Recommit "[VPlan] First step towards VPlan cost modeling. (#92555)"

This reverts commit 6f538f6.

A number of crashes have been fixed by separate fixes, including
ttps://github.com//pull/96622. This version of the
PR also pre-computes the costs for branches (except the latch) instead
of computing their costs as part of costing of replicate regions, as
there may not be a direct correspondence between original branches and
number of replicate regions.

Original message:
This adds a new interface to compute the cost of recipes, VPBasicBlocks,
VPRegionBlocks and VPlan, initially falling back to the legacy cost model
for all recipes. Follow-up patches will gradually migrate recipes to
compute their own costs step-by-step.

It also adds getBestPlan function to LVP which computes the cost of all
VPlans and picks the most profitable one together with the most
profitable VF.

The VPlan selected by the VPlan cost model is executed and there is an
assert to catch cases where the VPlan cost model and the legacy cost
model disagree. Even though I checked a number of different build
configurations on AArch64 and X86, there may be some differences
that have been missed.

Additional discussions and context can be found in @arcbbb's
#67647 and
#67934 which is an earlier
version of the current PR.

PR: #92555

llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h

@@ -334,6 +334,16 @@ class LoopVectorizationPlanner {
   /// A builder used to construct the current plan.
   VPBuilder Builder;
 
+  /// Computes the cost of \p Plan for vectorization factor \p VF.
+  ///
+  /// The current implementation requires access to the
+  /// LoopVectorizationLegality to handle inductions and reductions, which is
+  /// why it is kept separate from the VPlan-only cost infrastructure.
+  ///
+  /// TODO: Move to VPlan::cost once the use of LoopVectorizationLegality has
+  /// been retired.
+  InstructionCost cost(VPlan &Plan, ElementCount VF) const;
+
 public:
   LoopVectorizationPlanner(
       Loop *L, LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI,
@@ -355,6 +365,9 @@ class LoopVectorizationPlanner {
   /// Return the best VPlan for \p VF.
   VPlan &getBestPlanFor(ElementCount VF) const;
 
+  /// Return the most profitable plan and fix its VF to the most profitable one.
+  VPlan &getBestPlan() const;
+
   /// Generate the IR code for the vectorized loop captured in VPlan \p BestPlan
   /// according to the best selected \p VF and  \p UF.
   ///
@@ -434,6 +447,9 @@ class LoopVectorizationPlanner {
 
   /// \return The most profitable vectorization factor for the available VPlans
   /// and the cost of that VF.
+  /// This is now only used to verify the decisions by the new VPlan-based
+  /// cost-model and will be retired once the VPlan-based cost-model is
+  /// stabilized.
   VectorizationFactor selectVectorizationFactor();
 
   /// Returns true if the per-lane cost of VectorizationFactor A is lower than

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -290,7 +290,7 @@ static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(
     cl::desc("A flag that overrides the target's max interleave factor for "
              "vectorized loops."));
 
-static cl::opt<unsigned> ForceTargetInstructionCost(
+cl::opt<unsigned> ForceTargetInstructionCost(
     "force-target-instruction-cost", cl::init(0), cl::Hidden,
     cl::desc("A flag that overrides the target's expected cost for "
              "an instruction to a single constant value. Mostly "
@@ -412,14 +412,6 @@ static bool hasIrregularType(Type *Ty, const DataLayout &DL) {
   return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);
 }
 
-/// A helper function that returns the reciprocal of the block probability of
-/// predicated blocks. If we return X, we are assuming the predicated block
-/// will execute once for every X iterations of the loop header.
-///
-/// TODO: We should use actual block probability here, if available. Currently,
-///       we always assume predicated blocks have a 50% chance of executing.
-static unsigned getReciprocalPredBlockProb() { return 2; }
-
 /// Returns "best known" trip count for the specified loop \p L as defined by
 /// the following procedure:
 ///   1) Returns exact trip count if it is known.
@@ -1608,6 +1600,16 @@ class LoopVectorizationCostModel {
   /// \p VF is the vectorization factor chosen for the original loop.
   bool isEpilogueVectorizationProfitable(const ElementCount VF) const;
 
