1 //===- llvm/Analysis/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/IR/Intrinsics.h"
26 #include "llvm/Pass.h"
27 #include "llvm/Support/DataTypes.h"
36 /// TargetTransformInfo - This pass provides access to the codegen
37 /// interfaces that are needed for IR-level transformations.
38 class TargetTransformInfo {
40 /// \brief The TTI instance one level down the stack.
42 /// This is used to implement the default behavior all of the methods which
43 /// is to delegate up through the stack of TTIs until one can answer the
45 TargetTransformInfo *PrevTTI;
47 /// \brief The top of the stack of TTI analyses available.
49 /// This is a convenience routine maintained as TTI analyses become available
50 /// that complements the PrevTTI delegation chain. When one part of an
51 /// analysis pass wants to query another part of the analysis pass it can use
52 /// this to start back at the top of the stack.
53 TargetTransformInfo *TopTTI;
55 /// All pass subclasses must in their initializePass routine call
56 /// pushTTIStack with themselves to update the pointers tracking the previous
57 /// TTI instance in the analysis group's stack, and the top of the analysis
59 void pushTTIStack(Pass *P);
61 /// All pass subclasses must in their finalizePass routine call popTTIStack
62 /// to update the pointers tracking the previous TTI instance in the analysis
63 /// group's stack, and the top of the analysis group's stack.
66 /// All pass subclasses must call TargetTransformInfo::getAnalysisUsage.
67 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
70 /// This class is intended to be subclassed by real implementations.
71 virtual ~TargetTransformInfo() = 0;
73 /// \name Generic Target Information
76 /// \brief Underlying constants for 'cost' values in this interface.
78 /// Many APIs in this interface return a cost. This enum defines the
79 /// fundamental values that should be used to interpret (and produce) those
80 /// costs. The costs are returned as an unsigned rather than a member of this
81 /// enumeration because it is expected that the cost of one IR instruction
82 /// may have a multiplicative factor to it or otherwise won't fit directly
83 /// into the enum. Moreover, it is common to sum or average costs which works
84 /// better as simple integral values. Thus this enum only provides constants.
86 /// Note that these costs should usually reflect the intersection of code-size
87 /// cost and execution cost. A free instruction is typically one that folds
88 /// into another instruction. For example, reg-to-reg moves can often be
89 /// skipped by renaming the registers in the CPU, but they still are encoded
90 /// and thus wouldn't be considered 'free' here.
91 enum TargetCostConstants {
92 TCC_Free = 0, ///< Expected to fold away in lowering.
93 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
94 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
97 /// \brief Estimate the cost of a specific operation when lowered.
99 /// Note that this is designed to work on an arbitrary synthetic opcode, and
100 /// thus work for hypothetical queries before an instruction has even been
101 /// formed. However, this does *not* work for GEPs, and must not be called
102 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
103 /// analyzing a GEP's cost required more information.
105 /// Typically only the result type is required, and the operand type can be
106 /// omitted. However, if the opcode is one of the cast instructions, the
107 /// operand type is required.
109 /// The returned cost is defined in terms of \c TargetCostConstants, see its
110 /// comments for a detailed explanation of the cost values.
111 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty,
112 Type *OpTy = 0) const;
114 /// \brief Estimate the cost of a GEP operation when lowered.
116 /// The contract for this function is the same as \c getOperationCost except
117 /// that it supports an interface that provides extra information specific to
118 /// the GEP operation.
119 virtual unsigned getGEPCost(const Value *Ptr,
120 ArrayRef<const Value *> Operands) const;
122 /// \brief Estimate the cost of a function call when lowered.
124 /// The contract for this is the same as \c getOperationCost except that it
125 /// supports an interface that provides extra information specific to call
128 /// This is the most basic query for estimating call cost: it only knows the
129 /// function type and (potentially) the number of arguments at the call site.
130 /// The latter is only interesting for varargs function types.
131 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
133 /// \brief Estimate the cost of calling a specific function when lowered.
135 /// This overload adds the ability to reason about the particular function
136 /// being called in the event it is a library call with special lowering.
137 virtual unsigned getCallCost(const Function *F, int NumArgs = -1) const;
139 /// \brief Estimate the cost of calling a specific function when lowered.
141 /// This overload allows specifying a set of candidate argument values.
142 virtual unsigned getCallCost(const Function *F,
143 ArrayRef<const Value *> Arguments) const;
145 /// \brief Estimate the cost of an intrinsic when lowered.
147 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
148 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
149 ArrayRef<Type *> ParamTys) const;
151 /// \brief Estimate the cost of an intrinsic when lowered.
153 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
154 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
155 ArrayRef<const Value *> Arguments) const;
157 /// \brief Estimate the cost of a given IR user when lowered.
159 /// This can estimate the cost of either a ConstantExpr or Instruction when
160 /// lowered. It has two primary advantages over the \c getOperationCost and
161 /// \c getGEPCost above, and one significant disadvantage: it can only be
162 /// used when the IR construct has already been formed.
164 /// The advantages are that it can inspect the SSA use graph to reason more
165 /// accurately about the cost. For example, all-constant-GEPs can often be
166 /// folded into a load or other instruction, but if they are used in some
167 /// other context they may not be folded. This routine can distinguish such
170 /// The returned cost is defined in terms of \c TargetCostConstants, see its
171 /// comments for a detailed explanation of the cost values.
172 virtual unsigned getUserCost(const User *U) const;
174 /// \brief Test whether calls to a function lower to actual program function
177 /// The idea is to test whether the program is likely to require a 'call'
178 /// instruction or equivalent in order to call the given function.
