1 | //===- LoopVectorizationPlanner.h - Planner for LoopVectorization ---------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// |
9 | /// \file |
10 | /// This file provides a LoopVectorizationPlanner class. |
11 | /// InnerLoopVectorizer vectorizes loops which contain only one basic |
12 | /// LoopVectorizationPlanner - drives the vectorization process after having |
13 | /// passed Legality checks. |
14 | /// The planner builds and optimizes the Vectorization Plans which record the |
15 | /// decisions how to vectorize the given loop. In particular, represent the |
16 | /// control-flow of the vectorized version, the replication of instructions that |
17 | /// are to be scalarized, and interleave access groups. |
18 | /// |
19 | /// Also provides a VPlan-based builder utility analogous to IRBuilder. |
20 | /// It provides an instruction-level API for generating VPInstructions while |
21 | /// abstracting away the Recipe manipulation details. |
22 | //===----------------------------------------------------------------------===// |
23 | |
24 | #ifndef LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |
25 | #define LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |
26 | |
27 | #include "VPlan.h" |
28 | #include "llvm/ADT/SmallSet.h" |
29 | #include "llvm/Support/InstructionCost.h" |
30 | |
31 | namespace llvm { |
32 | |
33 | class LoopInfo; |
34 | class DominatorTree; |
35 | class LoopVectorizationLegality; |
36 | class LoopVectorizationCostModel; |
37 | class PredicatedScalarEvolution; |
38 | class LoopVectorizeHints; |
39 | class LoopVersioning; |
40 | class ; |
41 | class TargetTransformInfo; |
42 | class TargetLibraryInfo; |
43 | class VPRecipeBuilder; |
44 | struct VFRange; |
45 | |
46 | extern cl::opt<bool> EnableVPlanNativePath; |
47 | extern cl::opt<unsigned> ForceTargetInstructionCost; |
48 | |
49 | /// VPlan-based builder utility analogous to IRBuilder. |
50 | class VPBuilder { |
51 | VPBasicBlock *BB = nullptr; |
52 | VPBasicBlock::iterator InsertPt = VPBasicBlock::iterator(); |
53 | |
54 | /// Insert \p VPI in BB at InsertPt if BB is set. |
55 | template <typename T> T *tryInsertInstruction(T *R) { |
56 | if (BB) |
57 | BB->insert(Recipe: R, InsertPt); |
58 | return R; |
59 | } |
60 | |
61 | VPInstruction *createInstruction(unsigned Opcode, |
62 | ArrayRef<VPValue *> Operands, DebugLoc DL, |
63 | const Twine &Name = "" ) { |
64 | return tryInsertInstruction(R: new VPInstruction(Opcode, Operands, DL, Name)); |
65 | } |
66 | |
67 | VPInstruction *createInstruction(unsigned Opcode, |
68 | std::initializer_list<VPValue *> Operands, |
69 | DebugLoc DL, const Twine &Name = "" ) { |
70 | return createInstruction(Opcode, Operands: ArrayRef<VPValue *>(Operands), DL, Name); |
71 | } |
72 | |
73 | public: |
74 | VPBuilder() = default; |
75 | VPBuilder(VPBasicBlock *InsertBB) { setInsertPoint(InsertBB); } |
76 | VPBuilder(VPRecipeBase *InsertPt) { setInsertPoint(InsertPt); } |
77 | VPBuilder(VPBasicBlock *TheBB, VPBasicBlock::iterator IP) { |
78 | setInsertPoint(TheBB, IP); |
79 | } |
80 | |
81 | /// Clear the insertion point: created instructions will not be inserted into |
82 | /// a block. |
83 | void clearInsertionPoint() { |
84 | BB = nullptr; |
85 | InsertPt = VPBasicBlock::iterator(); |
86 | } |
87 | |
88 | VPBasicBlock *getInsertBlock() const { return BB; } |
89 | VPBasicBlock::iterator getInsertPoint() const { return InsertPt; } |
90 | |
91 | /// Create a VPBuilder to insert after \p R. |
92 | static VPBuilder getToInsertAfter(VPRecipeBase *R) { |
93 | VPBuilder B; |
