lib/Target/ARM/ARMTargetTransformInfo.h

   1 //===- ARMTargetTransformInfo.h - ARM 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 //
   9 /// \file
  10 /// This file a TargetTransformInfo::Concept conforming object specific to the
  11 /// ARM target machine. It uses the target's detailed information to
  12 /// provide more precise answers to certain TTI queries, while letting the
  13 /// target independent and default TTI implementations handle the rest.
  14 //
  15 //===----------------------------------------------------------------------===//
  16
  17 #ifndef LLVM_LIB_TARGET_ARM_ARMTARGETTRANSFORMINFO_H
  18 #define LLVM_LIB_TARGET_ARM_ARMTARGETTRANSFORMINFO_H
  19
  20 #include "ARM.h"
  21 #include "ARMSubtarget.h"
  22 #include "ARMTargetMachine.h"
  23 #include "llvm/ADT/ArrayRef.h"
  24 #include "llvm/Analysis/TargetTransformInfo.h"
  25 #include "llvm/CodeGen/BasicTTIImpl.h"
  26 #include "llvm/IR/Constant.h"
  27 #include "llvm/IR/Function.h"
  28 #include "llvm/MC/SubtargetFeature.h"
  29
  30 namespace llvm {
  31
  32 class APInt;
  33 class ARMTargetLowering;
  34 class Instruction;
  35 class Loop;
  36 class SCEV;
  37 class ScalarEvolution;
  38 class Type;
  39 class Value;
  40
  41 class ARMTTIImpl : public BasicTTIImplBase<ARMTTIImpl> {
  42   using BaseT = BasicTTIImplBase<ARMTTIImpl>;
  43   using TTI = TargetTransformInfo;
  44
  45   friend BaseT;
  46
  47   const ARMSubtarget *ST;
  48   const ARMTargetLowering *TLI;
  49
  50   // Currently the following features are excluded from InlineFeatureWhitelist.
  51   // ModeThumb, FeatureNoARM, ModeSoftFloat, FeatureFP64, FeatureD32
  52   // Depending on whether they are set or unset, different
  53   // instructions/registers are available. For example, inlining a callee with
  54   // -thumb-mode in a caller with +thumb-mode, may cause the assembler to
  55   // fail if the callee uses ARM only instructions, e.g. in inline asm.
  56   const FeatureBitset InlineFeatureWhitelist = {
  57       ARM::FeatureVFP2, ARM::FeatureVFP3, ARM::FeatureNEON, ARM::FeatureThumb2,
  58       ARM::FeatureFP16, ARM::FeatureVFP4, ARM::FeatureFPARMv8,
  59       ARM::FeatureFullFP16, ARM::FeatureFP16FML, ARM::FeatureHWDivThumb,
  60       ARM::FeatureHWDivARM, ARM::FeatureDB, ARM::FeatureV7Clrex,
  61       ARM::FeatureAcquireRelease, ARM::FeatureSlowFPBrcc,
  62       ARM::FeaturePerfMon, ARM::FeatureTrustZone, ARM::Feature8MSecExt,
  63       ARM::FeatureCrypto, ARM::FeatureCRC, ARM::FeatureRAS,
  64       ARM::FeatureFPAO, ARM::FeatureFuseAES, ARM::FeatureZCZeroing,
  65       ARM::FeatureProfUnpredicate, ARM::FeatureSlowVGETLNi32,
  66       ARM::FeatureSlowVDUP32, ARM::FeaturePreferVMOVSR,
  67       ARM::FeaturePrefISHSTBarrier, ARM::FeatureMuxedUnits,
  68       ARM::FeatureSlowOddRegister, ARM::FeatureSlowLoadDSubreg,
  69       ARM::FeatureDontWidenVMOVS, ARM::FeatureExpandMLx,
  70       ARM::FeatureHasVMLxHazards, ARM::FeatureNEONForFPMovs,
  71       ARM::FeatureNEONForFP, ARM::FeatureCheckVLDnAlign,
  72       ARM::FeatureHasSlowFPVMLx, ARM::FeatureVMLxForwarding,
  73       ARM::FeaturePref32BitThumb, ARM::FeatureAvoidPartialCPSR,
  74       ARM::FeatureCheapPredicableCPSR, ARM::FeatureAvoidMOVsShOp,
  75       ARM::FeatureHasRetAddrStack, ARM::FeatureHasNoBranchPredictor,
  76       ARM::FeatureDSP, ARM::FeatureMP, ARM::FeatureVirtualization,
  77       ARM::FeatureMClass, ARM::FeatureRClass, ARM::FeatureAClass,
  78       ARM::FeatureNaClTrap, ARM::FeatureStrictAlign, ARM::FeatureLongCalls,
  79       ARM::FeatureExecuteOnly, ARM::FeatureReserveR9, ARM::FeatureNoMovt,
  80       ARM::FeatureNoNegativeImmediates
  81   };
  82
  83   const ARMSubtarget *getST() const { return ST; }
  84   const ARMTargetLowering *getTLI() const { return TLI; }
  85
  86 public:
  87   explicit ARMTTIImpl(const ARMBaseTargetMachine *TM, const Function &F)
  88       : BaseT(TM, F.getParent()->getDataLayout()), ST(TM->getSubtargetImpl(F)),
  89         TLI(ST->getTargetLowering()) {}
  90
  91   bool areInlineCompatible(const Function *Caller,
  92                            const Function *Callee) const;
  93
  94   bool enableInterleavedAccessVectorization() { return true; }
  95
  96   bool shouldFavorBackedgeIndex(const Loop *L) const {
  97     if (L->getHeader()->getParent()->hasOptSize())
  98       return false;
  99     return ST->isMClass() && ST->isThumb2() && L->getNumBlocks() == 1;
 100   }
 101
 102   /// Floating-point computation using ARMv8 AArch32 Advanced
 103   /// SIMD instructions remains unchanged from ARMv7. Only AArch64 SIMD
 104   /// and Arm MVE are IEEE-754 compliant.
