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 hasBranchDivergence - Return true if branch divergence exists.
194 /// Branch divergence has a significantly negative impact on GPU performance
195 /// when threads in the same wavefront take different paths due to conditional
197 bool hasBranchDivergence() const;
199 /// \brief Test whether calls to a function lower to actual program function
202 /// The idea is to test whether the program is likely to require a 'call'
203 /// instruction or equivalent in order to call the given function.
205 /// FIXME: It's not clear that this is a good or useful query API. Client's
206 /// should probably move to simpler cost metrics using the above.
207 /// Alternatively, we could split the cost interface into distinct code-size
208 /// and execution-speed costs. This would allow modelling the core of this
209 /// query more accurately as a call is a single small instruction, but
210 /// incurs significant execution cost.
211 bool isLoweredToCall(const Function *F) const;
213 /// Parameters that control the generic loop unrolling transformation.
214 struct UnrollingPreferences {
215 /// The cost threshold for the unrolled loop, compared to
216 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
217 /// The unrolling factor is set such that the unrolled loop body does not
218 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
221 /// If complete unrolling could help other optimizations (e.g. InstSimplify)
222 /// to remove N% of instructions, then we can go beyond unroll threshold.
223 /// This value set the minimal percent for allowing that.
224 unsigned MinPercentOfOptimized;
225 /// The absolute cost threshold. We won't go beyond this even if complete
226 /// unrolling could result in optimizing out 90% of instructions.
227 unsigned AbsoluteThreshold;
228 /// The cost threshold for the unrolled loop when optimizing for size (set
229 /// to UINT_MAX to disable).
230 unsigned OptSizeThreshold;
231 /// The cost threshold for the unrolled loop, like Threshold, but used
232 /// for partial/runtime unrolling (set to UINT_MAX to disable).
233 unsigned PartialThreshold;
234 /// The cost threshold for the unrolled loop when optimizing for size, like
235 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
236 /// UINT_MAX to disable).
237 unsigned PartialOptSizeThreshold;
238 /// A forced unrolling factor (the number of concatenated bodies of the
239 /// original loop in the unrolled loop body). When set to 0, the unrolling
240 /// transformation will select an unrolling factor based on the current cost
241 /// threshold and other factors.
243 // Set the maximum unrolling factor. The unrolling factor may be selected
244 // using the appropriate cost threshold, but may not exceed this number
245 // (set to UINT_MAX to disable). This does not apply in cases where the
246 // loop is being fully unrolled.
248 /// Allow partial unrolling (unrolling of loops to expand the size of the
249 /// loop body, not only to eliminate small constant-trip-count loops).
251 /// Allow runtime unrolling (unrolling of loops to expand the size of the
252 /// loop body even when the number of loop iterations is not known at
257 /// \brief Get target-customized preferences for the generic loop unrolling
258 /// transformation. The caller will initialize UP with the current
259 /// target-independent defaults.
260 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
264 /// \name Scalar Target Information
267 /// \brief Flags indicating the kind of support for population count.
269 /// Compared to the SW implementation, HW support is supposed to
270 /// significantly boost the performance when the population is dense, and it
271 /// may or may not degrade performance if the population is sparse. A HW
272 /// support is considered as "Fast" if it can outperform, or is on a par
273 /// with, SW implementation when the population is sparse; otherwise, it is
274 /// considered as "Slow".
275 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
277 /// \brief Return true if the specified immediate is legal add immediate, that
278 /// is the target has add instructions which can add a register with the
279 /// immediate without having to materialize the immediate into a register.
280 bool isLegalAddImmediate(int64_t Imm) const;
282 /// \brief Return true if the specified immediate is legal icmp immediate,
283 /// that is the target has icmp instructions which can compare a register
284 /// against the immediate without having to materialize the immediate into a
286 bool isLegalICmpImmediate(int64_t Imm) const;
288 /// \brief Return true if the addressing mode represented by AM is legal for
289 /// this target, for a load/store of the specified type.
290 /// The type may be VoidTy, in which case only return true if the addressing
291 /// mode is legal for a load/store of any legal type.