+  /// Returns the execution time cost of an instruction for a given vector
+  /// width. Vector width of one means scalar.
+  InstructionCost getInstructionCost(Instruction *I, ElementCount VF);
+
+  /// Return the cost of instructions in an inloop reduction pattern, if I is
+  /// part of that pattern.
+  std::optional<InstructionCost>
+  getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy,
+                          TTI::TargetCostKind CostKind) const;
+
 private:
   unsigned NumPredStores = 0;
 
@@ -1633,16 +1635,6 @@ class LoopVectorizationCostModel {
   /// of elements.
   ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements);
 
-  /// Returns the execution time cost of an instruction for a given vector
-  /// width. Vector width of one means scalar.
-  InstructionCost getInstructionCost(Instruction *I, ElementCount VF);
-
-  /// Return the cost of instructions in an inloop reduction pattern, if I is
-  /// part of that pattern.
-  std::optional<InstructionCost>
-  getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy,
-                          TTI::TargetCostKind CostKind) const;
-
   /// Calculate vectorization cost of memory instruction \p I.
   InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF);
 
@@ -7288,7 +7280,10 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) {
              [](std::unique_ptr<VPlan> &P) { return P->hasScalarVFOnly(); }))
     return VectorizationFactor::Disabled();
 
-  // Select the optimal vectorization factor.
+  // Select the optimal vectorization factor according to the legacy cost-model.
+  // This is now only used to verify the decisions by the new VPlan-based
+  // cost-model and will be retired once the VPlan-based cost-model is
+  // stabilized.
   VectorizationFactor VF = selectVectorizationFactor();
   assert((VF.Width.isScalar() || VF.ScalarCost > 0) && "when vectorizing, the scalar cost must be non-zero.");
   if (!hasPlanWithVF(VF.Width)) {
@@ -7299,6 +7294,211 @@ LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) {
   return VF;
 }
 
