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

[Danila Malyutin 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;
----------------
danilaml wrote:

@fhahn That's the thing. The address IS NOT poison initially and it CAN'T happen in original IR for which `isSafeToSpeculativelyExecute` is called (if I understood your question correctly). Load from poison only happens due to what I believe incorrect lowering of VPlan to IR (e.g. it does `select cond, (op x, poison), y` instead of `op x, (select cond, x, z)` or something else valid). I'd try to dig further, but unfortunately I couldn't make heads and tails of VPlan<>IR interop (downside of separate IR).

In short, `isSafeToSpeculativelyExecute == true` is fine. The address being poison is the bug introduced by LoopVectorize/VPlan. This patch (or alternative patches described in the linked issue) are an attempt to reproduce this issue on current main since after some changes I was unable to reproduce this issue with the upstream code alone (no way to specify always-deferenceable loads that are not args).

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