incorrect `build-details.json` for `linux-ppc64le` `py314t` build

[lucascolley]

Hey @mgorny, I wonder whether you can help me understand this problem further?

In build-details.json, it seems the ppc64le py314t build is:

• missing abi.flags = ["t"]
• missing the t from the ".cpython-314t-x86_64-linux-gnu.so" entry of suffixes.extensions

Compare¹:

Footnotes

1. damn I love pixi-browse ↩

[lucascolley]

(this is blocking conda-forge/hypothesis-feedstock#1236)

[mgorny]

Interesting. I wonder if it's because we're cross-compiling there.

[mgorny]

Wait a minute, you are comparing two builds there, right? Could you be more specific because the screenshots are missing the build strings.

[mgorny]

Ah, sorry, that's ppc64le vs x86_64, nevermind.

[mgorny]

Okay, I see the problem:

https://github.com/python/cpython/blob/dd1194e2fc2c08c36f4fe7ee71f7ab0b82c09570/Tools/build/generate-build-details.py#L66
https://github.com/python/cpython/blob/dd1194e2fc2c08c36f4fe7ee71f7ab0b82c09570/Tools/build/generate-build-details.py#L68-L72

Note it's grabbing data from the running Python interpreter to fill that in, so these variables end up being filled with build-python rather than host-python (I should have noticed the x86_64 there). I wonder if this is properly fixable upstream. As a hack, we may use emulation to regenerate the file during the build.

[mgorny]

Apparently reported already: python/cpython#136267

[mgorny]

Okay, my obvious idea was wrong. With 3.15+ we're having multiple extensions, so reusing SOABI wouldn't work. I'm going to try using emulation to fix it locally in conda-forge then.