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"
39 class ScalarEvolution;
45 /// \brief Information about a load/store intrinsic defined by the target.
46 struct MemIntrinsicInfo {
48 : ReadMem(false), WriteMem(false), IsSimple(false), MatchingId(0),
49 NumMemRefs(0), PtrVal(nullptr) {}
52 /// True only if this memory operation is non-volatile, non-atomic, and
53 /// unordered. (See LoadInst/StoreInst for details on each)
55 // Same Id is set by the target for corresponding load/store intrinsics.
56 unsigned short MatchingId;
59 /// This is the pointer that the intrinsic is loading from or storing to.
60 /// If this is non-null, then analysis/optimization passes can assume that
61 /// this intrinsic is functionally equivalent to a load/store from this
66 /// \brief This pass provides access to the codegen interfaces that are needed
67 /// for IR-level transformations.
68 class TargetTransformInfo {
70 /// \brief Construct a TTI object using a type implementing the \c Concept
73 /// This is used by targets to construct a TTI wrapping their target-specific
74 /// implementaion that encodes appropriate costs for their target.
75 template <typename T> TargetTransformInfo(T Impl);
77 /// \brief Construct a baseline TTI object using a minimal implementation of
78 /// the \c Concept API below.
80 /// The TTI implementation will reflect the information in the DataLayout
81 /// provided if non-null.
82 explicit TargetTransformInfo(const DataLayout &DL);
84 // Provide move semantics.
85 TargetTransformInfo(TargetTransformInfo &&Arg);
86 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
88 // We need to define the destructor out-of-line to define our sub-classes
90 ~TargetTransformInfo();
92 /// \brief Handle the invalidation of this information.
94 /// When used as a result of \c TargetIRAnalysis this method will be called
95 /// when the function this was computed for changes. When it returns false,
96 /// the information is preserved across those changes.
97 bool invalidate(Function &, const PreservedAnalyses &,
98 FunctionAnalysisManager::Invalidator &) {
99 // FIXME: We should probably in some way ensure that the subtarget
100 // information for a function hasn't changed.
104 /// \name Generic Target Information
107 /// \brief Underlying constants for 'cost' values in this interface.
109 /// Many APIs in this interface return a cost. This enum defines the
110 /// fundamental values that should be used to interpret (and produce) those
111 /// costs. The costs are returned as an int rather than a member of this
112 /// enumeration because it is expected that the cost of one IR instruction
113 /// may have a multiplicative factor to it or otherwise won't fit directly
114 /// into the enum. Moreover, it is common to sum or average costs which works
115 /// better as simple integral values. Thus this enum only provides constants.
116 /// Also note that the returned costs are signed integers to make it natural
117 /// to add, subtract, and test with zero (a common boundary condition). It is
118 /// not expected that 2^32 is a realistic cost to be modeling at any point.
120 /// Note that these costs should usually reflect the intersection of code-size
121 /// cost and execution cost. A free instruction is typically one that folds
122 /// into another instruction. For example, reg-to-reg moves can often be
123 /// skipped by renaming the registers in the CPU, but they still are encoded
124 /// and thus wouldn't be considered 'free' here.
125 enum TargetCostConstants {
126 TCC_Free = 0, ///< Expected to fold away in lowering.
127 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
128 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
131 /// \brief Estimate the cost of a specific operation when lowered.
133 /// Note that this is designed to work on an arbitrary synthetic opcode, and
134 /// thus work for hypothetical queries before an instruction has even been
135 /// formed. However, this does *not* work for GEPs, and must not be called
136 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
137 /// analyzing a GEP's cost required more information.
139 /// Typically only the result type is required, and the operand type can be
140 /// omitted. However, if the opcode is one of the cast instructions, the
141 /// operand type is required.
143 /// The returned cost is defined in terms of \c TargetCostConstants, see its
144 /// comments for a detailed explanation of the cost values.
145 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy = nullptr) const;
147 /// \brief Estimate the cost of a GEP operation when lowered.
149 /// The contract for this function is the same as \c getOperationCost except
150 /// that it supports an interface that provides extra information specific to
151 /// the GEP operation.
152 int getGEPCost(Type *PointeeType, const Value *Ptr,
153 ArrayRef<const Value *> Operands) const;
155 /// \brief Estimate the cost of a function call when lowered.
157 /// The contract for this is the same as \c getOperationCost except that it
158 /// supports an interface that provides extra information specific to call
161 /// This is the most basic query for estimating call cost: it only knows the
162 /// function type and (potentially) the number of arguments at the call site.
163 /// The latter is only interesting for varargs function types.
164 int getCallCost(FunctionType *FTy, int NumArgs = -1) const;
166 /// \brief Estimate the cost of calling a specific function when lowered.
168 /// This overload adds the ability to reason about the particular function
169 /// being called in the event it is a library call with special lowering.
170 int getCallCost(const Function *F, int NumArgs = -1) const;
172 /// \brief Estimate the cost of calling a specific function when lowered.
174 /// This overload allows specifying a set of candidate argument values.
175 int getCallCost(const Function *F, ArrayRef<const Value *> Arguments) const;
177 /// \returns A value by which our inlining threshold should be multiplied.
178 /// This is primarily used to bump up the inlining threshold wholesale on
179 /// targets where calls are unusually expensive.
