[llvm] [VPlan] Port invalid cost remarks to VPlan. (PR #99322)

[via llvm-commits]

================
@@ -4350,24 +4347,38 @@ bool LoopVectorizationPlanner::isMoreProfitable(
   return CmpFn(RTCostA, RTCostB);
 }
 
-static void emitInvalidCostRemarks(SmallVector<InstructionVFPair> InvalidCosts,
-                                   OptimizationRemarkEmitter *ORE,
-                                   Loop *TheLoop) {
+void LoopVectorizationPlanner::emitInvalidCostRemarks(
+    OptimizationRemarkEmitter *ORE) {
+  using RecipeVFPair = std::pair<VPRecipeBase *, ElementCount>;
+  LLVMContext &LLVMCtx = OrigLoop->getHeader()->getContext();
+  SmallVector<RecipeVFPair> InvalidCosts;
+  for (const auto &Plan : VPlans) {
+    for (ElementCount VF : Plan->vectorFactors()) {
+      VPCostContext CostCtx(CM.TTI, Legal->getWidestInductionType(), LLVMCtx,
+                            CM);
----------------
ayalz wrote:

CostCtx is kept here rather than hoisting it alongside LLVMCtx above, due to its caching of SkipCostComputation. But the latter is initialized by calling LVP::cost(Plan, VF), whereas here all recipes are asked for their cost directly, w/o going through LVP::cost(). Should LVP::cost() be called first, and iff it returns invalid traverse the recipes? Or note that invalid costs cannot be skipped(?), so calling LVP::cost() is redundant when only invalid costs are sought, in which case CostCtx can be hoisted(?)

https://github.com/llvm/llvm-project/pull/99322