1//===- llvm/Analysis/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//
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/IR/Intrinsics.h"
26#include "llvm/Pass.h"
27#include "llvm/Support/DataTypes.h"
28
29namespace llvm {
30
31class GlobalValue;
32class Loop;
33class Type;
34class User;
35class Value;
36
37/// TargetTransformInfo - This pass provides access to the codegen
38/// interfaces that are needed for IR-level transformations.
39class TargetTransformInfo {
40protected:
41  /// \brief The TTI instance one level down the stack.
42  ///
43  /// This is used to implement the default behavior all of the methods which
44  /// is to delegate up through the stack of TTIs until one can answer the
45  /// query.
46  TargetTransformInfo *PrevTTI;
47
48  /// \brief The top of the stack of TTI analyses available.
49  ///
50  /// This is a convenience routine maintained as TTI analyses become available
51  /// that complements the PrevTTI delegation chain. When one part of an
52  /// analysis pass wants to query another part of the analysis pass it can use
53  /// this to start back at the top of the stack.
54  TargetTransformInfo *TopTTI;
55
56  /// All pass subclasses must in their initializePass routine call
57  /// pushTTIStack with themselves to update the pointers tracking the previous
58  /// TTI instance in the analysis group's stack, and the top of the analysis
59  /// group's stack.
60  void pushTTIStack(Pass *P);
61
62  /// All pass subclasses must call TargetTransformInfo::getAnalysisUsage.
63  virtual void getAnalysisUsage(AnalysisUsage &AU) const;
64
65public:
66  /// This class is intended to be subclassed by real implementations.
67  virtual ~TargetTransformInfo() = 0;
68
69  /// \name Generic Target Information
70  /// @{
71
72  /// \brief Underlying constants for 'cost' values in this interface.
73  ///
74  /// Many APIs in this interface return a cost. This enum defines the
75  /// fundamental values that should be used to interpret (and produce) those
76  /// costs. The costs are returned as an unsigned rather than a member of this
77  /// enumeration because it is expected that the cost of one IR instruction
78  /// may have a multiplicative factor to it or otherwise won't fit directly
79  /// into the enum. Moreover, it is common to sum or average costs which works
80  /// better as simple integral values. Thus this enum only provides constants.
81  ///
82  /// Note that these costs should usually reflect the intersection of code-size
83  /// cost and execution cost. A free instruction is typically one that folds
84  /// into another instruction. For example, reg-to-reg moves can often be
85  /// skipped by renaming the registers in the CPU, but they still are encoded
86  /// and thus wouldn't be considered 'free' here.
87  enum TargetCostConstants {
88    TCC_Free = 0,       ///< Expected to fold away in lowering.
89    TCC_Basic = 1,      ///< The cost of a typical 'add' instruction.
90    TCC_Expensive = 4   ///< The cost of a 'div' instruction on x86.
91  };
92
93  /// \brief Estimate the cost of a specific operation when lowered.
94  ///
95  /// Note that this is designed to work on an arbitrary synthetic opcode, and
96  /// thus work for hypothetical queries before an instruction has even been
97  /// formed. However, this does *not* work for GEPs, and must not be called
98  /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
99  /// analyzing a GEP's cost required more information.
100  ///
101  /// Typically only the result type is required, and the operand type can be
102  /// omitted. However, if the opcode is one of the cast instructions, the
103  /// operand type is required.
104  ///
105  /// The returned cost is defined in terms of \c TargetCostConstants, see its
106  /// comments for a detailed explanation of the cost values.
107  virtual unsigned getOperationCost(unsigned Opcode, Type *Ty,
108                                    Type *OpTy = nullptr) const;
109
110  /// \brief Estimate the cost of a GEP operation when lowered.
111  ///
112  /// The contract for this function is the same as \c getOperationCost except
113  /// that it supports an interface that provides extra information specific to
114  /// the GEP operation.
115  virtual unsigned getGEPCost(const Value *Ptr,
116                              ArrayRef<const Value *> Operands) const;
117
118  /// \brief Estimate the cost of a function call when lowered.
119  ///
120  /// The contract for this is the same as \c getOperationCost except that it
121  /// supports an interface that provides extra information specific to call
122  /// instructions.
