1 //===- TargetTransformInfo.h ------------------------------------*- C++ -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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
15 /// 3. Codegen-level implementation which uses target-specific hooks.
17 /// This file defines #2, which is the interface that IR-level transformations
18 /// use for querying the codegen.
20 //===----------------------------------------------------------------------===//
22 #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
23 #define LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
25 #include "llvm/IR/Intrinsics.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Support/DataTypes.h"
39 /// \brief Information about a load/store intrinsic defined by the target.
40 struct MemIntrinsicInfo {
42 : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
43 NumMemRefs(0), PtrVal(nullptr) {}
47 // Same Id is set by the target for corresponding load/store intrinsics.
48 unsigned short MatchingId;
53 /// \brief This pass provides access to the codegen interfaces that are needed
54 /// for IR-level transformations.
55 class TargetTransformInfo {
57 /// \brief Construct a TTI object using a type implementing the \c Concept
60 /// This is used by targets to construct a TTI wrapping their target-specific
61 /// implementaion that encodes appropriate costs for their target.
62 template <typename T> TargetTransformInfo(T Impl);
64 /// \brief Construct a baseline TTI object using a minimal implementation of
65 /// the \c Concept API below.
67 /// The TTI implementation will reflect the information in the DataLayout
68 /// provided if non-null.
69 explicit TargetTransformInfo(const DataLayout *DL);
71 // Provide move semantics.
72 TargetTransformInfo(TargetTransformInfo &&Arg);
73 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
75 // We need to define the destructor out-of-line to define our sub-classes
77 ~TargetTransformInfo();
79 /// \name Generic Target Information
82 /// \brief Underlying constants for 'cost' values in this interface.
84 /// Many APIs in this interface return a cost. This enum defines the
85 /// fundamental values that should be used to interpret (and produce) those
86 /// costs. The costs are returned as an unsigned rather than a member of this
87 /// enumeration because it is expected that the cost of one IR instruction
88 /// may have a multiplicative factor to it or otherwise won't fit directly
89 /// into the enum. Moreover, it is common to sum or average costs which works
90 /// better as simple integral values. Thus this enum only provides constants.
92 /// Note that these costs should usually reflect the intersection of code-size
93 /// cost and execution cost. A free instruction is typically one that folds
94 /// into another instruction. For example, reg-to-reg moves can often be
95 /// skipped by renaming the registers in the CPU, but they still are encoded
96 /// and thus wouldn't be considered 'free' here.
97 enum TargetCostConstants {
98 TCC_Free = 0, ///< Expected to fold away in lowering.
99 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
100 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
103 /// \brief Estimate the cost of a specific operation when lowered.
105 /// Note that this is designed to work on an arbitrary synthetic opcode, and
106 /// thus work for hypothetical queries before an instruction has even been
107 /// formed. However, this does *not* work for GEPs, and must not be called
108 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
109 /// analyzing a GEP's cost required more information.
111 /// Typically only the result type is required, and the operand type can be
112 /// omitted. However, if the opcode is one of the cast instructions, the
113 /// operand type is required.
115 /// The returned cost is defined in terms of \c TargetCostConstants, see its
116 /// comments for a detailed explanation of the cost values.
117 unsigned getOperationCost(unsigned Opcode, Type *Ty,
118 Type *OpTy = nullptr) const;
120 /// \brief Estimate the cost of a GEP operation when lowered.
122 /// The contract for this function is the same as \c getOperationCost except
123 /// that it supports an interface that provides extra information specific to
124 /// the GEP operation.
125 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
127 /// \brief Estimate the cost of a function call when lowered.
129 /// The contract for this is the same as \c getOperationCost except that it
130 /// supports an interface that provides extra information specific to call
133 /// This is the most basic query for estimating call cost: it only knows the
134 /// function type and (potentially) the number of arguments at the call site.
135 /// The latter is only interesting for varargs function types.
136 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
138 /// \brief Estimate the cost of calling a specific function when lowered.
140 /// This overload adds the ability to reason about the particular function
141 /// being called in the event it is a library call with special lowering.
142 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
144 /// \brief Estimate the cost of calling a specific function when lowered.
146 /// This overload allows specifying a set of candidate argument values.
147 unsigned getCallCost(const Function *F,
148 ArrayRef<const Value *> Arguments) const;
150 /// \brief Estimate the cost of an intrinsic when lowered.
152 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
153 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
154 ArrayRef<Type *> ParamTys) const;
156 /// \brief Estimate the cost of an intrinsic when lowered.
