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/IR/Intrinsics.h" 26 #include "llvm/Pass.h" 27 #include "llvm/Support/DataTypes.h" 28 29 namespace llvm { 30 31 class Function; 32 class GlobalValue; 33 class Loop; 34 class Type; 35 class User; 36 class Value; 37 38 /// TargetTransformInfo - This pass provides access to the codegen 39 /// interfaces that are needed for IR-level transformations. 40 class TargetTransformInfo { 41 protected: 42 /// \brief The TTI instance one level down the stack. 43 /// 44 /// This is used to implement the default behavior all of the methods which 45 /// is to delegate up through the stack of TTIs until one can answer the 46 /// query. 47 TargetTransformInfo *PrevTTI; 48 49 /// \brief The top of the stack of TTI analyses available. 50 /// 51 /// This is a convenience routine maintained as TTI analyses become available 52 /// that complements the PrevTTI delegation chain. When one part of an 53 /// analysis pass wants to query another part of the analysis pass it can use 54 /// this to start back at the top of the stack. 55 TargetTransformInfo *TopTTI; 56 57 /// All pass subclasses must in their initializePass routine call 58 /// pushTTIStack with themselves to update the pointers tracking the previous 59 /// TTI instance in the analysis group's stack, and the top of the analysis 60 /// group's stack. 61 void pushTTIStack(Pass *P); 62 63 /// All pass subclasses must call TargetTransformInfo::getAnalysisUsage. 64 virtual void getAnalysisUsage(AnalysisUsage &AU) const; 65 66 public: 67 /// This class is intended to be subclassed by real implementations. 68 virtual ~TargetTransformInfo() = 0; 69 70 /// \name Generic Target Information 71 /// @{ 72 73 /// \brief Underlying constants for 'cost' values in this interface. 74 /// 75 /// Many APIs in this interface return a cost. This enum defines the 76 /// fundamental values that should be used to interpret (and produce) those 77 /// costs. The costs are returned as an unsigned rather than a member of this 78 /// enumeration because it is expected that the cost of one IR instruction 79 /// may have a multiplicative factor to it or otherwise won't fit directly 80 /// into the enum. Moreover, it is common to sum or average costs which works 81 /// better as simple integral values. Thus this enum only provides constants. 82 /// 83 /// Note that these costs should usually reflect the intersection of code-size 84 /// cost and execution cost. A free instruction is typically one that folds 85 /// into another instruction. For example, reg-to-reg moves can often be 86 /// skipped by renaming the registers in the CPU, but they still are encoded 87 /// and thus wouldn't be considered 'free' here. 88 enum TargetCostConstants { 89 TCC_Free = 0, ///< Expected to fold away in lowering. 90 TCC_Basic = 1, ///< The cost of a typical 'add' instruction. 91 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86. 92 }; 93 94 /// \brief Estimate the cost of a specific operation when lowered. 95 /// 96 /// Note that this is designed to work on an arbitrary synthetic opcode, and 97 /// thus work for hypothetical queries before an instruction has even been 98 /// formed. However, this does *not* work for GEPs, and must not be called 99 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as 100 /// analyzing a GEP's cost required more information. 101 /// 102 /// Typically only the result type is required, and the operand type can be 103 /// omitted. However, if the opcode is one of the cast instructions, the 104 /// operand type is required. 105 /// 106 /// The returned cost is defined in terms of \c TargetCostConstants, see its 107 /// comments for a detailed explanation of the cost values. 108 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, 109 Type *OpTy = nullptr) const; 110 111 /// \brief Estimate the cost of a GEP operation when lowered. 112 /// 113 /// The contract for this function is the same as \c getOperationCost except 114 /// that it supports an interface that provides extra information specific to 115 /// the GEP operation. 116 virtual unsigned getGEPCost(const Value *Ptr, 117 ArrayRef<const Value *> Operands) const; 118 119 /// \brief Estimate the cost of a function call when lowered. 120 /// 121 /// The contract for this is the same as \c getOperationCost except that it 122 /// supports an interface that provides extra information specific to call 123 /// instructions. 124 /// 125 /// This is the most basic query for estimating call cost: it only knows the 126 /// function type and (potentially) the number of arguments at the call site. 127 /// The latter is only interesting for varargs function types. 128 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const; 129 130 /// \brief Estimate the cost of calling a specific function when lowered. 131 /// 132 /// This overload adds the ability to reason about the particular function 133 /// being called in the event it is a library call with special lowering. 