94 | B.setInsertPoint(TheBB: R->getParent(), IP: std::next(x: R->getIterator())); |
95 | return B; |
96 | } |
97 | |
98 | /// InsertPoint - A saved insertion point. |
99 | class VPInsertPoint { |
100 | VPBasicBlock *Block = nullptr; |
101 | VPBasicBlock::iterator Point; |
102 | |
103 | public: |
104 | /// Creates a new insertion point which doesn't point to anything. |
105 | VPInsertPoint() = default; |
106 | |
107 | /// Creates a new insertion point at the given location. |
108 | VPInsertPoint(VPBasicBlock *InsertBlock, VPBasicBlock::iterator InsertPoint) |
109 | : Block(InsertBlock), Point(InsertPoint) {} |
110 | |
111 | /// Returns true if this insert point is set. |
112 | bool isSet() const { return Block != nullptr; } |
113 | |
114 | VPBasicBlock *getBlock() const { return Block; } |
115 | VPBasicBlock::iterator getPoint() const { return Point; } |
116 | }; |
117 | |
118 | /// Sets the current insert point to a previously-saved location. |
119 | void restoreIP(VPInsertPoint IP) { |
120 | if (IP.isSet()) |
121 | setInsertPoint(TheBB: IP.getBlock(), IP: IP.getPoint()); |
122 | else |
123 | clearInsertionPoint(); |
124 | } |
125 | |
126 | /// This specifies that created VPInstructions should be appended to the end |
127 | /// of the specified block. |
128 | void setInsertPoint(VPBasicBlock *TheBB) { |
129 | assert(TheBB && "Attempting to set a null insert point" ); |
130 | BB = TheBB; |
131 | InsertPt = BB->end(); |
132 | } |
133 | |
134 | /// This specifies that created instructions should be inserted at the |
135 | /// specified point. |
136 | void setInsertPoint(VPBasicBlock *TheBB, VPBasicBlock::iterator IP) { |
137 | BB = TheBB; |
138 | InsertPt = IP; |
139 | } |
140 | |
141 | /// This specifies that created instructions should be inserted at the |
142 | /// specified point. |
143 | void setInsertPoint(VPRecipeBase *IP) { |
144 | BB = IP->getParent(); |
145 | InsertPt = IP->getIterator(); |
146 | } |
147 | |
148 | /// Insert \p R at the current insertion point. |
149 | void insert(VPRecipeBase *R) { BB->insert(Recipe: R, InsertPt); } |
150 | |
151 | /// Create an N-ary operation with \p Opcode, \p Operands and set \p Inst as |
152 | /// its underlying Instruction. |
153 | VPInstruction *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
154 | Instruction *Inst = nullptr, |
155 | const Twine &Name = "" ) { |
156 | DebugLoc DL = DebugLoc::getUnknown(); |
157 | if (Inst) |
158 | DL = Inst->getDebugLoc(); |
159 | VPInstruction *NewVPInst = createInstruction(Opcode, Operands, DL, Name); |
160 | NewVPInst->setUnderlyingValue(Inst); |
161 | return NewVPInst; |
162 | } |
163 | VPInstruction *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
164 | DebugLoc DL, const Twine &Name = "" ) { |
165 | return createInstruction(Opcode, Operands, DL, Name); |
166 | } |
167 | VPInstruction *createNaryOp(unsigned Opcode, ArrayRef<VPValue *> Operands, |
168 | const VPIRFlags &Flags, |
169 | DebugLoc DL = DebugLoc::getUnknown(), |
170 | const Twine &Name = "" ) { |
171 | return tryInsertInstruction( |
172 | R: new VPInstruction(Opcode, Operands, Flags, DL, Name)); |
173 | } |
174 | |
175 | VPInstruction *createNaryOp(unsigned Opcode, |
176 | std::initializer_list<VPValue *> Operands, |
177 | Type *ResultTy, const VPIRFlags &Flags = {}, |
178 | DebugLoc DL = DebugLoc::getUnknown(), |
179 | const Twine &Name = "" ) { |
180 | return tryInsertInstruction( |
181 | R: new VPInstructionWithType(Opcode, Operands, ResultTy, Flags, DL, Name)); |
182 | } |
183 | |
184 | VPInstruction *createOverflowingOp(unsigned Opcode, |
185 | std::initializer_list<VPValue *> Operands, |
