| 1 | //===- SLPCostAnalysis.cpp - SLP Vectorizer free cost helpers -------------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #include "SLPCostAnalysis.h" |
| 10 | |
| 11 | #include "llvm/ADT/STLExtras.h" |
| 12 | #include "llvm/ADT/SmallVector.h" |
| 13 | #include "llvm/IR/DerivedTypes.h" |
| 14 | #include "llvm/IR/Instructions.h" |
| 15 | #include "llvm/IR/Operator.h" |
| 16 | #include "llvm/IR/Value.h" |
| 17 | #include "llvm/Support/Casting.h" |
| 18 | |
| 19 | #include <utility> |
| 20 | |
| 21 | using namespace llvm; |
| 22 | |
| 23 | namespace llvm::slpvectorizer { |
| 24 | |
| 25 | InstructionCost getShuffleCost(const TargetTransformInfo &TTI, |
| 26 | TTI::ShuffleKind Kind, VectorType *Tp, |
| 27 | const TTI::TargetCostKind CostKind, |
| 28 | ArrayRef<int> Mask, int Index, VectorType *SubTp, |
| 29 | ArrayRef<const Value *> Args) { |
| 30 | VectorType *DstTy = Tp; |
| 31 | if (!Mask.empty()) |
| 32 | DstTy = FixedVectorType::get(ElementType: Tp->getScalarType(), NumElts: Mask.size()); |
| 33 | |
| 34 | if (Kind != TTI::SK_PermuteTwoSrc) |
| 35 | return TTI.getShuffleCost(Kind, DstTy, SrcTy: Tp, CostKind, Mask, Index, SubTp, |
| 36 | Args); |
| 37 | int NumSrcElts = Tp->getElementCount().getKnownMinValue(); |
| 38 | int NumSubElts; |
| 39 | if (Mask.size() > 2 && ShuffleVectorInst::isInsertSubvectorMask( |
| 40 | Mask, NumSrcElts, NumSubElts, Index)) { |
| 41 | if (Index + NumSubElts > NumSrcElts && |
| 42 | Index + NumSrcElts <= static_cast<int>(Mask.size())) |
| 43 | return TTI.getShuffleCost(Kind: TTI::SK_InsertSubvector, DstTy, SrcTy: Tp, CostKind, |
| 44 | Mask, Index, SubTp: Tp); |
| 45 | } |
| 46 | return TTI.getShuffleCost(Kind, DstTy, SrcTy: Tp, CostKind, Mask, Index, SubTp, |
| 47 | Args); |
| 48 | } |
| 49 | |
| 50 | std::pair<InstructionCost, InstructionCost> |
| 51 | getGEPCosts(const TargetTransformInfo &TTI, ArrayRef<Value *> Ptrs, |
| 52 | Value *BasePtr, unsigned Opcode, const TTI::TargetCostKind CostKind, |
| 53 | Type *ScalarTy, VectorType *VecTy) { |
| 54 | InstructionCost ScalarCost = 0; |
| 55 | InstructionCost VecCost = 0; |
| 56 | // Here we differentiate two cases: (1) when Ptrs represent a regular |
| 57 | // vectorization tree node (as they are pointer arguments of scattered |
| 58 | // loads) or (2) when Ptrs are the arguments of loads or stores being |
| 59 | // vectorized as plane wide unit-stride load/store since all the |
| 60 | // loads/stores are known to be from/to adjacent locations. |
| 61 | if (Opcode == Instruction::Load || Opcode == Instruction::Store) { |
| 62 | // Case 2: estimate costs for pointer related costs when vectorizing to |
| 63 | // a wide load/store. |
| 64 | // Scalar cost is estimated as a set of pointers with known relationship |
| 65 | // between them. |
| 66 | // For vector code we will use BasePtr as argument for the wide load/store |
| 67 | // but we also need to account all the instructions which are going to |
| 68 | // stay in vectorized code due to uses outside of these scalar |
| 69 | // loads/stores. |
| 70 | ScalarCost = TTI.getPointersChainCost( |
| 71 | Ptrs, Base: BasePtr, Info: TTI::PointersChainInfo::getUnitStride(), AccessTy: ScalarTy, |
| 72 | CostKind); |
| 73 | |
| 74 | SmallVector<const Value *> PtrsRetainedInVecCode; |
