[VPlan] Compute cost of more replicating loads/stores in ::computeCost.

Update VPReplicateRecipe::computeCost to compute costs of more
replicating loads/stores.

There are 2 cases that require extra checks to match the legacy cost
model:
 1. If the pointer is based on an induction, the legacy cost model
    passes its SCEV to getAddressComputationCost. In those cases, still
    fall back to the legacy cost. SCEV computations will be added  as
    follow-up
 2. If a load is used as part of an address of another load, the legacy
    cost model skips the scalarization overhead. Those cases are
    currently handled by a usedByLoadOrStore helper.

Note that getScalarizationOverhead also needs updating, because when
the legacy cost model computes the scalarization overhead, scalars have
not been collected yet, so we can't each for replicating recipes to skip
their cost, except other loads. This again can be further improved by
modeling inserts/extracts explicitly and consistently, and compute costs
for those operations directly where needed.

Author: fhahn

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -3925,7 +3925,8 @@ void LoopVectorizationPlanner::emitInvalidCostRemarks(
       if (VF.isScalar())
         continue;
 
-      VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind);
+      VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind,
+                            *CM.PSE.getSE());
       precomputeCosts(*Plan, VF, CostCtx);
       auto Iter = vp_depth_first_deep(Plan->getVectorLoopRegion()->getEntry());
       for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {
@@ -4182,7 +4183,8 @@ VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor() {
 
       // Add on other costs that are modelled in VPlan, but not in the legacy
       // cost model.
-      VPCostContext CostCtx(CM.TTI, *CM.TLI, *P, CM, CM.CostKind);
+      VPCostContext CostCtx(CM.TTI, *CM.TLI, *P, CM, CM.CostKind,
+                            *CM.PSE.getSE());
       VPRegionBlock *VectorRegion = P->getVectorLoopRegion();
       assert(VectorRegion && "Expected to have a vector region!");
       for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
@@ -6871,7 +6873,7 @@ LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,
 
 InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan,
                                                ElementCount VF) const {
-  VPCostContext CostCtx(CM.TTI, *CM.TLI, Plan, CM, CM.CostKind);
+  VPCostContext CostCtx(CM.TTI, *CM.TLI, Plan, CM, CM.CostKind, *PSE.getSE());
   InstructionCost Cost = precomputeCosts(Plan, VF, CostCtx);
 
   // Now compute and add the VPlan-based cost.
@@ -7082,7 +7084,8 @@ VectorizationFactor LoopVectorizationPlanner::computeBestVF() {
   // simplifications not accounted for in the legacy cost model. If that's the
   // case, don't trigger the assertion, as the extra simplifications may cause a
   // different VF to be picked by the VPlan-based cost model.
-  VPCostContext CostCtx(CM.TTI, *CM.TLI, BestPlan, CM, CM.CostKind);
+  VPCostContext CostCtx(CM.TTI, *CM.TLI, BestPlan, CM, CM.CostKind,
+                        *CM.PSE.getSE());
   precomputeCosts(BestPlan, BestFactor.Width, CostCtx);
   // Verify that the VPlan-based and legacy cost models agree, except for VPlans
   // with early exits and plans with additional VPlan simplifications. The
@@ -8704,7 +8707,8 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(
   // TODO: Enable following transform when the EVL-version of extended-reduction
   // and mulacc-reduction are implemented.
   if (!CM.foldTailWithEVL()) {
-    VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind);
+    VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind,
+                          *CM.PSE.getSE());
     VPlanTransforms::runPass(VPlanTransforms::convertToAbstractRecipes, *Plan,
                              CostCtx, Range);
   }
@@ -10043,7 +10047,7 @@ bool LoopVectorizePass::processLoop(Loop *L) {
     bool ForceVectorization =
         Hints.getForce() == LoopVectorizeHints::FK_Enabled;
     VPCostContext CostCtx(CM.TTI, *CM.TLI, LVP.getPlanFor(VF.Width), CM,
-                          CM.CostKind);
+                          CM.CostKind, *CM.PSE.getSE());
     if (!ForceVectorization &&
         !isOutsideLoopWorkProfitable(Checks, VF, L, PSE, CostCtx,
                                      LVP.getPlanFor(VF.Width), SEL,

llvm/lib/Transforms/Vectorize/VPlan.cpp

@@ -1778,8 +1778,10 @@ VPCostContext::getOperandInfo(VPValue *V) const {
   return TTI::getOperandInfo(V->getLiveInIRValue());
 }
 
-InstructionCost VPCostContext::getScalarizationOverhead(
-    Type *ResultTy, ArrayRef<const VPValue *> Operands, ElementCount VF) {
+InstructionCost
+VPCostContext::getScalarizationOverhead(Type *ResultTy,
+                                        ArrayRef<const VPValue *> Operands,
+                                        ElementCount VF, bool Skip) {
   if (VF.isScalar())
     return 0;
 