+InstructionCost VPCostContext::getLegacyCost(Instruction *UI,
+                                             ElementCount VF) const {
+  return CM.getInstructionCost(UI, VF);
+}
+
+bool VPCostContext::skipCostComputation(Instruction *UI, bool IsVector) const {
+  return CM.ValuesToIgnore.contains(UI) ||
+         (IsVector && CM.VecValuesToIgnore.contains(UI)) ||
+         SkipCostComputation.contains(UI);
+}
+
+InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan,
+                                               ElementCount VF) const {
+  InstructionCost Cost = 0;
+  LLVMContext &LLVMCtx = OrigLoop->getHeader()->getContext();
+  VPCostContext CostCtx(CM.TTI, Legal->getWidestInductionType(), LLVMCtx, CM);
+
+  // Cost modeling for inductions is inaccurate in the legacy cost model
+  // compared to the recipes that are generated. To match here initially during
+  // VPlan cost model bring up directly use the induction costs from the legacy
+  // cost model. Note that we do this as pre-processing; the VPlan may not have
+  // any recipes associated with the original induction increment instruction
+  // and may replace truncates with VPWidenIntOrFpInductionRecipe. We precompute
+  // the cost of induction phis and increments (both that are represented by
+  // recipes and those that are not), to avoid distinguishing between them here,
+  // and skip all recipes that represent induction phis and increments (the
+  // former case) later on, if they exist, to avoid counting them twice.
+  // Similarly we pre-compute the cost of any optimized truncates.
+  // TODO: Switch to more accurate costing based on VPlan.
+  for (const auto &[IV, IndDesc] : Legal->getInductionVars()) {
+    Instruction *IVInc = cast<Instruction>(
+        IV->getIncomingValueForBlock(OrigLoop->getLoopLatch()));
+    SmallVector<Instruction *> IVInsts = {IV, IVInc};
+    for (User *U : IV->users()) {
+      auto *CI = cast<Instruction>(U);
+      if (!CostCtx.CM.isOptimizableIVTruncate(CI, VF))
+        continue;
+      IVInsts.push_back(CI);
+    }
+    for (Instruction *IVInst : IVInsts) {
+      if (!CostCtx.SkipCostComputation.insert(IVInst).second)
+        continue;
+      InstructionCost InductionCost = CostCtx.getLegacyCost(IVInst, VF);
+      LLVM_DEBUG({
+        dbgs() << "Cost of " << InductionCost << " for VF " << VF
+               << ": induction instruction " << *IVInst << "\n";
+      });
+      Cost += InductionCost;
+    }
+  }
+
+  /// Compute the cost of all exiting conditions of the loop using the legacy
+  /// cost model. This is to match the legacy behavior, which adds the cost of
+  /// all exit conditions. Note that this over-estimates the cost, as there will
+  /// be a single condition to control the vector loop.
+  SmallVector<BasicBlock *> Exiting;
+  CM.TheLoop->getExitingBlocks(Exiting);
+  SetVector<Instruction *> ExitInstrs;
+  // Collect all exit conditions.
+  for (BasicBlock *EB : Exiting) {
+    auto *Term = dyn_cast<BranchInst>(EB->getTerminator());
+    if (!Term)
+      continue;
+    if (auto *CondI = dyn_cast<Instruction>(Term->getOperand(0))) {
+      ExitInstrs.insert(CondI);
+    }
+  }
+  // Compute the cost of all instructions only feeding the exit conditions.
+  for (unsigned I = 0; I != ExitInstrs.size(); ++I) {
+    Instruction *CondI = ExitInstrs[I];
+    if (!OrigLoop->contains(CondI) ||
+        !CostCtx.SkipCostComputation.insert(CondI).second)
+      continue;
+    Cost += CostCtx.getLegacyCost(CondI, VF);
+    for (Value *Op : CondI->operands()) {
+      auto *OpI = dyn_cast<Instruction>(Op);
+      if (!OpI || any_of(OpI->users(), [&ExitInstrs](User *U) {
+            return !ExitInstrs.contains(cast<Instruction>(U));
+          }))
+        continue;
+      ExitInstrs.insert(OpI);
+    }
+  }
+
+  // The legacy cost model has special logic to compute the cost of in-loop
+  // reductions, which may be smaller than the sum of all instructions involved
+  // in the reduction. For AnyOf reductions, VPlan codegen may remove the select
+  // which the legacy cost model uses to assign cost. Pre-compute their costs
+  // for now.
+  // TODO: Switch to costing based on VPlan once the logic has been ported.
+  for (const auto &[RedPhi, RdxDesc] : Legal->getReductionVars()) {
+    if (!CM.isInLoopReduction(RedPhi) &&
+        !RecurrenceDescriptor::isAnyOfRecurrenceKind(
+            RdxDesc.getRecurrenceKind()))
+      continue;
+
+    // AnyOf reduction codegen may remove the select. To match the legacy cost
+    // model, pre-compute the cost for AnyOf reductions here.
+    if (RecurrenceDescriptor::isAnyOfRecurrenceKind(
