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 {
47 /// This is the pointer that the intrinsic is loading from or storing to.
48 /// If this is non-null, then analysis/optimization passes can assume that
49 /// this intrinsic is functionally equivalent to a load/store from this
51 Value *PtrVal = nullptr;
53 // Ordering for atomic operations.
54 AtomicOrdering Ordering = AtomicOrdering::NotAtomic;
56 // Same Id is set by the target for corresponding load/store intrinsics.
57 unsigned short MatchingId = 0;
60 bool WriteMem = false;
61 bool IsVolatile = false;
63 bool isUnordered() const {
64 return (Ordering == AtomicOrdering::NotAtomic ||
65 Ordering == AtomicOrdering::Unordered) && !IsVolatile;
69 /// \brief This pass provides access to the codegen interfaces that are needed
70 /// for IR-level transformations.
71 class TargetTransformInfo {
73 /// \brief Construct a TTI object using a type implementing the \c Concept
76 /// This is used by targets to construct a TTI wrapping their target-specific
77 /// implementaion that encodes appropriate costs for their target.
78 template <typename T> TargetTransformInfo(T Impl);
80 /// \brief Construct a baseline TTI object using a minimal implementation of
81 /// the \c Concept API below.
83 /// The TTI implementation will reflect the information in the DataLayout
84 /// provided if non-null.
85 explicit TargetTransformInfo(const DataLayout &DL);
87 // Provide move semantics.
88 TargetTransformInfo(TargetTransformInfo &&Arg);
89 TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
91 // We need to define the destructor out-of-line to define our sub-classes
93 ~TargetTransformInfo();
95 /// \brief Handle the invalidation of this information.
97 /// When used as a result of \c TargetIRAnalysis this method will be called
98 /// when the function this was computed for changes. When it returns false,
99 /// the information is preserved across those changes.
100 bool invalidate(Function &, const PreservedAnalyses &,
101 FunctionAnalysisManager::Invalidator &) {
102 // FIXME: We should probably in some way ensure that the subtarget
103 // information for a function hasn't changed.
107 /// \name Generic Target Information
110 /// \brief Underlying constants for 'cost' values in this interface.
112 /// Many APIs in this interface return a cost. This enum defines the
113 /// fundamental values that should be used to interpret (and produce) those
114 /// costs. The costs are returned as an int rather than a member of this
115 /// enumeration because it is expected that the cost of one IR instruction
116 /// may have a multiplicative factor to it or otherwise won't fit directly
117 /// into the enum. Moreover, it is common to sum or average costs which works
118 /// better as simple integral values. Thus this enum only provides constants.
119 /// Also note that the returned costs are signed integers to make it natural
120 /// to add, subtract, and test with zero (a common boundary condition). It is
121 /// not expected that 2^32 is a realistic cost to be modeling at any point.
123 /// Note that these costs should usually reflect the intersection of code-size
124 /// cost and execution cost. A free instruction is typically one that folds
125 /// into another instruction. For example, reg-to-reg moves can often be
126 /// skipped by renaming the registers in the CPU, but they still are encoded
127 /// and thus wouldn't be considered 'free' here.
128 enum TargetCostConstants {
129 TCC_Free = 0, ///< Expected to fold away in lowering.
130 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
131 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
134 /// \brief Estimate the cost of a specific operation when lowered.
136 /// Note that this is designed to work on an arbitrary synthetic opcode, and
137 /// thus work for hypothetical queries before an instruction has even been
138 /// formed. However, this does *not* work for GEPs, and must not be called
139 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
140 /// analyzing a GEP's cost required more information.
142 /// Typically only the result type is required, and the operand type can be
143 /// omitted. However, if the opcode is one of the cast instructions, the
144 /// operand type is required.
146 /// The returned cost is defined in terms of \c TargetCostConstants, see its
147 /// comments for a detailed explanation of the cost values.
148 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy = nullptr) const;
150 /// \brief Estimate the cost of a GEP operation when lowered.
152 /// The contract for this function is the same as \c getOperationCost except
153 /// that it supports an interface that provides extra information specific to
154 /// the GEP operation.
155 int getGEPCost(Type *PointeeType, const Value *Ptr,
156 ArrayRef<const Value *> Operands) const;
158 /// \brief Estimate the cost of a function call when lowered.
160 /// The contract for this is the same as \c getOperationCost except that it
161 /// supports an interface that provides extra information specific to call
164 /// This is the most basic query for estimating call cost: it only knows the
165 /// function type and (potentially) the number of arguments at the call site.
166 /// The latter is only interesting for varargs function types.
167 int getCallCost(FunctionType *FTy, int NumArgs = -1) const;
169 /// \brief Estimate the cost of calling a specific function when lowered.
171 /// This overload adds the ability to reason about the particular function
172 /// being called in the event it is a library call with special lowering.
173 int getCallCost(const Function *F, int NumArgs = -1) const;
175 /// \brief Estimate the cost of calling a specific function when lowered.
177 /// This overload allows specifying a set of candidate argument values.
178 int getCallCost(const Function *F, ArrayRef<const Value *> Arguments) const;
180 /// \returns A value by which our inlining threshold should be multiplied.
181 /// This is primarily used to bump up the inlining threshold wholesale on
182 /// targets where calls are unusually expensive.
184 /// TODO: This is a rather blunt instrument. Perhaps altering the costs of
185 /// individual classes of instructions would be better.
186 unsigned getInliningThresholdMultiplier() const;
188 /// \brief Estimate the cost of an intrinsic when lowered.
190 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
191 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
192 ArrayRef<Type *> ParamTys) const;
194 /// \brief Estimate the cost of an intrinsic when lowered.
196 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
197 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
198 ArrayRef<const Value *> Arguments) const;
200 /// \return The estimated number of case clusters when lowering \p 'SI'.
201 /// \p JTSize Set a jump table size only when \p SI is suitable for a jump
203 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
204 unsigned &JTSize) const;
206 /// \brief Estimate the cost of a given IR user when lowered.
208 /// This can estimate the cost of either a ConstantExpr or Instruction when
209 /// lowered. It has two primary advantages over the \c getOperationCost and
210 /// \c getGEPCost above, and one significant disadvantage: it can only be
211 /// used when the IR construct has already been formed.
