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(Type *PointeeType, const Value *Ptr,
140 ArrayRef<const Value *> Operands) const;
142 /// \brief Estimate the cost of a function call when lowered.
144 /// The contract for this is the same as \c getOperationCost except that it
145 /// supports an interface that provides extra information specific to call
148 /// This is the most basic query for estimating call cost: it only knows the
149 /// function type and (potentially) the number of arguments at the call site.
150 /// The latter is only interesting for varargs function types.
151 unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
153 /// \brief Estimate the cost of calling a specific function when lowered.
155 /// This overload adds the ability to reason about the particular function
156 /// being called in the event it is a library call with special lowering.
157 unsigned getCallCost(const Function *F, int NumArgs = -1) const;
159 /// \brief Estimate the cost of calling a specific function when lowered.
161 /// This overload allows specifying a set of candidate argument values.
162 unsigned getCallCost(const Function *F,
163 ArrayRef<const Value *> Arguments) const;
165 /// \brief Estimate the cost of an intrinsic when lowered.
167 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
168 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
169 ArrayRef<Type *> ParamTys) const;
171 /// \brief Estimate the cost of an intrinsic when lowered.
173 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
174 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
175 ArrayRef<const Value *> Arguments) const;
177 /// \brief Estimate the cost of a given IR user when lowered.
179 /// This can estimate the cost of either a ConstantExpr or Instruction when
180 /// lowered. It has two primary advantages over the \c getOperationCost and
181 /// \c getGEPCost above, and one significant disadvantage: it can only be
182 /// used when the IR construct has already been formed.
184 /// The advantages are that it can inspect the SSA use graph to reason more
185 /// accurately about the cost. For example, all-constant-GEPs can often be
186 /// folded into a load or other instruction, but if they are used in some
187 /// other context they may not be folded. This routine can distinguish such
190 /// The returned cost is defined in terms of \c TargetCostConstants, see its
191 /// comments for a detailed explanation of the cost values.
192 unsigned getUserCost(const User *U) const;
194 /// \brief Return true if branch divergence exists.
196 /// Branch divergence has a significantly negative impact on GPU performance
197 /// when threads in the same wavefront take different paths due to conditional
199 bool hasBranchDivergence() const;
201 /// \brief Returns whether V is a source of divergence.
203 /// This function provides the target-dependent information for
204 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
205 /// builds the dependency graph, and then runs the reachability algorithm
206 /// starting with the sources of divergence.
207 bool isSourceOfDivergence(const Value *V) const;
209 /// \brief Test whether calls to a function lower to actual program function
212 /// The idea is to test whether the program is likely to require a 'call'
213 /// instruction or equivalent in order to call the given function.
215 /// FIXME: It's not clear that this is a good or useful query API. Client's
216 /// should probably move to simpler cost metrics using the above.
217 /// Alternatively, we could split the cost interface into distinct code-size
218 /// and execution-speed costs. This would allow modelling the core of this
219 /// query more accurately as a call is a single small instruction, but
220 /// incurs significant execution cost.
221 bool isLoweredToCall(const Function *F) const;
223 /// Parameters that control the generic loop unrolling transformation.
224 struct UnrollingPreferences {
225 /// The cost threshold for the unrolled loop. Should be relative to the
226 /// getUserCost values returned by this API, and the expectation is that
227 /// the unrolled loop's instructions when run through that interface should
228 /// not exceed this cost. However, this is only an estimate. Also, specific
229 /// loops may be unrolled even with a cost above this threshold if deemed
230 /// profitable. Set this to UINT_MAX to disable the loop body cost
233 /// If complete unrolling will reduce the cost of the loop below its
234 /// expected dynamic cost while rolled by this percentage, apply a discount
235 /// (below) to its unrolled cost.
236 unsigned PercentDynamicCostSavedThreshold;
237 /// The discount applied to the unrolled cost when the *dynamic* cost
238 /// savings of unrolling exceed the \c PercentDynamicCostSavedThreshold.
239 unsigned DynamicCostSavingsDiscount;
240 /// The cost threshold for the unrolled loop when optimizing for size (set
241 /// to UINT_MAX to disable).
242 unsigned OptSizeThreshold;
243 /// The cost threshold for the unrolled loop, like Threshold, but used
244 /// for partial/runtime unrolling (set to UINT_MAX to disable).