123  ///
124  /// This is the most basic query for estimating call cost: it only knows the
125  /// function type and (potentially) the number of arguments at the call site.
126  /// The latter is only interesting for varargs function types.
127  virtual unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
128
129  /// \brief Estimate the cost of calling a specific function when lowered.
130  ///
131  /// This overload adds the ability to reason about the particular function
132  /// being called in the event it is a library call with special lowering.
133  virtual unsigned getCallCost(const Function *F, int NumArgs = -1) const;
134
135  /// \brief Estimate the cost of calling a specific function when lowered.
136  ///
137  /// This overload allows specifying a set of candidate argument values.
138  virtual unsigned getCallCost(const Function *F,
139                               ArrayRef<const Value *> Arguments) const;
140
141  /// \brief Estimate the cost of an intrinsic when lowered.
142  ///
143  /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
144  virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
145                                    ArrayRef<Type *> ParamTys) const;
146
147  /// \brief Estimate the cost of an intrinsic when lowered.
148  ///
149  /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
150  virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
151                                    ArrayRef<const Value *> Arguments) const;
152
153  /// \brief Estimate the cost of a given IR user when lowered.
154  ///
155  /// This can estimate the cost of either a ConstantExpr or Instruction when
156  /// lowered. It has two primary advantages over the \c getOperationCost and
157  /// \c getGEPCost above, and one significant disadvantage: it can only be
158  /// used when the IR construct has already been formed.
159  ///
160  /// The advantages are that it can inspect the SSA use graph to reason more
161  /// accurately about the cost. For example, all-constant-GEPs can often be
162  /// folded into a load or other instruction, but if they are used in some
163  /// other context they may not be folded. This routine can distinguish such
164  /// cases.
165  ///
166  /// The returned cost is defined in terms of \c TargetCostConstants, see its
167  /// comments for a detailed explanation of the cost values.
168  virtual unsigned getUserCost(const User *U) const;
169
170  /// \brief hasBranchDivergence - Return true if branch divergence exists.
171  /// Branch divergence has a significantly negative impact on GPU performance
172  /// when threads in the same wavefront take different paths due to conditional
173  /// branches.
174  virtual bool hasBranchDivergence() const;
175
176  /// \brief Test whether calls to a function lower to actual program function
177  /// calls.
178  ///
179  /// The idea is to test whether the program is likely to require a 'call'
180  /// instruction or equivalent in order to call the given function.
181  ///
182  /// FIXME: It's not clear that this is a good or useful query API. Client's
183  /// should probably move to simpler cost metrics using the above.
184  /// Alternatively, we could split the cost interface into distinct code-size
185  /// and execution-speed costs. This would allow modelling the core of this
186  /// query more accurately as the a call is a single small instruction, but
187  /// incurs significant execution cost.
188  virtual bool isLoweredToCall(const Function *F) const;
189
190  /// Parameters that control the generic loop unrolling transformation.
191  struct UnrollingPreferences {
192    /// The cost threshold for the unrolled loop, compared to
193    /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
194    /// The unrolling factor is set such that the unrolled loop body does not
195    /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
196    /// restriction.
197    unsigned Threshold;
198    /// The cost threshold for the unrolled loop when optimizing for size (set
199    /// to UINT_MAX to disable).
200    unsigned OptSizeThreshold;
201    /// The cost threshold for the unrolled loop, like Threshold, but used
202    /// for partial/runtime unrolling (set to UINT_MAX to disable).
203    unsigned PartialThreshold;
204    /// The cost threshold for the unrolled loop when optimizing for size, like
205    /// OptSizeThreshold, but used for partial/runtime unrolling (set to UINT_MAX
206    /// to disable).
207    unsigned PartialOptSizeThreshold;
208    /// A forced unrolling factor (the number of concatenated bodies of the
209    /// original loop in the unrolled loop body). When set to 0, the unrolling
210    /// transformation will select an unrolling factor based on the current cost
211    /// threshold and other factors.
212    unsigned Count;
213    // Set the maximum unrolling factor. The unrolling factor may be selected
214    // using the appropriate cost threshold, but may not exceed this number
215    // (set to UINT_MAX to disable). This does not apply in cases where the
216    // loop is being fully unrolled.
