Andrew Scull | 5e1ddfa | 2018-08-14 10:06:54 +0100 | [diff] [blame^] | 1 | //===- TargetTransformInfoImpl.h --------------------------------*- C++ -*-===// |
| 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 provides helpers for the implementation of |
| 11 | /// a TargetTransformInfo-conforming class. |
| 12 | /// |
| 13 | //===----------------------------------------------------------------------===// |
| 14 | |
| 15 | #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H |
| 16 | #define LLVM_ANALYSIS_TARGETTRANSFORMINFOIMPL_H |
| 17 | |
| 18 | #include "llvm/Analysis/ScalarEvolutionExpressions.h" |
| 19 | #include "llvm/Analysis/TargetTransformInfo.h" |
| 20 | #include "llvm/Analysis/VectorUtils.h" |
| 21 | #include "llvm/IR/CallSite.h" |
| 22 | #include "llvm/IR/DataLayout.h" |
| 23 | #include "llvm/IR/Function.h" |
| 24 | #include "llvm/IR/GetElementPtrTypeIterator.h" |
| 25 | #include "llvm/IR/Operator.h" |
| 26 | #include "llvm/IR/Type.h" |
| 27 | |
| 28 | namespace llvm { |
| 29 | |
| 30 | /// \brief Base class for use as a mix-in that aids implementing |
| 31 | /// a TargetTransformInfo-compatible class. |
| 32 | class TargetTransformInfoImplBase { |
| 33 | protected: |
| 34 | typedef TargetTransformInfo TTI; |
| 35 | |
| 36 | const DataLayout &DL; |
| 37 | |
| 38 | explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {} |
| 39 | |
| 40 | public: |
| 41 | // Provide value semantics. MSVC requires that we spell all of these out. |
| 42 | TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg) |
| 43 | : DL(Arg.DL) {} |
| 44 | TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {} |
| 45 | |
| 46 | const DataLayout &getDataLayout() const { return DL; } |
| 47 | |
| 48 | unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) { |
| 49 | switch (Opcode) { |
| 50 | default: |
| 51 | // By default, just classify everything as 'basic'. |
| 52 | return TTI::TCC_Basic; |
| 53 | |
| 54 | case Instruction::GetElementPtr: |
| 55 | llvm_unreachable("Use getGEPCost for GEP operations!"); |
| 56 | |
| 57 | case Instruction::BitCast: |
| 58 | assert(OpTy && "Cast instructions must provide the operand type"); |
| 59 | if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy())) |
| 60 | // Identity and pointer-to-pointer casts are free. |
| 61 | return TTI::TCC_Free; |
| 62 | |
| 63 | // Otherwise, the default basic cost is used. |
| 64 | return TTI::TCC_Basic; |
| 65 | |
| 66 | case Instruction::FDiv: |
| 67 | case Instruction::FRem: |
| 68 | case Instruction::SDiv: |
| 69 | case Instruction::SRem: |
| 70 | case Instruction::UDiv: |
| 71 | case Instruction::URem: |
| 72 | return TTI::TCC_Expensive; |
| 73 | |
| 74 | case Instruction::IntToPtr: { |
| 75 | // An inttoptr cast is free so long as the input is a legal integer type |
| 76 | // which doesn't contain values outside the range of a pointer. |
| 77 | unsigned OpSize = OpTy->getScalarSizeInBits(); |
| 78 | if (DL.isLegalInteger(OpSize) && |
| 79 | OpSize <= DL.getPointerTypeSizeInBits(Ty)) |
| 80 | return TTI::TCC_Free; |
| 81 | |
| 82 | // Otherwise it's not a no-op. |
| 83 | return TTI::TCC_Basic; |
| 84 | } |
| 85 | case Instruction::PtrToInt: { |
| 86 | // A ptrtoint cast is free so long as the result is large enough to store |
| 87 | // the pointer, and a legal integer type. |
| 88 | unsigned DestSize = Ty->getScalarSizeInBits(); |
| 89 | if (DL.isLegalInteger(DestSize) && |
| 90 | DestSize >= DL.getPointerTypeSizeInBits(OpTy)) |
| 91 | return TTI::TCC_Free; |
| 92 | |
| 93 | // Otherwise it's not a no-op. |
| 94 | return TTI::TCC_Basic; |
| 95 | } |
| 96 | case Instruction::Trunc: |
| 97 | // trunc to a native type is free (assuming the target has compare and |
| 98 | // shift-right of the same width). |
| 99 | if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty))) |
| 100 | return TTI::TCC_Free; |
| 101 | |
| 102 | return TTI::TCC_Basic; |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | int getGEPCost(Type *PointeeType, const Value *Ptr, |
| 107 | ArrayRef<const Value *> Operands) { |
| 108 | // In the basic model, we just assume that all-constant GEPs will be folded |