| 75 | for (Value *V : Ptrs) { |
| 76 | if (V == BasePtr) { |
| 77 | PtrsRetainedInVecCode.push_back(Elt: V); |
| 78 | continue; |
| 79 | } |
| 80 | auto *Ptr = dyn_cast<GetElementPtrInst>(Val: V); |
| 81 | // For simplicity assume Ptr to stay in vectorized code if it's not a |
| 82 | // GEP instruction. We don't care since it's cost considered free. |
| 83 | // TODO: We should check for any uses outside of vectorizable tree |
| 84 | // rather than just single use. |
| 85 | if (!Ptr || !Ptr->hasOneUse()) |
| 86 | PtrsRetainedInVecCode.push_back(Elt: V); |
| 87 | } |
| 88 | |
| 89 | if (PtrsRetainedInVecCode.size() == Ptrs.size()) { |
| 90 | // If all pointers stay in vectorized code then we don't have |
| 91 | // any savings on that. |
| 92 | return std::make_pair(x: TTI::TCC_Free, y: TTI::TCC_Free); |
| 93 | } |
| 94 | VecCost = TTI.getPointersChainCost(Ptrs: PtrsRetainedInVecCode, Base: BasePtr, |
| 95 | Info: TTI::PointersChainInfo::getKnownStride(), |
| 96 | AccessTy: VecTy, CostKind); |
| 97 | } else { |
| 98 | // Case 1: Ptrs are the arguments of loads that we are going to transform |
| 99 | // into masked gather load intrinsic. |
| 100 | // All the scalar GEPs will be removed as a result of vectorization. |
| 101 | // For any external uses of some lanes extract element instructions will |
| 102 | // be generated (which cost is estimated separately). |
| 103 | TTI::PointersChainInfo PtrsInfo = |
| 104 | all_of(Range&: Ptrs, |
| 105 | P: [](const Value *V) { |
| 106 | auto *Ptr = dyn_cast<GetElementPtrInst>(Val: V); |
| 107 | return Ptr && !Ptr->hasAllConstantIndices(); |
| 108 | }) |
| 109 | ? TTI::PointersChainInfo::getUnknownStride() |
| 110 | : TTI::PointersChainInfo::getKnownStride(); |
| 111 | |
| 112 | ScalarCost = |
| 113 | TTI.getPointersChainCost(Ptrs, Base: BasePtr, Info: PtrsInfo, AccessTy: ScalarTy, CostKind); |
| 114 | auto *BaseGEP = dyn_cast<GEPOperator>(Val: BasePtr); |
| 115 | if (!BaseGEP) { |
| 116 | auto *It = find_if(Range&: Ptrs, P: IsaPred<GEPOperator>); |
| 117 | if (It != Ptrs.end()) |
| 118 | BaseGEP = cast<GEPOperator>(Val: *It); |
| 119 | } |
| 120 | if (BaseGEP) { |
| 121 | SmallVector<const Value *> Indices(BaseGEP->indices()); |
| 122 | VecCost = TTI.getGEPCost(PointeeType: BaseGEP->getSourceElementType(), |
| 123 | Ptr: BaseGEP->getPointerOperand(), Operands: Indices, CostKind, |
| 124 | AccessType: VecTy); |
| 125 | } |
| 126 | } |
| 127 | |
| 128 | return std::make_pair(x&: ScalarCost, y&: VecCost); |
| 129 | } |
| 130 | |
| 131 | InstructionCost getBlendedLoadCost(const TargetTransformInfo &TTI, Type *VecTy, |
| 132 | Align Alignment, unsigned AddressSpace, |
| 133 | const TTI::TargetCostKind CostKind) { |
| 134 | Type *CmpTy = CmpInst::makeCmpResultType(opnd_type: VecTy); |
| 135 | return 2 * TTI.getMemIntrinsicInstrCost( |
| 136 | MICA: MemIntrinsicCostAttributes(Intrinsic::masked_load, VecTy, |
| 137 | Alignment, AddressSpace), |
| 138 | CostKind) + |
| 139 | TTI.getArithmeticInstrCost(Opcode: Instruction::Xor, Ty: CmpTy, CostKind) + |
| 140 | TTI.getCmpSelInstrCost(Opcode: Instruction::Select, ValTy: VecTy, CondTy: CmpTy, |
| 141 | VecPred: CmpInst::BAD_ICMP_PREDICATE, CostKind); |
| 142 | } |
| 143 | |
| 144 | } // namespace llvm::slpvectorizer |
| 145 | |