245 unsigned PartialThreshold;
246 /// The cost threshold for the unrolled loop when optimizing for size, like
247 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
248 /// UINT_MAX to disable).
249 unsigned PartialOptSizeThreshold;
250 /// A forced unrolling factor (the number of concatenated bodies of the
251 /// original loop in the unrolled loop body). When set to 0, the unrolling
252 /// transformation will select an unrolling factor based on the current cost
253 /// threshold and other factors.
255 // Set the maximum unrolling factor. The unrolling factor may be selected
256 // using the appropriate cost threshold, but may not exceed this number
257 // (set to UINT_MAX to disable). This does not apply in cases where the
258 // loop is being fully unrolled.
260 /// Allow partial unrolling (unrolling of loops to expand the size of the
261 /// loop body, not only to eliminate small constant-trip-count loops).
263 /// Allow runtime unrolling (unrolling of loops to expand the size of the
264 /// loop body even when the number of loop iterations is not known at
267 /// Allow emitting expensive instructions (such as divisions) when computing
268 /// the trip count of a loop for runtime unrolling.
269 bool AllowExpensiveTripCount;
272 /// \brief Get target-customized preferences for the generic loop unrolling
273 /// transformation. The caller will initialize UP with the current
274 /// target-independent defaults.
275 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
279 /// \name Scalar Target Information
282 /// \brief Flags indicating the kind of support for population count.
284 /// Compared to the SW implementation, HW support is supposed to
285 /// significantly boost the performance when the population is dense, and it
286 /// may or may not degrade performance if the population is sparse. A HW
287 /// support is considered as "Fast" if it can outperform, or is on a par
288 /// with, SW implementation when the population is sparse; otherwise, it is
289 /// considered as "Slow".
290 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
292 /// \brief Return true if the specified immediate is legal add immediate, that
293 /// is the target has add instructions which can add a register with the
294 /// immediate without having to materialize the immediate into a register.
295 bool isLegalAddImmediate(int64_t Imm) const;
297 /// \brief Return true if the specified immediate is legal icmp immediate,
298 /// that is the target has icmp instructions which can compare a register
299 /// against the immediate without having to materialize the immediate into a
301 bool isLegalICmpImmediate(int64_t Imm) const;
303 /// \brief Return true if the addressing mode represented by AM is legal for
304 /// this target, for a load/store of the specified type.
305 /// The type may be VoidTy, in which case only return true if the addressing
306 /// mode is legal for a load/store of any legal type.
307 /// TODO: Handle pre/postinc as well.
308 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
309 bool HasBaseReg, int64_t Scale,
310 unsigned AddrSpace = 0) const;
312 /// \brief Return true if the target works with masked instruction
313 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
314 /// AVX-512 architecture will also allow masks for non-consecutive memory
316 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
317 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
319 /// \brief Return the cost of the scaling factor used in the addressing
320 /// mode represented by AM for this target, for a load/store
321 /// of the specified type.
322 /// If the AM is supported, the return value must be >= 0.
323 /// If the AM is not supported, it returns a negative value.
324 /// TODO: Handle pre/postinc as well.
325 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
326 bool HasBaseReg, int64_t Scale,
327 unsigned AddrSpace = 0) const;
329 /// \brief Return true if it's free to truncate a value of type Ty1 to type
330 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
331 /// by referencing its sub-register AX.
332 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
334 /// \brief Return true if it's free to zero extend a value of type Ty1 to type
335 /// Ty2. e.g. on x86-64, all instructions that define 32-bit values implicit
336 /// zero-extend the result out to 64 bits.
337 bool isZExtFree(Type *Ty1, Type *Ty2) const;
339 /// \brief Return true if it is profitable to hoist instruction in the
340 /// then/else to before if.
341 bool isProfitableToHoist(Instruction *I) const;
343 /// \brief Return true if this type is legal.
344 bool isTypeLegal(Type *Ty) const;
346 /// \brief Returns the target's jmp_buf alignment in bytes.
347 unsigned getJumpBufAlignment() const;
349 /// \brief Returns the target's jmp_buf size in bytes.
350 unsigned getJumpBufSize() const;
352 /// \brief Return true if switches should be turned into lookup tables for the
354 bool shouldBuildLookupTables() const;
356 /// \brief Don't restrict interleaved unrolling to small loops.
357 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
359 /// \brief Return hardware support for population count.
