[llvm] [AMDGPU][SIInsertWaitCnts] Use RegUnits-based tracking (PR #162077)

[Jay Foad via llvm-commits]

================
@@ -2421,21 +2391,20 @@ bool WaitcntBrackets::merge(const WaitcntBrackets &Other) {
       }
     }
 
-    for (int J = 0; J <= VgprUB; J++)
-      StrictDom |= mergeScore(M, VgprScores[T][J], Other.VgprScores[T][J]);
+    for (auto &[RegID, Info] : Other.VMem)
+      StrictDom |= mergeScore(M, VMem[RegID].Scores[T], Info.Scores[T]);
 
     if (isSmemCounter(T)) {
       unsigned Idx = getSgprScoresIdx(T);
-      for (int J = 0; J <= SgprUB; J++)
-        StrictDom |=
-            mergeScore(M, SgprScores[Idx][J], Other.SgprScores[Idx][J]);
+      for (auto &[RegID, Info] : Other.SGPRs)
+        StrictDom |= mergeScore(M, SGPRs[RegID].Scores[Idx], Info.Scores[Idx]);
     }
   }
 
-  for (int J = 0; J <= VgprUB; J++) {
-    unsigned char NewVmemTypes = VgprVmemTypes[J] | Other.VgprVmemTypes[J];
-    StrictDom |= NewVmemTypes != VgprVmemTypes[J];
-    VgprVmemTypes[J] = NewVmemTypes;
+  for (auto &[TID, Info] : Other.VMem) {
+    unsigned char NewVmemTypes = VMem[TID].VMEMTypes | Info.VMEMTypes;
+    StrictDom |= NewVmemTypes != VMem[TID].VMEMTypes;
+    VMem[TID].VMEMTypes = NewVmemTypes;
----------------
jayfoad wrote:

Can you avoid all these repeated lookups of the same key in the same map? Especially since we have already done this lookup in the other loop over Other.VMem on line 2394 above.

https://github.com/llvm/llvm-project/pull/162077