134 virtual unsigned getCallCost(const Function *F, int NumArgs = -1) const; 135 136 /// \brief Estimate the cost of calling a specific function when lowered. 137 /// 138 /// This overload allows specifying a set of candidate argument values. 139 virtual unsigned getCallCost(const Function *F, 140 ArrayRef<const Value *> Arguments) const; 141 142 /// \brief Estimate the cost of an intrinsic when lowered. 143 /// 144 /// Mirrors the \c getCallCost method but uses an intrinsic identifier. 145 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 146 ArrayRef<Type *> ParamTys) const; 147 148 /// \brief Estimate the cost of an intrinsic when lowered. 149 /// 150 /// Mirrors the \c getCallCost method but uses an intrinsic identifier. 151 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy, 152 ArrayRef<const Value *> Arguments) const; 153 154 /// \brief Estimate the cost of a given IR user when lowered. 155 /// 156 /// This can estimate the cost of either a ConstantExpr or Instruction when 157 /// lowered. It has two primary advantages over the \c getOperationCost and 158 /// \c getGEPCost above, and one significant disadvantage: it can only be 159 /// used when the IR construct has already been formed. 160 /// 161 /// The advantages are that it can inspect the SSA use graph to reason more 162 /// accurately about the cost. For example, all-constant-GEPs can often be 163 /// folded into a load or other instruction, but if they are used in some 164 /// other context they may not be folded. This routine can distinguish such 165 /// cases. 166 /// 167 /// The returned cost is defined in terms of \c TargetCostConstants, see its 168 /// comments for a detailed explanation of the cost values. 169 virtual unsigned getUserCost(const User *U) const; 170 171 /// \brief hasBranchDivergence - Return true if branch divergence exists. 172 /// Branch divergence has a significantly negative impact on GPU performance 173 /// when threads in the same wavefront take different paths due to conditional 174 /// branches. 175 virtual bool hasBranchDivergence() const; 176 177 /// \brief Test whether calls to a function lower to actual program function 178 /// calls. 179 /// 180 /// The idea is to test whether the program is likely to require a 'call' 181 /// instruction or equivalent in order to call the given function. 182 /// 183 /// FIXME: It's not clear that this is a good or useful query API. Client's 184 /// should probably move to simpler cost metrics using the above. 185 /// Alternatively, we could split the cost interface into distinct code-size 186 /// and execution-speed costs. This would allow modelling the core of this 187 /// query more accurately as a call is a single small instruction, but 188 /// incurs significant execution cost. 189 virtual bool isLoweredToCall(const Function *F) const; 190 191 /// Parameters that control the generic loop unrolling transformation. 192 struct UnrollingPreferences { 193 /// The cost threshold for the unrolled loop, compared to 194 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body. 195 /// The unrolling factor is set such that the unrolled loop body does not 196 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost 197 /// restriction. 198 unsigned Threshold; 199 /// The cost threshold for the unrolled loop when optimizing for size (set 200 /// to UINT_MAX to disable). 201 unsigned OptSizeThreshold; 202 /// The cost threshold for the unrolled loop, like Threshold, but used 203 /// for partial/runtime unrolling (set to UINT_MAX to disable). 204 unsigned PartialThreshold; 205 /// The cost threshold for the unrolled loop when optimizing for size, like 206 /// OptSizeThreshold, but used for partial/runtime unrolling (set to UINT_MAX 207 /// to disable). 208 unsigned PartialOptSizeThreshold; 209 /// A forced unrolling factor (the number of concatenated bodies of the 210 /// original loop in the unrolled loop body). When set to 0, the unrolling 211 /// transformation will select an unrolling factor based on the current cost 212 /// threshold and other factors. 213 unsigned Count; 214 // Set the maximum unrolling factor. The unrolling factor may be selected 215 // using the appropriate cost threshold, but may not exceed this number 216 // (set to UINT_MAX to disable). This does not apply in cases where the 217 // loop is being fully unrolled. 218 unsigned MaxCount; 219 /// Allow partial unrolling (unrolling of loops to expand the size of the 220 /// loop body, not only to eliminate small constant-trip-count loops). 221 bool Partial; 222 /// Allow runtime unrolling (unrolling of loops to expand the size of the 223 /// loop body even when the number of loop iterations is not known at compile 224 /// time). 225 bool Runtime; 226 }; 227 228 /// \brief Get target-customized preferences for the generic loop unrolling 229 /// transformation. The caller will initialize UP with the current 230 /// target-independent defaults. 