[llvm] [AMDGPU][SIInsertWaitcnts] Track SCC. Insert KM_CNT waits for SCC writes. (PR #157843)

[Jay Foad via llvm-commits]

================
@@ -2419,6 +2468,19 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
     if (T == DS_CNT)
       StrictDom |= mergeScore(M, LastGDS, Other.LastGDS);
 
+    if (T == KM_CNT) {
+      StrictDom |= mergeScore(M, SCCScore, Other.SCCScore);
+      if (Other.hasPendingEvent(SCC_WRITE)) {
+        unsigned OldEventsHasSCCWrite = OldEvents & (1 << SCC_WRITE);
+        if (!OldEventsHasSCCWrite) {
+          PendingSCCWrite = Other.PendingSCCWrite;
+        } else {
+          if (PendingSCCWrite != Other.PendingSCCWrite)
----------------
jayfoad wrote:

When control flow paths merge, each path could have a different unique scc-writing instruction. Then in the merged state PendingSCCWrite will be set to nullptr, indicating that we can no longer uniquely identify the scc-writing instruction.

https://github.com/llvm/llvm-project/pull/157843