@@ -1799,7 +1801,10 @@ InstructionCost VPCostContext::getScalarizationOverhead(
   SmallPtrSet<const VPValue *, 4> UniqueOperands;
   SmallVector<Type *> Tys;
   for (auto *Op : Operands) {
-    if (Op->isLiveIn() || isa<VPReplicateRecipe, VPPredInstPHIRecipe>(Op) ||
+    if (Op->isLiveIn() ||
+        (!Skip && isa<VPReplicateRecipe, VPPredInstPHIRecipe>(Op)) ||
+        (isa<VPReplicateRecipe>(Op) &&
+         cast<VPReplicateRecipe>(Op)->getOpcode() == Instruction::Load) ||
         !UniqueOperands.insert(Op).second)
       continue;
     Tys.push_back(toVectorizedTy(Types.inferScalarType(Op), VF));

llvm/lib/Transforms/Vectorize/VPlanHelpers.h

@@ -349,12 +349,14 @@ struct VPCostContext {
   LoopVectorizationCostModel &CM;
   SmallPtrSet<Instruction *, 8> SkipCostComputation;
   TargetTransformInfo::TargetCostKind CostKind;
+  ScalarEvolution &SE;
 
   VPCostContext(const TargetTransformInfo &TTI, const TargetLibraryInfo &TLI,
                 const VPlan &Plan, LoopVectorizationCostModel &CM,
-                TargetTransformInfo::TargetCostKind CostKind)
+                TargetTransformInfo::TargetCostKind CostKind,
+                ScalarEvolution &SE)
       : TTI(TTI), TLI(TLI), Types(Plan), LLVMCtx(Plan.getContext()), CM(CM),
-        CostKind(CostKind) {}
+        CostKind(CostKind), SE(SE) {}
 
   /// Return the cost for \p UI with \p VF using the legacy cost model as
   /// fallback until computing the cost of all recipes migrates to VPlan.
@@ -374,10 +376,12 @@ struct VPCostContext {
 
   /// Estimate the overhead of scalarizing a recipe with result type \p ResultTy
   /// and \p Operands with \p VF. This is a convenience wrapper for the
-  /// type-based getScalarizationOverhead API.
-  InstructionCost getScalarizationOverhead(Type *ResultTy,
-                                           ArrayRef<const VPValue *> Operands,
-                                           ElementCount VF);
+  /// type-based getScalarizationOverhead API. If \p AlwaysIncludeReplicatingR
+  /// is true, always compute the cost of scalarizing replicating operands.
+  InstructionCost
+  getScalarizationOverhead(Type *ResultTy, ArrayRef<const VPValue *> Operands,
+                           ElementCount VF,
+                           bool AlwaysIncludeReplicatingR = false);
 };
 
 /// This class can be used to assign names to VPValues. For VPValues without

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -3075,6 +3075,63 @@ bool VPReplicateRecipe::shouldPack() const {
   });
 }
 
