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/IR/Operator.h"
29 #include "llvm/IR/PassManager.h"
30 #include "llvm/Pass.h"
31 #include "llvm/Support/DataTypes.h"
43 /// \brief Information about a load/store intrinsic defined by the target.
44 struct MemIntrinsicInfo {
46 : ReadMem(false), WriteMem(false), IsSimple(false), MatchingId(0),
47 NumMemRefs(0), PtrVal(nullptr) {}
50 /// True only if this memory operation is non-volatile, non-atomic, and
51 /// unordered. (See LoadInst/StoreInst for details on each)
53 // Same Id is set by the target for corresponding load/store intrinsics.
54 unsigned short MatchingId;
59 /// \brief This pass provides access to the codegen interfaces that are needed
60 /// for IR-level transformations.
61 class TargetTransformInfo {
63 /// \brief Construct a TTI object using a type implementing the \c Concept
66 /// This is used by targets to construct a TTI wrapping their target-specific
67 /// implementaion that encodes appropriate costs for their target.
68 template <typename T> TargetTransformInfo(T Impl);
70 /// \brief Construct a baseline TTI object using a minimal implementation of
71 /// the \c Concept API below.
73 /// The TTI implementation will reflect the information in the DataLayout
74 /// provided if non-null.
75 explicit TargetTransformInfo(const DataLayout &DL);
77 // Provide move semantics.
78 TargetTransformInfo(TargetTransformInfo &&Arg);
79 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
81 // We need to define the destructor out-of-line to define our sub-classes
83 ~TargetTransformInfo();
85 /// \brief Handle the invalidation of this information.
87 /// When used as a result of \c TargetIRAnalysis this method will be called
88 /// when the function this was computed for changes. When it returns false,
89 /// the information is preserved across those changes.
90 bool invalidate(Function &, const PreservedAnalyses &,
91 FunctionAnalysisManager::Invalidator &) {
92 // FIXME: We should probably in some way ensure that the subtarget
93 // information for a function hasn't changed.
97 /// \name Generic Target Information
100 /// \brief Underlying constants for 'cost' values in this interface.
102 /// Many APIs in this interface return a cost. This enum defines the
103 /// fundamental values that should be used to interpret (and produce) those
104 /// costs. The costs are returned as an int rather than a member of this
105 /// enumeration because it is expected that the cost of one IR instruction
106 /// may have a multiplicative factor to it or otherwise won't fit directly
107 /// into the enum. Moreover, it is common to sum or average costs which works
108 /// better as simple integral values. Thus this enum only provides constants.
109 /// Also note that the returned costs are signed integers to make it natural
110 /// to add, subtract, and test with zero (a common boundary condition). It is
111 /// not expected that 2^32 is a realistic cost to be modeling at any point.
113 /// Note that these costs should usually reflect the intersection of code-size
114 /// cost and execution cost. A free instruction is typically one that folds
115 /// into another instruction. For example, reg-to-reg moves can often be
116 /// skipped by renaming the registers in the CPU, but they still are encoded
117 /// and thus wouldn't be considered 'free' here.
118 enum TargetCostConstants {
119 TCC_Free = 0, ///< Expected to fold away in lowering.
120 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
121 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
124 /// \brief Estimate the cost of a specific operation when lowered.
126 /// Note that this is designed to work on an arbitrary synthetic opcode, and
127 /// thus work for hypothetical queries before an instruction has even been
128 /// formed. However, this does *not* work for GEPs, and must not be called
129 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
130 /// analyzing a GEP's cost required more information.
132 /// Typically only the result type is required, and the operand type can be
133 /// omitted. However, if the opcode is one of the cast instructions, the
134 /// operand type is required.
136 /// The returned cost is defined in terms of \c TargetCostConstants, see its
137 /// comments for a detailed explanation of the cost values.
138 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy = nullptr) const;
140 /// \brief Estimate the cost of a GEP operation when lowered.
142 /// The contract for this function is the same as \c getOperationCost except
143 /// that it supports an interface that provides extra information specific to
144 /// the GEP operation.
145 int getGEPCost(Type *PointeeType, const Value *Ptr,
146 ArrayRef<const Value *> Operands) const;
148 /// \brief Estimate the cost of a function call when lowered.
150 /// The contract for this is the same as \c getOperationCost except that it
151 /// supports an interface that provides extra information specific to call
154 /// This is the most basic query for estimating call cost: it only knows the
155 /// function type and (potentially) the number of arguments at the call site.
156 /// The latter is only interesting for varargs function types.
157 int getCallCost(FunctionType *FTy, int NumArgs = -1) const;
159 /// \brief Estimate the cost of calling a specific function when lowered.
161 /// This overload adds the ability to reason about the particular function
162 /// being called in the event it is a library call with special lowering.
163 int getCallCost(const Function *F, int NumArgs = -1) const;
165 /// \brief Estimate the cost of calling a specific function when lowered.
167 /// This overload allows specifying a set of candidate argument values.
168 int getCallCost(const Function *F, ArrayRef<const Value *> Arguments) const;
170 /// \returns A value by which our inlining threshold should be multiplied.
171 /// This is primarily used to bump up the inlining threshold wholesale on
172 /// targets where calls are unusually expensive.
