xref: /freebsd/contrib/llvm-project/llvm/lib/Transforms/Vectorize/VPlan.h (revision a03411e84728e9b267056fd31c7d1d9d1dc1b01e)
1 //===- VPlan.h - Represent A Vectorizer Plan --------------------*- C++ -*-===//
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 contains the declarations of the Vectorization Plan base classes:
11 /// 1. VPBasicBlock and VPRegionBlock that inherit from a common pure virtual
12 ///    VPBlockBase, together implementing a Hierarchical CFG;
13 /// 2. Pure virtual VPRecipeBase serving as the base class for recipes contained
14 ///    within VPBasicBlocks;
15 /// 3. VPInstruction, a concrete Recipe and VPUser modeling a single planned
16 ///    instruction;
17 /// 4. The VPlan class holding a candidate for vectorization;
18 /// 5. The VPlanPrinter class providing a way to print a plan in dot format;
19 /// These are documented in docs/VectorizationPlan.rst.
20 //
21 //===----------------------------------------------------------------------===//
22 
23 #ifndef LLVM_TRANSFORMS_VECTORIZE_VPLAN_H
24 #define LLVM_TRANSFORMS_VECTORIZE_VPLAN_H
25 
26 #include "VPlanValue.h"
27 #include "llvm/ADT/DenseMap.h"
28 #include "llvm/ADT/MapVector.h"
29 #include "llvm/ADT/SmallBitVector.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallVector.h"
32 #include "llvm/ADT/Twine.h"
33 #include "llvm/ADT/ilist.h"
34 #include "llvm/ADT/ilist_node.h"
35 #include "llvm/Analysis/IVDescriptors.h"
36 #include "llvm/Analysis/LoopInfo.h"
37 #include "llvm/Analysis/VectorUtils.h"
38 #include "llvm/IR/DebugLoc.h"
39 #include "llvm/IR/FMF.h"
40 #include "llvm/IR/Operator.h"
41 #include <algorithm>
42 #include <cassert>
43 #include <cstddef>
44 #include <string>
45 
46 namespace llvm {
47 
48 class BasicBlock;
49 class DominatorTree;
50 class InnerLoopVectorizer;
51 class IRBuilderBase;
52 class LoopInfo;
53 class raw_ostream;
54 class RecurrenceDescriptor;
55 class SCEV;
56 class Type;
57 class VPBasicBlock;
58 class VPRegionBlock;
59 class VPlan;
60 class VPReplicateRecipe;
61 class VPlanSlp;
62 class Value;
63 class LoopVersioning;
64 
65 namespace Intrinsic {
66 typedef unsigned ID;
67 }
68 
69 /// Returns a calculation for the total number of elements for a given \p VF.
70 /// For fixed width vectors this value is a constant, whereas for scalable
71 /// vectors it is an expression determined at runtime.
72 Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF);
73 
74 /// Return a value for Step multiplied by VF.
75 Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF,
76                        int64_t Step);
77 
78 const SCEV *createTripCountSCEV(Type *IdxTy, PredicatedScalarEvolution &PSE,
79                                 Loop *CurLoop = nullptr);
80 
81 /// A range of powers-of-2 vectorization factors with fixed start and
82 /// adjustable end. The range includes start and excludes end, e.g.,:
83 /// [1, 16) = {1, 2, 4, 8}
84 struct VFRange {
85   // A power of 2.
86   const ElementCount Start;
87 
88   // A power of 2. If End <= Start range is empty.
89   ElementCount End;
90 
91   bool isEmpty() const {
92     return End.getKnownMinValue() <= Start.getKnownMinValue();
93   }
94 
95   VFRange(const ElementCount &Start, const ElementCount &End)
96       : Start(Start), End(End) {
97     assert(Start.isScalable() == End.isScalable() &&
98            "Both Start and End should have the same scalable flag");
99     assert(isPowerOf2_32(Start.getKnownMinValue()) &&
100            "Expected Start to be a power of 2");
101     assert(isPowerOf2_32(End.getKnownMinValue()) &&
102            "Expected End to be a power of 2");
103   }
104 
105   /// Iterator to iterate over vectorization factors in a VFRange.
106   class iterator
107       : public iterator_facade_base<iterator, std::forward_iterator_tag,
108                                     ElementCount> {
109     ElementCount VF;
110 
111   public:
112     iterator(ElementCount VF) : VF(VF) {}
113 
114     bool operator==(const iterator &Other) const { return VF == Other.VF; }
115 
116     ElementCount operator*() const { return VF; }
117 
118     iterator &operator++() {
119       VF *= 2;
120       return *this;
121     }
122   };
123 
124   iterator begin() { return iterator(Start); }
125   iterator end() {
126     assert(isPowerOf2_32(End.getKnownMinValue()));
127     return iterator(End);
128   }
129 };
130 
131 using VPlanPtr = std::unique_ptr<VPlan>;
132 
133 /// In what follows, the term "input IR" refers to code that is fed into the
134 /// vectorizer whereas the term "output IR" refers to code that is generated by
135 /// the vectorizer.
136 
137 /// VPLane provides a way to access lanes in both fixed width and scalable
138 /// vectors, where for the latter the lane index sometimes needs calculating
139 /// as a runtime expression.
140 class VPLane {
141 public:
142   /// Kind describes how to interpret Lane.
143   enum class Kind : uint8_t {
144     /// For First, Lane is the index into the first N elements of a
145     /// fixed-vector <N x <ElTy>> or a scalable vector <vscale x N x <ElTy>>.
146     First,
147     /// For ScalableLast, Lane is the offset from the start of the last
148     /// N-element subvector in a scalable vector <vscale x N x <ElTy>>. For
149     /// example, a Lane of 0 corresponds to lane `(vscale - 1) * N`, a Lane of
150     /// 1 corresponds to `((vscale - 1) * N) + 1`, etc.
151     ScalableLast
152   };
153 
154 private:
155   /// in [0..VF)
156   unsigned Lane;
157 
158   /// Indicates how the Lane should be interpreted, as described above.
159   Kind LaneKind;
160 
161 public:
162   VPLane(unsigned Lane, Kind LaneKind) : Lane(Lane), LaneKind(LaneKind) {}
163 
164   static VPLane getFirstLane() { return VPLane(0, VPLane::Kind::First); }
165 
166   static VPLane getLastLaneForVF(const ElementCount &VF) {
167     unsigned LaneOffset = VF.getKnownMinValue() - 1;
168     Kind LaneKind;
169     if (VF.isScalable())
170       // In this case 'LaneOffset' refers to the offset from the start of the
171       // last subvector with VF.getKnownMinValue() elements.
172       LaneKind = VPLane::Kind::ScalableLast;
173     else
174       LaneKind = VPLane::Kind::First;
175     return VPLane(LaneOffset, LaneKind);
176   }
177 
178   /// Returns a compile-time known value for the lane index and asserts if the
179   /// lane can only be calculated at runtime.
180   unsigned getKnownLane() const {
181     assert(LaneKind == Kind::First);
182     return Lane;
183   }
184 
185   /// Returns an expression describing the lane index that can be used at
186   /// runtime.
187   Value *getAsRuntimeExpr(IRBuilderBase &Builder, const ElementCount &VF) const;
188 
189   /// Returns the Kind of lane offset.
190   Kind getKind() const { return LaneKind; }
191 
192   /// Returns true if this is the first lane of the whole vector.
193   bool isFirstLane() const { return Lane == 0 && LaneKind == Kind::First; }
194 
195   /// Maps the lane to a cache index based on \p VF.
196   unsigned mapToCacheIndex(const ElementCount &VF) const {
197     switch (LaneKind) {
198     case VPLane::Kind::ScalableLast:
199       assert(VF.isScalable() && Lane < VF.getKnownMinValue());
200       return VF.getKnownMinValue() + Lane;
201     default:
202       assert(Lane < VF.getKnownMinValue());
203       return Lane;
204     }
205   }
206 
207   /// Returns the maxmimum number of lanes that we are able to consider
208   /// caching for \p VF.
209   static unsigned getNumCachedLanes(const ElementCount &VF) {
210     return VF.getKnownMinValue() * (VF.isScalable() ? 2 : 1);
211   }
212 };
213 
214 /// VPIteration represents a single point in the iteration space of the output
215 /// (vectorized and/or unrolled) IR loop.