181 /// TODO: This is a rather blunt instrument. Perhaps altering the costs of
182 /// individual classes of instructions would be better.
183 unsigned getInliningThresholdMultiplier() const;
185 /// \brief Estimate the cost of an intrinsic when lowered.
187 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
188 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
189 ArrayRef<Type *> ParamTys) const;
191 /// \brief Estimate the cost of an intrinsic when lowered.
193 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
194 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
195 ArrayRef<const Value *> Arguments) const;
197 /// \brief Estimate the cost of a given IR user when lowered.
199 /// This can estimate the cost of either a ConstantExpr or Instruction when
200 /// lowered. It has two primary advantages over the \c getOperationCost and
201 /// \c getGEPCost above, and one significant disadvantage: it can only be
202 /// used when the IR construct has already been formed.
204 /// The advantages are that it can inspect the SSA use graph to reason more
205 /// accurately about the cost. For example, all-constant-GEPs can often be
206 /// folded into a load or other instruction, but if they are used in some
207 /// other context they may not be folded. This routine can distinguish such
210 /// The returned cost is defined in terms of \c TargetCostConstants, see its
211 /// comments for a detailed explanation of the cost values.
212 int getUserCost(const User *U) const;
214 /// \brief Return true if branch divergence exists.
216 /// Branch divergence has a significantly negative impact on GPU performance
217 /// when threads in the same wavefront take different paths due to conditional
219 bool hasBranchDivergence() const;
221 /// \brief Returns whether V is a source of divergence.
223 /// This function provides the target-dependent information for
224 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
225 /// builds the dependency graph, and then runs the reachability algorithm
226 /// starting with the sources of divergence.
227 bool isSourceOfDivergence(const Value *V) const;
229 /// \brief Test whether calls to a function lower to actual program function
232 /// The idea is to test whether the program is likely to require a 'call'
233 /// instruction or equivalent in order to call the given function.
235 /// FIXME: It's not clear that this is a good or useful query API. Client's
236 /// should probably move to simpler cost metrics using the above.
237 /// Alternatively, we could split the cost interface into distinct code-size
238 /// and execution-speed costs. This would allow modelling the core of this
239 /// query more accurately as a call is a single small instruction, but
240 /// incurs significant execution cost.
241 bool isLoweredToCall(const Function *F) const;
243 /// Parameters that control the generic loop unrolling transformation.
244 struct UnrollingPreferences {
245 /// The cost threshold for the unrolled loop. Should be relative to the
246 /// getUserCost values returned by this API, and the expectation is that
247 /// the unrolled loop's instructions when run through that interface should
248 /// not exceed this cost. However, this is only an estimate. Also, specific
249 /// loops may be unrolled even with a cost above this threshold if deemed
250 /// profitable. Set this to UINT_MAX to disable the loop body cost
253 /// If complete unrolling will reduce the cost of the loop, we will boost
254 /// the Threshold by a certain percent to allow more aggressive complete
255 /// unrolling. This value provides the maximum boost percentage that we
256 /// can apply to Threshold (The value should be no less than 100).
257 /// BoostedThreshold = Threshold * min(RolledCost / UnrolledCost,
258 /// MaxPercentThresholdBoost / 100)
259 /// E.g. if complete unrolling reduces the loop execution time by 50%
260 /// then we boost the threshold by the factor of 2x. If unrolling is not
261 /// expected to reduce the running time, then we do not increase the
263 unsigned MaxPercentThresholdBoost;
264 /// The cost threshold for the unrolled loop when optimizing for size (set
265 /// to UINT_MAX to disable).
266 unsigned OptSizeThreshold;
267 /// The cost threshold for the unrolled loop, like Threshold, but used
268 /// for partial/runtime unrolling (set to UINT_MAX to disable).
269 unsigned PartialThreshold;
270 /// The cost threshold for the unrolled loop when optimizing for size, like
271 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
272 /// UINT_MAX to disable).
273 unsigned PartialOptSizeThreshold;
274 /// A forced unrolling factor (the number of concatenated bodies of the
275 /// original loop in the unrolled loop body). When set to 0, the unrolling
276 /// transformation will select an unrolling factor based on the current cost
277 /// threshold and other factors.
279 /// A forced peeling factor (the number of bodied of the original loop
280 /// that should be peeled off before the loop body). When set to 0, the
281 /// unrolling transformation will select a peeling factor based on profile
282 /// information and other factors.
284 /// Default unroll count for loops with run-time trip count.
285 unsigned DefaultUnrollRuntimeCount;
286 // Set the maximum unrolling factor. The unrolling factor may be selected
287 // using the appropriate cost threshold, but may not exceed this number
288 // (set to UINT_MAX to disable). This does not apply in cases where the
289 // loop is being fully unrolled.
291 /// Set the maximum unrolling factor for full unrolling. Like MaxCount, but
292 /// applies even if full unrolling is selected. This allows a target to fall
293 /// back to Partial unrolling if full unrolling is above FullUnrollMaxCount.
294 unsigned FullUnrollMaxCount;
295 // Represents number of instructions optimized when "back edge"
296 // becomes "fall through" in unrolled loop.
297 // For now we count a conditional branch on a backedge and a comparison
300 /// Allow partial unrolling (unrolling of loops to expand the size of the
301 /// loop body, not only to eliminate small constant-trip-count loops).
