TargetTransformInfo.h revision 13086a658ae06046ded902229f9918b8bad505bd
1//===- llvm/Analysis/TargetTransformInfo.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// 10// This pass exposes codegen information to IR-level passes. Every 11// transformation that uses codegen information is broken into three parts: 12// 1. The IR-level analysis pass. 13// 2. The IR-level transformation interface which provides the needed 14// information. 15// 3. Codegen-level implementation which uses target-specific hooks. 16// 17// This file defines #2, which is the interface that IR-level transformations 18// use for querying the codegen. 19// 20//===----------------------------------------------------------------------===// 21 22#ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFO_H 23#define LLVM_ANALYSIS_TARGETTRANSFORMINFO_H 24 25#include "llvm/CodeGen/ValueTypes.h" 26#include "llvm/IR/GlobalValue.h" 27#include "llvm/IR/Intrinsics.h" 28#include "llvm/IR/Type.h" 29#include "llvm/Pass.h" 30#include "llvm/Support/DataTypes.h" 31 32namespace llvm { 33 34/// TargetTransformInfo - This pass provides access to the codegen 35/// interfaces that are needed for IR-level transformations. 36class TargetTransformInfo { 37protected: 38 /// \brief The TTI instance one level down the stack. 39 /// 40 /// This is used to implement the default behavior all of the methods which 41 /// is to delegate up through the stack of TTIs until one can answer the 42 /// query. 43 TargetTransformInfo *PrevTTI; 44 45 /// \brief The top of the stack of TTI analyses available. 46 /// 47 /// This is a convenience routine maintained as TTI analyses become available 48 /// that complements the PrevTTI delegation chain. When one part of an 49 /// analysis pass wants to query another part of the analysis pass it can use 50 /// this to start back at the top of the stack. 51 TargetTransformInfo *TopTTI; 52 53 /// All pass subclasses must in their initializePass routine call 54 /// pushTTIStack with themselves to update the pointers tracking the previous 55 /// TTI instance in the analysis group's stack, and the top of the analysis 56 /// group's stack. 57 void pushTTIStack(Pass *P); 58 59 /// All pass subclasses must in their finalizePass routine call popTTIStack 60 /// to update the pointers tracking the previous TTI instance in the analysis 61 /// group's stack, and the top of the analysis group's stack. 62 void popTTIStack(); 63 64 /// All pass subclasses must call TargetTransformInfo::getAnalysisUsage. 65 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 66 67public: 68 /// This class is intended to be subclassed by real implementations. 69 virtual ~TargetTransformInfo() = 0; 70 71 /// \name Generic Target Information 72 /// @{ 73 74 /// \brief Underlying constants for 'cost' values in this interface. 75 /// 76 /// Many APIs in this interface return a cost. This enum defines the 77 /// fundamental values that should be used to interpret (and produce) those 78 /// costs. The costs are returned as an unsigned rather than a member of this 79 /// enumeration because it is expected that the cost of one IR instruction 80 /// may have a multiplicative factor to it or otherwise won't fit dircetly 81 /// into the enum. Moreover, it is common to sum or average costs which works 82 /// better as simple integral values. Thus this enum only provides constants. 83 /// 84 /// Note that these costs should usually reflect the intersection of code-size 85 /// cost and execution cost. A free instruction is typically one that folds 86 /// into another instruction. For example, reg-to-reg moves can often be 87 /// skipped by renaming the registers in the CPU, but they still are encoded 88 /// and thus wouldn't be considered 'free' here. 89 enum TargetCostConstants { 90 TCC_Free = 0, ///< Expected to fold away in lowering. 91 TCC_Basic = 1, ///< The cost of a typical 'add' instruction. 92 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86. 93 }; 94 95 /// \brief Estimate the cost of a specific operation when lowered. 96 /// 97 /// Note that this is designed to work on an arbitrary synthetic opcode, and 98 /// thus work for hypothetical queries before an instruction has even been 99 /// formed. However, this does *not* work for GEPs, and must not be called 100 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as 101 /// analyzing a GEP's cost required more information. 102 /// 103 /// Typically only the result type is required, and the operand type can be 104 /// omitted. However, if the opcode is one of the cast instructions, the 105 /// operand type is required. 106 /// 107 /// The returned cost is defined in terms of \c TargetCostConstants, see its 108 /// comments for a detailed explanation of the cost values. 109 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, 110 Type *OpTy = 0) const; 111 112 /// \brief Estimate the cost of a GEP operation when lowered. 