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/ADT/Optional.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/DataTypes.h"
37 class PreservedAnalyses;
42 /// \brief Information about a load/store intrinsic defined by the target.
43 struct MemIntrinsicInfo {
45 : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
46 NumMemRefs(0), PtrVal(nullptr) {}
50 // Same Id is set by the target for corresponding load/store intrinsics.
51 unsigned short MatchingId;
56 /// \brief This pass provides access to the codegen interfaces that are needed
57 /// for IR-level transformations.
58 class TargetTransformInfo {
60 /// \brief Construct a TTI object using a type implementing the \c Concept
63 /// This is used by targets to construct a TTI wrapping their target-specific
64 /// implementaion that encodes appropriate costs for their target.
65 template <typename T> TargetTransformInfo(T Impl);
67 /// \brief Construct a baseline TTI object using a minimal implementation of
68 /// the \c Concept API below.
70 /// The TTI implementation will reflect the information in the DataLayout
71 /// provided if non-null.
72 explicit TargetTransformInfo(const DataLayout &DL);
74 // Provide move semantics.
75 TargetTransformInfo(TargetTransformInfo &&Arg);
76 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
78 // We need to define the destructor out-of-line to define our sub-classes
80 ~TargetTransformInfo();
82 /// \brief Handle the invalidation of this information.
84 /// When used as a result of \c TargetIRAnalysis this method will be called
85 /// when the function this was computed for changes. When it returns false,
86 /// the information is preserved across those changes.
87 bool invalidate(Function &, const PreservedAnalyses &) {
88 // FIXME: We should probably in some way ensure that the subtarget
89 // information for a function hasn't changed.
93 /// \name Generic Target Information
96 /// \brief Underlying constants for 'cost' values in this interface.
98 /// Many APIs in this interface return a cost. This enum defines the
99 /// fundamental values that should be used to interpret (and produce) those
100 /// costs. The costs are returned as an unsigned rather than a member of this
101 /// enumeration because it is expected that the cost of one IR instruction
102 /// may have a multiplicative factor to it or otherwise won't fit directly
103 /// into the enum. Moreover, it is common to sum or average costs which works
104 /// better as simple integral values. Thus this enum only provides constants.
106 /// Note that these costs should usually reflect the intersection of code-size
107 /// cost and execution cost. A free instruction is typically one that folds
108 /// into another instruction. For example, reg-to-reg moves can often be
109 /// skipped by renaming the registers in the CPU, but they still are encoded
110 /// and thus wouldn't be considered 'free' here.
111 enum TargetCostConstants {
112 TCC_Free = 0, ///< Expected to fold away in lowering.
113 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
114 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
117 /// \brief Estimate the cost of a specific operation when lowered.
119 /// Note that this is designed to work on an arbitrary synthetic opcode, and
120 /// thus work for hypothetical queries before an instruction has even been
121 /// formed. However, this does *not* work for GEPs, and must not be called
122 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
123 /// analyzing a GEP's cost required more information.
125 /// Typically only the result type is required, and the operand type can be
126 /// omitted. However, if the opcode is one of the cast instructions, the
127 /// operand type is required.
129 /// The returned cost is defined in terms of \c TargetCostConstants, see its
130 /// comments for a detailed explanation of the cost values.
131 unsigned getOperationCost(unsigned Opcode, Type *Ty,
132 Type *OpTy = nullptr) const;
134 /// \brief Estimate the cost of a GEP operation when lowered.
136 /// The contract for this function is the same as \c getOperationCost except
137 /// that it supports an interface that provides extra information specific to
138 /// the GEP operation.
139 unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
141 /// \brief Estimate the cost of a function call when lowered.
143 /// The contract for this is the same as \c getOperationCost except that it
144 /// supports an interface that provides extra information specific to call
147 /// This is the most basic query for estimating call cost: it only knows the
148 /// function type and (potentially) the number of arguments at the call site.
149 /// The latter is only interesting for varargs function types.
150 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
152 /// \brief Estimate the cost of calling a specific function when lowered.
154 /// This overload adds the ability to reason about the particular function
155 /// being called in the event it is a library call with special lowering.
