X86TargetTransformInfo.cpp revision 83be7b0dd3ae9a3cb22d36ae4c1775972553b94b
1//===-- X86TargetTransformInfo.cpp - X86 specific TTI pass ----------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9/// \file 10/// This file implements a TargetTransformInfo analysis pass specific to the 11/// X86 target machine. It uses the target's detailed information to provide 12/// more precise answers to certain TTI queries, while letting the target 13/// independent and default TTI implementations handle the rest. 14/// 15//===----------------------------------------------------------------------===// 16 17#define DEBUG_TYPE "x86tti" 18#include "X86.h" 19#include "X86TargetMachine.h" 20#include "llvm/Analysis/TargetTransformInfo.h" 21#include "llvm/Support/Debug.h" 22#include "llvm/Target/TargetLowering.h" 23using namespace llvm; 24 25// Declare the pass initialization routine locally as target-specific passes 26// don't havve a target-wide initialization entry point, and so we rely on the 27// pass constructor initialization. 28namespace llvm { 29void initializeX86TTIPass(PassRegistry &); 30} 31 32namespace { 33 34class X86TTI : public ImmutablePass, public TargetTransformInfo { 35 const X86TargetMachine *TM; 36 const X86Subtarget *ST; 37 const X86TargetLowering *TLI; 38 39 /// Estimate the overhead of scalarizing an instruction. Insert and Extract 40 /// are set if the result needs to be inserted and/or extracted from vectors. 41 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const; 42 43public: 44 X86TTI() : ImmutablePass(ID), TM(0), ST(0), TLI(0) { 45 llvm_unreachable("This pass cannot be directly constructed"); 46 } 47 48 X86TTI(const X86TargetMachine *TM) 49 : ImmutablePass(ID), TM(TM), ST(TM->getSubtargetImpl()), 50 TLI(TM->getTargetLowering()) { 51 initializeX86TTIPass(*PassRegistry::getPassRegistry()); 52 } 53 54 virtual void initializePass() { 55 pushTTIStack(this); 56 } 57 58 virtual void finalizePass() { 59 popTTIStack(); 60 } 61 62 virtual void getAnalysisUsage(AnalysisUsage &AU) const { 63 TargetTransformInfo::getAnalysisUsage(AU); 64 } 65 66 /// Pass identification. 67 static char ID; 68 69 /// Provide necessary pointer adjustments for the two base classes. 70 virtual void *getAdjustedAnalysisPointer(const void *ID) { 71 if (ID == &TargetTransformInfo::ID) 72 return (TargetTransformInfo*)this; 73 return this; 74 } 75 76 /// \name Scalar TTI Implementations 77 /// @{ 78 virtual PopcntSupportKind getPopcntSupport(unsigned TyWidth) const; 79 80 /// @} 81 82 /// \name Vector TTI Implementations 83 /// @{ 84 85 virtual unsigned getNumberOfRegisters(bool Vector) const; 86 virtual unsigned getMaximumUnrollFactor() const; 87 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const; 88 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, 89 int Index, Type *SubTp) const; 90 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, 91 Type *Src) const; 92 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 93 Type *CondTy) const; 94 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val, 95 unsigned Index) const; 96 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src, 97 unsigned Alignment, 98 unsigned AddressSpace) const; 99 100 /// @} 101}; 102 103} // end anonymous namespace 104 105INITIALIZE_AG_PASS(X86TTI, TargetTransformInfo, "x86tti", 106 "X86 Target Transform Info", true, true, false) 107char X86TTI::ID = 0; 108 109ImmutablePass * 110llvm::createX86TargetTransformInfoPass(const X86TargetMachine *TM) { 111 return new X86TTI(TM); 112} 113 114 115//===----------------------------------------------------------------------===// 116// 117// X86 cost model. 118// 119//===----------------------------------------------------------------------===// 120 121namespace { 122struct X86CostTblEntry { 123 int ISD; 124 MVT Type; 125 unsigned Cost; 126}; 127} 128 129static int 130FindInTable(const X86CostTblEntry *Tbl, unsigned len, int ISD, MVT Ty) { 131 for (unsigned int i = 0; i < len; ++i) 132 if (Tbl[i].ISD == ISD && Tbl[i].Type == Ty) 133 return i; 134 135 // Could not find an entry. 