216 struct VPIteration {
217   /// in [0..UF)
218   unsigned Part;
219 
220   VPLane Lane;
221 
222   VPIteration(unsigned Part, unsigned Lane,
223               VPLane::Kind Kind = VPLane::Kind::First)
224       : Part(Part), Lane(Lane, Kind) {}
225 
226   VPIteration(unsigned Part, const VPLane &Lane) : Part(Part), Lane(Lane) {}
227 
228   bool isFirstIteration() const { return Part == 0 && Lane.isFirstLane(); }
229 };
230 
231 /// VPTransformState holds information passed down when "executing" a VPlan,
232 /// needed for generating the output IR.
233 struct VPTransformState {
234   VPTransformState(ElementCount VF, unsigned UF, LoopInfo *LI,
235                    DominatorTree *DT, IRBuilderBase &Builder,
236                    InnerLoopVectorizer *ILV, VPlan *Plan)
237       : VF(VF), UF(UF), LI(LI), DT(DT), Builder(Builder), ILV(ILV), Plan(Plan),
238         LVer(nullptr) {}
239 
240   /// The chosen Vectorization and Unroll Factors of the loop being vectorized.
241   ElementCount VF;
242   unsigned UF;
243 
244   /// Hold the indices to generate specific scalar instructions. Null indicates
245   /// that all instances are to be generated, using either scalar or vector
246   /// instructions.
247   std::optional<VPIteration> Instance;
248 
249   struct DataState {
250     /// A type for vectorized values in the new loop. Each value from the
251     /// original loop, when vectorized, is represented by UF vector values in
252     /// the new unrolled loop, where UF is the unroll factor.
253     typedef SmallVector<Value *, 2> PerPartValuesTy;
254 
255     DenseMap<VPValue *, PerPartValuesTy> PerPartOutput;
256 
257     using ScalarsPerPartValuesTy = SmallVector<SmallVector<Value *, 4>, 2>;
258     DenseMap<VPValue *, ScalarsPerPartValuesTy> PerPartScalars;
259   } Data;
260 
261   /// Get the generated Value for a given VPValue and a given Part. Note that
262   /// as some Defs are still created by ILV and managed in its ValueMap, this
263   /// method will delegate the call to ILV in such cases in order to provide
264   /// callers a consistent API.
265   /// \see set.
266   Value *get(VPValue *Def, unsigned Part);
267 
268   /// Get the generated Value for a given VPValue and given Part and Lane.
269   Value *get(VPValue *Def, const VPIteration &Instance);
270 
271   bool hasVectorValue(VPValue *Def, unsigned Part) {
272     auto I = Data.PerPartOutput.find(Def);
273     return I != Data.PerPartOutput.end() && Part < I->second.size() &&
274            I->second[Part];
275   }
276 
277   bool hasAnyVectorValue(VPValue *Def) const {
278     return Data.PerPartOutput.contains(Def);
279   }
280 
281   bool hasScalarValue(VPValue *Def, VPIteration Instance) {
282     auto I = Data.PerPartScalars.find(Def);
283     if (I == Data.PerPartScalars.end())
284       return false;
285     unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF);
286     return Instance.Part < I->second.size() &&
287            CacheIdx < I->second[Instance.Part].size() &&
288            I->second[Instance.Part][CacheIdx];
289   }
290 
291   /// Set the generated Value for a given VPValue and a given Part.
292   void set(VPValue *Def, Value *V, unsigned Part) {
293     if (!Data.PerPartOutput.count(Def)) {
294       DataState::PerPartValuesTy Entry(UF);
295       Data.PerPartOutput[Def] = Entry;
296     }
297     Data.PerPartOutput[Def][Part] = V;
298   }
299   /// Reset an existing vector value for \p Def and a given \p Part.
300   void reset(VPValue *Def, Value *V, unsigned Part) {
301     auto Iter = Data.PerPartOutput.find(Def);
302     assert(Iter != Data.PerPartOutput.end() &&
303            "need to overwrite existing value");
304     Iter->second[Part] = V;
305   }
306 
307   /// Set the generated scalar \p V for \p Def and the given \p Instance.
308   void set(VPValue *Def, Value *V, const VPIteration &Instance) {
309     auto Iter = Data.PerPartScalars.insert({Def, {}});
310     auto &PerPartVec = Iter.first->second;
311     while (PerPartVec.size() <= Instance.Part)
312       PerPartVec.emplace_back();
313     auto &Scalars = PerPartVec[Instance.Part];
314     unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF);
315     while (Scalars.size() <= CacheIdx)
316       Scalars.push_back(nullptr);
317     assert(!Scalars[CacheIdx] && "should overwrite existing value");
318     Scalars[CacheIdx] = V;
319   }
320 
321   /// Reset an existing scalar value for \p Def and a given \p Instance.
322   void reset(VPValue *Def, Value *V, const VPIteration &Instance) {
323     auto Iter = Data.PerPartScalars.find(Def);
324     assert(Iter != Data.PerPartScalars.end() &&
325            "need to overwrite existing value");
326     assert(Instance.Part < Iter->second.size() &&
327            "need to overwrite existing value");
328     unsigned CacheIdx = Instance.Lane.mapToCacheIndex(VF);
329     assert(CacheIdx < Iter->second[Instance.Part].size() &&
330            "need to overwrite existing value");
331     Iter->second[Instance.Part][CacheIdx] = V;
332   }
333 
334   /// Add additional metadata to \p To that was not present on \p Orig.
335   ///
336   /// Currently this is used to add the noalias annotations based on the
337   /// inserted memchecks.  Use this for instructions that are *cloned* into the
338   /// vector loop.
339   void addNewMetadata(Instruction *To, const Instruction *Orig);
340 
341   /// Add metadata from one instruction to another.
342   ///
343   /// This includes both the original MDs from \p From and additional ones (\see
344   /// addNewMetadata).  Use this for *newly created* instructions in the vector
345   /// loop.
346   void addMetadata(Instruction *To, Instruction *From);
347 
348   /// Similar to the previous function but it adds the metadata to a
349   /// vector of instructions.
350   void addMetadata(ArrayRef<Value *> To, Instruction *From);
351 
352   /// Set the debug location in the builder using the debug location in \p V.
353   void setDebugLocFromInst(const Value *V);
354 
355   /// Hold state information used when constructing the CFG of the output IR,
356   /// traversing the VPBasicBlocks and generating corresponding IR BasicBlocks.
357   struct CFGState {
358     /// The previous VPBasicBlock visited. Initially set to null.
359     VPBasicBlock *PrevVPBB = nullptr;
360 
361     /// The previous IR BasicBlock created or used. Initially set to the new
362     /// header BasicBlock.
363     BasicBlock *PrevBB = nullptr;
364 
365     /// The last IR BasicBlock in the output IR. Set to the exit block of the
366     /// vector loop.
367     BasicBlock *ExitBB = nullptr;
368 
369     /// A mapping of each VPBasicBlock to the corresponding BasicBlock. In case
370     /// of replication, maps the BasicBlock of the last replica created.
371     SmallDenseMap<VPBasicBlock *, BasicBlock *> VPBB2IRBB;
372 
373     CFGState() = default;
374 
375     /// Returns the BasicBlock* mapped to the pre-header of the loop region
376     /// containing \p R.
377     BasicBlock *getPreheaderBBFor(VPRecipeBase *R);
378   } CFG;
379 
380   /// Hold a pointer to LoopInfo to register new basic blocks in the loop.
381   LoopInfo *LI;
382 
383   /// Hold a pointer to Dominator Tree to register new basic blocks in the loop.
384   DominatorTree *DT;
385 
386   /// Hold a reference to the IRBuilder used to generate output IR code.
387   IRBuilderBase &Builder;
388 
389   VPValue2ValueTy VPValue2Value;
390 
391   /// Hold the canonical scalar IV of the vector loop (start=0, step=VF*UF).
392   Value *CanonicalIV = nullptr;
393 
394   /// Hold a pointer to InnerLoopVectorizer to reuse its IR generation methods.
395   InnerLoopVectorizer *ILV;
396 
397   /// Pointer to the VPlan code is generated for.
398   VPlan *Plan;
399 
400   /// The loop object for the current parent region, or nullptr.
401   Loop *CurrentVectorLoop = nullptr;
402 
403   /// LoopVersioning.  It's only set up (non-null) if memchecks were
404   /// used.
405   ///
406   /// This is currently only used to add no-alias metadata based on the
407   /// memchecks.  The actually versioning is performed manually.
408   LoopVersioning *LVer = nullptr;
409 
410   /// Map SCEVs to their expanded values. Populated when executing
411   /// VPExpandSCEVRecipes.
412   DenseMap<const SCEV *, Value *> ExpandedSCEVs;
413 };
414 
415 /// VPBlockBase is the building block of the Hierarchical Control-Flow Graph.
