[llvm] [AMDGPU] Add scheduling mask operand to S_SETPRIO (PR #204850)

[Frederik Harwath via llvm-commits]

frederik-h wrote:

> This should not be done. s_setprio is a real instruction, and the scheduling boundary masks are a huge hack to start with. Anything related to scheduling masks needs to be fully contained within the backend and not spread to any more interfaces. The intrinsics should generally correspond directly to a real instruction, without any meta-operands attached. I definitely don't want to have to start maintaining a whole extra set of intrinsics with sched mask pollution

But the current handling of the instruction is overly restrictive and @nhaehnle and I see a real need for a more flexible implementation. Can you suggest a better approach? Perhaps a different intrinsic with a more specific mask, a special intrinsic that only sets the salu bits in the mask internally etc.?

https://github.com/llvm/llvm-project/pull/204850