292 /// TODO: Handle pre/postinc as well.
293 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
294 bool HasBaseReg, int64_t Scale) const;
296 /// \brief Return true if the target works with masked instruction
297 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
298 /// AVX-512 architecture will also allow masks for non-consecutive memory
300 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
301 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
303 /// \brief Return the cost of the scaling factor used in the addressing
304 /// mode represented by AM for this target, for a load/store
305 /// of the specified type.
306 /// If the AM is supported, the return value must be >= 0.
307 /// If the AM is not supported, it returns a negative value.
308 /// TODO: Handle pre/postinc as well.
309 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
310 bool HasBaseReg, int64_t Scale) const;
312 /// \brief Return true if it's free to truncate a value of type Ty1 to type
313 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
314 /// by referencing its sub-register AX.
315 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
317 /// \brief Return true if it is profitable to hoist instruction in the
318 /// then/else to before if.
319 bool isProfitableToHoist(Instruction *I) const;
321 /// \brief Return true if this type is legal.
322 bool isTypeLegal(Type *Ty) const;
324 /// \brief Returns the target's jmp_buf alignment in bytes.
325 unsigned getJumpBufAlignment() const;
327 /// \brief Returns the target's jmp_buf size in bytes.
328 unsigned getJumpBufSize() const;
330 /// \brief Return true if switches should be turned into lookup tables for the
332 bool shouldBuildLookupTables() const;
334 /// \brief Return hardware support for population count.
335 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
337 /// \brief Return true if the hardware has a fast square-root instruction.
338 bool haveFastSqrt(Type *Ty) const;
340 /// \brief Return the expected cost of supporting the floating point operation
341 /// of the specified type.
342 unsigned getFPOpCost(Type *Ty) const;
344 /// \brief Return the expected cost of materializing for the given integer
345 /// immediate of the specified type.
346 unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
348 /// \brief Return the expected cost of materialization for the given integer
349 /// immediate of the specified type for a given instruction. The cost can be
350 /// zero if the immediate can be folded into the specified instruction.
351 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
353 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
357 /// \name Vector Target Information
360 /// \brief The various kinds of shuffle patterns for vector queries.
362 SK_Broadcast, ///< Broadcast element 0 to all other elements.
363 SK_Reverse, ///< Reverse the order of the vector.
364 SK_Alternate, ///< Choose alternate elements from vector.
365 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
366 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
369 /// \brief Additional information about an operand's possible values.
370 enum OperandValueKind {
371 OK_AnyValue, // Operand can have any value.
372 OK_UniformValue, // Operand is uniform (splat of a value).
373 OK_UniformConstantValue, // Operand is uniform constant.
374 OK_NonUniformConstantValue // Operand is a non uniform constant value.
377 /// \brief Additional properties of an operand's values.
378 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
380 /// \return The number of scalar or vector registers that the target has.
381 /// If 'Vectors' is true, it returns the number of vector registers. If it is
382 /// set to false, it returns the number of scalar registers.
383 unsigned getNumberOfRegisters(bool Vector) const;
385 /// \return The width of the largest scalar or vector register type.
386 unsigned getRegisterBitWidth(bool Vector) const;
388 /// \return The maximum interleave factor that any transform should try to
389 /// perform for this target. This number depends on the level of parallelism
390 /// and the number of execution units in the CPU.
391 unsigned getMaxInterleaveFactor() const;
393 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
395 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
396 OperandValueKind Opd1Info = OK_AnyValue,
397 OperandValueKind Opd2Info = OK_AnyValue,
398 OperandValueProperties Opd1PropInfo = OP_None,
399 OperandValueProperties Opd2PropInfo = OP_None) const;
401 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
402 /// The index and subtype parameters are used by the subvector insertion and
403 /// extraction shuffle kinds.
404 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
405 Type *SubTp = nullptr) const;
407 /// \return The expected cost of cast instructions, such as bitcast, trunc,
409 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
411 /// \return The expected cost of control-flow related instructions such as
413 unsigned getCFInstrCost(unsigned Opcode) const;
415 /// \returns The expected cost of compare and select instructions.