416 /// A VPBlockBase can be either a VPBasicBlock or a VPRegionBlock.
417 class VPBlockBase {
418   friend class VPBlockUtils;
419 
420   const unsigned char SubclassID; ///< Subclass identifier (for isa/dyn_cast).
421 
422   /// An optional name for the block.
423   std::string Name;
424 
425   /// The immediate VPRegionBlock which this VPBlockBase belongs to, or null if
426   /// it is a topmost VPBlockBase.
427   VPRegionBlock *Parent = nullptr;
428 
429   /// List of predecessor blocks.
430   SmallVector<VPBlockBase *, 1> Predecessors;
431 
432   /// List of successor blocks.
433   SmallVector<VPBlockBase *, 1> Successors;
434 
435   /// VPlan containing the block. Can only be set on the entry block of the
436   /// plan.
437   VPlan *Plan = nullptr;
438 
439   /// Add \p Successor as the last successor to this block.
440   void appendSuccessor(VPBlockBase *Successor) {
441     assert(Successor && "Cannot add nullptr successor!");
442     Successors.push_back(Successor);
443   }
444 
445   /// Add \p Predecessor as the last predecessor to this block.
446   void appendPredecessor(VPBlockBase *Predecessor) {
447     assert(Predecessor && "Cannot add nullptr predecessor!");
448     Predecessors.push_back(Predecessor);
449   }
450 
451   /// Remove \p Predecessor from the predecessors of this block.
452   void removePredecessor(VPBlockBase *Predecessor) {
453     auto Pos = find(Predecessors, Predecessor);
454     assert(Pos && "Predecessor does not exist");
455     Predecessors.erase(Pos);
456   }
457 
458   /// Remove \p Successor from the successors of this block.
459   void removeSuccessor(VPBlockBase *Successor) {
460     auto Pos = find(Successors, Successor);
461     assert(Pos && "Successor does not exist");
462     Successors.erase(Pos);
463   }
464 
465 protected:
466   VPBlockBase(const unsigned char SC, const std::string &N)
467       : SubclassID(SC), Name(N) {}
468 
469 public:
470   /// An enumeration for keeping track of the concrete subclass of VPBlockBase
471   /// that are actually instantiated. Values of this enumeration are kept in the
472   /// SubclassID field of the VPBlockBase objects. They are used for concrete
473   /// type identification.
474   using VPBlockTy = enum { VPBasicBlockSC, VPRegionBlockSC };
475 
476   using VPBlocksTy = SmallVectorImpl<VPBlockBase *>;
477 
478   virtual ~VPBlockBase() = default;
479 
480   const std::string &getName() const { return Name; }
481 
482   void setName(const Twine &newName) { Name = newName.str(); }
483 
484   /// \return an ID for the concrete type of this object.
485   /// This is used to implement the classof checks. This should not be used
486   /// for any other purpose, as the values may change as LLVM evolves.
487   unsigned getVPBlockID() const { return SubclassID; }
488 
489   VPRegionBlock *getParent() { return Parent; }
490   const VPRegionBlock *getParent() const { return Parent; }
491 
492   /// \return A pointer to the plan containing the current block.
493   VPlan *getPlan();
494   const VPlan *getPlan() const;
495 
496   /// Sets the pointer of the plan containing the block. The block must be the
497   /// entry block into the VPlan.
498   void setPlan(VPlan *ParentPlan);
499 
500   void setParent(VPRegionBlock *P) { Parent = P; }
501 
502   /// \return the VPBasicBlock that is the entry of this VPBlockBase,
503   /// recursively, if the latter is a VPRegionBlock. Otherwise, if this
504   /// VPBlockBase is a VPBasicBlock, it is returned.
505   const VPBasicBlock *getEntryBasicBlock() const;
506   VPBasicBlock *getEntryBasicBlock();
507 
508   /// \return the VPBasicBlock that is the exiting this VPBlockBase,
509   /// recursively, if the latter is a VPRegionBlock. Otherwise, if this
510   /// VPBlockBase is a VPBasicBlock, it is returned.
511   const VPBasicBlock *getExitingBasicBlock() const;
512   VPBasicBlock *getExitingBasicBlock();
513 
514   const VPBlocksTy &getSuccessors() const { return Successors; }
515   VPBlocksTy &getSuccessors() { return Successors; }
516 
517   iterator_range<VPBlockBase **> successors() { return Successors; }
518 
519   const VPBlocksTy &getPredecessors() const { return Predecessors; }
520   VPBlocksTy &getPredecessors() { return Predecessors; }
521 
522   /// \return the successor of this VPBlockBase if it has a single successor.
523   /// Otherwise return a null pointer.
524   VPBlockBase *getSingleSuccessor() const {
525     return (Successors.size() == 1 ? *Successors.begin() : nullptr);
526   }
527 
528   /// \return the predecessor of this VPBlockBase if it has a single
529   /// predecessor. Otherwise return a null pointer.
530   VPBlockBase *getSinglePredecessor() const {
531     return (Predecessors.size() == 1 ? *Predecessors.begin() : nullptr);
532   }
533 
534   size_t getNumSuccessors() const { return Successors.size(); }
535   size_t getNumPredecessors() const { return Predecessors.size(); }
536 
537   /// An Enclosing Block of a block B is any block containing B, including B
538   /// itself. \return the closest enclosing block starting from "this", which
539   /// has successors. \return the root enclosing block if all enclosing blocks
540   /// have no successors.
541   VPBlockBase *getEnclosingBlockWithSuccessors();
542 
543   /// \return the closest enclosing block starting from "this", which has
544   /// predecessors. \return the root enclosing block if all enclosing blocks
545   /// have no predecessors.
546   VPBlockBase *getEnclosingBlockWithPredecessors();
547 
548   /// \return the successors either attached directly to this VPBlockBase or, if
549   /// this VPBlockBase is the exit block of a VPRegionBlock and has no
550   /// successors of its own, search recursively for the first enclosing
551   /// VPRegionBlock that has successors and return them. If no such
552   /// VPRegionBlock exists, return the (empty) successors of the topmost
553   /// VPBlockBase reached.
554   const VPBlocksTy &getHierarchicalSuccessors() {
555     return getEnclosingBlockWithSuccessors()->getSuccessors();
556   }
557 
558   /// \return the hierarchical successor of this VPBlockBase if it has a single
559   /// hierarchical successor. Otherwise return a null pointer.
560   VPBlockBase *getSingleHierarchicalSuccessor() {
561     return getEnclosingBlockWithSuccessors()->getSingleSuccessor();
562   }
563 
564   /// \return the predecessors either attached directly to this VPBlockBase or,
565   /// if this VPBlockBase is the entry block of a VPRegionBlock and has no
566   /// predecessors of its own, search recursively for the first enclosing
567   /// VPRegionBlock that has predecessors and return them. If no such
568   /// VPRegionBlock exists, return the (empty) predecessors of the topmost
569   /// VPBlockBase reached.
570   const VPBlocksTy &getHierarchicalPredecessors() {
571     return getEnclosingBlockWithPredecessors()->getPredecessors();
572   }
573 
574   /// \return the hierarchical predecessor of this VPBlockBase if it has a
575   /// single hierarchical predecessor. Otherwise return a null pointer.
576   VPBlockBase *getSingleHierarchicalPredecessor() {
577     return getEnclosingBlockWithPredecessors()->getSinglePredecessor();
578   }
579 
580   /// Set a given VPBlockBase \p Successor as the single successor of this
581   /// VPBlockBase. This VPBlockBase is not added as predecessor of \p Successor.
582   /// This VPBlockBase must have no successors.
583   void setOneSuccessor(VPBlockBase *Successor) {
584     assert(Successors.empty() && "Setting one successor when others exist.");
585     appendSuccessor(Successor);
586   }
587 
588   /// Set two given VPBlockBases \p IfTrue and \p IfFalse to be the two
589   /// successors of this VPBlockBase. This VPBlockBase is not added as
590   /// predecessor of \p IfTrue or \p IfFalse. This VPBlockBase must have no
591   /// successors.
592   void setTwoSuccessors(VPBlockBase *IfTrue, VPBlockBase *IfFalse) {
593     assert(Successors.empty() && "Setting two successors when others exist.");
594     appendSuccessor(IfTrue);
595     appendSuccessor(IfFalse);
596   }
597 
598   /// Set each VPBasicBlock in \p NewPreds as predecessor of this VPBlockBase.
599   /// This VPBlockBase must have no predecessors. This VPBlockBase is not added
600   /// as successor of any VPBasicBlock in \p NewPreds.
