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"
35 class PreservedAnalyses;
40 /// \brief Information about a load/store intrinsic defined by the target.
41 struct MemIntrinsicInfo {
43 : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
44 NumMemRefs(0), PtrVal(nullptr) {}
48 // Same Id is set by the target for corresponding load/store intrinsics.
49 unsigned short MatchingId;
54 /// \brief This pass provides access to the codegen interfaces that are needed
55 /// for IR-level transformations.
56 class TargetTransformInfo {
58 /// \brief Construct a TTI object using a type implementing the \c Concept
61 /// This is used by targets to construct a TTI wrapping their target-specific
62 /// implementaion that encodes appropriate costs for their target.
63 template <typename T> TargetTransformInfo(T Impl);
65 /// \brief Construct a baseline TTI object using a minimal implementation of
66 /// the \c Concept API below.
68 /// The TTI implementation will reflect the information in the DataLayout
69 /// provided if non-null.
70 explicit TargetTransformInfo(const DataLayout *DL);
72 // Provide move semantics.
73 TargetTransformInfo(TargetTransformInfo &&Arg);
74 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
76 // We need to define the destructor out-of-line to define our sub-classes
78 ~TargetTransformInfo();
80 /// \brief Handle the invalidation of this information.
82 /// When used as a result of \c TargetIRAnalysis this method will be called
83 /// when the function this was computed for changes. When it returns false,
84 /// the information is preserved across those changes.
85 bool invalidate(Function &, const PreservedAnalyses &) {
86 // FIXME: We should probably in some way ensure that the subtarget
87 // information for a function hasn't changed.
91 /// \name Generic Target Information
94 /// \brief Underlying constants for 'cost' values in this interface.
96 /// Many APIs in this interface return a cost. This enum defines the
97 /// fundamental values that should be used to interpret (and produce) those
98 /// costs. The costs are returned as an unsigned rather than a member of this
99 /// enumeration because it is expected that the cost of one IR instruction
100 /// may have a multiplicative factor to it or otherwise won't fit directly
101 /// into the enum. Moreover, it is common to sum or average costs which works
102 /// better as simple integral values. Thus this enum only provides constants.
104 /// Note that these costs should usually reflect the intersection of code-size
105 /// cost and execution cost. A free instruction is typically one that folds
106 /// into another instruction. For example, reg-to-reg moves can often be
107 /// skipped by renaming the registers in the CPU, but they still are encoded
108 /// and thus wouldn't be considered 'free' here.
109 enum TargetCostConstants {
110 TCC_Free = 0, ///< Expected to fold away in lowering.
111 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
112 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
115 /// \brief Estimate the cost of a specific operation when lowered.
117 /// Note that this is designed to work on an arbitrary synthetic opcode, and
118 /// thus work for hypothetical queries before an instruction has even been
119 /// formed. However, this does *not* work for GEPs, and must not be called
120 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
121 /// analyzing a GEP's cost required more information.
123 /// Typically only the result type is required, and the operand type can be
124 /// omitted. However, if the opcode is one of the cast instructions, the
125 /// operand type is required.
127 /// The returned cost is defined in terms of \c TargetCostConstants, see its
128 /// comments for a detailed explanation of the cost values.
129 unsigned getOperationCost(unsigned Opcode, Type *Ty,
130 Type *OpTy = nullptr) const;
132 /// \brief Estimate the cost of a GEP operation when lowered.
134 /// The contract for this function is the same as \c getOperationCost except
135 /// that it supports an interface that provides extra information specific to
136 /// the GEP operation.
137 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
139 /// \brief Estimate the cost of a function call when lowered.
141 /// The contract for this is the same as \c getOperationCost except that it
142 /// supports an interface that provides extra information specific to call
145 /// This is the most basic query for estimating call cost: it only knows the
146 /// function type and (potentially) the number of arguments at the call site.
147 /// The latter is only interesting for varargs function types.
148 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
150 /// \brief Estimate the cost of calling a specific function when lowered.
152 /// This overload adds the ability to reason about the particular function
153 /// being called in the event it is a library call with special lowering.
154 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
156 /// \brief Estimate the cost of calling a specific function when lowered.
158 /// This overload allows specifying a set of candidate argument values.