+            RdxDesc.getRecurrenceKind())) {
+      auto *Select = cast<SelectInst>(*find_if(
+          RedPhi->users(), [](User *U) { return isa<SelectInst>(U); }));
+      assert(!CostCtx.SkipCostComputation.contains(Select) &&
+             "reduction op visited multiple times");
+      CostCtx.SkipCostComputation.insert(Select);
+      auto ReductionCost = CostCtx.getLegacyCost(Select, VF);
+      LLVM_DEBUG(dbgs() << "Cost of " << ReductionCost << " for VF " << VF
+                        << ":\n any-of reduction " << *Select << "\n");
+      Cost += ReductionCost;
+      continue;
+    }
+
+    const auto &ChainOps = RdxDesc.getReductionOpChain(RedPhi, OrigLoop);
+    SetVector<Instruction *> ChainOpsAndOperands(ChainOps.begin(),
+                                                 ChainOps.end());
+    // Also include the operands of instructions in the chain, as the cost-model
+    // may mark extends as free.
+    for (auto *ChainOp : ChainOps) {
+      for (Value *Op : ChainOp->operands()) {
+        if (auto *I = dyn_cast<Instruction>(Op))
+          ChainOpsAndOperands.insert(I);
+      }
+    }
+
+    // Pre-compute the cost for I, if it has a reduction pattern cost.
+    for (Instruction *I : ChainOpsAndOperands) {
+      auto ReductionCost = CM.getReductionPatternCost(
+          I, VF, ToVectorTy(I->getType(), VF), TTI::TCK_RecipThroughput);
+      if (!ReductionCost)
+        continue;
+
+      assert(!CostCtx.SkipCostComputation.contains(I) &&
+             "reduction op visited multiple times");
+      CostCtx.SkipCostComputation.insert(I);
+      LLVM_DEBUG(dbgs() << "Cost of " << ReductionCost << " for VF " << VF
+                        << ":\n in-loop reduction " << *I << "\n");
+      Cost += *ReductionCost;
+    }
+  }
+
+  // Pre-compute the costs for branches except for the backedge, as the number
+  // of replicate regions in a VPlan may not directly match the number of
+  // branches, which would lead to different decisions.
+  // TODO: Compute cost of branches for each replicate region in the VPlan,
+  // which is more accurate than the legacy cost model.
+  for (BasicBlock *BB : OrigLoop->blocks()) {
+    if (BB == OrigLoop->getLoopLatch())
+      continue;
+    CostCtx.SkipCostComputation.insert(BB->getTerminator());
+    auto BranchCost = CostCtx.getLegacyCost(BB->getTerminator(), VF);
+    Cost += BranchCost;
+  }
+  // Now compute and add the VPlan-based cost.
+  Cost += Plan.cost(VF, CostCtx);
+  LLVM_DEBUG(dbgs() << "Cost for VF " << VF << ": " << Cost << "\n");
+  return Cost;
+}
+
+VPlan &LoopVectorizationPlanner::getBestPlan() const {
+  // If there is a single VPlan with a single VF, return it directly.
+  VPlan &FirstPlan = *VPlans[0];
+  if (VPlans.size() == 1 && size(FirstPlan.vectorFactors()) == 1)
+    return FirstPlan;
+
+  VPlan *BestPlan = &FirstPlan;
+  ElementCount ScalarVF = ElementCount::getFixed(1);
+  assert(hasPlanWithVF(ScalarVF) &&
+         "More than a single plan/VF w/o any plan having scalar VF");
+
+  // TODO: Compute scalar cost using VPlan-based cost model.
+  InstructionCost ScalarCost = CM.expectedCost(ScalarVF);
+  VectorizationFactor BestFactor(ScalarVF, ScalarCost, ScalarCost);
+
+  bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled;
+  if (ForceVectorization) {
+    // Ignore scalar width, because the user explicitly wants vectorization.
+    // Initialize cost to max so that VF = 2 is, at least, chosen during cost
+    // evaluation.
+    BestFactor.Cost = InstructionCost::getMax();
+  }
+
+  for (auto &P : VPlans) {
+    for (ElementCount VF : P->vectorFactors()) {
+      if (VF.isScalar())
+        continue;
+      if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) {
+        LLVM_DEBUG(
+            dbgs()
+            << "LV: Not considering vector loop of width " << VF
+            << " because it will not generate any vector instructions.\n");
+        continue;
+      }
+
+      InstructionCost Cost = cost(*P, VF);
+      VectorizationFactor CurrentFactor(VF, Cost, ScalarCost);
+      if (isMoreProfitable(CurrentFactor, BestFactor)) {
+        BestFactor = CurrentFactor;
+        BestPlan = &*P;
+      }
+    }
+  }
+  BestPlan->setVF(BestFactor.Width);
+  return *BestPlan;
+}
+
 VPlan &LoopVectorizationPlanner::getBestPlanFor(ElementCount VF) const {
   assert(count_if(VPlans,
                   [VF](const VPlanPtr &Plan) { return Plan->hasVF(VF); }) ==
@@ -10169,8 +10369,15 @@ bool LoopVectorizePass::processLoop(Loop *L) {
                                VF.MinProfitableTripCount, IC, &LVL, &CM, BFI,
                                PSI, Checks);
 