360 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
362 /// \brief Return true if the hardware has a fast square-root instruction.
363 bool haveFastSqrt(Type *Ty) const;
365 /// \brief Return the expected cost of supporting the floating point operation
366 /// of the specified type.
367 unsigned getFPOpCost(Type *Ty) const;
369 /// \brief Return the expected cost of materializing for the given integer
370 /// immediate of the specified type.
371 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
373 /// \brief Return the expected cost of materialization for the given integer
374 /// immediate of the specified type for a given instruction. The cost can be
375 /// zero if the immediate can be folded into the specified instruction.
376 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
378 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
382 /// \name Vector Target Information
385 /// \brief The various kinds of shuffle patterns for vector queries.
387 SK_Broadcast, ///< Broadcast element 0 to all other elements.
388 SK_Reverse, ///< Reverse the order of the vector.
389 SK_Alternate, ///< Choose alternate elements from vector.
390 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
391 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
394 /// \brief Additional information about an operand's possible values.
395 enum OperandValueKind {
396 OK_AnyValue, // Operand can have any value.
397 OK_UniformValue, // Operand is uniform (splat of a value).
398 OK_UniformConstantValue, // Operand is uniform constant.
399 OK_NonUniformConstantValue // Operand is a non uniform constant value.
402 /// \brief Additional properties of an operand's values.
403 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
405 /// \return The number of scalar or vector registers that the target has.
406 /// If 'Vectors' is true, it returns the number of vector registers. If it is
407 /// set to false, it returns the number of scalar registers.
408 unsigned getNumberOfRegisters(bool Vector) const;
410 /// \return The width of the largest scalar or vector register type.
411 unsigned getRegisterBitWidth(bool Vector) const;
413 /// \return The maximum interleave factor that any transform should try to
414 /// perform for this target. This number depends on the level of parallelism
415 /// and the number of execution units in the CPU.
416 unsigned getMaxInterleaveFactor(unsigned VF) const;
418 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
420 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
421 OperandValueKind Opd1Info = OK_AnyValue,
422 OperandValueKind Opd2Info = OK_AnyValue,
423 OperandValueProperties Opd1PropInfo = OP_None,
424 OperandValueProperties Opd2PropInfo = OP_None) const;
426 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
427 /// The index and subtype parameters are used by the subvector insertion and
428 /// extraction shuffle kinds.
429 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
430 Type *SubTp = nullptr) const;
432 /// \return The expected cost of cast instructions, such as bitcast, trunc,
434 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
436 /// \return The expected cost of control-flow related instructions such as
438 unsigned getCFInstrCost(unsigned Opcode) const;
440 /// \returns The expected cost of compare and select instructions.
441 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
442 Type *CondTy = nullptr) const;
444 /// \return The expected cost of vector Insert and Extract.
445 /// Use -1 to indicate that there is no information on the index value.
446 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
447 unsigned Index = -1) const;
449 /// \return The cost of Load and Store instructions.
450 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
451 unsigned AddressSpace) const;
453 /// \return The cost of masked Load and Store instructions.
454 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
455 unsigned AddressSpace) const;
457 /// \return The cost of the interleaved memory operation.
458 /// \p Opcode is the memory operation code
459 /// \p VecTy is the vector type of the interleaved access.
460 /// \p Factor is the interleave factor
461 /// \p Indices is the indices for interleaved load members (as interleaved
462 /// load allows gaps)
463 /// \p Alignment is the alignment of the memory operation
464 /// \p AddressSpace is address space of the pointer.
465 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
467 ArrayRef<unsigned> Indices,
469 unsigned AddressSpace) const;
471 /// \brief Calculate the cost of performing a vector reduction.
473 /// This is the cost of reducing the vector value of type \p Ty to a scalar
474 /// value using the operation denoted by \p Opcode. The form of the reduction
475 /// can either be a pairwise reduction or a reduction that splits the vector
476 /// at every reduction level.
480 /// ((v0+v1), (v2, v3), undef, undef)
483 /// ((v0+v2), (v1+v3), undef, undef)
484 unsigned getReductionCost(unsigned Opcode, Type *Ty,
485 bool IsPairwiseForm) const;
487 /// \returns The cost of Intrinsic instructions.
488 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
489 ArrayRef<Type *> Tys) const;
491 /// \returns The cost of Call instructions.