159 unsigned getCallCost(const Function *F,
160 ArrayRef<const Value *> Arguments) const;
162 /// \brief Estimate the cost of an intrinsic when lowered.
164 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
165 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
166 ArrayRef<Type *> ParamTys) const;
168 /// \brief Estimate the cost of an intrinsic when lowered.
170 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
171 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
172 ArrayRef<const Value *> Arguments) const;
174 /// \brief Estimate the cost of a given IR user when lowered.
176 /// This can estimate the cost of either a ConstantExpr or Instruction when
177 /// lowered. It has two primary advantages over the \c getOperationCost and
178 /// \c getGEPCost above, and one significant disadvantage: it can only be
179 /// used when the IR construct has already been formed.
181 /// The advantages are that it can inspect the SSA use graph to reason more
182 /// accurately about the cost. For example, all-constant-GEPs can often be
183 /// folded into a load or other instruction, but if they are used in some
184 /// other context they may not be folded. This routine can distinguish such
187 /// The returned cost is defined in terms of \c TargetCostConstants, see its
188 /// comments for a detailed explanation of the cost values.
189 unsigned getUserCost(const User *U) const;
191 /// \brief hasBranchDivergence - Return true if branch divergence exists.
192 /// Branch divergence has a significantly negative impact on GPU performance
193 /// when threads in the same wavefront take different paths due to conditional
195 bool hasBranchDivergence() const;
197 /// \brief Test whether calls to a function lower to actual program function
200 /// The idea is to test whether the program is likely to require a 'call'
201 /// instruction or equivalent in order to call the given function.
203 /// FIXME: It's not clear that this is a good or useful query API. Client's
204 /// should probably move to simpler cost metrics using the above.
205 /// Alternatively, we could split the cost interface into distinct code-size
206 /// and execution-speed costs. This would allow modelling the core of this
207 /// query more accurately as a call is a single small instruction, but
208 /// incurs significant execution cost.
209 bool isLoweredToCall(const Function *F) const;
211 /// Parameters that control the generic loop unrolling transformation.
212 struct UnrollingPreferences {
213 /// The cost threshold for the unrolled loop, compared to
214 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
215 /// The unrolling factor is set such that the unrolled loop body does not
216 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
219 /// The cost threshold for the unrolled loop when optimizing for size (set
220 /// to UINT_MAX to disable).
221 unsigned OptSizeThreshold;
222 /// The cost threshold for the unrolled loop, like Threshold, but used
223 /// for partial/runtime unrolling (set to UINT_MAX to disable).
224 unsigned PartialThreshold;
225 /// The cost threshold for the unrolled loop when optimizing for size, like
226 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
227 /// UINT_MAX to disable).
228 unsigned PartialOptSizeThreshold;
229 /// A forced unrolling factor (the number of concatenated bodies of the
230 /// original loop in the unrolled loop body). When set to 0, the unrolling
231 /// transformation will select an unrolling factor based on the current cost
232 /// threshold and other factors.
234 // Set the maximum unrolling factor. The unrolling factor may be selected
235 // using the appropriate cost threshold, but may not exceed this number
236 // (set to UINT_MAX to disable). This does not apply in cases where the
237 // loop is being fully unrolled.
239 /// Allow partial unrolling (unrolling of loops to expand the size of the
240 /// loop body, not only to eliminate small constant-trip-count loops).
242 /// Allow runtime unrolling (unrolling of loops to expand the size of the
243 /// loop body even when the number of loop iterations is not known at
248 /// \brief Get target-customized preferences for the generic loop unrolling
249 /// transformation. The caller will initialize UP with the current
250 /// target-independent defaults.
251 void getUnrollingPreferences(const Function *F, Loop *L,
252 UnrollingPreferences &UP) const;
256 /// \name Scalar Target Information
259 /// \brief Flags indicating the kind of support for population count.
261 /// Compared to the SW implementation, HW support is supposed to
262 /// significantly boost the performance when the population is dense, and it
263 /// may or may not degrade performance if the population is sparse. A HW
264 /// support is considered as "Fast" if it can outperform, or is on a par
265 /// with, SW implementation when the population is sparse; otherwise, it is
266 /// considered as "Slow".
