Add more Fortran compiler testing in CI

[foxtran]

It is important to test SW with all compilers which are using in production.

At least these compilers are available kinda freely and can be used in CI:

• nvfortran
• armflang
• LLVM flang

Comes from #1255

[ACSimon33]

Good idea! Intel compilers (oneAPI) could also be added to the list. ICC/ifort 2023.2.0 and icx/ifx 2025.3.0 should work now on Linux and Windows as well.

I'll also think about how I can provide you with a NAG Fortran license or a Docker container with NAG Fortran. We at nAG have nagfor pipelines for the LAPACK repo, but those won't run for PRs unfortunately. So I can only report back after something's already broken on master. @langou @weslleyspereira how are the action runners currently setup? Do we have custom Docker images? Are we using our own action runners or some shared ones from GitHub?

[foxtran]

how are the action runners currently setup?

mostly GitHub CI + extra something.

Do we have custom Docker images?

you can provide one

Are we using our own action runners or some shared ones from GitHub?

No idea, but looks like shared from GitHub.

I'll also think about how I can provide you with a NAG Fortran license or a Docker container with NAG Fortran.

Good question, technically, in GitHub, there is an encrypted storage for credentials which can be passed into CI image, but I have no idea how to steal them to prevent possible attacks. For Docker container with internal keys one can extract the key by getting the container.

[martin-frbg]

nvfortran could be a bit tedious to set up in CI, as the NVIDIA compiler comes as a kitchensink package with CUDA and whatnot. Even getting it downloaded into an image turned out to be impossible (within the usual constraints of free or budget CI jobs) when I tried to wire it up for OpenBLAS a while ago.

armflang from the Arm Linux Toolchain appears to behave just like a repackaged LLVM flang-new

LLVM flang is what is now used in the Appveyor CI job - ironically, the outdated flang-11 it replaced apparently displayed pretty much the same (segfault) behaviour) w.r.t the reverted PR as nvfortran and the deprecated Arm Fortran compiler (shared heritage ?), so there might even be some merit to resurrecting that

[foxtran]

armflang from the Arm Linux Toolchain appears to behave just like a repackaged LLVM flang-new

Nope, like flang classic as of 2024-2025. And NVfortran is also flang-classic.

[martin-frbg]

Current Arm Toolchain definitely identifies (and behaves) as LLVMFlang i.e. flang-new. The deprecated Arm Fortran appears to have been based on classic though