303 /// Allow runtime unrolling (unrolling of loops to expand the size of the
304 /// loop body even when the number of loop iterations is not known at
307 /// Allow generation of a loop remainder (extra iterations after unroll).
309 /// Allow emitting expensive instructions (such as divisions) when computing
310 /// the trip count of a loop for runtime unrolling.
311 bool AllowExpensiveTripCount;
312 /// Apply loop unroll on any kind of loop
313 /// (mainly to loops that fail runtime unrolling).
315 /// Allow using trip count upper bound to unroll loops.
317 /// Allow peeling off loop iterations for loops with low dynamic tripcount.
321 /// \brief Get target-customized preferences for the generic loop unrolling
322 /// transformation. The caller will initialize UP with the current
323 /// target-independent defaults.
324 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
328 /// \name Scalar Target Information
331 /// \brief Flags indicating the kind of support for population count.
333 /// Compared to the SW implementation, HW support is supposed to
334 /// significantly boost the performance when the population is dense, and it
335 /// may or may not degrade performance if the population is sparse. A HW
336 /// support is considered as "Fast" if it can outperform, or is on a par
337 /// with, SW implementation when the population is sparse; otherwise, it is
338 /// considered as "Slow".
339 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
341 /// \brief Return true if the specified immediate is legal add immediate, that
342 /// is the target has add instructions which can add a register with the
343 /// immediate without having to materialize the immediate into a register.
344 bool isLegalAddImmediate(int64_t Imm) const;
346 /// \brief Return true if the specified immediate is legal icmp immediate,
347 /// that is the target has icmp instructions which can compare a register
348 /// against the immediate without having to materialize the immediate into a
350 bool isLegalICmpImmediate(int64_t Imm) const;
352 /// \brief Return true if the addressing mode represented by AM is legal for
353 /// this target, for a load/store of the specified type.
354 /// The type may be VoidTy, in which case only return true if the addressing
355 /// mode is legal for a load/store of any legal type.
356 /// TODO: Handle pre/postinc as well.
357 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
358 bool HasBaseReg, int64_t Scale,
359 unsigned AddrSpace = 0) const;
361 /// \brief Return true if the target supports masked load/store
362 /// AVX2 and AVX-512 targets allow masks for consecutive load and store
363 bool isLegalMaskedStore(Type *DataType) const;
364 bool isLegalMaskedLoad(Type *DataType) const;
366 /// \brief Return true if the target supports masked gather/scatter
367 /// AVX-512 fully supports gather and scatter for vectors with 32 and 64
368 /// bits scalar type.
369 bool isLegalMaskedScatter(Type *DataType) const;
370 bool isLegalMaskedGather(Type *DataType) const;
372 /// \brief Return the cost of the scaling factor used in the addressing
373 /// mode represented by AM for this target, for a load/store
374 /// of the specified type.
375 /// If the AM is supported, the return value must be >= 0.
376 /// If the AM is not supported, it returns a negative value.
377 /// TODO: Handle pre/postinc as well.
378 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
379 bool HasBaseReg, int64_t Scale,
380 unsigned AddrSpace = 0) const;
382 /// \brief Return true if target supports the load / store
383 /// instruction with the given Offset on the form reg + Offset. It
384 /// may be that Offset is too big for a certain type (register
386 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) const;
388 /// \brief Return true if it's free to truncate a value of type Ty1 to type
389 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
390 /// by referencing its sub-register AX.
391 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
393 /// \brief Return true if it is profitable to hoist instruction in the
394 /// then/else to before if.
395 bool isProfitableToHoist(Instruction *I) const;
397 /// \brief Return true if this type is legal.
398 bool isTypeLegal(Type *Ty) const;
400 /// \brief Returns the target's jmp_buf alignment in bytes.
401 unsigned getJumpBufAlignment() const;
403 /// \brief Returns the target's jmp_buf size in bytes.
404 unsigned getJumpBufSize() const;
406 /// \brief Return true if switches should be turned into lookup tables for the
408 bool shouldBuildLookupTables() const;
410 /// \brief Return true if switches should be turned into lookup tables
411 /// containing this constant value for the target.
412 bool shouldBuildLookupTablesForConstant(Constant *C) const;
414 /// \brief Don't restrict interleaved unrolling to small loops.
415 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
417 /// \brief Enable matching of interleaved access groups.
418 bool enableInterleavedAccessVectorization() const;
420 /// \brief Indicate that it is potentially unsafe to automatically vectorize
421 /// floating-point operations because the semantics of vector and scalar
422 /// floating-point semantics may differ. For example, ARM NEON v7 SIMD math
423 /// does not support IEEE-754 denormal numbers, while depending on the
424 /// platform, scalar floating-point math does.
425 /// This applies to floating-point math operations and calls, not memory
426 /// operations, shuffles, or casts.
427 bool isFPVectorizationPotentiallyUnsafe() const;
429 /// \brief Determine if the target supports unaligned memory accesses.
430 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
431 unsigned BitWidth, unsigned AddressSpace = 0,
432 unsigned Alignment = 1,
433 bool *Fast = nullptr) const;
435 /// \brief Return hardware support for population count.
436 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
438 /// \brief Return true if the hardware has a fast square-root instruction.