| 109 | // into their uses via addressing modes. |
| 110 | for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx) |
| 111 | if (!isa<Constant>(Operands[Idx])) |
| 112 | return TTI::TCC_Basic; |
| 113 | |
| 114 | return TTI::TCC_Free; |
| 115 | } |
| 116 | |
| 117 | unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI, |
| 118 | unsigned &JTSize) { |
| 119 | JTSize = 0; |
| 120 | return SI.getNumCases(); |
| 121 | } |
| 122 | |
| 123 | int getExtCost(const Instruction *I, const Value *Src) { |
| 124 | return TTI::TCC_Basic; |
| 125 | } |
| 126 | |
| 127 | unsigned getCallCost(FunctionType *FTy, int NumArgs) { |
| 128 | assert(FTy && "FunctionType must be provided to this routine."); |
| 129 | |
| 130 | // The target-independent implementation just measures the size of the |
| 131 | // function by approximating that each argument will take on average one |
| 132 | // instruction to prepare. |
| 133 | |
| 134 | if (NumArgs < 0) |
| 135 | // Set the argument number to the number of explicit arguments in the |
| 136 | // function. |
| 137 | NumArgs = FTy->getNumParams(); |
| 138 | |
| 139 | return TTI::TCC_Basic * (NumArgs + 1); |
| 140 | } |
| 141 | |
| 142 | unsigned getInliningThresholdMultiplier() { return 1; } |
| 143 | |
| 144 | unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, |
| 145 | ArrayRef<Type *> ParamTys) { |
| 146 | switch (IID) { |
| 147 | default: |
| 148 | // Intrinsics rarely (if ever) have normal argument setup constraints. |
| 149 | // Model them as having a basic instruction cost. |
| 150 | // FIXME: This is wrong for libc intrinsics. |
| 151 | return TTI::TCC_Basic; |
| 152 | |
| 153 | case Intrinsic::annotation: |
| 154 | case Intrinsic::assume: |
| 155 | case Intrinsic::sideeffect: |
| 156 | case Intrinsic::dbg_declare: |
| 157 | case Intrinsic::dbg_value: |
| 158 | case Intrinsic::invariant_start: |
| 159 | case Intrinsic::invariant_end: |
| 160 | case Intrinsic::lifetime_start: |
| 161 | case Intrinsic::lifetime_end: |
| 162 | case Intrinsic::objectsize: |
| 163 | case Intrinsic::ptr_annotation: |
| 164 | case Intrinsic::var_annotation: |
| 165 | case Intrinsic::experimental_gc_result: |
| 166 | case Intrinsic::experimental_gc_relocate: |
| 167 | case Intrinsic::coro_alloc: |
| 168 | case Intrinsic::coro_begin: |
| 169 | case Intrinsic::coro_free: |
| 170 | case Intrinsic::coro_end: |
| 171 | case Intrinsic::coro_frame: |
| 172 | case Intrinsic::coro_size: |
| 173 | case Intrinsic::coro_suspend: |
| 174 | case Intrinsic::coro_param: |
| 175 | case Intrinsic::coro_subfn_addr: |
| 176 | // These intrinsics don't actually represent code after lowering. |
| 177 | return TTI::TCC_Free; |
| 178 | } |
| 179 | } |
| 180 | |
| 181 | bool hasBranchDivergence() { return false; } |
| 182 | |
| 183 | bool isSourceOfDivergence(const Value *V) { return false; } |
| 184 | |
| 185 | bool isAlwaysUniform(const Value *V) { return false; } |
| 186 | |
| 187 | unsigned getFlatAddressSpace () { |
| 188 | return -1; |
| 189 | } |
| 190 | |
| 191 | bool isLoweredToCall(const Function *F) { |
| 192 | assert(F && "A concrete function must be provided to this routine."); |
| 193 | |
| 194 | // FIXME: These should almost certainly not be handled here, and instead |
| 195 | // handled with the help of TLI or the target itself. This was largely |
| 196 | // ported from existing analysis heuristics here so that such refactorings |
| 197 | // can take place in the future. |
| 198 | |
| 199 | if (F->isIntrinsic()) |
| 200 | return false; |
| 201 | |
| 202 | if (F->hasLocalLinkage() || !F->hasName()) |
| 203 | return true; |
| 204 | |
| 205 | StringRef Name = F->getName(); |
| 206 | |
| 207 | // These will all likely lower to a single selection DAG node. |
| 208 | if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" || |
| 209 | Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" || |
| 210 | Name == "fmin" || Name == "fminf" || Name == "fminl" || |
| 211 | Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" || |
| 212 | Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" || |
