1 //===- TargetTransformInfo.h ------------------------------------*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 /// \file
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
14 ///    information.
15 /// 3. Codegen-level implementation which uses target-specific hooks.
16 ///
17 /// This file defines #2, which is the interface that IR-level transformations
18 /// use for querying the codegen.
19 ///
20 //===----------------------------------------------------------------------===//
21 
22 #ifndef LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
23 #define LLVM_ANALYSIS_TARGETTRANSFORMINFO_H
24 
25 #include "llvm/ADT/Optional.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/IR/Intrinsics.h"
28 #include "llvm/Pass.h"
29 #include "llvm/Support/DataTypes.h"
30 #include <functional>
31 
32 namespace llvm {
33 
34 class Function;
35 class GlobalValue;
36 class Loop;
37 class PreservedAnalyses;
38 class Type;
39 class User;
40 class Value;
41 
42 /// \brief Information about a load/store intrinsic defined by the target.
43 struct MemIntrinsicInfo {
MemIntrinsicInfoMemIntrinsicInfo44   MemIntrinsicInfo()
45       : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
46         NumMemRefs(0), PtrVal(nullptr) {}
47   bool ReadMem;
48   bool WriteMem;
49   bool Vol;
50   // Same Id is set by the target for corresponding load/store intrinsics.
51   unsigned short MatchingId;
52   int NumMemRefs;
53   Value *PtrVal;
54 };
55 
56 /// \brief This pass provides access to the codegen interfaces that are needed
57 /// for IR-level transformations.
58 class TargetTransformInfo {
59 public:
60   /// \brief Construct a TTI object using a type implementing the \c Concept
61   /// API below.
62   ///
63   /// This is used by targets to construct a TTI wrapping their target-specific
64   /// implementaion that encodes appropriate costs for their target.
65   template <typename T> TargetTransformInfo(T Impl);
66 
67   /// \brief Construct a baseline TTI object using a minimal implementation of
68   /// the \c Concept API below.
69   ///
70   /// The TTI implementation will reflect the information in the DataLayout
71   /// provided if non-null.
72   explicit TargetTransformInfo(const DataLayout *DL);
73 
74   // Provide move semantics.
75   TargetTransformInfo(TargetTransformInfo &&Arg);
76   TargetTransformInfo &operator=(TargetTransformInfo &&RHS);
77 
78   // We need to define the destructor out-of-line to define our sub-classes
79   // out-of-line.
80   ~TargetTransformInfo();
81 
82   /// \brief Handle the invalidation of this information.
83   ///
84   /// When used as a result of \c TargetIRAnalysis this method will be called
85   /// when the function this was computed for changes. When it returns false,
86   /// the information is preserved across those changes.
invalidate(Function &,const PreservedAnalyses &)87   bool invalidate(Function &, const PreservedAnalyses &) {
88     // FIXME: We should probably in some way ensure that the subtarget
89     // information for a function hasn't changed.
90     return false;
91   }
92 
93   /// \name Generic Target Information
94   /// @{
95 
96   /// \brief Underlying constants for 'cost' values in this interface.
97   ///
98   /// Many APIs in this interface return a cost. This enum defines the
99   /// fundamental values that should be used to interpret (and produce) those
100   /// costs. The costs are returned as an unsigned rather than a member of this
101   /// enumeration because it is expected that the cost of one IR instruction
102   /// may have a multiplicative factor to it or otherwise won't fit directly
103   /// into the enum. Moreover, it is common to sum or average costs which works
104   /// better as simple integral values. Thus this enum only provides constants.
105   ///
106   /// Note that these costs should usually reflect the intersection of code-size
107   /// cost and execution cost. A free instruction is typically one that folds
108   /// into another instruction. For example, reg-to-reg moves can often be
109   /// skipped by renaming the registers in the CPU, but they still are encoded
110   /// and thus wouldn't be considered 'free' here.
111   enum TargetCostConstants {
112     TCC_Free = 0,     ///< Expected to fold away in lowering.
113     TCC_Basic = 1,    ///< The cost of a typical 'add' instruction.
114     TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
115   };
116 
117   /// \brief Estimate the cost of a specific operation when lowered.
118   ///
119   /// Note that this is designed to work on an arbitrary synthetic opcode, and
120   /// thus work for hypothetical queries before an instruction has even been
121   /// formed. However, this does *not* work for GEPs, and must not be called
122   /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
123   /// analyzing a GEP's cost required more information.
124   ///
125   /// Typically only the result type is required, and the operand type can be
126   /// omitted. However, if the opcode is one of the cast instructions, the
127   /// operand type is required.
128   ///
129   /// The returned cost is defined in terms of \c TargetCostConstants, see its
130   /// comments for a detailed explanation of the cost values.
131   unsigned getOperationCost(unsigned Opcode, Type *Ty,
132                             Type *OpTy = nullptr) const;
133 
134   /// \brief Estimate the cost of a GEP operation when lowered.
135   ///
136   /// The contract for this function is the same as \c getOperationCost except
137   /// that it supports an interface that provides extra information specific to
138   /// the GEP operation.