158 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
159 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
160 ArrayRef<const Value *> Arguments) const;
162 /// \brief Estimate the cost of a given IR user when lowered.
164 /// This can estimate the cost of either a ConstantExpr or Instruction when
165 /// lowered. It has two primary advantages over the \c getOperationCost and
166 /// \c getGEPCost above, and one significant disadvantage: it can only be
167 /// used when the IR construct has already been formed.
169 /// The advantages are that it can inspect the SSA use graph to reason more
170 /// accurately about the cost. For example, all-constant-GEPs can often be
171 /// folded into a load or other instruction, but if they are used in some
172 /// other context they may not be folded. This routine can distinguish such
175 /// The returned cost is defined in terms of \c TargetCostConstants, see its
176 /// comments for a detailed explanation of the cost values.
177 unsigned getUserCost(const User *U) const;
179 /// \brief hasBranchDivergence - Return true if branch divergence exists.
180 /// Branch divergence has a significantly negative impact on GPU performance
181 /// when threads in the same wavefront take different paths due to conditional
183 bool hasBranchDivergence() const;
185 /// \brief Test whether calls to a function lower to actual program function
188 /// The idea is to test whether the program is likely to require a 'call'
189 /// instruction or equivalent in order to call the given function.
191 /// FIXME: It's not clear that this is a good or useful query API. Client's
192 /// should probably move to simpler cost metrics using the above.
193 /// Alternatively, we could split the cost interface into distinct code-size
194 /// and execution-speed costs. This would allow modelling the core of this
195 /// query more accurately as a call is a single small instruction, but
196 /// incurs significant execution cost.
197 bool isLoweredToCall(const Function *F) const;
199 /// Parameters that control the generic loop unrolling transformation.
200 struct UnrollingPreferences {
201 /// The cost threshold for the unrolled loop, compared to
202 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
203 /// The unrolling factor is set such that the unrolled loop body does not
204 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
207 /// The cost threshold for the unrolled loop when optimizing for size (set
208 /// to UINT_MAX to disable).
209 unsigned OptSizeThreshold;
210 /// The cost threshold for the unrolled loop, like Threshold, but used
211 /// for partial/runtime unrolling (set to UINT_MAX to disable).
212 unsigned PartialThreshold;
213 /// The cost threshold for the unrolled loop when optimizing for size, like
214 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
215 /// UINT_MAX to disable).
216 unsigned PartialOptSizeThreshold;
217 /// A forced unrolling factor (the number of concatenated bodies of the
218 /// original loop in the unrolled loop body). When set to 0, the unrolling
219 /// transformation will select an unrolling factor based on the current cost
220 /// threshold and other factors.
222 // Set the maximum unrolling factor. The unrolling factor may be selected
223 // using the appropriate cost threshold, but may not exceed this number
224 // (set to UINT_MAX to disable). This does not apply in cases where the
225 // loop is being fully unrolled.
227 /// Allow partial unrolling (unrolling of loops to expand the size of the
228 /// loop body, not only to eliminate small constant-trip-count loops).
230 /// Allow runtime unrolling (unrolling of loops to expand the size of the
231 /// loop body even when the number of loop iterations is not known at
236 /// \brief Get target-customized preferences for the generic loop unrolling
237 /// transformation. The caller will initialize UP with the current
238 /// target-independent defaults.
239 void getUnrollingPreferences(const Function *F, Loop *L,
240 UnrollingPreferences &UP) const;
244 /// \name Scalar Target Information
247 /// \brief Flags indicating the kind of support for population count.
249 /// Compared to the SW implementation, HW support is supposed to
250 /// significantly boost the performance when the population is dense, and it
251 /// may or may not degrade performance if the population is sparse. A HW
252 /// support is considered as "Fast" if it can outperform, or is on a par
253 /// with, SW implementation when the population is sparse; otherwise, it is
254 /// considered as "Slow".
255 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
257 /// \brief Return true if the specified immediate is legal add immediate, that
258 /// is the target has add instructions which can add a register with the
259 /// immediate without having to materialize the immediate into a register.
260 bool isLegalAddImmediate(int64_t Imm) const;
262 /// \brief Return true if the specified immediate is legal icmp immediate,
263 /// that is the target has icmp instructions which can compare a register
264 /// against the immediate without having to materialize the immediate into a
266 bool isLegalICmpImmediate(int64_t Imm) const;
268 /// \brief Return true if the addressing mode represented by AM is legal for
269 /// this target, for a load/store of the specified type.
270 /// The type may be VoidTy, in which case only return true if the addressing
271 /// mode is legal for a load/store of any legal type.