| 213 | Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") |
| 214 | return false; |
| 215 | |
| 216 | // These are all likely to be optimized into something smaller. |
| 217 | if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" || |
| 218 | Name == "exp2l" || Name == "exp2f" || Name == "floor" || |
| 219 | Name == "floorf" || Name == "ceil" || Name == "round" || |
| 220 | Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" || |
| 221 | Name == "llabs") |
| 222 | return false; |
| 223 | |
| 224 | return true; |
| 225 | } |
| 226 | |
| 227 | void getUnrollingPreferences(Loop *, ScalarEvolution &, |
| 228 | TTI::UnrollingPreferences &) {} |
| 229 | |
| 230 | bool isLegalAddImmediate(int64_t Imm) { return false; } |
| 231 | |
| 232 | bool isLegalICmpImmediate(int64_t Imm) { return false; } |
| 233 | |
| 234 | bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, |
| 235 | bool HasBaseReg, int64_t Scale, |
| 236 | unsigned AddrSpace, Instruction *I = nullptr) { |
| 237 | // Guess that only reg and reg+reg addressing is allowed. This heuristic is |
| 238 | // taken from the implementation of LSR. |
| 239 | return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1); |
| 240 | } |
| 241 | |
| 242 | bool isLSRCostLess(TTI::LSRCost &C1, TTI::LSRCost &C2) { |
| 243 | return std::tie(C1.NumRegs, C1.AddRecCost, C1.NumIVMuls, C1.NumBaseAdds, |
| 244 | C1.ScaleCost, C1.ImmCost, C1.SetupCost) < |
| 245 | std::tie(C2.NumRegs, C2.AddRecCost, C2.NumIVMuls, C2.NumBaseAdds, |
| 246 | C2.ScaleCost, C2.ImmCost, C2.SetupCost); |
| 247 | } |
| 248 | |
| 249 | bool canMacroFuseCmp() { return false; } |
| 250 | |
| 251 | bool shouldFavorPostInc() const { return false; } |
| 252 | |
| 253 | bool isLegalMaskedStore(Type *DataType) { return false; } |
| 254 | |
| 255 | bool isLegalMaskedLoad(Type *DataType) { return false; } |
| 256 | |
| 257 | bool isLegalMaskedScatter(Type *DataType) { return false; } |
| 258 | |
| 259 | bool isLegalMaskedGather(Type *DataType) { return false; } |
| 260 | |
| 261 | bool hasDivRemOp(Type *DataType, bool IsSigned) { return false; } |
| 262 | |
| 263 | bool hasVolatileVariant(Instruction *I, unsigned AddrSpace) { return false; } |
| 264 | |
| 265 | bool prefersVectorizedAddressing() { return true; } |
| 266 | |
| 267 | int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, |
| 268 | bool HasBaseReg, int64_t Scale, unsigned AddrSpace) { |
| 269 | // Guess that all legal addressing mode are free. |
| 270 | if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, |
| 271 | Scale, AddrSpace)) |
| 272 | return 0; |
| 273 | return -1; |
| 274 | } |
| 275 | |
| 276 | bool LSRWithInstrQueries() { return false; } |
| 277 | |
| 278 | bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; } |
| 279 | |
| 280 | bool isProfitableToHoist(Instruction *I) { return true; } |
| 281 | |
| 282 | bool useAA() { return false; } |
| 283 | |
| 284 | bool isTypeLegal(Type *Ty) { return false; } |
| 285 | |
| 286 | unsigned getJumpBufAlignment() { return 0; } |
| 287 | |
| 288 | unsigned getJumpBufSize() { return 0; } |
| 289 | |
| 290 | bool shouldBuildLookupTables() { return true; } |
| 291 | bool shouldBuildLookupTablesForConstant(Constant *C) { return true; } |
| 292 | |
| 293 | bool useColdCCForColdCall(Function &F) { return false; } |
| 294 | |
| 295 | unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) { |
| 296 | return 0; |
| 297 | } |
| 298 | |
| 299 | unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args, |
| 300 | unsigned VF) { return 0; } |
| 301 | |
| 302 | bool supportsEfficientVectorElementLoadStore() { return false; } |
| 303 | |
| 304 | bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; } |
| 305 | |
| 306 | const TTI::MemCmpExpansionOptions *enableMemCmpExpansion( |
| 307 | bool IsZeroCmp) const { |
| 308 | return nullptr; |
| 309 | } |
| 310 | |
| 311 | bool enableInterleavedAccessVectorization() { return false; } |
| 312 | |
| 313 | bool isFPVectorizationPotentiallyUnsafe() { return false; } |
| 314 | |
| 315 | bool allowsMisalignedMemoryAccesses(LLVMContext &Context, |
