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/TargetTransformInfo.h" 19 #include "llvm/IR/CallSite.h" 20 #include "llvm/IR/DataLayout.h" 21 #include "llvm/IR/Function.h" 22 #include "llvm/IR/GetElementPtrTypeIterator.h" 23 #include "llvm/IR/Operator.h" 24 #include "llvm/IR/Type.h" 25 #include "llvm/Analysis/VectorUtils.h" 26 27 namespace llvm { 28 29 /// \brief Base class for use as a mix-in that aids implementing 30 /// a TargetTransformInfo-compatible class. 31 class TargetTransformInfoImplBase { 32 protected: 33 typedef TargetTransformInfo TTI; 34 35 const DataLayout &DL; 36 37 explicit TargetTransformInfoImplBase(const DataLayout &DL) : DL(DL) {} 38 39 public: 40 // Provide value semantics. MSVC requires that we spell all of these out. 41 TargetTransformInfoImplBase(const TargetTransformInfoImplBase &Arg) 42 : DL(Arg.DL) {} 43 TargetTransformInfoImplBase(TargetTransformInfoImplBase &&Arg) : DL(Arg.DL) {} 44 45 const DataLayout &getDataLayout() const { return DL; } 46 47 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) { 48 switch (Opcode) { 49 default: 50 // By default, just classify everything as 'basic'. 51 return TTI::TCC_Basic; 52 53 case Instruction::GetElementPtr: 54 llvm_unreachable("Use getGEPCost for GEP operations!"); 55 56 case Instruction::BitCast: 57 assert(OpTy && "Cast instructions must provide the operand type"); 58 if (Ty == OpTy || (Ty->isPointerTy() && OpTy->isPointerTy())) 59 // Identity and pointer-to-pointer casts are free. 60 return TTI::TCC_Free; 61 62 // Otherwise, the default basic cost is used. 63 return TTI::TCC_Basic; 64 65 case Instruction::FDiv: 66 case Instruction::FRem: 67 case Instruction::SDiv: 68 case Instruction::SRem: 69 case Instruction::UDiv: 70 case Instruction::URem: 71 return TTI::TCC_Expensive; 72 73 case Instruction::IntToPtr: { 74 // An inttoptr cast is free so long as the input is a legal integer type 75 // which doesn't contain values outside the range of a pointer. 76 unsigned OpSize = OpTy->getScalarSizeInBits(); 77 if (DL.isLegalInteger(OpSize) && 78 OpSize <= DL.getPointerTypeSizeInBits(Ty)) 79 return TTI::TCC_Free; 80 81 // Otherwise it's not a no-op. 82 return TTI::TCC_Basic; 83 } 84 case Instruction::PtrToInt: { 85 // A ptrtoint cast is free so long as the result is large enough to store 86 // the pointer, and a legal integer type. 87 unsigned DestSize = Ty->getScalarSizeInBits(); 88 if (DL.isLegalInteger(DestSize) && 89 DestSize >= DL.getPointerTypeSizeInBits(OpTy)) 90 return TTI::TCC_Free; 91 92 // Otherwise it's not a no-op. 93 return TTI::TCC_Basic; 94 } 95 case Instruction::Trunc: 96 // trunc to a native type is free (assuming the target has compare and 97 // shift-right of the same width). 98 if (DL.isLegalInteger(DL.getTypeSizeInBits(Ty))) 99 return TTI::TCC_Free; 100 101 return TTI::TCC_Basic; 102 } 103 } 104 105 unsigned getGEPCost(Type *PointeeType, const Value *Ptr, 106 ArrayRef<const Value *> Operands) { 107 // In the basic model, we just assume that all-constant GEPs will be folded 108 // into their uses via addressing modes. 109 for (unsigned Idx = 0, Size = Operands.size(); Idx != Size; ++Idx) 110 if (!isa<Constant>(Operands[Idx])) 111 return TTI::TCC_Basic; 112 113 return TTI::TCC_Free; 114 } 115 116 unsigned getCallCost(FunctionType *FTy, int NumArgs) { 117 assert(FTy && "FunctionType must be provided to this routine."); 118 119 // The target-independent implementation just measures the size of the 120 // function by approximating that each argument will take on average one 121 // instruction to prepare. 122 123 if (NumArgs < 0) 124 // Set the argument number to the number of explicit arguments in the 125 // function. 126 NumArgs = FTy->getNumParams(); 127 128 return TTI::TCC_Basic * (NumArgs + 1); 129 } 130 131 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 132 ArrayRef<Type *> ParamTys) { 133 switch (IID) { 134 default: 135 // Intrinsics rarely (if ever) have normal argument setup constraints. 136 // Model them as having a basic instruction cost. 137 // FIXME: This is wrong for libc intrinsics. 