139   unsigned getGEPCost(const Value *Ptr, ArrayRef<const Value *> Operands) const;
140 
141   /// \brief Estimate the cost of a function call when lowered.
142   ///
143   /// The contract for this is the same as \c getOperationCost except that it
144   /// supports an interface that provides extra information specific to call
145   /// instructions.
146   ///
147   /// This is the most basic query for estimating call cost: it only knows the
148   /// function type and (potentially) the number of arguments at the call site.
149   /// The latter is only interesting for varargs function types.
150   unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
151 
152   /// \brief Estimate the cost of calling a specific function when lowered.
153   ///
154   /// This overload adds the ability to reason about the particular function
155   /// being called in the event it is a library call with special lowering.
156   unsigned getCallCost(const Function *F, int NumArgs = -1) const;
157 
158   /// \brief Estimate the cost of calling a specific function when lowered.
159   ///
160   /// This overload allows specifying a set of candidate argument values.
161   unsigned getCallCost(const Function *F,
162                        ArrayRef<const Value *> Arguments) const;
163 
164   /// \brief Estimate the cost of an intrinsic when lowered.
165   ///
166   /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
167   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
168                             ArrayRef<Type *> ParamTys) const;
169 
170   /// \brief Estimate the cost of an intrinsic when lowered.
171   ///
172   /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
173   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
174                             ArrayRef<const Value *> Arguments) const;
175 
176   /// \brief Estimate the cost of a given IR user when lowered.
177   ///
178   /// This can estimate the cost of either a ConstantExpr or Instruction when
179   /// lowered. It has two primary advantages over the \c getOperationCost and
180   /// \c getGEPCost above, and one significant disadvantage: it can only be
181   /// used when the IR construct has already been formed.
182   ///
183   /// The advantages are that it can inspect the SSA use graph to reason more
184   /// accurately about the cost. For example, all-constant-GEPs can often be
185   /// folded into a load or other instruction, but if they are used in some
186   /// other context they may not be folded. This routine can distinguish such
187   /// cases.
188   ///
189   /// The returned cost is defined in terms of \c TargetCostConstants, see its
190   /// comments for a detailed explanation of the cost values.
191   unsigned getUserCost(const User *U) const;
192 
193   /// \brief Return true if branch divergence exists.
194   ///
195   /// Branch divergence has a significantly negative impact on GPU performance
196   /// when threads in the same wavefront take different paths due to conditional
197   /// branches.
198   bool hasBranchDivergence() const;
199 
200   /// \brief Returns whether V is a source of divergence.
201   ///
202   /// This function provides the target-dependent information for
203   /// the target-independent DivergenceAnalysis. DivergenceAnalysis first
204   /// builds the dependency graph, and then runs the reachability algorithm
205   /// starting with the sources of divergence.
206   bool isSourceOfDivergence(const Value *V) const;
207 
208   /// \brief Test whether calls to a function lower to actual program function
209   /// calls.
210   ///
211   /// The idea is to test whether the program is likely to require a 'call'
212   /// instruction or equivalent in order to call the given function.
213   ///
214   /// FIXME: It's not clear that this is a good or useful query API. Client's
215   /// should probably move to simpler cost metrics using the above.
216   /// Alternatively, we could split the cost interface into distinct code-size
217   /// and execution-speed costs. This would allow modelling the core of this
218   /// query more accurately as a call is a single small instruction, but
219   /// incurs significant execution cost.
220   bool isLoweredToCall(const Function *F) const;
221 
222   /// Parameters that control the generic loop unrolling transformation.
223   struct UnrollingPreferences {
224     /// The cost threshold for the unrolled loop, compared to
225     /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
226     /// The unrolling factor is set such that the unrolled loop body does not
227     /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
228     /// restriction.
229     unsigned Threshold;
230     /// If complete unrolling could help other optimizations (e.g. InstSimplify)
231     /// to remove N% of instructions, then we can go beyond unroll threshold.
232     /// This value set the minimal percent for allowing that.
233     unsigned MinPercentOfOptimized;
234     /// The absolute cost threshold. We won't go beyond this even if complete
235     /// unrolling could result in optimizing out 90% of instructions.
236     unsigned AbsoluteThreshold;
237     /// The cost threshold for the unrolled loop when optimizing for size (set
238     /// to UINT_MAX to disable).
239     unsigned OptSizeThreshold;
240     /// The cost threshold for the unrolled loop, like Threshold, but used
241     /// for partial/runtime unrolling (set to UINT_MAX to disable).
242     unsigned PartialThreshold;
243     /// The cost threshold for the unrolled loop when optimizing for size, like
244     /// OptSizeThreshold, but used for partial/runtime unrolling (set to
245     /// UINT_MAX to disable).
246     unsigned PartialOptSizeThreshold;
247     /// A forced unrolling factor (the number of concatenated bodies of the
248     /// original loop in the unrolled loop body). When set to 0, the unrolling
249     /// transformation will select an unrolling factor based on the current cost
250     /// threshold and other factors.