113 /// 114 /// The contract for this function is the same as \c getOperationCost except 115 /// that it supports an interface that provides extra information specific to 116 /// the GEP operation. 117 virtual unsigned getGEPCost(const Value *Ptr, 118 ArrayRef<const Value *> Operands) const; 119 120 /// \brief Estimate the cost of a function call when lowered. 121 /// 122 /// The contract for this is the same as \c getOperationCost except that it 123 /// supports an interface that provides extra information specific to call 124 /// instructions. 125 /// 126 /// This is the most basic query for estimating call cost: it only knows the 127 /// function type and (potentially) the number of arguments at the call site. 128 /// The latter is only interesting for varargs function types. 129 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const; 130 131 /// \brief Estimate the cost of calling a specific function when lowered. 132 /// 133 /// This overload adds the ability to reason about the particular function 134 /// being called in the event it is a library call with special lowering. 135 virtual unsigned getCallCost(const Function *F, int NumArgs = -1) const; 136 137 /// \brief Estimate the cost of calling a specific function when lowered. 138 /// 139 /// This overload allows specifying a set of candidate argument values. 140 virtual unsigned getCallCost(const Function *F, 141 ArrayRef<const Value *> Arguments) const; 142 143 /// \brief Estimate the cost of an intrinsic when lowered. 144 /// 145 /// Mirrors the \c getCallCost method but uses an intrinsic identifier. 146 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 147 ArrayRef<Type *> ParamTys) const; 148 149 /// \brief Estimate the cost of an intrinsic when lowered. 150 /// 151 /// Mirrors the \c getCallCost method but uses an intrinsic identifier. 152 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 153 ArrayRef<const Value *> Arguments) const; 154 155 /// \brief Estimate the cost of a given IR user when lowered. 156 /// 157 /// This can estimate the cost of either a ConstantExpr or Instruction when 158 /// lowered. It has two primary advantages over the \c getOperationCost and 159 /// \c getGEPCost above, and one significant disadvantage: it can only be 160 /// used when the IR construct has already been formed. 161 /// 162 /// The advantages are that it can inspect the SSA use graph to reason more 163 /// accurately about the cost. For example, all-constant-GEPs can often be 164 /// folded into a load or other instruction, but if they are used in some 165 /// other context they may not be folded. This routine can distinguish such 166 /// cases. 167 /// 168 /// The returned cost is defined in terms of \c TargetCostConstants, see its 169 /// comments for a detailed explanation of the cost values. 170 virtual unsigned getUserCost(const User *U) const; 171 172 /// \brief Test whether calls to a function lower to actual program function 173 /// calls. 174 /// 175 /// The idea is to test whether the program is likely to require a 'call' 176 /// instruction or equivalent in order to call the given function. 177 /// 178 /// FIXME: It's not clear that this is a good or useful query API. Client's 179 /// should probably move to simpler cost metrics using the above. 180 /// Alternatively, we could split the cost interface into distinct code-size 181 /// and execution-speed costs. This would allow modelling the core of this 182 /// query more accurately as the a call is a single small instruction, but 183 /// incurs significant execution cost. 184 virtual bool isLoweredToCall(const Function *F) const; 185 186 /// @} 187 188 /// \name Scalar Target Information 189 /// @{ 190 191 /// \brief Flags indicating the kind of support for population count. 192 /// 193 /// Compared to the SW implementation, HW support is supposed to 194 /// significantly boost the performance when the population is dense, and it 195 /// may or may not degrade performance if the population is sparse. A HW 196 /// support is considered as "Fast" if it can outperform, or is on a par 197 /// with, SW implementaion when the population is sparse; otherwise, it is 198 /// considered as "Slow". 199 enum PopcntSupportKind { 200 PSK_Software, 201 PSK_SlowHardware, 202 PSK_FastHardware 203 }; 204 205 /// isLegalAddImmediate - Return true if the specified immediate is legal 206 /// add immediate, that is the target has add instructions which can add 207 /// a register with the immediate without having to materialize the 208 /// immediate into a register. 209 virtual bool isLegalAddImmediate(int64_t Imm) const; 210 211 /// isLegalICmpImmediate - Return true if the specified immediate is legal 212 /// icmp immediate, that is the target has icmp instructions which can compare 213 /// a register against the immediate without having to materialize the 214 /// immediate into a register. 