apply OpenBLAS_jll v0.3.23+4 patch

[pablosanjose]

Closes #53054

CC: @giordano @ViralBShah

[pablosanjose]

Two questions:
(1) is CI working correctly? If so, I cannot make heads or tails of the test failures
(2) should we add the MWE of #53054 as a test?

[giordano]

First of all, have you verified the bug is indeed gone on this branch? Regarding a regression test, that'd be a good idea, but a test only on this branch would be quite useless as it'd be lost forever in future releases. You want to submit a separate PR for master, perhaps to be backported to 1.10 (but the backport doesn't seem massively useful).

[ViralBShah]

On 1.10 branch, we also need to add the patch for a source based build. So the patch should also get added to deps/patches, and it should be applied in deps/openblas.mk.

[ViralBShah]

The best way to confirm the bug is indeed fixed is to use the build artifact produced by this PR:

https://buildkite.com/organizations/julialang/pipelines/julia-release-1-dot-10/builds/191/jobs/018d4778-8e0d-4db8-8f73-bd65ca4992ae/artifacts/018d477e-8eb1-4b6f-a13c-df29e0d4c1fe

[imciner2]

(2) should we add the MWE of #53054 as a test?

I'd say the regression test should at a minimum go into master (we added one for the other OpenBLAS patch we are using for M1 just recently in #53059). That should be a separate PR though.

[pablosanjose]

First of all, have you verified the bug is indeed gone on this branch?

I can confirm that for me this branch builds from source using BinaryBuilder, and the result is free of the bug.

[pablosanjose]

Regarding the test failures, at least some of them are also appearing in buildkite runs of other branches, such as avi/53062, so they should be unrelated to the patch.

Regarding a regression test, that'd be a good idea, but a test only on this branch would be quite useless as it'd be lost forever in future releases. You want to submit a separate PR for master, perhaps to be backported to 1.10 (but the backport doesn't seem massively useful).

I'd say the regression test should at a minimum go into master (we added one for the other OpenBLAS patch we are using for M1 just recently in #53059). That should be a separate PR though.

I'll add a test here and submit it also to master as a separate PR.

[pablosanjose]

There is a potential issue I'm trying to understand: this patch is for OpenBLAS (JuliaRegistries/General#99646) which is indeed v0.3.23+4, but the OpenBLAS32 version (JuliaRegistries/General#99647) is v0.3.23+3. Is this a problem?

[giordano]

Is this a problem?

No.

[KristofferC]

So this should be merged in addition to 3c73873?

[imciner2]

So this should be merged in addition to 3c73873?

Yes, because this updates the local build for OpenBLAS in case people are not using the BinaryBuilder version.

[KristofferC]

Test error is for another change on the backport branch.