492 unsigned getCallInstrCost(Function *F, Type *RetTy,
493 ArrayRef<Type *> Tys) const;
495 /// \returns The number of pieces into which the provided type must be
496 /// split during legalization. Zero is returned when the answer is unknown.
497 unsigned getNumberOfParts(Type *Tp) const;
499 /// \returns The cost of the address computation. For most targets this can be
500 /// merged into the instruction indexing mode. Some targets might want to
501 /// distinguish between address computation for memory operations on vector
502 /// types and scalar types. Such targets should override this function.
503 /// The 'IsComplex' parameter is a hint that the address computation is likely
504 /// to involve multiple instructions and as such unlikely to be merged into
505 /// the address indexing mode.
506 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
508 /// \returns The cost, if any, of keeping values of the given types alive
511 /// Some types may require the use of register classes that do not have
512 /// any callee-saved registers, so would require a spill and fill.
513 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
515 /// \returns True if the intrinsic is a supported memory intrinsic. Info
516 /// will contain additional information - whether the intrinsic may write
517 /// or read to memory, volatility and the pointer. Info is undefined
518 /// if false is returned.
519 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
521 /// \returns A value which is the result of the given memory intrinsic. New
522 /// instructions may be created to extract the result from the given intrinsic
523 /// memory operation. Returns nullptr if the target cannot create a result
524 /// from the given intrinsic.
525 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
526 Type *ExpectedType) const;
528 /// \returns True if the two functions have compatible attributes for inlining
530 bool hasCompatibleFunctionAttributes(const Function *Caller,
531 const Function *Callee) const;
536 /// \brief The abstract base class used to type erase specific TTI
540 /// \brief The template model for the base class which wraps a concrete
541 /// implementation in a type erased interface.
542 template <typename T> class Model;
544 std::unique_ptr<Concept> TTIImpl;
547 class TargetTransformInfo::Concept {
549 virtual ~Concept() = 0;
550 virtual const DataLayout &getDataLayout() const = 0;
551 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
552 virtual unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
553 ArrayRef<const Value *> Operands) = 0;
554 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
555 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
556 virtual unsigned getCallCost(const Function *F,
557 ArrayRef<const Value *> Arguments) = 0;
558 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
559 ArrayRef<Type *> ParamTys) = 0;
560 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
561 ArrayRef<const Value *> Arguments) = 0;
562 virtual unsigned getUserCost(const User *U) = 0;
563 virtual bool hasBranchDivergence() = 0;
564 virtual bool isSourceOfDivergence(const Value *V) = 0;
565 virtual bool isLoweredToCall(const Function *F) = 0;
566 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
567 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
568 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
569 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
570 int64_t BaseOffset, bool HasBaseReg,
572 unsigned AddrSpace) = 0;
573 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
574 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
575 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
576 int64_t BaseOffset, bool HasBaseReg,
577 int64_t Scale, unsigned AddrSpace) = 0;
578 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
579 virtual bool isZExtFree(Type *Ty1, Type *Ty2) = 0;
580 virtual bool isProfitableToHoist(Instruction *I) = 0;
581 virtual bool isTypeLegal(Type *Ty) = 0;
582 virtual unsigned getJumpBufAlignment() = 0;
583 virtual unsigned getJumpBufSize() = 0;
584 virtual bool shouldBuildLookupTables() = 0;
585 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
586 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
587 virtual bool haveFastSqrt(Type *Ty) = 0;
588 virtual unsigned getFPOpCost(Type *Ty) = 0;
589 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
590 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
592 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
593 const APInt &Imm, Type *Ty) = 0;
594 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
595 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
596 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
598 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
599 OperandValueKind Opd2Info,
600 OperandValueProperties Opd1PropInfo,
601 OperandValueProperties Opd2PropInfo) = 0;
602 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
604 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
605 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
606 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
608 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
610 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
612 unsigned AddressSpace) = 0;
613 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
615 unsigned AddressSpace) = 0;
616 virtual unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
618 ArrayRef<unsigned> Indices,
620 unsigned AddressSpace) = 0;
621 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
622 bool IsPairwiseForm) = 0;
623 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
624 ArrayRef<Type *> Tys) = 0;
625 virtual unsigned getCallInstrCost(Function *F, Type *RetTy,
626 ArrayRef<Type *> Tys) = 0;
627 virtual unsigned getNumberOfParts(Type *Tp) = 0;
628 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
629 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
630 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
631 MemIntrinsicInfo &Info) = 0;
632 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
633 Type *ExpectedType) = 0;
634 virtual bool hasCompatibleFunctionAttributes(const Function *Caller,
635 const Function *Callee) const = 0;
638 template <typename T>
639 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
643 Model(T Impl) : Impl(std::move(Impl)) {}
646 const DataLayout &getDataLayout() const override {
647 return Impl.getDataLayout();
650 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
