Not sure this will work, we will see.

That looks right to me. How does the RTOOLS40_HOME environment variable get created?

How does the RTOOLS40_HOME environment variable get created?

Looks like it is set by rtools during installation. Maybe that is why not all the windows tests were failing?

So maybe using RTOOLS43_HOME would also fix things?

Darn, still failing.

Note that full tests on 64 bit windows passed.

Those tests look as if they're using gfortran to compile code. If you look at https://dev.azure.com/numpy/numpy/_build/results?buildId=29282&view=logs&j=d02055d6-a30d-5402-19aa-ca2310eeee52&t=3b8a6fc4-0b0a-5f32-7dd9-04a675f63363&l=1691 I think the 64 bit compiler is being picked up, rather than the 32 bit. The 32 bit is in $env:RTOOLS40_HOME\mingw32\bin.

The path and flags look correct:

- powershell: |
    If ($(BITS) -eq 32) {
        $env:CFLAGS = "-m32"
        $env:LDFLAGS = "-m32"
        $env:PATH = "$env:RTOOLS40_HOME\\mingw32\\bin;$env:PATH"
    }
    Else
    {
        $env:PATH = "$env:RTOOLS40_HOME\\mingw64\\bin;$env:PATH"
    }
    python runtests.py -n --show-build-log --mode=$(TEST_MODE) -- -rsx --junitxml=junit/test-results.xml

What I cannot determine is what changed to make the tests suddenly fail. Do we need some flags for gfortran?

Seems like adding the path worked, just harder to do then one would think. Is this the right thing then (with maybe tiny cleanup?).
To me it seems like we might as well give it a try, the constant failure isn't helpful after all...

EDIT: Is the echo "c:\rtools40\ucrt64\bin;" >> $env:GITHUB_PATH a way to add things to the path that is more reliable maybe?!

Is the echo "c:\rtools40\ucrt64\bin;" >> $env:GITHUB_PATH a way to add things to the path

I'm giving it a shot, let's see how it does. I would prefer that as it is hopefully global.

It turns out that rtools 42+ do not support 32 bit builds, so we are stuck with pinning rtools until we drop the 32 bit Windows builds.

I'm also curious to see if the 32 bit wheel builds work tonight.

Is the echo "c:\rtools40\ucrt64\bin;" >> $env:GITHUB_PATH a way to add things to the path

Yes, I think that's the way to add things permanently to path. If you add an environment variable in one step they're not carried through to subsequent steps.

See https://docs.github.com/en/actions/using-workflows/workflow-commands-for-github-actions#setting-an-environment-variable. GITHUB_ENV does environment variables, GITHUB_PATH does path I think.

See https://docs.github.com/en/actions/using-workflows/workflow-commands-for-github-actions#adding-a-system-path as well.

(I did try making a PR against @charris fork to try to illustrate this)

Used the (different) method we used in numpy-wheels for setting global variables and it worked. There was an unrelated test failure that turns up now and then and that can be ignored. I'm going to put this in.

Thanks for the help @andyfaff @seberg.