601   void setPredecessors(ArrayRef<VPBlockBase *> NewPreds) {
602     assert(Predecessors.empty() && "Block predecessors already set.");
603     for (auto *Pred : NewPreds)
604       appendPredecessor(Pred);
605   }
606 
607   /// Remove all the predecessor of this block.
608   void clearPredecessors() { Predecessors.clear(); }
609 
610   /// Remove all the successors of this block.
611   void clearSuccessors() { Successors.clear(); }
612 
613   /// The method which generates the output IR that correspond to this
614   /// VPBlockBase, thereby "executing" the VPlan.
615   virtual void execute(VPTransformState *State) = 0;
616 
617   /// Delete all blocks reachable from a given VPBlockBase, inclusive.
618   static void deleteCFG(VPBlockBase *Entry);
619 
620   /// Return true if it is legal to hoist instructions into this block.
621   bool isLegalToHoistInto() {
622     // There are currently no constraints that prevent an instruction to be
623     // hoisted into a VPBlockBase.
624     return true;
625   }
626 
627   /// Replace all operands of VPUsers in the block with \p NewValue and also
628   /// replaces all uses of VPValues defined in the block with NewValue.
629   virtual void dropAllReferences(VPValue *NewValue) = 0;
630 
631 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
632   void printAsOperand(raw_ostream &OS, bool PrintType) const {
633     OS << getName();
634   }
635 
636   /// Print plain-text dump of this VPBlockBase to \p O, prefixing all lines
637   /// with \p Indent. \p SlotTracker is used to print unnamed VPValue's using
638   /// consequtive numbers.
639   ///
640   /// Note that the numbering is applied to the whole VPlan, so printing
641   /// individual blocks is consistent with the whole VPlan printing.
642   virtual void print(raw_ostream &O, const Twine &Indent,
643                      VPSlotTracker &SlotTracker) const = 0;
644 
645   /// Print plain-text dump of this VPlan to \p O.
646   void print(raw_ostream &O) const {
647     VPSlotTracker SlotTracker(getPlan());
648     print(O, "", SlotTracker);
649   }
650 
651   /// Print the successors of this block to \p O, prefixing all lines with \p
652   /// Indent.
653   void printSuccessors(raw_ostream &O, const Twine &Indent) const;
654 
655   /// Dump this VPBlockBase to dbgs().
656   LLVM_DUMP_METHOD void dump() const { print(dbgs()); }
657 #endif
658 };
659 
660 /// A value that is used outside the VPlan. The operand of the user needs to be
661 /// added to the associated LCSSA phi node.
662 class VPLiveOut : public VPUser {
663   PHINode *Phi;
664 
665 public:
666   VPLiveOut(PHINode *Phi, VPValue *Op)
667       : VPUser({Op}, VPUser::VPUserID::LiveOut), Phi(Phi) {}
668 
669   static inline bool classof(const VPUser *U) {
670     return U->getVPUserID() == VPUser::VPUserID::LiveOut;
671   }
672 
673   /// Fixup the wrapped LCSSA phi node in the unique exit block.  This simply
674   /// means we need to add the appropriate incoming value from the middle
675   /// block as exiting edges from the scalar epilogue loop (if present) are
676   /// already in place, and we exit the vector loop exclusively to the middle
677   /// block.
678   void fixPhi(VPlan &Plan, VPTransformState &State);
679 
680   /// Returns true if the VPLiveOut uses scalars of operand \p Op.
681   bool usesScalars(const VPValue *Op) const override {
682     assert(is_contained(operands(), Op) &&
683            "Op must be an operand of the recipe");
684     return true;
685   }
686 
687   PHINode *getPhi() const { return Phi; }
688 
689 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
690   /// Print the VPLiveOut to \p O.
691   void print(raw_ostream &O, VPSlotTracker &SlotTracker) const;
692 #endif
693 };
694 
695 /// VPRecipeBase is a base class modeling a sequence of one or more output IR
696 /// instructions. VPRecipeBase owns the the VPValues it defines through VPDef
697 /// and is responsible for deleting its defined values. Single-value
698 /// VPRecipeBases that also inherit from VPValue must make sure to inherit from
699 /// VPRecipeBase before VPValue.
700 class VPRecipeBase : public ilist_node_with_parent<VPRecipeBase, VPBasicBlock>,
701                      public VPDef,
702                      public VPUser {
703   friend VPBasicBlock;
704   friend class VPBlockUtils;
705 
706   /// Each VPRecipe belongs to a single VPBasicBlock.
707   VPBasicBlock *Parent = nullptr;
708 
709 public:
710   VPRecipeBase(const unsigned char SC, ArrayRef<VPValue *> Operands)
711       : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {}
712 
713   template <typename IterT>
714   VPRecipeBase(const unsigned char SC, iterator_range<IterT> Operands)
715       : VPDef(SC), VPUser(Operands, VPUser::VPUserID::Recipe) {}
716   virtual ~VPRecipeBase() = default;
717 
718   /// \return the VPBasicBlock which this VPRecipe belongs to.
719   VPBasicBlock *getParent() { return Parent; }
720   const VPBasicBlock *getParent() const { return Parent; }
721 
722   /// The method which generates the output IR instructions that correspond to
723   /// this VPRecipe, thereby "executing" the VPlan.
724   virtual void execute(VPTransformState &State) = 0;
725 
726   /// Insert an unlinked recipe into a basic block immediately before
727   /// the specified recipe.
728   void insertBefore(VPRecipeBase *InsertPos);
729   /// Insert an unlinked recipe into \p BB immediately before the insertion
730   /// point \p IP;
731   void insertBefore(VPBasicBlock &BB, iplist<VPRecipeBase>::iterator IP);
732 
733   /// Insert an unlinked Recipe into a basic block immediately after
734   /// the specified Recipe.
735   void insertAfter(VPRecipeBase *InsertPos);
736 
737   /// Unlink this recipe from its current VPBasicBlock and insert it into
738   /// the VPBasicBlock that MovePos lives in, right after MovePos.
739   void moveAfter(VPRecipeBase *MovePos);
740 
741   /// Unlink this recipe and insert into BB before I.
742   ///
743   /// \pre I is a valid iterator into BB.
744   void moveBefore(VPBasicBlock &BB, iplist<VPRecipeBase>::iterator I);
745 
746   /// This method unlinks 'this' from the containing basic block, but does not
747   /// delete it.
748   void removeFromParent();
749 
750   /// This method unlinks 'this' from the containing basic block and deletes it.
751   ///
752   /// \returns an iterator pointing to the element after the erased one
753   iplist<VPRecipeBase>::iterator eraseFromParent();
754 
755   /// Returns the underlying instruction, if the recipe is a VPValue or nullptr
756   /// otherwise.
757   Instruction *getUnderlyingInstr() {
758     return cast<Instruction>(getVPSingleValue()->getUnderlyingValue());
759   }
760   const Instruction *getUnderlyingInstr() const {
761     return cast<Instruction>(getVPSingleValue()->getUnderlyingValue());
762   }
763 
764   /// Method to support type inquiry through isa, cast, and dyn_cast.
765   static inline bool classof(const VPDef *D) {
766     // All VPDefs are also VPRecipeBases.
767     return true;
768   }
769 
770   static inline bool classof(const VPUser *U) {
771     return U->getVPUserID() == VPUser::VPUserID::Recipe;
772   }
773 
774   /// Returns true if the recipe may have side-effects.
775   bool mayHaveSideEffects() const;
776 
777   /// Returns true for PHI-like recipes.
778   bool isPhi() const {
779     return getVPDefID() >= VPFirstPHISC && getVPDefID() <= VPLastPHISC;
780   }
781 
782   /// Returns true if the recipe may read from memory.
783   bool mayReadFromMemory() const;
784 
785   /// Returns true if the recipe may write to memory.
786   bool mayWriteToMemory() const;
787 
788   /// Returns true if the recipe may read from or write to memory.
789   bool mayReadOrWriteMemory() const {
790     return mayReadFromMemory() || mayWriteToMemory();
791   }
792 };
793 
794 // Helper macro to define common classof implementations for recipes.
795 #define VP_CLASSOF_IMPL(VPDefID)                                               \
796   static inline bool classof(const VPDef *D) {                                 \
797     return D->getVPDefID() == VPDefID;                                         \
798   }                                                                            \
799   static inline bool classof(const VPValue *V) {                               \
800     auto *R = V->getDefiningRecipe();                                          \
801     return R && R->getVPDefID() == VPDefID;                                    \
802   }                                                                            \
803   static inline bool classof(const VPUser *U) {                                \
804     auto *R = dyn_cast<VPRecipeBase>(U);                                       \
805     return R && R->getVPDefID() == VPDefID;                                    \
806   }                                                                            \
807   static inline bool classof(const VPRecipeBase *R) {                          \
808     return R->getVPDefID() == VPDefID;                                         \
809   }
810 
811 /// This is a concrete Recipe that models a single VPlan-level instruction.