174 /// TODO: This is a rather blunt instrument. Perhaps altering the costs of
175 /// individual classes of instructions would be better.
176 unsigned getInliningThresholdMultiplier() const;
178 /// \brief Estimate the cost of an intrinsic when lowered.
180 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
181 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
182 ArrayRef<Type *> ParamTys) const;
184 /// \brief Estimate the cost of an intrinsic when lowered.
186 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
187 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
188 ArrayRef<const Value *> Arguments) const;
190 /// \brief Estimate the cost of a given IR user when lowered.
192 /// This can estimate the cost of either a ConstantExpr or Instruction when
193 /// lowered. It has two primary advantages over the \c getOperationCost and
194 /// \c getGEPCost above, and one significant disadvantage: it can only be
195 /// used when the IR construct has already been formed.
197 /// The advantages are that it can inspect the SSA use graph to reason more
198 /// accurately about the cost. For example, all-constant-GEPs can often be
199 /// folded into a load or other instruction, but if they are used in some
200 /// other context they may not be folded. This routine can distinguish such
203 /// The returned cost is defined in terms of \c TargetCostConstants, see its
204 /// comments for a detailed explanation of the cost values.
205 int getUserCost(const User *U) const;
207 /// \brief Return true if branch divergence exists.
209 /// Branch divergence has a significantly negative impact on GPU performance
210 /// when threads in the same wavefront take different paths due to conditional
212 bool hasBranchDivergence() const;
214 /// \brief Returns whether V is a source of divergence.
216 /// This function provides the target-dependent information for
217 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
218 /// builds the dependency graph, and then runs the reachability algorithm
219 /// starting with the sources of divergence.
220 bool isSourceOfDivergence(const Value *V) const;
222 /// \brief Test whether calls to a function lower to actual program function
225 /// The idea is to test whether the program is likely to require a 'call'
226 /// instruction or equivalent in order to call the given function.
228 /// FIXME: It's not clear that this is a good or useful query API. Client's
229 /// should probably move to simpler cost metrics using the above.
230 /// Alternatively, we could split the cost interface into distinct code-size
231 /// and execution-speed costs. This would allow modelling the core of this
232 /// query more accurately as a call is a single small instruction, but
233 /// incurs significant execution cost.
234 bool isLoweredToCall(const Function *F) const;
236 /// Parameters that control the generic loop unrolling transformation.
237 struct UnrollingPreferences {
238 /// The cost threshold for the unrolled loop. Should be relative to the
239 /// getUserCost values returned by this API, and the expectation is that
240 /// the unrolled loop's instructions when run through that interface should
241 /// not exceed this cost. However, this is only an estimate. Also, specific
242 /// loops may be unrolled even with a cost above this threshold if deemed
243 /// profitable. Set this to UINT_MAX to disable the loop body cost
246 /// If complete unrolling will reduce the cost of the loop, we will boost
247 /// the Threshold by a certain percent to allow more aggressive complete
248 /// unrolling. This value provides the maximum boost percentage that we
249 /// can apply to Threshold (The value should be no less than 100).
250 /// BoostedThreshold = Threshold * min(RolledCost / UnrolledCost,
251 /// MaxPercentThresholdBoost / 100)
252 /// E.g. if complete unrolling reduces the loop execution time by 50%
253 /// then we boost the threshold by the factor of 2x. If unrolling is not
254 /// expected to reduce the running time, then we do not increase the
256 unsigned MaxPercentThresholdBoost;
257 /// The cost threshold for the unrolled loop when optimizing for size (set
258 /// to UINT_MAX to disable).
259 unsigned OptSizeThreshold;
260 /// The cost threshold for the unrolled loop, like Threshold, but used
261 /// for partial/runtime unrolling (set to UINT_MAX to disable).
262 unsigned PartialThreshold;
263 /// The cost threshold for the unrolled loop when optimizing for size, like
264 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
265 /// UINT_MAX to disable).
266 unsigned PartialOptSizeThreshold;
267 /// A forced unrolling factor (the number of concatenated bodies of the
268 /// original loop in the unrolled loop body). When set to 0, the unrolling
269 /// transformation will select an unrolling factor based on the current cost
270 /// threshold and other factors.
272 /// A forced peeling factor (the number of bodied of the original loop
273 /// that should be peeled off before the loop body). When set to 0, the
274 /// unrolling transformation will select a peeling factor based on profile
275 /// information and other factors.
277 /// Default unroll count for loops with run-time trip count.
278 unsigned DefaultUnrollRuntimeCount;
279 // Set the maximum unrolling factor. The unrolling factor may be selected
280 // using the appropriate cost threshold, but may not exceed this number
281 // (set to UINT_MAX to disable). This does not apply in cases where the
282 // loop is being fully unrolled.
284 /// Set the maximum unrolling factor for full unrolling. Like MaxCount, but
285 /// applies even if full unrolling is selected. This allows a target to fall
286 /// back to Partial unrolling if full unrolling is above FullUnrollMaxCount.
287 unsigned FullUnrollMaxCount;
288 // Represents number of instructions optimized when "back edge"
289 // becomes "fall through" in unrolled loop.
