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 int 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.
105 /// Also note that the returned costs are signed integers to make it natural
106 /// to add, subtract, and test with zero (a common boundary condition). It is
107 /// not expected that 2^32 is a realistic cost to be modeling at any point.
109 /// Note that these costs should usually reflect the intersection of code-size
110 /// cost and execution cost. A free instruction is typically one that folds
111 /// into another instruction. For example, reg-to-reg moves can often be
112 /// skipped by renaming the registers in the CPU, but they still are encoded
113 /// and thus wouldn't be considered 'free' here.
114 enum TargetCostConstants {
115 TCC_Free = 0, ///< Expected to fold away in lowering.
116 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
117 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
120 /// \brief Estimate the cost of a specific operation when lowered.
122 /// Note that this is designed to work on an arbitrary synthetic opcode, and
123 /// thus work for hypothetical queries before an instruction has even been
124 /// formed. However, this does *not* work for GEPs, and must not be called
125 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
126 /// analyzing a GEP's cost required more information.
128 /// Typically only the result type is required, and the operand type can be
129 /// omitted. However, if the opcode is one of the cast instructions, the
130 /// operand type is required.
132 /// The returned cost is defined in terms of \c TargetCostConstants, see its
133 /// comments for a detailed explanation of the cost values.
134 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy = nullptr) const;
136 /// \brief Estimate the cost of a GEP operation when lowered.
138 /// The contract for this function is the same as \c getOperationCost except
139 /// that it supports an interface that provides extra information specific to
140 /// the GEP operation.
141 int getGEPCost(Type *PointeeType, const Value *Ptr,
142 ArrayRef<const Value *> Operands) const;
144 /// \brief Estimate the cost of a function call when lowered.
146 /// The contract for this is the same as \c getOperationCost except that it
147 /// supports an interface that provides extra information specific to call
150 /// This is the most basic query for estimating call cost: it only knows the
151 /// function type and (potentially) the number of arguments at the call site.
152 /// The latter is only interesting for varargs function types.
153 int getCallCost(FunctionType *FTy, int NumArgs = -1) const;
155 /// \brief Estimate the cost of calling a specific function when lowered.
157 /// This overload adds the ability to reason about the particular function
158 /// being called in the event it is a library call with special lowering.
159 int getCallCost(const Function *F, int NumArgs = -1) const;
161 /// \brief Estimate the cost of calling a specific function when lowered.
163 /// This overload allows specifying a set of candidate argument values.
164 int getCallCost(const Function *F, ArrayRef<const Value *> Arguments) const;
166 /// \brief Estimate the cost of an intrinsic when lowered.
168 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
169 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
170 ArrayRef<Type *> ParamTys) const;
172 /// \brief Estimate the cost of an intrinsic when lowered.
174 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
175 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
176 ArrayRef<const Value *> Arguments) const;
178 /// \brief Estimate the cost of a given IR user when lowered.
180 /// This can estimate the cost of either a ConstantExpr or Instruction when
181 /// lowered. It has two primary advantages over the \c getOperationCost and
182 /// \c getGEPCost above, and one significant disadvantage: it can only be
183 /// used when the IR construct has already been formed.
185 /// The advantages are that it can inspect the SSA use graph to reason more
186 /// accurately about the cost. For example, all-constant-GEPs can often be
187 /// folded into a load or other instruction, but if they are used in some
188 /// other context they may not be folded. This routine can distinguish such
191 /// The returned cost is defined in terms of \c TargetCostConstants, see its
192 /// comments for a detailed explanation of the cost values.
193 int getUserCost(const User *U) const;
195 /// \brief Return true if branch divergence exists.
197 /// Branch divergence has a significantly negative impact on GPU performance
198 /// when threads in the same wavefront take different paths due to conditional
200 bool hasBranchDivergence() const;
202 /// \brief Returns whether V is a source of divergence.
204 /// This function provides the target-dependent information for
205 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
206 /// builds the dependency graph, and then runs the reachability algorithm
207 /// starting with the sources of divergence.