251     unsigned Count;
252     // Set the maximum unrolling factor. The unrolling factor may be selected
253     // using the appropriate cost threshold, but may not exceed this number
254     // (set to UINT_MAX to disable). This does not apply in cases where the
255     // loop is being fully unrolled.
256     unsigned MaxCount;
257     /// Allow partial unrolling (unrolling of loops to expand the size of the
258     /// loop body, not only to eliminate small constant-trip-count loops).
259     bool Partial;
260     /// Allow runtime unrolling (unrolling of loops to expand the size of the
261     /// loop body even when the number of loop iterations is not known at
262     /// compile time).
263     bool Runtime;
264     /// Allow emitting expensive instructions (such as divisions) when computing
265     /// the trip count of a loop for runtime unrolling.
266     bool AllowExpensiveTripCount;
267   };
268 
269   /// \brief Get target-customized preferences for the generic loop unrolling
270   /// transformation. The caller will initialize UP with the current
271   /// target-independent defaults.
272   void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
273 
274   /// @}
275 
276   /// \name Scalar Target Information
277   /// @{
278 
279   /// \brief Flags indicating the kind of support for population count.
280   ///
281   /// Compared to the SW implementation, HW support is supposed to
282   /// significantly boost the performance when the population is dense, and it
283   /// may or may not degrade performance if the population is sparse. A HW
284   /// support is considered as "Fast" if it can outperform, or is on a par
285   /// with, SW implementation when the population is sparse; otherwise, it is
286   /// considered as "Slow".
287   enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
288 
289   /// \brief Return true if the specified immediate is legal add immediate, that
290   /// is the target has add instructions which can add a register with the
291   /// immediate without having to materialize the immediate into a register.
292   bool isLegalAddImmediate(int64_t Imm) const;
293 
294   /// \brief Return true if the specified immediate is legal icmp immediate,
295   /// that is the target has icmp instructions which can compare a register
296   /// against the immediate without having to materialize the immediate into a
297   /// register.
298   bool isLegalICmpImmediate(int64_t Imm) const;
299 
300   /// \brief Return true if the addressing mode represented by AM is legal for
301   /// this target, for a load/store of the specified type.
302   /// The type may be VoidTy, in which case only return true if the addressing
303   /// mode is legal for a load/store of any legal type.
304   /// TODO: Handle pre/postinc as well.
305   bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
306                              bool HasBaseReg, int64_t Scale) const;
307 
308   /// \brief Return true if the target works with masked instruction
309   /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
310   /// AVX-512 architecture will also allow masks for non-consecutive memory
311   /// accesses.
312   bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
313   bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
314 
315   /// \brief Return the cost of the scaling factor used in the addressing
316   /// mode represented by AM for this target, for a load/store
317   /// of the specified type.
318   /// If the AM is supported, the return value must be >= 0.
319   /// If the AM is not supported, it returns a negative value.
320   /// TODO: Handle pre/postinc as well.
321   int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
322                            bool HasBaseReg, int64_t Scale) const;
323 
324   /// \brief Return true if it's free to truncate a value of type Ty1 to type
325   /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
326   /// by referencing its sub-register AX.
327   bool isTruncateFree(Type *Ty1, Type *Ty2) const;
328 
329   /// \brief Return true if it is profitable to hoist instruction in the
330   /// then/else to before if.
331   bool isProfitableToHoist(Instruction *I) const;
332 
333   /// \brief Return true if this type is legal.
334   bool isTypeLegal(Type *Ty) const;
335 
336   /// \brief Returns the target's jmp_buf alignment in bytes.
337   unsigned getJumpBufAlignment() const;
338 
339   /// \brief Returns the target's jmp_buf size in bytes.
340   unsigned getJumpBufSize() const;
341 
342   /// \brief Return true if switches should be turned into lookup tables for the
343   /// target.
344   bool shouldBuildLookupTables() const;
345 
346   /// \brief Don't restrict interleaved unrolling to small loops.
347   bool enableAggressiveInterleaving(bool LoopHasReductions) const;
348 
349   /// \brief Return hardware support for population count.
350   PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
351 
352   /// \brief Return true if the hardware has a fast square-root instruction.
353   bool haveFastSqrt(Type *Ty) const;
354 
355   /// \brief Return the expected cost of supporting the floating point operation
356   /// of the specified type.
357   unsigned getFPOpCost(Type *Ty) const;
358 
359   /// \brief Return the expected cost of materializing for the given integer
360   /// immediate of the specified type.
361   unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
362 
363   /// \brief Return the expected cost of materialization for the given integer
364   /// immediate of the specified type for a given instruction. The cost can be
365   /// zero if the immediate can be folded into the specified instruction.
366   unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
367                          Type *Ty) const;
368   unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
369                          Type *Ty) const;
370   /// @}
371 
372   /// \name Vector Target Information
373   /// @{
374 
375   /// \brief The various kinds of shuffle patterns for vector queries.