213 /// The advantages are that it can inspect the SSA use graph to reason more
214 /// accurately about the cost. For example, all-constant-GEPs can often be
215 /// folded into a load or other instruction, but if they are used in some
216 /// other context they may not be folded. This routine can distinguish such
219 /// The returned cost is defined in terms of \c TargetCostConstants, see its
220 /// comments for a detailed explanation of the cost values.
221 int getUserCost(const User *U) const;
223 /// \brief Return true if branch divergence exists.
225 /// Branch divergence has a significantly negative impact on GPU performance
226 /// when threads in the same wavefront take different paths due to conditional
228 bool hasBranchDivergence() const;
230 /// \brief Returns whether V is a source of divergence.
232 /// This function provides the target-dependent information for
233 /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
234 /// builds the dependency graph, and then runs the reachability algorithm
235 /// starting with the sources of divergence.
236 bool isSourceOfDivergence(const Value *V) const;
238 /// Returns the address space ID for a target's 'flat' address space. Note
239 /// this is not necessarily the same as addrspace(0), which LLVM sometimes
240 /// refers to as the generic address space. The flat address space is a
241 /// generic address space that can be used access multiple segments of memory
242 /// with different address spaces. Access of a memory location through a
243 /// pointer with this address space is expected to be legal but slower
244 /// compared to the same memory location accessed through a pointer with a
245 /// different address space.
247 /// This is for for targets with different pointer representations which can
248 /// be converted with the addrspacecast instruction. If a pointer is converted
249 /// to this address space, optimizations should attempt to replace the access
250 /// with the source address space.
252 /// \returns ~0u if the target does not have such a flat address space to
254 unsigned getFlatAddressSpace() const;
256 /// \brief Test whether calls to a function lower to actual program function
259 /// The idea is to test whether the program is likely to require a 'call'
260 /// instruction or equivalent in order to call the given function.
262 /// FIXME: It's not clear that this is a good or useful query API. Client's
263 /// should probably move to simpler cost metrics using the above.
264 /// Alternatively, we could split the cost interface into distinct code-size
265 /// and execution-speed costs. This would allow modelling the core of this
266 /// query more accurately as a call is a single small instruction, but
267 /// incurs significant execution cost.
268 bool isLoweredToCall(const Function *F) const;
270 /// Parameters that control the generic loop unrolling transformation.
271 struct UnrollingPreferences {
272 /// The cost threshold for the unrolled loop. Should be relative to the
273 /// getUserCost values returned by this API, and the expectation is that
274 /// the unrolled loop's instructions when run through that interface should
275 /// not exceed this cost. However, this is only an estimate. Also, specific
276 /// loops may be unrolled even with a cost above this threshold if deemed
277 /// profitable. Set this to UINT_MAX to disable the loop body cost
280 /// If complete unrolling will reduce the cost of the loop, we will boost
281 /// the Threshold by a certain percent to allow more aggressive complete
282 /// unrolling. This value provides the maximum boost percentage that we
283 /// can apply to Threshold (The value should be no less than 100).
284 /// BoostedThreshold = Threshold * min(RolledCost / UnrolledCost,
285 /// MaxPercentThresholdBoost / 100)
286 /// E.g. if complete unrolling reduces the loop execution time by 50%
287 /// then we boost the threshold by the factor of 2x. If unrolling is not
288 /// expected to reduce the running time, then we do not increase the
290 unsigned MaxPercentThresholdBoost;
291 /// The cost threshold for the unrolled loop when optimizing for size (set
292 /// to UINT_MAX to disable).
293 unsigned OptSizeThreshold;
294 /// The cost threshold for the unrolled loop, like Threshold, but used
295 /// for partial/runtime unrolling (set to UINT_MAX to disable).
296 unsigned PartialThreshold;
297 /// The cost threshold for the unrolled loop when optimizing for size, like
298 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
299 /// UINT_MAX to disable).
300 unsigned PartialOptSizeThreshold;
301 /// A forced unrolling factor (the number of concatenated bodies of the
302 /// original loop in the unrolled loop body). When set to 0, the unrolling
303 /// transformation will select an unrolling factor based on the current cost
304 /// threshold and other factors.
306 /// A forced peeling factor (the number of bodied of the original loop
307 /// that should be peeled off before the loop body). When set to 0, the
308 /// unrolling transformation will select a peeling factor based on profile
309 /// information and other factors.
311 /// Default unroll count for loops with run-time trip count.
312 unsigned DefaultUnrollRuntimeCount;
313 // Set the maximum unrolling factor. The unrolling factor may be selected
314 // using the appropriate cost threshold, but may not exceed this number
315 // (set to UINT_MAX to disable). This does not apply in cases where the
316 // loop is being fully unrolled.
318 /// Set the maximum unrolling factor for full unrolling. Like MaxCount, but
319 /// applies even if full unrolling is selected. This allows a target to fall
320 /// back to Partial unrolling if full unrolling is above FullUnrollMaxCount.
321 unsigned FullUnrollMaxCount;
322 // Represents number of instructions optimized when "back edge"
323 // becomes "fall through" in unrolled loop.
324 // For now we count a conditional branch on a backedge and a comparison
327 /// Allow partial unrolling (unrolling of loops to expand the size of the
328 /// loop body, not only to eliminate small constant-trip-count loops).
330 /// Allow runtime unrolling (unrolling of loops to expand the size of the
331 /// loop body even when the number of loop iterations is not known at
334 /// Allow generation of a loop remainder (extra iterations after unroll).
336 /// Allow emitting expensive instructions (such as divisions) when computing
337 /// the trip count of a loop for runtime unrolling.
338 bool AllowExpensiveTripCount;
339 /// Apply loop unroll on any kind of loop
340 /// (mainly to loops that fail runtime unrolling).
342 /// Allow using trip count upper bound to unroll loops.
344 /// Allow peeling off loop iterations for loops with low dynamic tripcount.
348 /// \brief Get target-customized preferences for the generic loop unrolling
349 /// transformation. The caller will initialize UP with the current
350 /// target-independent defaults.
