[llvm] [AMDGPU] Iterative scheduling must behave the same with/without debug (PR #178460)

[via llvm-commits]

================
@@ -381,10 +381,14 @@ void GCNIterativeScheduler::scheduleRegion(Region &R, Range &&Schedule,
   auto Top = R.Begin;
   for (const auto &I : Schedule) {
     auto MI = getMachineInstr(I);
-    if (MI != &*Top) {
+
+    MachineBasicBlock::iterator MII = MI->getIterator();
+    if (MII != Top) {
+      bool NonDebugReordered =
+          !MI->isDebugInstr() && skipDebugInstructionsForward(Top, MII) != MII;
       BB->remove(MI);
       BB->insert(Top, MI);
-      if (!MI->isDebugInstr())
+      if (NonDebugReordered)
----------------
LU-JOHN wrote:

> Why does this need to be explicitly checking for debug instructions? Just do skipDebugInstructionsForward and then check if the position changed?

`skipDebugInstructionsForward()!=MII` alone does not guarantee that LiveIntervals needs to be updated.  We can have MI is a debug instruction and `skipDebugInstructionsForward()!=MII`, but moving MI will not change the order of non-debug instructions. 

https://github.com/llvm/llvm-project/pull/178460