[llvm] [LoopVectorize] Add an option to test potential issues with VPlan (PR #121154)

[Florian Hahn via llvm-commits]

================
@@ -3448,8 +3453,9 @@ bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const {
   // If predication is not needed, avoid it.
   // TODO: We can use the loop-preheader as context point here and get
   // context sensitive reasoning for isSafeToSpeculativelyExecute.
+  Instruction *CtxI = CheckSpeculatabilityAtUse ? I->getNextNode() : nullptr;
----------------
fhahn wrote:

Hmm one thing I am missing is how `isSafeToSpeculativelyExecute == true` and the address being poison could happen. Even if the load is from a special address space or something like that, `isSafeToSpeculativelyExecute` should still consider the fact that the address could be poison?

https://github.com/llvm/llvm-project/pull/121154