351 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
355 /// \name Scalar Target Information
358 /// \brief Flags indicating the kind of support for population count.
360 /// Compared to the SW implementation, HW support is supposed to
361 /// significantly boost the performance when the population is dense, and it
362 /// may or may not degrade performance if the population is sparse. A HW
363 /// support is considered as "Fast" if it can outperform, or is on a par
364 /// with, SW implementation when the population is sparse; otherwise, it is
365 /// considered as "Slow".
366 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
368 /// \brief Return true if the specified immediate is legal add immediate, that
369 /// is the target has add instructions which can add a register with the
370 /// immediate without having to materialize the immediate into a register.
371 bool isLegalAddImmediate(int64_t Imm) const;
373 /// \brief Return true if the specified immediate is legal icmp immediate,
374 /// that is the target has icmp instructions which can compare a register
375 /// against the immediate without having to materialize the immediate into a
377 bool isLegalICmpImmediate(int64_t Imm) const;
379 /// \brief Return true if the addressing mode represented by AM is legal for
380 /// this target, for a load/store of the specified type.
381 /// The type may be VoidTy, in which case only return true if the addressing
382 /// mode is legal for a load/store of any legal type.
383 /// TODO: Handle pre/postinc as well.
384 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
385 bool HasBaseReg, int64_t Scale,
386 unsigned AddrSpace = 0) const;
388 /// \brief Return true if the target supports masked load/store
389 /// AVX2 and AVX-512 targets allow masks for consecutive load and store
390 bool isLegalMaskedStore(Type *DataType) const;
391 bool isLegalMaskedLoad(Type *DataType) const;
393 /// \brief Return true if the target supports masked gather/scatter
394 /// AVX-512 fully supports gather and scatter for vectors with 32 and 64
395 /// bits scalar type.
396 bool isLegalMaskedScatter(Type *DataType) const;
397 bool isLegalMaskedGather(Type *DataType) const;
399 /// \brief Return the cost of the scaling factor used in the addressing
400 /// mode represented by AM for this target, for a load/store
401 /// of the specified type.
402 /// If the AM is supported, the return value must be >= 0.
403 /// If the AM is not supported, it returns a negative value.
404 /// TODO: Handle pre/postinc as well.
405 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
406 bool HasBaseReg, int64_t Scale,
407 unsigned AddrSpace = 0) const;
409 /// \brief Return true if target supports the load / store
410 /// instruction with the given Offset on the form reg + Offset. It
411 /// may be that Offset is too big for a certain type (register
413 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) const;
415 /// \brief Return true if it's free to truncate a value of type Ty1 to type
416 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
417 /// by referencing its sub-register AX.
418 bool isTruncateFree(Type *Ty1, Type *Ty2) const;
420 /// \brief Return true if it is profitable to hoist instruction in the
421 /// then/else to before if.
422 bool isProfitableToHoist(Instruction *I) const;
424 /// \brief Return true if this type is legal.
425 bool isTypeLegal(Type *Ty) const;
427 /// \brief Returns the target's jmp_buf alignment in bytes.
428 unsigned getJumpBufAlignment() const;
430 /// \brief Returns the target's jmp_buf size in bytes.
431 unsigned getJumpBufSize() const;
433 /// \brief Return true if switches should be turned into lookup tables for the
435 bool shouldBuildLookupTables() const;
437 /// \brief Return true if switches should be turned into lookup tables
438 /// containing this constant value for the target.
439 bool shouldBuildLookupTablesForConstant(Constant *C) const;
441 unsigned getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) const;
443 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
446 /// If target has efficient vector element load/store instructions, it can
447 /// return true here so that insertion/extraction costs are not added to
448 /// the scalarization cost of a load/store.
449 bool supportsEfficientVectorElementLoadStore() const;
451 /// \brief Don't restrict interleaved unrolling to small loops.
452 bool enableAggressiveInterleaving(bool LoopHasReductions) const;
454 /// \brief Enable matching of interleaved access groups.
455 bool enableInterleavedAccessVectorization() const;
457 /// \brief Indicate that it is potentially unsafe to automatically vectorize
458 /// floating-point operations because the semantics of vector and scalar
459 /// floating-point semantics may differ. For example, ARM NEON v7 SIMD math
460 /// does not support IEEE-754 denormal numbers, while depending on the
461 /// platform, scalar floating-point math does.
462 /// This applies to floating-point math operations and calls, not memory
463 /// operations, shuffles, or casts.
464 bool isFPVectorizationPotentiallyUnsafe() const;
466 /// \brief Determine if the target supports unaligned memory accesses.
467 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
468 unsigned BitWidth, unsigned AddressSpace = 0,
469 unsigned Alignment = 1,
470 bool *Fast = nullptr) const;
472 /// \brief Return hardware support for population count.
473 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
475 /// \brief Return true if the hardware has a fast square-root instruction.
476 bool haveFastSqrt(Type *Ty) const;
478 /// \brief Return the expected cost of supporting the floating point operation
479 /// of the specified type.
480 int getFPOpCost(Type *Ty) const;
482 /// \brief Return the expected cost of materializing for the given integer
483 /// immediate of the specified type.
484 int getIntImmCost(const APInt &Imm, Type *Ty) const;
486 /// \brief Return the expected cost of materialization for the given integer
487 /// immediate of the specified type for a given instruction. The cost can be
488 /// zero if the immediate can be folded into the specified instruction.
489 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
491 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
494 /// \brief Return the expected cost for the given integer when optimising
495 /// for size. This is different than the other integer immediate cost
496 /// functions in that it is subtarget agnostic. This is useful when you e.g.
497 /// target one ISA such as Aarch32 but smaller encodings could be possible
498 /// with another such as Thumb. This return value is used as a penalty when
499 /// the total costs for a constant is calculated (the bigger the cost, the
500 /// more beneficial constant hoisting is).
501 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
505 /// \name Vector Target Information
508 /// \brief The various kinds of shuffle patterns for vector queries.