416 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
417 Type *CondTy = nullptr) const;
419 /// \return The expected cost of vector Insert and Extract.
420 /// Use -1 to indicate that there is no information on the index value.
421 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
422 unsigned Index = -1) const;
424 /// \return The cost of Load and Store instructions.
425 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
426 unsigned AddressSpace) const;
428 /// \return The cost of masked Load and Store instructions.
429 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
430 unsigned AddressSpace) const;
432 /// \brief Calculate the cost of performing a vector reduction.
434 /// This is the cost of reducing the vector value of type \p Ty to a scalar
435 /// value using the operation denoted by \p Opcode. The form of the reduction
436 /// can either be a pairwise reduction or a reduction that splits the vector
437 /// at every reduction level.
441 /// ((v0+v1), (v2, v3), undef, undef)
444 /// ((v0+v2), (v1+v3), undef, undef)
445 unsigned getReductionCost(unsigned Opcode, Type *Ty,
446 bool IsPairwiseForm) const;
448 /// \returns The cost of Intrinsic instructions.
449 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
450 ArrayRef<Type *> Tys) const;
452 /// \returns The number of pieces into which the provided type must be
453 /// split during legalization. Zero is returned when the answer is unknown.
454 unsigned getNumberOfParts(Type *Tp) const;
456 /// \returns The cost of the address computation. For most targets this can be
457 /// merged into the instruction indexing mode. Some targets might want to
458 /// distinguish between address computation for memory operations on vector
459 /// types and scalar types. Such targets should override this function.
460 /// The 'IsComplex' parameter is a hint that the address computation is likely
461 /// to involve multiple instructions and as such unlikely to be merged into
462 /// the address indexing mode.
463 unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
465 /// \returns The cost, if any, of keeping values of the given types alive
468 /// Some types may require the use of register classes that do not have
469 /// any callee-saved registers, so would require a spill and fill.
470 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
472 /// \returns True if the intrinsic is a supported memory intrinsic. Info
473 /// will contain additional information - whether the intrinsic may write
474 /// or read to memory, volatility and the pointer. Info is undefined
475 /// if false is returned.
476 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
478 /// \returns A value which is the result of the given memory intrinsic. New
479 /// instructions may be created to extract the result from the given intrinsic
480 /// memory operation. Returns nullptr if the target cannot create a result
481 /// from the given intrinsic.
482 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
483 Type *ExpectedType) const;
488 /// \brief The abstract base class used to type erase specific TTI
492 /// \brief The template model for the base class which wraps a concrete
493 /// implementation in a type erased interface.
494 template <typename T> class Model;
496 std::unique_ptr<Concept> TTIImpl;
499 class TargetTransformInfo::Concept {
501 virtual ~Concept() = 0;
503 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
504 virtual unsigned getGEPCost(const Value *Ptr,
505 ArrayRef<const Value *> Operands) = 0;
506 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
507 virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
508 virtual unsigned getCallCost(const Function *F,
509 ArrayRef<const Value *> Arguments) = 0;
510 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
511 ArrayRef<Type *> ParamTys) = 0;
512 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
513 ArrayRef<const Value *> Arguments) = 0;
514 virtual unsigned getUserCost(const User *U) = 0;
515 virtual bool hasBranchDivergence() = 0;
516 virtual bool isLoweredToCall(const Function *F) = 0;
517 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
518 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
519 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
520 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
521 int64_t BaseOffset, bool HasBaseReg,
523 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
524 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
525 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
526 int64_t BaseOffset, bool HasBaseReg,
528 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
529 virtual bool isProfitableToHoist(Instruction *I) = 0;
530 virtual bool isTypeLegal(Type *Ty) = 0;
531 virtual unsigned getJumpBufAlignment() = 0;
532 virtual unsigned getJumpBufSize() = 0;
533 virtual bool shouldBuildLookupTables() = 0;
534 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
535 virtual bool haveFastSqrt(Type *Ty) = 0;
536 virtual unsigned getFPOpCost(Type *Ty) = 0;
537 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
538 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
540 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
541 const APInt &Imm, Type *Ty) = 0;
542 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
543 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
544 virtual unsigned getMaxInterleaveFactor() = 0;
546 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
547 OperandValueKind Opd2Info,
548 OperandValueProperties Opd1PropInfo,
549 OperandValueProperties Opd2PropInfo) = 0;
550 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
552 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
553 virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
554 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
556 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
558 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
560 unsigned AddressSpace) = 0;
561 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
563 unsigned AddressSpace) = 0;
564 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
565 bool IsPairwiseForm) = 0;
566 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
