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/IR/Intrinsics.h"
26 #include "llvm/IR/IntrinsicInst.h"
27 #include "llvm/Pass.h"
28 #include "llvm/Support/DataTypes.h"
39 /// \brief Information about a load/store intrinsic defined by the target.
40 struct MemIntrinsicInfo {
42 : ReadMem(false), WriteMem(false), Vol(false), MatchingId(0),
43 NumMemRefs(0), PtrVal(nullptr) {}
47 // Same Id is set by the target for corresponding load/store intrinsics.
48 unsigned short MatchingId;
53 /// \brief This pass provides access to the codegen interfaces that are needed
54 /// for IR-level transformations.
55 class TargetTransformInfo {
57 /// \brief The TTI instance one level down the stack.
59 /// This is used to implement the default behavior all of the methods which
60 /// is to delegate up through the stack of TTIs until one can answer the
62 TargetTransformInfo *PrevTTI;
64 /// \brief The top of the stack of TTI analyses available.
66 /// This is a convenience routine maintained as TTI analyses become available
67 /// that complements the PrevTTI delegation chain. When one part of an
68 /// analysis pass wants to query another part of the analysis pass it can use
69 /// this to start back at the top of the stack.
70 TargetTransformInfo *TopTTI;
72 /// All pass subclasses must in their initializePass routine call
73 /// pushTTIStack with themselves to update the pointers tracking the previous
74 /// TTI instance in the analysis group's stack, and the top of the analysis
76 void pushTTIStack(Pass *P);
78 /// All pass subclasses must call TargetTransformInfo::getAnalysisUsage.
79 virtual void getAnalysisUsage(AnalysisUsage &AU) const;
82 /// This class is intended to be subclassed by real implementations.
83 virtual ~TargetTransformInfo() = 0;
85 /// \name Generic Target Information
88 /// \brief Underlying constants for 'cost' values in this interface.
90 /// Many APIs in this interface return a cost. This enum defines the
91 /// fundamental values that should be used to interpret (and produce) those
92 /// costs. The costs are returned as an unsigned rather than a member of this
93 /// enumeration because it is expected that the cost of one IR instruction
94 /// may have a multiplicative factor to it or otherwise won't fit directly
95 /// into the enum. Moreover, it is common to sum or average costs which works
96 /// better as simple integral values. Thus this enum only provides constants.
98 /// Note that these costs should usually reflect the intersection of code-size
99 /// cost and execution cost. A free instruction is typically one that folds
100 /// into another instruction. For example, reg-to-reg moves can often be
101 /// skipped by renaming the registers in the CPU, but they still are encoded
102 /// and thus wouldn't be considered 'free' here.
103 enum TargetCostConstants {
104 TCC_Free = 0, ///< Expected to fold away in lowering.
105 TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
106 TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
109 /// \brief Estimate the cost of a specific operation when lowered.
111 /// Note that this is designed to work on an arbitrary synthetic opcode, and
112 /// thus work for hypothetical queries before an instruction has even been
113 /// formed. However, this does *not* work for GEPs, and must not be called
114 /// for a GEP instruction. Instead, use the dedicated getGEPCost interface as
115 /// analyzing a GEP's cost required more information.
117 /// Typically only the result type is required, and the operand type can be
118 /// omitted. However, if the opcode is one of the cast instructions, the
119 /// operand type is required.
121 /// The returned cost is defined in terms of \c TargetCostConstants, see its
122 /// comments for a detailed explanation of the cost values.
123 virtual unsigned getOperationCost(unsigned Opcode, Type *Ty,
124 Type *OpTy = nullptr) const;
126 /// \brief Estimate the cost of a GEP operation when lowered.
128 /// The contract for this function is the same as \c getOperationCost except
129 /// that it supports an interface that provides extra information specific to
130 /// the GEP operation.
131 virtual unsigned getGEPCost(const Value *Ptr,
132 ArrayRef<const Value *> Operands) const;
134 /// \brief Estimate the cost of a function call when lowered.
136 /// The contract for this is the same as \c getOperationCost except that it
137 /// supports an interface that provides extra information specific to call
140 /// This is the most basic query for estimating call cost: it only knows the
141 /// function type and (potentially) the number of arguments at the call site.
142 /// The latter is only interesting for varargs function types.
143 virtual unsigned getCallCost(FunctionType *FTy, int NumArgs = -1) const;
145 /// \brief Estimate the cost of calling a specific function when lowered.
147 /// This overload adds the ability to reason about the particular function
148 /// being called in the event it is a library call with special lowering.
149 virtual unsigned getCallCost(const Function *F, int NumArgs = -1) const;
151 /// \brief Estimate the cost of calling a specific function when lowered.
153 /// This overload allows specifying a set of candidate argument values.