510 SK_Broadcast, ///< Broadcast element 0 to all other elements.
511 SK_Reverse, ///< Reverse the order of the vector.
512 SK_Alternate, ///< Choose alternate elements from vector.
513 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
514 SK_ExtractSubvector,///< ExtractSubvector Index indicates start offset.
515 SK_PermuteTwoSrc, ///< Merge elements from two source vectors into one
516 ///< with any shuffle mask.
517 SK_PermuteSingleSrc ///< Shuffle elements of single source vector with any
521 /// \brief Additional information about an operand's possible values.
522 enum OperandValueKind {
523 OK_AnyValue, // Operand can have any value.
524 OK_UniformValue, // Operand is uniform (splat of a value).
525 OK_UniformConstantValue, // Operand is uniform constant.
526 OK_NonUniformConstantValue // Operand is a non uniform constant value.
529 /// \brief Additional properties of an operand's values.
530 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
532 /// \return The number of scalar or vector registers that the target has.
533 /// If 'Vectors' is true, it returns the number of vector registers. If it is
534 /// set to false, it returns the number of scalar registers.
535 unsigned getNumberOfRegisters(bool Vector) const;
537 /// \return The width of the largest scalar or vector register type.
538 unsigned getRegisterBitWidth(bool Vector) const;
540 /// \return True if it should be considered for address type promotion.
541 /// \p AllowPromotionWithoutCommonHeader Set true if promoting \p I is
542 /// profitable without finding other extensions fed by the same input.
543 bool shouldConsiderAddressTypePromotion(
544 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const;
546 /// \return The size of a cache line in bytes.
547 unsigned getCacheLineSize() const;
549 /// \return How much before a load we should place the prefetch instruction.
550 /// This is currently measured in number of instructions.
551 unsigned getPrefetchDistance() const;
553 /// \return Some HW prefetchers can handle accesses up to a certain constant
554 /// stride. This is the minimum stride in bytes where it makes sense to start
555 /// adding SW prefetches. The default is 1, i.e. prefetch with any stride.
556 unsigned getMinPrefetchStride() const;
558 /// \return The maximum number of iterations to prefetch ahead. If the
559 /// required number of iterations is more than this number, no prefetching is
561 unsigned getMaxPrefetchIterationsAhead() const;
563 /// \return The maximum interleave factor that any transform should try to
564 /// perform for this target. This number depends on the level of parallelism
565 /// and the number of execution units in the CPU.
566 unsigned getMaxInterleaveFactor(unsigned VF) const;
568 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
569 /// \p Args is an optional argument which holds the instruction operands
570 /// values so the TTI can analyize those values searching for special
571 /// cases\optimizations based on those values.
572 int getArithmeticInstrCost(
573 unsigned Opcode, Type *Ty, OperandValueKind Opd1Info = OK_AnyValue,
574 OperandValueKind Opd2Info = OK_AnyValue,
575 OperandValueProperties Opd1PropInfo = OP_None,
576 OperandValueProperties Opd2PropInfo = OP_None,
577 ArrayRef<const Value *> Args = ArrayRef<const Value *>()) const;
579 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
580 /// The index and subtype parameters are used by the subvector insertion and
581 /// extraction shuffle kinds.
582 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
583 Type *SubTp = nullptr) const;
585 /// \return The expected cost of cast instructions, such as bitcast, trunc,
586 /// zext, etc. If there is an existing instruction that holds Opcode, it
587 /// may be passed in the 'I' parameter.
588 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
589 const Instruction *I = nullptr) const;
591 /// \return The expected cost of a sign- or zero-extended vector extract. Use
592 /// -1 to indicate that there is no information about the index value.
593 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
594 unsigned Index = -1) const;
596 /// \return The expected cost of control-flow related instructions such as
598 int getCFInstrCost(unsigned Opcode) const;
600 /// \returns The expected cost of compare and select instructions. If there
601 /// is an existing instruction that holds Opcode, it may be passed in the
603 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
604 Type *CondTy = nullptr, const Instruction *I = nullptr) const;
606 /// \return The expected cost of vector Insert and Extract.
607 /// Use -1 to indicate that there is no information on the index value.
608 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index = -1) const;
610 /// \return The cost of Load and Store instructions.
611 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
612 unsigned AddressSpace, const Instruction *I = nullptr) const;
614 /// \return The cost of masked Load and Store instructions.
615 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
616 unsigned AddressSpace) const;
618 /// \return The cost of Gather or Scatter operation
619 /// \p Opcode - is a type of memory access Load or Store
620 /// \p DataTy - a vector type of the data to be loaded or stored
621 /// \p Ptr - pointer [or vector of pointers] - address[es] in memory
622 /// \p VariableMask - true when the memory access is predicated with a mask
623 /// that is not a compile-time constant
624 /// \p Alignment - alignment of single element
625 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy, Value *Ptr,
626 bool VariableMask, unsigned Alignment) const;
628 /// \return The cost of the interleaved memory operation.
629 /// \p Opcode is the memory operation code
630 /// \p VecTy is the vector type of the interleaved access.
631 /// \p Factor is the interleave factor
632 /// \p Indices is the indices for interleaved load members (as interleaved
633 /// load allows gaps)
634 /// \p Alignment is the alignment of the memory operation
635 /// \p AddressSpace is address space of the pointer.
636 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
637 ArrayRef<unsigned> Indices, unsigned Alignment,
638 unsigned AddressSpace) const;
640 /// \brief Calculate the cost of performing a vector reduction.
642 /// This is the cost of reducing the vector value of type \p Ty to a scalar
643 /// value using the operation denoted by \p Opcode. The form of the reduction
644 /// can either be a pairwise reduction or a reduction that splits the vector
645 /// at every reduction level.
649 /// ((v0+v1), (v2, v3), undef, undef)
652 /// ((v0+v2), (v1+v3), undef, undef)
653 int getReductionCost(unsigned Opcode, Type *Ty, bool IsPairwiseForm) const;
655 /// \returns The cost of Intrinsic instructions. Analyses the real arguments.
656 /// Three cases are handled: 1. scalar instruction 2. vector instruction
657 /// 3. scalar instruction which is to be vectorized with VF.
658 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
659 ArrayRef<Value *> Args, FastMathFlags FMF,
660 unsigned VF = 1) const;
662 /// \returns The cost of Intrinsic instructions. Types analysis only.
