[llvm-branch-commits] [llvm] AMDGPU: Back-propagate wqm for sources of side-effect instruction (PR #193395)

[via llvm-branch-commits]

================
@@ -725,6 +741,19 @@ char SIWholeQuadMode::analyzeFunction(MachineFunction &MF) {
       propagateInstruction(*WI.MI, Worklist);
     else
       propagateBlock(*WI.MBB, Worklist);
+
+    if (Worklist.empty()) {
----------------
ruiling wrote:

On a second thinking, I think even theoretically, we only need to process the `ExeczSideEffectInstrs` once. We only go propagating the wqm-ness if it is in the wqm-region. The wqm backward propagation of the side-effect instruction should not expand the wqm-region. So we are ok to clear the `ExeczSideEffectInstrs`.

Yes, I agree it's odd to see it in the loop, as it needs to redo the loop. But I don't have a good idea to improve this.

https://github.com/llvm/llvm-project/pull/193395