I can't get this to work. It seems somehow specifying CC messes with the cross-compilation.

Can I ask what motivated the attempt to try making this switch? Using Clang seems healthier on macOS, since that's how everything gets compiled and it avoids pulling in C/C++ runtime libs.

There is a floating point error emmited in zgetrf on macos-arm64 that I am trying to track down. It does not appear on the cirrusCI builds, but can be reproduced by rebuilding locally. I am trying to determine what combination of compiler + processor triggers the FPE. Locally, when I build OpenBLAS with GCC I do not see the FPE, so I was trying to replicate that success here.

It seems this is not the correct path, closing.

I can't see so quickly if zgetrf uses a BLAS function that uses SME, but if so it could be similar to numpy/numpy#29223.

I also noticed that the macos-arm64 builds on this repo still use macos-13 x86-64 runners, which probably doesn't make too much sense anymore and will make it hard or impossible to see this in CI. Switching to macos-latest would be useful either way.