215 virtual bool isLegalICmpImmediate(int64_t Imm) const; 216 217 /// isLegalAddressingMode - Return true if the addressing mode represented by 218 /// AM is legal for this target, for a load/store of the specified type. 219 /// The type may be VoidTy, in which case only return true if the addressing 220 /// mode is legal for a load/store of any legal type. 221 /// TODO: Handle pre/postinc as well. 222 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 223 int64_t BaseOffset, bool HasBaseReg, 224 int64_t Scale) const; 225 226 /// isTruncateFree - Return true if it's free to truncate a value of 227 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in 228 /// register EAX to i16 by referencing its sub-register AX. 229 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const; 230 231 /// Is this type legal. 232 virtual bool isTypeLegal(Type *Ty) const; 233 234 /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes 235 virtual unsigned getJumpBufAlignment() const; 236 237 /// getJumpBufSize - returns the target's jmp_buf size in bytes. 238 virtual unsigned getJumpBufSize() const; 239 240 /// shouldBuildLookupTables - Return true if switches should be turned into 241 /// lookup tables for the target. 242 virtual bool shouldBuildLookupTables() const; 243 244 /// getPopcntSupport - Return hardware support for population count. 245 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const; 246 247 /// getIntImmCost - Return the expected cost of materializing the given 248 /// integer immediate of the specified type. 249 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const; 250 251 /// @} 252 253 /// \name Vector Target Information 254 /// @{ 255 256 /// \brief The various kinds of shuffle patterns for vector queries. 257 enum ShuffleKind { 258 SK_Broadcast, ///< Broadcast element 0 to all other elements. 259 SK_Reverse, ///< Reverse the order of the vector. 260 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset. 261 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset. 262 }; 263 264 /// \return The number of scalar or vector registers that the target has. 265 /// If 'Vectors' is true, it returns the number of vector registers. If it is 266 /// set to false, it returns the number of scalar registers. 267 virtual unsigned getNumberOfRegisters(bool Vector) const; 268 269 /// \return The width of the largest scalar or vector register type. 270 virtual unsigned getRegisterBitWidth(bool Vector) const; 271 272 /// \return The maximum unroll factor that the vectorizer should try to 273 /// perform for this target. This number depends on the level of parallelism 274 /// and the number of execution units in the CPU. 275 virtual unsigned getMaximumUnrollFactor() const; 276 277 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc. 278 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const; 279 280 /// \return The cost of a shuffle instruction of kind Kind and of type Tp. 281 /// The index and subtype parameters are used by the subvector insertion and 282 /// extraction shuffle kinds. 283 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0, 284 Type *SubTp = 0) const; 285 286 /// \return The expected cost of cast instructions, such as bitcast, trunc, 287 /// zext, etc. 288 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, 289 Type *Src) const; 290 291 /// \return The expected cost of control-flow related instrutctions such as 292 /// Phi, Ret, Br. 293 virtual unsigned getCFInstrCost(unsigned Opcode) const; 294 295 /// \returns The expected cost of compare and select instructions. 296 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 297 Type *CondTy = 0) const; 298 299 /// \return The expected cost of vector Insert and Extract. 300 /// Use -1 to indicate that there is no information on the index value. 301 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val, 302 unsigned Index = -1) const; 303 304 /// \return The cost of Load and Store instructions. 305 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src, 306 unsigned Alignment, 307 unsigned AddressSpace) const; 308 309 /// \returns The cost of Intrinsic instructions. 310 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 311 ArrayRef<Type *> Tys) const; 312 313 /// \returns The number of pieces into which the provided type must be 314 /// split during legalization. Zero is returned when the answer is unknown. 315 virtual unsigned getNumberOfParts(Type *Tp) const; 316 317 /// @} 318 319 /// Analysis group identification. 320 static char ID; 321}; 322 323/// \brief Create the base case instance of a pass in the TTI analysis group. 324/// 325/// This class provides the base case for the stack of TTI analyses. It doesn't 326/// delegate to anything and uses the STTI and VTTI objects passed in to 327/// satisfy the queries. 328ImmutablePass *createNoTargetTransformInfoPass(); 329 330} // End llvm namespace 331 332#endif 333