376   enum ShuffleKind {
377     SK_Broadcast,       ///< Broadcast element 0 to all other elements.
378     SK_Reverse,         ///< Reverse the order of the vector.
379     SK_Alternate,       ///< Choose alternate elements from vector.
380     SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
381     SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
382   };
383 
384   /// \brief Additional information about an operand's possible values.
385   enum OperandValueKind {
386     OK_AnyValue,               // Operand can have any value.
387     OK_UniformValue,           // Operand is uniform (splat of a value).
388     OK_UniformConstantValue,   // Operand is uniform constant.
389     OK_NonUniformConstantValue // Operand is a non uniform constant value.
390   };
391 
392   /// \brief Additional properties of an operand's values.
393   enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
394 
395   /// \return The number of scalar or vector registers that the target has.
396   /// If 'Vectors' is true, it returns the number of vector registers. If it is
397   /// set to false, it returns the number of scalar registers.
398   unsigned getNumberOfRegisters(bool Vector) const;
399 
400   /// \return The width of the largest scalar or vector register type.
401   unsigned getRegisterBitWidth(bool Vector) const;
402 
403   /// \return The maximum interleave factor that any transform should try to
404   /// perform for this target. This number depends on the level of parallelism
405   /// and the number of execution units in the CPU.
406   unsigned getMaxInterleaveFactor() const;
407 
408   /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
409   unsigned
410   getArithmeticInstrCost(unsigned Opcode, Type *Ty,
411                          OperandValueKind Opd1Info = OK_AnyValue,
412                          OperandValueKind Opd2Info = OK_AnyValue,
413                          OperandValueProperties Opd1PropInfo = OP_None,
414                          OperandValueProperties Opd2PropInfo = OP_None) const;
415 
416   /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
417   /// The index and subtype parameters are used by the subvector insertion and
418   /// extraction shuffle kinds.
419   unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
420                           Type *SubTp = nullptr) const;
421 
422   /// \return The expected cost of cast instructions, such as bitcast, trunc,
423   /// zext, etc.
424   unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) const;
425 
426   /// \return The expected cost of control-flow related instructions such as
427   /// Phi, Ret, Br.
428   unsigned getCFInstrCost(unsigned Opcode) const;
429 
430   /// \returns The expected cost of compare and select instructions.
431   unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
432                               Type *CondTy = nullptr) const;
433 
434   /// \return The expected cost of vector Insert and Extract.
435   /// Use -1 to indicate that there is no information on the index value.
436   unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
437                               unsigned Index = -1) const;
438 
439   /// \return The cost of Load and Store instructions.
440   unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
441                            unsigned AddressSpace) const;
442 
443   /// \return The cost of masked Load and Store instructions.
444   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
445                                  unsigned AddressSpace) const;
446 
447   /// \brief Calculate the cost of performing a vector reduction.
448   ///
449   /// This is the cost of reducing the vector value of type \p Ty to a scalar
450   /// value using the operation denoted by \p Opcode. The form of the reduction
451   /// can either be a pairwise reduction or a reduction that splits the vector
452   /// at every reduction level.
453   ///
454   /// Pairwise:
455   ///  (v0, v1, v2, v3)
456   ///  ((v0+v1), (v2, v3), undef, undef)
457   /// Split:
458   ///  (v0, v1, v2, v3)
459   ///  ((v0+v2), (v1+v3), undef, undef)
460   unsigned getReductionCost(unsigned Opcode, Type *Ty,
461                             bool IsPairwiseForm) const;
462 
463   /// \returns The cost of Intrinsic instructions.
464   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
465                                  ArrayRef<Type *> Tys) const;
466 
467   /// \returns The cost of Call instructions.
468   unsigned getCallInstrCost(Function *F, Type *RetTy,
469                             ArrayRef<Type *> Tys) const;
470 
471   /// \returns The number of pieces into which the provided type must be
472   /// split during legalization. Zero is returned when the answer is unknown.
473   unsigned getNumberOfParts(Type *Tp) const;
474 
475   /// \returns The cost of the address computation. For most targets this can be
476   /// merged into the instruction indexing mode. Some targets might want to
477   /// distinguish between address computation for memory operations on vector
478   /// types and scalar types. Such targets should override this function.
479   /// The 'IsComplex' parameter is a hint that the address computation is likely
480   /// to involve multiple instructions and as such unlikely to be merged into
481   /// the address indexing mode.
482   unsigned getAddressComputationCost(Type *Ty, bool IsComplex = false) const;
483 
484   /// \returns The cost, if any, of keeping values of the given types alive
485   /// over a callsite.
486   ///
487   /// Some types may require the use of register classes that do not have
488   /// any callee-saved registers, so would require a spill and fill.
489   unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
490 
491   /// \returns True if the intrinsic is a supported memory intrinsic.  Info
492   /// will contain additional information - whether the intrinsic may write
493   /// or read to memory, volatility and the pointer.  Info is undefined
494   /// if false is returned.
495   bool getTgtMemIntrinsic(IntrinsicInst *Inst, MemIntrinsicInfo &Info) const;
496 
497   /// \returns A value which is the result of the given memory intrinsic.  New
498   /// instructions may be created to extract the result from the given intrinsic
499   /// memory operation.  Returns nullptr if the target cannot create a result
500   /// from the given intrinsic.