812 /// While as any Recipe it may generate a sequence of IR instructions when
813 /// executed, these instructions would always form a single-def expression as
814 /// the VPInstruction is also a single def-use vertex.
815 class VPInstruction : public VPRecipeBase, public VPValue {
816   friend class VPlanSlp;
817 
818 public:
819   /// VPlan opcodes, extending LLVM IR with idiomatics instructions.
820   enum {
821     FirstOrderRecurrenceSplice =
822         Instruction::OtherOpsEnd + 1, // Combines the incoming and previous
823                                       // values of a first-order recurrence.
824     Not,
825     ICmpULE,
826     SLPLoad,
827     SLPStore,
828     ActiveLaneMask,
829     CalculateTripCountMinusVF,
830     CanonicalIVIncrement,
831     CanonicalIVIncrementNUW,
832     // The next two are similar to the above, but instead increment the
833     // canonical IV separately for each unrolled part.
834     CanonicalIVIncrementForPart,
835     CanonicalIVIncrementForPartNUW,
836     BranchOnCount,
837     BranchOnCond
838   };
839 
840 private:
841   typedef unsigned char OpcodeTy;
842   OpcodeTy Opcode;
843   FastMathFlags FMF;
844   DebugLoc DL;
845 
846   /// An optional name that can be used for the generated IR instruction.
847   const std::string Name;
848 
849   /// Utility method serving execute(): generates a single instance of the
850   /// modeled instruction. \returns the generated value for \p Part.
851   /// In some cases an existing value is returned rather than a generated
852   /// one.
853   Value *generateInstruction(VPTransformState &State, unsigned Part);
854 
855 protected:
856   void setUnderlyingInstr(Instruction *I) { setUnderlyingValue(I); }
857 
858 public:
859   VPInstruction(unsigned Opcode, ArrayRef<VPValue *> Operands, DebugLoc DL,
860                 const Twine &Name = "")
861       : VPRecipeBase(VPDef::VPInstructionSC, Operands), VPValue(this),
862         Opcode(Opcode), DL(DL), Name(Name.str()) {}
863 
864   VPInstruction(unsigned Opcode, std::initializer_list<VPValue *> Operands,
865                 DebugLoc DL = {}, const Twine &Name = "")
866       : VPInstruction(Opcode, ArrayRef<VPValue *>(Operands), DL, Name) {}
867 
868   VP_CLASSOF_IMPL(VPDef::VPInstructionSC)
869 
870   VPInstruction *clone() const {
871     SmallVector<VPValue *, 2> Operands(operands());
872     return new VPInstruction(Opcode, Operands, DL, Name);
873   }
874 
875   unsigned getOpcode() const { return Opcode; }
876 
877   /// Generate the instruction.
878   /// TODO: We currently execute only per-part unless a specific instance is
879   /// provided.
880   void execute(VPTransformState &State) override;
881 
882 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
883   /// Print the VPInstruction to \p O.
884   void print(raw_ostream &O, const Twine &Indent,
885              VPSlotTracker &SlotTracker) const override;
886 
887   /// Print the VPInstruction to dbgs() (for debugging).
888   LLVM_DUMP_METHOD void dump() const;
889 #endif
890 
891   /// Return true if this instruction may modify memory.
892   bool mayWriteToMemory() const {
893     // TODO: we can use attributes of the called function to rule out memory
894     //       modifications.
895     return Opcode == Instruction::Store || Opcode == Instruction::Call ||
896            Opcode == Instruction::Invoke || Opcode == SLPStore;
897   }
898 
899   bool hasResult() const {
900     // CallInst may or may not have a result, depending on the called function.
901     // Conservatively return calls have results for now.
902     switch (getOpcode()) {
903     case Instruction::Ret:
904     case Instruction::Br:
905     case Instruction::Store:
906     case Instruction::Switch:
907     case Instruction::IndirectBr:
908     case Instruction::Resume:
909     case Instruction::CatchRet:
910     case Instruction::Unreachable:
911     case Instruction::Fence:
912     case Instruction::AtomicRMW:
913     case VPInstruction::BranchOnCond:
914     case VPInstruction::BranchOnCount:
915       return false;
916     default:
917       return true;
918     }
919   }
920 
921   /// Set the fast-math flags.
922   void setFastMathFlags(FastMathFlags FMFNew);
923 
924   /// Returns true if the recipe only uses the first lane of operand \p Op.
925   bool onlyFirstLaneUsed(const VPValue *Op) const override {
926     assert(is_contained(operands(), Op) &&
927            "Op must be an operand of the recipe");
928     if (getOperand(0) != Op)
929       return false;
930     switch (getOpcode()) {
931     default:
932       return false;
933     case VPInstruction::ActiveLaneMask:
934     case VPInstruction::CalculateTripCountMinusVF:
935     case VPInstruction::CanonicalIVIncrement:
936     case VPInstruction::CanonicalIVIncrementNUW:
937     case VPInstruction::CanonicalIVIncrementForPart:
938     case VPInstruction::CanonicalIVIncrementForPartNUW:
939     case VPInstruction::BranchOnCount:
940       return true;
941     };
942     llvm_unreachable("switch should return");
943   }
944 };
945 
946 /// Class to record LLVM IR flag for a recipe along with it.
947 class VPRecipeWithIRFlags : public VPRecipeBase {
948   enum class OperationType : unsigned char {
949     OverflowingBinOp,
950     PossiblyExactOp,
951     GEPOp,
952     FPMathOp,
953     Other
954   };
955   struct WrapFlagsTy {
956     char HasNUW : 1;
957     char HasNSW : 1;
958   };
959   struct ExactFlagsTy {
960     char IsExact : 1;
961   };
962   struct GEPFlagsTy {
963     char IsInBounds : 1;
964   };
965   struct FastMathFlagsTy {
966     char AllowReassoc : 1;
967     char NoNaNs : 1;
968     char NoInfs : 1;
969     char NoSignedZeros : 1;
970     char AllowReciprocal : 1;
971     char AllowContract : 1;
972     char ApproxFunc : 1;
973   };
974 
975   OperationType OpType;
976 
977   union {
978     WrapFlagsTy WrapFlags;
979     ExactFlagsTy ExactFlags;
980     GEPFlagsTy GEPFlags;
981     FastMathFlagsTy FMFs;
982     unsigned char AllFlags;
983   };
984 
985 public:
986   template <typename IterT>
987   VPRecipeWithIRFlags(const unsigned char SC, iterator_range<IterT> Operands)
988       : VPRecipeBase(SC, Operands) {
989     OpType = OperationType::Other;
990     AllFlags = 0;
991   }
992 
993   template <typename IterT>
994   VPRecipeWithIRFlags(const unsigned char SC, iterator_range<IterT> Operands,
995                       Instruction &I)
996       : VPRecipeWithIRFlags(SC, Operands) {
997     if (auto *Op = dyn_cast<OverflowingBinaryOperator>(&I)) {
998       OpType = OperationType::OverflowingBinOp;
999       WrapFlags.HasNUW = Op->hasNoUnsignedWrap();
1000       WrapFlags.HasNSW = Op->hasNoSignedWrap();
1001     } else if (auto *Op = dyn_cast<PossiblyExactOperator>(&I)) {
1002       OpType = OperationType::PossiblyExactOp;
1003       ExactFlags.IsExact = Op->isExact();
1004     } else if (auto *GEP = dyn_cast<GetElementPtrInst>(&I)) {
1005       OpType = OperationType::GEPOp;
1006       GEPFlags.IsInBounds = GEP->isInBounds();
1007     } else if (auto *Op = dyn_cast<FPMathOperator>(&I)) {
1008       OpType = OperationType::FPMathOp;
1009       FastMathFlags FMF = Op->getFastMathFlags();
1010       FMFs.AllowReassoc = FMF.allowReassoc();
1011       FMFs.NoNaNs = FMF.noNaNs();
1012       FMFs.NoInfs = FMF.noInfs();
1013       FMFs.NoSignedZeros = FMF.noSignedZeros();
1014       FMFs.AllowReciprocal = FMF.allowReciprocal();
1015       FMFs.AllowContract = FMF.allowContract();
1016       FMFs.ApproxFunc = FMF.approxFunc();
1017     }
1018   }
1019 
1020   static inline bool classof(const VPRecipeBase *R) {
1021     return R->getVPDefID() == VPRecipeBase::VPWidenSC ||
1022            R->getVPDefID() == VPRecipeBase::VPWidenGEPSC ||
1023            R->getVPDefID() == VPRecipeBase::VPReplicateSC;
1024   }
1025 
1026   /// Drop all poison-generating flags.
