I wonder how good a LLVM backend for these rare architectures would have to be for that to be “good enough” for the kernel team. Obviously correctness should be non-negotiable, but how important is it that the generated e.g. Alpha code is performant for somebody’s hobby?