LLVM  10.0.0svn
ARMTargetTransformInfo.h
Go to the documentation of this file.
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"
26 #include "llvm/IR/Constant.h"
27 #include "llvm/IR/Function.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> {
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.
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 
175  const SCEV *Ptr);
176 
178  unsigned Opcode, Type *Ty,
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);
196  AssumptionCache &AC,
197  TargetLibraryInfo *LibInfo,
198  HardwareLoopInfo &HWLoopInfo);
199 
202 
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
uint64_t CallInst * C
int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, const Instruction *I=nullptr)
bool enableInterleavedAccessVectorization()
Type
MessagePack types as defined in the standard, with the exception of Integer being divided into a sign...
Definition: MsgPackReader.h:48
This class represents lattice values for constants.
Definition: AllocatorList.h:23
unsigned getMaxInterleaveFactor(unsigned VF)
bool hasOptSize() const
Optimize this function for size (-Os) or minimum size (-Oz).
Definition: Function.h:622
bool needsRelocation() const
This method classifies the entry according to whether or not it may generate a relocation entry...
Definition: Constants.cpp:513
int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, const Instruction *I=nullptr)
int getArithmeticInstrCost(unsigned Opcode, Type *Ty, TTI::OperandValueKind Op1Info=TTI::OK_AnyValue, TTI::OperandValueKind Op2Info=TTI::OK_AnyValue, TTI::OperandValueProperties Opd1PropInfo=TTI::OP_None, TTI::OperandValueProperties Opd2PropInfo=TTI::OP_None, ArrayRef< const Value *> Args=ArrayRef< const Value *>())
The main scalar evolution driver.
A cache of @llvm.assume calls within a function.
F(f)
bool isThumb1Only() const
Definition: ARMSubtarget.h:757
Base class which can be used to help build a TTI implementation.
Definition: BasicTTIImpl.h:77
unsigned getIntImmCost(const APInt &Imm, Type *Ty)
int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef< unsigned > Indices, unsigned Alignment, unsigned AddressSpace, bool UseMaskForCond=false, bool UseMaskForGaps=false)
unsigned getMaxInterleaveFactor() const
Definition: ARMSubtarget.h:823
bool isLoweredToCall(const Function *F)
int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, unsigned AddressSpace, const Instruction *I=nullptr)
BlockT * getHeader() const
Definition: LoopInfo.h:105
int getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, Type *SubTp)
bool useReductionIntrinsic(unsigned Opcode, Type *Ty, TTI::ReductionFlags Flags) const
unsigned getNumberOfRegisters(bool Vector)
ArrayRef - Represent a constant reference to an array (0 or more elements consecutively in memory)...
Definition: APInt.h:32
bool isTargetDarwin() const
Definition: ARMSubtarget.h:690
bool shouldExpandReduction(const IntrinsicInst *II) const
ARMTTIImpl(const ARMBaseTargetMachine *TM, const Function &F)
Container class for subtarget features.
Flags describing the kind of vector reduction.
The instances of the Type class are immutable: once they are created, they are never changed...
Definition: Type.h:45
bool isMClass() const
Definition: ARMSubtarget.h:760
This is an important base class in LLVM.
Definition: Constant.h:41
bool shouldBuildLookupTablesForConstant(Constant *C) const
This file provides a helper that implements much of the TTI interface in terms of the target-independ...
Attributes of a target dependent hardware loop.
OperandValueProperties
Additional properties of an operand&#39;s values.
bool shouldFavorBackedgeIndex(const Loop *L) const
bool isHardwareLoopProfitable(Loop *L, ScalarEvolution &SE, AssumptionCache &AC, TargetLibraryInfo *LibInfo, HardwareLoopInfo &HWLoopInfo)
This pass provides access to the codegen interfaces that are needed for IR-level transformations.
bool hasMVEFloatOps() const
Definition: ARMSubtarget.h:587
Provides information about what library functions are available for the current target.
AddressSpace
Definition: NVPTXBaseInfo.h:21
int getAddressComputationCost(Type *Val, ScalarEvolution *SE, const SCEV *Ptr)
bool hasNEON() const
Definition: ARMSubtarget.h:613
Class for arbitrary precision integers.
Definition: APInt.h:69
amdgpu Simplify well known AMD library false FunctionCallee Callee
bool isROPI() const
bool isFPVectorizationPotentiallyUnsafe()
Floating-point computation using ARMv8 AArch32 Advanced SIMD instructions remains unchanged from ARMv...
bool isThumb2() const
Definition: ARMSubtarget.h:758
unsigned getNumBlocks() const
Get the number of blocks in this loop in constant time.
Definition: LoopInfo.h:168
bool areInlineCompatible(const Function *Caller, const Function *Callee) const
int getMemcpyCost(const Instruction *I)
This class represents an analyzed expression in the program.
Represents a single loop in the control flow graph.
Definition: LoopInfo.h:509
Parameters that control the generic loop unrolling transformation.
const Function * getParent() const
Return the enclosing method, or null if none.
Definition: BasicBlock.h:106
int getIntImmCost(const APInt &Imm, Type *Ty)
#define I(x, y, z)
Definition: MD5.cpp:58
void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP)
static const Function * getParent(const Value *V)
int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm, Type *Ty)
bool hasMVEIntegerOps() const
Definition: ARMSubtarget.h:586
bool isRWPI() const
unsigned getRegisterBitWidth(bool Vector) const
int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index)
const DataLayout & getDataLayout() const
OperandValueKind
Additional information about an operand&#39;s possible values.
This pass exposes codegen information to IR-level passes.
constexpr char Args[]
Key for Kernel::Metadata::mArgs.
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:43
ShuffleKind
The various kinds of shuffle patterns for vector queries.