| 316 | unsigned BitWidth, |
| 317 | unsigned AddressSpace, |
| 318 | unsigned Alignment, |
| 319 | bool *Fast) { return false; } |
| 320 | |
| 321 | TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) { |
| 322 | return TTI::PSK_Software; |
| 323 | } |
| 324 | |
| 325 | bool haveFastSqrt(Type *Ty) { return false; } |
| 326 | |
| 327 | bool isFCmpOrdCheaperThanFCmpZero(Type *Ty) { return true; } |
| 328 | |
| 329 | unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; } |
| 330 | |
| 331 | int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm, |
| 332 | Type *Ty) { |
| 333 | return 0; |
| 334 | } |
| 335 | |
| 336 | unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; } |
| 337 | |
| 338 | unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, |
| 339 | Type *Ty) { |
| 340 | return TTI::TCC_Free; |
| 341 | } |
| 342 | |
| 343 | unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, |
| 344 | Type *Ty) { |
| 345 | return TTI::TCC_Free; |
| 346 | } |
| 347 | |
| 348 | unsigned getNumberOfRegisters(bool Vector) { return 8; } |
| 349 | |
| 350 | unsigned getRegisterBitWidth(bool Vector) const { return 32; } |
| 351 | |
| 352 | unsigned getMinVectorRegisterBitWidth() { return 128; } |
| 353 | |
| 354 | bool shouldMaximizeVectorBandwidth(bool OptSize) const { return false; } |
| 355 | |
| 356 | bool |
| 357 | shouldConsiderAddressTypePromotion(const Instruction &I, |
| 358 | bool &AllowPromotionWithoutCommonHeader) { |
| 359 | AllowPromotionWithoutCommonHeader = false; |
| 360 | return false; |
| 361 | } |
| 362 | |
| 363 | unsigned getCacheLineSize() { return 0; } |
| 364 | |
| 365 | llvm::Optional<unsigned> getCacheSize(TargetTransformInfo::CacheLevel Level) { |
| 366 | switch (Level) { |
| 367 | case TargetTransformInfo::CacheLevel::L1D: |
| 368 | LLVM_FALLTHROUGH; |
| 369 | case TargetTransformInfo::CacheLevel::L2D: |
| 370 | return llvm::Optional<unsigned>(); |
| 371 | } |
| 372 | |
| 373 | llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); |
| 374 | } |
| 375 | |
| 376 | llvm::Optional<unsigned> getCacheAssociativity( |
| 377 | TargetTransformInfo::CacheLevel Level) { |
| 378 | switch (Level) { |
| 379 | case TargetTransformInfo::CacheLevel::L1D: |
| 380 | LLVM_FALLTHROUGH; |
| 381 | case TargetTransformInfo::CacheLevel::L2D: |
| 382 | return llvm::Optional<unsigned>(); |
| 383 | } |
| 384 | |
| 385 | llvm_unreachable("Unknown TargetTransformInfo::CacheLevel"); |
| 386 | } |
| 387 | |
| 388 | unsigned getPrefetchDistance() { return 0; } |
| 389 | |
| 390 | unsigned getMinPrefetchStride() { return 1; } |
| 391 | |
| 392 | unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX; } |
| 393 | |
| 394 | unsigned getMaxInterleaveFactor(unsigned VF) { return 1; } |
| 395 | |
| 396 | unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, |
| 397 | TTI::OperandValueKind Opd1Info, |
| 398 | TTI::OperandValueKind Opd2Info, |
| 399 | TTI::OperandValueProperties Opd1PropInfo, |
| 400 | TTI::OperandValueProperties Opd2PropInfo, |
| 401 | ArrayRef<const Value *> Args) { |
| 402 | return 1; |
| 403 | } |
| 404 | |
| 405 | unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index, |
| 406 | Type *SubTp) { |
| 407 | return 1; |
| 408 | } |
| 409 | |
| 410 | unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src, |
| 411 | const Instruction *I) { return 1; } |
| 412 | |
| 413 | unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst, |
| 414 | VectorType *VecTy, unsigned Index) { |
| 415 | return 1; |
| 416 | } |
| 417 | |
| 418 | unsigned getCFInstrCost(unsigned Opcode) { return 1; } |
| 419 | |
| 420 | unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy, |
| 421 | const Instruction *I) { |
| 422 | return 1; |
| 423 | } |
| 424 | |
| 425 | unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { |
| 426 | return 1; |
| 427 | } |
| 428 | |
| 429 | unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, |
| 430 | unsigned AddressSpace, const Instruction *I) { |
| 431 | return 1; |
| 432 | } |
| 433 | |
| 434 | unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, |