208 bool isSourceOfDivergence(const Value *V) const;
210 /// \brief Test whether calls to a function lower to actual program function
213 /// The idea is to test whether the program is likely to require a 'call'
214 /// instruction or equivalent in order to call the given function.
216 /// FIXME: It's not clear that this is a good or useful query API. Client's
217 /// should probably move to simpler cost metrics using the above.
218 /// Alternatively, we could split the cost interface into distinct code-size
219 /// and execution-speed costs. This would allow modelling the core of this
220 /// query more accurately as a call is a single small instruction, but
221 /// incurs significant execution cost.
222 bool isLoweredToCall(const Function *F) const;
224 /// Parameters that control the generic loop unrolling transformation.
225 struct UnrollingPreferences {
226 /// The cost threshold for the unrolled loop. Should be relative to the
227 /// getUserCost values returned by this API, and the expectation is that
228 /// the unrolled loop's instructions when run through that interface should
229 /// not exceed this cost. However, this is only an estimate. Also, specific
230 /// loops may be unrolled even with a cost above this threshold if deemed
231 /// profitable. Set this to UINT_MAX to disable the loop body cost
234 /// If complete unrolling will reduce the cost of the loop below its
235 /// expected dynamic cost while rolled by this percentage, apply a discount
236 /// (below) to its unrolled cost.
237 unsigned PercentDynamicCostSavedThreshold;
238 /// The discount applied to the unrolled cost when the *dynamic* cost
239 /// savings of unrolling exceed the \c PercentDynamicCostSavedThreshold.
240 unsigned DynamicCostSavingsDiscount;
241 /// The cost threshold for the unrolled loop when optimizing for size (set
242 /// to UINT_MAX to disable).
243 unsigned OptSizeThreshold;
244 /// The cost threshold for the unrolled loop, like Threshold, but used
245 /// for partial/runtime unrolling (set to UINT_MAX to disable).
246 unsigned PartialThreshold;
247 /// The cost threshold for the unrolled loop when optimizing for size, like
248 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
249 /// UINT_MAX to disable).
250 unsigned PartialOptSizeThreshold;
251 /// A forced unrolling factor (the number of concatenated bodies of the
252 /// original loop in the unrolled loop body). When set to 0, the unrolling
253 /// transformation will select an unrolling factor based on the current cost
254 /// threshold and other factors.
256 // Set the maximum unrolling factor. The unrolling factor may be selected
257 // using the appropriate cost threshold, but may not exceed this number
258 // (set to UINT_MAX to disable). This does not apply in cases where the
259 // loop is being fully unrolled.
261 /// Allow partial unrolling (unrolling of loops to expand the size of the
262 /// loop body, not only to eliminate small constant-trip-count loops).
264 /// Allow runtime unrolling (unrolling of loops to expand the size of the
265 /// loop body even when the number of loop iterations is not known at
268 /// Allow emitting expensive instructions (such as divisions) when computing
269 /// the trip count of a loop for runtime unrolling.
270 bool AllowExpensiveTripCount;
273 /// \brief Get target-customized preferences for the generic loop unrolling
274 /// transformation. The caller will initialize UP with the current
275 /// target-independent defaults.
276 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
280 /// \name Scalar Target Information
283 /// \brief Flags indicating the kind of support for population count.
285 /// Compared to the SW implementation, HW support is supposed to
286 /// significantly boost the performance when the population is dense, and it
287 /// may or may not degrade performance if the population is sparse. A HW
288 /// support is considered as "Fast" if it can outperform, or is on a par
289 /// with, SW implementation when the population is sparse; otherwise, it is
290 /// considered as "Slow".
291 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
293 /// \brief Return true if the specified immediate is legal add immediate, that
294 /// is the target has add instructions which can add a register with the
295 /// immediate without having to materialize the immediate into a register.