567 ArrayRef<Type *> Tys) = 0;
568 virtual unsigned getNumberOfParts(Type *Tp) = 0;
569 virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
570 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
571 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
572 MemIntrinsicInfo &Info) = 0;
573 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
574 Type *ExpectedType) = 0;
577 template <typename T>
578 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
582 Model(T Impl) : Impl(std::move(Impl)) {}
585 unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
586 return Impl.getOperationCost(Opcode, Ty, OpTy);
588 unsigned getGEPCost(const Value *Ptr,
589 ArrayRef<const Value *> Operands) override {
590 return Impl.getGEPCost(Ptr, Operands);
592 unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
593 return Impl.getCallCost(FTy, NumArgs);
595 unsigned getCallCost(const Function *F, int NumArgs) override {
596 return Impl.getCallCost(F, NumArgs);
598 unsigned getCallCost(const Function *F,
599 ArrayRef<const Value *> Arguments) override {
600 return Impl.getCallCost(F, Arguments);
602 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
603 ArrayRef<Type *> ParamTys) override {
604 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
606 unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
607 ArrayRef<const Value *> Arguments) override {
608 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
610 unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
611 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
612 bool isLoweredToCall(const Function *F) override {
613 return Impl.isLoweredToCall(F);
615 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
616 return Impl.getUnrollingPreferences(L, UP);
618 bool isLegalAddImmediate(int64_t Imm) override {
619 return Impl.isLegalAddImmediate(Imm);
621 bool isLegalICmpImmediate(int64_t Imm) override {
622 return Impl.isLegalICmpImmediate(Imm);
624 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
625 bool HasBaseReg, int64_t Scale) override {
626 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
629 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
630 return Impl.isLegalMaskedStore(DataType, Consecutive);
632 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
633 return Impl.isLegalMaskedLoad(DataType, Consecutive);
635 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
636 bool HasBaseReg, int64_t Scale) override {
637 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
639 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
640 return Impl.isTruncateFree(Ty1, Ty2);
642 bool isProfitableToHoist(Instruction *I) override {
643 return Impl.isProfitableToHoist(I);
645 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
646 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
647 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
648 bool shouldBuildLookupTables() override {
649 return Impl.shouldBuildLookupTables();
651 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
652 return Impl.getPopcntSupport(IntTyWidthInBit);
654 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
656 unsigned getFPOpCost(Type *Ty) override {
657 return Impl.getFPOpCost(Ty);
660 unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
661 return Impl.getIntImmCost(Imm, Ty);
663 unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
665 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
667 unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
669 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
671 unsigned getNumberOfRegisters(bool Vector) override {
672 return Impl.getNumberOfRegisters(Vector);
674 unsigned getRegisterBitWidth(bool Vector) override {
675 return Impl.getRegisterBitWidth(Vector);
677 unsigned getMaxInterleaveFactor() override {
678 return Impl.getMaxInterleaveFactor();
681 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
682 OperandValueKind Opd2Info,
683 OperandValueProperties Opd1PropInfo,
684 OperandValueProperties Opd2PropInfo) override {
685 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
686 Opd1PropInfo, Opd2PropInfo);
688 unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
689 Type *SubTp) override {
690 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
692 unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
693 return Impl.getCastInstrCost(Opcode, Dst, Src);
695 unsigned getCFInstrCost(unsigned Opcode) override {
696 return Impl.getCFInstrCost(Opcode);
698 unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
699 Type *CondTy) override {
700 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
702 unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
703 unsigned Index) override {
704 return Impl.getVectorInstrCost(Opcode, Val, Index);
706 unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
707 unsigned AddressSpace) override {
708 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
710 unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
711 unsigned AddressSpace) override {
712 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
714 unsigned getReductionCost(unsigned Opcode, Type *Ty,
715 bool IsPairwiseForm) override {
716 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
718 unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
719 ArrayRef<Type *> Tys) override {
720 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
722 unsigned getNumberOfParts(Type *Tp) override {
723 return Impl.getNumberOfParts(Tp);
725 unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
726 return Impl.getAddressComputationCost(Ty, IsComplex);
728 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
729 return Impl.getCostOfKeepingLiveOverCall(Tys);
731 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
732 MemIntrinsicInfo &Info) override {
733 return Impl.getTgtMemIntrinsic(Inst, Info);
735 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
736 Type *ExpectedType) override {
737 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
741 template <typename T>
742 TargetTransformInfo::TargetTransformInfo(T Impl)
743 : TTIImpl(new Model<T>(Impl)) {}
745 /// \brief Analysis pass providing the \c TargetTransformInfo.