| 435 | unsigned AddressSpace) { |
| 436 | return 1; |
| 437 | } |
| 438 | |
| 439 | unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr, |
| 440 | bool VariableMask, |
| 441 | unsigned Alignment) { |
| 442 | return 1; |
| 443 | } |
| 444 | |
| 445 | unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, |
| 446 | unsigned Factor, |
| 447 | ArrayRef<unsigned> Indices, |
| 448 | unsigned Alignment, |
| 449 | unsigned AddressSpace) { |
| 450 | return 1; |
| 451 | } |
| 452 | |
| 453 | unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, |
| 454 | ArrayRef<Type *> Tys, FastMathFlags FMF, |
| 455 | unsigned ScalarizationCostPassed) { |
| 456 | return 1; |
| 457 | } |
| 458 | unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, |
| 459 | ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) { |
| 460 | return 1; |
| 461 | } |
| 462 | |
| 463 | unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) { |
| 464 | return 1; |
| 465 | } |
| 466 | |
| 467 | unsigned getNumberOfParts(Type *Tp) { return 0; } |
| 468 | |
| 469 | unsigned getAddressComputationCost(Type *Tp, ScalarEvolution *, |
| 470 | const SCEV *) { |
| 471 | return 0; |
| 472 | } |
| 473 | |
| 474 | unsigned getArithmeticReductionCost(unsigned, Type *, bool) { return 1; } |
| 475 | |
| 476 | unsigned getMinMaxReductionCost(Type *, Type *, bool, bool) { return 1; } |
| 477 | |
| 478 | unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; } |
| 479 | |
| 480 | bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) { |
| 481 | return false; |
| 482 | } |
| 483 | |
| 484 | unsigned getAtomicMemIntrinsicMaxElementSize() const { |
| 485 | // Note for overrides: You must ensure for all element unordered-atomic |
| 486 | // memory intrinsics that all power-of-2 element sizes up to, and |
| 487 | // including, the return value of this method have a corresponding |
| 488 | // runtime lib call. These runtime lib call definitions can be found |
| 489 | // in RuntimeLibcalls.h |
| 490 | return 0; |
| 491 | } |
| 492 | |
| 493 | Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst, |
| 494 | Type *ExpectedType) { |
| 495 | return nullptr; |
| 496 | } |
| 497 | |
| 498 | Type *getMemcpyLoopLoweringType(LLVMContext &Context, Value *Length, |
| 499 | unsigned SrcAlign, unsigned DestAlign) const { |
| 500 | return Type::getInt8Ty(Context); |
| 501 | } |
| 502 | |
| 503 | void getMemcpyLoopResidualLoweringType(SmallVectorImpl<Type *> &OpsOut, |
| 504 | LLVMContext &Context, |
| 505 | unsigned RemainingBytes, |
| 506 | unsigned SrcAlign, |
| 507 | unsigned DestAlign) const { |
| 508 | for (unsigned i = 0; i != RemainingBytes; ++i) |
| 509 | OpsOut.push_back(Type::getInt8Ty(Context)); |
| 510 | } |
| 511 | |
| 512 | bool areInlineCompatible(const Function *Caller, |
| 513 | const Function *Callee) const { |
| 514 | return (Caller->getFnAttribute("target-cpu") == |
| 515 | Callee->getFnAttribute("target-cpu")) && |
| 516 | (Caller->getFnAttribute("target-features") == |
| 517 | Callee->getFnAttribute("target-features")); |
| 518 | } |
| 519 | |
| 520 | bool isIndexedLoadLegal(TTI::MemIndexedMode Mode, Type *Ty, |
| 521 | const DataLayout &DL) const { |
| 522 | return false; |
| 523 | } |
| 524 | |
| 525 | bool isIndexedStoreLegal(TTI::MemIndexedMode Mode, Type *Ty, |
| 526 | const DataLayout &DL) const { |
| 527 | return false; |
| 528 | } |
| 529 | |
| 530 | unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const { return 128; } |
| 531 | |
| 532 | bool isLegalToVectorizeLoad(LoadInst *LI) const { return true; } |
| 533 | |
| 534 | bool isLegalToVectorizeStore(StoreInst *SI) const { return true; } |
| 535 | |
| 536 | bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes, |
| 537 | unsigned Alignment, |
| 538 | unsigned AddrSpace) const { |
| 539 | return true; |
| 540 | } |
| 541 | |
| 542 | bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes, |
| 543 | unsigned Alignment, |
| 544 | unsigned AddrSpace) const { |
| 545 | return true; |
| 546 | } |
| 547 | |
| 548 | unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize, |
| 549 | unsigned ChainSizeInBytes, |