138 return TTI::TCC_Basic; 139 140 case Intrinsic::annotation: 141 case Intrinsic::assume: 142 case Intrinsic::dbg_declare: 143 case Intrinsic::dbg_value: 144 case Intrinsic::invariant_start: 145 case Intrinsic::invariant_end: 146 case Intrinsic::lifetime_start: 147 case Intrinsic::lifetime_end: 148 case Intrinsic::objectsize: 149 case Intrinsic::ptr_annotation: 150 case Intrinsic::var_annotation: 151 case Intrinsic::experimental_gc_result_int: 152 case Intrinsic::experimental_gc_result_float: 153 case Intrinsic::experimental_gc_result_ptr: 154 case Intrinsic::experimental_gc_result: 155 case Intrinsic::experimental_gc_relocate: 156 // These intrinsics don't actually represent code after lowering. 157 return TTI::TCC_Free; 158 } 159 } 160 161 bool hasBranchDivergence() { return false; } 162 163 bool isSourceOfDivergence(const Value *V) { return false; } 164 165 bool isLoweredToCall(const Function *F) { 166 // FIXME: These should almost certainly not be handled here, and instead 167 // handled with the help of TLI or the target itself. This was largely 168 // ported from existing analysis heuristics here so that such refactorings 169 // can take place in the future. 170 171 if (F->isIntrinsic()) 172 return false; 173 174 if (F->hasLocalLinkage() || !F->hasName()) 175 return true; 176 177 StringRef Name = F->getName(); 178 179 // These will all likely lower to a single selection DAG node. 180 if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" || 181 Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" || 182 Name == "fmin" || Name == "fminf" || Name == "fminl" || 183 Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" || 184 Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" || 185 Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl") 186 return false; 187 188 // These are all likely to be optimized into something smaller. 189 if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" || 190 Name == "exp2l" || Name == "exp2f" || Name == "floor" || 191 Name == "floorf" || Name == "ceil" || Name == "round" || 192 Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" || 193 Name == "llabs") 194 return false; 195 196 return true; 197 } 198 199 void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {} 200 201 bool isLegalAddImmediate(int64_t Imm) { return false; } 202 203 bool isLegalICmpImmediate(int64_t Imm) { return false; } 204 205 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, 206 bool HasBaseReg, int64_t Scale, 207 unsigned AddrSpace) { 208 // Guess that only reg and reg+reg addressing is allowed. This heuristic is 209 // taken from the implementation of LSR. 210 return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1); 211 } 212 213 bool isLegalMaskedStore(Type *DataType) { return false; } 214 215 bool isLegalMaskedLoad(Type *DataType) { return false; } 216 217 bool isLegalMaskedScatter(Type *DataType) { return false; } 218 219 bool isLegalMaskedGather(Type *DataType) { return false; } 220 221 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset, 222 bool HasBaseReg, int64_t Scale, unsigned AddrSpace) { 223 // Guess that all legal addressing mode are free. 224 if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg, 225 Scale, AddrSpace)) 226 return 0; 227 return -1; 228 } 229 230 bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; } 231 232 bool isProfitableToHoist(Instruction *I) { return true; } 233 234 bool isTypeLegal(Type *Ty) { return false; } 235 236 unsigned getJumpBufAlignment() { return 0; } 237 238 unsigned getJumpBufSize() { return 0; } 239 240 bool shouldBuildLookupTables() { return true; } 241 242 bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; } 243 244 bool enableInterleavedAccessVectorization() { return false; } 245 246 TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) { 247 return TTI::PSK_Software; 248 } 249 250 bool haveFastSqrt(Type *Ty) { return false; } 251 252 unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; } 253 254 unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; } 255 256 unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm, 257 Type *Ty) { 258 return TTI::TCC_Free; 259 } 260 261 