1027   void dropPoisonGeneratingFlags() {
1028     // NOTE: This needs to be kept in-sync with
1029     // Instruction::dropPoisonGeneratingFlags.
1030     switch (OpType) {
1031     case OperationType::OverflowingBinOp:
1032       WrapFlags.HasNUW = false;
1033       WrapFlags.HasNSW = false;
1034       break;
1035     case OperationType::PossiblyExactOp:
1036       ExactFlags.IsExact = false;
1037       break;
1038     case OperationType::GEPOp:
1039       GEPFlags.IsInBounds = false;
1040       break;
1041     case OperationType::FPMathOp:
1042       FMFs.NoNaNs = false;
1043       FMFs.NoInfs = false;
1044       break;
1045     case OperationType::Other:
1046       break;
1047     }
1048   }
1049 
1050   /// Set the IR flags for \p I.
1051   void setFlags(Instruction *I) const {
1052     switch (OpType) {
1053     case OperationType::OverflowingBinOp:
1054       I->setHasNoUnsignedWrap(WrapFlags.HasNUW);
1055       I->setHasNoSignedWrap(WrapFlags.HasNSW);
1056       break;
1057     case OperationType::PossiblyExactOp:
1058       I->setIsExact(ExactFlags.IsExact);
1059       break;
1060     case OperationType::GEPOp:
1061       cast<GetElementPtrInst>(I)->setIsInBounds(GEPFlags.IsInBounds);
1062       break;
1063     case OperationType::FPMathOp:
1064       I->setHasAllowReassoc(FMFs.AllowReassoc);
1065       I->setHasNoNaNs(FMFs.NoNaNs);
1066       I->setHasNoInfs(FMFs.NoInfs);
1067       I->setHasNoSignedZeros(FMFs.NoSignedZeros);
1068       I->setHasAllowReciprocal(FMFs.AllowReciprocal);
1069       I->setHasAllowContract(FMFs.AllowContract);
1070       I->setHasApproxFunc(FMFs.ApproxFunc);
1071       break;
1072     case OperationType::Other:
1073       break;
1074     }
1075   }
1076 
1077   bool isInBounds() const {
1078     assert(OpType == OperationType::GEPOp &&
1079            "recipe doesn't have inbounds flag");
1080     return GEPFlags.IsInBounds;
1081   }
1082 
1083 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1084   FastMathFlags getFastMathFlags() const {
1085     FastMathFlags Res;
1086     Res.setAllowReassoc(FMFs.AllowReassoc);
1087     Res.setNoNaNs(FMFs.NoNaNs);
1088     Res.setNoInfs(FMFs.NoInfs);
1089     Res.setNoSignedZeros(FMFs.NoSignedZeros);
1090     Res.setAllowReciprocal(FMFs.AllowReciprocal);
1091     Res.setAllowContract(FMFs.AllowContract);
1092     Res.setApproxFunc(FMFs.ApproxFunc);
1093     return Res;
1094   }
1095 
1096   void printFlags(raw_ostream &O) const;
1097 #endif
1098 };
1099 
1100 /// VPWidenRecipe is a recipe for producing a copy of vector type its
1101 /// ingredient. This recipe covers most of the traditional vectorization cases
1102 /// where each ingredient transforms into a vectorized version of itself.
1103 class VPWidenRecipe : public VPRecipeWithIRFlags, public VPValue {
1104 
1105 public:
1106   template <typename IterT>
1107   VPWidenRecipe(Instruction &I, iterator_range<IterT> Operands)
1108       : VPRecipeWithIRFlags(VPDef::VPWidenSC, Operands, I), VPValue(this, &I) {}
1109 
1110   ~VPWidenRecipe() override = default;
1111 
1112   VP_CLASSOF_IMPL(VPDef::VPWidenSC)
1113 
1114   /// Produce widened copies of all Ingredients.
1115   void execute(VPTransformState &State) override;
1116 
1117 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1118   /// Print the recipe.
1119   void print(raw_ostream &O, const Twine &Indent,
1120              VPSlotTracker &SlotTracker) const override;
1121 #endif
1122 };
1123 
1124 /// VPWidenCastRecipe is a recipe to create vector cast instructions.
1125 class VPWidenCastRecipe : public VPRecipeBase, public VPValue {
1126   /// Cast instruction opcode.
1127   Instruction::CastOps Opcode;
1128 
1129   /// Result type for the cast.
1130   Type *ResultTy;
1131 
1132 public:
1133   VPWidenCastRecipe(Instruction::CastOps Opcode, VPValue *Op, Type *ResultTy,
1134                     CastInst *UI = nullptr)
1135       : VPRecipeBase(VPDef::VPWidenCastSC, Op), VPValue(this, UI),
1136         Opcode(Opcode), ResultTy(ResultTy) {
1137     assert((!UI || UI->getOpcode() == Opcode) &&
1138            "opcode of underlying cast doesn't match");
1139     assert((!UI || UI->getType() == ResultTy) &&
1140            "result type of underlying cast doesn't match");
1141   }
1142 
1143   ~VPWidenCastRecipe() override = default;
1144 
1145   VP_CLASSOF_IMPL(VPDef::VPWidenCastSC)
1146 
1147   /// Produce widened copies of the cast.
1148   void execute(VPTransformState &State) override;
1149 
1150 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1151   /// Print the recipe.
1152   void print(raw_ostream &O, const Twine &Indent,
1153              VPSlotTracker &SlotTracker) const override;
1154 #endif
1155 
1156   Instruction::CastOps getOpcode() const { return Opcode; }
1157 
1158   /// Returns the result type of the cast.
1159   Type *getResultType() const { return ResultTy; }
1160 };
1161 
1162 /// A recipe for widening Call instructions.
1163 class VPWidenCallRecipe : public VPRecipeBase, public VPValue {
1164   /// ID of the vector intrinsic to call when widening the call. If set the
1165   /// Intrinsic::not_intrinsic, a library call will be used instead.
1166   Intrinsic::ID VectorIntrinsicID;
1167   /// If this recipe represents a library call, Variant stores a pointer to
1168   /// the chosen function. There is a 1:1 mapping between a given VF and the
1169   /// chosen vectorized variant, so there will be a different vplan for each
1170   /// VF with a valid variant.
1171   Function *Variant;
1172 
1173 public:
1174   template <typename IterT>
1175   VPWidenCallRecipe(CallInst &I, iterator_range<IterT> CallArguments,
1176                     Intrinsic::ID VectorIntrinsicID,
1177                     Function *Variant = nullptr)
1178       : VPRecipeBase(VPDef::VPWidenCallSC, CallArguments), VPValue(this, &I),
1179         VectorIntrinsicID(VectorIntrinsicID), Variant(Variant) {}
1180 
1181   ~VPWidenCallRecipe() override = default;
1182 
1183   VP_CLASSOF_IMPL(VPDef::VPWidenCallSC)
1184 
1185   /// Produce a widened version of the call instruction.
1186   void execute(VPTransformState &State) override;
1187 
1188 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1189   /// Print the recipe.
1190   void print(raw_ostream &O, const Twine &Indent,
1191              VPSlotTracker &SlotTracker) const override;
1192 #endif
1193 };
1194 
1195 /// A recipe for widening select instructions.
1196 struct VPWidenSelectRecipe : public VPRecipeBase, public VPValue {
1197   template <typename IterT>
1198   VPWidenSelectRecipe(SelectInst &I, iterator_range<IterT> Operands)
1199       : VPRecipeBase(VPDef::VPWidenSelectSC, Operands), VPValue(this, &I) {}
1200 
1201   ~VPWidenSelectRecipe() override = default;
1202 
1203   VP_CLASSOF_IMPL(VPDef::VPWidenSelectSC)
1204 
1205   /// Produce a widened version of the select instruction.
1206   void execute(VPTransformState &State) override;
1207 
1208 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1209   /// Print the recipe.
1210   void print(raw_ostream &O, const Twine &Indent,
1211              VPSlotTracker &SlotTracker) const override;
1212 #endif
1213 
1214   VPValue *getCond() const {
1215     return getOperand(0);
1216   }
1217 
1218   bool isInvariantCond() const {
1219     return getCond()->isDefinedOutsideVectorRegions();
1220   }
1221 };
1222 
1223 /// A recipe for handling GEP instructions.