501   Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
502                                            Type *ExpectedType) const;
503 
504   /// @}
505 
506 private:
507   /// \brief The abstract base class used to type erase specific TTI
508   /// implementations.
509   class Concept;
510 
511   /// \brief The template model for the base class which wraps a concrete
512   /// implementation in a type erased interface.
513   template <typename T> class Model;
514 
515   std::unique_ptr<Concept> TTIImpl;
516 };
517 
518 class TargetTransformInfo::Concept {
519 public:
520   virtual ~Concept() = 0;
521 
522   virtual unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) = 0;
523   virtual unsigned getGEPCost(const Value *Ptr,
524                               ArrayRef<const Value *> Operands) = 0;
525   virtual unsigned getCallCost(FunctionType *FTy, int NumArgs) = 0;
526   virtual unsigned getCallCost(const Function *F, int NumArgs) = 0;
527   virtual unsigned getCallCost(const Function *F,
528                                ArrayRef<const Value *> Arguments) = 0;
529   virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
530                                     ArrayRef<Type *> ParamTys) = 0;
531   virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
532                                     ArrayRef<const Value *> Arguments) = 0;
533   virtual unsigned getUserCost(const User *U) = 0;
534   virtual bool hasBranchDivergence() = 0;
535   virtual bool isSourceOfDivergence(const Value *V) = 0;
536   virtual bool isLoweredToCall(const Function *F) = 0;
537   virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) = 0;
538   virtual bool isLegalAddImmediate(int64_t Imm) = 0;
539   virtual bool isLegalICmpImmediate(int64_t Imm) = 0;
540   virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
541                                      int64_t BaseOffset, bool HasBaseReg,
542                                      int64_t Scale) = 0;
543   virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) = 0;
544   virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) = 0;
545   virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
546                                    int64_t BaseOffset, bool HasBaseReg,
547                                    int64_t Scale) = 0;
548   virtual bool isTruncateFree(Type *Ty1, Type *Ty2) = 0;
549   virtual bool isProfitableToHoist(Instruction *I) = 0;
550   virtual bool isTypeLegal(Type *Ty) = 0;
551   virtual unsigned getJumpBufAlignment() = 0;
552   virtual unsigned getJumpBufSize() = 0;
553   virtual bool shouldBuildLookupTables() = 0;
554   virtual bool enableAggressiveInterleaving(bool LoopHasReductions) = 0;
555   virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) = 0;
556   virtual bool haveFastSqrt(Type *Ty) = 0;
557   virtual unsigned getFPOpCost(Type *Ty) = 0;
558   virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) = 0;
559   virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
560                                  Type *Ty) = 0;
561   virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
562                                  const APInt &Imm, Type *Ty) = 0;
563   virtual unsigned getNumberOfRegisters(bool Vector) = 0;
564   virtual unsigned getRegisterBitWidth(bool Vector) = 0;
565   virtual unsigned getMaxInterleaveFactor() = 0;
566   virtual unsigned
567   getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
568                          OperandValueKind Opd2Info,
569                          OperandValueProperties Opd1PropInfo,
570                          OperandValueProperties Opd2PropInfo) = 0;
571   virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
572                                   Type *SubTp) = 0;
573   virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) = 0;
574   virtual unsigned getCFInstrCost(unsigned Opcode) = 0;
575   virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
576                                       Type *CondTy) = 0;
577   virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
578                                       unsigned Index) = 0;
579   virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
580                                    unsigned Alignment,
581                                    unsigned AddressSpace) = 0;
582   virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
583                                          unsigned Alignment,
584                                          unsigned AddressSpace) = 0;
585   virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
586                                     bool IsPairwiseForm) = 0;
587   virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
588                                          ArrayRef<Type *> Tys) = 0;
589   virtual unsigned getCallInstrCost(Function *F, Type *RetTy,
590                                     ArrayRef<Type *> Tys) = 0;
591   virtual unsigned getNumberOfParts(Type *Tp) = 0;
592   virtual unsigned getAddressComputationCost(Type *Ty, bool IsComplex) = 0;
593   virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) = 0;
594   virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
595                                   MemIntrinsicInfo &Info) = 0;
596   virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
597                                                    Type *ExpectedType) = 0;