296 bool isLegalAddImmediate(int64_t Imm) const;
298 /// \brief Return true if the specified immediate is legal icmp immediate,
299 /// that is the target has icmp instructions which can compare a register
300 /// against the immediate without having to materialize the immediate into a
302 bool isLegalICmpImmediate(int64_t Imm) const;
304 /// \brief Return true if the addressing mode represented by AM is legal for
305 /// this target, for a load/store of the specified type.
306 /// The type may be VoidTy, in which case only return true if the addressing
307 /// mode is legal for a load/store of any legal type.
308 /// TODO: Handle pre/postinc as well.
309 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
310 bool HasBaseReg, int64_t Scale,
311 unsigned AddrSpace = 0) const;
313 /// \brief Return true if the target works with masked instruction
314 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
315 /// AVX-512 architecture will also allow masks for non-consecutive memory
317 bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
318 bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
320 /// \brief Return the cost of the scaling factor used in the addressing
321 /// mode represented by AM for this target, for a load/store
322 /// of the specified type.
323 /// If the AM is supported, the return value must be >= 0.
324 /// If the AM is not supported, it returns a negative value.
325 /// TODO: Handle pre/postinc as well.
326 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
327 bool HasBaseReg, int64_t Scale,
328 unsigned AddrSpace = 0) const;
330 /// \brief Return true if it's free to truncate a value of type Ty1 to type
331 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
332 /// by referencing its sub-register AX.
333 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
335 /// \brief Return true if it's free to zero extend a value of type Ty1 to type
336 /// Ty2. e.g. on x86-64, all instructions that define 32-bit values implicit
337 /// zero-extend the result out to 64 bits.
338 bool isZExtFree(Type *Ty1, Type *Ty2) const;
340 /// \brief Return true if it is profitable to hoist instruction in the
341 /// then/else to before if.
342 bool isProfitableToHoist(Instruction *I) const;
344 /// \brief Return true if this type is legal.
345 bool isTypeLegal(Type *Ty) const;
347 /// \brief Returns the target's jmp_buf alignment in bytes.
348 unsigned getJumpBufAlignment() const;
350 /// \brief Returns the target's jmp_buf size in bytes.
351 unsigned getJumpBufSize() const;
353 /// \brief Return true if switches should be turned into lookup tables for the
355 bool shouldBuildLookupTables() const;
357 /// \brief Don't restrict interleaved unrolling to small loops.
358 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
360 /// \brief Enable matching of interleaved access groups.
361 bool enableInterleavedAccessVectorization() const;
363 /// \brief Return hardware support for population count.
364 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
366 /// \brief Return true if the hardware has a fast square-root instruction.
367 bool haveFastSqrt(Type *Ty) const;
369 /// \brief Return the expected cost of supporting the floating point operation
370 /// of the specified type.
371 int getFPOpCost(Type *Ty) const;
373 /// \brief Return the expected cost of materializing for the given integer
374 /// immediate of the specified type.
375 int getIntImmCost(const APInt &Imm, Type *Ty) const;
377 /// \brief Return the expected cost of materialization for the given integer
378 /// immediate of the specified type for a given instruction. The cost can be
379 /// zero if the immediate can be folded into the specified instruction.
380 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
382 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
386 /// \name Vector Target Information
389 /// \brief The various kinds of shuffle patterns for vector queries.
391 SK_Broadcast, ///< Broadcast element 0 to all other elements.
392 SK_Reverse, ///< Reverse the order of the vector.
393 SK_Alternate, ///< Choose alternate elements from vector.
394 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
395 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
398 /// \brief Additional information about an operand's possible values.
399 enum OperandValueKind {
400 OK_AnyValue, // Operand can have any value.
401 OK_UniformValue, // Operand is uniform (splat of a value).
402 OK_UniformConstantValue, // Operand is uniform constant.
403 OK_NonUniformConstantValue // Operand is a non uniform constant value.
406 /// \brief Additional properties of an operand's values.
407 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
409 /// \return The number of scalar or vector registers that the target has.
410 /// If 'Vectors' is true, it returns the number of vector registers. If it is
411 /// set to false, it returns the number of scalar registers.
412 unsigned getNumberOfRegisters(bool Vector) const;
414 /// \return The width of the largest scalar or vector register type.
