Description of problem:

The attached test program fails on aarch64 on RHEL 8 and 9.
The failure in not present on other architectures.
It is not present on Fedora 36 and Fedora Rawhide on any architecture including aarch64.

Version-Release number of selected component (if applicable):

Failure seen with:

gcc-8.5.0-17.el8.aarch64
openblas-devel-0.3.15-4.el8.aarch64

gcc-11.3.1-2.1.el9.aarch64
openblas-devel-0.3.15-3.el9.aarch64

No failure seen with:

gcc-8.5.0-17.el8.x86_64
openblas-devel-0.3.15-4.el8.x86_64

gcc-11.3.1-2.1.el9.x86_64
openblas-devel-0.3.15-3.el9.x86_64

gcc-12.2.1-2.fc36.aarch64
openblas-devel-0.3.19-3.fc36.aarch64

gcc-12.2.1-2.fc36.x86_64
openblas-devel-0.3.19-3.fc36.x86_64

gcc-12.2.1-3.fc38.aarch64
openblas-devel-0.3.21-3.fc38.aarch64

gcc-12.2.1-3.fc38.x86_64
openblas-devel-0.3.21-3.fc38.x86_64

How reproducible:

Always on RHEL 8 aarch64 and RHEL 9 aarch64

Steps to Reproduce:
1. Compile the test program: gcc -o test test.c -lopenblaso
2. Run the test program: ./test

Actual results:

On RHEL 8 aarch64 and RHEL 9 aarch64 the 2nd and 3rd row of the matrix a[200] - a[599] consists of inf and nan:

a[200] = -inf;
a[201] = nan;
a[202] = -nan;
a[203] = -nan;

Expected results:

No inf and nan:

a[200] = -0.433977;
a[201] = 9.420676;
a[202] = -0.000003;
a[203] = 0.021688;

Additional info:

This causes build failures for the R-qtl package in EPEL 8 and EPEL 9