| 550 | VectorType *VecTy) const { |
| 551 | return VF; |
| 552 | } |
| 553 | |
| 554 | unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize, |
| 555 | unsigned ChainSizeInBytes, |
| 556 | VectorType *VecTy) const { |
| 557 | return VF; |
| 558 | } |
| 559 | |
| 560 | bool useReductionIntrinsic(unsigned Opcode, Type *Ty, |
| 561 | TTI::ReductionFlags Flags) const { |
| 562 | return false; |
| 563 | } |
| 564 | |
| 565 | bool shouldExpandReduction(const IntrinsicInst *II) const { |
| 566 | return true; |
| 567 | } |
| 568 | |
| 569 | protected: |
| 570 | // Obtain the minimum required size to hold the value (without the sign) |
| 571 | // In case of a vector it returns the min required size for one element. |
| 572 | unsigned minRequiredElementSize(const Value* Val, bool &isSigned) { |
| 573 | if (isa<ConstantDataVector>(Val) || isa<ConstantVector>(Val)) { |
| 574 | const auto* VectorValue = cast<Constant>(Val); |
| 575 | |
| 576 | // In case of a vector need to pick the max between the min |
| 577 | // required size for each element |
| 578 | auto *VT = cast<VectorType>(Val->getType()); |
| 579 | |
| 580 | // Assume unsigned elements |
| 581 | isSigned = false; |
| 582 | |
| 583 | // The max required size is the total vector width divided by num |
| 584 | // of elements in the vector |
| 585 | unsigned MaxRequiredSize = VT->getBitWidth() / VT->getNumElements(); |
| 586 | |
| 587 | unsigned MinRequiredSize = 0; |
| 588 | for(unsigned i = 0, e = VT->getNumElements(); i < e; ++i) { |
| 589 | if (auto* IntElement = |
| 590 | dyn_cast<ConstantInt>(VectorValue->getAggregateElement(i))) { |
| 591 | bool signedElement = IntElement->getValue().isNegative(); |
| 592 | // Get the element min required size. |
| 593 | unsigned ElementMinRequiredSize = |
| 594 | IntElement->getValue().getMinSignedBits() - 1; |
| 595 | // In case one element is signed then all the vector is signed. |
| 596 | isSigned |= signedElement; |
| 597 | // Save the max required bit size between all the elements. |
| 598 | MinRequiredSize = std::max(MinRequiredSize, ElementMinRequiredSize); |
| 599 | } |
| 600 | else { |
| 601 | // not an int constant element |
| 602 | return MaxRequiredSize; |
| 603 | } |
| 604 | } |
| 605 | return MinRequiredSize; |
| 606 | } |
| 607 | |
| 608 | if (const auto* CI = dyn_cast<ConstantInt>(Val)) { |
| 609 | isSigned = CI->getValue().isNegative(); |
| 610 | return CI->getValue().getMinSignedBits() - 1; |
| 611 | } |
| 612 | |
| 613 | if (const auto* Cast = dyn_cast<SExtInst>(Val)) { |
| 614 | isSigned = true; |
| 615 | return Cast->getSrcTy()->getScalarSizeInBits() - 1; |
| 616 | } |
| 617 | |
| 618 | if (const auto* Cast = dyn_cast<ZExtInst>(Val)) { |
| 619 | isSigned = false; |
| 620 | return Cast->getSrcTy()->getScalarSizeInBits(); |
| 621 | } |
| 622 | |
| 623 | isSigned = false; |
| 624 | return Val->getType()->getScalarSizeInBits(); |
| 625 | } |
| 626 | |
| 627 | bool isStridedAccess(const SCEV *Ptr) { |
| 628 | return Ptr && isa<SCEVAddRecExpr>(Ptr); |
| 629 | } |
| 630 | |
| 631 | const SCEVConstant *getConstantStrideStep(ScalarEvolution *SE, |
| 632 | const SCEV *Ptr) { |
| 633 | if (!isStridedAccess(Ptr)) |
| 634 | return nullptr; |
| 635 | const SCEVAddRecExpr *AddRec = cast<SCEVAddRecExpr>(Ptr); |
| 636 | return dyn_cast<SCEVConstant>(AddRec->getStepRecurrence(*SE)); |
| 637 | } |
| 638 | |
| 639 | bool isConstantStridedAccessLessThan(ScalarEvolution *SE, const SCEV *Ptr, |
| 640 | int64_t MergeDistance) { |
| 641 | const SCEVConstant *Step = getConstantStrideStep(SE, Ptr); |
| 642 | if (!Step) |
| 643 | return false; |
| 644 | APInt StrideVal = Step->getAPInt(); |
| 645 | if (StrideVal.getBitWidth() > 64) |
| 646 | return false; |
| 647 | // FIXME: Need to take absolute value for negative stride case. |
| 648 | return StrideVal.getSExtValue() < MergeDistance; |
| 649 | } |
| 650 | }; |
| 651 | |
| 652 | /// \brief CRTP base class for use as a mix-in that aids implementing |
| 653 | /// a TargetTransformInfo-compatible class. |
| 654 | template <typename T> |