156 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
158 /// \brief Estimate the cost of calling a specific function when lowered.
160 /// This overload allows specifying a set of candidate argument values.
161 unsigned getCallCost(const Function *F,
162 ArrayRef<const Value *> Arguments) const;
164 /// \brief Estimate the cost of an intrinsic when lowered.
166 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
167 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
168 ArrayRef<Type *> ParamTys) const;
170 /// \brief Estimate the cost of an intrinsic when lowered.
172 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
173 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
174 ArrayRef<const Value *> Arguments) const;
176 /// \brief Estimate the cost of a given IR user when lowered.
178 /// This can estimate the cost of either a ConstantExpr or Instruction when
179 /// lowered. It has two primary advantages over the \c getOperationCost and
180 /// \c getGEPCost above, and one significant disadvantage: it can only be
181 /// used when the IR construct has already been formed.
183 /// The advantages are that it can inspect the SSA use graph to reason more
184 /// accurately about the cost. For example, all-constant-GEPs can often be
185 /// folded into a load or other instruction, but if they are used in some
186 /// other context they may not be folded. This routine can distinguish such
189 /// The returned cost is defined in terms of \c TargetCostConstants, see its
190 /// comments for a detailed explanation of the cost values.
191 unsigned getUserCost(const User *U) const;
193 /// \brief Return true if branch divergence exists.
195 /// Branch divergence has a significantly negative impact on GPU performance
196 /// when threads in the same wavefront take different paths due to conditional
198 bool hasBranchDivergence() const;
200 /// \brief Returns whether V is a source of divergence.
202 /// This function provides the target-dependent information for
203 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
204 /// builds the dependency graph, and then runs the reachability algorithm
205 /// starting with the sources of divergence.
206 bool isSourceOfDivergence(const Value *V) const;
208 /// \brief Test whether calls to a function lower to actual program function
211 /// The idea is to test whether the program is likely to require a 'call'
212 /// instruction or equivalent in order to call the given function.
214 /// FIXME: It's not clear that this is a good or useful query API. Client's
215 /// should probably move to simpler cost metrics using the above.
216 /// Alternatively, we could split the cost interface into distinct code-size
217 /// and execution-speed costs. This would allow modelling the core of this
218 /// query more accurately as a call is a single small instruction, but
219 /// incurs significant execution cost.
220 bool isLoweredToCall(const Function *F) const;
222 /// Parameters that control the generic loop unrolling transformation.
223 struct UnrollingPreferences {
224 /// The cost threshold for the unrolled loop. Should be relative to the
225 /// getUserCost values returned by this API, and the expectation is that
226 /// the unrolled loop's instructions when run through that interface should
227 /// not exceed this cost. However, this is only an estimate. Also, specific
228 /// loops may be unrolled even with a cost above this threshold if deemed
229 /// profitable. Set this to UINT_MAX to disable the loop body cost
232 /// If complete unrolling will reduce the cost of the loop below its
233 /// expected dynamic cost while rolled by this percentage, apply a discount
234 /// (below) to its unrolled cost.
235 unsigned PercentDynamicCostSavedThreshold;
236 /// The discount applied to the unrolled cost when the *dynamic* cost
237 /// savings of unrolling exceed the \c PercentDynamicCostSavedThreshold.
238 unsigned DynamicCostSavingsDiscount;
239 /// The cost threshold for the unrolled loop when optimizing for size (set
240 /// to UINT_MAX to disable).
241 unsigned OptSizeThreshold;
242 /// The cost threshold for the unrolled loop, like Threshold, but used
243 /// for partial/runtime unrolling (set to UINT_MAX to disable).
244 unsigned PartialThreshold;
245 /// The cost threshold for the unrolled loop when optimizing for size, like
246 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
247 /// UINT_MAX to disable).
248 unsigned PartialOptSizeThreshold;
249 /// A forced unrolling factor (the number of concatenated bodies of the
250 /// original loop in the unrolled loop body). When set to 0, the unrolling
251 /// transformation will select an unrolling factor based on the current cost
252 /// threshold and other factors.
