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 // Provide move semantics.
65 TargetTransformInfo(TargetTransformInfo &&Arg);
66 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
68 // We need to define the destructor out-of-line to define our sub-classes
70 ~TargetTransformInfo();
72 /// \name Generic Target Information
75 /// \brief Underlying constants for 'cost' values in this interface.
77 /// Many APIs in this interface return a cost. This enum defines the
78 /// fundamental values that should be used to interpret (and produce) those
79 /// costs. The costs are returned as an unsigned rather than a member of this
80 /// enumeration because it is expected that the cost of one IR instruction
81 /// may have a multiplicative factor to it or otherwise won't fit directly
82 /// into the enum. Moreover, it is common to sum or average costs which works
83 /// better as simple integral values. Thus this enum only provides constants.
85 /// Note that these costs should usually reflect the intersection of code-size
86 /// cost and execution cost. A free instruction is typically one that folds
87 /// into another instruction. For example, reg-to-reg moves can often be
88 /// skipped by renaming the registers in the CPU, but they still are encoded
89 /// and thus wouldn't be considered 'free' here.
90 enum TargetCostConstants {
91 TCC_Free = 0, ///< Expected to fold away in lowering.
92 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
93 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
96 /// \brief Estimate the cost of a specific operation when lowered.
98 /// Note that this is designed to work on an arbitrary synthetic opcode, and
99 /// thus work for hypothetical queries before an instruction has even been
100 /// formed. However, this does *not* work for GEPs, and must not be called
101 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
102 /// analyzing a GEP's cost required more information.
104 /// Typically only the result type is required, and the operand type can be
105 /// omitted. However, if the opcode is one of the cast instructions, the
106 /// operand type is required.
108 /// The returned cost is defined in terms of \c TargetCostConstants, see its
109 /// comments for a detailed explanation of the cost values.
110 unsigned getOperationCost(unsigned Opcode, Type *Ty,
111 Type *OpTy = nullptr) const;
113 /// \brief Estimate the cost of a GEP operation when lowered.
115 /// The contract for this function is the same as \c getOperationCost except
116 /// that it supports an interface that provides extra information specific to
117 /// the GEP operation.
118 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
120 /// \brief Estimate the cost of a function call when lowered.
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
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 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
131 /// \brief Estimate the cost of calling a specific function when lowered.
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 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
137 /// \brief Estimate the cost of calling a specific function when lowered.
139 /// This overload allows specifying a set of candidate argument values.
140 unsigned getCallCost(const Function *F,
141 ArrayRef<const Value *> Arguments) const;
143 /// \brief Estimate the cost of an intrinsic when lowered.
145 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
146 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
147 ArrayRef<Type *> ParamTys) const;
149 /// \brief Estimate the cost of an intrinsic when lowered.
151 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
152 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
153 ArrayRef<const Value *> Arguments) const;
155 /// \brief Estimate the cost of a given IR user when lowered.
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.
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
168 /// The returned cost is defined in terms of \c TargetCostConstants, see its
169 /// comments for a detailed explanation of the cost values.
170 unsigned getUserCost(const User *U) const;
172 /// \brief hasBranchDivergence - Return true if branch divergence exists.
173 /// Branch divergence has a significantly negative impact on GPU performance
174 /// when threads in the same wavefront take different paths due to conditional
176 bool hasBranchDivergence() const;
178 /// \brief Test whether calls to a function lower to actual program function
181 /// The idea is to test whether the program is likely to require a 'call'
182 /// instruction or equivalent in order to call the given function.
184 /// FIXME: It's not clear that this is a good or useful query API. Client's
185 /// should probably move to simpler cost metrics using the above.
186 /// Alternatively, we could split the cost interface into distinct code-size
187 /// and execution-speed costs. This would allow modelling the core of this
188 /// query more accurately as a call is a single small instruction, but
189 /// incurs significant execution cost.
190 bool isLoweredToCall(const Function *F) const;
192 /// Parameters that control the generic loop unrolling transformation.