-        VPlan &BestPlan = LVP.getBestPlanFor(VF.Width);
-        LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false);
+        VPlan &BestPlan = LVP.getBestPlan();
+        assert(size(BestPlan.vectorFactors()) == 1 &&
+               "Plan should have a single VF");
+        ElementCount Width = *BestPlan.vectorFactors().begin();
+        LLVM_DEBUG(dbgs() << "VF picked by VPlan cost model: " << Width
+                          << "\n");
+        assert(VF.Width == Width &&
+               "VPlan cost model and legacy cost model disagreed");
+        LVP.executePlan(Width, IC, BestPlan, LB, DT, false);
         ++LoopsVectorized;
 
         // Add metadata to disable runtime unrolling a scalar loop when there

llvm/lib/Transforms/Vectorize/VPlan.cpp

@@ -776,6 +776,48 @@ void VPRegionBlock::execute(VPTransformState *State) {
   State->Instance.reset();
 }
 
+InstructionCost VPBasicBlock::cost(ElementCount VF, VPCostContext &Ctx) {
+  InstructionCost Cost = 0;
+  for (VPRecipeBase &R : Recipes)
+    Cost += R.cost(VF, Ctx);
+  return Cost;
+}
+
+InstructionCost VPRegionBlock::cost(ElementCount VF, VPCostContext &Ctx) {
+  if (!isReplicator()) {
+    InstructionCost Cost = 0;
+    for (VPBlockBase *Block : vp_depth_first_shallow(getEntry()))
+      Cost += Block->cost(VF, Ctx);
+    InstructionCost BackedgeCost =
+        Ctx.TTI.getCFInstrCost(Instruction::Br, TTI::TCK_RecipThroughput);
+    LLVM_DEBUG(dbgs() << "Cost of " << BackedgeCost << " for VF " << VF
+                      << ": vector loop backedge\n");
+    Cost += BackedgeCost;
+    return Cost;
+  }
+
+  // Compute the cost of a replicate region. Replicating isn't supported for
+  // scalable vectors, return an invalid cost for them.
+  // TODO: Discard scalable VPlans with replicate recipes earlier after
+  // construction.
+  if (VF.isScalable())
+    return InstructionCost::getInvalid();
+
+  // First compute the cost of the conditionally executed recipes, followed by
+  // account for the branching cost, except if the mask is a header mask or
+  // uniform condition.
+  using namespace llvm::VPlanPatternMatch;
+  VPBasicBlock *Then = cast<VPBasicBlock>(getEntry()->getSuccessors()[0]);
+  InstructionCost ThenCost = Then->cost(VF, Ctx);
+
+  // For the scalar case, we may not always execute the original predicated
+  // block, Thus, scale the block's cost by the probability of executing it.
+  if (VF.isScalar())
+    return ThenCost / getReciprocalPredBlockProb();
+
+  return ThenCost;
+}
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 void VPRegionBlock::print(raw_ostream &O, const Twine &Indent,
                           VPSlotTracker &SlotTracker) const {
@@ -1040,6 +1082,12 @@ void VPlan::execute(VPTransformState *State) {
          "DT not preserved correctly");
 }
 
+InstructionCost VPlan::cost(ElementCount VF, VPCostContext &Ctx) {
+  // For now only return the cost of the vector loop region, ignoring any other
+  // blocks, like the preheader or middle blocks.
+  return getVectorLoopRegion()->cost(VF, Ctx);
+}
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 void VPlan::printLiveIns(raw_ostream &O) const {
   VPSlotTracker SlotTracker(this);