| 655 | class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase { |
| 656 | private: |
| 657 | typedef TargetTransformInfoImplBase BaseT; |
| 658 | |
| 659 | protected: |
| 660 | explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {} |
| 661 | |
| 662 | public: |
| 663 | using BaseT::getCallCost; |
| 664 | |
| 665 | unsigned getCallCost(const Function *F, int NumArgs) { |
| 666 | assert(F && "A concrete function must be provided to this routine."); |
| 667 | |
| 668 | if (NumArgs < 0) |
| 669 | // Set the argument number to the number of explicit arguments in the |
| 670 | // function. |
| 671 | NumArgs = F->arg_size(); |
| 672 | |
| 673 | if (Intrinsic::ID IID = F->getIntrinsicID()) { |
| 674 | FunctionType *FTy = F->getFunctionType(); |
| 675 | SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end()); |
| 676 | return static_cast<T *>(this) |
| 677 | ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys); |
| 678 | } |
| 679 | |
| 680 | if (!static_cast<T *>(this)->isLoweredToCall(F)) |
| 681 | return TTI::TCC_Basic; // Give a basic cost if it will be lowered |
| 682 | // directly. |
| 683 | |
| 684 | return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs); |
| 685 | } |
| 686 | |
| 687 | unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) { |
| 688 | // Simply delegate to generic handling of the call. |
| 689 | // FIXME: We should use instsimplify or something else to catch calls which |
| 690 | // will constant fold with these arguments. |
| 691 | return static_cast<T *>(this)->getCallCost(F, Arguments.size()); |
| 692 | } |
| 693 | |
| 694 | using BaseT::getGEPCost; |
| 695 | |
| 696 | int getGEPCost(Type *PointeeType, const Value *Ptr, |
| 697 | ArrayRef<const Value *> Operands) { |
| 698 | const GlobalValue *BaseGV = nullptr; |
| 699 | if (Ptr != nullptr) { |
| 700 | // TODO: will remove this when pointers have an opaque type. |
| 701 | assert(Ptr->getType()->getScalarType()->getPointerElementType() == |
| 702 | PointeeType && |
| 703 | "explicit pointee type doesn't match operand's pointee type"); |
| 704 | BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts()); |
| 705 | } |
| 706 | bool HasBaseReg = (BaseGV == nullptr); |
| 707 | |
| 708 | auto PtrSizeBits = DL.getPointerTypeSizeInBits(Ptr->getType()); |
| 709 | APInt BaseOffset(PtrSizeBits, 0); |
| 710 | int64_t Scale = 0; |
| 711 | |
| 712 | auto GTI = gep_type_begin(PointeeType, Operands); |
| 713 | Type *TargetType = nullptr; |
| 714 | |
| 715 | // Handle the case where the GEP instruction has a single operand, |
| 716 | // the basis, therefore TargetType is a nullptr. |
| 717 | if (Operands.empty()) |
| 718 | return !BaseGV ? TTI::TCC_Free : TTI::TCC_Basic; |
| 719 | |
| 720 | for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) { |
| 721 | TargetType = GTI.getIndexedType(); |
| 722 | // We assume that the cost of Scalar GEP with constant index and the |
| 723 | // cost of Vector GEP with splat constant index are the same. |
| 724 | const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I); |
| 725 | if (!ConstIdx) |
| 726 | if (auto Splat = getSplatValue(*I)) |
| 727 | ConstIdx = dyn_cast<ConstantInt>(Splat); |
| 728 | if (StructType *STy = GTI.getStructTypeOrNull()) { |
| 729 | // For structures the index is always splat or scalar constant |
| 730 | assert(ConstIdx && "Unexpected GEP index"); |
| 731 | uint64_t Field = ConstIdx->getZExtValue(); |
| 732 | BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field); |
| 733 | } else { |
| 734 | int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType()); |
| 735 | if (ConstIdx) { |
| 736 | BaseOffset += |
| 737 | ConstIdx->getValue().sextOrTrunc(PtrSizeBits) * ElementSize; |
| 738 | } else { |
| 739 | // Needs scale register. |
| 740 | if (Scale != 0) |
| 741 | // No addressing mode takes two scale registers. |
| 742 | return TTI::TCC_Basic; |
| 743 | Scale = ElementSize; |
| 744 | } |
| 745 | } |
| 746 | } |
| 747 | |
| 748 | // Assumes the address space is 0 when Ptr is nullptr. |
| 749 | unsigned AS = |
| 750 | (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace()); |