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm, 262 Type *Ty) { 263 return TTI::TCC_Free; 264 } 265 266 unsigned getNumberOfRegisters(bool Vector) { return 8; } 267 268 unsigned getRegisterBitWidth(bool Vector) { return 32; } 269 270 unsigned getMaxInterleaveFactor(unsigned VF) { return 1; } 271 272 unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty, 273 TTI::OperandValueKind Opd1Info, 274 TTI::OperandValueKind Opd2Info, 275 TTI::OperandValueProperties Opd1PropInfo, 276 TTI::OperandValueProperties Opd2PropInfo) { 277 return 1; 278 } 279 280 unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index, 281 Type *SubTp) { 282 return 1; 283 } 284 285 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { return 1; } 286 287 unsigned getCFInstrCost(unsigned Opcode) { return 1; } 288 289 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) { 290 return 1; 291 } 292 293 unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) { 294 return 1; 295 } 296 297 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 298 unsigned AddressSpace) { 299 return 1; 300 } 301 302 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment, 303 unsigned AddressSpace) { 304 return 1; 305 } 306 307 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, 308 unsigned Factor, 309 ArrayRef<unsigned> Indices, 310 unsigned Alignment, 311 unsigned AddressSpace) { 312 return 1; 313 } 314 315 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 316 ArrayRef<Type *> Tys) { 317 return 1; 318 } 319 320 unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) { 321 return 1; 322 } 323 324 unsigned getNumberOfParts(Type *Tp) { return 0; } 325 326 unsigned getAddressComputationCost(Type *Tp, bool) { return 0; } 327 328 unsigned getReductionCost(unsigned, Type *, bool) { return 1; } 329 330 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; } 331 332 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) { 333 return false; 334 } 335 336 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst, 337 Type *ExpectedType) { 338 return nullptr; 339 } 340 341 bool areInlineCompatible(const Function *Caller, 342 const Function *Callee) const { 343 return (Caller->getFnAttribute("target-cpu") == 344 Callee->getFnAttribute("target-cpu")) && 345 (Caller->getFnAttribute("target-features") == 346 Callee->getFnAttribute("target-features")); 347 } 348 }; 349 350 /// \brief CRTP base class for use as a mix-in that aids implementing 351 /// a TargetTransformInfo-compatible class. 352 template <typename T> 353 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase { 354 private: 355 typedef TargetTransformInfoImplBase BaseT; 356 357 protected: 358 explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {} 359 360 public: 361 // Provide value semantics. MSVC requires that we spell all of these out. 362 TargetTransformInfoImplCRTPBase(const TargetTransformInfoImplCRTPBase &Arg) 363 : BaseT(static_cast<const BaseT &>(Arg)) {} 364 TargetTransformInfoImplCRTPBase(TargetTransformInfoImplCRTPBase &&Arg) 365 : BaseT(std::move(static_cast<BaseT &>(Arg))) {} 366 367 using BaseT::getCallCost; 368 369 unsigned getCallCost(const Function *F, int NumArgs) { 370 assert(F && "A concrete function must be provided to this routine."); 371 372 if (NumArgs < 0) 373 // Set the argument number to the number of explicit arguments in the 374 // function. 375 NumArgs = F->arg_size(); 376 377 if (Intrinsic::ID IID = F->getIntrinsicID()) { 378 FunctionType *FTy = F->getFunctionType(); 379 SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end()); 380 return static_cast<T *>(this) 381 ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys); 382 } 383 384 if (!static_cast<T *>(this)->isLoweredToCall(F)) 385 return TTI::TCC_Basic; // Give a basic cost if it will be lowered 386 // directly. 387 388 return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs); 389 } 390 391 unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) { 392 // Simply delegate to generic handling of the call. 393 // FIXME: We should use instsimplify or something else to catch calls which 394 // will constant fold with these arguments. 