290 // For now we count a conditional branch on a backedge and a comparison
293 /// Allow partial unrolling (unrolling of loops to expand the size of the
294 /// loop body, not only to eliminate small constant-trip-count loops).
296 /// Allow runtime unrolling (unrolling of loops to expand the size of the
297 /// loop body even when the number of loop iterations is not known at
300 /// Allow generation of a loop remainder (extra iterations after unroll).
302 /// Allow emitting expensive instructions (such as divisions) when computing
303 /// the trip count of a loop for runtime unrolling.
304 bool AllowExpensiveTripCount;
305 /// Apply loop unroll on any kind of loop
306 /// (mainly to loops that fail runtime unrolling).
308 /// Allow using trip count upper bound to unroll loops.
310 /// Allow peeling off loop iterations for loops with low dynamic tripcount.
314 /// \brief Get target-customized preferences for the generic loop unrolling
315 /// transformation. The caller will initialize UP with the current
316 /// target-independent defaults.
317 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
321 /// \name Scalar Target Information
324 /// \brief Flags indicating the kind of support for population count.
326 /// Compared to the SW implementation, HW support is supposed to
327 /// significantly boost the performance when the population is dense, and it
328 /// may or may not degrade performance if the population is sparse. A HW
329 /// support is considered as "Fast" if it can outperform, or is on a par
330 /// with, SW implementation when the population is sparse; otherwise, it is
331 /// considered as "Slow".
332 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
334 /// \brief Return true if the specified immediate is legal add immediate, that
335 /// is the target has add instructions which can add a register with the
336 /// immediate without having to materialize the immediate into a register.
337 bool isLegalAddImmediate(int64_t Imm) const;
339 /// \brief Return true if the specified immediate is legal icmp immediate,
340 /// that is the target has icmp instructions which can compare a register
341 /// against the immediate without having to materialize the immediate into a
343 bool isLegalICmpImmediate(int64_t Imm) const;
345 /// \brief Return true if the addressing mode represented by AM is legal for
346 /// this target, for a load/store of the specified type.
347 /// The type may be VoidTy, in which case only return true if the addressing
348 /// mode is legal for a load/store of any legal type.
349 /// TODO: Handle pre/postinc as well.
350 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
351 bool HasBaseReg, int64_t Scale,
352 unsigned AddrSpace = 0) const;
354 /// \brief Return true if the target supports masked load/store
355 /// AVX2 and AVX-512 targets allow masks for consecutive load and store
356 bool isLegalMaskedStore(Type *DataType) const;
357 bool isLegalMaskedLoad(Type *DataType) const;
359 /// \brief Return true if the target supports masked gather/scatter
360 /// AVX-512 fully supports gather and scatter for vectors with 32 and 64
361 /// bits scalar type.
362 bool isLegalMaskedScatter(Type *DataType) const;
363 bool isLegalMaskedGather(Type *DataType) const;
365 /// \brief Return the cost of the scaling factor used in the addressing
366 /// mode represented by AM for this target, for a load/store
367 /// of the specified type.
368 /// If the AM is supported, the return value must be >= 0.
369 /// If the AM is not supported, it returns a negative value.
370 /// TODO: Handle pre/postinc as well.
371 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
372 bool HasBaseReg, int64_t Scale,
373 unsigned AddrSpace = 0) const;
375 /// \brief Return true if target supports the load / store
376 /// instruction with the given Offset on the form reg + Offset. It
377 /// may be that Offset is too big for a certain type (register
379 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) const;
381 /// \brief Return true if it's free to truncate a value of type Ty1 to type
382 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
383 /// by referencing its sub-register AX.
384 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
386 /// \brief Return true if it is profitable to hoist instruction in the
387 /// then/else to before if.
388 bool isProfitableToHoist(Instruction *I) const;
390 /// \brief Return true if this type is legal.
391 bool isTypeLegal(Type *Ty) const;
393 /// \brief Returns the target's jmp_buf alignment in bytes.
394 unsigned getJumpBufAlignment() const;
396 /// \brief Returns the target's jmp_buf size in bytes.
397 unsigned getJumpBufSize() const;
399 /// \brief Return true if switches should be turned into lookup tables for the
401 bool shouldBuildLookupTables() const;
403 /// \brief Return true if switches should be turned into lookup tables
404 /// containing this constant value for the target.
405 bool shouldBuildLookupTablesForConstant(Constant *C) const;
407 /// \brief Don't restrict interleaved unrolling to small loops.
408 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
410 /// \brief Enable matching of interleaved access groups.
411 bool enableInterleavedAccessVectorization() const;
413 /// \brief Indicate that it is potentially unsafe to automatically vectorize
414 /// floating-point operations because the semantics of vector and scalar
415 /// floating-point semantics may differ. For example, ARM NEON v7 SIMD math
416 /// does not support IEEE-754 denormal numbers, while depending on the
417 /// platform, scalar floating-point math does.
418 /// This applies to floating-point math operations and calls, not memory
419 /// operations, shuffles, or casts.
420 bool isFPVectorizationPotentiallyUnsafe() const;
422 /// \brief Determine if the target supports unaligned memory accesses.
423 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
424 unsigned BitWidth, unsigned AddressSpace = 0,
425 unsigned Alignment = 1,
426 bool *Fast = nullptr) const;
428 /// \brief Return hardware support for population count.
