Home | History | Annotate | Download | only in Analysis 
      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   int 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 getInliningThresholdMultiplier() { return 1; }
    132 
    133   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
    134                             ArrayRef<Type *> ParamTys) {
    135     switch (IID) {
    136     default:
    137       // Intrinsics rarely (if ever) have normal argument setup constraints.
    138       // Model them as having a basic instruction cost.
    139       // FIXME: This is wrong for libc intrinsics.
    140       return TTI::TCC_Basic;
    141 
    142     case Intrinsic::annotation:
    143     case Intrinsic::assume:
    144     case Intrinsic::dbg_declare:
    145     case Intrinsic::dbg_value:
    146     case Intrinsic::invariant_start:
    147     case Intrinsic::invariant_end:
    148     case Intrinsic::lifetime_start:
    149     case Intrinsic::lifetime_end:
    150     case Intrinsic::objectsize:
    151     case Intrinsic::ptr_annotation:
    152     case Intrinsic::var_annotation:
    153     case Intrinsic::experimental_gc_result:
    154     case Intrinsic::experimental_gc_relocate:
    155       // These intrinsics don't actually represent code after lowering.
    156       return TTI::TCC_Free;
    157     }
    158   }
    159 
    160   bool hasBranchDivergence() { return false; }
    161 
    162   bool isSourceOfDivergence(const Value *V) { return false; }
    163 
    164   bool isLoweredToCall(const Function *F) {
    165     // FIXME: These should almost certainly not be handled here, and instead
    166     // handled with the help of TLI or the target itself. This was largely
    167     // ported from existing analysis heuristics here so that such refactorings
    168     // can take place in the future.
    169 
    170     if (F->isIntrinsic())
    171       return false;
    172 
    173     if (F->hasLocalLinkage() || !F->hasName())
    174       return true;
    175 
    176     StringRef Name = F->getName();
    177 
    178     // These will all likely lower to a single selection DAG node.
    179     if (Name == "copysign" || Name == "copysignf" || Name == "copysignl" ||
    180         Name == "fabs" || Name == "fabsf" || Name == "fabsl" || Name == "sin" ||
    181         Name == "fmin" || Name == "fminf" || Name == "fminl" ||
    182         Name == "fmax" || Name == "fmaxf" || Name == "fmaxl" ||
    183         Name == "sinf" || Name == "sinl" || Name == "cos" || Name == "cosf" ||
    184         Name == "cosl" || Name == "sqrt" || Name == "sqrtf" || Name == "sqrtl")
    185       return false;
    186 
    187     // These are all likely to be optimized into something smaller.
    188     if (Name == "pow" || Name == "powf" || Name == "powl" || Name == "exp2" ||
    189         Name == "exp2l" || Name == "exp2f" || Name == "floor" ||
    190         Name == "floorf" || Name == "ceil" || Name == "round" ||
    191         Name == "ffs" || Name == "ffsl" || Name == "abs" || Name == "labs" ||
    192         Name == "llabs")
    193       return false;
    194 
    195     return true;
    196   }
    197 
    198   void getUnrollingPreferences(Loop *, TTI::UnrollingPreferences &) {}
    199 
    200   bool isLegalAddImmediate(int64_t Imm) { return false; }
    201 
    202   bool isLegalICmpImmediate(int64_t Imm) { return false; }
    203 
    204   bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
    205                              bool HasBaseReg, int64_t Scale,
    206                              unsigned AddrSpace) {
    207     // Guess that only reg and reg+reg addressing is allowed. This heuristic is
    208     // taken from the implementation of LSR.
    209     return !BaseGV && BaseOffset == 0 && (Scale == 0 || Scale == 1);
    210   }
    211 
    212   bool isLegalMaskedStore(Type *DataType) { return false; }
    213 
    214   bool isLegalMaskedLoad(Type *DataType) { return false; }
    215 
    216   bool isLegalMaskedScatter(Type *DataType) { return false; }
    217 
    218   bool isLegalMaskedGather(Type *DataType) { return false; }
    219 
    220   int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
    221                            bool HasBaseReg, int64_t Scale, unsigned AddrSpace) {
    222     // Guess that all legal addressing mode are free.