651 return Impl.getOperationCost(Opcode, Ty, OpTy);
653 unsigned getGEPCost(Type *PointeeType, const Value *Ptr,
654 ArrayRef<const Value *> Operands) override {
655 return Impl.getGEPCost(PointeeType, Ptr, Operands);
657 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
658 return Impl.getCallCost(FTy, NumArgs);
660 unsigned getCallCost(const Function *F, int NumArgs) override {
661 return Impl.getCallCost(F, NumArgs);
663 unsigned getCallCost(const Function *F,
664 ArrayRef<const Value *> Arguments) override {
665 return Impl.getCallCost(F, Arguments);
667 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
668 ArrayRef<Type *> ParamTys) override {
669 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
671 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
672 ArrayRef<const Value *> Arguments) override {
673 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
675 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
676 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
677 bool isSourceOfDivergence(const Value *V) override {
678 return Impl.isSourceOfDivergence(V);
680 bool isLoweredToCall(const Function *F) override {
681 return Impl.isLoweredToCall(F);
683 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
684 return Impl.getUnrollingPreferences(L, UP);
686 bool isLegalAddImmediate(int64_t Imm) override {
687 return Impl.isLegalAddImmediate(Imm);
689 bool isLegalICmpImmediate(int64_t Imm) override {
690 return Impl.isLegalICmpImmediate(Imm);
692 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
693 bool HasBaseReg, int64_t Scale,
694 unsigned AddrSpace) override {
695 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
698 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
699 return Impl.isLegalMaskedStore(DataType, Consecutive);
701 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
702 return Impl.isLegalMaskedLoad(DataType, Consecutive);
704 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
705 bool HasBaseReg, int64_t Scale,
706 unsigned AddrSpace) override {
707 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
710 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
711 return Impl.isTruncateFree(Ty1, Ty2);
713 bool isZExtFree(Type *Ty1, Type *Ty2) override {
714 return Impl.isZExtFree(Ty1, Ty2);
716 bool isProfitableToHoist(Instruction *I) override {
717 return Impl.isProfitableToHoist(I);
719 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
720 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
721 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
722 bool shouldBuildLookupTables() override {
723 return Impl.shouldBuildLookupTables();
725 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
726 return Impl.enableAggressiveInterleaving(LoopHasReductions);
728 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
729 return Impl.getPopcntSupport(IntTyWidthInBit);
731 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
733 unsigned getFPOpCost(Type *Ty) override {
734 return Impl.getFPOpCost(Ty);
737 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
738 return Impl.getIntImmCost(Imm, Ty);
740 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
742 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
744 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
746 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
748 unsigned getNumberOfRegisters(bool Vector) override {
749 return Impl.getNumberOfRegisters(Vector);
751 unsigned getRegisterBitWidth(bool Vector) override {
752 return Impl.getRegisterBitWidth(Vector);
754 unsigned getMaxInterleaveFactor(unsigned VF) override {
755 return Impl.getMaxInterleaveFactor(VF);
758 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
759 OperandValueKind Opd2Info,
760 OperandValueProperties Opd1PropInfo,
761 OperandValueProperties Opd2PropInfo) override {
762 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
763 Opd1PropInfo, Opd2PropInfo);
765 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
766 Type *SubTp) override {
767 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
769 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
770 return Impl.getCastInstrCost(Opcode, Dst, Src);
772 unsigned getCFInstrCost(unsigned Opcode) override {
773 return Impl.getCFInstrCost(Opcode);
775 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
776 Type *CondTy) override {
777 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
779 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
780 unsigned Index) override {
781 return Impl.getVectorInstrCost(Opcode, Val, Index);
783 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
784 unsigned AddressSpace) override {
785 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
787 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
788 unsigned AddressSpace) override {
789 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
791 unsigned getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
793 ArrayRef<unsigned> Indices,
795 unsigned AddressSpace) override {
796 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
797 Alignment, AddressSpace);
799 unsigned getReductionCost(unsigned Opcode, Type *Ty,
800 bool IsPairwiseForm) override {
801 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
803 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
804 ArrayRef<Type *> Tys) override {
805 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
807 unsigned getCallInstrCost(Function *F, Type *RetTy,
808 ArrayRef<Type *> Tys) override {
809 return Impl.getCallInstrCost(F, RetTy, Tys);
811 unsigned getNumberOfParts(Type *Tp) override {
812 return Impl.getNumberOfParts(Tp);
814 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
815 return Impl.getAddressComputationCost(Ty, IsComplex);
817 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
818 return Impl.getCostOfKeepingLiveOverCall(Tys);
820 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
821 MemIntrinsicInfo &Info) override {
822 return Impl.getTgtMemIntrinsic(Inst, Info);
824 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
825 Type *ExpectedType) override {
826 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
828 bool hasCompatibleFunctionAttributes(const Function *Caller,
829 const Function *Callee) const override {
830 return Impl.hasCompatibleFunctionAttributes(Caller, Callee);
834 template <typename T>
835 TargetTransformInfo::TargetTransformInfo(T Impl)
836 : TTIImpl(new Model<T>(Impl)) {}
838 /// \brief Analysis pass providing the \c TargetTransformInfo.