439 bool haveFastSqrt(Type *Ty) const;
441 /// \brief Return the expected cost of supporting the floating point operation
442 /// of the specified type.
443 int getFPOpCost(Type *Ty) const;
445 /// \brief Return the expected cost of materializing for the given integer
446 /// immediate of the specified type.
447 int getIntImmCost(const APInt &Imm, Type *Ty) const;
449 /// \brief Return the expected cost of materialization for the given integer
450 /// immediate of the specified type for a given instruction. The cost can be
451 /// zero if the immediate can be folded into the specified instruction.
452 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
454 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
457 /// \brief Return the expected cost for the given integer when optimising
458 /// for size. This is different than the other integer immediate cost
459 /// functions in that it is subtarget agnostic. This is useful when you e.g.
460 /// target one ISA such as Aarch32 but smaller encodings could be possible
461 /// with another such as Thumb. This return value is used as a penalty when
462 /// the total costs for a constant is calculated (the bigger the cost, the
463 /// more beneficial constant hoisting is).
464 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
468 /// \name Vector Target Information
471 /// \brief The various kinds of shuffle patterns for vector queries.
473 SK_Broadcast, ///< Broadcast element 0 to all other elements.
474 SK_Reverse, ///< Reverse the order of the vector.
475 SK_Alternate, ///< Choose alternate elements from vector.
476 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
477 SK_ExtractSubvector,///< ExtractSubvector Index indicates start offset.
478 SK_PermuteTwoSrc, ///< Merge elements from two source vectors into one
479 ///< with any shuffle mask.
480 SK_PermuteSingleSrc ///< Shuffle elements of single source vector with any
484 /// \brief Additional information about an operand's possible values.
485 enum OperandValueKind {
486 OK_AnyValue, // Operand can have any value.
487 OK_UniformValue, // Operand is uniform (splat of a value).
488 OK_UniformConstantValue, // Operand is uniform constant.
489 OK_NonUniformConstantValue // Operand is a non uniform constant value.
492 /// \brief Additional properties of an operand's values.
493 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
495 /// \return The number of scalar or vector registers that the target has.
496 /// If 'Vectors' is true, it returns the number of vector registers. If it is
497 /// set to false, it returns the number of scalar registers.
498 unsigned getNumberOfRegisters(bool Vector) const;
500 /// \return The width of the largest scalar or vector register type.
501 unsigned getRegisterBitWidth(bool Vector) const;
503 /// \return The size of a cache line in bytes.
504 unsigned getCacheLineSize() const;
506 /// \return How much before a load we should place the prefetch instruction.
507 /// This is currently measured in number of instructions.
508 unsigned getPrefetchDistance() const;
510 /// \return Some HW prefetchers can handle accesses up to a certain constant
511 /// stride. This is the minimum stride in bytes where it makes sense to start
512 /// adding SW prefetches. The default is 1, i.e. prefetch with any stride.
513 unsigned getMinPrefetchStride() const;
515 /// \return The maximum number of iterations to prefetch ahead. If the
516 /// required number of iterations is more than this number, no prefetching is
518 unsigned getMaxPrefetchIterationsAhead() const;
520 /// \return The maximum interleave factor that any transform should try to
521 /// perform for this target. This number depends on the level of parallelism
522 /// and the number of execution units in the CPU.
523 unsigned getMaxInterleaveFactor(unsigned VF) const;
525 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
526 /// \p Args is an optional argument which holds the instruction operands
527 /// values so the TTI can analyize those values searching for special
528 /// cases\optimizations based on those values.
529 int getArithmeticInstrCost(
530 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info = OK_AnyValue,
531 OperandValueKind Opd2Info = OK_AnyValue,
532 OperandValueProperties Opd1PropInfo = OP_None,
533 OperandValueProperties Opd2PropInfo = OP_None,
534 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) const;
536 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
537 /// The index and subtype parameters are used by the subvector insertion and
538 /// extraction shuffle kinds.
539 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
540 Type *SubTp = nullptr) const;
542 /// \return The expected cost of cast instructions, such as bitcast, trunc,
544 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
546 /// \return The expected cost of a sign- or zero-extended vector extract. Use
547 /// -1 to indicate that there is no information about the index value.
548 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
549 unsigned Index = -1) const;
551 /// \return The expected cost of control-flow related instructions such as
553 int getCFInstrCost(unsigned Opcode) const;
555 /// \returns The expected cost of compare and select instructions.
556 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
557 Type *CondTy = nullptr) const;
559 /// \return The expected cost of vector Insert and Extract.
560 /// Use -1 to indicate that there is no information on the index value.
561 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index = -1) const;
563 /// \return The cost of Load and Store instructions.
564 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
565 unsigned AddressSpace) const;
567 /// \return The cost of masked Load and Store instructions.
568 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
569 unsigned AddressSpace) const;
571 /// \return The cost of Gather or Scatter operation
572 /// \p Opcode - is a type of memory access Load or Store
573 /// \p DataTy - a vector type of the data to be loaded or stored
574 /// \p Ptr - pointer [or vector of pointers] - address[es] in memory
575 /// \p VariableMask - true when the memory access is predicated with a mask
576 /// that is not a compile-time constant
577 /// \p Alignment - alignment of single element
578 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
579 bool VariableMask, unsigned Alignment) const;
581 /// \return The cost of the interleaved memory operation.