254 // Set the maximum unrolling factor. The unrolling factor may be selected
255 // using the appropriate cost threshold, but may not exceed this number
256 // (set to UINT_MAX to disable). This does not apply in cases where the
257 // loop is being fully unrolled.
259 /// Allow partial unrolling (unrolling of loops to expand the size of the
260 /// loop body, not only to eliminate small constant-trip-count loops).
262 /// Allow runtime unrolling (unrolling of loops to expand the size of the
263 /// loop body even when the number of loop iterations is not known at
266 /// Allow emitting expensive instructions (such as divisions) when computing
267 /// the trip count of a loop for runtime unrolling.
268 bool AllowExpensiveTripCount;
271 /// \brief Get target-customized preferences for the generic loop unrolling
272 /// transformation. The caller will initialize UP with the current
273 /// target-independent defaults.
274 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
278 /// \name Scalar Target Information
281 /// \brief Flags indicating the kind of support for population count.
283 /// Compared to the SW implementation, HW support is supposed to
284 /// significantly boost the performance when the population is dense, and it
285 /// may or may not degrade performance if the population is sparse. A HW
286 /// support is considered as "Fast" if it can outperform, or is on a par
287 /// with, SW implementation when the population is sparse; otherwise, it is
288 /// considered as "Slow".
289 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
291 /// \brief Return true if the specified immediate is legal add immediate, that
292 /// is the target has add instructions which can add a register with the
293 /// immediate without having to materialize the immediate into a register.
294 bool isLegalAddImmediate(int64_t Imm) const;
296 /// \brief Return true if the specified immediate is legal icmp immediate,
297 /// that is the target has icmp instructions which can compare a register
298 /// against the immediate without having to materialize the immediate into a
300 bool isLegalICmpImmediate(int64_t Imm) const;
302 /// \brief Return true if the addressing mode represented by AM is legal for
303 /// this target, for a load/store of the specified type.
304 /// The type may be VoidTy, in which case only return true if the addressing
305 /// mode is legal for a load/store of any legal type.
306 /// TODO: Handle pre/postinc as well.
307 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
308 bool HasBaseReg, int64_t Scale,
309 unsigned AddrSpace = 0) const;
311 /// \brief Return true if the target works with masked instruction
312 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
313 /// AVX-512 architecture will also allow masks for non-consecutive memory
315 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
316 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
318 /// \brief Return the cost of the scaling factor used in the addressing
319 /// mode represented by AM for this target, for a load/store
320 /// of the specified type.
321 /// If the AM is supported, the return value must be >= 0.
322 /// If the AM is not supported, it returns a negative value.
323 /// TODO: Handle pre/postinc as well.
324 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
325 bool HasBaseReg, int64_t Scale,
326 unsigned AddrSpace = 0) const;
328 /// \brief Return true if it's free to truncate a value of type Ty1 to type
329 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
330 /// by referencing its sub-register AX.
331 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
333 /// \brief Return true if it is profitable to hoist instruction in the
334 /// then/else to before if.
335 bool isProfitableToHoist(Instruction *I) const;
337 /// \brief Return true if this type is legal.
338 bool isTypeLegal(Type *Ty) const;
340 /// \brief Returns the target's jmp_buf alignment in bytes.
341 unsigned getJumpBufAlignment() const;
343 /// \brief Returns the target's jmp_buf size in bytes.
344 unsigned getJumpBufSize() const;
346 /// \brief Return true if switches should be turned into lookup tables for the
348 bool shouldBuildLookupTables() const;
350 /// \brief Don't restrict interleaved unrolling to small loops.
351 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
353 /// \brief Return hardware support for population count.
354 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
356 /// \brief Return true if the hardware has a fast square-root instruction.
357 bool haveFastSqrt(Type *Ty) const;
359 /// \brief Return the expected cost of supporting the floating point operation
360 /// of the specified type.
361 unsigned getFPOpCost(Type *Ty) const;
363 /// \brief Return the expected cost of materializing for the given integer
364 /// immediate of the specified type.
365 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
367 /// \brief Return the expected cost of materialization for the given integer
368 /// immediate of the specified type for a given instruction. The cost can be
369 /// zero if the immediate can be folded into the specified instruction.