429 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
431 /// \brief Return true if the hardware has a fast square-root instruction.
432 bool haveFastSqrt(Type *Ty) const;
434 /// \brief Return the expected cost of supporting the floating point operation
435 /// of the specified type.
436 int getFPOpCost(Type *Ty) const;
438 /// \brief Return the expected cost of materializing for the given integer
439 /// immediate of the specified type.
440 int getIntImmCost(const APInt &Imm, Type *Ty) const;
442 /// \brief Return the expected cost of materialization for the given integer
443 /// immediate of the specified type for a given instruction. The cost can be
444 /// zero if the immediate can be folded into the specified instruction.
445 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
447 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
450 /// \brief Return the expected cost for the given integer when optimising
451 /// for size. This is different than the other integer immediate cost
452 /// functions in that it is subtarget agnostic. This is useful when you e.g.
453 /// target one ISA such as Aarch32 but smaller encodings could be possible
454 /// with another such as Thumb. This return value is used as a penalty when
455 /// the total costs for a constant is calculated (the bigger the cost, the
456 /// more beneficial constant hoisting is).
457 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
461 /// \name Vector Target Information
464 /// \brief The various kinds of shuffle patterns for vector queries.
466 SK_Broadcast, ///< Broadcast element 0 to all other elements.
467 SK_Reverse, ///< Reverse the order of the vector.
468 SK_Alternate, ///< Choose alternate elements from vector.
469 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
470 SK_ExtractSubvector,///< ExtractSubvector Index indicates start offset.
471 SK_PermuteTwoSrc, ///< Merge elements from two source vectors into one
472 ///< with any shuffle mask.
473 SK_PermuteSingleSrc ///< Shuffle elements of single source vector with any
477 /// \brief Additional information about an operand's possible values.
478 enum OperandValueKind {
479 OK_AnyValue, // Operand can have any value.
480 OK_UniformValue, // Operand is uniform (splat of a value).
481 OK_UniformConstantValue, // Operand is uniform constant.
482 OK_NonUniformConstantValue // Operand is a non uniform constant value.
485 /// \brief Additional properties of an operand's values.
486 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
488 /// \return The number of scalar or vector registers that the target has.
489 /// If 'Vectors' is true, it returns the number of vector registers. If it is
490 /// set to false, it returns the number of scalar registers.
491 unsigned getNumberOfRegisters(bool Vector) const;
493 /// \return The width of the largest scalar or vector register type.
494 unsigned getRegisterBitWidth(bool Vector) const;
496 /// \return The size of a cache line in bytes.
497 unsigned getCacheLineSize() const;
499 /// \return How much before a load we should place the prefetch instruction.
500 /// This is currently measured in number of instructions.
501 unsigned getPrefetchDistance() const;
503 /// \return Some HW prefetchers can handle accesses up to a certain constant
504 /// stride. This is the minimum stride in bytes where it makes sense to start
505 /// adding SW prefetches. The default is 1, i.e. prefetch with any stride.
506 unsigned getMinPrefetchStride() const;
508 /// \return The maximum number of iterations to prefetch ahead. If the
509 /// required number of iterations is more than this number, no prefetching is
511 unsigned getMaxPrefetchIterationsAhead() const;
513 /// \return The maximum interleave factor that any transform should try to
514 /// perform for this target. This number depends on the level of parallelism
515 /// and the number of execution units in the CPU.
516 unsigned getMaxInterleaveFactor(unsigned VF) const;
518 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
519 int getArithmeticInstrCost(
520 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info = OK_AnyValue,
521 OperandValueKind Opd2Info = OK_AnyValue,
522 OperandValueProperties Opd1PropInfo = OP_None,
523 OperandValueProperties Opd2PropInfo = OP_None) const;
525 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
526 /// The index and subtype parameters are used by the subvector insertion and
527 /// extraction shuffle kinds.
528 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
529 Type *SubTp = nullptr) const;
531 /// \return The expected cost of cast instructions, such as bitcast, trunc,
533 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
535 /// \return The expected cost of a sign- or zero-extended vector extract. Use
536 /// -1 to indicate that there is no information about the index value.
537 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
538 unsigned Index = -1) const;
540 /// \return The expected cost of control-flow related instructions such as
542 int getCFInstrCost(unsigned Opcode) const;
544 /// \returns The expected cost of compare and select instructions.
545 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
546 Type *CondTy = nullptr) const;
548 /// \return The expected cost of vector Insert and Extract.
549 /// Use -1 to indicate that there is no information on the index value.
550 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index = -1) const;
552 /// \return The cost of Load and Store instructions.
553 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
554 unsigned AddressSpace) const;
556 /// \return The cost of masked Load and Store instructions.
557 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
558 unsigned AddressSpace) const;
560 /// \return The cost of Gather or Scatter operation
561 /// \p Opcode - is a type of memory access Load or Store
562 /// \p DataTy - a vector type of the data to be loaded or stored
563 /// \p Ptr - pointer [or vector of pointers] - address[es] in memory
564 /// \p VariableMask - true when the memory access is predicated with a mask
565 /// that is not a compile-time constant
566 /// \p Alignment - alignment of single element
567 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
568 bool VariableMask, unsigned Alignment) const;
570 /// \return The cost of the interleaved memory operation.