582 /// \p Opcode is the memory operation code
583 /// \p VecTy is the vector type of the interleaved access.
584 /// \p Factor is the interleave factor
585 /// \p Indices is the indices for interleaved load members (as interleaved
586 /// load allows gaps)
587 /// \p Alignment is the alignment of the memory operation
588 /// \p AddressSpace is address space of the pointer.
589 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
590 ArrayRef<unsigned> Indices, unsigned Alignment,
591 unsigned AddressSpace) const;
593 /// \brief Calculate the cost of performing a vector reduction.
595 /// This is the cost of reducing the vector value of type \p Ty to a scalar
596 /// value using the operation denoted by \p Opcode. The form of the reduction
597 /// can either be a pairwise reduction or a reduction that splits the vector
598 /// at every reduction level.
602 /// ((v0+v1), (v2, v3), undef, undef)
605 /// ((v0+v2), (v1+v3), undef, undef)
606 int getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwiseForm) const;
608 /// \returns The cost of Intrinsic instructions. Types analysis only.
609 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
610 ArrayRef<Type *> Tys, FastMathFlags FMF) const;
612 /// \returns The cost of Intrinsic instructions. Analyses the real arguments.
613 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
614 ArrayRef<Value *> Args, FastMathFlags FMF) const;
616 /// \returns The cost of Call instructions.
617 int getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) const;
619 /// \returns The number of pieces into which the provided type must be
620 /// split during legalization. Zero is returned when the answer is unknown.
621 unsigned getNumberOfParts(Type *Tp) const;
623 /// \returns The cost of the address computation. For most targets this can be
624 /// merged into the instruction indexing mode. Some targets might want to
625 /// distinguish between address computation for memory operations on vector
626 /// types and scalar types. Such targets should override this function.
627 /// The 'SE' parameter holds pointer for the scalar evolution object which
628 /// is used in order to get the Ptr step value in case of constant stride.
629 /// The 'Ptr' parameter holds SCEV of the access pointer.
630 int getAddressComputationCost(Type *Ty, ScalarEvolution *SE = nullptr,
631 const SCEV *Ptr = nullptr) const;
633 /// \returns The cost, if any, of keeping values of the given types alive
636 /// Some types may require the use of register classes that do not have
637 /// any callee-saved registers, so would require a spill and fill.
638 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
640 /// \returns True if the intrinsic is a supported memory intrinsic. Info
641 /// will contain additional information - whether the intrinsic may write
642 /// or read to memory, volatility and the pointer. Info is undefined
643 /// if false is returned.
644 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
646 /// \returns A value which is the result of the given memory intrinsic. New
647 /// instructions may be created to extract the result from the given intrinsic
648 /// memory operation. Returns nullptr if the target cannot create a result
649 /// from the given intrinsic.
650 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
651 Type *ExpectedType) const;
653 /// \returns True if the two functions have compatible attributes for inlining
655 bool areInlineCompatible(const Function *Caller,
656 const Function *Callee) const;
658 /// \returns The bitwidth of the largest vector type that should be used to
659 /// load/store in the given address space.
660 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
662 /// \returns True if the load instruction is legal to vectorize.
663 bool isLegalToVectorizeLoad(LoadInst *LI) const;
665 /// \returns True if the store instruction is legal to vectorize.
666 bool isLegalToVectorizeStore(StoreInst *SI) const;
668 /// \returns True if it is legal to vectorize the given load chain.
669 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
671 unsigned AddrSpace) const;
673 /// \returns True if it is legal to vectorize the given store chain.
674 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
676 unsigned AddrSpace) const;
678 /// \returns The new vector factor value if the target doesn't support \p
679 /// SizeInBytes loads or has a better vector factor.
680 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
681 unsigned ChainSizeInBytes,
682 VectorType *VecTy) const;
684 /// \returns The new vector factor value if the target doesn't support \p
685 /// SizeInBytes stores or has a better vector factor.
686 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
687 unsigned ChainSizeInBytes,
688 VectorType *VecTy) const;
693 /// \brief The abstract base class used to type erase specific TTI
697 /// \brief The template model for the base class which wraps a concrete
698 /// implementation in a type erased interface.