267 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
269 /// \brief Return true if the specified immediate is legal add immediate, that
270 /// is the target has add instructions which can add a register with the
271 /// immediate without having to materialize the immediate into a register.
272 bool isLegalAddImmediate(int64_t Imm) const;
274 /// \brief Return true if the specified immediate is legal icmp immediate,
275 /// that is the target has icmp instructions which can compare a register
276 /// against the immediate without having to materialize the immediate into a
278 bool isLegalICmpImmediate(int64_t Imm) const;
280 /// \brief Return true if the addressing mode represented by AM is legal for
281 /// this target, for a load/store of the specified type.
282 /// The type may be VoidTy, in which case only return true if the addressing
283 /// mode is legal for a load/store of any legal type.
284 /// TODO: Handle pre/postinc as well.
285 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
286 bool HasBaseReg, int64_t Scale) const;
288 /// \brief Return true if the target works with masked instruction
289 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
290 /// AVX-512 architecture will also allow masks for non-consecutive memory
292 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
293 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
295 /// \brief Return the cost of the scaling factor used in the addressing
296 /// mode represented by AM for this target, for a load/store
297 /// of the specified type.
298 /// If the AM is supported, the return value must be >= 0.
299 /// If the AM is not supported, it returns a negative value.
300 /// TODO: Handle pre/postinc as well.
301 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
302 bool HasBaseReg, int64_t Scale) const;
304 /// \brief Return true if it's free to truncate a value of type Ty1 to type
305 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
306 /// by referencing its sub-register AX.
307 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
309 /// \brief Return true if this type is legal.
310 bool isTypeLegal(Type *Ty) const;
312 /// \brief Returns the target's jmp_buf alignment in bytes.
313 unsigned getJumpBufAlignment() const;
315 /// \brief Returns the target's jmp_buf size in bytes.
316 unsigned getJumpBufSize() const;
318 /// \brief Return true if switches should be turned into lookup tables for the
320 bool shouldBuildLookupTables() const;
322 /// \brief Return hardware support for population count.
323 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
325 /// \brief Return true if the hardware has a fast square-root instruction.
326 bool haveFastSqrt(Type *Ty) const;
328 /// \brief Return the expected cost of materializing for the given integer
329 /// immediate of the specified type.
330 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
332 /// \brief Return the expected cost of materialization for the given integer
333 /// immediate of the specified type for a given instruction. The cost can be
334 /// zero if the immediate can be folded into the specified instruction.
335 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
337 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
341 /// \name Vector Target Information
344 /// \brief The various kinds of shuffle patterns for vector queries.
346 SK_Broadcast, ///< Broadcast element 0 to all other elements.
347 SK_Reverse, ///< Reverse the order of the vector.
348 SK_Alternate, ///< Choose alternate elements from vector.
349 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
350 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
353 /// \brief Additional information about an operand's possible values.
354 enum OperandValueKind {
355 OK_AnyValue, // Operand can have any value.
356 OK_UniformValue, // Operand is uniform (splat of a value).
357 OK_UniformConstantValue, // Operand is uniform constant.
358 OK_NonUniformConstantValue // Operand is a non uniform constant value.
361 /// \brief Additional properties of an operand's values.
362 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
364 /// \return The number of scalar or vector registers that the target has.
365 /// If 'Vectors' is true, it returns the number of vector registers. If it is
366 /// set to false, it returns the number of scalar registers.
367 unsigned getNumberOfRegisters(bool Vector) const;
369 /// \return The width of the largest scalar or vector register type.
370 unsigned getRegisterBitWidth(bool Vector) const;
372 /// \return The maximum interleave factor that any transform should try to
373 /// perform for this target. This number depends on the level of parallelism
374 /// and the number of execution units in the CPU.
375 unsigned getMaxInterleaveFactor() const;
377 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
379 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
380 OperandValueKind Opd1Info = OK_AnyValue,
381 OperandValueKind Opd2Info = OK_AnyValue,
382 OperandValueProperties Opd1PropInfo = OP_None,
383 OperandValueProperties Opd2PropInfo = OP_None) const;
385 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
386 /// The index and subtype parameters are used by the subvector insertion and
387 /// extraction shuffle kinds.