231 virtual void getUnrollingPreferences(const Function *F, Loop *L, 232 UnrollingPreferences &UP) const; 233 234 /// @} 235 236 /// \name Scalar Target Information 237 /// @{ 238 239 /// \brief Flags indicating the kind of support for population count. 240 /// 241 /// Compared to the SW implementation, HW support is supposed to 242 /// significantly boost the performance when the population is dense, and it 243 /// may or may not degrade performance if the population is sparse. A HW 244 /// support is considered as "Fast" if it can outperform, or is on a par 245 /// with, SW implementation when the population is sparse; otherwise, it is 246 /// considered as "Slow". 247 enum PopcntSupportKind { 248 PSK_Software, 249 PSK_SlowHardware, 250 PSK_FastHardware 251 }; 252 253 /// \brief Return true if the specified immediate is legal add immediate, that 254 /// is the target has add instructions which can add a register with the 255 /// immediate without having to materialize the immediate into a register. 256 virtual bool isLegalAddImmediate(int64_t Imm) const; 257 258 /// \brief Return true if the specified immediate is legal icmp immediate, 259 /// that is the target has icmp instructions which can compare a register 260 /// against the immediate without having to materialize the immediate into a 261 /// register. 262 virtual bool isLegalICmpImmediate(int64_t Imm) const; 263 264 /// \brief Return true if the addressing mode represented by AM is legal for 265 /// this target, for a load/store of the specified type. 266 /// The type may be VoidTy, in which case only return true if the addressing 267 /// mode is legal for a load/store of any legal type. 268 /// TODO: Handle pre/postinc as well. 269 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, 270 int64_t BaseOffset, bool HasBaseReg, 271 int64_t Scale) const; 272 273 /// \brief Return true if the target works with masked instruction 274 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements. 275 /// AVX-512 architecture will also allow masks for non-consecutive memory 276 /// accesses. 277 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) const; 278 virtual bool isLegalMaskedLoad (Type *DataType, int Consecutive) const; 279 280 /// \brief Return the cost of the scaling factor used in the addressing 281 /// mode represented by AM for this target, for a load/store 282 /// of the specified type. 283 /// If the AM is supported, the return value must be >= 0. 284 /// If the AM is not supported, it returns a negative value. 285 /// TODO: Handle pre/postinc as well. 286 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, 287 int64_t BaseOffset, bool HasBaseReg, 288 int64_t Scale) const; 289 290 /// \brief Return true if it's free to truncate a value of type Ty1 to type 291 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16 292 /// by referencing its sub-register AX. 293 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const; 294 295 /// \brief Return true if this type is legal. 296 virtual bool isTypeLegal(Type *Ty) const; 297 298 /// \brief Returns the target's jmp_buf alignment in bytes. 299 virtual unsigned getJumpBufAlignment() const; 300 301 /// \brief Returns the target's jmp_buf size in bytes. 302 virtual unsigned getJumpBufSize() const; 303 304 /// \brief Return true if switches should be turned into lookup tables for the 305 /// target. 306 virtual bool shouldBuildLookupTables() const; 307 308 /// \brief Return hardware support for population count. 309 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const; 310 311 /// \brief Return true if the hardware has a fast square-root instruction. 312 virtual bool haveFastSqrt(Type *Ty) const; 313 314 /// \brief Return the expected cost of materializing for the given integer 315 /// immediate of the specified type. 316 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const; 317 318 /// \brief Return the expected cost of materialization for the given integer 319 /// immediate of the specified type for a given instruction. The cost can be 320 /// zero if the immediate can be folded into the specified instruction. 321 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm, 322 Type *Ty) const; 323 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, 324 const APInt &Imm, Type *Ty) const; 325 /// @} 326 327 /// \name Vector Target Information 328 /// @{ 329 330 /// \brief The various kinds of shuffle patterns for vector queries. 331 enum ShuffleKind { 332 SK_Broadcast, ///< Broadcast element 0 to all other elements. 333 SK_Reverse, ///< Reverse the order of the vector. 334 SK_Alternate, ///< Choose alternate elements from vector. 335 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset. 336 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset. 337 }; 338 339 /// \brief Additional information about an operand's possible values. 340 enum OperandValueKind { 341 OK_AnyValue, // Operand can have any value. 342 OK_UniformValue, // Operand is uniform (splat of a value). 