415 unsigned getRegisterBitWidth(bool Vector) const;
417 /// \return The maximum interleave factor that any transform should try to
418 /// perform for this target. This number depends on the level of parallelism
419 /// and the number of execution units in the CPU.
420 unsigned getMaxInterleaveFactor(unsigned VF) const;
422 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
423 int getArithmeticInstrCost(
424 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info = OK_AnyValue,
425 OperandValueKind Opd2Info = OK_AnyValue,
426 OperandValueProperties Opd1PropInfo = OP_None,
427 OperandValueProperties Opd2PropInfo = OP_None) const;
429 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
430 /// The index and subtype parameters are used by the subvector insertion and
431 /// extraction shuffle kinds.
432 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
433 Type *SubTp = nullptr) const;
435 /// \return The expected cost of cast instructions, such as bitcast, trunc,
437 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
439 /// \return The expected cost of control-flow related instructions such as
441 int getCFInstrCost(unsigned Opcode) const;
443 /// \returns The expected cost of compare and select instructions.
444 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
445 Type *CondTy = nullptr) const;
447 /// \return The expected cost of vector Insert and Extract.
448 /// Use -1 to indicate that there is no information on the index value.
449 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index = -1) const;
451 /// \return The cost of Load and Store instructions.
452 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
453 unsigned AddressSpace) const;
455 /// \return The cost of masked Load and Store instructions.
456 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
457 unsigned AddressSpace) const;
459 /// \return The cost of the interleaved memory operation.
460 /// \p Opcode is the memory operation code
461 /// \p VecTy is the vector type of the interleaved access.
462 /// \p Factor is the interleave factor
463 /// \p Indices is the indices for interleaved load members (as interleaved
464 /// load allows gaps)
465 /// \p Alignment is the alignment of the memory operation
466 /// \p AddressSpace is address space of the pointer.
467 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
468 ArrayRef<unsigned> Indices, unsigned Alignment,
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 int getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwiseForm) const;
486 /// \returns The cost of Intrinsic instructions.
487 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
488 ArrayRef<Type *> Tys) const;
490 /// \returns The cost of Call instructions.
491 int getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) const;
493 /// \returns The number of pieces into which the provided type must be
494 /// split during legalization. Zero is returned when the answer is unknown.
495 unsigned getNumberOfParts(Type *Tp) const;
497 /// \returns The cost of the address computation. For most targets this can be
498 /// merged into the instruction indexing mode. Some targets might want to
499 /// distinguish between address computation for memory operations on vector
500 /// types and scalar types. Such targets should override this function.
501 /// The 'IsComplex' parameter is a hint that the address computation is likely
502 /// to involve multiple instructions and as such unlikely to be merged into
503 /// the address indexing mode.
504 int getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
506 /// \returns The cost, if any, of keeping values of the given types alive
509 /// Some types may require the use of register classes that do not have
510 /// any callee-saved registers, so would require a spill and fill.
511 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
513 /// \returns True if the intrinsic is a supported memory intrinsic. Info
514 /// will contain additional information - whether the intrinsic may write
515 /// or read to memory, volatility and the pointer. Info is undefined
516 /// if false is returned.
517 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
519 /// \returns A value which is the result of the given memory intrinsic. New
520 /// instructions may be created to extract the result from the given intrinsic
521 /// memory operation. Returns nullptr if the target cannot create a result
522 /// from the given intrinsic.
523 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
524 Type *ExpectedType) const;
526 /// \returns True if the two functions have compatible attributes for inlining
528 bool areInlineCompatible(const Function *Caller,
529 const Function *Callee) const;
534 /// \brief The abstract base class used to type erase specific TTI
538 /// \brief The template model for the base class which wraps a concrete
539 /// implementation in a type erased interface.