571 /// \p Opcode is the memory operation code
572 /// \p VecTy is the vector type of the interleaved access.
573 /// \p Factor is the interleave factor
574 /// \p Indices is the indices for interleaved load members (as interleaved
575 /// load allows gaps)
576 /// \p Alignment is the alignment of the memory operation
577 /// \p AddressSpace is address space of the pointer.
578 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
579 ArrayRef<unsigned> Indices, unsigned Alignment,
580 unsigned AddressSpace) const;
582 /// \brief Calculate the cost of performing a vector reduction.
584 /// This is the cost of reducing the vector value of type \p Ty to a scalar
585 /// value using the operation denoted by \p Opcode. The form of the reduction
586 /// can either be a pairwise reduction or a reduction that splits the vector
587 /// at every reduction level.
591 /// ((v0+v1), (v2, v3), undef, undef)
594 /// ((v0+v2), (v1+v3), undef, undef)
595 int getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwiseForm) const;
597 /// \returns The cost of Intrinsic instructions. Types analysis only.
598 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
599 ArrayRef<Type *> Tys, FastMathFlags FMF) const;
601 /// \returns The cost of Intrinsic instructions. Analyses the real arguments.
602 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
603 ArrayRef<Value *> Args, FastMathFlags FMF) const;
605 /// \returns The cost of Call instructions.
606 int getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) const;
608 /// \returns The number of pieces into which the provided type must be
609 /// split during legalization. Zero is returned when the answer is unknown.
610 unsigned getNumberOfParts(Type *Tp) const;
612 /// \returns The cost of the address computation. For most targets this can be
613 /// merged into the instruction indexing mode. Some targets might want to
614 /// distinguish between address computation for memory operations on vector
615 /// types and scalar types. Such targets should override this function.
616 /// The 'IsComplex' parameter is a hint that the address computation is likely
617 /// to involve multiple instructions and as such unlikely to be merged into
618 /// the address indexing mode.
619 int getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
621 /// \returns The cost, if any, of keeping values of the given types alive
624 /// Some types may require the use of register classes that do not have
625 /// any callee-saved registers, so would require a spill and fill.
626 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
628 /// \returns True if the intrinsic is a supported memory intrinsic. Info
629 /// will contain additional information - whether the intrinsic may write
630 /// or read to memory, volatility and the pointer. Info is undefined
631 /// if false is returned.
632 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
634 /// \returns A value which is the result of the given memory intrinsic. New
635 /// instructions may be created to extract the result from the given intrinsic
636 /// memory operation. Returns nullptr if the target cannot create a result
637 /// from the given intrinsic.
638 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
639 Type *ExpectedType) const;
641 /// \returns True if the two functions have compatible attributes for inlining
643 bool areInlineCompatible(const Function *Caller,
644 const Function *Callee) const;
646 /// \returns The bitwidth of the largest vector type that should be used to
647 /// load/store in the given address space.
648 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
650 /// \returns True if the load instruction is legal to vectorize.
651 bool isLegalToVectorizeLoad(LoadInst *LI) const;
653 /// \returns True if the store instruction is legal to vectorize.
654 bool isLegalToVectorizeStore(StoreInst *SI) const;
656 /// \returns True if it is legal to vectorize the given load chain.
657 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
659 unsigned AddrSpace) const;
661 /// \returns True if it is legal to vectorize the given store chain.
662 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
664 unsigned AddrSpace) const;
666 /// \returns The new vector factor value if the target doesn't support \p
667 /// SizeInBytes loads or has a better vector factor.
668 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
669 unsigned ChainSizeInBytes,
670 VectorType *VecTy) const;
672 /// \returns The new vector factor value if the target doesn't support \p
673 /// SizeInBytes stores or has a better vector factor.
674 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
675 unsigned ChainSizeInBytes,
676 VectorType *VecTy) const;
681 /// \brief The abstract base class used to type erase specific TTI
685 /// \brief The template model for the base class which wraps a concrete
686 /// implementation in a type erased interface.