272 /// TODO: Handle pre/postinc as well.
273 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
274 bool HasBaseReg, int64_t Scale) const;
276 /// \brief Return true if the target works with masked instruction
277 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
278 /// AVX-512 architecture will also allow masks for non-consecutive memory
280 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
281 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
283 /// \brief Return the cost of the scaling factor used in the addressing
284 /// mode represented by AM for this target, for a load/store
285 /// of the specified type.
286 /// If the AM is supported, the return value must be >= 0.
287 /// If the AM is not supported, it returns a negative value.
288 /// TODO: Handle pre/postinc as well.
289 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
290 bool HasBaseReg, int64_t Scale) const;
292 /// \brief Return true if it's free to truncate a value of type Ty1 to type
293 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
294 /// by referencing its sub-register AX.
295 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
297 /// \brief Return true if this type is legal.
298 bool isTypeLegal(Type *Ty) const;
300 /// \brief Returns the target's jmp_buf alignment in bytes.
301 unsigned getJumpBufAlignment() const;
303 /// \brief Returns the target's jmp_buf size in bytes.
304 unsigned getJumpBufSize() const;
306 /// \brief Return true if switches should be turned into lookup tables for the
308 bool shouldBuildLookupTables() const;
310 /// \brief Return hardware support for population count.
311 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
313 /// \brief Return true if the hardware has a fast square-root instruction.
314 bool haveFastSqrt(Type *Ty) const;
316 /// \brief Return the expected cost of materializing for the given integer
317 /// immediate of the specified type.
318 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
320 /// \brief Return the expected cost of materialization for the given integer
321 /// immediate of the specified type for a given instruction. The cost can be
322 /// zero if the immediate can be folded into the specified instruction.
323 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
325 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
329 /// \name Vector Target Information
332 /// \brief The various kinds of shuffle patterns for vector queries.
334 SK_Broadcast, ///< Broadcast element 0 to all other elements.
335 SK_Reverse, ///< Reverse the order of the vector.
336 SK_Alternate, ///< Choose alternate elements from vector.
337 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
338 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
341 /// \brief Additional information about an operand's possible values.
342 enum OperandValueKind {
343 OK_AnyValue, // Operand can have any value.
344 OK_UniformValue, // Operand is uniform (splat of a value).
345 OK_UniformConstantValue, // Operand is uniform constant.
346 OK_NonUniformConstantValue // Operand is a non uniform constant value.
349 /// \brief Additional properties of an operand's values.
350 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
352 /// \return The number of scalar or vector registers that the target has.
353 /// If 'Vectors' is true, it returns the number of vector registers. If it is
354 /// set to false, it returns the number of scalar registers.
355 unsigned getNumberOfRegisters(bool Vector) const;
357 /// \return The width of the largest scalar or vector register type.
358 unsigned getRegisterBitWidth(bool Vector) const;
360 /// \return The maximum interleave factor that any transform should try to
361 /// perform for this target. This number depends on the level of parallelism
362 /// and the number of execution units in the CPU.
363 unsigned getMaxInterleaveFactor() const;
365 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
367 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
368 OperandValueKind Opd1Info = OK_AnyValue,
369 OperandValueKind Opd2Info = OK_AnyValue,
370 OperandValueProperties Opd1PropInfo = OP_None,
371 OperandValueProperties Opd2PropInfo = OP_None) const;
373 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
374 /// The index and subtype parameters are used by the subvector insertion and
375 /// extraction shuffle kinds.
376 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
377 Type *SubTp = nullptr) const;
379 /// \return The expected cost of cast instructions, such as bitcast, trunc,
381 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
383 /// \return The expected cost of control-flow related instructions such as
385 unsigned getCFInstrCost(unsigned Opcode) const;
387 /// \returns The expected cost of compare and select instructions.
388 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
389 Type *CondTy = nullptr) const;
391 /// \return The expected cost of vector Insert and Extract.
392 /// Use -1 to indicate that there is no information on the index value.
393 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
394 unsigned Index = -1) const;
396 /// \return The cost of Load and Store instructions.
397 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
398 unsigned AddressSpace) const;
400 /// \return The cost of masked Load and Store instructions.
401 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
402 unsigned AddressSpace) const;
404 /// \brief Calculate the cost of performing a vector reduction.
406 /// This is the cost of reducing the vector value of type \p Ty to a scalar
407 /// value using the operation denoted by \p Opcode. The form of the reduction
408 /// can either be a pairwise reduction or a reduction that splits the vector
409 /// at every reduction level.