193 struct UnrollingPreferences {
194 /// The cost threshold for the unrolled loop, compared to
195 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
196 /// The unrolling factor is set such that the unrolled loop body does not
197 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
200 /// The cost threshold for the unrolled loop when optimizing for size (set
201 /// to UINT_MAX to disable).
202 unsigned OptSizeThreshold;
203 /// The cost threshold for the unrolled loop, like Threshold, but used
204 /// for partial/runtime unrolling (set to UINT_MAX to disable).
205 unsigned PartialThreshold;
206 /// The cost threshold for the unrolled loop when optimizing for size, like
207 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
208 /// UINT_MAX to disable).
209 unsigned PartialOptSizeThreshold;
210 /// A forced unrolling factor (the number of concatenated bodies of the
211 /// original loop in the unrolled loop body). When set to 0, the unrolling
212 /// transformation will select an unrolling factor based on the current cost
213 /// threshold and other factors.
215 // Set the maximum unrolling factor. The unrolling factor may be selected
216 // using the appropriate cost threshold, but may not exceed this number
217 // (set to UINT_MAX to disable). This does not apply in cases where the
218 // loop is being fully unrolled.
220 /// Allow partial unrolling (unrolling of loops to expand the size of the
221 /// loop body, not only to eliminate small constant-trip-count loops).
223 /// Allow runtime unrolling (unrolling of loops to expand the size of the
224 /// loop body even when the number of loop iterations is not known at
229 /// \brief Get target-customized preferences for the generic loop unrolling
230 /// transformation. The caller will initialize UP with the current
231 /// target-independent defaults.
232 void getUnrollingPreferences(const Function *F, Loop *L,
233 UnrollingPreferences &UP) const;
237 /// \name Scalar Target Information
240 /// \brief Flags indicating the kind of support for population count.
242 /// Compared to the SW implementation, HW support is supposed to
243 /// significantly boost the performance when the population is dense, and it
244 /// may or may not degrade performance if the population is sparse. A HW
245 /// support is considered as "Fast" if it can outperform, or is on a par
246 /// with, SW implementation when the population is sparse; otherwise, it is
247 /// considered as "Slow".
248 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
250 /// \brief Return true if the specified immediate is legal add immediate, that
251 /// is the target has add instructions which can add a register with the
252 /// immediate without having to materialize the immediate into a register.
253 bool isLegalAddImmediate(int64_t Imm) const;
255 /// \brief Return true if the specified immediate is legal icmp immediate,
256 /// that is the target has icmp instructions which can compare a register
257 /// against the immediate without having to materialize the immediate into a
259 bool isLegalICmpImmediate(int64_t Imm) const;
261 /// \brief Return true if the addressing mode represented by AM is legal for
262 /// this target, for a load/store of the specified type.
263 /// The type may be VoidTy, in which case only return true if the addressing
264 /// mode is legal for a load/store of any legal type.
265 /// TODO: Handle pre/postinc as well.
266 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
267 bool HasBaseReg, int64_t Scale) const;
269 /// \brief Return true if the target works with masked instruction
270 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
271 /// AVX-512 architecture will also allow masks for non-consecutive memory
273 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
274 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
276 /// \brief Return the cost of the scaling factor used in the addressing
277 /// mode represented by AM for this target, for a load/store
278 /// of the specified type.
279 /// If the AM is supported, the return value must be >= 0.
280 /// If the AM is not supported, it returns a negative value.
281 /// TODO: Handle pre/postinc as well.
282 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
283 bool HasBaseReg, int64_t Scale) const;
285 /// \brief Return true if it's free to truncate a value of type Ty1 to type
286 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
287 /// by referencing its sub-register AX.
288 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
290 /// \brief Return true if this type is legal.
291 bool isTypeLegal(Type *Ty) const;
293 /// \brief Returns the target's jmp_buf alignment in bytes.
294 unsigned getJumpBufAlignment() const;
296 /// \brief Returns the target's jmp_buf size in bytes.
297 unsigned getJumpBufSize() const;
299 /// \brief Return true if switches should be turned into lookup tables for the
301 bool shouldBuildLookupTables() const;
303 /// \brief Return hardware support for population count.
304 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
306 /// \brief Return true if the hardware has a fast square-root instruction.