186 | VPRecipeWithIRFlags::WrapFlagsTy WrapFlags, |
187 | DebugLoc DL = DebugLoc::getUnknown(), |
188 | const Twine &Name = "" ) { |
189 | return tryInsertInstruction( |
190 | R: new VPInstruction(Opcode, Operands, WrapFlags, DL, Name)); |
191 | } |
192 | |
193 | VPInstruction *createNot(VPValue *Operand, |
194 | DebugLoc DL = DebugLoc::getUnknown(), |
195 | const Twine &Name = "" ) { |
196 | return createInstruction(Opcode: VPInstruction::Not, Operands: {Operand}, DL, Name); |
197 | } |
198 | |
199 | VPInstruction *createAnd(VPValue *LHS, VPValue *RHS, |
200 | DebugLoc DL = DebugLoc::getUnknown(), |
201 | const Twine &Name = "" ) { |
202 | return createInstruction(Opcode: Instruction::BinaryOps::And, Operands: {LHS, RHS}, DL, Name); |
203 | } |
204 | |
205 | VPInstruction *createOr(VPValue *LHS, VPValue *RHS, |
206 | DebugLoc DL = DebugLoc::getUnknown(), |
207 | const Twine &Name = "" ) { |
208 | |
209 | return tryInsertInstruction(R: new VPInstruction( |
210 | Instruction::BinaryOps::Or, {LHS, RHS}, |
211 | VPRecipeWithIRFlags::DisjointFlagsTy(false), DL, Name)); |
212 | } |
213 | |
214 | VPInstruction *createLogicalAnd(VPValue *LHS, VPValue *RHS, |
215 | DebugLoc DL = DebugLoc::getUnknown(), |
216 | const Twine &Name = "" ) { |
217 | return tryInsertInstruction( |
218 | R: new VPInstruction(VPInstruction::LogicalAnd, {LHS, RHS}, DL, Name)); |
219 | } |
220 | |
221 | VPInstruction * |
222 | createSelect(VPValue *Cond, VPValue *TrueVal, VPValue *FalseVal, |
223 | DebugLoc DL = DebugLoc::getUnknown(), const Twine &Name = "" , |
224 | std::optional<FastMathFlags> FMFs = std::nullopt) { |
225 | auto *Select = |
226 | FMFs ? new VPInstruction(Instruction::Select, {Cond, TrueVal, FalseVal}, |
227 | *FMFs, DL, Name) |
228 | : new VPInstruction(Instruction::Select, {Cond, TrueVal, FalseVal}, |
229 | DL, Name); |
230 | return tryInsertInstruction(R: Select); |
231 | } |
232 | |
233 | /// Create a new ICmp VPInstruction with predicate \p Pred and operands \p A |
234 | /// and \p B. |
235 | /// TODO: add createFCmp when needed. |
236 | VPInstruction *createICmp(CmpInst::Predicate Pred, VPValue *A, VPValue *B, |
237 | DebugLoc DL = DebugLoc::getUnknown(), |
238 | const Twine &Name = "" ) { |
239 | assert(Pred >= CmpInst::FIRST_ICMP_PREDICATE && |
240 | Pred <= CmpInst::LAST_ICMP_PREDICATE && "invalid predicate" ); |
241 | return tryInsertInstruction( |
242 | R: new VPInstruction(Instruction::ICmp, {A, B}, Pred, DL, Name)); |
243 | } |
244 | |
245 | VPInstruction *createPtrAdd(VPValue *Ptr, VPValue *Offset, |
246 | DebugLoc DL = DebugLoc::getUnknown(), |
247 | const Twine &Name = "" ) { |
248 | return tryInsertInstruction( |
249 | R: new VPInstruction(VPInstruction::PtrAdd, {Ptr, Offset}, |
250 | GEPNoWrapFlags::none(), DL, Name)); |
251 | } |
252 | VPInstruction *createInBoundsPtrAdd(VPValue *Ptr, VPValue *Offset, |
253 | DebugLoc DL = DebugLoc::getUnknown(), |
254 | const Twine &Name = "" ) { |
255 | return tryInsertInstruction( |
256 | R: new VPInstruction(VPInstruction::PtrAdd, {Ptr, Offset}, |
257 | GEPNoWrapFlags::inBounds(), DL, Name)); |
258 | } |
259 | |
260 | VPPhi *createScalarPhi(ArrayRef<VPValue *> IncomingValues, DebugLoc DL, |
261 | const Twine &Name = "" ) { |
262 | return tryInsertInstruction(R: new VPPhi(IncomingValues, DL, Name)); |
263 | } |
264 | |
265 | /// Convert the input value \p Current to the corresponding value of an |
266 | /// induction with \p Start and \p Step values, using \p Start + \p Current * |
267 | /// \p Step. |
268 | VPDerivedIVRecipe *createDerivedIV(InductionDescriptor::InductionKind Kind, |