370 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
372 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
376 /// \name Vector Target Information
379 /// \brief The various kinds of shuffle patterns for vector queries.
381 SK_Broadcast, ///< Broadcast element 0 to all other elements.
382 SK_Reverse, ///< Reverse the order of the vector.
383 SK_Alternate, ///< Choose alternate elements from vector.
384 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
385 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
388 /// \brief Additional information about an operand's possible values.
389 enum OperandValueKind {
390 OK_AnyValue, // Operand can have any value.
391 OK_UniformValue, // Operand is uniform (splat of a value).
392 OK_UniformConstantValue, // Operand is uniform constant.
393 OK_NonUniformConstantValue // Operand is a non uniform constant value.
396 /// \brief Additional properties of an operand's values.
397 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
399 /// \return The number of scalar or vector registers that the target has.
400 /// If 'Vectors' is true, it returns the number of vector registers. If it is
401 /// set to false, it returns the number of scalar registers.
402 unsigned getNumberOfRegisters(bool Vector) const;
404 /// \return The width of the largest scalar or vector register type.
405 unsigned getRegisterBitWidth(bool Vector) const;
407 /// \return The maximum interleave factor that any transform should try to
408 /// perform for this target. This number depends on the level of parallelism
409 /// and the number of execution units in the CPU.
410 unsigned getMaxInterleaveFactor(unsigned VF) const;
412 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
414 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
415 OperandValueKind Opd1Info = OK_AnyValue,
416 OperandValueKind Opd2Info = OK_AnyValue,
417 OperandValueProperties Opd1PropInfo = OP_None,
418 OperandValueProperties Opd2PropInfo = OP_None) const;
420 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
421 /// The index and subtype parameters are used by the subvector insertion and
422 /// extraction shuffle kinds.
423 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
424 Type *SubTp = nullptr) const;
426 /// \return The expected cost of cast instructions, such as bitcast, trunc,
428 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
430 /// \return The expected cost of control-flow related instructions such as
432 unsigned getCFInstrCost(unsigned Opcode) const;
434 /// \returns The expected cost of compare and select instructions.
435 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
436 Type *CondTy = nullptr) const;
438 /// \return The expected cost of vector Insert and Extract.
439 /// Use -1 to indicate that there is no information on the index value.
440 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
441 unsigned Index = -1) const;
443 /// \return The cost of Load and Store instructions.
444 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
445 unsigned AddressSpace) const;
447 /// \return The cost of masked Load and Store instructions.
448 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
449 unsigned AddressSpace) const;
451 /// \return The cost of the interleaved memory operation.
452 /// \p Opcode is the memory operation code
453 /// \p VecTy is the vector type of the interleaved access.
454 /// \p Factor is the interleave factor
455 /// \p Indices is the indices for interleaved load members (as interleaved
456 /// load allows gaps)
457 /// \p Alignment is the alignment of the memory operation
458 /// \p AddressSpace is address space of the pointer.
459 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
461 ArrayRef<unsigned> Indices,
463 unsigned AddressSpace) const;
465 /// \brief Calculate the cost of performing a vector reduction.
467 /// This is the cost of reducing the vector value of type \p Ty to a scalar
468 /// value using the operation denoted by \p Opcode. The form of the reduction
469 /// can either be a pairwise reduction or a reduction that splits the vector
470 /// at every reduction level.
474 /// ((v0+v1), (v2, v3), undef, undef)
477 /// ((v0+v2), (v1+v3), undef, undef)
478 unsigned getReductionCost(unsigned Opcode, Type *Ty,
479 bool IsPairwiseForm) const;
481 /// \returns The cost of Intrinsic instructions.
482 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
483 ArrayRef<Type *> Tys) const;
485 /// \returns The cost of Call instructions.
486 unsigned getCallInstrCost(Function *F, Type *RetTy,
487 ArrayRef<Type *> Tys) const;
489 /// \returns The number of pieces into which the provided type must be
490 /// split during legalization. Zero is returned when the answer is unknown.
491 unsigned getNumberOfParts(Type *Tp) const;
493 /// \returns The cost of the address computation. For most targets this can be
494 /// merged into the instruction indexing mode. Some targets might want to
495 /// distinguish between address computation for memory operations on vector
496 /// types and scalar types. Such targets should override this function.