1224 class VPWidenGEPRecipe : public VPRecipeWithIRFlags, public VPValue {
1225   bool isPointerLoopInvariant() const {
1226     return getOperand(0)->isDefinedOutsideVectorRegions();
1227   }
1228 
1229   bool isIndexLoopInvariant(unsigned I) const {
1230     return getOperand(I + 1)->isDefinedOutsideVectorRegions();
1231   }
1232 
1233   bool areAllOperandsInvariant() const {
1234     return all_of(operands(), [](VPValue *Op) {
1235       return Op->isDefinedOutsideVectorRegions();
1236     });
1237   }
1238 
1239 public:
1240   template <typename IterT>
1241   VPWidenGEPRecipe(GetElementPtrInst *GEP, iterator_range<IterT> Operands)
1242       : VPRecipeWithIRFlags(VPDef::VPWidenGEPSC, Operands, *GEP),
1243         VPValue(this, GEP) {}
1244 
1245   ~VPWidenGEPRecipe() override = default;
1246 
1247   VP_CLASSOF_IMPL(VPDef::VPWidenGEPSC)
1248 
1249   /// Generate the gep nodes.
1250   void execute(VPTransformState &State) override;
1251 
1252 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1253   /// Print the recipe.
1254   void print(raw_ostream &O, const Twine &Indent,
1255              VPSlotTracker &SlotTracker) const override;
1256 #endif
1257 };
1258 
1259 /// A pure virtual base class for all recipes modeling header phis, including
1260 /// phis for first order recurrences, pointer inductions and reductions. The
1261 /// start value is the first operand of the recipe and the incoming value from
1262 /// the backedge is the second operand.
1263 ///
1264 /// Inductions are modeled using the following sub-classes:
1265 ///  * VPCanonicalIVPHIRecipe: Canonical scalar induction of the vector loop,
1266 ///    starting at a specified value (zero for the main vector loop, the resume
1267 ///    value for the epilogue vector loop) and stepping by 1. The induction
1268 ///    controls exiting of the vector loop by comparing against the vector trip
1269 ///    count. Produces a single scalar PHI for the induction value per
1270 ///    iteration.
1271 ///  * VPWidenIntOrFpInductionRecipe: Generates vector values for integer and
1272 ///    floating point inductions with arbitrary start and step values. Produces
1273 ///    a vector PHI per-part.
1274 ///  * VPDerivedIVRecipe: Converts the canonical IV value to the corresponding
1275 ///    value of an IV with different start and step values. Produces a single
1276 ///    scalar value per iteration
1277 ///  * VPScalarIVStepsRecipe: Generates scalar values per-lane based on a
1278 ///    canonical or derived induction.
1279 ///  * VPWidenPointerInductionRecipe: Generate vector and scalar values for a
1280 ///    pointer induction. Produces either a vector PHI per-part or scalar values
1281 ///    per-lane based on the canonical induction.
1282 class VPHeaderPHIRecipe : public VPRecipeBase, public VPValue {
1283 protected:
1284   VPHeaderPHIRecipe(unsigned char VPDefID, Instruction *UnderlyingInstr,
1285                     VPValue *Start = nullptr)
1286       : VPRecipeBase(VPDefID, {}), VPValue(this, UnderlyingInstr) {
1287     if (Start)
1288       addOperand(Start);
1289   }
1290 
1291 public:
1292   ~VPHeaderPHIRecipe() override = default;
1293 
1294   /// Method to support type inquiry through isa, cast, and dyn_cast.
1295   static inline bool classof(const VPRecipeBase *B) {
1296     return B->getVPDefID() >= VPDef::VPFirstHeaderPHISC &&
1297            B->getVPDefID() <= VPDef::VPLastHeaderPHISC;
1298   }
1299   static inline bool classof(const VPValue *V) {
1300     auto *B = V->getDefiningRecipe();
1301     return B && B->getVPDefID() >= VPRecipeBase::VPFirstHeaderPHISC &&
1302            B->getVPDefID() <= VPRecipeBase::VPLastHeaderPHISC;
1303   }
1304 
1305   /// Generate the phi nodes.
1306   void execute(VPTransformState &State) override = 0;
1307 
1308 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1309   /// Print the recipe.
1310   void print(raw_ostream &O, const Twine &Indent,
1311              VPSlotTracker &SlotTracker) const override = 0;
1312 #endif
1313 
1314   /// Returns the start value of the phi, if one is set.
1315   VPValue *getStartValue() {
1316     return getNumOperands() == 0 ? nullptr : getOperand(0);
1317   }
1318   VPValue *getStartValue() const {
1319     return getNumOperands() == 0 ? nullptr : getOperand(0);
1320   }
1321 
1322   /// Update the start value of the recipe.
1323   void setStartValue(VPValue *V) { setOperand(0, V); }
1324 
1325   /// Returns the incoming value from the loop backedge.
1326   virtual VPValue *getBackedgeValue() {
1327     return getOperand(1);
1328   }
1329 
1330   /// Returns the backedge value as a recipe. The backedge value is guaranteed
1331   /// to be a recipe.
1332   virtual VPRecipeBase &getBackedgeRecipe() {
1333     return *getBackedgeValue()->getDefiningRecipe();
1334   }
1335 };
1336 
1337 /// A recipe for handling phi nodes of integer and floating-point inductions,
1338 /// producing their vector values.
1339 class VPWidenIntOrFpInductionRecipe : public VPHeaderPHIRecipe {
1340   PHINode *IV;
1341   TruncInst *Trunc;
1342   const InductionDescriptor &IndDesc;
1343 
1344 public:
1345   VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, VPValue *Step,
1346                                 const InductionDescriptor &IndDesc)
1347       : VPHeaderPHIRecipe(VPDef::VPWidenIntOrFpInductionSC, IV, Start), IV(IV),
1348         Trunc(nullptr), IndDesc(IndDesc) {
1349     addOperand(Step);
1350   }
1351 
1352   VPWidenIntOrFpInductionRecipe(PHINode *IV, VPValue *Start, VPValue *Step,
1353                                 const InductionDescriptor &IndDesc,
1354                                 TruncInst *Trunc)
1355       : VPHeaderPHIRecipe(VPDef::VPWidenIntOrFpInductionSC, Trunc, Start),
1356         IV(IV), Trunc(Trunc), IndDesc(IndDesc) {
1357     addOperand(Step);
1358   }
1359 
1360   ~VPWidenIntOrFpInductionRecipe() override = default;
1361 
1362   VP_CLASSOF_IMPL(VPDef::VPWidenIntOrFpInductionSC)
1363 
1364   /// Generate the vectorized and scalarized versions of the phi node as
1365   /// needed by their users.
1366   void execute(VPTransformState &State) override;
1367 
1368 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1369   /// Print the recipe.
1370   void print(raw_ostream &O, const Twine &Indent,
1371              VPSlotTracker &SlotTracker) const override;
1372 #endif
1373 
1374   VPValue *getBackedgeValue() override {
1375     // TODO: All operands of base recipe must exist and be at same index in
1376     // derived recipe.
1377     llvm_unreachable(
1378         "VPWidenIntOrFpInductionRecipe generates its own backedge value");
1379   }
1380 
1381   VPRecipeBase &getBackedgeRecipe() override {
1382     // TODO: All operands of base recipe must exist and be at same index in
1383     // derived recipe.
1384     llvm_unreachable(
1385         "VPWidenIntOrFpInductionRecipe generates its own backedge value");
1386   }
1387 
1388   /// Returns the step value of the induction.
1389   VPValue *getStepValue() { return getOperand(1); }
1390   const VPValue *getStepValue() const { return getOperand(1); }
1391 
1392   /// Returns the first defined value as TruncInst, if it is one or nullptr
1393   /// otherwise.
1394   TruncInst *getTruncInst() { return Trunc; }
1395   const TruncInst *getTruncInst() const { return Trunc; }
1396 
1397   PHINode *getPHINode() { return IV; }
1398 
1399   /// Returns the induction descriptor for the recipe.
1400   const InductionDescriptor &getInductionDescriptor() const { return IndDesc; }
1401 
1402   /// Returns true if the induction is canonical, i.e. starting at 0 and
1403   /// incremented by UF * VF (= the original IV is incremented by 1).
1404   bool isCanonical() const;
1405 
1406   /// Returns the scalar type of the induction.