687 template <typename T> class Model;
689 std::unique_ptr<Concept> TTIImpl;
692 class TargetTransformInfo::Concept {
694 virtual ~Concept() = 0;
695 virtual const DataLayout &getDataLayout() const = 0;
696 virtual int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
697 virtual int getGEPCost(Type *PointeeType, const Value *Ptr,
698 ArrayRef<const Value *> Operands) = 0;
699 virtual int getCallCost(FunctionType *FTy, int NumArgs) = 0;
700 virtual int getCallCost(const Function *F, int NumArgs) = 0;
701 virtual int getCallCost(const Function *F,
702 ArrayRef<const Value *> Arguments) = 0;
703 virtual unsigned getInliningThresholdMultiplier() = 0;
704 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
705 ArrayRef<Type *> ParamTys) = 0;
706 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
707 ArrayRef<const Value *> Arguments) = 0;
708 virtual int getUserCost(const User *U) = 0;
709 virtual bool hasBranchDivergence() = 0;
710 virtual bool isSourceOfDivergence(const Value *V) = 0;
711 virtual bool isLoweredToCall(const Function *F) = 0;
712 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
713 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
714 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
715 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
716 int64_t BaseOffset, bool HasBaseReg,
718 unsigned AddrSpace) = 0;
719 virtual bool isLegalMaskedStore(Type *DataType) = 0;
720 virtual bool isLegalMaskedLoad(Type *DataType) = 0;
721 virtual bool isLegalMaskedScatter(Type *DataType) = 0;
722 virtual bool isLegalMaskedGather(Type *DataType) = 0;
723 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
724 int64_t BaseOffset, bool HasBaseReg,
725 int64_t Scale, unsigned AddrSpace) = 0;
726 virtual bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) = 0;
727 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
728 virtual bool isProfitableToHoist(Instruction *I) = 0;
729 virtual bool isTypeLegal(Type *Ty) = 0;
730 virtual unsigned getJumpBufAlignment() = 0;
731 virtual unsigned getJumpBufSize() = 0;
732 virtual bool shouldBuildLookupTables() = 0;
733 virtual bool shouldBuildLookupTablesForConstant(Constant *C) = 0;
734 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
735 virtual bool enableInterleavedAccessVectorization() = 0;
736 virtual bool isFPVectorizationPotentiallyUnsafe() = 0;
737 virtual bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
739 unsigned AddressSpace,
742 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
743 virtual bool haveFastSqrt(Type *Ty) = 0;
744 virtual int getFPOpCost(Type *Ty) = 0;
745 virtual int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
747 virtual int getIntImmCost(const APInt &Imm, Type *Ty) = 0;
748 virtual int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
750 virtual int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
752 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
753 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
754 virtual unsigned getCacheLineSize() = 0;
755 virtual unsigned getPrefetchDistance() = 0;
756 virtual unsigned getMinPrefetchStride() = 0;
757 virtual unsigned getMaxPrefetchIterationsAhead() = 0;
758 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
760 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
761 OperandValueKind Opd2Info,
762 OperandValueProperties Opd1PropInfo,
763 OperandValueProperties Opd2PropInfo) = 0;
764 virtual int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
766 virtual int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
767 virtual int getExtractWithExtendCost(unsigned Opcode, Type *Dst,
768 VectorType *VecTy, unsigned Index) = 0;
769 virtual int getCFInstrCost(unsigned Opcode) = 0;
770 virtual int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
772 virtual int getVectorInstrCost(unsigned Opcode, Type *Val,
774 virtual int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
775 unsigned AddressSpace) = 0;
776 virtual int getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
778 unsigned AddressSpace) = 0;
779 virtual int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
780 Value *Ptr, bool VariableMask,
781 unsigned Alignment) = 0;
782 virtual int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
784 ArrayRef<unsigned> Indices,
786 unsigned AddressSpace) = 0;
787 virtual int getReductionCost(unsigned Opcode, Type *Ty,
788 bool IsPairwiseForm) = 0;
789 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
790 ArrayRef<Type *> Tys,
791 FastMathFlags FMF) = 0;
792 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
793 ArrayRef<Value *> Args,
794 FastMathFlags FMF) = 0;
795 virtual int getCallInstrCost(Function *F, Type *RetTy,
796 ArrayRef<Type *> Tys) = 0;
797 virtual unsigned getNumberOfParts(Type *Tp) = 0;
798 virtual int getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
799 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
800 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
801 MemIntrinsicInfo &Info) = 0;
802 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
803 Type *ExpectedType) = 0;
804 virtual bool areInlineCompatible(const Function *Caller,
805 const Function *Callee) const = 0;
806 virtual unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const = 0;
807 virtual bool isLegalToVectorizeLoad(LoadInst *LI) const = 0;
808 virtual bool isLegalToVectorizeStore(StoreInst *SI) const = 0;
809 virtual bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
811 unsigned AddrSpace) const = 0;
812 virtual bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
814 unsigned AddrSpace) const = 0;
815 virtual unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
816 unsigned ChainSizeInBytes,
817 VectorType *VecTy) const = 0;
818 virtual unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
819 unsigned ChainSizeInBytes,
820 VectorType *VecTy) const = 0;
823 template <typename T>
824 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
828 Model(T Impl) : Impl(std::move(Impl)) {}
831 const DataLayout &getDataLayout() const override {
832 return Impl.getDataLayout();
835 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
836 return Impl.getOperationCost(Opcode, Ty, OpTy);
838 int getGEPCost(Type *PointeeType, const Value *Ptr,
839 ArrayRef<const Value *> Operands) override {
840 return Impl.getGEPCost(PointeeType, Ptr, Operands);
