[llvm] [VPlan] Add transform to replace VPWidenCanonicalIV with wide IV. (PR #194267)

[Florian Hahn via llvm-commits]

================
@@ -749,6 +749,56 @@ static void removeRedundantCanonicalIVs(VPlan &Plan) {
   WidenNewIV->eraseFromParent();
 }
 
+void VPlanTransforms::replaceWidenCanonicalIVWithWidenIV(
+    VPlan &Plan, ScalarEvolution &SE, const TargetTransformInfo &TTI,
+    TargetTransformInfo::TargetCostKind CostKind, ElementCount VF, unsigned UF,
+    const SmallPtrSetImpl<const Value *> &ValuesToIgnore) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  if (!LoopRegion || Plan.hasScalarVFOnly())
+    return;
+
+  VPValue *CanonicalIV = LoopRegion->getCanonicalIV();
+  auto *WideCanIV = vputils::findUserOf<VPWidenCanonicalIVRecipe>(CanonicalIV);
+  if (!WideCanIV || vputils::onlyFirstLaneUsed(WideCanIV) ||
+      vputils::onlyScalarValuesUsed(WideCanIV))
+    return;
+
+  // Introduce a new VPWidenIntOrFpInductionRecipe if profitable.
+  Type *CanIVTy = LoopRegion->getCanonicalIVType();
+  auto *VecTy = VectorType::get(CanIVTy, VF);
+  InstructionCost BroadcastCost = TTI.getShuffleCost(
+      TargetTransformInfo::SK_Broadcast, VecTy, VecTy, {}, CostKind);
+  InstructionCost PHICost = TTI.getCFInstrCost(Instruction::PHI, CostKind);
+  if (PHICost > BroadcastCost)
+    return;
+
+  // Bail out if the additional wide induction phi increase the expected spill
+  // cost.
+  VPRegisterUsage UnrolledBase =
+      calculateRegisterUsageForPlan(Plan, VF, TTI, ValuesToIgnore)[0];
+  for (unsigned &NumUsers : make_second_range(UnrolledBase.MaxLocalUsers))
+    NumUsers *= UF;
+  unsigned RegClass = TTI.getRegisterClassForType(/*Vector=*/true, VecTy);
+  VPRegisterUsage Projected = UnrolledBase;
+  Projected.MaxLocalUsers[RegClass] += 1;
+  if (Projected.spillCost(TTI, CostKind) >
+      UnrolledBase.spillCost(TTI, CostKind))
+    return;
+
+  Constant *Zero = Constant::getNullValue(CanIVTy);
+  InductionDescriptor ID(Zero, InductionDescriptor::IK_IntInduction,
+                         SE.getOne(CanIVTy));
+  VPValue *StepV = Plan.getConstantInt(CanIVTy, 1);
+  auto *NewWideIV = new VPWidenIntOrFpInductionRecipe(
+      /*IV=*/nullptr, Plan.getOrAddLiveIn(Zero), StepV, &Plan.getVF(), ID,
+      VPIRFlags::WrapFlagsTy(/*HasNUW=*/false, /*HasNSW=*/false),
----------------
fhahn wrote:

updated thanks

https://github.com/llvm/llvm-project/pull/194267