307 bool haveFastSqrt(Type *Ty) const;
309 /// \brief Return the expected cost of materializing for the given integer
310 /// immediate of the specified type.
311 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
313 /// \brief Return the expected cost of materialization for the given integer
314 /// immediate of the specified type for a given instruction. The cost can be
315 /// zero if the immediate can be folded into the specified instruction.
316 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
318 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
322 /// \name Vector Target Information
325 /// \brief The various kinds of shuffle patterns for vector queries.
327 SK_Broadcast, ///< Broadcast element 0 to all other elements.
328 SK_Reverse, ///< Reverse the order of the vector.
329 SK_Alternate, ///< Choose alternate elements from vector.
330 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
331 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
334 /// \brief Additional information about an operand's possible values.
335 enum OperandValueKind {
336 OK_AnyValue, // Operand can have any value.
337 OK_UniformValue, // Operand is uniform (splat of a value).
338 OK_UniformConstantValue, // Operand is uniform constant.
339 OK_NonUniformConstantValue // Operand is a non uniform constant value.
342 /// \brief Additional properties of an operand's values.
343 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
345 /// \return The number of scalar or vector registers that the target has.
346 /// If 'Vectors' is true, it returns the number of vector registers. If it is
347 /// set to false, it returns the number of scalar registers.
348 unsigned getNumberOfRegisters(bool Vector) const;
350 /// \return The width of the largest scalar or vector register type.
351 unsigned getRegisterBitWidth(bool Vector) const;
353 /// \return The maximum interleave factor that any transform should try to
354 /// perform for this target. This number depends on the level of parallelism
355 /// and the number of execution units in the CPU.
356 unsigned getMaxInterleaveFactor() const;
358 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
360 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
361 OperandValueKind Opd1Info = OK_AnyValue,
362 OperandValueKind Opd2Info = OK_AnyValue,
363 OperandValueProperties Opd1PropInfo = OP_None,
364 OperandValueProperties Opd2PropInfo = OP_None) const;
366 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
367 /// The index and subtype parameters are used by the subvector insertion and
368 /// extraction shuffle kinds.
369 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
370 Type *SubTp = nullptr) const;
372 /// \return The expected cost of cast instructions, such as bitcast, trunc,
374 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
376 /// \return The expected cost of control-flow related instructions such as
378 unsigned getCFInstrCost(unsigned Opcode) const;
380 /// \returns The expected cost of compare and select instructions.
381 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
382 Type *CondTy = nullptr) const;
384 /// \return The expected cost of vector Insert and Extract.
385 /// Use -1 to indicate that there is no information on the index value.
386 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
387 unsigned Index = -1) const;
389 /// \return The cost of Load and Store instructions.
390 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
391 unsigned AddressSpace) const;
393 /// \return The cost of masked Load and Store instructions.
394 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
395 unsigned AddressSpace) const;
397 /// \brief Calculate the cost of performing a vector reduction.
399 /// This is the cost of reducing the vector value of type \p Ty to a scalar
400 /// value using the operation denoted by \p Opcode. The form of the reduction
401 /// can either be a pairwise reduction or a reduction that splits the vector
402 /// at every reduction level.
406 /// ((v0+v1), (v2, v3), undef, undef)
409 /// ((v0+v2), (v1+v3), undef, undef)
410 unsigned getReductionCost(unsigned Opcode, Type *Ty,
411 bool IsPairwiseForm) const;
413 /// \returns The cost of Intrinsic instructions.
414 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
415 ArrayRef<Type *> Tys) const;
417 /// \returns The number of pieces into which the provided type must be
418 /// split during legalization. Zero is returned when the answer is unknown.
419 unsigned getNumberOfParts(Type *Tp) const;
421 /// \returns The cost of the address computation. For most targets this can be
422 /// merged into the instruction indexing mode. Some targets might want to
423 /// distinguish between address computation for memory operations on vector
424 /// types and scalar types. Such targets should override this function.
425 /// The 'IsComplex' parameter is a hint that the address computation is likely
426 /// to involve multiple instructions and as such unlikely to be merged into
427 /// the address indexing mode.