388 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
389 Type *SubTp = nullptr) const;
391 /// \return The expected cost of cast instructions, such as bitcast, trunc,
393 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
395 /// \return The expected cost of control-flow related instructions such as
397 unsigned getCFInstrCost(unsigned Opcode) const;
399 /// \returns The expected cost of compare and select instructions.
400 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
401 Type *CondTy = nullptr) const;
403 /// \return The expected cost of vector Insert and Extract.
404 /// Use -1 to indicate that there is no information on the index value.
405 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
406 unsigned Index = -1) const;
408 /// \return The cost of Load and Store instructions.
409 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
410 unsigned AddressSpace) const;
412 /// \return The cost of masked Load and Store instructions.
413 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
414 unsigned AddressSpace) const;
416 /// \brief Calculate the cost of performing a vector reduction.
418 /// This is the cost of reducing the vector value of type \p Ty to a scalar
419 /// value using the operation denoted by \p Opcode. The form of the reduction
420 /// can either be a pairwise reduction or a reduction that splits the vector
421 /// at every reduction level.
425 /// ((v0+v1), (v2, v3), undef, undef)
428 /// ((v0+v2), (v1+v3), undef, undef)
429 unsigned getReductionCost(unsigned Opcode, Type *Ty,
430 bool IsPairwiseForm) const;
432 /// \returns The cost of Intrinsic instructions.
433 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
434 ArrayRef<Type *> Tys) const;
436 /// \returns The number of pieces into which the provided type must be
437 /// split during legalization. Zero is returned when the answer is unknown.
438 unsigned getNumberOfParts(Type *Tp) const;
440 /// \returns The cost of the address computation. For most targets this can be
441 /// merged into the instruction indexing mode. Some targets might want to
442 /// distinguish between address computation for memory operations on vector
443 /// types and scalar types. Such targets should override this function.
444 /// The 'IsComplex' parameter is a hint that the address computation is likely
445 /// to involve multiple instructions and as such unlikely to be merged into
446 /// the address indexing mode.
447 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
449 /// \returns The cost, if any, of keeping values of the given types alive
452 /// Some types may require the use of register classes that do not have
453 /// any callee-saved registers, so would require a spill and fill.
454 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
456 /// \returns True if the intrinsic is a supported memory intrinsic. Info
457 /// will contain additional information - whether the intrinsic may write
458 /// or read to memory, volatility and the pointer. Info is undefined
459 /// if false is returned.
460 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
462 /// \returns A value which is the result of the given memory intrinsic. New
463 /// instructions may be created to extract the result from the given intrinsic
464 /// memory operation. Returns nullptr if the target cannot create a result
465 /// from the given intrinsic.
466 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
467 Type *ExpectedType) const;
472 /// \brief The abstract base class used to type erase specific TTI
476 /// \brief The template model for the base class which wraps a concrete
477 /// implementation in a type erased interface.
478 template <typename T> class Model;
480 std::unique_ptr<Concept> TTIImpl;
483 class TargetTransformInfo::Concept {
485 virtual ~Concept() = 0;
487 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
488 virtual unsigned getGEPCost(const Value *Ptr,
489 ArrayRef<const Value *> Operands) = 0;
490 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
491 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
492 virtual unsigned getCallCost(const Function *F,
493 ArrayRef<const Value *> Arguments) = 0;
494 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
495 ArrayRef<Type *> ParamTys) = 0;
496 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
497 ArrayRef<const Value *> Arguments) = 0;
498 virtual unsigned getUserCost(const User *U) = 0;
499 virtual bool hasBranchDivergence() = 0;
500 virtual bool isLoweredToCall(const Function *F) = 0;
501 virtual void getUnrollingPreferences(const Function *F, Loop *L,
502 UnrollingPreferences &UP) = 0;
503 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
504 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
505 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
506 int64_t BaseOffset, bool HasBaseReg,
508 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
509 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
510 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
511 int64_t BaseOffset, bool HasBaseReg,
513 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
514 virtual bool isTypeLegal(Type *Ty) = 0;
515 virtual unsigned getJumpBufAlignment() = 0;
516 virtual unsigned getJumpBufSize() = 0;
517 virtual bool shouldBuildLookupTables() = 0;
518 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
519 virtual bool haveFastSqrt(Type *Ty) = 0;
520 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
521 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
523 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
524 const APInt &Imm, Type *Ty) = 0;
525 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
526 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
527 virtual unsigned getMaxInterleaveFactor() = 0;
529 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
530 OperandValueKind Opd2Info,
531 OperandValueProperties Opd1PropInfo,
532 OperandValueProperties Opd2PropInfo) = 0;
533 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
535 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
536 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
537 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
539 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
541 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
