"spin build" errors on windows

[ngoldbaum]

When I try to use spin to build numpy on windows, I get:

meson.build:1:0: ERROR: prefix balue '/usr' must be an absolute path

It looks like this is hard-coded in the build command. I've never had an issue with it before so no idea why it's problematic now.

[ngoldbaum]

It looks like deleting the hard-coded prefix lets me get past the error.

[stefanv]

Is this a bug in spin then?

I've been trying to get numpy onto 0.12 but I guess you've seen that Matti had concerns about the test modifications.

[ngoldbaum]

Is this a bug in spin then?

Well, I was able to build numpy with spin after making this change to spin:

diff --git a/spin/cmds/meson.py b/spin/cmds/meson.py
index 366cd60..b2d8149 100644
--- a/spin/cmds/meson.py
+++ b/spin/cmds/meson.py
@@ -266,7 +266,7 @@ def build(meson_args, jobs=None, clean=False, verbose=False>
     if gcov:
         meson_args = meson_args + ["-Db_coverage=true"]

-    setup_cmd = _meson_cli() + ["setup", build_dir, "--prefix=/usr"] + meson_a>
+    setup_cmd = _meson_cli() + ["setup", build_dir] + meson_args

     if clean:
         print(f"Removing `{build_dir}`")

I have no idea why the --prefix=/usr is necessary so following the principle of Chesterton's fence, I didn't want to put in a PR to do something about it without knowing more.

I haven't been following the issues with spin in NumPy's CI, this was all just to test a NumPy PR on my Windows PC.

[stefanv]

The Windows tests for spin run in bash, I think. Which environment do you use?

[ngoldbaum]

This is using powershell via windows terminal.

[stefanv]

Ah, OK, that's wild west territory for us. Maybe I can get a subset of tests to run on that platform. This is what you've been using all along, without any difficulty?

[ngoldbaum]

Yes, it should be possible to build numpy using powershell. I'm not sure when the last time I tried was, it may predate switching to spin.

[ngoldbaum]

I just hit this again in CI, trying to update the numpy tests on windows to install python with uv instead of setup-python. For whatever reason, this setup hits the same error:

https://github.com/ngoldbaum/numpy/actions/runs/11373312391/job/31639601441#step:10:25

You never explained earlier - is there any particular reason why --prefix=/usr is hard-coded like that? What would break if I deleted it?

[ngoldbaum]

Ah, there's this note in the meson docs:

prefix defaults to C:/ on Windows, and /usr/local otherwise. You should always override this value.

So clearly --prefix=/usr is a bad default on windows.

[rgommers]

I don't see a problem with using C:/ on Windows and /usr on other platforms. The path is still going to be reproducible, and it has the upside of actually existing.

Note that to actually have a uniform install path within the source tree, the recommended way is to use meson install --destdir (see https://mesonbuild.com/Installing.html#destdir-support).

[ngoldbaum]

The other thing that annoys me about how it's currently set up is there's no way to override the default. If you try to manually pass --prefix spin doesn't notice that and happily passes two --prefix arguments to meson, which meson complains about.

So, I think the solution is probably to use C:\ on windows, /usr on unix-ish platforms, and also only pass in a --prefix if the user didn't explicitly pass it.

I'll try to do that later today.