842 int getCallCost(FunctionType *FTy, int NumArgs) override {
843 return Impl.getCallCost(FTy, NumArgs);
845 int getCallCost(const Function *F, int NumArgs) override {
846 return Impl.getCallCost(F, NumArgs);
848 int getCallCost(const Function *F,
849 ArrayRef<const Value *> Arguments) override {
850 return Impl.getCallCost(F, Arguments);
852 unsigned getInliningThresholdMultiplier() override {
853 return Impl.getInliningThresholdMultiplier();
855 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
856 ArrayRef<Type *> ParamTys) override {
857 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
859 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
860 ArrayRef<const Value *> Arguments) override {
861 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
863 int getUserCost(const User *U) override { return Impl.getUserCost(U); }
864 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
865 bool isSourceOfDivergence(const Value *V) override {
866 return Impl.isSourceOfDivergence(V);
868 bool isLoweredToCall(const Function *F) override {
869 return Impl.isLoweredToCall(F);
871 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
872 return Impl.getUnrollingPreferences(L, UP);
874 bool isLegalAddImmediate(int64_t Imm) override {
875 return Impl.isLegalAddImmediate(Imm);
877 bool isLegalICmpImmediate(int64_t Imm) override {
878 return Impl.isLegalICmpImmediate(Imm);
880 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
881 bool HasBaseReg, int64_t Scale,
882 unsigned AddrSpace) override {
883 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
886 bool isLegalMaskedStore(Type *DataType) override {
887 return Impl.isLegalMaskedStore(DataType);
889 bool isLegalMaskedLoad(Type *DataType) override {
890 return Impl.isLegalMaskedLoad(DataType);
892 bool isLegalMaskedScatter(Type *DataType) override {
893 return Impl.isLegalMaskedScatter(DataType);
895 bool isLegalMaskedGather(Type *DataType) override {
896 return Impl.isLegalMaskedGather(DataType);
898 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
899 bool HasBaseReg, int64_t Scale,
900 unsigned AddrSpace) override {
901 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
904 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) override {
905 return Impl.isFoldableMemAccessOffset(I, Offset);
907 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
908 return Impl.isTruncateFree(Ty1, Ty2);
910 bool isProfitableToHoist(Instruction *I) override {
911 return Impl.isProfitableToHoist(I);
913 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
914 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
915 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
916 bool shouldBuildLookupTables() override {
917 return Impl.shouldBuildLookupTables();
919 bool shouldBuildLookupTablesForConstant(Constant *C) override {
920 return Impl.shouldBuildLookupTablesForConstant(C);
922 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
923 return Impl.enableAggressiveInterleaving(LoopHasReductions);
925 bool enableInterleavedAccessVectorization() override {
926 return Impl.enableInterleavedAccessVectorization();
928 bool isFPVectorizationPotentiallyUnsafe() override {
929 return Impl.isFPVectorizationPotentiallyUnsafe();
931 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
932 unsigned BitWidth, unsigned AddressSpace,
933 unsigned Alignment, bool *Fast) override {
934 return Impl.allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace,
937 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
938 return Impl.getPopcntSupport(IntTyWidthInBit);
940 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
942 int getFPOpCost(Type *Ty) override { return Impl.getFPOpCost(Ty); }
944 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
946 return Impl.getIntImmCodeSizeCost(Opc, Idx, Imm, Ty);
948 int getIntImmCost(const APInt &Imm, Type *Ty) override {
949 return Impl.getIntImmCost(Imm, Ty);
951 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
953 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
955 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
957 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
959 unsigned getNumberOfRegisters(bool Vector) override {
960 return Impl.getNumberOfRegisters(Vector);
962 unsigned getRegisterBitWidth(bool Vector) override {
963 return Impl.getRegisterBitWidth(Vector);
966 unsigned getCacheLineSize() override {
967 return Impl.getCacheLineSize();
969 unsigned getPrefetchDistance() override { return Impl.getPrefetchDistance(); }
970 unsigned getMinPrefetchStride() override {
971 return Impl.getMinPrefetchStride();
973 unsigned getMaxPrefetchIterationsAhead() override {
974 return Impl.getMaxPrefetchIterationsAhead();
976 unsigned getMaxInterleaveFactor(unsigned VF) override {
977 return Impl.getMaxInterleaveFactor(VF);
980 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
981 OperandValueKind Opd2Info,
982 OperandValueProperties Opd1PropInfo,
983 OperandValueProperties Opd2PropInfo) override {
984 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
985 Opd1PropInfo, Opd2PropInfo);
987 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
988 Type *SubTp) override {
989 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
991 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
992 return Impl.getCastInstrCost(Opcode, Dst, Src);
994 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
995 unsigned Index) override {
996 return Impl.getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
998 int getCFInstrCost(unsigned Opcode) override {
999 return Impl.getCFInstrCost(Opcode);
1001 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) override {
1002 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
1004 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) override {
1005 return Impl.getVectorInstrCost(Opcode, Val, Index);
1007 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1008 unsigned AddressSpace) override {
1009 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
1011 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1012 unsigned AddressSpace) override {
1013 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
1015 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
1016 Value *Ptr, bool VariableMask,
1017 unsigned Alignment) override {
1018 return Impl.getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