    223     if (isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
    224                               Scale, AddrSpace))
    225       return 0;
    226     return -1;
    227   }
    228 
    229   bool isTruncateFree(Type *Ty1, Type *Ty2) { return false; }
    230 
    231   bool isProfitableToHoist(Instruction *I) { return true; }
    232 
    233   bool isTypeLegal(Type *Ty) { return false; }
    234 
    235   unsigned getJumpBufAlignment() { return 0; }
    236 
    237   unsigned getJumpBufSize() { return 0; }
    238 
    239   bool shouldBuildLookupTables() { return true; }
    240 
    241   bool enableAggressiveInterleaving(bool LoopHasReductions) { return false; }
    242 
    243   bool enableInterleavedAccessVectorization() { return false; }
    244 
    245   bool isFPVectorizationPotentiallyUnsafe() { return false; }
    246 
    247   bool allowsMisalignedMemoryAccesses(unsigned BitWidth,
    248                                       unsigned AddressSpace,
    249                                       unsigned Alignment,
    250                                       bool *Fast) { return false; }
    251 
    252   TTI::PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) {
    253     return TTI::PSK_Software;
    254   }
    255 
    256   bool haveFastSqrt(Type *Ty) { return false; }
    257 
    258   unsigned getFPOpCost(Type *Ty) { return TargetTransformInfo::TCC_Basic; }
    259 
    260   int getIntImmCodeSizeCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
    261                             Type *Ty) {
    262     return 0;
    263   }
    264 
    265   unsigned getIntImmCost(const APInt &Imm, Type *Ty) { return TTI::TCC_Basic; }
    266 
    267   unsigned getIntImmCost(unsigned Opcode, unsigned Idx, const APInt &Imm,
    268                          Type *Ty) {
    269     return TTI::TCC_Free;
    270   }
    271 
    272   unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
    273                          Type *Ty) {
    274     return TTI::TCC_Free;
    275   }
    276 
    277   unsigned getNumberOfRegisters(bool Vector) { return 8; }
    278 
    279   unsigned getRegisterBitWidth(bool Vector) { return 32; }
    280 
    281   unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) { return 128; }
    282 
    283   unsigned getCacheLineSize() { return 0; }
    284 
    285   unsigned getPrefetchDistance() { return 0; }
    286 
    287   unsigned getMinPrefetchStride() { return 1; }
    288 
    289   unsigned getMaxPrefetchIterationsAhead() { return UINT_MAX; }
    290 
    291   unsigned getMaxInterleaveFactor(unsigned VF) { return 1; }
    292 
    293   unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
    294                                   TTI::OperandValueKind Opd1Info,
    295                                   TTI::OperandValueKind Opd2Info,
    296                                   TTI::OperandValueProperties Opd1PropInfo,
    297                                   TTI::OperandValueProperties Opd2PropInfo) {
    298     return 1;
    299   }
    300 
    301   unsigned getShuffleCost(TTI::ShuffleKind Kind, Type *Ty, int Index,
    302                           Type *SubTp) {
    303     return 1;
    304   }
    305 
    306   unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) { return 1; }
    307 
    308   unsigned getExtractWithExtendCost(unsigned Opcode, Type *Dst,
    309                                     VectorType *VecTy, unsigned Index) {
    310     return 1;
    311   }
    312 
    313   unsigned getCFInstrCost(unsigned Opcode) { return 1; }
    314 
    315   unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) {
    316     return 1;
    317   }
    318 
    319   unsigned getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) {
    320     return 1;
    321   }
    322 
    323   unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
    324                            unsigned AddressSpace) {
    325     return 1;
    326   }
    327 
    328   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
    329                                  unsigned AddressSpace) {
    330     return 1;
    331   }
    332 
    333   unsigned getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
    334                                   bool VariableMask,
    335                                   unsigned Alignment) {
    336     return 1;
    337   }
    338 
    339   unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
    340                                       unsigned Factor,
    341                                       ArrayRef<unsigned> Indices,
    342                                       unsigned Alignment,
    343                                       unsigned AddressSpace) {
    344     return 1;
    345   }
    346 
    347   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
    348                                  ArrayRef<Type *> Tys, FastMathFlags FMF) {
    349     return 1;
    350   }
    351   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
    352                                  ArrayRef<Value *> Args, FastMathFlags FMF) {
    353     return 1;
    354   }
    355 
    356   unsigned getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) {
    357     return 1;
    358   }
    359 
    360   unsigned getNumberOfParts(Type *Tp) { return 0; }
    361 
    362   unsigned getAddressComputationCost(Type *Tp, bool) { return 0; }
    363 
    364   unsigned getReductionCost(unsigned, Type *, bool) { return 1; }
    365 
    366   unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) { return 0; }
    367 
    368   bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) {
    369     return false;
    370   }
    371 
    372   Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
    373                                            Type *ExpectedType) {
    374     return nullptr;
    375   }
    376 
    377   bool areInlineCompatible(const Function *Caller,
    378                            const Function *Callee) const {
    379     return (Caller->getFnAttribute("target-cpu") ==
    380             Callee->getFnAttribute("target-cpu")) &&
    381            (Caller->getFnAttribute("target-features") ==
    382             Callee->getFnAttribute("target-features"));
    383   }
    384 };
    385 
    386 /// \brief CRTP base class for use as a mix-in that aids implementing
    387 /// a TargetTransformInfo-compatible class.
