[VPlan] Compute cost for most opcodes in VPWidenRecipe (NFCI). (#98764)

Implement VPWidenRecipe::computeCost for most cases (except 
UDiv,SDiv,URem,SRem which require additional logic).

Note that this specializes `::computeCost` instead of `::cost`, as
`VPRecipeBase::cost` is responsible for skipping cost-computations
for pre-computed recipes for now.

The most recent version of the VPlan-based cost model introduction 
has been committed on Jul 10 (b841e2e) and we should
probably give it at least a week in case additional mismatches surface.

PR: #98764

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -4354,8 +4354,8 @@ void LoopVectorizationPlanner::emitInvalidCostRemarks(
   SmallVector<RecipeVFPair> InvalidCosts;
   for (const auto &Plan : VPlans) {
     for (ElementCount VF : Plan->vectorFactors()) {
-      VPCostContext CostCtx(CM.TTI, Legal->getWidestInductionType(), LLVMCtx,
-                            CM);
+      VPCostContext CostCtx(CM.TTI, *CM.TLI, Legal->getWidestInductionType(),
+                            LLVMCtx, CM);
       auto Iter = vp_depth_first_deep(Plan->getVectorLoopRegion()->getEntry());
       for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {
         for (auto &R : *VPBB) {
@@ -7062,7 +7062,8 @@ InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan,
                                                ElementCount VF) const {
   InstructionCost Cost = 0;
   LLVMContext &LLVMCtx = OrigLoop->getHeader()->getContext();
-  VPCostContext CostCtx(CM.TTI, Legal->getWidestInductionType(), LLVMCtx, CM);
+  VPCostContext CostCtx(CM.TTI, *CM.TLI, 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

llvm/lib/Transforms/Vectorize/VPlan.h

@@ -736,14 +736,16 @@ class VPLiveOut : public VPUser {
 /// Struct to hold various analysis needed for cost computations.
 struct VPCostContext {
   const TargetTransformInfo &TTI;
+  const TargetLibraryInfo &TLI;
   VPTypeAnalysis Types;
   LLVMContext &LLVMCtx;
   LoopVectorizationCostModel &CM;
   SmallPtrSet<Instruction *, 8> SkipCostComputation;
 
-  VPCostContext(const TargetTransformInfo &TTI, Type *CanIVTy,
-                LLVMContext &LLVMCtx, LoopVectorizationCostModel &CM)
-      : TTI(TTI), Types(CanIVTy, LLVMCtx), LLVMCtx(LLVMCtx), CM(CM) {}
+  VPCostContext(const TargetTransformInfo &TTI, const TargetLibraryInfo &TLI,
+                Type *CanIVTy, LLVMContext &LLVMCtx,
+                LoopVectorizationCostModel &CM)
+      : TTI(TTI), TLI(TLI), Types(CanIVTy, LLVMCtx), LLVMCtx(LLVMCtx), CM(CM) {}
 
   /// 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.
@@ -796,7 +798,7 @@ class VPRecipeBase : public ilist_node_with_parent<VPRecipeBase, VPBasicBlock>,
   /// Return the cost of this recipe, taking into account if the cost
   /// computation should be skipped and the ForceTargetInstructionCost flag.
   /// Also takes care of printing the cost for debugging.
-  virtual InstructionCost cost(ElementCount VF, VPCostContext &Ctx);
+  InstructionCost cost(ElementCount VF, VPCostContext &Ctx);
 
   /// Insert an unlinked recipe into a basic block immediately before
   /// the specified recipe.
@@ -860,9 +862,11 @@ class VPRecipeBase : public ilist_node_with_parent<VPRecipeBase, VPBasicBlock>,
   DebugLoc getDebugLoc() const { return DL; }
 
 protected:
-  /// Compute the cost of this recipe using the legacy cost model and the
-  /// underlying instructions.
-  InstructionCost computeCost(ElementCount VF, VPCostContext &Ctx) const;
+  /// Compute the cost of this recipe either using a recipe's specialized
+  /// implementation or using the legacy cost model and the underlying
+  /// instructions.
+  virtual InstructionCost computeCost(ElementCount VF,
+                                      VPCostContext &Ctx) const;
 };
 