269 | FPMathOperator *FPBinOp, VPValue *Start, |
270 | VPValue *Current, VPValue *Step, |
271 | const Twine &Name = "" ) { |
272 | return tryInsertInstruction( |
273 | R: new VPDerivedIVRecipe(Kind, FPBinOp, Start, Current, Step, Name)); |
274 | } |
275 | |
276 | VPInstruction *createScalarCast(Instruction::CastOps Opcode, VPValue *Op, |
277 | Type *ResultTy, DebugLoc DL) { |
278 | return tryInsertInstruction( |
279 | R: new VPInstructionWithType(Opcode, Op, ResultTy, {}, DL)); |
280 | } |
281 | |
282 | VPValue *createScalarZExtOrTrunc(VPValue *Op, Type *ResultTy, Type *SrcTy, |
283 | DebugLoc DL) { |
284 | if (ResultTy == SrcTy) |
285 | return Op; |
286 | Instruction::CastOps CastOp = |
287 | ResultTy->getScalarSizeInBits() < SrcTy->getScalarSizeInBits() |
288 | ? Instruction::Trunc |
289 | : Instruction::ZExt; |
290 | return createScalarCast(Opcode: CastOp, Op, ResultTy, DL); |
291 | } |
292 | |
293 | VPWidenCastRecipe *createWidenCast(Instruction::CastOps Opcode, VPValue *Op, |
294 | Type *ResultTy) { |
295 | return tryInsertInstruction(R: new VPWidenCastRecipe(Opcode, Op, ResultTy)); |
296 | } |
297 | |
298 | VPScalarIVStepsRecipe * |
299 | createScalarIVSteps(Instruction::BinaryOps InductionOpcode, |
300 | FPMathOperator *FPBinOp, VPValue *IV, VPValue *Step, |
301 | VPValue *VF, DebugLoc DL) { |
302 | return tryInsertInstruction(R: new VPScalarIVStepsRecipe( |
303 | IV, Step, VF, InductionOpcode, |
304 | FPBinOp ? FPBinOp->getFastMathFlags() : FastMathFlags(), DL)); |
305 | } |
306 | |
307 | //===--------------------------------------------------------------------===// |
308 | // RAII helpers. |
309 | //===--------------------------------------------------------------------===// |
310 | |
311 | /// RAII object that stores the current insertion point and restores it when |
312 | /// the object is destroyed. |
313 | class InsertPointGuard { |
314 | VPBuilder &Builder; |
315 | VPBasicBlock *Block; |
316 | VPBasicBlock::iterator Point; |
317 | |
318 | public: |
319 | InsertPointGuard(VPBuilder &B) |
320 | : Builder(B), Block(B.getInsertBlock()), Point(B.getInsertPoint()) {} |
321 | |
322 | InsertPointGuard(const InsertPointGuard &) = delete; |
323 | InsertPointGuard &operator=(const InsertPointGuard &) = delete; |
324 | |
325 | ~InsertPointGuard() { Builder.restoreIP(IP: VPInsertPoint(Block, Point)); } |
326 | }; |
327 | }; |
328 | |
329 | /// TODO: The following VectorizationFactor was pulled out of |
330 | /// LoopVectorizationCostModel class. LV also deals with |
331 | /// VectorizerParams::VectorizationFactor. |
332 | /// We need to streamline them. |
333 | |
334 | /// Information about vectorization costs. |
335 | struct VectorizationFactor { |
336 | /// Vector width with best cost. |
337 | ElementCount Width; |
338 | |
339 | /// Cost of the loop with that width. |
340 | InstructionCost Cost; |
341 | |
342 | /// Cost of the scalar loop. |
343 | InstructionCost ScalarCost; |
344 | |
345 | /// The minimum trip count required to make vectorization profitable, e.g. due |
346 | /// to runtime checks. |
347 | ElementCount MinProfitableTripCount; |
348 | |
349 | VectorizationFactor(ElementCount Width, InstructionCost Cost, |
350 | InstructionCost ScalarCost) |
351 | : Width(Width), Cost(Cost), ScalarCost(ScalarCost) {} |
352 | |
353 | /// Width 1 means no vectorization, cost 0 means uncomputed cost. |
354 | static VectorizationFactor Disabled() { |
355 | return {ElementCount::getFixed(MinVal: 1), 0, 0}; |
356 | } |
357 | |
358 | bool operator==(const VectorizationFactor &rhs) const { |