154 virtual unsigned getCallCost(const Function *F,
155 ArrayRef<const Value *> Arguments) const;
157 /// \brief Estimate the cost of an intrinsic when lowered.
159 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
160 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
161 ArrayRef<Type *> ParamTys) const;
163 /// \brief Estimate the cost of an intrinsic when lowered.
165 /// Mirrors the \c getCallCost method but uses an intrinsic identifier.
166 virtual unsigned getIntrinsicCost(Intrinsic::ID IID, Type *RetTy,
167 ArrayRef<const Value *> Arguments) const;
169 /// \brief Estimate the cost of a given IR user when lowered.
171 /// This can estimate the cost of either a ConstantExpr or Instruction when
172 /// lowered. It has two primary advantages over the \c getOperationCost and
173 /// \c getGEPCost above, and one significant disadvantage: it can only be
174 /// used when the IR construct has already been formed.
176 /// The advantages are that it can inspect the SSA use graph to reason more
177 /// accurately about the cost. For example, all-constant-GEPs can often be
178 /// folded into a load or other instruction, but if they are used in some
179 /// other context they may not be folded. This routine can distinguish such
182 /// The returned cost is defined in terms of \c TargetCostConstants, see its
183 /// comments for a detailed explanation of the cost values.
184 virtual unsigned getUserCost(const User *U) const;
186 /// \brief hasBranchDivergence - Return true if branch divergence exists.
187 /// Branch divergence has a significantly negative impact on GPU performance
188 /// when threads in the same wavefront take different paths due to conditional
190 virtual bool hasBranchDivergence() const;
192 /// \brief Test whether calls to a function lower to actual program function
195 /// The idea is to test whether the program is likely to require a 'call'
196 /// instruction or equivalent in order to call the given function.
198 /// FIXME: It's not clear that this is a good or useful query API. Client's
199 /// should probably move to simpler cost metrics using the above.
200 /// Alternatively, we could split the cost interface into distinct code-size
201 /// and execution-speed costs. This would allow modelling the core of this
202 /// query more accurately as a call is a single small instruction, but
203 /// incurs significant execution cost.
204 virtual bool isLoweredToCall(const Function *F) const;
206 /// Parameters that control the generic loop unrolling transformation.
207 struct UnrollingPreferences {
208 /// The cost threshold for the unrolled loop, compared to
209 /// CodeMetrics.NumInsts aggregated over all basic blocks in the loop body.
210 /// The unrolling factor is set such that the unrolled loop body does not
211 /// exceed this cost. Set this to UINT_MAX to disable the loop body cost
214 /// The cost threshold for the unrolled loop when optimizing for size (set
215 /// to UINT_MAX to disable).
216 unsigned OptSizeThreshold;
217 /// The cost threshold for the unrolled loop, like Threshold, but used
218 /// for partial/runtime unrolling (set to UINT_MAX to disable).
219 unsigned PartialThreshold;
220 /// The cost threshold for the unrolled loop when optimizing for size, like
221 /// OptSizeThreshold, but used for partial/runtime unrolling (set to
222 /// UINT_MAX to disable).
223 unsigned PartialOptSizeThreshold;
224 /// A forced unrolling factor (the number of concatenated bodies of the
225 /// original loop in the unrolled loop body). When set to 0, the unrolling
226 /// transformation will select an unrolling factor based on the current cost
227 /// threshold and other factors.
229 // Set the maximum unrolling factor. The unrolling factor may be selected
230 // using the appropriate cost threshold, but may not exceed this number
231 // (set to UINT_MAX to disable). This does not apply in cases where the
232 // loop is being fully unrolled.
234 /// Allow partial unrolling (unrolling of loops to expand the size of the
235 /// loop body, not only to eliminate small constant-trip-count loops).
237 /// Allow runtime unrolling (unrolling of loops to expand the size of the
238 /// loop body even when the number of loop iterations is not known at
243 /// \brief Get target-customized preferences for the generic loop unrolling
244 /// transformation. The caller will initialize UP with the current
245 /// target-independent defaults.
246 virtual void getUnrollingPreferences(const Function *F, Loop *L,
247 UnrollingPreferences &UP) const;
251 /// \name Scalar Target Information
254 /// \brief Flags indicating the kind of support for population count.
256 /// Compared to the SW implementation, HW support is supposed to
257 /// significantly boost the performance when the population is dense, and it
258 /// may or may not degrade performance if the population is sparse. A HW
259 /// support is considered as "Fast" if it can outperform, or is on a par
260 /// with, SW implementation when the population is sparse; otherwise, it is
261 /// considered as "Slow".