| 751 | |
| 752 | if (static_cast<T *>(this)->isLegalAddressingMode( |
| 753 | TargetType, const_cast<GlobalValue *>(BaseGV), |
| 754 | BaseOffset.sextOrTrunc(64).getSExtValue(), HasBaseReg, Scale, AS)) |
| 755 | return TTI::TCC_Free; |
| 756 | return TTI::TCC_Basic; |
| 757 | } |
| 758 | |
| 759 | using BaseT::getIntrinsicCost; |
| 760 | |
| 761 | unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, |
| 762 | ArrayRef<const Value *> Arguments) { |
| 763 | // Delegate to the generic intrinsic handling code. This mostly provides an |
| 764 | // opportunity for targets to (for example) special case the cost of |
| 765 | // certain intrinsics based on constants used as arguments. |
| 766 | SmallVector<Type *, 8> ParamTys; |
| 767 | ParamTys.reserve(Arguments.size()); |
| 768 | for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx) |
| 769 | ParamTys.push_back(Arguments[Idx]->getType()); |
| 770 | return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys); |
| 771 | } |
| 772 | |
| 773 | unsigned getUserCost(const User *U, ArrayRef<const Value *> Operands) { |
| 774 | if (isa<PHINode>(U)) |
| 775 | return TTI::TCC_Free; // Model all PHI nodes as free. |
| 776 | |
| 777 | // Static alloca doesn't generate target instructions. |
| 778 | if (auto *A = dyn_cast<AllocaInst>(U)) |
| 779 | if (A->isStaticAlloca()) |
| 780 | return TTI::TCC_Free; |
| 781 | |
| 782 | if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) { |
| 783 | return static_cast<T *>(this)->getGEPCost(GEP->getSourceElementType(), |
| 784 | GEP->getPointerOperand(), |
| 785 | Operands.drop_front()); |
| 786 | } |
| 787 | |
| 788 | if (auto CS = ImmutableCallSite(U)) { |
| 789 | const Function *F = CS.getCalledFunction(); |
| 790 | if (!F) { |
| 791 | // Just use the called value type. |
| 792 | Type *FTy = CS.getCalledValue()->getType()->getPointerElementType(); |
| 793 | return static_cast<T *>(this) |
| 794 | ->getCallCost(cast<FunctionType>(FTy), CS.arg_size()); |
| 795 | } |
| 796 | |
| 797 | SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end()); |
| 798 | return static_cast<T *>(this)->getCallCost(F, Arguments); |
| 799 | } |
| 800 | |
| 801 | if (const CastInst *CI = dyn_cast<CastInst>(U)) { |
| 802 | // Result of a cmp instruction is often extended (to be used by other |
| 803 | // cmp instructions, logical or return instructions). These are usually |
| 804 | // nop on most sane targets. |
| 805 | if (isa<CmpInst>(CI->getOperand(0))) |
| 806 | return TTI::TCC_Free; |
| 807 | if (isa<SExtInst>(CI) || isa<ZExtInst>(CI) || isa<FPExtInst>(CI)) |
| 808 | return static_cast<T *>(this)->getExtCost(CI, Operands.back()); |
| 809 | } |
| 810 | |
| 811 | return static_cast<T *>(this)->getOperationCost( |
| 812 | Operator::getOpcode(U), U->getType(), |
| 813 | U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr); |
| 814 | } |
| 815 | |
| 816 | int getInstructionLatency(const Instruction *I) { |
| 817 | SmallVector<const Value *, 4> Operands(I->value_op_begin(), |
| 818 | I->value_op_end()); |
| 819 | if (getUserCost(I, Operands) == TTI::TCC_Free) |
| 820 | return 0; |
| 821 | |
| 822 | if (isa<LoadInst>(I)) |
| 823 | return 4; |
| 824 | |
| 825 | Type *DstTy = I->getType(); |
| 826 | |
| 827 | // Usually an intrinsic is a simple instruction. |
| 828 | // A real function call is much slower. |
| 829 | if (auto *CI = dyn_cast<CallInst>(I)) { |
| 830 | const Function *F = CI->getCalledFunction(); |
| 831 | if (!F || static_cast<T *>(this)->isLoweredToCall(F)) |
| 832 | return 40; |
| 833 | // Some intrinsics return a value and a flag, we use the value type |
| 834 | // to decide its latency. |
| 835 | if (StructType* StructTy = dyn_cast<StructType>(DstTy)) |
| 836 | DstTy = StructTy->getElementType(0); |
| 837 | // Fall through to simple instructions. |
| 838 | } |
| 839 | |
| 840 | if (VectorType *VectorTy = dyn_cast<VectorType>(DstTy)) |
| 841 | DstTy = VectorTy->getElementType(); |
| 842 | if (DstTy->isFloatingPointTy()) |
| 843 | return 3; |
| 844 | |
| 845 | return 1; |
| 846 | } |
| 847 | }; |
| 848 | } |
| 849 | |
| 850 | #endif |