699 template <typename T> class Model;
701 std::unique_ptr<Concept> TTIImpl;
704 class TargetTransformInfo::Concept {
706 virtual ~Concept() = 0;
707 virtual const DataLayout &getDataLayout() const = 0;
708 virtual int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
709 virtual int getGEPCost(Type *PointeeType, const Value *Ptr,
710 ArrayRef<const Value *> Operands) = 0;
711 virtual int getCallCost(FunctionType *FTy, int NumArgs) = 0;
712 virtual int getCallCost(const Function *F, int NumArgs) = 0;
713 virtual int getCallCost(const Function *F,
714 ArrayRef<const Value *> Arguments) = 0;
715 virtual unsigned getInliningThresholdMultiplier() = 0;
716 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
717 ArrayRef<Type *> ParamTys) = 0;
718 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
719 ArrayRef<const Value *> Arguments) = 0;
720 virtual int getUserCost(const User *U) = 0;
721 virtual bool hasBranchDivergence() = 0;
722 virtual bool isSourceOfDivergence(const Value *V) = 0;
723 virtual bool isLoweredToCall(const Function *F) = 0;
724 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
725 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
726 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
727 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
728 int64_t BaseOffset, bool HasBaseReg,
730 unsigned AddrSpace) = 0;
731 virtual bool isLegalMaskedStore(Type *DataType) = 0;
732 virtual bool isLegalMaskedLoad(Type *DataType) = 0;
733 virtual bool isLegalMaskedScatter(Type *DataType) = 0;
734 virtual bool isLegalMaskedGather(Type *DataType) = 0;
735 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
736 int64_t BaseOffset, bool HasBaseReg,
737 int64_t Scale, unsigned AddrSpace) = 0;
738 virtual bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) = 0;
739 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
740 virtual bool isProfitableToHoist(Instruction *I) = 0;
741 virtual bool isTypeLegal(Type *Ty) = 0;
742 virtual unsigned getJumpBufAlignment() = 0;
743 virtual unsigned getJumpBufSize() = 0;
744 virtual bool shouldBuildLookupTables() = 0;
745 virtual bool shouldBuildLookupTablesForConstant(Constant *C) = 0;
746 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
747 virtual bool enableInterleavedAccessVectorization() = 0;
748 virtual bool isFPVectorizationPotentiallyUnsafe() = 0;
749 virtual bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
751 unsigned AddressSpace,
754 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
755 virtual bool haveFastSqrt(Type *Ty) = 0;
756 virtual int getFPOpCost(Type *Ty) = 0;
757 virtual int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
759 virtual int getIntImmCost(const APInt &Imm, Type *Ty) = 0;
760 virtual int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
762 virtual int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
764 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
765 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
766 virtual unsigned getCacheLineSize() = 0;
767 virtual unsigned getPrefetchDistance() = 0;
768 virtual unsigned getMinPrefetchStride() = 0;
769 virtual unsigned getMaxPrefetchIterationsAhead() = 0;
770 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
772 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
773 OperandValueKind Opd2Info,
774 OperandValueProperties Opd1PropInfo,
775 OperandValueProperties Opd2PropInfo,
776 ArrayRef<const Value *> Args) = 0;
777 virtual int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
779 virtual int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
780 virtual int getExtractWithExtendCost(unsigned Opcode, Type *Dst,
781 VectorType *VecTy, unsigned Index) = 0;
782 virtual int getCFInstrCost(unsigned Opcode) = 0;
783 virtual int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
785 virtual int getVectorInstrCost(unsigned Opcode, Type *Val,
787 virtual int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
788 unsigned AddressSpace) = 0;
789 virtual int getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
791 unsigned AddressSpace) = 0;
792 virtual int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
793 Value *Ptr, bool VariableMask,
794 unsigned Alignment) = 0;
795 virtual int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
797 ArrayRef<unsigned> Indices,
799 unsigned AddressSpace) = 0;
800 virtual int getReductionCost(unsigned Opcode, Type *Ty,
801 bool IsPairwiseForm) = 0;
802 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
803 ArrayRef<Type *> Tys,
804 FastMathFlags FMF) = 0;
805 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
806 ArrayRef<Value *> Args,
807 FastMathFlags FMF) = 0;
808 virtual int getCallInstrCost(Function *F, Type *RetTy,
809 ArrayRef<Type *> Tys) = 0;
810 virtual unsigned getNumberOfParts(Type *Tp) = 0;
811 virtual int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
812 const SCEV *Ptr) = 0;
813 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
814 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
815 MemIntrinsicInfo &Info) = 0;
816 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
817 Type *ExpectedType) = 0;
818 virtual bool areInlineCompatible(const Function *Caller,
819 const Function *Callee) const = 0;
820 virtual unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const = 0;
821 virtual bool isLegalToVectorizeLoad(LoadInst *LI) const = 0;
822 virtual bool isLegalToVectorizeStore(StoreInst *SI) const = 0;
823 virtual bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
825 unsigned AddrSpace) const = 0;
826 virtual bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
828 unsigned AddrSpace) const = 0;
829 virtual unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
830 unsigned ChainSizeInBytes,
831 VectorType *VecTy) const = 0;
832 virtual unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
833 unsigned ChainSizeInBytes,
834 VectorType *VecTy) const = 0;
837 template <typename T>
838 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
842 Model(T Impl) : Impl(std::move(Impl)) {}
845 const DataLayout &getDataLayout() const override {
846 return Impl.getDataLayout();
849 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
850 return Impl.getOperationCost(Opcode, Ty, OpTy);
852 int getGEPCost(Type *PointeeType, const Value *Ptr,
853 ArrayRef<const Value *> Operands) override {
854 return Impl.getGEPCost(PointeeType, Ptr, Operands);
