1 //===-- X86TargetTransformInfo.h - X86 specific TTI -------------*- C++ -*-===// 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 /// \file 9 /// This file a TargetTransformInfo::Concept conforming object specific to the 10 /// X86 target machine. It uses the target's detailed information to 11 /// provide more precise answers to certain TTI queries, while letting the 12 /// target independent and default TTI implementations handle the rest. 13 /// 14 //===----------------------------------------------------------------------===// 15 16 #ifndef LLVM_LIB_TARGET_X86_X86TARGETTRANSFORMINFO_H 17 #define LLVM_LIB_TARGET_X86_X86TARGETTRANSFORMINFO_H 18 19 #include "X86.h" 20 #include "X86TargetMachine.h" 21 #include "llvm/Analysis/TargetTransformInfo.h" 22 #include "llvm/CodeGen/BasicTTIImpl.h" 23 #include "llvm/CodeGen/TargetLowering.h" 24 25 namespace llvm { 26 27 class X86TTIImpl : public BasicTTIImplBase<X86TTIImpl> { 28 typedef BasicTTIImplBase<X86TTIImpl> BaseT; 29 typedef TargetTransformInfo TTI; 30 friend BaseT; 31 32 const X86Subtarget *ST; 33 const X86TargetLowering *TLI; 34 35 const X86Subtarget *getST() const { return ST; } 36 const X86TargetLowering *getTLI() const { return TLI; } 37 38 const FeatureBitset InlineFeatureIgnoreList = { 39 // This indicates the CPU is 64 bit capable not that we are in 64-bit 40 // mode. 41 X86::Feature64Bit, 42 43 // These features don't have any intrinsics or ABI effect. 44 X86::FeatureNOPL, 45 X86::FeatureCMPXCHG16B, 46 X86::FeatureLAHFSAHF, 47 48 // Codegen control options. 49 X86::FeatureFast11ByteNOP, 50 X86::FeatureFast15ByteNOP, 51 X86::FeatureFastBEXTR, 52 X86::FeatureFastHorizontalOps, 53 X86::FeatureFastLZCNT, 54 X86::FeatureFastScalarFSQRT, 55 X86::FeatureFastSHLDRotate, 56 X86::FeatureFastScalarShiftMasks, 57 X86::FeatureFastVectorShiftMasks, 58 X86::FeatureFastVariableShuffle, 59 X86::FeatureFastVectorFSQRT, 60 X86::FeatureLEAForSP, 61 X86::FeatureLEAUsesAG, 62 X86::FeatureLZCNTFalseDeps, 63 X86::FeatureBranchFusion, 64 X86::FeatureMacroFusion, 65 X86::FeatureMergeToThreeWayBranch, 66 X86::FeaturePadShortFunctions, 67 X86::FeaturePOPCNTFalseDeps, 68 X86::FeatureSSEUnalignedMem, 69 X86::FeatureSlow3OpsLEA, 70 X86::FeatureSlowDivide32, 71 X86::FeatureSlowDivide64, 72 X86::FeatureSlowIncDec, 73 X86::FeatureSlowLEA, 74 X86::FeatureSlowPMADDWD, 75 X86::FeatureSlowPMULLD, 76 X86::FeatureSlowSHLD, 77 X86::FeatureSlowTwoMemOps, 78 X86::FeatureSlowUAMem16, 79 X86::FeaturePreferMaskRegisters, 80 X86::FeatureInsertVZEROUPPER, 81 X86::FeatureUseGLMDivSqrtCosts, 82 83 // Perf-tuning flags. 84 X86::FeatureHasFastGather, 85 X86::FeatureSlowUAMem32, 86 87 // Based on whether user set the -mprefer-vector-width command line. 88 X86::FeaturePrefer128Bit, 89 X86::FeaturePrefer256Bit, 90 91 // CPU name enums. These just follow CPU string. 92 X86::ProcIntelAtom, 93 X86::ProcIntelSLM, 94 }; 95 96 public: 97 explicit X86TTIImpl(const X86TargetMachine *TM, const Function &F) 98 : BaseT(TM, F.getParent()->getDataLayout()), ST(TM->getSubtargetImpl(F)), 99 TLI(ST->getTargetLowering()) {} 100 101 /// \name Scalar TTI Implementations 102 /// @{ 103 TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth); 104 105 /// @} 106 107 /// \name Cache TTI Implementation 108 /// @{ 109 llvm::Optional<unsigned> getCacheSize( 110 TargetTransformInfo::CacheLevel Level) const; 111 llvm::Optional<unsigned> getCacheAssociativity( 112 TargetTransformInfo::CacheLevel Level) const; 113 /// @} 114 115 /// \name Vector TTI Implementations 116 /// @{ 117 118 unsigned getNumberOfRegisters(unsigned ClassID) const; 119 unsigned getRegisterBitWidth(bool Vector) const; 120 unsigned getLoadStoreVecRegBitWidth(unsigned AS) const; 121 unsigned getMaxInterleaveFactor(unsigned VF); 122 int getArithmeticInstrCost( 123 unsigned Opcode, Type *Ty, 124 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue, 125 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue, 126 