428 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
430 /// \returns The cost, if any, of keeping values of the given types alive
433 /// Some types may require the use of register classes that do not have
434 /// any callee-saved registers, so would require a spill and fill.
435 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
437 /// \returns True if the intrinsic is a supported memory intrinsic. Info
438 /// will contain additional information - whether the intrinsic may write
439 /// or read to memory, volatility and the pointer. Info is undefined
440 /// if false is returned.
441 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
443 /// \returns A value which is the result of the given memory intrinsic. New
444 /// instructions may be created to extract the result from the given intrinsic
445 /// memory operation. Returns nullptr if the target cannot create a result
446 /// from the given intrinsic.
447 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
448 Type *ExpectedType) const;
453 /// \brief The abstract base class used to type erase specific TTI
457 /// \brief The template model for the base class which wraps a concrete
458 /// implementation in a type erased interface.
459 template <typename T> class Model;
461 std::unique_ptr<Concept> TTIImpl;
464 class TargetTransformInfo::Concept {
466 virtual ~Concept() = 0;
468 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
469 virtual unsigned getGEPCost(const Value *Ptr,
470 ArrayRef<const Value *> Operands) = 0;
471 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
472 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
473 virtual unsigned getCallCost(const Function *F,
474 ArrayRef<const Value *> Arguments) = 0;
475 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
476 ArrayRef<Type *> ParamTys) = 0;
477 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
478 ArrayRef<const Value *> Arguments) = 0;
479 virtual unsigned getUserCost(const User *U) = 0;
480 virtual bool hasBranchDivergence() = 0;
481 virtual bool isLoweredToCall(const Function *F) = 0;
482 virtual void getUnrollingPreferences(const Function *F, Loop *L,
483 UnrollingPreferences &UP) = 0;
484 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
485 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
486 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
487 int64_t BaseOffset, bool HasBaseReg,
489 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
490 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
491 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
492 int64_t BaseOffset, bool HasBaseReg,
494 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
495 virtual bool isTypeLegal(Type *Ty) = 0;
496 virtual unsigned getJumpBufAlignment() = 0;
497 virtual unsigned getJumpBufSize() = 0;
498 virtual bool shouldBuildLookupTables() = 0;
499 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
500 virtual bool haveFastSqrt(Type *Ty) = 0;
501 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
502 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
504 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
505 const APInt &Imm, Type *Ty) = 0;
506 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
507 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
508 virtual unsigned getMaxInterleaveFactor() = 0;
510 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
511 OperandValueKind Opd2Info,
512 OperandValueProperties Opd1PropInfo,
513 OperandValueProperties Opd2PropInfo) = 0;
514 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
516 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
517 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
518 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
520 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
522 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
524 unsigned AddressSpace) = 0;
525 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
527 unsigned AddressSpace) = 0;
528 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
529 bool IsPairwiseForm) = 0;
530 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
531 ArrayRef<Type *> Tys) = 0;
532 virtual unsigned getNumberOfParts(Type *Tp) = 0;
533 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
534 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
535 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
536 MemIntrinsicInfo &Info) = 0;
537 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
538 Type *ExpectedType) = 0;
541 template <typename T>
542 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
546 Model(T Impl) : Impl(std::move(Impl)) {}
549 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
550 return Impl.getOperationCost(Opcode, Ty, OpTy);
552 unsigned getGEPCost(const Value *Ptr,
553 ArrayRef<const Value *> Operands) override {
554 return Impl.getGEPCost(Ptr, Operands);
556 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
557 return Impl.getCallCost(FTy, NumArgs);
559 unsigned getCallCost(const Function *F, int NumArgs) override {
560 return Impl.getCallCost(F, NumArgs);
562 unsigned getCallCost(const Function *F,
563 ArrayRef<const Value *> Arguments) override {
564 return Impl.getCallCost(F, Arguments);
566 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
567 ArrayRef<Type *> ParamTys) override {
568 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
570 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
571 ArrayRef<const Value *> Arguments) override {
572 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
