Axel Huebl

Ah no, it's a global - this one need export: https://github.com/AMReX-Codes/amrex/blob/4f639294606d47185d31eaee4af66fc6b590e5a2/Src/Base/AMReX_GpuDevice.H#L172

@robertmaynard not sure if relevant to your tasks, but we enabled `.dll` support in #1847 . With this CI build (Windows + CUDA), I am still puzzled.

I tried `cmd` terminal and `-G Ninja` in an earlier commit (2nd last), but that did not work either.

> May be add a cron option to on: to compile the Windows+CUDA build on a weekly basis? I am caching the download, but if it gets to build at...

I think this PR should cache it's download well after the first build. We can re-work it in case it does not and causes overly long CI times.

re-opening to trigger CI

cc @PhilMiller might have some insight for macOS here, due to his work in https://github.com/ECP-WarpX/WarpX/pull/2896

loosely related, thus x-ref: #1969 (@sayerhs) - that PR is just a compile error on Apple M1 CPUs for signal handling of the same functionality.

@RTSandberg for the test, set break points around/int the backtrace writing function and compile with `-DCMAKE_BUILD_TYPE=Debug`. If you know `gdb` already, here is a good cheat sheet to use `lldb`...

We debugged this today and it looks like although ```C++ _MM_SET_EXCEPTION_MASK(prev_fpe_mask & ~curr_fpe_excepts); prev_handler_sigfpe = signal(SIGFPE, BLBackTrace::handler); ``` is called on macOS (x64), the signal handler `BLBackTrace::handler` is not called...