While this does the bare minimum to get anv compiling on non-x86 platforms, it is a start.
The problem coems with anv sprinkling memory fences and cacheline flushes around the codebase. This is architecture specific how this should be done.
Theoretically memory fencing can be implemented with c11 and atomic_thread_fence but this will be left as an exercise for the future.
No one is running these Intel dGPUs on non-x86 currently anyway.