747 /// The core idea of the TargetIRAnalysis is to expose an interface through
748 /// which LLVM targets can analyze and provide information about the middle
749 /// end's target-independent IR. This supports use cases such as target-aware
750 /// cost modeling of IR constructs.
752 /// This is a function analysis because much of the cost modeling for targets
753 /// is done in a subtarget specific way and LLVM supports compiling different
754 /// functions targeting different subtargets in order to support runtime
755 /// dispatch according to the observed subtarget.
756 class TargetIRAnalysis {
758 typedef TargetTransformInfo Result;
760 /// \brief Opaque, unique identifier for this analysis pass.
761 static void *ID() { return (void *)&PassID; }
763 /// \brief Provide access to a name for this pass for debugging purposes.
764 static StringRef name() { return "TargetIRAnalysis"; }
766 /// \brief Default construct a target IR analysis.
768 /// This will use the module's datalayout to construct a baseline
769 /// conservative TTI result.
772 /// \brief Construct an IR analysis pass around a target-provide callback.
774 /// The callback will be called with a particular function for which the TTI
775 /// is needed and must return a TTI object for that function.
776 TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
778 // Value semantics. We spell out the constructors for MSVC.
779 TargetIRAnalysis(const TargetIRAnalysis &Arg)
780 : TTICallback(Arg.TTICallback) {}
781 TargetIRAnalysis(TargetIRAnalysis &&Arg)
782 : TTICallback(std::move(Arg.TTICallback)) {}
783 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
784 TTICallback = RHS.TTICallback;
787 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
788 TTICallback = std::move(RHS.TTICallback);
792 Result run(Function &F);
797 /// \brief The callback used to produce a result.
799 /// We use a completely opaque callback so that targets can provide whatever
800 /// mechanism they desire for constructing the TTI for a given function.
802 /// FIXME: Should we really use std::function? It's relatively inefficient.
803 /// It might be possible to arrange for even stateful callbacks to outlive
804 /// the analysis and thus use a function_ref which would be lighter weight.
805 /// This may also be less error prone as the callback is likely to reference
806 /// the external TargetMachine, and that reference needs to never dangle.
807 std::function<Result(Function &)> TTICallback;
809 /// \brief Helper function used as the callback in the default constructor.
810 static Result getDefaultTTI(Function &F);
813 /// \brief Wrapper pass for TargetTransformInfo.
815 /// This pass can be constructed from a TTI object which it stores internally
816 /// and is queried by passes.
817 class TargetTransformInfoWrapperPass : public ImmutablePass {
818 TargetIRAnalysis TIRA;
819 Optional<TargetTransformInfo> TTI;
821 virtual void anchor();
826 /// \brief We must provide a default constructor for the pass but it should
829 /// Use the constructor below or call one of the creation routines.
830 TargetTransformInfoWrapperPass();
832 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
834 TargetTransformInfo &getTTI(Function &F);
837 /// \brief Create an analysis pass wrapper around a TTI object.
839 /// This analysis pass just holds the TTI instance and makes it available to
841 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
843 } // End llvm namespace