[llvm] [AMDGPU] V_SET_INACTIVE optimizations (PR #98864)

[Jay Foad via llvm-commits]

================
@@ -555,16 +559,24 @@ char SIWholeQuadMode::scanInstructions(MachineFunction &MF,
         GlobalFlags |= StateStrictWQM;
       } else if (Opcode == AMDGPU::V_SET_INACTIVE_B32 ||
                  Opcode == AMDGPU::V_SET_INACTIVE_B64) {
-        III.Disabled = StateStrict;
-        MachineOperand &Inactive = MI.getOperand(2);
-        if (Inactive.isReg()) {
-          if (Inactive.isUndef()) {
-            LowerToCopyInstrs.push_back(&MI);
-          } else {
-            markOperand(MI, Inactive, StateStrictWWM, Worklist);
+        // Ignore these if V_SET_INACTIVE which already has exec src register.
+        // These are generated by an earlier pass which has seperately ensured
+        // WWM and provided a mask of inactive lanes.
----------------
jayfoad wrote:

@perlfu what's the point of this? As far as I can tell it only happens with the hand written MIR test that you added (preloaded_set_inactive in wqm.mir).

https://github.com/llvm/llvm-project/pull/98864