543 unsigned AddressSpace) = 0;
544 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
546 unsigned AddressSpace) = 0;
547 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
548 bool IsPairwiseForm) = 0;
549 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
550 ArrayRef<Type *> Tys) = 0;
551 virtual unsigned getNumberOfParts(Type *Tp) = 0;
552 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
553 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
554 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
555 MemIntrinsicInfo &Info) = 0;
556 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
557 Type *ExpectedType) = 0;
560 template <typename T>
561 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
565 Model(T Impl) : Impl(std::move(Impl)) {}
568 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
569 return Impl.getOperationCost(Opcode, Ty, OpTy);
571 unsigned getGEPCost(const Value *Ptr,
572 ArrayRef<const Value *> Operands) override {
573 return Impl.getGEPCost(Ptr, Operands);
575 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
576 return Impl.getCallCost(FTy, NumArgs);
578 unsigned getCallCost(const Function *F, int NumArgs) override {
579 return Impl.getCallCost(F, NumArgs);
581 unsigned getCallCost(const Function *F,
582 ArrayRef<const Value *> Arguments) override {
583 return Impl.getCallCost(F, Arguments);
585 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
586 ArrayRef<Type *> ParamTys) override {
587 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
589 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
590 ArrayRef<const Value *> Arguments) override {
591 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
593 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
594 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
595 bool isLoweredToCall(const Function *F) override {
596 return Impl.isLoweredToCall(F);
598 void getUnrollingPreferences(const Function *F, Loop *L,
599 UnrollingPreferences &UP) override {
600 return Impl.getUnrollingPreferences(F, L, UP);
602 bool isLegalAddImmediate(int64_t Imm) override {
603 return Impl.isLegalAddImmediate(Imm);
605 bool isLegalICmpImmediate(int64_t Imm) override {
606 return Impl.isLegalICmpImmediate(Imm);
608 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
609 bool HasBaseReg, int64_t Scale) override {
610 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
613 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
614 return Impl.isLegalMaskedStore(DataType, Consecutive);
616 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
617 return Impl.isLegalMaskedLoad(DataType, Consecutive);
619 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
620 bool HasBaseReg, int64_t Scale) override {
621 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
623 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
624 return Impl.isTruncateFree(Ty1, Ty2);
626 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
627 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
628 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
629 bool shouldBuildLookupTables() override {
630 return Impl.shouldBuildLookupTables();
632 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
633 return Impl.getPopcntSupport(IntTyWidthInBit);
635 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
636 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
637 return Impl.getIntImmCost(Imm, Ty);
639 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
641 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
643 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
645 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
647 unsigned getNumberOfRegisters(bool Vector) override {
648 return Impl.getNumberOfRegisters(Vector);
650 unsigned getRegisterBitWidth(bool Vector) override {
651 return Impl.getRegisterBitWidth(Vector);
653 unsigned getMaxInterleaveFactor() override {
654 return Impl.getMaxInterleaveFactor();
657 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
658 OperandValueKind Opd2Info,
659 OperandValueProperties Opd1PropInfo,
660 OperandValueProperties Opd2PropInfo) override {
661 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
662 Opd1PropInfo, Opd2PropInfo);
664 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
665 Type *SubTp) override {
666 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
668 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
669 return Impl.getCastInstrCost(Opcode, Dst, Src);
671 unsigned getCFInstrCost(unsigned Opcode) override {
672 return Impl.getCFInstrCost(Opcode);
674 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
675 Type *CondTy) override {
676 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
678 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
679 unsigned Index) override {
680 return Impl.getVectorInstrCost(Opcode, Val, Index);
682 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
683 unsigned AddressSpace) override {
684 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
686 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
687 unsigned AddressSpace) override {
688 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
690 unsigned getReductionCost(unsigned Opcode, Type *Ty,
691 bool IsPairwiseForm) override {
692 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
694 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
695 ArrayRef<Type *> Tys) override {
696 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
698 unsigned getNumberOfParts(Type *Tp) override {
699 return Impl.getNumberOfParts(Tp);
701 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
702 return Impl.getAddressComputationCost(Ty, IsComplex);
704 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
705 return Impl.getCostOfKeepingLiveOverCall(Tys);
707 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
708 MemIntrinsicInfo &Info) override {
709 return Impl.getTgtMemIntrinsic(Inst, Info);
711 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
712 Type *ExpectedType) override {
713 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
717 template <typename T>
718 TargetTransformInfo::TargetTransformInfo(T Impl)
719 : TTIImpl(new Model<T>(Impl)) {}
721 /// \brief Analysis pass providing the \c TargetTransformInfo.