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None, 127 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None, 128 ArrayRef<const Value *> Args = ArrayRef<const Value *>(), 129 const Instruction *CxtI = nullptr); 130 int getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, Type *SubTp); 131 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, 132 const Instruction *I = nullptr); 133 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, 134 const Instruction *I = nullptr); 135 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index); 136 int getMemoryOpCost(unsigned Opcode, Type *Src, MaybeAlign Alignment, 137 unsigned AddressSpace, const Instruction *I = nullptr); 138 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 139 unsigned AddressSpace); 140 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr, 141 bool VariableMask, unsigned Alignment); 142 int getAddressComputationCost(Type *PtrTy, ScalarEvolution *SE, 143 const SCEV *Ptr); 144 145 unsigned getAtomicMemIntrinsicMaxElementSize() const; 146 147 int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 148 ArrayRef<Type *> Tys, FastMathFlags FMF, 149 unsigned ScalarizationCostPassed = UINT_MAX); 150 int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 151 ArrayRef<Value *> Args, FastMathFlags FMF, 152 unsigned VF = 1); 153 154 int getArithmeticReductionCost(unsigned Opcode, Type *Ty, 155 bool IsPairwiseForm); 156 157 int getMinMaxReductionCost(Type *Ty, Type *CondTy, bool IsPairwiseForm, 158 bool IsUnsigned); 159 160 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, 161 unsigned Factor, ArrayRef<unsigned> Indices, 162 unsigned Alignment, unsigned AddressSpace, 163 bool UseMaskForCond = false, 164 bool UseMaskForGaps = false); 165 int getInterleavedMemoryOpCostAVX512(unsigned Opcode, Type *VecTy, 166 unsigned Factor, ArrayRef<unsigned> Indices, 167 unsigned Alignment, unsigned AddressSpace, 168 bool UseMaskForCond = false, 169 bool UseMaskForGaps = false); 170 int getInterleavedMemoryOpCostAVX2(unsigned Opcode, Type *VecTy, 171 unsigned Factor, ArrayRef<unsigned> Indices, 172 unsigned Alignment, unsigned AddressSpace, 173 bool UseMaskForCond = false, 174 bool UseMaskForGaps = false); 175 176 int getIntImmCost(int64_t); 177 178 int getIntImmCost(const APInt &Imm, Type *Ty); 179 180 unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands); 181 182 int getIntImmCostInst(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty); 183 int getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, 184 Type *Ty); 185 bool isLSRCostLess(TargetTransformInfo::LSRCost &C1, 186 TargetTransformInfo::LSRCost &C2); 187 bool canMacroFuseCmp(); 188 bool isLegalMaskedLoad(Type *DataType, MaybeAlign Alignment); 189 bool isLegalMaskedStore(Type *DataType, MaybeAlign Alignment); 190 bool isLegalNTLoad(Type *DataType, Align Alignment); 191 bool isLegalNTStore(Type *DataType, Align Alignment); 192 bool isLegalMaskedGather(Type *DataType, MaybeAlign Alignment); 193 bool isLegalMaskedScatter(Type *DataType, MaybeAlign Alignment); 194 bool isLegalMaskedExpandLoad(Type *DataType); 195 bool isLegalMaskedCompressStore(Type *DataType); 196 bool hasDivRemOp(Type *DataType, bool IsSigned); 197 bool isFCmpOrdCheaperThanFCmpZero(Type *Ty); 198 bool areInlineCompatible(const Function *Caller, 199 const Function *Callee) const; 200 bool areFunctionArgsABICompatible(const Function *Caller, 201 const Function *Callee, 202 SmallPtrSetImpl<Argument *> &Args) const; 203 TTI::MemCmpExpansionOptions enableMemCmpExpansion(bool OptSize, 204 bool IsZeroCmp) const; 205 bool enableInterleavedAccessVectorization(); 206 private: 207 int getGSScalarCost(unsigned Opcode, Type *DataTy, bool VariableMask, 208 unsigned Alignment, unsigned AddressSpace); 209 int getGSVectorCost(unsigned Opcode, Type *DataTy, Value *Ptr, 210 unsigned Alignment, unsigned AddressSpace); 211 212 /// @} 213 }; 214 215 } // end namespace llvm 216 217 #endif 218