OpenBLAS 0.3.21 works perfectly with GNU Octave, but 0.3.22 causes several test failures in matrix inverses, eigenvector calculations, and sparse matrix operations. These include returning Inf instead of NaN and giving an answer that is wrong by several orders of magnitude. Affected Octave functions include inv(), eigs(), residue(), and det(), and any user code using those functions.

Thread: https://octave.discourse.group/t/make-check-fails-with-openblas-0-3-22/4289

How to reproduce:

1. Download & build a current version of GNU Octave from source: https://octave.org/download.
2. Run make check.
3. Observation: OpenBLAS 0.3.22 will give 10 test failures in the above functions. 0.3.21 will give zero test failures.

Workaround: I have reverted my system OpenBLAS to 0.3.21 for now, and Octave works properly again.

EDIT: Build settings for OpenBLAS, both versions:

FC=gfortran USE_OPENMP=1 USE_THREAD=1 USE_TLS=0 NO_STATIC=1 CPP_THREAD_SAFETY_TEST=1