395 return static_cast<T *>(this)->getCallCost(F, Arguments.size()); 396 } 397 398 using BaseT::getGEPCost; 399 400 unsigned getGEPCost(Type *PointeeType, const Value *Ptr, 401 ArrayRef<const Value *> Operands) { 402 const GlobalValue *BaseGV = nullptr; 403 if (Ptr != nullptr) { 404 // TODO: will remove this when pointers have an opaque type. 405 assert(Ptr->getType()->getScalarType()->getPointerElementType() == 406 PointeeType && 407 "explicit pointee type doesn't match operand's pointee type"); 408 BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts()); 409 } 410 bool HasBaseReg = (BaseGV == nullptr); 411 int64_t BaseOffset = 0; 412 int64_t Scale = 0; 413 414 // Assumes the address space is 0 when Ptr is nullptr. 415 unsigned AS = 416 (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace()); 417 auto GTI = gep_type_begin(PointerType::get(PointeeType, AS), Operands); 418 for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) { 419 // We assume that the cost of Scalar GEP with constant index and the 420 // cost of Vector GEP with splat constant index are the same. 421 const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I); 422 if (!ConstIdx) 423 if (auto Splat = getSplatValue(*I)) 424 ConstIdx = dyn_cast<ConstantInt>(Splat); 425 if (isa<SequentialType>(*GTI)) { 426 int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType()); 427 if (ConstIdx) 428 BaseOffset += ConstIdx->getSExtValue() * ElementSize; 429 else { 430 // Needs scale register. 431 if (Scale != 0) 432 // No addressing mode takes two scale registers. 433 return TTI::TCC_Basic; 434 Scale = ElementSize; 435 } 436 } else { 437 StructType *STy = cast<StructType>(*GTI); 438 // For structures the index is always splat or scalar constant 439 assert(ConstIdx && "Unexpected GEP index"); 440 uint64_t Field = ConstIdx->getZExtValue(); 441 BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field); 442 } 443 } 444 445 if (static_cast<T *>(this)->isLegalAddressingMode( 446 PointerType::get(*GTI, AS), const_cast<GlobalValue *>(BaseGV), 447 BaseOffset, HasBaseReg, Scale, AS)) { 448 return TTI::TCC_Free; 449 } 450 return TTI::TCC_Basic; 451 } 452 453 using BaseT::getIntrinsicCost; 454 455 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 456 ArrayRef<const Value *> Arguments) { 457 // Delegate to the generic intrinsic handling code. This mostly provides an 458 // opportunity for targets to (for example) special case the cost of 459 // certain intrinsics based on constants used as arguments. 460 SmallVector<Type *, 8> ParamTys; 461 ParamTys.reserve(Arguments.size()); 462 for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx) 463 ParamTys.push_back(Arguments[Idx]->getType()); 464 return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys); 465 } 466 467 unsigned getUserCost(const User *U) { 468 if (isa<PHINode>(U)) 469 return TTI::TCC_Free; // Model all PHI nodes as free. 470 471 if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) { 472 SmallVector<Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end()); 473 return static_cast<T *>(this)->getGEPCost( 474 GEP->getSourceElementType(), GEP->getPointerOperand(), Indices); 475 } 476 477 if (auto CS = ImmutableCallSite(U)) { 478 const Function *F = CS.getCalledFunction(); 479 if (!F) { 480 // Just use the called value type. 481 Type *FTy = CS.getCalledValue()->getType()->getPointerElementType(); 482 return static_cast<T *>(this) 483 ->getCallCost(cast<FunctionType>(FTy), CS.arg_size()); 484 } 485 486 SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end()); 487 return static_cast<T *>(this)->getCallCost(F, Arguments); 488 } 489 490 if (const CastInst *CI = dyn_cast<CastInst>(U)) { 491 // Result of a cmp instruction is often extended (to be used by other 492 // cmp instructions, logical or return instructions). These are usually 493 // nop on most sane targets. 494 if (isa<CmpInst>(CI->getOperand(0))) 495 return TTI::TCC_Free; 496 } 497 498 return static_cast<T *>(this)->getOperationCost( 499 Operator::getOpcode(U), U->getType(), 500 U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr); 501 } 502 }; 503 } 504 505 #endif 506