540 template <typename T> class Model;
542 std::unique_ptr<Concept> TTIImpl;
545 class TargetTransformInfo::Concept {
547 virtual ~Concept() = 0;
548 virtual const DataLayout &getDataLayout() const = 0;
549 virtual int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
550 virtual int getGEPCost(Type *PointeeType, const Value *Ptr,
551 ArrayRef<const Value *> Operands) = 0;
552 virtual int getCallCost(FunctionType *FTy, int NumArgs) = 0;
553 virtual int getCallCost(const Function *F, int NumArgs) = 0;
554 virtual int getCallCost(const Function *F,
555 ArrayRef<const Value *> Arguments) = 0;
556 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
557 ArrayRef<Type *> ParamTys) = 0;
558 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
559 ArrayRef<const Value *> Arguments) = 0;
560 virtual int getUserCost(const User *U) = 0;
561 virtual bool hasBranchDivergence() = 0;
562 virtual bool isSourceOfDivergence(const Value *V) = 0;
563 virtual bool isLoweredToCall(const Function *F) = 0;
564 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
565 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
566 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
567 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
568 int64_t BaseOffset, bool HasBaseReg,
570 unsigned AddrSpace) = 0;
571 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
572 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
573 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
574 int64_t BaseOffset, bool HasBaseReg,
575 int64_t Scale, unsigned AddrSpace) = 0;
576 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
577 virtual bool isZExtFree(Type *Ty1, Type *Ty2) = 0;
578 virtual bool isProfitableToHoist(Instruction *I) = 0;
579 virtual bool isTypeLegal(Type *Ty) = 0;
580 virtual unsigned getJumpBufAlignment() = 0;
581 virtual unsigned getJumpBufSize() = 0;
582 virtual bool shouldBuildLookupTables() = 0;
583 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
584 virtual bool enableInterleavedAccessVectorization() = 0;
585 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
586 virtual bool haveFastSqrt(Type *Ty) = 0;
587 virtual int getFPOpCost(Type *Ty) = 0;
588 virtual int getIntImmCost(const APInt &Imm, Type *Ty) = 0;
589 virtual int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
591 virtual int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
593 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
594 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
595 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
597 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
598 OperandValueKind Opd2Info,
599 OperandValueProperties Opd1PropInfo,
600 OperandValueProperties Opd2PropInfo) = 0;
601 virtual int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
603 virtual int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
604 virtual int getCFInstrCost(unsigned Opcode) = 0;
605 virtual int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
607 virtual int getVectorInstrCost(unsigned Opcode, Type *Val,
609 virtual int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
610 unsigned AddressSpace) = 0;
611 virtual int getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
613 unsigned AddressSpace) = 0;
614 virtual int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
616 ArrayRef<unsigned> Indices,
618 unsigned AddressSpace) = 0;
619 virtual int getReductionCost(unsigned Opcode, Type *Ty,
620 bool IsPairwiseForm) = 0;
621 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
622 ArrayRef<Type *> Tys) = 0;
623 virtual int getCallInstrCost(Function *F, Type *RetTy,
624 ArrayRef<Type *> Tys) = 0;
625 virtual unsigned getNumberOfParts(Type *Tp) = 0;
626 virtual int getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
627 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
628 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
629 MemIntrinsicInfo &Info) = 0;
630 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
631 Type *ExpectedType) = 0;
632 virtual bool areInlineCompatible(const Function *Caller,
633 const Function *Callee) const = 0;
636 template <typename T>
637 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
641 Model(T Impl) : Impl(std::move(Impl)) {}
644 const DataLayout &getDataLayout() const override {
645 return Impl.getDataLayout();
648 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
649 return Impl.getOperationCost(Opcode, Ty, OpTy);
651 int getGEPCost(Type *PointeeType, const Value *Ptr,
652 ArrayRef<const Value *> Operands) override {
653 return Impl.getGEPCost(PointeeType, Ptr, Operands);
655 int getCallCost(FunctionType *FTy, int NumArgs) override {
656 return Impl.getCallCost(FTy, NumArgs);
658 int getCallCost(const Function *F, int NumArgs) override {
659 return Impl.getCallCost(F, NumArgs);
661 int getCallCost(const Function *F,
662 ArrayRef<const Value *> Arguments) override {
663 return Impl.getCallCost(F, Arguments);
665 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
666 ArrayRef<Type *> ParamTys) override {
667 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
669 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
670 ArrayRef<const Value *> Arguments) override {
671 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
673 int getUserCost(const User *U) override { return Impl.getUserCost(U); }
674 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
675 bool isSourceOfDivergence(const Value *V) override {
676 return Impl.isSourceOfDivergence(V);