    388 template <typename T>
    389 class TargetTransformInfoImplCRTPBase : public TargetTransformInfoImplBase {
    390 private:
    391   typedef TargetTransformInfoImplBase BaseT;
    392 
    393 protected:
    394   explicit TargetTransformInfoImplCRTPBase(const DataLayout &DL) : BaseT(DL) {}
    395 
    396 public:
    397   // Provide value semantics. MSVC requires that we spell all of these out.
    398   TargetTransformInfoImplCRTPBase(const TargetTransformInfoImplCRTPBase &Arg)
    399       : BaseT(static_cast<const BaseT &>(Arg)) {}
    400   TargetTransformInfoImplCRTPBase(TargetTransformInfoImplCRTPBase &&Arg)
    401       : BaseT(std::move(static_cast<BaseT &>(Arg))) {}
    402 
    403   using BaseT::getCallCost;
    404 
    405   unsigned getCallCost(const Function *F, int NumArgs) {
    406     assert(F && "A concrete function must be provided to this routine.");
    407 
    408     if (NumArgs < 0)
    409       // Set the argument number to the number of explicit arguments in the
    410       // function.
    411       NumArgs = F->arg_size();
    412 
    413     if (Intrinsic::ID IID = F->getIntrinsicID()) {
    414       FunctionType *FTy = F->getFunctionType();
    415       SmallVector<Type *, 8> ParamTys(FTy->param_begin(), FTy->param_end());
    416       return static_cast<T *>(this)
    417           ->getIntrinsicCost(IID, FTy->getReturnType(), ParamTys);
    418     }
    419 
    420     if (!static_cast<T *>(this)->isLoweredToCall(F))
    421       return TTI::TCC_Basic; // Give a basic cost if it will be lowered
    422                              // directly.
    423 
    424     return static_cast<T *>(this)->getCallCost(F->getFunctionType(), NumArgs);
    425   }
    426 
    427   unsigned getCallCost(const Function *F, ArrayRef<const Value *> Arguments) {
    428     // Simply delegate to generic handling of the call.
    429     // FIXME: We should use instsimplify or something else to catch calls which
    430     // will constant fold with these arguments.
    431     return static_cast<T *>(this)->getCallCost(F, Arguments.size());
    432   }
    433 
    434   using BaseT::getGEPCost;
    435 
    436   int getGEPCost(Type *PointeeType, const Value *Ptr,
    437                  ArrayRef<const Value *> Operands) {
    438     const GlobalValue *BaseGV = nullptr;
    439     if (Ptr != nullptr) {
    440       // TODO: will remove this when pointers have an opaque type.
    441       assert(Ptr->getType()->getScalarType()->getPointerElementType() ==
    442                  PointeeType &&
    443              "explicit pointee type doesn't match operand's pointee type");
    444       BaseGV = dyn_cast<GlobalValue>(Ptr->stripPointerCasts());
    445     }
    446     bool HasBaseReg = (BaseGV == nullptr);
    447     int64_t BaseOffset = 0;
    448     int64_t Scale = 0;
    449 
    450     // Assumes the address space is 0 when Ptr is nullptr.