497 /// The 'IsComplex' parameter is a hint that the address computation is likely
498 /// to involve multiple instructions and as such unlikely to be merged into
499 /// the address indexing mode.
500 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
502 /// \returns The cost, if any, of keeping values of the given types alive
505 /// Some types may require the use of register classes that do not have
506 /// any callee-saved registers, so would require a spill and fill.
507 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
509 /// \returns True if the intrinsic is a supported memory intrinsic. Info
510 /// will contain additional information - whether the intrinsic may write
511 /// or read to memory, volatility and the pointer. Info is undefined
512 /// if false is returned.
513 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
515 /// \returns A value which is the result of the given memory intrinsic. New
516 /// instructions may be created to extract the result from the given intrinsic
517 /// memory operation. Returns nullptr if the target cannot create a result
518 /// from the given intrinsic.
519 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
520 Type *ExpectedType) const;
522 /// \returns True if the two functions have compatible attributes for inlining
524 bool hasCompatibleFunctionAttributes(const Function *Caller,
525 const Function *Callee) const;
530 /// \brief The abstract base class used to type erase specific TTI
534 /// \brief The template model for the base class which wraps a concrete
535 /// implementation in a type erased interface.
536 template <typename T> class Model;
538 std::unique_ptr<Concept> TTIImpl;
541 class TargetTransformInfo::Concept {
543 virtual ~Concept() = 0;
544 virtual const DataLayout &getDataLayout() const = 0;
545 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
546 virtual unsigned getGEPCost(const Value *Ptr,
547 ArrayRef<const Value *> Operands) = 0;
548 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
549 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
550 virtual unsigned getCallCost(const Function *F,
551 ArrayRef<const Value *> Arguments) = 0;
552 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
553 ArrayRef<Type *> ParamTys) = 0;
554 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
555 ArrayRef<const Value *> Arguments) = 0;
556 virtual unsigned getUserCost(const User *U) = 0;
557 virtual bool hasBranchDivergence() = 0;
558 virtual bool isSourceOfDivergence(const Value *V) = 0;
559 virtual bool isLoweredToCall(const Function *F) = 0;
560 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
561 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
562 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
563 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
564 int64_t BaseOffset, bool HasBaseReg,
566 unsigned AddrSpace) = 0;
567 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
568 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
569 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
570 int64_t BaseOffset, bool HasBaseReg,
571 int64_t Scale, unsigned AddrSpace) = 0;
572 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
573 virtual bool isProfitableToHoist(Instruction *I) = 0;
574 virtual bool isTypeLegal(Type *Ty) = 0;
575 virtual unsigned getJumpBufAlignment() = 0;
576 virtual unsigned getJumpBufSize() = 0;
577 virtual bool shouldBuildLookupTables() = 0;
578 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
579 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
580 virtual bool haveFastSqrt(Type *Ty) = 0;
581 virtual unsigned getFPOpCost(Type *Ty) = 0;
582 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
583 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
585 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
586 const APInt &Imm, Type *Ty) = 0;
587 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
588 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
589 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
591 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
592 OperandValueKind Opd2Info,
593 OperandValueProperties Opd1PropInfo,
594 OperandValueProperties Opd2PropInfo) = 0;
595 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
597 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
598 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
599 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
601 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
603 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
605 unsigned AddressSpace) = 0;
606 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
608 unsigned AddressSpace) = 0;
609 virtual unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
611 ArrayRef<unsigned> Indices,
613 unsigned AddressSpace) = 0;
614 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
615 bool IsPairwiseForm) = 0;
616 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
617 ArrayRef<Type *> Tys) = 0;
618 virtual unsigned getCallInstrCost(Function *F, Type *RetTy,
619 ArrayRef<Type *> Tys) = 0;
620 virtual unsigned getNumberOfParts(Type *Tp) = 0;
621 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
622 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
623 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
624 MemIntrinsicInfo &Info) = 0;
625 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
626 Type *ExpectedType) = 0;
627 virtual bool hasCompatibleFunctionAttributes(const Function *Caller,
628 const Function *Callee) const = 0;
631 template <typename T>
632 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
636 Model(T Impl) : Impl(std::move(Impl)) {}
639 const DataLayout &getDataLayout() const override {
640 return Impl.getDataLayout();
643 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
644 return Impl.getOperationCost(Opcode, Ty, OpTy);
646 unsigned getGEPCost(const Value *Ptr,
647 ArrayRef<const Value *> Operands) override {
648 return Impl.getGEPCost(Ptr, Operands);
650 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
651 return Impl.getCallCost(FTy, NumArgs);