359 | return Width == rhs.Width && Cost == rhs.Cost; |
360 | } |
361 | |
362 | bool operator!=(const VectorizationFactor &rhs) const { |
363 | return !(*this == rhs); |
364 | } |
365 | }; |
366 | |
367 | /// A class that represents two vectorization factors (initialized with 0 by |
368 | /// default). One for fixed-width vectorization and one for scalable |
369 | /// vectorization. This can be used by the vectorizer to choose from a range of |
370 | /// fixed and/or scalable VFs in order to find the most cost-effective VF to |
371 | /// vectorize with. |
372 | struct FixedScalableVFPair { |
373 | ElementCount FixedVF; |
374 | ElementCount ScalableVF; |
375 | |
376 | FixedScalableVFPair() |
377 | : FixedVF(ElementCount::getFixed(MinVal: 0)), |
378 | ScalableVF(ElementCount::getScalable(MinVal: 0)) {} |
379 | FixedScalableVFPair(const ElementCount &Max) : FixedScalableVFPair() { |
380 | *(Max.isScalable() ? &ScalableVF : &FixedVF) = Max; |
381 | } |
382 | FixedScalableVFPair(const ElementCount &FixedVF, |
383 | const ElementCount &ScalableVF) |
384 | : FixedVF(FixedVF), ScalableVF(ScalableVF) { |
385 | assert(!FixedVF.isScalable() && ScalableVF.isScalable() && |
386 | "Invalid scalable properties" ); |
387 | } |
388 | |
389 | static FixedScalableVFPair getNone() { return FixedScalableVFPair(); } |
390 | |
391 | /// \return true if either fixed- or scalable VF is non-zero. |
392 | explicit operator bool() const { return FixedVF || ScalableVF; } |
393 | |
394 | /// \return true if either fixed- or scalable VF is a valid vector VF. |
395 | bool hasVector() const { return FixedVF.isVector() || ScalableVF.isVector(); } |
396 | }; |
397 | |
398 | /// Planner drives the vectorization process after having passed |
399 | /// Legality checks. |
400 | class LoopVectorizationPlanner { |
401 | /// The loop that we evaluate. |
402 | Loop *OrigLoop; |
403 | |
404 | /// Loop Info analysis. |
405 | LoopInfo *LI; |
406 | |
407 | /// The dominator tree. |
408 | DominatorTree *DT; |
409 | |
410 | /// Target Library Info. |
411 | const TargetLibraryInfo *TLI; |
412 | |
413 | /// Target Transform Info. |
414 | const TargetTransformInfo &TTI; |
415 | |
416 | /// The legality analysis. |
417 | LoopVectorizationLegality *Legal; |
418 | |
419 | /// The profitability analysis. |
420 | LoopVectorizationCostModel &CM; |
421 | |
422 | /// The interleaved access analysis. |
423 | InterleavedAccessInfo &IAI; |
424 | |
425 | PredicatedScalarEvolution &PSE; |
426 | |
427 | const LoopVectorizeHints &Hints; |
428 | |
429 | OptimizationRemarkEmitter *ORE; |
430 | |
431 | SmallVector<VPlanPtr, 4> VPlans; |
432 | |
433 | /// Profitable vector factors. |
434 | SmallVector<VectorizationFactor, 8> ProfitableVFs; |
435 | |
436 | /// A builder used to construct the current plan. |
437 | VPBuilder Builder; |
438 | |
439 | /// Computes the cost of \p Plan for vectorization factor \p VF. |
440 | /// |
441 | /// The current implementation requires access to the |
442 | /// LoopVectorizationLegality to handle inductions and reductions, which is |
443 | /// why it is kept separate from the VPlan-only cost infrastructure. |
444 | /// |
445 | /// TODO: Move to VPlan::cost once the use of LoopVectorizationLegality has |
446 | /// been retired. |
447 | InstructionCost cost(VPlan &Plan, ElementCount VF) const; |
448 | |
449 | /// Precompute costs for certain instructions using the legacy cost model. The |
450 | /// function is used to bring up the VPlan-based cost model to initially avoid |
451 | /// taking different decisions due to inaccuracies in the legacy cost model. |
