[llvm] [UndefOrPoison] [CompileTime] Avoid IDom walk unless required. NFC (PR #90092)

[Nikita Popov via llvm-commits]

================
@@ -7283,31 +7283,35 @@ static bool isGuaranteedNotToBeUndefOrPoison(
   // BB1:
   //   CtxI ; V cannot be undef or poison here
   auto *Dominator = DNode->getIDom();
-  while (Dominator) {
-    auto *TI = Dominator->getBlock()->getTerminator();
-
-    Value *Cond = nullptr;
-    if (auto BI = dyn_cast_or_null<BranchInst>(TI)) {
-      if (BI->isConditional())
-        Cond = BI->getCondition();
-    } else if (auto SI = dyn_cast_or_null<SwitchInst>(TI)) {
-      Cond = SI->getCondition();
-    }
+  // This check is purely for compile time reasons: we can skip the IDom walk
+  // if what we are checking for includes undef and the value is not a boolean.
+  if (!includesUndef(Kind) || V->getType()->isIntOrIntVectorTy(1))
+    while (Dominator) {
+      auto *TI = Dominator->getBlock()->getTerminator();
+
+      Value *Cond = nullptr;
+      if (auto BI = dyn_cast_or_null<BranchInst>(TI)) {
+        if (BI->isConditional())
+          Cond = BI->getCondition();
+      } else if (auto SI = dyn_cast_or_null<SwitchInst>(TI)) {
----------------
nikic wrote:

Hm, I just realized that the switch case is applicable to all integer types, unlike the branch case, which only works for i1.

Maybe rather than trying to skip this code entirely we should add a limit on how many idoms we inspect? For practical purposes, limiting the walk to 10 idoms is probably quite sufficient.

https://github.com/llvm/llvm-project/pull/90092