 105   bool isFPVectorizationPotentiallyUnsafe() {
 106     return !ST->isTargetDarwin() && !ST->hasMVEFloatOps();
 107   }
 108
 109   /// \name Scalar TTI Implementations
 110   /// @{
 111
 112   int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
 113                             Type *Ty);
 114
 115   using BaseT::getIntImmCost;
 116   int getIntImmCost(const APInt &Imm, Type *Ty);
 117
 118   int getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty);
 119
 120   /// @}
 121
 122   /// \name Vector TTI Implementations
 123   /// @{
 124
 125   unsigned getNumberOfRegisters(bool Vector) {
 126     if (Vector) {
 127       if (ST->hasNEON())
 128         return 16;
 129       if (ST->hasMVEIntegerOps())
 130         return 8;
 131       return 0;
 132     }
 133
 134     if (ST->isThumb1Only())
 135       return 8;
 136     return 13;
 137   }
 138
 139   unsigned getRegisterBitWidth(bool Vector) const {
 140     if (Vector) {
 141       if (ST->hasNEON())
 142         return 128;
 143       if (ST->hasMVEIntegerOps())
 144         return 128;
 145       return 0;
 146     }
 147
 148     return 32;
 149   }
 150
 151   unsigned getMaxInterleaveFactor(unsigned VF) {
 152     return ST->getMaxInterleaveFactor();
 153   }
 154
 155   int getMemcpyCost(const Instruction *I);
 156
 157   int getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, Type *SubTp);
 158
 159   bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
 160                              TTI::ReductionFlags Flags) const;
 161
 162   bool shouldExpandReduction(const IntrinsicInst *II) const {
 163     return false;
 164   }
 165
 166   int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
 167                        const Instruction *I = nullptr);
 168
 169   int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
 170                          const Instruction *I = nullptr);
 171
 172   int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index);
 173
 174   int getAddressComputationCost(Type *Val, ScalarEvolution *SE,
 175                                 const SCEV *Ptr);
 176
 177   int getArithmeticInstrCost(
 178       unsigned Opcode, Type *Ty,
 179       TTI::OperandValueKind Op1Info = TTI::OK_AnyValue,
 180       TTI::OperandValueKind Op2Info = TTI::OK_AnyValue,
 181       TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None,
 182       TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None,
 183       ArrayRef<const Value *> Args = ArrayRef<const Value *>());
 184
 185   int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
 186                       unsigned AddressSpace, const Instruction *I = nullptr);
 187
 188   int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
 189                                  ArrayRef<unsigned> Indices, unsigned Alignment,
 190                                  unsigned AddressSpace,
 191                                  bool UseMaskForCond = false,
 192                                  bool UseMaskForGaps = false);
 193
 194   bool isLoweredToCall(const Function *F);
 195   bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE,
 196                                 AssumptionCache &AC,
 197                                 TargetLibraryInfo *LibInfo,
 198                                 HardwareLoopInfo &HWLoopInfo);
 199
 200   void getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
 201                                TTI::UnrollingPreferences &UP);
 202
 203   bool shouldBuildLookupTablesForConstant(Constant *C) const {
 204     // In the ROPI and RWPI relocation models we can't have pointers to global
 205     // variables or functions in constant data, so don't convert switches to
 206     // lookup tables if any of the values would need relocation.
 207     if (ST->isROPI() || ST->isRWPI())
 208       return !C->needsRelocation();
 209
 210     return true;
 211   }
 212   /// @}
 213 };
 214
 215 } // end namespace llvm
 216
 217 #endif // LLVM_LIB_TARGET_ARM_ARMTARGETTRANSFORMINFO_H