723 /// The core idea of the TargetIRAnalysis is to expose an interface through
724 /// which LLVM targets can analyze and provide information about the middle
725 /// end's target-independent IR. This supports use cases such as target-aware
726 /// cost modeling of IR constructs.
728 /// This is a function analysis because much of the cost modeling for targets
729 /// is done in a subtarget specific way and LLVM supports compiling different
730 /// functions targeting different subtargets in order to support runtime
731 /// dispatch according to the observed subtarget.
732 class TargetIRAnalysis {
734 typedef TargetTransformInfo Result;
736 /// \brief Opaque, unique identifier for this analysis pass.
737 static void *ID() { return (void *)&PassID; }
739 /// \brief Provide access to a name for this pass for debugging purposes.
740 static StringRef name() { return "TargetIRAnalysis"; }
742 /// \brief Default construct a target IR analysis.
744 /// This will use the module's datalayout to construct a baseline
745 /// conservative TTI result.
748 /// \brief Construct an IR analysis pass around a target-provide callback.
750 /// The callback will be called with a particular function for which the TTI
751 /// is needed and must return a TTI object for that function.
752 TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
754 // Value semantics. We spell out the constructors for MSVC.
755 TargetIRAnalysis(const TargetIRAnalysis &Arg)
756 : TTICallback(Arg.TTICallback) {}
757 TargetIRAnalysis(TargetIRAnalysis &&Arg)
758 : TTICallback(std::move(Arg.TTICallback)) {}
759 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
760 TTICallback = RHS.TTICallback;
763 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
764 TTICallback = std::move(RHS.TTICallback);
768 Result run(Function &F);
773 /// \brief The callback used to produce a result.
775 /// We use a completely opaque callback so that targets can provide whatever
776 /// mechanism they desire for constructing the TTI for a given function.
778 /// FIXME: Should we really use std::function? It's relatively inefficient.
779 /// It might be possible to arrange for even stateful callbacks to outlive
780 /// the analysis and thus use a function_ref which would be lighter weight.
781 /// This may also be less error prone as the callback is likely to reference
782 /// the external TargetMachine, and that reference needs to never dangle.
783 std::function<Result(Function &)> TTICallback;
785 /// \brief Helper function used as the callback in the default constructor.
786 static Result getDefaultTTI(Function &F);
789 /// \brief Wrapper pass for TargetTransformInfo.
791 /// This pass can be constructed from a TTI object which it stores internally
792 /// and is queried by passes.
793 class TargetTransformInfoWrapperPass : public ImmutablePass {
794 TargetTransformInfo TTI;
796 virtual void anchor();
801 /// \brief We must provide a default constructor for the pass but it should
804 /// Use the constructor below or call one of the creation routines.
805 TargetTransformInfoWrapperPass();
807 explicit TargetTransformInfoWrapperPass(TargetTransformInfo TTI);
809 TargetTransformInfo &getTTI() { return TTI; }
810 const TargetTransformInfo &getTTI() const { return TTI; }
813 /// \brief Create an analysis pass wrapper around a TTI object.
815 /// This analysis pass just holds the TTI instance and makes it available to
817 ImmutablePass *createTargetTransformInfoWrapperPass(TargetTransformInfo TTI);
819 } // End llvm namespace