452 | InstructionCost precomputeCosts(VPlan &Plan, ElementCount VF, |
453 | VPCostContext &CostCtx) const; |
454 | |
455 | public: |
456 | LoopVectorizationPlanner( |
457 | Loop *L, LoopInfo *LI, DominatorTree *DT, const TargetLibraryInfo *TLI, |
458 | const TargetTransformInfo &TTI, LoopVectorizationLegality *Legal, |
459 | LoopVectorizationCostModel &CM, InterleavedAccessInfo &IAI, |
460 | PredicatedScalarEvolution &PSE, const LoopVectorizeHints &Hints, |
461 | OptimizationRemarkEmitter *ORE) |
462 | : OrigLoop(L), LI(LI), DT(DT), TLI(TLI), TTI(TTI), Legal(Legal), CM(CM), |
463 | IAI(IAI), PSE(PSE), Hints(Hints), ORE(ORE) {} |
464 | |
465 | /// Build VPlans for the specified \p UserVF and \p UserIC if they are |
466 | /// non-zero or all applicable candidate VFs otherwise. If vectorization and |
467 | /// interleaving should be avoided up-front, no plans are generated. |
468 | void plan(ElementCount UserVF, unsigned UserIC); |
469 | |
470 | /// Use the VPlan-native path to plan how to best vectorize, return the best |
471 | /// VF and its cost. |
472 | VectorizationFactor planInVPlanNativePath(ElementCount UserVF); |
473 | |
474 | /// Return the VPlan for \p VF. At the moment, there is always a single VPlan |
475 | /// for each VF. |
476 | VPlan &getPlanFor(ElementCount VF) const; |
477 | |
478 | /// Compute and return the most profitable vectorization factor. Also collect |
479 | /// all profitable VFs in ProfitableVFs. |
480 | VectorizationFactor computeBestVF(); |
481 | |
482 | /// Generate the IR code for the vectorized loop captured in VPlan \p BestPlan |
483 | /// according to the best selected \p VF and \p UF. |
484 | /// |
485 | /// TODO: \p VectorizingEpilogue indicates if the executed VPlan is for the |
486 | /// epilogue vector loop. It should be removed once the re-use issue has been |
487 | /// fixed. |
488 | /// |
489 | /// Returns a mapping of SCEVs to their expanded IR values. |
490 | /// Note that this is a temporary workaround needed due to the current |
491 | /// epilogue handling. |
492 | DenseMap<const SCEV *, Value *> executePlan(ElementCount VF, unsigned UF, |
493 | VPlan &BestPlan, |
494 | InnerLoopVectorizer &LB, |
495 | DominatorTree *DT, |
496 | bool VectorizingEpilogue); |
497 | |
498 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
499 | void printPlans(raw_ostream &O); |
500 | #endif |
501 | |
502 | /// Look through the existing plans and return true if we have one with |
503 | /// vectorization factor \p VF. |
504 | bool hasPlanWithVF(ElementCount VF) const { |
505 | return any_of(Range: VPlans, |
506 | P: [&](const VPlanPtr &Plan) { return Plan->hasVF(VF); }); |
507 | } |
508 | |
509 | /// Test a \p Predicate on a \p Range of VF's. Return the value of applying |
510 | /// \p Predicate on Range.Start, possibly decreasing Range.End such that the |
511 | /// returned value holds for the entire \p Range. |
512 | static bool |
513 | getDecisionAndClampRange(const std::function<bool(ElementCount)> &Predicate, |
514 | VFRange &Range); |
515 | |
516 | /// \return The most profitable vectorization factor and the cost of that VF |
517 | /// for vectorizing the epilogue. Returns VectorizationFactor::Disabled if |
518 | /// epilogue vectorization is not supported for the loop. |
519 | VectorizationFactor |
520 | selectEpilogueVectorizationFactor(const ElementCount MaxVF, unsigned IC); |
521 | |
522 | /// Emit remarks for recipes with invalid costs in the available VPlans. |
523 | void (OptimizationRemarkEmitter *ORE); |
524 | |
525 | protected: |
