[AMDGPU] Waterfall loop codegen improvement in SIInstrInfo (#192415)
When generating waterfall loops, use the instructions `v_cmpx_eq_*` and
`s_andn2_wrexec_*` as recommended for recent architectures, instead of
`v_cmp_eq_*` and `s_and saveexec`.
This PR only updates waterfall loop code generation in
`SIInstrInfo.cpp`. Other places that generated waterfall loops can be
handled separately.
- Add new lane mask constant for `s_andn2_wrexec`
- Set `isTerminator` for `v_cmpx_eq_{u32,u64}_e32`
- Fix test `mubuf-legalize-operands.mir` to track liveness needed for
verifying phi nodes
- Update .ll and .mir tests to accept the new instruction sequences
Assisted-by: Claude