663 /// If ScalarizationCostPassed is UINT_MAX, the cost of scalarizing the
664 /// arguments and the return value will be computed based on types.
665 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
666 ArrayRef<Type *> Tys, FastMathFlags FMF,
667 unsigned ScalarizationCostPassed = UINT_MAX) const;
669 /// \returns The cost of Call instructions.
670 int getCallInstrCost(Function *F, Type *RetTy, ArrayRef<Type *> Tys) const;
672 /// \returns The number of pieces into which the provided type must be
673 /// split during legalization. Zero is returned when the answer is unknown.
674 unsigned getNumberOfParts(Type *Tp) const;
676 /// \returns The cost of the address computation. For most targets this can be
677 /// merged into the instruction indexing mode. Some targets might want to
678 /// distinguish between address computation for memory operations on vector
679 /// types and scalar types. Such targets should override this function.
680 /// The 'SE' parameter holds pointer for the scalar evolution object which
681 /// is used in order to get the Ptr step value in case of constant stride.
682 /// The 'Ptr' parameter holds SCEV of the access pointer.
683 int getAddressComputationCost(Type *Ty, ScalarEvolution *SE = nullptr,
684 const SCEV *Ptr = nullptr) const;
686 /// \returns The cost, if any, of keeping values of the given types alive
689 /// Some types may require the use of register classes that do not have
690 /// any callee-saved registers, so would require a spill and fill.
691 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
693 /// \returns True if the intrinsic is a supported memory intrinsic. Info
694 /// will contain additional information - whether the intrinsic may write
695 /// or read to memory, volatility and the pointer. Info is undefined
696 /// if false is returned.
697 bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
699 /// \returns A value which is the result of the given memory intrinsic. New
700 /// instructions may be created to extract the result from the given intrinsic
701 /// memory operation. Returns nullptr if the target cannot create a result
702 /// from the given intrinsic.
703 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
704 Type *ExpectedType) const;
706 /// \returns True if the two functions have compatible attributes for inlining
708 bool areInlineCompatible(const Function *Caller,
709 const Function *Callee) const;
711 /// \returns The bitwidth of the largest vector type that should be used to
712 /// load/store in the given address space.
713 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const;
715 /// \returns True if the load instruction is legal to vectorize.
716 bool isLegalToVectorizeLoad(LoadInst *LI) const;
718 /// \returns True if the store instruction is legal to vectorize.
719 bool isLegalToVectorizeStore(StoreInst *SI) const;
721 /// \returns True if it is legal to vectorize the given load chain.
722 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
724 unsigned AddrSpace) const;
726 /// \returns True if it is legal to vectorize the given store chain.
727 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
729 unsigned AddrSpace) const;
731 /// \returns The new vector factor value if the target doesn't support \p
732 /// SizeInBytes loads or has a better vector factor.
733 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
734 unsigned ChainSizeInBytes,
735 VectorType *VecTy) const;
737 /// \returns The new vector factor value if the target doesn't support \p
738 /// SizeInBytes stores or has a better vector factor.
739 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
740 unsigned ChainSizeInBytes,
741 VectorType *VecTy) const;
746 /// \brief The abstract base class used to type erase specific TTI
750 /// \brief The template model for the base class which wraps a concrete
751 /// implementation in a type erased interface.
752 template <typename T> class Model;
754 std::unique_ptr<Concept> TTIImpl;
757 class TargetTransformInfo::Concept {
759 virtual ~Concept() = 0;
760 virtual const DataLayout &getDataLayout() const = 0;
761 virtual int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
762 virtual int getGEPCost(Type *PointeeType, const Value *Ptr,
763 ArrayRef<const Value *> Operands) = 0;
764 virtual int getCallCost(FunctionType *FTy, int NumArgs) = 0;
765 virtual int getCallCost(const Function *F, int NumArgs) = 0;
766 virtual int getCallCost(const Function *F,
767 ArrayRef<const Value *> Arguments) = 0;
768 virtual unsigned getInliningThresholdMultiplier() = 0;
769 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
770 ArrayRef<Type *> ParamTys) = 0;
771 virtual int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
772 ArrayRef<const Value *> Arguments) = 0;
773 virtual unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
774 unsigned &JTSize) = 0;
775 virtual int getUserCost(const User *U) = 0;
776 virtual bool hasBranchDivergence() = 0;
777 virtual bool isSourceOfDivergence(const Value *V) = 0;
778 virtual unsigned getFlatAddressSpace() = 0;
779 virtual bool isLoweredToCall(const Function *F) = 0;
780 virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
781 virtual bool isLegalAddImmediate(int64_t Imm) = 0;
782 virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
783 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
784 int64_t BaseOffset, bool HasBaseReg,
786 unsigned AddrSpace) = 0;
787 virtual bool isLegalMaskedStore(Type *DataType) = 0;
788 virtual bool isLegalMaskedLoad(Type *DataType) = 0;
789 virtual bool isLegalMaskedScatter(Type *DataType) = 0;
790 virtual bool isLegalMaskedGather(Type *DataType) = 0;
791 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
792 int64_t BaseOffset, bool HasBaseReg,
793 int64_t Scale, unsigned AddrSpace) = 0;
794 virtual bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) = 0;
795 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
796 virtual bool isProfitableToHoist(Instruction *I) = 0;
797 virtual bool isTypeLegal(Type *Ty) = 0;
798 virtual unsigned getJumpBufAlignment() = 0;
799 virtual unsigned getJumpBufSize() = 0;
800 virtual bool shouldBuildLookupTables() = 0;
801 virtual bool shouldBuildLookupTablesForConstant(Constant *C) = 0;
803 getScalarizationOverhead(Type *Ty, bool Insert, bool Extract) = 0;
804 virtual unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
