was pleasantly surprised that quite a few of our changes were in 3.9

19:56 <mattip> Corbin: can you try to mimic an example use of stb_image_write.h somewhere so you know what has to be define/undefined?

19:56 <cfbolz> mattip: eg almost all gc.collects in tests

19:56 <mattip> cool

20:00 <Corbin> mattip: I think I got it. I used a separate module which effectively just becomes the instantiated header; https://bpa.st/P3XA translates.

20:00 <mattip> +1

20:02 <Corbin> And it produces a PNG! Progress. Sorry for rubber-ducking.

20:30 <mattip> :)