1 //===- AMDGPUTargetTransformInfo.h - AMDGPU 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 /// AMDGPU 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_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H 18 #define LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H 19 20 #include "AMDGPU.h" 21 #include "AMDGPUSubtarget.h" 22 #include "AMDGPUTargetMachine.h" 23 #include "MCTargetDesc/AMDGPUMCTargetDesc.h" 24 #include "Utils/AMDGPUBaseInfo.h" 25 #include "llvm/ADT/ArrayRef.h" 26 #include "llvm/Analysis/TargetTransformInfo.h" 27 #include "llvm/CodeGen/BasicTTIImpl.h" 28 #include "llvm/IR/Function.h" 29 #include "llvm/MC/SubtargetFeature.h" 30 #include "llvm/Support/MathExtras.h" 31 #include <cassert> 32 33 namespace llvm { 34 35 class AMDGPUTargetLowering; 36 class Loop; 37 class ScalarEvolution; 38 class Type; 39 class Value; 40 41 class AMDGPUTTIImpl final : public BasicTTIImplBase<AMDGPUTTIImpl> { 42 using BaseT = BasicTTIImplBase<AMDGPUTTIImpl>; 43 using TTI = TargetTransformInfo; 44 45 friend BaseT; 46 47 Triple TargetTriple; 48 49 const GCNSubtarget *ST; 50 const TargetLoweringBase *TLI; 51 52 const TargetSubtargetInfo *getST() const { return ST; } 53 const TargetLoweringBase *getTLI() const { return TLI; } 54 55 public: 56 explicit AMDGPUTTIImpl(const AMDGPUTargetMachine *TM, const Function &F) 57 : BaseT(TM, F.getParent()->getDataLayout()), 58 TargetTriple(TM->getTargetTriple()), 59 ST(static_cast<const GCNSubtarget *>(TM->getSubtargetImpl(F))), 60 TLI(ST->getTargetLowering()) {} 61 62 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 63 TTI::UnrollingPreferences &UP); 64 }; 65 66 class GCNTTIImpl final : public BasicTTIImplBase<GCNTTIImpl> { 67 using BaseT = BasicTTIImplBase<GCNTTIImpl>; 68 using TTI = TargetTransformInfo; 69 70 friend BaseT; 71 72 const GCNSubtarget *ST; 73 const SITargetLowering *TLI; 74 AMDGPUTTIImpl CommonTTI; 75 bool IsGraphicsShader; 76 bool HasFP32Denormals; 77 78 const FeatureBitset InlineFeatureIgnoreList = { 79 // Codegen control options which don't matter. 80 AMDGPU::FeatureEnableLoadStoreOpt, 81 AMDGPU::FeatureEnableSIScheduler, 82 AMDGPU::FeatureEnableUnsafeDSOffsetFolding, 83 AMDGPU::FeatureFlatForGlobal, 84 AMDGPU::FeaturePromoteAlloca, 85 AMDGPU::FeatureUnalignedBufferAccess, 86 AMDGPU::FeatureUnalignedScratchAccess, 87 88 AMDGPU::FeatureAutoWaitcntBeforeBarrier, 89 90 // Property of the kernel/environment which can't actually differ. 91 AMDGPU::FeatureSGPRInitBug, 92 AMDGPU::FeatureXNACK, 93 AMDGPU::FeatureTrapHandler, 94 AMDGPU::FeatureCodeObjectV3, 95 96 // The default assumption needs to be ecc is enabled, but no directly 97 // exposed operations depend on it, so it can be safely inlined. 98 AMDGPU::FeatureSRAMECC, 99 100 // Perf-tuning features 101 AMDGPU::FeatureFastFMAF32, 102 AMDGPU::HalfRate64Ops 103 }; 104 105 const GCNSubtarget *getST() const { return ST; } 106 const AMDGPUTargetLowering *getTLI() const { return TLI; } 107 108 static inline int getFullRateInstrCost() { 109 return TargetTransformInfo::TCC_Basic; 110 } 111 112 static inline int getHalfRateInstrCost() { 113 return 2 * TargetTransformInfo::TCC_Basic; 114 } 115 116 // TODO: The size is usually 8 bytes, but takes 4x as many cycles. Maybe 117 // should be 2 or 4. 118 static inline int getQuarterRateInstrCost() { 119 return 3 * TargetTransformInfo::TCC_Basic; 120 } 121 122 // On some parts, normal fp64 operations are half rate, and others 123 // quarter. This also applies to some integer operations. 124 inline int get64BitInstrCost() const { 125 return ST->hasHalfRate64Ops() ? 126 getHalfRateInstrCost() : getQuarterRateInstrCost(); 127 } 128 129 public: 130 explicit GCNTTIImpl(const AMDGPUTargetMachine *TM, const Function &F) 131 : BaseT(TM, F.getParent()->getDataLayout()), 132 ST(static_cast<const GCNSubtarget*>(TM->getSubtargetImpl(F))), 133 TLI(ST->getTargetLowering()), 134 CommonTTI(TM, F), 135 IsGraphicsShader(AMDGPU::isShader(F.getCallingConv())), 136 HasFP32Denormals(ST->hasFP32Denormals(F)) { } 137 138 bool hasBranchDivergence() { return true; } 139 bool useGPUDivergenceAnalysis() const; 140 141 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 142 TTI::UnrollingPreferences &UP); 143 144 TTI::PopcntSupportKind getPopcntSupport(unsigned TyWidth) { 145 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 146 return TTI::PSK_FastHardware; 147 } 148 149 unsigned getHardwareNumberOfRegisters(bool Vector) const; 150 unsigned getNumberOfRegisters(bool Vector) const; 151 unsigned getRegisterBitWidth(bool Vector) const; 152 unsigned getMinVectorRegisterBitWidth() const; 153 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize, 