598 };
599 
600 template <typename T>
601 class TargetTransformInfo::Model final : public TargetTransformInfo::Concept {
602   T Impl;
603 
604 public:
Model(T Impl)605   Model(T Impl) : Impl(std::move(Impl)) {}
~Model()606   ~Model() override {}
607 
getOperationCost(unsigned Opcode,Type * Ty,Type * OpTy)608   unsigned getOperationCost(unsigned Opcode, Type *Ty, Type *OpTy) override {
609     return Impl.getOperationCost(Opcode, Ty, OpTy);
610   }
getGEPCost(const Value * Ptr,ArrayRef<const Value * > Operands)611   unsigned getGEPCost(const Value *Ptr,
612                       ArrayRef<const Value *> Operands) override {
613     return Impl.getGEPCost(Ptr, Operands);
614   }
getCallCost(FunctionType * FTy,int NumArgs)615   unsigned getCallCost(FunctionType *FTy, int NumArgs) override {
616     return Impl.getCallCost(FTy, NumArgs);
617   }
getCallCost(const Function * F,int NumArgs)618   unsigned getCallCost(const Function *F, int NumArgs) override {
619     return Impl.getCallCost(F, NumArgs);
620   }
getCallCost(const Function * F,ArrayRef<const Value * > Arguments)621   unsigned getCallCost(const Function *F,
622                        ArrayRef<const Value *> Arguments) override {
623     return Impl.getCallCost(F, Arguments);
624   }
getIntrinsicCost(Intrinsic::ID IID,Type * RetTy,ArrayRef<Type * > ParamTys)625   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
626                             ArrayRef<Type *> ParamTys) override {
627     return Impl.getIntrinsicCost(IID, RetTy, ParamTys);
628   }
getIntrinsicCost(Intrinsic::ID IID,Type * RetTy,ArrayRef<const Value * > Arguments)629   unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
630                             ArrayRef<const Value *> Arguments) override {
631     return Impl.getIntrinsicCost(IID, RetTy, Arguments);
632   }
getUserCost(const User * U)633   unsigned getUserCost(const User *U) override { return Impl.getUserCost(U); }
hasBranchDivergence()634   bool hasBranchDivergence() override { return Impl.hasBranchDivergence(); }
isSourceOfDivergence(const Value * V)635   bool isSourceOfDivergence(const Value *V) override {
636     return Impl.isSourceOfDivergence(V);
637   }
isLoweredToCall(const Function * F)638   bool isLoweredToCall(const Function *F) override {
639     return Impl.isLoweredToCall(F);
640   }
getUnrollingPreferences(Loop * L,UnrollingPreferences & UP)641   void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) override {
642     return Impl.getUnrollingPreferences(L, UP);
643   }
isLegalAddImmediate(int64_t Imm)644   bool isLegalAddImmediate(int64_t Imm) override {
645     return Impl.isLegalAddImmediate(Imm);
646   }
isLegalICmpImmediate(int64_t Imm)647   bool isLegalICmpImmediate(int64_t Imm) override {
648     return Impl.isLegalICmpImmediate(Imm);
649   }
isLegalAddressingMode(Type * Ty,GlobalValue * BaseGV,int64_t BaseOffset,bool HasBaseReg,int64_t Scale)650   bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
651                              bool HasBaseReg, int64_t Scale) override {
652     return Impl.isLegalAddressingMode(Ty, BaseGV, BaseOffset, HasBaseReg,
653                                       Scale);
654   }
isLegalMaskedStore(Type * DataType,int Consecutive)655   bool isLegalMaskedStore(Type *DataType, int Consecutive) override {
656     return Impl.isLegalMaskedStore(DataType, Consecutive);
657   }
isLegalMaskedLoad(Type * DataType,int Consecutive)658   bool isLegalMaskedLoad(Type *DataType, int Consecutive) override {
659     return Impl.isLegalMaskedLoad(DataType, Consecutive);
660   }
getScalingFactorCost(Type * Ty,GlobalValue * BaseGV,int64_t BaseOffset,bool HasBaseReg,int64_t Scale)661   int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV, int64_t BaseOffset,
662                            bool HasBaseReg, int64_t Scale) override {
663     return Impl.getScalingFactorCost(Ty, BaseGV, BaseOffset, HasBaseReg, Scale);
664   }
isTruncateFree(Type * Ty1,Type * Ty2)665   bool isTruncateFree(Type *Ty1, Type *Ty2) override {
666     return Impl.isTruncateFree(Ty1, Ty2);
667   }
isProfitableToHoist(Instruction * I)668   bool isProfitableToHoist(Instruction *I) override {
669     return Impl.isProfitableToHoist(I);
670   }
isTypeLegal(Type * Ty)671   bool isTypeLegal(Type *Ty) override { return Impl.isTypeLegal(Ty); }
getJumpBufAlignment()672   unsigned getJumpBufAlignment() override { return Impl.getJumpBufAlignment(); }
getJumpBufSize()673   unsigned getJumpBufSize() override { return Impl.getJumpBufSize(); }
shouldBuildLookupTables()674   bool shouldBuildLookupTables() override {
675     return Impl.shouldBuildLookupTables();
676   }
enableAggressiveInterleaving(bool LoopHasReductions)677   bool enableAggressiveInterleaving(bool LoopHasReductions) override {
678     return Impl.enableAggressiveInterleaving(LoopHasReductions);
679   }
getPopcntSupport(unsigned IntTyWidthInBit)680   PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) override {
681     return Impl.getPopcntSupport(IntTyWidthInBit);
682   }
haveFastSqrt(Type * Ty)683   bool haveFastSqrt(Type *Ty) override { return Impl.haveFastSqrt(Ty); }
684 
getFPOpCost(Type * Ty)685   unsigned getFPOpCost(Type *Ty) override {
686     return Impl.getFPOpCost(Ty);
687   }
688 
getIntImmCost(const APInt & Imm,Type * Ty)689   unsigned getIntImmCost(const APInt &Imm, Type *Ty) override {
690     return Impl.getIntImmCost(Imm, Ty);