856 int getCallCost(FunctionType *FTy, int NumArgs) override {
857 return Impl.getCallCost(FTy, NumArgs);
859 int getCallCost(const Function *F, int NumArgs) override {
860 return Impl.getCallCost(F, NumArgs);
862 int getCallCost(const Function *F,
863 ArrayRef<const Value *> Arguments) override {
864 return Impl.getCallCost(F, Arguments);
866 unsigned getInliningThresholdMultiplier() override {
867 return Impl.getInliningThresholdMultiplier();
869 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
870 ArrayRef<Type *> ParamTys) override {
871 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
873 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
874 ArrayRef<const Value *> Arguments) override {
875 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
877 int getUserCost(const User *U) override { return Impl.getUserCost(U); }
878 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
879 bool isSourceOfDivergence(const Value *V) override {
880 return Impl.isSourceOfDivergence(V);
882 bool isLoweredToCall(const Function *F) override {
883 return Impl.isLoweredToCall(F);
885 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
886 return Impl.getUnrollingPreferences(L, UP);
888 bool isLegalAddImmediate(int64_t Imm) override {
889 return Impl.isLegalAddImmediate(Imm);
891 bool isLegalICmpImmediate(int64_t Imm) override {
892 return Impl.isLegalICmpImmediate(Imm);
894 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
895 bool HasBaseReg, int64_t Scale,
896 unsigned AddrSpace) override {
897 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
900 bool isLegalMaskedStore(Type *DataType) override {
901 return Impl.isLegalMaskedStore(DataType);
903 bool isLegalMaskedLoad(Type *DataType) override {
904 return Impl.isLegalMaskedLoad(DataType);
906 bool isLegalMaskedScatter(Type *DataType) override {
907 return Impl.isLegalMaskedScatter(DataType);
909 bool isLegalMaskedGather(Type *DataType) override {
910 return Impl.isLegalMaskedGather(DataType);
912 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
913 bool HasBaseReg, int64_t Scale,
914 unsigned AddrSpace) override {
915 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
918 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) override {
919 return Impl.isFoldableMemAccessOffset(I, Offset);
921 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
922 return Impl.isTruncateFree(Ty1, Ty2);
924 bool isProfitableToHoist(Instruction *I) override {
925 return Impl.isProfitableToHoist(I);
927 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
928 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
929 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
930 bool shouldBuildLookupTables() override {
931 return Impl.shouldBuildLookupTables();
933 bool shouldBuildLookupTablesForConstant(Constant *C) override {
934 return Impl.shouldBuildLookupTablesForConstant(C);
936 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
937 return Impl.enableAggressiveInterleaving(LoopHasReductions);
939 bool enableInterleavedAccessVectorization() override {
940 return Impl.enableInterleavedAccessVectorization();
942 bool isFPVectorizationPotentiallyUnsafe() override {
943 return Impl.isFPVectorizationPotentiallyUnsafe();
945 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
946 unsigned BitWidth, unsigned AddressSpace,
947 unsigned Alignment, bool *Fast) override {
948 return Impl.allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace,
951 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
952 return Impl.getPopcntSupport(IntTyWidthInBit);
954 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
956 int getFPOpCost(Type *Ty) override { return Impl.getFPOpCost(Ty); }
958 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
960 return Impl.getIntImmCodeSizeCost(Opc, Idx, Imm, Ty);
962 int getIntImmCost(const APInt &Imm, Type *Ty) override {
963 return Impl.getIntImmCost(Imm, Ty);
965 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
967 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
969 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
971 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
973 unsigned getNumberOfRegisters(bool Vector) override {
974 return Impl.getNumberOfRegisters(Vector);
976 unsigned getRegisterBitWidth(bool Vector) override {
977 return Impl.getRegisterBitWidth(Vector);
980 unsigned getCacheLineSize() override {
981 return Impl.getCacheLineSize();
983 unsigned getPrefetchDistance() override { return Impl.getPrefetchDistance(); }
984 unsigned getMinPrefetchStride() override {
985 return Impl.getMinPrefetchStride();
987 unsigned getMaxPrefetchIterationsAhead() override {
988 return Impl.getMaxPrefetchIterationsAhead();
990 unsigned getMaxInterleaveFactor(unsigned VF) override {
991 return Impl.getMaxInterleaveFactor(VF);
994 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
995 OperandValueKind Opd2Info,
996 OperandValueProperties Opd1PropInfo,
997 OperandValueProperties Opd2PropInfo,
998 ArrayRef<const Value *> Args) override {
999 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
1000 Opd1PropInfo, Opd2PropInfo, Args);
1002 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
1003 Type *SubTp) override {
1004 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
1006 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
1007 return Impl.getCastInstrCost(Opcode, Dst, Src);
1009 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
1010 unsigned Index) override {
1011 return Impl.getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
1013 int getCFInstrCost(unsigned Opcode) override {
1014 return Impl.getCFInstrCost(Opcode);
1016 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy) override {
1017 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
1019 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) override {
1020 return Impl.getVectorInstrCost(Opcode, Val, Index);
1022 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1023 unsigned AddressSpace) override {
1024 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
1026 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1027 unsigned AddressSpace) override {
1028 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
1030 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
1031 Value *Ptr, bool VariableMask,
1032 unsigned Alignment) override {
1033 return Impl.getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