678 bool isLoweredToCall(const Function *F) override {
679 return Impl.isLoweredToCall(F);
681 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
682 return Impl.getUnrollingPreferences(L, UP);
684 bool isLegalAddImmediate(int64_t Imm) override {
685 return Impl.isLegalAddImmediate(Imm);
687 bool isLegalICmpImmediate(int64_t Imm) override {
688 return Impl.isLegalICmpImmediate(Imm);
690 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
691 bool HasBaseReg, int64_t Scale,
692 unsigned AddrSpace) override {
693 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
696 bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
697 return Impl.isLegalMaskedStore(DataType, Consecutive);
699 bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
700 return Impl.isLegalMaskedLoad(DataType, Consecutive);
702 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
703 bool HasBaseReg, int64_t Scale,
704 unsigned AddrSpace) override {
705 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
708 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
709 return Impl.isTruncateFree(Ty1, Ty2);
711 bool isZExtFree(Type *Ty1, Type *Ty2) override {
712 return Impl.isZExtFree(Ty1, Ty2);
714 bool isProfitableToHoist(Instruction *I) override {
715 return Impl.isProfitableToHoist(I);
717 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
718 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
719 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
720 bool shouldBuildLookupTables() override {
721 return Impl.shouldBuildLookupTables();
723 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
724 return Impl.enableAggressiveInterleaving(LoopHasReductions);
726 bool enableInterleavedAccessVectorization() override {
727 return Impl.enableInterleavedAccessVectorization();
729 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
730 return Impl.getPopcntSupport(IntTyWidthInBit);
732 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
734 int getFPOpCost(Type *Ty) override { return Impl.getFPOpCost(Ty); }
736 int getIntImmCost(const APInt &Imm, Type *Ty) override {
737 return Impl.getIntImmCost(Imm, Ty);
739 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
741 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
743 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
745 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
747 unsigned getNumberOfRegisters(bool Vector) override {
748 return Impl.getNumberOfRegisters(Vector);
750 unsigned getRegisterBitWidth(bool Vector) override {
751 return Impl.getRegisterBitWidth(Vector);
753 unsigned getMaxInterleaveFactor(unsigned VF) override {
754 return Impl.getMaxInterleaveFactor(VF);
757 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
758 OperandValueKind Opd2Info,
759 OperandValueProperties Opd1PropInfo,
760 OperandValueProperties Opd2PropInfo) override {
761 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
762 Opd1PropInfo, Opd2PropInfo);
764 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
765 Type *SubTp) override {
766 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
768 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
769 return Impl.getCastInstrCost(Opcode, Dst, Src);
771 int getCFInstrCost(unsigned Opcode) override {
772 return Impl.getCFInstrCost(Opcode);
774 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) override {
775 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
777 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) override {
778 return Impl.getVectorInstrCost(Opcode, Val, Index);
780 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
781 unsigned AddressSpace) override {
782 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
784 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
785 unsigned AddressSpace) override {
786 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
788 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
789 ArrayRef<unsigned> Indices, unsigned Alignment,
790 unsigned AddressSpace) override {
791 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
792 Alignment, AddressSpace);
794 int getReductionCost(unsigned Opcode, Type *Ty,
795 bool IsPairwiseForm) override {
796 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
798 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
799 ArrayRef<Type *> Tys) override {
800 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
802 int getCallInstrCost(Function *F, Type *RetTy,
803 ArrayRef<Type *> Tys) override {
804 return Impl.getCallInstrCost(F, RetTy, Tys);
806 unsigned getNumberOfParts(Type *Tp) override {
807 return Impl.getNumberOfParts(Tp);
809 int getAddressComputationCost(Type *Ty, bool IsComplex) override {
810 return Impl.getAddressComputationCost(Ty, IsComplex);
812 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
813 return Impl.getCostOfKeepingLiveOverCall(Tys);
815 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
816 MemIntrinsicInfo &Info) override {
817 return Impl.getTgtMemIntrinsic(Inst, Info);
819 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
820 Type *ExpectedType) override {
821 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
823 bool areInlineCompatible(const Function *Caller,
824 const Function *Callee) const override {
825 return Impl.areInlineCompatible(Caller, Callee);
829 template <typename T>
830 TargetTransformInfo::TargetTransformInfo(T Impl)
831 : TTIImpl(new Model<T>(Impl)) {}
833 /// \brief Analysis pass providing the \c TargetTransformInfo.