413 /// ((v0+v1), (v2, v3), undef, undef)
416 /// ((v0+v2), (v1+v3), undef, undef)
417 unsigned getReductionCost(unsigned Opcode, Type *Ty,
418 bool IsPairwiseForm) const;
420 /// \returns The cost of Intrinsic instructions.
421 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
422 ArrayRef<Type *> Tys) const;
424 /// \returns The number of pieces into which the provided type must be
425 /// split during legalization. Zero is returned when the answer is unknown.
426 unsigned getNumberOfParts(Type *Tp) const;
428 /// \returns The cost of the address computation. For most targets this can be
429 /// merged into the instruction indexing mode. Some targets might want to
430 /// distinguish between address computation for memory operations on vector
431 /// types and scalar types. Such targets should override this function.
432 /// The 'IsComplex' parameter is a hint that the address computation is likely
433 /// to involve multiple instructions and as such unlikely to be merged into
434 /// the address indexing mode.
435 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
437 /// \returns The cost, if any, of keeping values of the given types alive
440 /// Some types may require the use of register classes that do not have
441 /// any callee-saved registers, so would require a spill and fill.
442 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
444 /// \returns True if the intrinsic is a supported memory intrinsic. Info
445 /// will contain additional information - whether the intrinsic may write
446 /// or read to memory, volatility and the pointer. Info is undefined
447 /// if false is returned.
448 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
450 /// \returns A value which is the result of the given memory intrinsic. New
451 /// instructions may be created to extract the result from the given intrinsic
452 /// memory operation. Returns nullptr if the target cannot create a result
453 /// from the given intrinsic.
454 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
455 Type *ExpectedType) const;
460 /// \brief The abstract base class used to type erase specific TTI
464 /// \brief The template model for the base class which wraps a concrete
465 /// implementation in a type erased interface.
466 template <typename T> class Model;
468 std::unique_ptr<Concept> TTIImpl;
471 class TargetTransformInfo::Concept {
473 virtual ~Concept() = 0;
475 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
476 virtual unsigned getGEPCost(const Value *Ptr,
477 ArrayRef<const Value *> Operands) = 0;
478 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
479 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
480 virtual unsigned getCallCost(const Function *F,
481 ArrayRef<const Value *> Arguments) = 0;
482 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
483 ArrayRef<Type *> ParamTys) = 0;
484 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
485 ArrayRef<const Value *> Arguments) = 0;
486 virtual unsigned getUserCost(const User *U) = 0;
487 virtual bool hasBranchDivergence() = 0;
488 virtual bool isLoweredToCall(const Function *F) = 0;
489 virtual void getUnrollingPreferences(const Function *F, Loop *L,
490 UnrollingPreferences &UP) = 0;
491 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
492 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
493 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
494 int64_t BaseOffset, bool HasBaseReg,
496 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
497 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
498 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
499 int64_t BaseOffset, bool HasBaseReg,
501 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
502 virtual bool isTypeLegal(Type *Ty) = 0;
503 virtual unsigned getJumpBufAlignment() = 0;
504 virtual unsigned getJumpBufSize() = 0;
505 virtual bool shouldBuildLookupTables() = 0;
506 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
507 virtual bool haveFastSqrt(Type *Ty) = 0;
508 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
509 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
511 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
512 const APInt &Imm, Type *Ty) = 0;
513 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
514 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
515 virtual unsigned getMaxInterleaveFactor() = 0;
517 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
518 OperandValueKind Opd2Info,
519 OperandValueProperties Opd1PropInfo,
520 OperandValueProperties Opd2PropInfo) = 0;
521 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
523 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
524 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
525 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
527 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
529 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
531 unsigned AddressSpace) = 0;
532 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
534 unsigned AddressSpace) = 0;
535 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
536 bool IsPairwiseForm) = 0;
537 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
538 ArrayRef<Type *> Tys) = 0;
539 virtual unsigned getNumberOfParts(Type *Tp) = 0;
540 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
541 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
542 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
543 MemIntrinsicInfo &Info) = 0;
544 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
545 Type *ExpectedType) = 0;
548 template <typename T>
549 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
553 Model(T Impl) : Impl(std::move(Impl)) {}
556 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
557 return Impl.getOperationCost(Opcode, Ty, OpTy);
559 unsigned getGEPCost(const Value *Ptr,
560 ArrayRef<const Value *> Operands) override {
561 return Impl.getGEPCost(Ptr, Operands);
563 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
564 return Impl.getCallCost(FTy, NumArgs);
