ci: add builds for windows 11 arm

[davidhewitt]

Closes #209

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

CodSpeed Performance Report

Merging #210 will not alter performance

_{Comparing dh/windows-arm-builds (32d89e0) with main (c6758ed)}

Summary

✅ 70 untouched

[jessekrubin]

hi @davidhewitt,

I have a possibly stupid question...

I was poking around and have noticed that in this repo as well as others the pydantic-et-al team does "PGO" builds.

What/why is that? When should it be used? When should it not be used?

-jesse

[davidhewitt]

PGO is profile guided optimization. By running the test suite with special instrumentation the compiler can record the hot paths and then optimize the final artifact to make those as efficient as possible.

[davidhewitt]

On pydantic-core we saw up to 30% performance gains in benchmarks by enabling.

[MugundanMCW]

Hi @davidhewitt,

• The wheel build is currently failing on windows-11-arm GA runners because the cachedir contains x64 Python installations, which causes linker conflicts.
• I resolved this issue by explicitly removing all cached x64 Python variants before the build, which allowed the build to succeed. Here’s the step I added in CI:

        - name: Remove cached x64 Python for windows-11-arm
          if: matrix.os == 'windows-11-arm'
          run: Remove-Item -Recurse -Force "C:\hostedtoolcache\windows\Python\*\x64"
          shell: pwsh

• Note: This issue only occurs on GitHub’s windows-11-arm runners. It does not happen on a local Windows ARM64 setup, since no x64 Python installations are present there.

[jessekrubin]

Hi @davidhewitt!

Thanks for the explanation! Does that mean you are getting a build that is optimized for your test suite? Ie your test suite should ideally be representative of real world use cases?

How difficult is it to set up pgo builds?

[davidhewitt]

Does that mean you are getting a build that is optimized for your test suite? Ie your test suite should ideally be representative of real world use cases?

Yep exactly. The more comprehensive your test suite the better :)

How difficult is it to set up pgo builds?

At the moment, quite painful - the best examples are the build scripts here, in pydantic core, and also my discussion in PyO3/maturin#1840