691   }
getIntImmCost(unsigned Opc,unsigned Idx,const APInt & Imm,Type * Ty)692   unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
693                          Type *Ty) override {
694     return Impl.getIntImmCost(Opc, Idx, Imm, Ty);
695   }
getIntImmCost(Intrinsic::ID IID,unsigned Idx,const APInt & Imm,Type * Ty)696   unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx, const APInt &Imm,
697                          Type *Ty) override {
698     return Impl.getIntImmCost(IID, Idx, Imm, Ty);
699   }
getNumberOfRegisters(bool Vector)700   unsigned getNumberOfRegisters(bool Vector) override {
701     return Impl.getNumberOfRegisters(Vector);
702   }
getRegisterBitWidth(bool Vector)703   unsigned getRegisterBitWidth(bool Vector) override {
704     return Impl.getRegisterBitWidth(Vector);
705   }
getMaxInterleaveFactor()706   unsigned getMaxInterleaveFactor() override {
707     return Impl.getMaxInterleaveFactor();
708   }
709   unsigned
getArithmeticInstrCost(unsigned Opcode,Type * Ty,OperandValueKind Opd1Info,OperandValueKind Opd2Info,OperandValueProperties Opd1PropInfo,OperandValueProperties Opd2PropInfo)710   getArithmeticInstrCost(unsigned Opcode, Type *Ty, OperandValueKind Opd1Info,
711                          OperandValueKind Opd2Info,
712                          OperandValueProperties Opd1PropInfo,
713                          OperandValueProperties Opd2PropInfo) override {
714     return Impl.getArithmeticInstrCost(Opcode, Ty, Opd1Info, Opd2Info,
715                                        Opd1PropInfo, Opd2PropInfo);
716   }
getShuffleCost(ShuffleKind Kind,Type * Tp,int Index,Type * SubTp)717   unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
718                           Type *SubTp) override {
719     return Impl.getShuffleCost(Kind, Tp, Index, SubTp);
720   }
getCastInstrCost(unsigned Opcode,Type * Dst,Type * Src)721   unsigned getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src) override {
722     return Impl.getCastInstrCost(Opcode, Dst, Src);
723   }
getCFInstrCost(unsigned Opcode)724   unsigned getCFInstrCost(unsigned Opcode) override {
725     return Impl.getCFInstrCost(Opcode);
726   }
getCmpSelInstrCost(unsigned Opcode,Type * ValTy,Type * CondTy)727   unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
728                               Type *CondTy) override {
729     return Impl.getCmpSelInstrCost(Opcode, ValTy, CondTy);
730   }
getVectorInstrCost(unsigned Opcode,Type * Val,unsigned Index)731   unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
732                               unsigned Index) override {
733     return Impl.getVectorInstrCost(Opcode, Val, Index);
734   }
getMemoryOpCost(unsigned Opcode,Type * Src,unsigned Alignment,unsigned AddressSpace)735   unsigned getMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
736                            unsigned AddressSpace) override {
737     return Impl.getMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
738   }
getMaskedMemoryOpCost(unsigned Opcode,Type * Src,unsigned Alignment,unsigned AddressSpace)739   unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src, unsigned Alignment,
740                                  unsigned AddressSpace) override {
741     return Impl.getMaskedMemoryOpCost(Opcode, Src, Alignment, AddressSpace);
742   }
getReductionCost(unsigned Opcode,Type * Ty,bool IsPairwiseForm)743   unsigned getReductionCost(unsigned Opcode, Type *Ty,
744                             bool IsPairwiseForm) override {
745     return Impl.getReductionCost(Opcode, Ty, IsPairwiseForm);
746   }
getIntrinsicInstrCost(Intrinsic::ID ID,Type * RetTy,ArrayRef<Type * > Tys)747   unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
748                                  ArrayRef<Type *> Tys) override {
749     return Impl.getIntrinsicInstrCost(ID, RetTy, Tys);
750   }
getCallInstrCost(Function * F,Type * RetTy,ArrayRef<Type * > Tys)751   unsigned getCallInstrCost(Function *F, Type *RetTy,
752                             ArrayRef<Type *> Tys) override {
753     return Impl.getCallInstrCost(F, RetTy, Tys);
754   }
getNumberOfParts(Type * Tp)755   unsigned getNumberOfParts(Type *Tp) override {
756     return Impl.getNumberOfParts(Tp);
757   }
getAddressComputationCost(Type * Ty,bool IsComplex)758   unsigned getAddressComputationCost(Type *Ty, bool IsComplex) override {
759     return Impl.getAddressComputationCost(Ty, IsComplex);
760   }
getCostOfKeepingLiveOverCall(ArrayRef<Type * > Tys)761   unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) override {
762     return Impl.getCostOfKeepingLiveOverCall(Tys);
763   }
getTgtMemIntrinsic(IntrinsicInst * Inst,MemIntrinsicInfo & Info)764   bool getTgtMemIntrinsic(IntrinsicInst *Inst,
765                           MemIntrinsicInfo &Info) override {
766     return Impl.getTgtMemIntrinsic(Inst, Info);
767   }
getOrCreateResultFromMemIntrinsic(IntrinsicInst * Inst,Type * ExpectedType)768   Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
769                                            Type *ExpectedType) override {
770     return Impl.getOrCreateResultFromMemIntrinsic(Inst, ExpectedType);
771   }
772 };
773 
774 template <typename T>
TargetTransformInfo(T Impl)775 TargetTransformInfo::TargetTransformInfo(T Impl)
776     : TTIImpl(new Model<T>(Impl)) {}
777 
778 /// \brief Analysis pass providing the \c TargetTransformInfo.