262 enum PopcntSupportKind { PSK_Software, PSK_SlowHardware, PSK_FastHardware };
264 /// \brief Return true if the specified immediate is legal add immediate, that
265 /// is the target has add instructions which can add a register with the
266 /// immediate without having to materialize the immediate into a register.
267 virtual bool isLegalAddImmediate(int64_t Imm) const;
269 /// \brief Return true if the specified immediate is legal icmp immediate,
270 /// that is the target has icmp instructions which can compare a register
271 /// against the immediate without having to materialize the immediate into a
273 virtual bool isLegalICmpImmediate(int64_t Imm) const;
275 /// \brief Return true if the addressing mode represented by AM is legal for
276 /// this target, for a load/store of the specified type.
277 /// The type may be VoidTy, in which case only return true if the addressing
278 /// mode is legal for a load/store of any legal type.
279 /// TODO: Handle pre/postinc as well.
280 virtual bool isLegalAddressingMode(Type *Ty, GlobalValue *BaseGV,
281 int64_t BaseOffset, bool HasBaseReg,
282 int64_t Scale) const;
284 /// \brief Return true if the target works with masked instruction
285 /// AVX2 allows masks for consecutive load and store for i32 and i64 elements.
286 /// AVX-512 architecture will also allow masks for non-consecutive memory
288 virtual bool isLegalMaskedStore(Type *DataType, int Consecutive) const;
289 virtual bool isLegalMaskedLoad(Type *DataType, int Consecutive) const;
291 /// \brief Return the cost of the scaling factor used in the addressing
292 /// mode represented by AM for this target, for a load/store
293 /// of the specified type.
294 /// If the AM is supported, the return value must be >= 0.
295 /// If the AM is not supported, it returns a negative value.
296 /// TODO: Handle pre/postinc as well.
297 virtual int getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
298 int64_t BaseOffset, bool HasBaseReg,
299 int64_t Scale) const;
301 /// \brief Return true if it's free to truncate a value of type Ty1 to type
302 /// Ty2. e.g. On x86 it's free to truncate a i32 value in register EAX to i16
303 /// by referencing its sub-register AX.
304 virtual bool isTruncateFree(Type *Ty1, Type *Ty2) const;
306 /// \brief Return true if this type is legal.
307 virtual bool isTypeLegal(Type *Ty) const;
309 /// \brief Returns the target's jmp_buf alignment in bytes.
310 virtual unsigned getJumpBufAlignment() const;
312 /// \brief Returns the target's jmp_buf size in bytes.
313 virtual unsigned getJumpBufSize() const;
315 /// \brief Return true if switches should be turned into lookup tables for the
317 virtual bool shouldBuildLookupTables() const;
319 /// \brief Return hardware support for population count.
320 virtual PopcntSupportKind getPopcntSupport(unsigned IntTyWidthInBit) const;
322 /// \brief Return true if the hardware has a fast square-root instruction.
323 virtual bool haveFastSqrt(Type *Ty) const;
325 /// \brief Return the expected cost of materializing for the given integer
326 /// immediate of the specified type.
327 virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
329 /// \brief Return the expected cost of materialization for the given integer
330 /// immediate of the specified type for a given instruction. The cost can be
331 /// zero if the immediate can be folded into the specified instruction.
332 virtual unsigned getIntImmCost(unsigned Opc, unsigned Idx, const APInt &Imm,
334 virtual unsigned getIntImmCost(Intrinsic::ID IID, unsigned Idx,
335 const APInt &Imm, Type *Ty) const;
338 /// \name Vector Target Information
341 /// \brief The various kinds of shuffle patterns for vector queries.
343 SK_Broadcast, ///< Broadcast element 0 to all other elements.
344 SK_Reverse, ///< Reverse the order of the vector.
345 SK_Alternate, ///< Choose alternate elements from vector.
346 SK_InsertSubvector, ///< InsertSubvector. Index indicates start offset.
347 SK_ExtractSubvector ///< ExtractSubvector Index indicates start offset.
350 /// \brief Additional information about an operand's possible values.
351 enum OperandValueKind {
352 OK_AnyValue, // Operand can have any value.
353 OK_UniformValue, // Operand is uniform (splat of a value).
354 OK_UniformConstantValue, // Operand is uniform constant.
355 OK_NonUniformConstantValue // Operand is a non uniform constant value.
358 /// \brief Additional properties of an operand's values.
359 enum OperandValueProperties { OP_None = 0, OP_PowerOf2 = 1 };
361 /// \return The number of scalar or vector registers that the target has.
362 /// If 'Vectors' is true, it returns the number of vector registers. If it is
363 /// set to false, it returns the number of scalar registers.
364 virtual unsigned getNumberOfRegisters(bool Vector) const;
366 /// \return The width of the largest scalar or vector register type.