217    unsigned MaxCount;
218    /// Allow partial unrolling (unrolling of loops to expand the size of the
219    /// loop body, not only to eliminate small constant-trip-count loops).
220    bool     Partial;
221    /// Allow runtime unrolling (unrolling of loops to expand the size of the
222    /// loop body even when the number of loop iterations is not known at compile
223    /// time).
224    bool     Runtime;
225  };
226
227  /// \brief Get target-customized preferences for the generic loop unrolling
228  /// transformation. The caller will initialize UP with the current
229  /// target-independent defaults.
230  virtual void getUnrollingPreferences(Loop *L, UnrollingPreferences &UP) const;
231
232  /// @}
233
234  /// \name Scalar Target Information
235  /// @{
236
237  /// \brief Flags indicating the kind of support for population count.
238  ///
239  /// Compared to the SW implementation, HW support is supposed to
240  /// significantly boost the performance when the population is dense, and it
241  /// may or may not degrade performance if the population is sparse. A HW
242  /// support is considered as "Fast" if it can outperform, or is on a par
243  /// with, SW implementation when the population is sparse; otherwise, it is
244  /// considered as "Slow".
245  enum PopcntSupportKind {
246    PSK_Software,
247    PSK_SlowHardware,
248    PSK_FastHardware
249  };
250
251  /// \brief Return true if the specified immediate is legal add immediate, that
252  /// is the target has add instructions which can add a register with the
253  /// immediate without having to materialize the immediate into a register.
254  virtual bool isLegalAddImmediate(int64_t Imm) const;
255
256  /// \brief Return true if the specified immediate is legal icmp immediate,
257  /// that is the target has icmp instructions which can compare a register
258  /// against the immediate without having to materialize the immediate into a
259  /// register.
260  virtual bool isLegalICmpImmediate(int64_t Imm) const;
261
262  /// \brief Return true if the addressing mode represented by AM is legal for
263  /// this target, for a load/store of the specified type.
264  /// The type may be VoidTy, in which case only return true if the addressing
265  /// mode is legal for a load/store of any legal type.
266  /// TODO: Handle pre/postinc as well.
267  virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
268                                     int64_t BaseOffset, bool HasBaseReg,
269                                     int64_t Scale) const;
270
271  /// \brief Return the cost of the scaling factor used in the addressing
272  /// mode represented by AM for this target, for a load/store
273  /// of the specified type.
274  /// If the AM is supported, the return value must be >= 0.
275  /// If the AM is not supported, it returns a negative value.
276  /// TODO: Handle pre/postinc as well.
277  virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
278                                   int64_t BaseOffset, bool HasBaseReg,
279                                   int64_t Scale) const;
280
281  /// \brief Return true if it's free to truncate a value of type Ty1 to type
282  /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
283  /// by referencing its sub-register AX.
284  virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
285
286  /// \brief Return true if this type is legal.
287  virtual bool isTypeLegal(Type *Ty) const;
288
289  /// \brief Returns the target's jmp_buf alignment in bytes.
290  virtual unsigned getJumpBufAlignment() const;
291
292  /// \brief Returns the target's jmp_buf size in bytes.
293  virtual unsigned getJumpBufSize() const;
294
295  /// \brief Return true if switches should be turned into lookup tables for the
296  /// target.
297  virtual bool shouldBuildLookupTables() const;
298
299  /// \brief Return hardware support for population count.
300  virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
301
302  /// \brief Return true if the hardware has a fast square-root instruction.
303  virtual bool haveFastSqrt(Type *Ty) const;
304
305  /// \brief Return the expected cost of materializing for the given integer
306  /// immediate of the specified type.
307  virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
308
309  /// \brief Return the expected cost of materialization for the given integer
310  /// immediate of the specified type for a given instruction. The cost can be
311  /// zero if the immediate can be folded into the specified instruction.
312  virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
313                                 Type *Ty) const;
314  virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
315                                 const APInt &Imm, Type *Ty) const;
316  /// @}
317
318  /// \name Vector Target Information
319  /// @{
320
321  /// \brief The various kinds of shuffle patterns for vector queries.
322  enum ShuffleKind {
323    SK_Broadcast,       ///< Broadcast element 0 to all other elements.
324    SK_Reverse,         ///< Reverse the order of the vector.
325    SK_Alternate,       ///< Choose alternate elements from vector.
326    SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
327    SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
328  };
329
330  /// \brief Additional information about an operand's possible values.