806 virtual bool supportsEfficientVectorElementLoadStore() = 0;
807 virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
808 virtual bool enableInterleavedAccessVectorization() = 0;
809 virtual bool isFPVectorizationPotentiallyUnsafe() = 0;
810 virtual bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
812 unsigned AddressSpace,
815 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
816 virtual bool haveFastSqrt(Type *Ty) = 0;
817 virtual int getFPOpCost(Type *Ty) = 0;
818 virtual int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
820 virtual int getIntImmCost(const APInt &Imm, Type *Ty) = 0;
821 virtual int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
823 virtual int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
825 virtual unsigned getNumberOfRegisters(bool Vector) = 0;
826 virtual unsigned getRegisterBitWidth(bool Vector) = 0;
827 virtual bool shouldConsiderAddressTypePromotion(
828 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) = 0;
829 virtual unsigned getCacheLineSize() = 0;
830 virtual unsigned getPrefetchDistance() = 0;
831 virtual unsigned getMinPrefetchStride() = 0;
832 virtual unsigned getMaxPrefetchIterationsAhead() = 0;
833 virtual unsigned getMaxInterleaveFactor(unsigned VF) = 0;
835 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
836 OperandValueKind Opd2Info,
837 OperandValueProperties Opd1PropInfo,
838 OperandValueProperties Opd2PropInfo,
839 ArrayRef<const Value *> Args) = 0;
840 virtual int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
842 virtual int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
843 const Instruction *I) = 0;
844 virtual int getExtractWithExtendCost(unsigned Opcode, Type *Dst,
845 VectorType *VecTy, unsigned Index) = 0;
846 virtual int getCFInstrCost(unsigned Opcode) = 0;
847 virtual int getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
848 Type *CondTy, const Instruction *I) = 0;
849 virtual int getVectorInstrCost(unsigned Opcode, Type *Val,
851 virtual int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
852 unsigned AddressSpace, const Instruction *I) = 0;
853 virtual int getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
855 unsigned AddressSpace) = 0;
856 virtual int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
857 Value *Ptr, bool VariableMask,
858 unsigned Alignment) = 0;
859 virtual int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy,
861 ArrayRef<unsigned> Indices,
863 unsigned AddressSpace) = 0;
864 virtual int getReductionCost(unsigned Opcode, Type *Ty,
865 bool IsPairwiseForm) = 0;
866 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
867 ArrayRef<Type *> Tys, FastMathFlags FMF,
868 unsigned ScalarizationCostPassed) = 0;
869 virtual int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
870 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) = 0;
871 virtual int getCallInstrCost(Function *F, Type *RetTy,
872 ArrayRef<Type *> Tys) = 0;
873 virtual unsigned getNumberOfParts(Type *Tp) = 0;
874 virtual int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
875 const SCEV *Ptr) = 0;
876 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
877 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
878 MemIntrinsicInfo &Info) = 0;
879 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
880 Type *ExpectedType) = 0;
881 virtual bool areInlineCompatible(const Function *Caller,
882 const Function *Callee) const = 0;
883 virtual unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const = 0;
884 virtual bool isLegalToVectorizeLoad(LoadInst *LI) const = 0;
885 virtual bool isLegalToVectorizeStore(StoreInst *SI) const = 0;
886 virtual bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
888 unsigned AddrSpace) const = 0;
889 virtual bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
891 unsigned AddrSpace) const = 0;
892 virtual unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
893 unsigned ChainSizeInBytes,
894 VectorType *VecTy) const = 0;
895 virtual unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
896 unsigned ChainSizeInBytes,
897 VectorType *VecTy) const = 0;
900 template <typename T>
901 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
905 Model(T Impl) : Impl(std::move(Impl)) {}
908 const DataLayout &getDataLayout() const override {
909 return Impl.getDataLayout();
912 int getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
913 return Impl.getOperationCost(Opcode, Ty, OpTy);
915 int getGEPCost(Type *PointeeType, const Value *Ptr,
916 ArrayRef<const Value *> Operands) override {
917 return Impl.getGEPCost(PointeeType, Ptr, Operands);
919 int getCallCost(FunctionType *FTy, int NumArgs) override {
920 return Impl.getCallCost(FTy, NumArgs);
922 int getCallCost(const Function *F, int NumArgs) override {
923 return Impl.getCallCost(F, NumArgs);
925 int getCallCost(const Function *F,
926 ArrayRef<const Value *> Arguments) override {
927 return Impl.getCallCost(F, Arguments);
929 unsigned getInliningThresholdMultiplier() override {
930 return Impl.getInliningThresholdMultiplier();
932 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
933 ArrayRef<Type *> ParamTys) override {
934 return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
936 int getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
937 ArrayRef<const Value *> Arguments) override {
938 return Impl.getIntrinsicCost(IID, RetTy, Arguments);
940 int getUserCost(const User *U) override { return Impl.getUserCost(U); }
941 bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
942 bool isSourceOfDivergence(const Value *V) override {
943 return Impl.isSourceOfDivergence(V);
946 unsigned getFlatAddressSpace() override {
947 return Impl.getFlatAddressSpace();
950 bool isLoweredToCall(const Function *F) override {
951 return Impl.isLoweredToCall(F);
953 void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
954 return Impl.getUnrollingPreferences(L, UP);
956 bool isLegalAddImmediate(int64_t Imm) override {
957 return Impl.isLegalAddImmediate(Imm);
959 bool isLegalICmpImmediate(int64_t Imm) override {
960 return Impl.isLegalICmpImmediate(Imm);
962 bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
963 bool HasBaseReg, int64_t Scale,
964 unsigned AddrSpace) override {
965 return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
