[llvm] [LV][VPlan] Not adding shuffle cost when store loop invariant value. (PR #109644)

[Elvis Wang via llvm-commits]

================
@@ -5666,7 +5666,10 @@ LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,
   }
 
   bool Reverse = ConsecutiveStride < 0;
-  if (Reverse)
+  const StoreInst *SI = dyn_cast<StoreInst>(I);
+  bool IsLoopInvariantStoreValue =
+      SI && Legal->isInvariant(const_cast<StoreInst *>(SI)->getValueOperand());
----------------
ElvisWang123 wrote:

You are right, using this implementation may introduce extra divergences. 
I think the implementation in the legacy cost model is what we want but we cannot get the analysis in the VPlanRecipes.

Do you have other better methods to figure out if the value is loop invariant in the VPlanRecipes?

https://github.com/llvm/llvm-project/pull/109644