367 virtual unsigned getRegisterBitWidth(bool Vector) const;
369 /// \return The maximum interleave factor that any transform should try to
370 /// perform for this target. This number depends on the level of parallelism
371 /// and the number of execution units in the CPU.
372 virtual unsigned getMaxInterleaveFactor() const;
374 /// \return The expected cost of arithmetic ops, such as mul, xor, fsub, etc.
376 getArithmeticInstrCost(unsigned Opcode, Type *Ty,
377 OperandValueKind Opd1Info = OK_AnyValue,
378 OperandValueKind Opd2Info = OK_AnyValue,
379 OperandValueProperties Opd1PropInfo = OP_None,
380 OperandValueProperties Opd2PropInfo = OP_None) const;
382 /// \return The cost of a shuffle instruction of kind Kind and of type Tp.
383 /// The index and subtype parameters are used by the subvector insertion and
384 /// extraction shuffle kinds.
385 virtual unsigned getShuffleCost(ShuffleKind Kind, Type *Tp, int Index = 0,
386 Type *SubTp = nullptr) const;
388 /// \return The expected cost of cast instructions, such as bitcast, trunc,
390 virtual unsigned getCastInstrCost(unsigned Opcode, Type *Dst,
393 /// \return The expected cost of control-flow related instructions such as
395 virtual unsigned getCFInstrCost(unsigned Opcode) const;
397 /// \returns The expected cost of compare and select instructions.
398 virtual unsigned getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
399 Type *CondTy = nullptr) const;
401 /// \return The expected cost of vector Insert and Extract.
402 /// Use -1 to indicate that there is no information on the index value.
403 virtual unsigned getVectorInstrCost(unsigned Opcode, Type *Val,
404 unsigned Index = -1) const;
406 /// \return The cost of Load and Store instructions.
407 virtual unsigned getMemoryOpCost(unsigned Opcode, Type *Src,
409 unsigned AddressSpace) const;
411 /// \return The cost of masked Load and Store instructions.
412 virtual unsigned getMaskedMemoryOpCost(unsigned Opcode, Type *Src,
414 unsigned AddressSpace) const;
416 /// \brief Calculate the cost of performing a vector reduction.
418 /// This is the cost of reducing the vector value of type \p Ty to a scalar
419 /// value using the operation denoted by \p Opcode. The form of the reduction
420 /// can either be a pairwise reduction or a reduction that splits the vector
421 /// at every reduction level.
425 /// ((v0+v1), (v2, v3), undef, undef)
428 /// ((v0+v2), (v1+v3), undef, undef)
429 virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
430 bool IsPairwiseForm) const;
432 /// \returns The cost of Intrinsic instructions.
433 virtual unsigned getIntrinsicInstrCost(Intrinsic::ID ID, Type *RetTy,
434 ArrayRef<Type *> Tys) const;
436 /// \returns The number of pieces into which the provided type must be
437 /// split during legalization. Zero is returned when the answer is unknown.
438 virtual unsigned getNumberOfParts(Type *Tp) const;
440 /// \returns The cost of the address computation. For most targets this can be
441 /// merged into the instruction indexing mode. Some targets might want to
442 /// distinguish between address computation for memory operations on vector
443 /// types and scalar types. Such targets should override this function.
444 /// The 'IsComplex' parameter is a hint that the address computation is likely
445 /// to involve multiple instructions and as such unlikely to be merged into
446 /// the address indexing mode.
447 virtual unsigned getAddressComputationCost(Type *Ty,
448 bool IsComplex = false) const;
450 /// \returns The cost, if any, of keeping values of the given types alive
453 /// Some types may require the use of register classes that do not have
454 /// any callee-saved registers, so would require a spill and fill.
455 virtual unsigned getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const;
457 /// \returns True if the intrinsic is a supported memory intrinsic. Info
458 /// will contain additional information - whether the intrinsic may write
459 /// or read to memory, volatility and the pointer. Info is undefined
460 /// if false is returned.
461 virtual bool getTgtMemIntrinsic(IntrinsicInst *Inst,
462 MemIntrinsicInfo &Info) const;
464 /// \returns A value which is the result of the given memory intrinsic. New
465 /// instructions may be created to extract the result from the given intrinsic
466 /// memory operation. Returns nullptr if the target cannot create a result
467 /// from the given intrinsic.
468 virtual Value *getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
469 Type *ExpectedType) const;
473 /// Analysis group identification.
477 /// \brief Create the base case instance of a pass in the TTI analysis group.
479 /// This class provides the base case for the stack of TTI analyzes. It doesn't
480 /// delegate to anything and uses the STTI and VTTI objects passed in to
481 /// satisfy the queries.
482 ImmutablePass *createNoTargetTransformInfoPass();
484 } // End llvm namespace