1036 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
1037 ArrayRef<unsigned> Indices, unsigned Alignment,
1038 unsigned AddressSpace) override {
1039 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
1040 Alignment, AddressSpace);
1042 int getReductionCost(unsigned Opcode, Type *Ty,
1043 bool IsPairwiseForm) override {
1044 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
1046 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
1047 FastMathFlags FMF) override {
1048 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys, FMF);
1050 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
1051 ArrayRef<Value *> Args,
1052 FastMathFlags FMF) override {
1053 return Impl.getIntrinsicInstrCost(ID, RetTy, Args, FMF);
1055 int getCallInstrCost(Function *F, Type *RetTy,
1056 ArrayRef<Type *> Tys) override {
1057 return Impl.getCallInstrCost(F, RetTy, Tys);
1059 unsigned getNumberOfParts(Type *Tp) override {
1060 return Impl.getNumberOfParts(Tp);
1062 int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
1063 const SCEV *Ptr) override {
1064 return Impl.getAddressComputationCost(Ty, SE, Ptr);
1066 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
1067 return Impl.getCostOfKeepingLiveOverCall(Tys);
1069 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
1070 MemIntrinsicInfo &Info) override {
1071 return Impl.getTgtMemIntrinsic(Inst, Info);
1073 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
1074 Type *ExpectedType) override {
1075 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
1077 bool areInlineCompatible(const Function *Caller,
1078 const Function *Callee) const override {
1079 return Impl.areInlineCompatible(Caller, Callee);
1081 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const override {
1082 return Impl.getLoadStoreVecRegBitWidth(AddrSpace);
1084 bool isLegalToVectorizeLoad(LoadInst *LI) const override {
1085 return Impl.isLegalToVectorizeLoad(LI);
1087 bool isLegalToVectorizeStore(StoreInst *SI) const override {
1088 return Impl.isLegalToVectorizeStore(SI);
1090 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
1092 unsigned AddrSpace) const override {
1093 return Impl.isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1096 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
1098 unsigned AddrSpace) const override {
1099 return Impl.isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1102 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
1103 unsigned ChainSizeInBytes,
1104 VectorType *VecTy) const override {
1105 return Impl.getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1107 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
1108 unsigned ChainSizeInBytes,
1109 VectorType *VecTy) const override {
1110 return Impl.getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1114 template <typename T>
1115 TargetTransformInfo::TargetTransformInfo(T Impl)
1116 : TTIImpl(new Model<T>(Impl)) {}
1118 /// \brief Analysis pass providing the \c TargetTransformInfo.
1120 /// The core idea of the TargetIRAnalysis is to expose an interface through
1121 /// which LLVM targets can analyze and provide information about the middle
1122 /// end's target-independent IR. This supports use cases such as target-aware
1123 /// cost modeling of IR constructs.
1125 /// This is a function analysis because much of the cost modeling for targets
1126 /// is done in a subtarget specific way and LLVM supports compiling different
1127 /// functions targeting different subtargets in order to support runtime
1128 /// dispatch according to the observed subtarget.
1129 class TargetIRAnalysis : public AnalysisInfoMixin<TargetIRAnalysis> {
1131 typedef TargetTransformInfo Result;
1133 /// \brief Default construct a target IR analysis.
1135 /// This will use the module's datalayout to construct a baseline
1136 /// conservative TTI result.
1139 /// \brief Construct an IR analysis pass around a target-provide callback.
1141 /// The callback will be called with a particular function for which the TTI
1142 /// is needed and must return a TTI object for that function.
1143 TargetIRAnalysis(std::function<Result(const Function &)> TTICallback);
1145 // Value semantics. We spell out the constructors for MSVC.
1146 TargetIRAnalysis(const TargetIRAnalysis &Arg)
1147 : TTICallback(Arg.TTICallback) {}
1148 TargetIRAnalysis(TargetIRAnalysis &&Arg)
1149 : TTICallback(std::move(Arg.TTICallback)) {}
1150 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
1151 TTICallback = RHS.TTICallback;
1154 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
1155 TTICallback = std::move(RHS.TTICallback);
1159 Result run(const Function &F, FunctionAnalysisManager &);
1162 friend AnalysisInfoMixin<TargetIRAnalysis>;
1163 static AnalysisKey Key;
1165 /// \brief The callback used to produce a result.
1167 /// We use a completely opaque callback so that targets can provide whatever
1168 /// mechanism they desire for constructing the TTI for a given function.
1170 /// FIXME: Should we really use std::function? It's relatively inefficient.
1171 /// It might be possible to arrange for even stateful callbacks to outlive
1172 /// the analysis and thus use a function_ref which would be lighter weight.
1173 /// This may also be less error prone as the callback is likely to reference
1174 /// the external TargetMachine, and that reference needs to never dangle.
1175 std::function<Result(const Function &)> TTICallback;
1177 /// \brief Helper function used as the callback in the default constructor.
1178 static Result getDefaultTTI(const Function &F);
1181 /// \brief Wrapper pass for TargetTransformInfo.
1183 /// This pass can be constructed from a TTI object which it stores internally
1184 /// and is queried by passes.
1185 class TargetTransformInfoWrapperPass : public ImmutablePass {
1186 TargetIRAnalysis TIRA;
1187 Optional<TargetTransformInfo> TTI;
1189 virtual void anchor();
1194 /// \brief We must provide a default constructor for the pass but it should
1197 /// Use the constructor below or call one of the creation routines.
1198 TargetTransformInfoWrapperPass();
1200 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1202 TargetTransformInfo &getTTI(const Function &F);
1205 /// \brief Create an analysis pass wrapper around a TTI object.
1207 /// This analysis pass just holds the TTI instance and makes it available to
1209 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1211 } // End llvm namespace