566 unsigned getCallCost(const Function *F, int NumArgs) override {
567 return Impl.getCallCost(F, NumArgs);
569 unsigned getCallCost(const Function *F,
570 ArrayRef<const Value *> Arguments) override {
571 return Impl.getCallCost(F, Arguments);
573 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
574 ArrayRef<Type *> ParamTys) override {
575 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
577 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
578 ArrayRef<const Value *> Arguments) override {
579 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
581 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
582 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
583 bool isLoweredToCall(const Function *F) override {
584 return Impl.isLoweredToCall(F);
586 void getUnrollingPreferences(const Function *F, Loop *L,
587 UnrollingPreferences &UP) override {
588 return Impl.getUnrollingPreferences(F, L, UP);
590 bool isLegalAddImmediate(int64_t Imm) override {
591 return Impl.isLegalAddImmediate(Imm);
593 bool isLegalICmpImmediate(int64_t Imm) override {
594 return Impl.isLegalICmpImmediate(Imm);
596 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
597 bool HasBaseReg, int64_t Scale) override {
598 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
601 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
602 return Impl.isLegalMaskedStore(DataType, Consecutive);
604 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
605 return Impl.isLegalMaskedLoad(DataType, Consecutive);
607 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
608 bool HasBaseReg, int64_t Scale) override {
609 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
611 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
612 return Impl.isTruncateFree(Ty1, Ty2);
614 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
615 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
616 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
617 bool shouldBuildLookupTables() override {
618 return Impl.shouldBuildLookupTables();
620 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
621 return Impl.getPopcntSupport(IntTyWidthInBit);
623 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
624 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
625 return Impl.getIntImmCost(Imm, Ty);
627 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
629 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
631 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
633 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
635 unsigned getNumberOfRegisters(bool Vector) override {
636 return Impl.getNumberOfRegisters(Vector);
638 unsigned getRegisterBitWidth(bool Vector) override {
639 return Impl.getRegisterBitWidth(Vector);
641 unsigned getMaxInterleaveFactor() override {
642 return Impl.getMaxInterleaveFactor();
645 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
646 OperandValueKind Opd2Info,
647 OperandValueProperties Opd1PropInfo,
648 OperandValueProperties Opd2PropInfo) override {
649 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
650 Opd1PropInfo, Opd2PropInfo);
652 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
653 Type *SubTp) override {
654 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
656 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
657 return Impl.getCastInstrCost(Opcode, Dst, Src);
659 unsigned getCFInstrCost(unsigned Opcode) override {
660 return Impl.getCFInstrCost(Opcode);
662 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
663 Type *CondTy) override {
664 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
666 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
667 unsigned Index) override {
668 return Impl.getVectorInstrCost(Opcode, Val, Index);
670 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
671 unsigned AddressSpace) override {
672 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
674 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
675 unsigned AddressSpace) override {
676 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
678 unsigned getReductionCost(unsigned Opcode, Type *Ty,
679 bool IsPairwiseForm) override {
680 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
682 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
683 ArrayRef<Type *> Tys) override {
684 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
686 unsigned getNumberOfParts(Type *Tp) override {
687 return Impl.getNumberOfParts(Tp);
689 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
690 return Impl.getAddressComputationCost(Ty, IsComplex);
692 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
693 return Impl.getCostOfKeepingLiveOverCall(Tys);
695 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
696 MemIntrinsicInfo &Info) override {
697 return Impl.getTgtMemIntrinsic(Inst, Info);
699 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
700 Type *ExpectedType) override {
701 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
705 template <typename T>
706 TargetTransformInfo::TargetTransformInfo(T Impl)
707 : TTIImpl(new Model<T>(Impl)) {}
709 /// \brief Wrapper pass for TargetTransformInfo.
711 /// This pass can be constructed from a TTI object which it stores internally
712 /// and is queried by passes.
713 class TargetTransformInfoWrapperPass : public ImmutablePass {
714 TargetTransformInfo TTI;
716 virtual void anchor();
721 /// \brief We must provide a default constructor for the pass but it should
724 /// Use the constructor below or call one of the creation routines.
725 TargetTransformInfoWrapperPass();
727 explicit TargetTransformInfoWrapperPass(TargetTransformInfo TTI);
729 TargetTransformInfo &getTTI() { return TTI; }
730 const TargetTransformInfo &getTTI() const { return TTI; }
733 /// \brief Create an analysis pass wrapper around a TTI object.
735 /// This analysis pass just holds the TTI instance and makes it available to
737 ImmutablePass *createTargetTransformInfoWrapperPass(TargetTransformInfo TTI);
739 } // End llvm namespace