840 /// The core idea of the TargetIRAnalysis is to expose an interface through
841 /// which LLVM targets can analyze and provide information about the middle
842 /// end's target-independent IR. This supports use cases such as target-aware
843 /// cost modeling of IR constructs.
845 /// This is a function analysis because much of the cost modeling for targets
846 /// is done in a subtarget specific way and LLVM supports compiling different
847 /// functions targeting different subtargets in order to support runtime
848 /// dispatch according to the observed subtarget.
849 class TargetIRAnalysis {
851 typedef TargetTransformInfo Result;
853 /// \brief Opaque, unique identifier for this analysis pass.
854 static void *ID() { return (void *)&PassID; }
856 /// \brief Provide access to a name for this pass for debugging purposes.
857 static StringRef name() { return "TargetIRAnalysis"; }
859 /// \brief Default construct a target IR analysis.
861 /// This will use the module's datalayout to construct a baseline
862 /// conservative TTI result.
865 /// \brief Construct an IR analysis pass around a target-provide callback.
867 /// The callback will be called with a particular function for which the TTI
868 /// is needed and must return a TTI object for that function.
869 TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
871 // Value semantics. We spell out the constructors for MSVC.
872 TargetIRAnalysis(const TargetIRAnalysis &Arg)
873 : TTICallback(Arg.TTICallback) {}
874 TargetIRAnalysis(TargetIRAnalysis &&Arg)
875 : TTICallback(std::move(Arg.TTICallback)) {}
876 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
877 TTICallback = RHS.TTICallback;
880 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
881 TTICallback = std::move(RHS.TTICallback);
885 Result run(Function &F);
890 /// \brief The callback used to produce a result.
892 /// We use a completely opaque callback so that targets can provide whatever
893 /// mechanism they desire for constructing the TTI for a given function.
895 /// FIXME: Should we really use std::function? It's relatively inefficient.
896 /// It might be possible to arrange for even stateful callbacks to outlive
897 /// the analysis and thus use a function_ref which would be lighter weight.
898 /// This may also be less error prone as the callback is likely to reference
899 /// the external TargetMachine, and that reference needs to never dangle.
900 std::function<Result(Function &)> TTICallback;
902 /// \brief Helper function used as the callback in the default constructor.
903 static Result getDefaultTTI(Function &F);
906 /// \brief Wrapper pass for TargetTransformInfo.
908 /// This pass can be constructed from a TTI object which it stores internally
909 /// and is queried by passes.
910 class TargetTransformInfoWrapperPass : public ImmutablePass {
911 TargetIRAnalysis TIRA;
912 Optional<TargetTransformInfo> TTI;
914 virtual void anchor();
919 /// \brief We must provide a default constructor for the pass but it should
922 /// Use the constructor below or call one of the creation routines.
923 TargetTransformInfoWrapperPass();
925 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
927 TargetTransformInfo &getTTI(Function &F);
930 /// \brief Create an analysis pass wrapper around a TTI object.
932 /// This analysis pass just holds the TTI instance and makes it available to
934 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
936 } // End llvm namespace