574 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
575 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
576 bool isLoweredToCall(const Function *F) override {
577 return Impl.isLoweredToCall(F);
579 void getUnrollingPreferences(const Function *F, Loop *L,
580 UnrollingPreferences &UP) override {
581 return Impl.getUnrollingPreferences(F, L, UP);
583 bool isLegalAddImmediate(int64_t Imm) override {
584 return Impl.isLegalAddImmediate(Imm);
586 bool isLegalICmpImmediate(int64_t Imm) override {
587 return Impl.isLegalICmpImmediate(Imm);
589 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
590 bool HasBaseReg, int64_t Scale) override {
591 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
594 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
595 return Impl.isLegalMaskedStore(DataType, Consecutive);
597 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
598 return Impl.isLegalMaskedLoad(DataType, Consecutive);
600 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
601 bool HasBaseReg, int64_t Scale) override {
602 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
604 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
605 return Impl.isTruncateFree(Ty1, Ty2);
607 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
608 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
609 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
610 bool shouldBuildLookupTables() override {
611 return Impl.shouldBuildLookupTables();
613 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
614 return Impl.getPopcntSupport(IntTyWidthInBit);
616 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
617 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
618 return Impl.getIntImmCost(Imm, Ty);
620 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
622 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
624 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
626 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
628 unsigned getNumberOfRegisters(bool Vector) override {
629 return Impl.getNumberOfRegisters(Vector);
631 unsigned getRegisterBitWidth(bool Vector) override {
632 return Impl.getRegisterBitWidth(Vector);
634 unsigned getMaxInterleaveFactor() override {
635 return Impl.getMaxInterleaveFactor();
638 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
639 OperandValueKind Opd2Info,
640 OperandValueProperties Opd1PropInfo,
641 OperandValueProperties Opd2PropInfo) override {
642 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
643 Opd1PropInfo, Opd2PropInfo);
645 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
646 Type *SubTp) override {
647 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
649 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
650 return Impl.getCastInstrCost(Opcode, Dst, Src);
652 unsigned getCFInstrCost(unsigned Opcode) override {
653 return Impl.getCFInstrCost(Opcode);
655 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
656 Type *CondTy) override {
657 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
659 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
660 unsigned Index) override {
661 return Impl.getVectorInstrCost(Opcode, Val, Index);
663 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
664 unsigned AddressSpace) override {
665 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
667 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
668 unsigned AddressSpace) override {
669 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
671 unsigned getReductionCost(unsigned Opcode, Type *Ty,
672 bool IsPairwiseForm) override {
673 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
675 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
676 ArrayRef<Type *> Tys) override {
677 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
679 unsigned getNumberOfParts(Type *Tp) override {
680 return Impl.getNumberOfParts(Tp);
682 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
683 return Impl.getAddressComputationCost(Ty, IsComplex);
685 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
686 return Impl.getCostOfKeepingLiveOverCall(Tys);
688 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
689 MemIntrinsicInfo &Info) override {
690 return Impl.getTgtMemIntrinsic(Inst, Info);
692 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
693 Type *ExpectedType) override {
694 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
698 template <typename T>
699 TargetTransformInfo::TargetTransformInfo(T Impl)
700 : TTIImpl(new Model<T>(Impl)) {}
702 /// \brief Wrapper pass for TargetTransformInfo.
704 /// This pass can be constructed from a TTI object which it stores internally
705 /// and is queried by passes.
706 class TargetTransformInfoWrapperPass : public ImmutablePass {
707 TargetTransformInfo TTI;
709 virtual void anchor();
714 /// \brief We must provide a default constructor for the pass but it should
717 /// Use the constructor below or call one of the creation routines.
718 TargetTransformInfoWrapperPass();
720 explicit TargetTransformInfoWrapperPass(TargetTransformInfo TTI);
722 TargetTransformInfo &getTTI() { return TTI; }
723 const TargetTransformInfo &getTTI() const { return TTI; }
726 /// \brief Create the base case instance of a pass in the TTI analysis group.
728 /// This class provides the base case for the stack of TTI analyzes. It doesn't
729 /// delegate to anything and uses the STTI and VTTI objects passed in to
730 /// satisfy the queries.
731 ImmutablePass *createNoTargetTransformInfoPass(const DataLayout *DL);
733 } // End llvm namespace