154 unsigned ChainSizeInBytes, 155 VectorType *VecTy) const; 156 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize, 157 unsigned ChainSizeInBytes, 158 VectorType *VecTy) const; 159 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const; 160 161 bool isLegalToVectorizeMemChain(unsigned ChainSizeInBytes, 162 unsigned Alignment, 163 unsigned AddrSpace) const; 164 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, 165 unsigned Alignment, 166 unsigned AddrSpace) const; 167 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, 168 unsigned Alignment, 169 unsigned AddrSpace) const; 170 171 unsigned getMaxInterleaveFactor(unsigned VF); 172 173 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const; 174 175 int getArithmeticInstrCost( 176 unsigned Opcode, Type *Ty, 177 TTI::OperandValueKind Opd1Info = TTI::OK_AnyValue, 178 TTI::OperandValueKind Opd2Info = TTI::OK_AnyValue, 179 TTI::OperandValueProperties Opd1PropInfo = TTI::OP_None, 180 TTI::OperandValueProperties Opd2PropInfo = TTI::OP_None, 181 ArrayRef<const Value *> Args = ArrayRef<const Value *>(), 182 const Instruction *CxtI = nullptr); 183 184 unsigned getCFInstrCost(unsigned Opcode); 185 186 bool isInlineAsmSourceOfDivergence(const CallInst *CI, 187 ArrayRef<unsigned> Indices = {}) const; 188 189 int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index); 190 bool isSourceOfDivergence(const Value *V) const; 191 bool isAlwaysUniform(const Value *V) const; 192 193 unsigned getFlatAddressSpace() const { 194 // Don't bother running InferAddressSpaces pass on graphics shaders which 195 // don't use flat addressing. 196 if (IsGraphicsShader) 197 return -1; 198 return AMDGPUAS::FLAT_ADDRESS; 199 } 200 201 bool collectFlatAddressOperands(SmallVectorImpl<int> &OpIndexes, 202 Intrinsic::ID IID) const; 203 bool rewriteIntrinsicWithAddressSpace(IntrinsicInst *II, 204 Value *OldV, Value *NewV) const; 205 206 unsigned getVectorSplitCost() { return 0; } 207 208 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, 209 Type *SubTp); 210 211 bool areInlineCompatible(const Function *Caller, 212 const Function *Callee) const; 213 214 unsigned getInliningThresholdMultiplier() { return 11; } 215 216 int getInlinerVectorBonusPercent() { return 0; } 217 218 int getArithmeticReductionCost(unsigned Opcode, 219 Type *Ty, 220 bool IsPairwise); 221 template <typename T> 222 int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 223 ArrayRef<T *> Args, FastMathFlags FMF, 224 unsigned VF); 225 int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 226 ArrayRef<Type *> Tys, FastMathFlags FMF, 227 unsigned ScalarizationCostPassed = UINT_MAX); 228 int getIntrinsicInstrCost(Intrinsic::ID IID, Type *RetTy, 229 ArrayRef<Value *> Args, FastMathFlags FMF, 230 unsigned VF = 1); 231 int getMinMaxReductionCost(Type *Ty, Type *CondTy, 232 bool IsPairwiseForm, 233 bool IsUnsigned); 234 unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands); 235 }; 236 237 class R600TTIImpl final : public BasicTTIImplBase<R600TTIImpl> { 238 using BaseT = BasicTTIImplBase<R600TTIImpl>; 239 using TTI = TargetTransformInfo; 240 241 friend BaseT; 242 243 const R600Subtarget *ST; 244 const AMDGPUTargetLowering *TLI; 245 AMDGPUTTIImpl CommonTTI; 246 247 public: 248 explicit R600TTIImpl(const AMDGPUTargetMachine *TM, const Function &F) 249 : BaseT(TM, F.getParent()->getDataLayout()), 250 ST(static_cast<const R600Subtarget*>(TM->getSubtargetImpl(F))), 251 TLI(ST->getTargetLowering()), 252 CommonTTI(TM, F) {} 253 254 const R600Subtarget *getST() const { return ST; } 255 const AMDGPUTargetLowering *getTLI() const { return TLI; } 256 257 void getUnrollingPreferences(Loop *L, ScalarEvolution &SE, 258 TTI::UnrollingPreferences &UP); 259 unsigned getHardwareNumberOfRegisters(bool Vec) const; 260 unsigned getNumberOfRegisters(bool Vec) const; 261 unsigned getRegisterBitWidth(bool Vector) const; 262 unsigned getMinVectorRegisterBitWidth() const; 263 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const; 264 bool isLegalToVectorizeMemChain(unsigned ChainSizeInBytes, unsigned Alignment, 265 unsigned AddrSpace) const; 266 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, 267 unsigned Alignment, 268 unsigned AddrSpace) const; 269 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, 270 unsigned Alignment, 271 unsigned AddrSpace) const; 272 unsigned getMaxInterleaveFactor(unsigned VF); 273 unsigned getCFInstrCost(unsigned Opcode); 274 int getVectorInstrCost(unsigned Opcode, Type *ValTy, unsigned Index); 275 }; 276 277 } // end namespace llvm 278 279 #endif // LLVM_LIB_TARGET_AMDGPU_AMDGPUTARGETTRANSFORMINFO_H 280