Might also be worth looking at https://docs.travis-ci.com/user/multi-cpu-architectures/\|https://docs.travis-ci.com/user/multi-cpu-architectures/ since it includes arm64

@samth thanks for the reference, I will take a look at it soon.

Initial results: The bytecode interpreter backend can be up to 15x as slow, but the difference is much smaller (maybe 3x) for the bytecode interpreter compiling itself. So, the approach looks practical, so far.