 // Helper macro to define common classof implementations for recipes.
@@ -1426,6 +1430,10 @@ class VPWidenRecipe : public VPRecipeWithIRFlags {
   /// processing State.VF elements.
   void execute(VPTransformState &State) override;
 
+  /// Return the cost of this VPWidenRecipe.
+  InstructionCost computeCost(ElementCount VF,
+                              VPCostContext &Ctx) const override;
+
   unsigned getOpcode() const { return Opcode; }
 
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)

llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp

@@ -1140,6 +1140,80 @@ void VPWidenRecipe::execute(VPTransformState &State) {
 #endif
 }
 
+InstructionCost VPWidenRecipe::computeCost(ElementCount VF,
+                                           VPCostContext &Ctx) const {
+  TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
+  switch (Opcode) {
+  case Instruction::FNeg: {
+    Type *VectorTy =
+        ToVectorTy(Ctx.Types.inferScalarType(this->getVPSingleValue()), VF);
+    return Ctx.TTI.getArithmeticInstrCost(
+        Opcode, VectorTy, CostKind,
+        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},
+        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None});
+  }
+
+  case Instruction::UDiv:
+  case Instruction::SDiv:
+  case Instruction::SRem:
+  case Instruction::URem:
+    // More complex computation, let the legacy cost-model handle this for now.
+    return Ctx.getLegacyCost(cast<Instruction>(getUnderlyingValue()), VF);
+  case Instruction::Add:
+  case Instruction::FAdd:
+  case Instruction::Sub:
+  case Instruction::FSub:
+  case Instruction::Mul:
+  case Instruction::FMul:
+  case Instruction::FDiv:
+  case Instruction::FRem:
+  case Instruction::Shl:
+  case Instruction::LShr:
+  case Instruction::AShr:
+  case Instruction::And:
+  case Instruction::Or:
+  case Instruction::Xor: {
+    VPValue *RHS = getOperand(1);
+    // Certain instructions can be cheaper to vectorize if they have a constant
+    // second vector operand. One example of this are shifts on x86.
+    TargetTransformInfo::OperandValueInfo RHSInfo = {
+        TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None};
+    if (RHS->isLiveIn())
+      RHSInfo = Ctx.TTI.getOperandInfo(RHS->getLiveInIRValue());
+
+    if (RHSInfo.Kind == TargetTransformInfo::OK_AnyValue &&
+        getOperand(1)->isDefinedOutsideVectorRegions())
+      RHSInfo.Kind = TargetTransformInfo::OK_UniformValue;
+    Type *VectorTy =
+        ToVectorTy(Ctx.Types.inferScalarType(this->getVPSingleValue()), VF);
+    Instruction *CtxI = dyn_cast_or_null<Instruction>(getUnderlyingValue());
+
+    SmallVector<const Value *, 4> Operands;
+    if (CtxI)
+      Operands.append(CtxI->value_op_begin(), CtxI->value_op_end());
+    return Ctx.TTI.getArithmeticInstrCost(
+        Opcode, VectorTy, CostKind,
+        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},
+        RHSInfo, Operands, CtxI, &Ctx.TLI);
+  }
+  case Instruction::Freeze: {
+    // This opcode is unknown. Assume that it is the same as 'mul'.
+    Type *VectorTy =
+        ToVectorTy(Ctx.Types.inferScalarType(this->getVPSingleValue()), VF);
+    return Ctx.TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);
+  }
+  case Instruction::ICmp:
+  case Instruction::FCmp: {
+    Instruction *CtxI = dyn_cast_or_null<Instruction>(getUnderlyingValue());
+    Type *VectorTy = ToVectorTy(Ctx.Types.inferScalarType(getOperand(0)), VF);
+    return Ctx.TTI.getCmpSelInstrCost(Opcode, VectorTy, nullptr, getPredicate(),
+                                      CostKind, CtxI);
+  }
+  default:
+    llvm_unreachable("Unsupported opcode for instruction");
+  }
+}
+
 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 void VPWidenRecipe::print(raw_ostream &O, const Twine &Indent,
                           VPSlotTracker &SlotTracker) const {