526 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
527 | /// according to the information gathered by Legal when it checked if it is |
528 | /// legal to vectorize the loop. |
529 | void buildVPlans(ElementCount MinVF, ElementCount MaxVF); |
530 | |
531 | private: |
532 | /// Build a VPlan according to the information gathered by Legal. \return a |
533 | /// VPlan for vectorization factors \p Range.Start and up to \p Range.End |
534 | /// exclusive, possibly decreasing \p Range.End. If no VPlan can be built for |
535 | /// the input range, set the largest included VF to the maximum VF for which |
536 | /// no plan could be built. |
537 | VPlanPtr tryToBuildVPlan(VFRange &Range); |
538 | |
539 | /// Build a VPlan using VPRecipes according to the information gather by |
540 | /// Legal. This method is only used for the legacy inner loop vectorizer. |
541 | /// \p Range's largest included VF is restricted to the maximum VF the |
542 | /// returned VPlan is valid for. If no VPlan can be built for the input range, |
543 | /// set the largest included VF to the maximum VF for which no plan could be |
544 | /// built. Each VPlan is built starting from a copy of \p InitialPlan, which |
545 | /// is a plain CFG VPlan wrapping the original scalar loop. |
546 | VPlanPtr tryToBuildVPlanWithVPRecipes(VPlanPtr InitialPlan, VFRange &Range, |
547 | LoopVersioning *LVer); |
548 | |
549 | /// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive, |
550 | /// according to the information gathered by Legal when it checked if it is |
551 | /// legal to vectorize the loop. This method creates VPlans using VPRecipes. |
552 | void buildVPlansWithVPRecipes(ElementCount MinVF, ElementCount MaxVF); |
553 | |
554 | // Adjust the recipes for reductions. For in-loop reductions the chain of |
555 | // instructions leading from the loop exit instr to the phi need to be |
556 | // converted to reductions, with one operand being vector and the other being |
557 | // the scalar reduction chain. For other reductions, a select is introduced |
558 | // between the phi and users outside the vector region when folding the tail. |
559 | void adjustRecipesForReductions(VPlanPtr &Plan, |
560 | VPRecipeBuilder &RecipeBuilder, |
561 | ElementCount MinVF); |
562 | |
563 | #ifndef NDEBUG |
564 | /// \return The most profitable vectorization factor for the available VPlans |
565 | /// and the cost of that VF. |
566 | /// This is now only used to verify the decisions by the new VPlan-based |
567 | /// cost-model and will be retired once the VPlan-based cost-model is |
568 | /// stabilized. |
569 | VectorizationFactor selectVectorizationFactor(); |
570 | #endif |
571 | |
572 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than |
573 | /// that of B. |
574 | bool isMoreProfitable(const VectorizationFactor &A, |
575 | const VectorizationFactor &B, bool HasTail) const; |
576 | |
577 | /// Returns true if the per-lane cost of VectorizationFactor A is lower than |
578 | /// that of B in the context of vectorizing a loop with known \p MaxTripCount. |
579 | bool isMoreProfitable(const VectorizationFactor &A, |
580 | const VectorizationFactor &B, |
581 | const unsigned MaxTripCount, bool HasTail) const; |
582 | |
583 | /// Determines if we have the infrastructure to vectorize the loop and its |
584 | /// epilogue, assuming the main loop is vectorized by \p VF. |
585 | bool isCandidateForEpilogueVectorization(const ElementCount VF) const; |
586 | }; |
587 | |
588 | } // namespace llvm |
589 | |
590 | #endif // LLVM_TRANSFORMS_VECTORIZE_LOOPVECTORIZATIONPLANNER_H |
591 | |