653 unsigned getCallCost(const Function *F, int NumArgs) override {
654 return Impl.getCallCost(F, NumArgs);
656 unsigned getCallCost(const Function *F,
657 ArrayRef<const Value *> Arguments) override {
658 return Impl.getCallCost(F, Arguments);
660 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
661 ArrayRef<Type *> ParamTys) override {
662 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
664 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
665 ArrayRef<const Value *> Arguments) override {
666 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
668 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
669 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
670 bool isSourceOfDivergence(const Value *V) override {
671 return Impl.isSourceOfDivergence(V);
673 bool isLoweredToCall(const Function *F) override {
674 return Impl.isLoweredToCall(F);
676 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
677 return Impl.getUnrollingPreferences(L, UP);
679 bool isLegalAddImmediate(int64_t Imm) override {
680 return Impl.isLegalAddImmediate(Imm);
682 bool isLegalICmpImmediate(int64_t Imm) override {
683 return Impl.isLegalICmpImmediate(Imm);
685 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
686 bool HasBaseReg, int64_t Scale,
687 unsigned AddrSpace) override {
688 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
691 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
692 return Impl.isLegalMaskedStore(DataType, Consecutive);
694 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
695 return Impl.isLegalMaskedLoad(DataType, Consecutive);
697 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
698 bool HasBaseReg, int64_t Scale,
699 unsigned AddrSpace) override {
700 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
703 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
704 return Impl.isTruncateFree(Ty1, Ty2);
706 bool isProfitableToHoist(Instruction *I) override {
707 return Impl.isProfitableToHoist(I);
709 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
710 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
711 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
712 bool shouldBuildLookupTables() override {
713 return Impl.shouldBuildLookupTables();
715 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
716 return Impl.enableAggressiveInterleaving(LoopHasReductions);
718 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
719 return Impl.getPopcntSupport(IntTyWidthInBit);
721 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
723 unsigned getFPOpCost(Type *Ty) override {
724 return Impl.getFPOpCost(Ty);
727 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
728 return Impl.getIntImmCost(Imm, Ty);
730 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
732 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
734 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
736 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
738 unsigned getNumberOfRegisters(bool Vector) override {
739 return Impl.getNumberOfRegisters(Vector);
741 unsigned getRegisterBitWidth(bool Vector) override {
742 return Impl.getRegisterBitWidth(Vector);
744 unsigned getMaxInterleaveFactor(unsigned VF) override {
745 return Impl.getMaxInterleaveFactor(VF);
748 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
749 OperandValueKind Opd2Info,
750 OperandValueProperties Opd1PropInfo,
751 OperandValueProperties Opd2PropInfo) override {
752 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
753 Opd1PropInfo, Opd2PropInfo);
755 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
756 Type *SubTp) override {
757 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
759 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
760 return Impl.getCastInstrCost(Opcode, Dst, Src);
762 unsigned getCFInstrCost(unsigned Opcode) override {
763 return Impl.getCFInstrCost(Opcode);
765 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
766 Type *CondTy) override {
767 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
769 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
770 unsigned Index) override {
771 return Impl.getVectorInstrCost(Opcode, Val, Index);
773 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
774 unsigned AddressSpace) override {
775 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
777 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
778 unsigned AddressSpace) override {
779 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
781 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
783 ArrayRef<unsigned> Indices,
785 unsigned AddressSpace) override {
786 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
787 Alignment, AddressSpace);
789 unsigned getReductionCost(unsigned Opcode, Type *Ty,
790 bool IsPairwiseForm) override {
791 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
793 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
794 ArrayRef<Type *> Tys) override {
795 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
797 unsigned getCallInstrCost(Function *F, Type *RetTy,
798 ArrayRef<Type *> Tys) override {
799 return Impl.getCallInstrCost(F, RetTy, Tys);
801 unsigned getNumberOfParts(Type *Tp) override {
802 return Impl.getNumberOfParts(Tp);
804 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
805 return Impl.getAddressComputationCost(Ty, IsComplex);
807 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
808 return Impl.getCostOfKeepingLiveOverCall(Tys);
810 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
811 MemIntrinsicInfo &Info) override {
812 return Impl.getTgtMemIntrinsic(Inst, Info);
814 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
815 Type *ExpectedType) override {
816 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
818 bool hasCompatibleFunctionAttributes(const Function *Caller,
819 const Function *Callee) const override {
820 return Impl.hasCompatibleFunctionAttributes(Caller, Callee);
824 template <typename T>
825 TargetTransformInfo::TargetTransformInfo(T Impl)
826 : TTIImpl(new Model<T>(Impl)) {}
828 /// \brief Analysis pass providing the \c TargetTransformInfo.
