LGTM. I am trying to run the benchmark comparison on this, but something seems off

python runtests.py --bench-compare HEAD main
...
·· Building 1d0487b5 <getenv-slow-warn-mem-policy> for virtualenv-py3.11-Cython-packaging-setuptools59.2.0..........................................................
·· Installing 1d0487b5 <getenv-slow-warn-mem-policy> into virtualenv-py3.11-Cython-packaging-setuptools59.2.0
·· Failed to build the project and import the benchmark suite.

It would be nice to see that this has some echo in the benchmarks.

Hmmm, good question, I do think a few of our benchmarks should notice this, although I don't think we have dealloc benchmarks specifically.

spin bench --compare -t bench_scalar

works for me on main on MacOS, but I don't have access to my windows test machine on which I dug into this right now.

(scalar, because I suspect a few of them create temporary 0-D arrays making up a big chunk of their overhead. Another benchmark to try could be ArrayCoercionSmall)

I cannot see a performance difference on MacOS. I suspect the only reason why this getenv call was there in the first place is because we couldn't measure its performance impact on linux/mac.

Once I worked out that I had pkg-config set up to use OpenBLAS on windows, which does not work (see #24251), I could disable it and run benchmarks. It seems benchmarks for bench_scalar on windows are quite random. Running the reproducer at the top of the issue printed, for 1.25.1

Function completed in: 0.64 seconds
Function returned after 2.4 seconds
Function completed in: 6.30 seconds
Function returned after 24.0 seconds

For this PR

Function completed in: 0.53 seconds
Function returned after 0.6 seconds
Function completed in: 5.66 seconds
Function returned after 6.7 seconds

Looks good!

Thanks @seberg