    451     unsigned AS =
    452         (Ptr == nullptr ? 0 : Ptr->getType()->getPointerAddressSpace());
    453     auto GTI = gep_type_begin(PointeeType, AS, Operands);
    454     for (auto I = Operands.begin(); I != Operands.end(); ++I, ++GTI) {
    455       // We assume that the cost of Scalar GEP with constant index and the
    456       // cost of Vector GEP with splat constant index are the same.
    457       const ConstantInt *ConstIdx = dyn_cast<ConstantInt>(*I);
    458       if (!ConstIdx)
    459         if (auto Splat = getSplatValue(*I))
    460           ConstIdx = dyn_cast<ConstantInt>(Splat);
    461       if (isa<SequentialType>(*GTI)) {
    462         int64_t ElementSize = DL.getTypeAllocSize(GTI.getIndexedType());
    463         if (ConstIdx)
    464           BaseOffset += ConstIdx->getSExtValue() * ElementSize;
    465         else {
    466           // Needs scale register.
    467           if (Scale != 0)
    468             // No addressing mode takes two scale registers.
    469             return TTI::TCC_Basic;
    470           Scale = ElementSize;
    471         }
    472       } else {
    473         StructType *STy = cast<StructType>(*GTI);
    474         // For structures the index is always splat or scalar constant
    475         assert(ConstIdx && "Unexpected GEP index");
    476         uint64_t Field = ConstIdx->getZExtValue();
    477         BaseOffset += DL.getStructLayout(STy)->getElementOffset(Field);
    478       }
    479     }
    480 
    481     if (static_cast<T *>(this)->isLegalAddressingMode(
    482             PointerType::get(*GTI, AS), const_cast<GlobalValue *>(BaseGV),
    483             BaseOffset, HasBaseReg, Scale, AS)) {
    484       return TTI::TCC_Free;
    485     }
    486     return TTI::TCC_Basic;
    487   }
    488 
    489   using BaseT::getIntrinsicCost;
    490 
    491   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
    492                             ArrayRef<const Value *> Arguments) {
    493     // Delegate to the generic intrinsic handling code. This mostly provides an
    494     // opportunity for targets to (for example) special case the cost of
    495     // certain intrinsics based on constants used as arguments.
    496     SmallVector<Type *, 8> ParamTys;
    497     ParamTys.reserve(Arguments.size());
    498     for (unsigned Idx = 0, Size = Arguments.size(); Idx != Size; ++Idx)
    499       ParamTys.push_back(Arguments[Idx]->getType());
    500     return static_cast<T *>(this)->getIntrinsicCost(IID, RetTy, ParamTys);
    501   }
    502 
    503   unsigned getUserCost(const User *U) {
    504     if (isa<PHINode>(U))
    505       return TTI::TCC_Free; // Model all PHI nodes as free.
    506 
    507     if (const GEPOperator *GEP = dyn_cast<GEPOperator>(U)) {
    508       SmallVector<Value *, 4> Indices(GEP->idx_begin(), GEP->idx_end());
    509       return static_cast<T *>(this)->getGEPCost(
    510           GEP->getSourceElementType(), GEP->getPointerOperand(), Indices);
    511     }
    512 
    513     if (auto CS = ImmutableCallSite(U)) {
    514       const Function *F = CS.getCalledFunction();
    515       if (!F) {
    516         // Just use the called value type.
    517         Type *FTy = CS.getCalledValue()->getType()->getPointerElementType();
    518         return static_cast<T *>(this)
    519             ->getCallCost(cast<FunctionType>(FTy), CS.arg_size());
    520       }
    521 
    522       SmallVector<const Value *, 8> Arguments(CS.arg_begin(), CS.arg_end());
    523       return static_cast<T *>(this)->getCallCost(F, Arguments);
    524     }
    525 
    526     if (const CastInst *CI = dyn_cast<CastInst>(U)) {
    527       // Result of a cmp instruction is often extended (to be used by other
    528       // cmp instructions, logical or return instructions). These are usually
    529       // nop on most sane targets.
    530       if (isa<CmpInst>(CI->getOperand(0)))
    531         return TTI::TCC_Free;
    532     }
    533 
    534     return static_cast<T *>(this)->getOperationCost(
    535         Operator::getOpcode(U), U->getType(),
    536         U->getNumOperands() == 1 ? U->getOperand(0)->getType() : nullptr);
    537   }
    538 };
    539 }
    540 
    541 #endif
    542