830 /// The core idea of the TargetIRAnalysis is to expose an interface through
831 /// which LLVM targets can analyze and provide information about the middle
832 /// end's target-independent IR. This supports use cases such as target-aware
833 /// cost modeling of IR constructs.
835 /// This is a function analysis because much of the cost modeling for targets
836 /// is done in a subtarget specific way and LLVM supports compiling different
837 /// functions targeting different subtargets in order to support runtime
838 /// dispatch according to the observed subtarget.
839 class TargetIRAnalysis {
841 typedef TargetTransformInfo Result;
843 /// \brief Opaque, unique identifier for this analysis pass.
844 static void *ID() { return (void *)&PassID; }
846 /// \brief Provide access to a name for this pass for debugging purposes.
847 static StringRef name() { return "TargetIRAnalysis"; }
849 /// \brief Default construct a target IR analysis.
851 /// This will use the module's datalayout to construct a baseline
852 /// conservative TTI result.
855 /// \brief Construct an IR analysis pass around a target-provide callback.
857 /// The callback will be called with a particular function for which the TTI
858 /// is needed and must return a TTI object for that function.
859 TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
861 // Value semantics. We spell out the constructors for MSVC.
862 TargetIRAnalysis(const TargetIRAnalysis &Arg)
863 : TTICallback(Arg.TTICallback) {}
864 TargetIRAnalysis(TargetIRAnalysis &&Arg)
865 : TTICallback(std::move(Arg.TTICallback)) {}
866 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
867 TTICallback = RHS.TTICallback;
870 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
871 TTICallback = std::move(RHS.TTICallback);
875 Result run(Function &F);
880 /// \brief The callback used to produce a result.
882 /// We use a completely opaque callback so that targets can provide whatever
883 /// mechanism they desire for constructing the TTI for a given function.
885 /// FIXME: Should we really use std::function? It's relatively inefficient.
886 /// It might be possible to arrange for even stateful callbacks to outlive
887 /// the analysis and thus use a function_ref which would be lighter weight.
888 /// This may also be less error prone as the callback is likely to reference
889 /// the external TargetMachine, and that reference needs to never dangle.
890 std::function<Result(Function &)> TTICallback;
892 /// \brief Helper function used as the callback in the default constructor.
893 static Result getDefaultTTI(Function &F);
896 /// \brief Wrapper pass for TargetTransformInfo.
898 /// This pass can be constructed from a TTI object which it stores internally
899 /// and is queried by passes.
900 class TargetTransformInfoWrapperPass : public ImmutablePass {
901 TargetIRAnalysis TIRA;
902 Optional<TargetTransformInfo> TTI;
904 virtual void anchor();
909 /// \brief We must provide a default constructor for the pass but it should
912 /// Use the constructor below or call one of the creation routines.
913 TargetTransformInfoWrapperPass();
915 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
917 TargetTransformInfo &getTTI(Function &F);
920 /// \brief Create an analysis pass wrapper around a TTI object.
922 /// This analysis pass just holds the TTI instance and makes it available to
924 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
926 } // End llvm namespace