180 /// FIXME: It's not clear that this is a good or useful query API. Client's
181 /// should probably move to simpler cost metrics using the above.
182 /// Alternatively, we could split the cost interface into distinct code-size
183 /// and execution-speed costs. This would allow modelling the core of this
184 /// query more accurately as the a call is a single small instruction, but
185 /// incurs significant execution cost.
186 virtual bool isLoweredToCall(const Function *F) const;
190 /// \name Scalar Target Information
193 /// \brief Flags indicating the kind of support for population count.
195 /// Compared to the SW implementation, HW support is supposed to
196 /// significantly boost the performance when the population is dense, and it
197 /// may or may not degrade performance if the population is sparse. A HW
198 /// support is considered as "Fast" if it can outperform, or is on a par
199 /// with, SW implementation when the population is sparse; otherwise, it is
200 /// considered as "Slow".
201 enum PopcntSupportKind {
207 /// isLegalAddImmediate - Return true if the specified immediate is legal
208 /// add immediate, that is the target has add instructions which can add
209 /// a register with the immediate without having to materialize the
210 /// immediate into a register.
211 virtual bool isLegalAddImmediate(int64_t Imm) const;
213 /// isLegalICmpImmediate - Return true if the specified immediate is legal
214 /// icmp immediate, that is the target has icmp instructions which can compare
215 /// a register against the immediate without having to materialize the
216 /// immediate into a register.
217 virtual bool isLegalICmpImmediate(int64_t Imm) const;
219 /// isLegalAddressingMode - Return true if the addressing mode represented by
220 /// AM is legal for this target, for a load/store of the specified type.
221 /// The type may be VoidTy, in which case only return true if the addressing
222 /// mode is legal for a load/store of any legal type.
223 /// TODO: Handle pre/postinc as well.
224 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
225 int64_t BaseOffset, bool HasBaseReg,
226 int64_t Scale) const;
228 /// isTruncateFree - Return true if it's free to truncate a value of
229 /// type Ty1 to type Ty2. e.g. On x86 it's free to truncate a i32 value in
230 /// register EAX to i16 by referencing its sub-register AX.
231 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
233 /// Is this type legal.
234 virtual bool isTypeLegal(Type *Ty) const;
236 /// getJumpBufAlignment - returns the target's jmp_buf alignment in bytes
237 virtual unsigned getJumpBufAlignment() const;
239 /// getJumpBufSize - returns the target's jmp_buf size in bytes.
240 virtual unsigned getJumpBufSize() const;
242 /// shouldBuildLookupTables - Return true if switches should be turned into
243 /// lookup tables for the target.
244 virtual bool shouldBuildLookupTables() const;
246 /// getPopcntSupport - Return hardware support for population count.
247 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
249 /// getIntImmCost - Return the expected cost of materializing the given
250 /// integer immediate of the specified type.
251 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
255 /// \name Vector Target Information
258 /// \brief The various kinds of shuffle patterns for vector queries.
260 SK_Broadcast, ///< Broadcast element 0 to all other elements.
261 SK_Reverse, ///< Reverse the order of the vector.
262 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
263 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
266 /// \return The number of scalar or vector registers that the target has.
267 /// If 'Vectors' is true, it returns the number of vector registers. If it is
268 /// set to false, it returns the number of scalar registers.
269 virtual unsigned getNumberOfRegisters(bool Vector) const;
271 /// \return The width of the largest scalar or vector register type.
272 virtual unsigned getRegisterBitWidth(bool Vector) const;
274 /// \return The maximum unroll factor that the vectorizer should try to
275 /// perform for this target. This number depends on the level of parallelism
276 /// and the number of execution units in the CPU.
277 virtual unsigned getMaximumUnrollFactor() const;
279 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
280 virtual unsigned getArithmeticInstrCost(unsigned Opcode, Type *Ty) const;
282 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
283 /// The index and subtype parameters are used by the subvector insertion and
284 /// extraction shuffle kinds.
285 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
286 Type *SubTp = 0) const;
288 /// \return The expected cost of cast instructions, such as bitcast, trunc,
290 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
293 /// \return The expected cost of control-flow related instructions such as
295 virtual unsigned getCFInstrCost(unsigned Opcode) const;
297 /// \returns The expected cost of compare and select instructions.
298 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
299 Type *CondTy = 0) const;
301 /// \return The expected cost of vector Insert and Extract.
302 /// Use -1 to indicate that there is no information on the index value.
303 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
304 unsigned Index = -1) const;
306 /// \return The cost of Load and Store instructions.
307 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
309 unsigned AddressSpace) const;
311 /// \returns The cost of Intrinsic instructions.
312 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
313 ArrayRef<Type *> Tys) const;
315 /// \returns The number of pieces into which the provided type must be
316 /// split during legalization. Zero is returned when the answer is unknown.
317 virtual unsigned getNumberOfParts(Type *Tp) const;
319 /// \returns The cost of the address computation. For most targets this can be
320 /// merged into the instruction indexing mode. Some targets might want to
321 /// distinguish between address computation for memory operations on vector
322 /// types and scalar types. Such targets should override this function.
323 virtual unsigned getAddressComputationCost(Type *Ty) const;
327 /// Analysis group identification.
331 /// \brief Create the base case instance of a pass in the TTI analysis group.
333 /// This class provides the base case for the stack of TTI analyzes. It doesn't
334 /// delegate to anything and uses the STTI and VTTI objects passed in to
335 /// satisfy the queries.
336 ImmutablePass *createNoTargetTransformInfoPass();
338 } // End llvm namespace