1021 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
1022 ArrayRef<unsigned> Indices, unsigned Alignment,
1023 unsigned AddressSpace) override {
1024 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
1025 Alignment, AddressSpace);
1027 int getReductionCost(unsigned Opcode, Type *Ty,
1028 bool IsPairwiseForm) override {
1029 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
1031 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
1032 FastMathFlags FMF) override {
1033 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
1035 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
1036 ArrayRef<Value *> Args,
1037 FastMathFlags FMF) override {
1038 return Impl.getIntrinsicInstrCost(ID, RetTy, Args, FMF);
1040 int getCallInstrCost(Function *F, Type *RetTy,
1041 ArrayRef<Type *> Tys) override {
1042 return Impl.getCallInstrCost(F, RetTy, Tys);
1044 unsigned getNumberOfParts(Type *Tp) override {
1045 return Impl.getNumberOfParts(Tp);
1047 int getAddressComputationCost(Type *Ty, bool IsComplex) override {
1048 return Impl.getAddressComputationCost(Ty, IsComplex);
1050 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
1051 return Impl.getCostOfKeepingLiveOverCall(Tys);
1053 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
1054 MemIntrinsicInfo &Info) override {
1055 return Impl.getTgtMemIntrinsic(Inst, Info);
1057 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
1058 Type *ExpectedType) override {
1059 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
1061 bool areInlineCompatible(const Function *Caller,
1062 const Function *Callee) const override {
1063 return Impl.areInlineCompatible(Caller, Callee);
1065 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const override {
1066 return Impl.getLoadStoreVecRegBitWidth(AddrSpace);
1068 bool isLegalToVectorizeLoad(LoadInst *LI) const override {
1069 return Impl.isLegalToVectorizeLoad(LI);
1071 bool isLegalToVectorizeStore(StoreInst *SI) const override {
1072 return Impl.isLegalToVectorizeStore(SI);
1074 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
1076 unsigned AddrSpace) const override {
1077 return Impl.isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1080 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
1082 unsigned AddrSpace) const override {
1083 return Impl.isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1086 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
1087 unsigned ChainSizeInBytes,
1088 VectorType *VecTy) const override {
1089 return Impl.getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1091 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
1092 unsigned ChainSizeInBytes,
1093 VectorType *VecTy) const override {
1094 return Impl.getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1098 template <typename T>
1099 TargetTransformInfo::TargetTransformInfo(T Impl)
1100 : TTIImpl(new Model<T>(Impl)) {}
1102 /// \brief Analysis pass providing the \c TargetTransformInfo.
1104 /// The core idea of the TargetIRAnalysis is to expose an interface through
1105 /// which LLVM targets can analyze and provide information about the middle
1106 /// end's target-independent IR. This supports use cases such as target-aware
1107 /// cost modeling of IR constructs.
1109 /// This is a function analysis because much of the cost modeling for targets
1110 /// is done in a subtarget specific way and LLVM supports compiling different
1111 /// functions targeting different subtargets in order to support runtime
1112 /// dispatch according to the observed subtarget.
1113 class TargetIRAnalysis : public AnalysisInfoMixin<TargetIRAnalysis> {
1115 typedef TargetTransformInfo Result;
1117 /// \brief Default construct a target IR analysis.
1119 /// This will use the module's datalayout to construct a baseline
1120 /// conservative TTI result.
1123 /// \brief Construct an IR analysis pass around a target-provide callback.
1125 /// The callback will be called with a particular function for which the TTI
1126 /// is needed and must return a TTI object for that function.
1127 TargetIRAnalysis(std::function<Result(const Function &)> TTICallback);
1129 // Value semantics. We spell out the constructors for MSVC.
1130 TargetIRAnalysis(const TargetIRAnalysis &Arg)
1131 : TTICallback(Arg.TTICallback) {}
1132 TargetIRAnalysis(TargetIRAnalysis &&Arg)
1133 : TTICallback(std::move(Arg.TTICallback)) {}
1134 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
1135 TTICallback = RHS.TTICallback;
1138 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
1139 TTICallback = std::move(RHS.TTICallback);
1143 Result run(const Function &F, FunctionAnalysisManager &);
1146 friend AnalysisInfoMixin<TargetIRAnalysis>;
1147 static AnalysisKey Key;
1149 /// \brief The callback used to produce a result.
1151 /// We use a completely opaque callback so that targets can provide whatever
1152 /// mechanism they desire for constructing the TTI for a given function.
1154 /// FIXME: Should we really use std::function? It's relatively inefficient.
1155 /// It might be possible to arrange for even stateful callbacks to outlive
1156 /// the analysis and thus use a function_ref which would be lighter weight.
1157 /// This may also be less error prone as the callback is likely to reference
1158 /// the external TargetMachine, and that reference needs to never dangle.
1159 std::function<Result(const Function &)> TTICallback;
1161 /// \brief Helper function used as the callback in the default constructor.
1162 static Result getDefaultTTI(const Function &F);
1165 /// \brief Wrapper pass for TargetTransformInfo.
1167 /// This pass can be constructed from a TTI object which it stores internally
1168 /// and is queried by passes.
1169 class TargetTransformInfoWrapperPass : public ImmutablePass {
1170 TargetIRAnalysis TIRA;
1171 Optional<TargetTransformInfo> TTI;
1173 virtual void anchor();
1178 /// \brief We must provide a default constructor for the pass but it should
1181 /// Use the constructor below or call one of the creation routines.
1182 TargetTransformInfoWrapperPass();
1184 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1186 TargetTransformInfo &getTTI(const Function &F);
1189 /// \brief Create an analysis pass wrapper around a TTI object.
1191 /// This analysis pass just holds the TTI instance and makes it available to
1193 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1195 } // End llvm namespace