136 return -1; 137} 138 139namespace { 140struct X86TypeConversionCostTblEntry { 141 int ISD; 142 MVT Dst; 143 MVT Src; 144 unsigned Cost; 145}; 146} 147 148static int 149FindInConvertTable(const X86TypeConversionCostTblEntry *Tbl, unsigned len, 150 int ISD, MVT Dst, MVT Src) { 151 for (unsigned int i = 0; i < len; ++i) 152 if (Tbl[i].ISD == ISD && Tbl[i].Src == Src && Tbl[i].Dst == Dst) 153 return i; 154 155 // Could not find an entry. 156 return -1; 157} 158 159X86TTI::PopcntSupportKind X86TTI::getPopcntSupport(unsigned TyWidth) const { 160 assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2"); 161 // TODO: Currently the __builtin_popcount() implementation using SSE3 162 // instructions is inefficient. Once the problem is fixed, we should 163 // call ST->hasSSE3() instead of ST->hasSSE4(). 164 return ST->hasSSE41() ? PSK_FastHardware : PSK_Software; 165} 166 167unsigned X86TTI::getNumberOfRegisters(bool Vector) const { 168 if (ST->is64Bit()) 169 return 16; 170 return 8; 171} 172 173unsigned X86TTI::getMaximumUnrollFactor() const { 174 if (ST->isAtom()) 175 return 1; 176 177 // Sandybridge and Haswell have multiple execution ports and pipelined 178 // vector units. 179 if (ST->hasAVX()) 180 return 4; 181 182 return 2; 183} 184 185unsigned X86TTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty) const { 186 // Legalize the type. 187 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty); 188 189 int ISD = TLI->InstructionOpcodeToISD(Opcode); 190 assert(ISD && "Invalid opcode"); 191 192 static const X86CostTblEntry AVX1CostTable[] = { 193 // We don't have to scalarize unsupported ops. We can issue two half-sized 194 // operations and we only need to extract the upper YMM half. 195 // Two ops + 1 extract + 1 insert = 4. 196 { ISD::MUL, MVT::v8i32, 4 }, 197 { ISD::SUB, MVT::v8i32, 4 }, 198 { ISD::ADD, MVT::v8i32, 4 }, 199 { ISD::MUL, MVT::v4i64, 4 }, 200 { ISD::SUB, MVT::v4i64, 4 }, 201 { ISD::ADD, MVT::v4i64, 4 }, 202 }; 203 204 // Look for AVX1 lowering tricks. 205 if (ST->hasAVX()) { 206 int Idx = FindInTable(AVX1CostTable, array_lengthof(AVX1CostTable), ISD, 207 LT.second); 208 if (Idx != -1) 209 return LT.first * AVX1CostTable[Idx].Cost; 210 } 211 // Fallback to the default implementation. 212 return TargetTransformInfo::getArithmeticInstrCost(Opcode, Ty); 213} 214 215unsigned X86TTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index, 216 Type *SubTp) const { 217 // We only estimate the cost of reverse shuffles. 218 if (Kind != SK_Reverse) 219 return TargetTransformInfo::getShuffleCost(Kind, Tp, Index, SubTp); 220 221 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Tp); 222 unsigned Cost = 1; 223 if (LT.second.getSizeInBits() > 128) 224 Cost = 3; // Extract + insert + copy. 225 226 // Multiple by the number of parts. 227 return Cost * LT.first; 228} 229 230unsigned X86TTI::getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const { 231 int ISD = TLI->InstructionOpcodeToISD(Opcode); 232 assert(ISD && "Invalid opcode"); 233 234 EVT SrcTy = TLI->getValueType(Src); 235 EVT DstTy = TLI->getValueType(Dst); 236 237 if (!SrcTy.isSimple() || !DstTy.isSimple()) 238 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src); 239 240 static const X86TypeConversionCostTblEntry AVXConversionTbl[] = { 241 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 242 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 1 }, 243 { ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 244 { ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 1 }, 245 { ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1 }, 246 { ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1 }, 247 { ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 }, 248 { ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 }, 249 { ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 1 }, 250 { ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 1 }, 251 { ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f32, 1 }, 252 { ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 1 }, 253 { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i1, 6 }, 254 { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i1, 9 }, 255 { ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 3 }, 256 }; 257 258 if (ST->hasAVX()) { 259 int Idx = FindInConvertTable(AVXConversionTbl, 260 array_lengthof(AVXConversionTbl), 261 ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()); 262 if (Idx != -1) 263 return AVXConversionTbl[Idx].Cost; 264 } 265 266 return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src); 267} 268 269unsigned X86TTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 270 Type *CondTy) const { 271 // Legalize the type. 272 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy); 273 274 MVT MTy = LT.second; 275 276 int ISD = TLI->InstructionOpcodeToISD(Opcode); 277 assert(ISD && "Invalid opcode"); 278 279 static const X86CostTblEntry SSE42CostTbl[] = { 280 { ISD::SETCC, MVT::v2f64, 1 }, 281 { ISD::SETCC, MVT::v4f32, 1 }, 282 { ISD::SETCC, MVT::v2i64, 1 }, 283 { ISD::SETCC, MVT::v4i32, 1 }, 284 { ISD::SETCC, MVT::v8i16, 1 }, 285 { ISD::SETCC, MVT::v16i8, 1 }, 286 }; 287 288 static const X86CostTblEntry AVX1CostTbl[] = { 289 { ISD::SETCC, MVT::v4f64, 1 }, 290 { ISD::SETCC, MVT::v8f32, 1 }, 291 // AVX1 does not support 8-wide integer compare. 292 { ISD::SETCC, MVT::v4i64, 4 }, 293 { ISD::SETCC, MVT::v8i32, 4 }, 294 { ISD::SETCC, MVT::v16i16, 4 }, 295 { ISD::SETCC, MVT::v32i8, 4 }, 296 }; 297 298 static const X86CostTblEntry AVX2CostTbl[] = { 299 { ISD::SETCC, MVT::v4i64, 1 }, 300 { ISD::SETCC, MVT::v8i32, 1 }, 301 { ISD::SETCC, MVT::v16i16, 1 }, 302 { ISD::SETCC, MVT::v32i8, 1 }, 303 }; 304 305 if (ST->hasAVX2()) { 306 int Idx = FindInTable(AVX2CostTbl, array_lengthof(AVX2CostTbl), ISD, MTy); 307 if (Idx != -1) 308 return LT.first * AVX2CostTbl[Idx].Cost; 309 } 310 311 if (ST->hasAVX()) { 312 int Idx = FindInTable(AVX1CostTbl, array_lengthof(AVX1CostTbl), ISD, MTy); 313 if (Idx != -1) 314 return LT.first * AVX1CostTbl[Idx].Cost; 315 } 316 317 if (ST->hasSSE42()) { 318 int Idx = FindInTable(SSE42CostTbl, array_lengthof(SSE42CostTbl), ISD, MTy); 319 if (Idx != -1) 320 return LT.first * SSE42CostTbl[Idx].Cost; 321 } 322 323 return TargetTransformInfo::getCmpSelInstrCost(Opcode, ValTy, CondTy); 324} 325 326unsigned X86TTI::getVectorInstrCost(unsigned Opcode, Type *Val, 327 unsigned Index) const { 328 assert(Val->isVectorTy() && "This must be a vector type"); 329 330 if (Index != -1U) { 331 // Legalize the type. 332 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Val); 333 334 // This type is legalized to a scalar type. 335 if (!LT.second.isVector()) 336 return 0; 337 338 // The type may be split. Normalize the index to the new type. 339 unsigned Width = LT.second.getVectorNumElements(); 340 Index = Index % Width; 341 342 // Floating point scalars are already located in index #0. 343 if (Val->getScalarType()->isFloatingPointTy() && Index == 0) 344 return 0; 345 } 346 347 return TargetTransformInfo::getVectorInstrCost(Opcode, Val, Index); 348} 349 350unsigned X86TTI::getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 351 unsigned AddressSpace) const { 352 // Legalize the type. 353 std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src); 354 assert((Opcode == Instruction::Load || Opcode == Instruction::Store) && 355 "Invalid Opcode"); 356 357 // Each load/store unit costs 1. 358 unsigned Cost = LT.first * 1; 359 360 // On Sandybridge 256bit load/stores are double pumped 361 // (but not on Haswell). 362 if (LT.second.getSizeInBits() > 128 && !ST->hasAVX2()) 363 Cost*=2; 364 365 return Cost; 366} 367