968 bool isLegalMaskedStore(Type *DataType) override {
969 return Impl.isLegalMaskedStore(DataType);
971 bool isLegalMaskedLoad(Type *DataType) override {
972 return Impl.isLegalMaskedLoad(DataType);
974 bool isLegalMaskedScatter(Type *DataType) override {
975 return Impl.isLegalMaskedScatter(DataType);
977 bool isLegalMaskedGather(Type *DataType) override {
978 return Impl.isLegalMaskedGather(DataType);
980 int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
981 bool HasBaseReg, int64_t Scale,
982 unsigned AddrSpace) override {
983 return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg,
986 bool isFoldableMemAccessOffset(Instruction *I, int64_t Offset) override {
987 return Impl.isFoldableMemAccessOffset(I, Offset);
989 bool isTruncateFree(Type *Ty1, Type *Ty2) override {
990 return Impl.isTruncateFree(Ty1, Ty2);
992 bool isProfitableToHoist(Instruction *I) override {
993 return Impl.isProfitableToHoist(I);
995 bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
996 unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
997 unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
998 bool shouldBuildLookupTables() override {
999 return Impl.shouldBuildLookupTables();
1001 bool shouldBuildLookupTablesForConstant(Constant *C) override {
1002 return Impl.shouldBuildLookupTablesForConstant(C);
1004 unsigned getScalarizationOverhead(Type *Ty, bool Insert,
1005 bool Extract) override {
1006 return Impl.getScalarizationOverhead(Ty, Insert, Extract);
1008 unsigned getOperandsScalarizationOverhead(ArrayRef<const Value *> Args,
1009 unsigned VF) override {
1010 return Impl.getOperandsScalarizationOverhead(Args, VF);
1013 bool supportsEfficientVectorElementLoadStore() override {
1014 return Impl.supportsEfficientVectorElementLoadStore();
1017 bool enableAggressiveInterleaving(bool LoopHasReductions) override {
1018 return Impl.enableAggressiveInterleaving(LoopHasReductions);
1020 bool enableInterleavedAccessVectorization() override {
1021 return Impl.enableInterleavedAccessVectorization();
1023 bool isFPVectorizationPotentiallyUnsafe() override {
1024 return Impl.isFPVectorizationPotentiallyUnsafe();
1026 bool allowsMisalignedMemoryAccesses(LLVMContext &Context,
1027 unsigned BitWidth, unsigned AddressSpace,
1028 unsigned Alignment, bool *Fast) override {
1029 return Impl.allowsMisalignedMemoryAccesses(Context, BitWidth, AddressSpace,
1032 PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
1033 return Impl.getPopcntSupport(IntTyWidthInBit);
1035 bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
1037 int getFPOpCost(Type *Ty) override { return Impl.getFPOpCost(Ty); }
1039 int getIntImmCodeSizeCost(unsigned Opc, unsigned Idx, const APInt &Imm,
1040 Type *Ty) override {
1041 return Impl.getIntImmCodeSizeCost(Opc, Idx, Imm, Ty);
1043 int getIntImmCost(const APInt &Imm, Type *Ty) override {
1044 return Impl.getIntImmCost(Imm, Ty);
1046 int getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
1047 Type *Ty) override {
1048 return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
1050 int getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
1051 Type *Ty) override {
1052 return Impl.getIntImmCost(IID, Idx, Imm, Ty);
1054 unsigned getNumberOfRegisters(bool Vector) override {
1055 return Impl.getNumberOfRegisters(Vector);
1057 unsigned getRegisterBitWidth(bool Vector) override {
1058 return Impl.getRegisterBitWidth(Vector);
1060 bool shouldConsiderAddressTypePromotion(
1061 const Instruction &I, bool &AllowPromotionWithoutCommonHeader) override {
1062 return Impl.shouldConsiderAddressTypePromotion(
1063 I, AllowPromotionWithoutCommonHeader);
1065 unsigned getCacheLineSize() override {
1066 return Impl.getCacheLineSize();
1068 unsigned getPrefetchDistance() override { return Impl.getPrefetchDistance(); }
1069 unsigned getMinPrefetchStride() override {
1070 return Impl.getMinPrefetchStride();
1072 unsigned getMaxPrefetchIterationsAhead() override {
1073 return Impl.getMaxPrefetchIterationsAhead();
1075 unsigned getMaxInterleaveFactor(unsigned VF) override {
1076 return Impl.getMaxInterleaveFactor(VF);
1078 unsigned getEstimatedNumberOfCaseClusters(const SwitchInst &SI,
1079 unsigned &JTSize) override {
1080 return Impl.getEstimatedNumberOfCaseClusters(SI, JTSize);
1083 getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
1084 OperandValueKind Opd2Info,
1085 OperandValueProperties Opd1PropInfo,
1086 OperandValueProperties Opd2PropInfo,
1087 ArrayRef<const Value *> Args) override {
1088 return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
1089 Opd1PropInfo, Opd2PropInfo, Args);
1091 int getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
1092 Type *SubTp) override {
1093 return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
1095 int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
1096 const Instruction *I) override {
1097 return Impl.getCastInstrCost(Opcode, Dst, Src, I);
1099 int getExtractWithExtendCost(unsigned Opcode, Type *Dst, VectorType *VecTy,
1100 unsigned Index) override {
1101 return Impl.getExtractWithExtendCost(Opcode, Dst, VecTy, Index);
1103 int getCFInstrCost(unsigned Opcode) override {
1104 return Impl.getCFInstrCost(Opcode);
1106 int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
1107 const Instruction *I) override {
1108 return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy, I);
1110 int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index) override {
1111 return Impl.getVectorInstrCost(Opcode, Val, Index);
1113 int getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1114 unsigned AddressSpace, const Instruction *I) override {
1115 return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace, I);
1117 int getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
1118 unsigned AddressSpace) override {
1119 return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
1121 int getGatherScatterOpCost(unsigned Opcode, Type *DataTy,
1122 Value *Ptr, bool VariableMask,
1123 unsigned Alignment) override {
1124 return Impl.getGatherScatterOpCost(Opcode, DataTy, Ptr, VariableMask,
1127 int getInterleavedMemoryOpCost(unsigned Opcode, Type *VecTy, unsigned Factor,
1128 ArrayRef<unsigned> Indices, unsigned Alignment,
1129 unsigned AddressSpace) override {
1130 return Impl.getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
1131 Alignment, AddressSpace);