343 OK_UniformConstantValue, // Operand is uniform constant. 344 OK_NonUniformConstantValue // Operand is a non uniform constant value. 345 }; 346 347 /// \brief Additional properties of an operand's values. 348 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 }; 349 350 /// \return The number of scalar or vector registers that the target has. 351 /// If 'Vectors' is true, it returns the number of vector registers. If it is 352 /// set to false, it returns the number of scalar registers. 353 virtual unsigned getNumberOfRegisters(bool Vector) const; 354 355 /// \return The width of the largest scalar or vector register type. 356 virtual unsigned getRegisterBitWidth(bool Vector) const; 357 358 /// \return The maximum interleave factor that any transform should try to 359 /// perform for this target. This number depends on the level of parallelism 360 /// and the number of execution units in the CPU. 361 virtual unsigned getMaxInterleaveFactor() const; 362 363 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc. 364 virtual unsigned 365 getArithmeticInstrCost(unsigned Opcode, Type *Ty, 366 OperandValueKind Opd1Info = OK_AnyValue, 367 OperandValueKind Opd2Info = OK_AnyValue, 368 OperandValueProperties Opd1PropInfo = OP_None, 369 OperandValueProperties Opd2PropInfo = OP_None) const; 370 371 /// \return The cost of a shuffle instruction of kind Kind and of type Tp. 372 /// The index and subtype parameters are used by the subvector insertion and 373 /// extraction shuffle kinds. 374 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0, 375 Type *SubTp = nullptr) const; 376 377 /// \return The expected cost of cast instructions, such as bitcast, trunc, 378 /// zext, etc. 379 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, 380 Type *Src) const; 381 382 /// \return The expected cost of control-flow related instructions such as 383 /// Phi, Ret, Br. 384 virtual unsigned getCFInstrCost(unsigned Opcode) const; 385 386 /// \returns The expected cost of compare and select instructions. 387 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy, 388 Type *CondTy = nullptr) const; 389 390 /// \return The expected cost of vector Insert and Extract. 391 /// Use -1 to indicate that there is no information on the index value. 392 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val, 393 unsigned Index = -1) const; 394 395 /// \return The cost of Load and Store instructions. 396 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src, 397 unsigned Alignment, 398 unsigned AddressSpace) const; 399 400 /// \brief Calculate the cost of performing a vector reduction. 401 /// 402 /// This is the cost of reducing the vector value of type \p Ty to a scalar 403 /// value using the operation denoted by \p Opcode. The form of the reduction 404 /// can either be a pairwise reduction or a reduction that splits the vector 405 /// at every reduction level. 406 /// 407 /// Pairwise: 408 /// (v0, v1, v2, v3) 409 /// ((v0+v1), (v2, v3), undef, undef) 410 /// Split: 411 /// (v0, v1, v2, v3) 412 /// ((v0+v2), (v1+v3), undef, undef) 413 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty, 414 bool IsPairwiseForm) const; 415 416 /// \returns The cost of Intrinsic instructions. 417 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, 418 ArrayRef<Type *> Tys) const; 419 420 /// \returns The number of pieces into which the provided type must be 421 /// split during legalization. Zero is returned when the answer is unknown. 422 virtual unsigned getNumberOfParts(Type *Tp) const; 423 424 /// \returns The cost of the address computation. For most targets this can be 425 /// merged into the instruction indexing mode. Some targets might want to 426 /// distinguish between address computation for memory operations on vector 427 /// types and scalar types. Such targets should override this function. 428 /// The 'IsComplex' parameter is a hint that the address computation is likely 429 /// to involve multiple instructions and as such unlikely to be merged into 430 /// the address indexing mode. 431 virtual unsigned getAddressComputationCost(Type *Ty, 432 bool IsComplex = false) const; 433 434 /// \returns The cost, if any, of keeping values of the given types alive 435 /// over a callsite. 436 /// 437 /// Some types may require the use of register classes that do not have 438 /// any callee-saved registers, so would require a spill and fill. 439 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type*> Tys) const; 440 441 /// @} 442 443 /// Analysis group identification. 444 static char ID; 445 }; 446 447 /// \brief Create the base case instance of a pass in the TTI analysis group. 448 /// 449 /// This class provides the base case for the stack of TTI analyzes. It doesn't 450 /// delegate to anything and uses the STTI and VTTI objects passed in to 451 /// satisfy the queries. 452 ImmutablePass *createNoTargetTransformInfoPass(); 453 454 } // End llvm namespace 455 456 #endif 457