+/// Returns true if \p Ptr is a pointer computation for which the legacy cost
+/// model computes a SCEV expression when comping the address cost.
+static bool shouldUseAddressAccessSCEV(VPValue *Ptr) {
+  auto *PtrR = Ptr->getDefiningRecipe();
+  if (!PtrR || !((isa<VPReplicateRecipe>(PtrR) &&
+                  cast<VPReplicateRecipe>(PtrR)->getOpcode() ==
+                      Instruction::GetElementPtr) ||
+                 isa<VPWidenGEPRecipe>(PtrR)))
+    return false;
+
+  // We are looking for a gep with all loop invariant indices except for one
+  // which should be an induction variable.
+  unsigned NumOperands = PtrR->getNumOperands();
+  for (unsigned Idx = 1; Idx < NumOperands; ++Idx) {
+    VPValue *Opd = PtrR->getOperand(Idx);
+    if (!(Opd->isDefinedOutsideLoopRegions()) &&
+        !isa<VPScalarIVStepsRecipe, VPWidenIntOrFpInductionRecipe>(Opd))
+      return false;
+  }
+
+  return true;
+}
+
+/// Returns true of \p V is used as part of the address of another load or
+/// store.
+static bool isUsedByLoadStoreAddress(const VPUser *V) {
+  SmallPtrSet<const VPUser *, 4> Seen;
+  SmallVector<const VPUser *> WorkList = {V};
+
+  while (!WorkList.empty()) {
+    auto *Cur = dyn_cast<VPSingleDefRecipe>(WorkList.pop_back_val());
+    if (!Cur || !Seen.insert(Cur).second)
+      continue;
+
+    for (VPUser *U : Cur->users()) {
+      if (auto *InterleaveR = dyn_cast<VPInterleaveRecipe>(U))
+        if (InterleaveR->getAddr() == Cur)
+          return true;
+      if (auto *RepR = dyn_cast<VPReplicateRecipe>(U)) {
+        if (RepR->getOpcode() == Instruction::Load &&
+            RepR->getOperand(0) == Cur)
+          return true;
+        if (RepR->getOpcode() == Instruction::Store &&
+            RepR->getOperand(1) == Cur)
+          return true;
+      }
+      if (auto *MemR = dyn_cast<VPWidenMemoryRecipe>(U)) {
+        if (MemR->getAddr() == Cur && MemR->isConsecutive())
+          return true;
+      }
+    }
+
+    append_range(WorkList, cast<VPSingleDefRecipe>(Cur)->users());
+  }
+  return false;
+}
+
 InstructionCost VPReplicateRecipe::computeCost(ElementCount VF,
                                                VPCostContext &Ctx) const {
   Instruction *UI = cast<Instruction>(getUnderlyingValue());
@@ -3182,21 +3239,54 @@ InstructionCost VPReplicateRecipe::computeCost(ElementCount VF,
   }
   case Instruction::Load:
   case Instruction::Store: {
-    if (isSingleScalar()) {
-      bool IsLoad = UI->getOpcode() == Instruction::Load;
-      Type *ValTy = Ctx.Types.inferScalarType(IsLoad ? this : getOperand(0));
-      Type *ScalarPtrTy = Ctx.Types.inferScalarType(getOperand(IsLoad ? 0 : 1));
-      const Align Alignment = getLoadStoreAlignment(UI);
-      unsigned AS = getLoadStoreAddressSpace(UI);
-      TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(UI->getOperand(0));
-      InstructionCost ScalarMemOpCost = Ctx.TTI.getMemoryOpCost(
-          UI->getOpcode(), ValTy, Alignment, AS, Ctx.CostKind, OpInfo, UI);
-      return ScalarMemOpCost + Ctx.TTI.getAddressComputationCost(
-                                   ScalarPtrTy, nullptr, nullptr, Ctx.CostKind);
-    }
+    if (VF.isScalable() && !isSingleScalar())
+      return InstructionCost::getInvalid();
+
     // TODO: See getMemInstScalarizationCost for how to handle replicating and
     // predicated cases.
-    break;
+    if (getParent()->getParent() && getParent()->getParent()->isReplicator())
+      break;
+
+    bool IsLoad = UI->getOpcode() == Instruction::Load;
+    // TODO: Handle cases where we need to pass a SCEV to
+    // getAddressComputationCost.
+    if (shouldUseAddressAccessSCEV(getOperand(!IsLoad)))
+      break;
+
+    Type *ValTy = Ctx.Types.inferScalarType(IsLoad ? this : getOperand(0));
+    Type *ScalarPtrTy = Ctx.Types.inferScalarType(getOperand(IsLoad ? 0 : 1));
+    const Align Alignment = getLoadStoreAlignment(UI);
+    unsigned AS = getLoadStoreAddressSpace(UI);
+    TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(UI->getOperand(0));
+    InstructionCost ScalarMemOpCost = Ctx.TTI.getMemoryOpCost(
+        UI->getOpcode(), ValTy, Alignment, AS, Ctx.CostKind, OpInfo);
+
+    Type *PtrTy = isSingleScalar() ? ScalarPtrTy : toVectorTy(ScalarPtrTy, VF);
+
+    InstructionCost ScalarCost =
+        ScalarMemOpCost + Ctx.TTI.getAddressComputationCost(
+                              PtrTy, &Ctx.SE, nullptr, Ctx.CostKind);
+    if (isSingleScalar())
+      return ScalarCost;
+
+    SmallVector<const VPValue *> OpsToScalarize;
+    Type *ResultTy = Type::getVoidTy(getParent()->getPlan()->getContext());
+    // Set ResultTy and OpsToScalarize, if scalarization is needed. Currently we
+    // don't assign scalarization overhead in general, if the target prefers
+    // vectorized addressing or the loaded value is used as part of an address
+    // of another load or store.
+    if (Ctx.TTI.prefersVectorizedAddressing() ||
+        !isUsedByLoadStoreAddress(this)) {
+      if (!(IsLoad && !Ctx.TTI.prefersVectorizedAddressing()) &&
+          !(!IsLoad && Ctx.TTI.supportsEfficientVectorElementLoadStore()))
+        append_range(OpsToScalarize, operands());
+
+      if (!Ctx.TTI.supportsEfficientVectorElementLoadStore())
+        ResultTy = Ctx.Types.inferScalarType(this);
+    }
+
+    return (ScalarCost * VF.getFixedValue()) +
+           Ctx.getScalarizationOverhead(ResultTy, OpsToScalarize, VF, true);
   }
   }