331  enum OperandValueKind {
332    OK_AnyValue,                 // Operand can have any value.
333    OK_UniformValue,             // Operand is uniform (splat of a value).
334    OK_UniformConstantValue,     // Operand is uniform constant.
335    OK_NonUniformConstantValue   // Operand is a non uniform constant value.
336  };
337
338  /// \return The number of scalar or vector registers that the target has.
339  /// If 'Vectors' is true, it returns the number of vector registers. If it is
340  /// set to false, it returns the number of scalar registers.
341  virtual unsigned getNumberOfRegisters(bool Vector) const;
342
343  /// \return The width of the largest scalar or vector register type.
344  virtual unsigned getRegisterBitWidth(bool Vector) const;
345
346  /// \return The maximum unroll factor that the vectorizer should try to
347  /// perform for this target. This number depends on the level of parallelism
348  /// and the number of execution units in the CPU.
349  virtual unsigned getMaximumUnrollFactor() const;
350
351  /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
352  virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty,
353                                  OperandValueKind Opd1Info = OK_AnyValue,
354                                  OperandValueKind Opd2Info = OK_AnyValue) const;
355
356  /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
357  /// The index and subtype parameters are used by the subvector insertion and
358  /// extraction shuffle kinds.
359  virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
360                                  Type *SubTp = nullptr) const;
361
362  /// \return The expected cost of cast instructions, such as bitcast, trunc,
363  /// zext, etc.
364  virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
365                                    Type *Src) const;
366
367  /// \return The expected cost of control-flow related instructions such as
368  /// Phi, Ret, Br.
369  virtual unsigned getCFInstrCost(unsigned Opcode) const;
370
371  /// \returns The expected cost of compare and select instructions.
372  virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
373                                      Type *CondTy = nullptr) const;
374
375  /// \return The expected cost of vector Insert and Extract.
376  /// Use -1 to indicate that there is no information on the index value.
377  virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
378                                      unsigned Index = -1) const;
379
380  /// \return The cost of Load and Store instructions.
381  virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
382                                   unsigned Alignment,
383                                   unsigned AddressSpace) const;
384
385  /// \brief Calculate the cost of performing a vector reduction.
386  ///
387  /// This is the cost of reducing the vector value of type \p Ty to a scalar
388  /// value using the operation denoted by \p Opcode. The form of the reduction
389  /// can either be a pairwise reduction or a reduction that splits the vector
390  /// at every reduction level.
391  ///
392  /// Pairwise:
393  ///  (v0, v1, v2, v3)
394  ///  ((v0+v1), (v2, v3), undef, undef)
395  /// Split:
396  ///  (v0, v1, v2, v3)
397  ///  ((v0+v2), (v1+v3), undef, undef)
398  virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
399                                    bool IsPairwiseForm) const;
400
401  /// \returns The cost of Intrinsic instructions.
402  virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
403                                         ArrayRef<Type *> Tys) const;
404
405  /// \returns The number of pieces into which the provided type must be
406  /// split during legalization. Zero is returned when the answer is unknown.
407  virtual unsigned getNumberOfParts(Type *Tp) const;
408
409  /// \returns The cost of the address computation. For most targets this can be
410  /// merged into the instruction indexing mode. Some targets might want to
411  /// distinguish between address computation for memory operations on vector
412  /// types and scalar types. Such targets should override this function.
413  /// The 'IsComplex' parameter is a hint that the address computation is likely
414  /// to involve multiple instructions and as such unlikely to be merged into
415  /// the address indexing mode.
416  virtual unsigned getAddressComputationCost(Type *Ty,
417                                             bool IsComplex = false) const;
418
419  /// @}
420
421  /// Analysis group identification.
422  static char ID;
423};
424
425/// \brief Create the base case instance of a pass in the TTI analysis group.
426///
427/// This class provides the base case for the stack of TTI analyzes. It doesn't
428/// delegate to anything and uses the STTI and VTTI objects passed in to
429/// satisfy the queries.
430ImmutablePass *createNoTargetTransformInfoPass();
431
432} // End llvm namespace
433
434#endif
435