1133 int getReductionCost(unsigned Opcode, Type *Ty,
1134 bool IsPairwiseForm) override {
1135 return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
1137 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy, ArrayRef<Type *> Tys,
1138 FastMathFlags FMF, unsigned ScalarizationCostPassed) override {
1139 return Impl.getIntrinsicInstrCost(ID, RetTy, Tys, FMF,
1140 ScalarizationCostPassed);
1142 int getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
1143 ArrayRef<Value *> Args, FastMathFlags FMF, unsigned VF) override {
1144 return Impl.getIntrinsicInstrCost(ID, RetTy, Args, FMF, VF);
1146 int getCallInstrCost(Function *F, Type *RetTy,
1147 ArrayRef<Type *> Tys) override {
1148 return Impl.getCallInstrCost(F, RetTy, Tys);
1150 unsigned getNumberOfParts(Type *Tp) override {
1151 return Impl.getNumberOfParts(Tp);
1153 int getAddressComputationCost(Type *Ty, ScalarEvolution *SE,
1154 const SCEV *Ptr) override {
1155 return Impl.getAddressComputationCost(Ty, SE, Ptr);
1157 unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
1158 return Impl.getCostOfKeepingLiveOverCall(Tys);
1160 bool getTgtMemIntrinsic(IntrinsicInst *Inst,
1161 MemIntrinsicInfo &Info) override {
1162 return Impl.getTgtMemIntrinsic(Inst, Info);
1164 Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
1165 Type *ExpectedType) override {
1166 return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
1168 bool areInlineCompatible(const Function *Caller,
1169 const Function *Callee) const override {
1170 return Impl.areInlineCompatible(Caller, Callee);
1172 unsigned getLoadStoreVecRegBitWidth(unsigned AddrSpace) const override {
1173 return Impl.getLoadStoreVecRegBitWidth(AddrSpace);
1175 bool isLegalToVectorizeLoad(LoadInst *LI) const override {
1176 return Impl.isLegalToVectorizeLoad(LI);
1178 bool isLegalToVectorizeStore(StoreInst *SI) const override {
1179 return Impl.isLegalToVectorizeStore(SI);
1181 bool isLegalToVectorizeLoadChain(unsigned ChainSizeInBytes,
1183 unsigned AddrSpace) const override {
1184 return Impl.isLegalToVectorizeLoadChain(ChainSizeInBytes, Alignment,
1187 bool isLegalToVectorizeStoreChain(unsigned ChainSizeInBytes,
1189 unsigned AddrSpace) const override {
1190 return Impl.isLegalToVectorizeStoreChain(ChainSizeInBytes, Alignment,
1193 unsigned getLoadVectorFactor(unsigned VF, unsigned LoadSize,
1194 unsigned ChainSizeInBytes,
1195 VectorType *VecTy) const override {
1196 return Impl.getLoadVectorFactor(VF, LoadSize, ChainSizeInBytes, VecTy);
1198 unsigned getStoreVectorFactor(unsigned VF, unsigned StoreSize,
1199 unsigned ChainSizeInBytes,
1200 VectorType *VecTy) const override {
1201 return Impl.getStoreVectorFactor(VF, StoreSize, ChainSizeInBytes, VecTy);
1205 template <typename T>
1206 TargetTransformInfo::TargetTransformInfo(T Impl)
1207 : TTIImpl(new Model<T>(Impl)) {}
1209 /// \brief Analysis pass providing the \c TargetTransformInfo.
1211 /// The core idea of the TargetIRAnalysis is to expose an interface through
1212 /// which LLVM targets can analyze and provide information about the middle
1213 /// end's target-independent IR. This supports use cases such as target-aware
1214 /// cost modeling of IR constructs.
1216 /// This is a function analysis because much of the cost modeling for targets
1217 /// is done in a subtarget specific way and LLVM supports compiling different
1218 /// functions targeting different subtargets in order to support runtime
1219 /// dispatch according to the observed subtarget.
1220 class TargetIRAnalysis : public AnalysisInfoMixin<TargetIRAnalysis> {
1222 typedef TargetTransformInfo Result;
1224 /// \brief Default construct a target IR analysis.
1226 /// This will use the module's datalayout to construct a baseline
1227 /// conservative TTI result.
1230 /// \brief Construct an IR analysis pass around a target-provide callback.
1232 /// The callback will be called with a particular function for which the TTI
1233 /// is needed and must return a TTI object for that function.
1234 TargetIRAnalysis(std::function<Result(const Function &)> TTICallback);
1236 // Value semantics. We spell out the constructors for MSVC.
1237 TargetIRAnalysis(const TargetIRAnalysis &Arg)
1238 : TTICallback(Arg.TTICallback) {}
1239 TargetIRAnalysis(TargetIRAnalysis &&Arg)
1240 : TTICallback(std::move(Arg.TTICallback)) {}
1241 TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
1242 TTICallback = RHS.TTICallback;
1245 TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
1246 TTICallback = std::move(RHS.TTICallback);
1250 Result run(const Function &F, FunctionAnalysisManager &);
1253 friend AnalysisInfoMixin<TargetIRAnalysis>;
1254 static AnalysisKey Key;
1256 /// \brief The callback used to produce a result.
1258 /// We use a completely opaque callback so that targets can provide whatever
1259 /// mechanism they desire for constructing the TTI for a given function.
1261 /// FIXME: Should we really use std::function? It's relatively inefficient.
1262 /// It might be possible to arrange for even stateful callbacks to outlive
1263 /// the analysis and thus use a function_ref which would be lighter weight.
1264 /// This may also be less error prone as the callback is likely to reference
1265 /// the external TargetMachine, and that reference needs to never dangle.
1266 std::function<Result(const Function &)> TTICallback;
1268 /// \brief Helper function used as the callback in the default constructor.
1269 static Result getDefaultTTI(const Function &F);
1272 /// \brief Wrapper pass for TargetTransformInfo.
1274 /// This pass can be constructed from a TTI object which it stores internally
1275 /// and is queried by passes.
1276 class TargetTransformInfoWrapperPass : public ImmutablePass {
1277 TargetIRAnalysis TIRA;
1278 Optional<TargetTransformInfo> TTI;
1280 virtual void anchor();
1285 /// \brief We must provide a default constructor for the pass but it should
1288 /// Use the constructor below or call one of the creation routines.
1289 TargetTransformInfoWrapperPass();
1291 explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1293 TargetTransformInfo &getTTI(const Function &F);
1296 /// \brief Create an analysis pass wrapper around a TTI object.
1298 /// This analysis pass just holds the TTI instance and makes it available to
1300 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
1302 } // End llvm namespace