[llvm] [FixIrreducible][UnifyLoopExits] Support switch statements (PR #206567)

[Robert Imschweiler via llvm-commits]

================
@@ -309,20 +309,23 @@ static bool fixIrreducible(Cycle &C, CycleInfo &CI, DominatorTree &DT,
                         << " -> " << printBasicBlock(Succ0)
                         << (Succ0 && Succ1 ? " " : "") << printBasicBlock(Succ1)
                         << '\n');
-    } else if (CallBrInst *CallBr = dyn_cast<CallBrInst>(P->getTerminator())) {
-      for (unsigned I = 0; I < CallBr->getNumSuccessors(); ++I) {
-        BasicBlock *Succ = CallBr->getSuccessor(I);
+    } else if (isa<CallBrInst>(P->getTerminator()) ||
+               isa<SwitchInst>(P->getTerminator())) {
+      Instruction *Term = P->getTerminator();
+      for (unsigned I = 0; I < Term->getNumSuccessors(); ++I) {
+        BasicBlock *Succ = Term->getSuccessor(I);
         if (Succ != Header)
           continue;
-        BasicBlock *NewSucc = SplitCallBrEdge(P, Succ, I, &DTU, &CI, LI);
+        BasicBlock *NewSucc = SplitMultiBrEdge(P, Succ, I, &DTU, &CI, LI);
         CHub.addBranch(NewSucc, Succ);
----------------
ro-i wrote:

uhm, yeah. It's not exactly the same issue as in the other PR since that only dealt with conditional branches (sorry, I mixed that up in my head), but it's also related to a subset of the duplicate successor situation, I think. I guess that there is a testing gap (for callbr/switch) for the duplicate successor + phi node situation. Do you want to handle this (probably best in a separate PR)? (Otherwise, I can also fix the callbr situation first if you want.)

https://github.com/llvm/llvm-project/pull/206567