835 /// The core idea of the TargetIRAnalysis is to expose an interface through
836 /// which LLVM targets can analyze and provide information about the middle
837 /// end's target-independent IR. This supports use cases such as target-aware
838 /// cost modeling of IR constructs.
840 /// This is a function analysis because much of the cost modeling for targets
841 /// is done in a subtarget specific way and LLVM supports compiling different
842 /// functions targeting different subtargets in order to support runtime
843 /// dispatch according to the observed subtarget.
844 class TargetIRAnalysis {
846 typedef TargetTransformInfo Result;
848 /// \brief Opaque, unique identifier for this analysis pass.
849 static void *ID() { return (void *)&PassID; }
851 /// \brief Provide access to a name for this pass for debugging purposes.
852 static StringRef name() { return "TargetIRAnalysis"; }
854 /// \brief Default construct a target IR analysis.
856 /// This will use the module's datalayout to construct a baseline
857 /// conservative TTI result.
860 /// \brief Construct an IR analysis pass around a target-provide callback.
862 /// The callback will be called with a particular function for which the TTI
863 /// is needed and must return a TTI object for that function.
864 TargetIRAnalysis(std::function<Result(const Function &)> TTICallback);
866 // Value semantics. We spell out the constructors for MSVC.
867 TargetIRAnalysis(const TargetIRAnalysis &Arg)
868 : TTICallback(Arg.TTICallback) {}
869 TargetIRAnalysis(TargetIRAnalysis &&Arg)
870 : TTICallback(std::move(Arg.TTICallback)) {}
871 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
872 TTICallback = RHS.TTICallback;
875 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
876 TTICallback = std::move(RHS.TTICallback);
880 Result run(const Function &F);
885 /// \brief The callback used to produce a result.
887 /// We use a completely opaque callback so that targets can provide whatever
888 /// mechanism they desire for constructing the TTI for a given function.
890 /// FIXME: Should we really use std::function? It's relatively inefficient.
891 /// It might be possible to arrange for even stateful callbacks to outlive
892 /// the analysis and thus use a function_ref which would be lighter weight.
893 /// This may also be less error prone as the callback is likely to reference
894 /// the external TargetMachine, and that reference needs to never dangle.
895 std::function<Result(const Function &)> TTICallback;
897 /// \brief Helper function used as the callback in the default constructor.
898 static Result getDefaultTTI(const Function &F);
901 /// \brief Wrapper pass for TargetTransformInfo.
903 /// This pass can be constructed from a TTI object which it stores internally
904 /// and is queried by passes.
905 class TargetTransformInfoWrapperPass : public ImmutablePass {
906 TargetIRAnalysis TIRA;
907 Optional<TargetTransformInfo> TTI;
909 virtual void anchor();
914 /// \brief We must provide a default constructor for the pass but it should
917 /// Use the constructor below or call one of the creation routines.
918 TargetTransformInfoWrapperPass();
920 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
922 TargetTransformInfo &getTTI(const Function &F);
925 /// \brief Create an analysis pass wrapper around a TTI object.
927 /// This analysis pass just holds the TTI instance and makes it available to
929 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
931 } // End llvm namespace