These are scalar instructions that change vector instructions, so they should not be executed without any active lanes. The implementation of -amdgpu-skip-threshold also seem to be backwards from expected, since decreasing it prevents removal.