779 ///
780 /// The core idea of the TargetIRAnalysis is to expose an interface through
781 /// which LLVM targets can analyze and provide information about the middle
782 /// end's target-independent IR. This supports use cases such as target-aware
783 /// cost modeling of IR constructs.
784 ///
785 /// This is a function analysis because much of the cost modeling for targets
786 /// is done in a subtarget specific way and LLVM supports compiling different
787 /// functions targeting different subtargets in order to support runtime
788 /// dispatch according to the observed subtarget.
789 class TargetIRAnalysis {
790 public:
791   typedef TargetTransformInfo Result;
792 
793   /// \brief Opaque, unique identifier for this analysis pass.
ID()794   static void *ID() { return (void *)&PassID; }
795 
796   /// \brief Provide access to a name for this pass for debugging purposes.
name()797   static StringRef name() { return "TargetIRAnalysis"; }
798 
799   /// \brief Default construct a target IR analysis.
800   ///
801   /// This will use the module's datalayout to construct a baseline
802   /// conservative TTI result.
803   TargetIRAnalysis();
804 
805   /// \brief Construct an IR analysis pass around a target-provide callback.
806   ///
807   /// The callback will be called with a particular function for which the TTI
808   /// is needed and must return a TTI object for that function.
809   TargetIRAnalysis(std::function<Result(Function &)> TTICallback);
810 
811   // Value semantics. We spell out the constructors for MSVC.
TargetIRAnalysis(const TargetIRAnalysis & Arg)812   TargetIRAnalysis(const TargetIRAnalysis &Arg)
813       : TTICallback(Arg.TTICallback) {}
TargetIRAnalysis(TargetIRAnalysis && Arg)814   TargetIRAnalysis(TargetIRAnalysis &&Arg)
815       : TTICallback(std::move(Arg.TTICallback)) {}
816   TargetIRAnalysis &operator=(const TargetIRAnalysis &RHS) {
817     TTICallback = RHS.TTICallback;
818     return *this;
819   }
820   TargetIRAnalysis &operator=(TargetIRAnalysis &&RHS) {
821     TTICallback = std::move(RHS.TTICallback);
822     return *this;
823   }
824 
825   Result run(Function &F);
826 
827 private:
828   static char PassID;
829 
830   /// \brief The callback used to produce a result.
831   ///
832   /// We use a completely opaque callback so that targets can provide whatever
833   /// mechanism they desire for constructing the TTI for a given function.
834   ///
835   /// FIXME: Should we really use std::function? It's relatively inefficient.
836   /// It might be possible to arrange for even stateful callbacks to outlive
837   /// the analysis and thus use a function_ref which would be lighter weight.
838   /// This may also be less error prone as the callback is likely to reference
839   /// the external TargetMachine, and that reference needs to never dangle.
840   std::function<Result(Function &)> TTICallback;
841 
842   /// \brief Helper function used as the callback in the default constructor.
843   static Result getDefaultTTI(Function &F);
844 };
845 
846 /// \brief Wrapper pass for TargetTransformInfo.
847 ///
848 /// This pass can be constructed from a TTI object which it stores internally
849 /// and is queried by passes.
850 class TargetTransformInfoWrapperPass : public ImmutablePass {
851   TargetIRAnalysis TIRA;
852   Optional<TargetTransformInfo> TTI;
853 
854   virtual void anchor();
855 
856 public:
857   static char ID;
858 
859   /// \brief We must provide a default constructor for the pass but it should
860   /// never be used.
861   ///
862   /// Use the constructor below or call one of the creation routines.
863   TargetTransformInfoWrapperPass();
864 
865   explicit TargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
866 
867   TargetTransformInfo &getTTI(Function &F);
868 };
869 
870 /// \brief Create an analysis pass wrapper around a TTI object.
871 ///
872 /// This analysis pass just holds the TTI instance and makes it available to
873 /// clients.
874 ImmutablePass *createTargetTransformInfoWrapperPass(TargetIRAnalysis TIRA);
875 
876 } // End llvm namespace
877 
878 #endif
879