1407   const Type *getScalarType() const {
1408     return Trunc ? Trunc->getType() : IV->getType();
1409   }
1410 };
1411 
1412 class VPWidenPointerInductionRecipe : public VPHeaderPHIRecipe {
1413   const InductionDescriptor &IndDesc;
1414 
1415   bool IsScalarAfterVectorization;
1416 
1417 public:
1418   /// Create a new VPWidenPointerInductionRecipe for \p Phi with start value \p
1419   /// Start.
1420   VPWidenPointerInductionRecipe(PHINode *Phi, VPValue *Start, VPValue *Step,
1421                                 const InductionDescriptor &IndDesc,
1422                                 bool IsScalarAfterVectorization)
1423       : VPHeaderPHIRecipe(VPDef::VPWidenPointerInductionSC, Phi),
1424         IndDesc(IndDesc),
1425         IsScalarAfterVectorization(IsScalarAfterVectorization) {
1426     addOperand(Start);
1427     addOperand(Step);
1428   }
1429 
1430   ~VPWidenPointerInductionRecipe() override = default;
1431 
1432   VP_CLASSOF_IMPL(VPDef::VPWidenPointerInductionSC)
1433 
1434   /// Generate vector values for the pointer induction.
1435   void execute(VPTransformState &State) override;
1436 
1437   /// Returns true if only scalar values will be generated.
1438   bool onlyScalarsGenerated(ElementCount VF);
1439 
1440   /// Returns the induction descriptor for the recipe.
1441   const InductionDescriptor &getInductionDescriptor() const { return IndDesc; }
1442 
1443 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1444   /// Print the recipe.
1445   void print(raw_ostream &O, const Twine &Indent,
1446              VPSlotTracker &SlotTracker) const override;
1447 #endif
1448 };
1449 
1450 /// A recipe for handling header phis that are widened in the vector loop.
1451 /// In the VPlan native path, all incoming VPValues & VPBasicBlock pairs are
1452 /// managed in the recipe directly.
1453 class VPWidenPHIRecipe : public VPHeaderPHIRecipe {
1454   /// List of incoming blocks. Only used in the VPlan native path.
1455   SmallVector<VPBasicBlock *, 2> IncomingBlocks;
1456 
1457 public:
1458   /// Create a new VPWidenPHIRecipe for \p Phi with start value \p Start.
1459   VPWidenPHIRecipe(PHINode *Phi, VPValue *Start = nullptr)
1460       : VPHeaderPHIRecipe(VPDef::VPWidenPHISC, Phi) {
1461     if (Start)
1462       addOperand(Start);
1463   }
1464 
1465   ~VPWidenPHIRecipe() override = default;
1466 
1467   VP_CLASSOF_IMPL(VPDef::VPWidenPHISC)
1468 
1469   /// Generate the phi/select nodes.
1470   void execute(VPTransformState &State) override;
1471 
1472 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1473   /// Print the recipe.
1474   void print(raw_ostream &O, const Twine &Indent,
1475              VPSlotTracker &SlotTracker) const override;
1476 #endif
1477 
1478   /// Adds a pair (\p IncomingV, \p IncomingBlock) to the phi.
1479   void addIncoming(VPValue *IncomingV, VPBasicBlock *IncomingBlock) {
1480     addOperand(IncomingV);
1481     IncomingBlocks.push_back(IncomingBlock);
1482   }
1483 
1484   /// Returns the \p I th incoming VPBasicBlock.
1485   VPBasicBlock *getIncomingBlock(unsigned I) { return IncomingBlocks[I]; }
1486 
1487   /// Returns the \p I th incoming VPValue.
1488   VPValue *getIncomingValue(unsigned I) { return getOperand(I); }
1489 };
1490 
1491 /// A recipe for handling first-order recurrence phis. The start value is the
1492 /// first operand of the recipe and the incoming value from the backedge is the
1493 /// second operand.
1494 struct VPFirstOrderRecurrencePHIRecipe : public VPHeaderPHIRecipe {
1495   VPFirstOrderRecurrencePHIRecipe(PHINode *Phi, VPValue &Start)
1496       : VPHeaderPHIRecipe(VPDef::VPFirstOrderRecurrencePHISC, Phi, &Start) {}
1497 
1498   VP_CLASSOF_IMPL(VPDef::VPFirstOrderRecurrencePHISC)
1499 
1500   static inline bool classof(const VPHeaderPHIRecipe *R) {
1501     return R->getVPDefID() == VPDef::VPFirstOrderRecurrencePHISC;
1502   }
1503 
1504   void execute(VPTransformState &State) override;
1505 
1506 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1507   /// Print the recipe.
1508   void print(raw_ostream &O, const Twine &Indent,
1509              VPSlotTracker &SlotTracker) const override;
1510 #endif
1511 };
1512 
1513 /// A recipe for handling reduction phis. The start value is the first operand
1514 /// of the recipe and the incoming value from the backedge is the second
1515 /// operand.
1516 class VPReductionPHIRecipe : public VPHeaderPHIRecipe {
1517   /// Descriptor for the reduction.
1518   const RecurrenceDescriptor &RdxDesc;
1519 
1520   /// The phi is part of an in-loop reduction.
1521   bool IsInLoop;
1522 
1523   /// The phi is part of an ordered reduction. Requires IsInLoop to be true.
1524   bool IsOrdered;
1525 
1526 public:
1527   /// Create a new VPReductionPHIRecipe for the reduction \p Phi described by \p
1528   /// RdxDesc.
1529   VPReductionPHIRecipe(PHINode *Phi, const RecurrenceDescriptor &RdxDesc,
1530                        VPValue &Start, bool IsInLoop = false,
1531                        bool IsOrdered = false)
1532       : VPHeaderPHIRecipe(VPDef::VPReductionPHISC, Phi, &Start),
1533         RdxDesc(RdxDesc), IsInLoop(IsInLoop), IsOrdered(IsOrdered) {
1534     assert((!IsOrdered || IsInLoop) && "IsOrdered requires IsInLoop");
1535   }
1536 
1537   ~VPReductionPHIRecipe() override = default;
1538 
1539   VP_CLASSOF_IMPL(VPDef::VPReductionPHISC)
1540 
1541   static inline bool classof(const VPHeaderPHIRecipe *R) {
1542     return R->getVPDefID() == VPDef::VPReductionPHISC;
1543   }
1544 
1545   /// Generate the phi/select nodes.
1546   void execute(VPTransformState &State) override;
1547 
1548 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1549   /// Print the recipe.
1550   void print(raw_ostream &O, const Twine &Indent,
1551              VPSlotTracker &SlotTracker) const override;
1552 #endif
1553 
1554   const RecurrenceDescriptor &getRecurrenceDescriptor() const {
1555     return RdxDesc;
1556   }
1557 
1558   /// Returns true, if the phi is part of an ordered reduction.
1559   bool isOrdered() const { return IsOrdered; }
1560 
1561   /// Returns true, if the phi is part of an in-loop reduction.
1562   bool isInLoop() const { return IsInLoop; }
1563 };
1564 
1565 /// A recipe for vectorizing a phi-node as a sequence of mask-based select
1566 /// instructions.
1567 class VPBlendRecipe : public VPRecipeBase, public VPValue {
1568   PHINode *Phi;
1569 
1570 public:
1571   /// The blend operation is a User of the incoming values and of their
1572   /// respective masks, ordered [I0, M0, I1, M1, ...]. Note that a single value
1573   /// might be incoming with a full mask for which there is no VPValue.
1574   VPBlendRecipe(PHINode *Phi, ArrayRef<VPValue *> Operands)
1575       : VPRecipeBase(VPDef::VPBlendSC, Operands), VPValue(this, Phi), Phi(Phi) {
1576     assert(Operands.size() > 0 &&
1577            ((Operands.size() == 1) || (Operands.size() % 2 == 0)) &&
1578            "Expected either a single incoming value or a positive even number "
1579            "of operands");
1580   }
1581 
1582   VP_CLASSOF_IMPL(VPDef::VPBlendSC)
1583 
1584   /// Return the number of incoming values, taking into account that a single
1585   /// incoming value has no mask.
1586   unsigned getNumIncomingValues() const { return (getNumOperands() + 1) / 2; }
1587 
1588   /// Return incoming value number \p Idx.
1589   VPValue *getIncomingValue(unsigned Idx) const { return getOperand(Idx * 2); }
1590 
1591   /// Return mask number \p Idx.
1592   VPValue *getMask(unsigned Idx) const { return getOperand(Idx * 2 + 1); }
1593 
1594   /// Generate the phi/select nodes.
1595   void execute(VPTransformState &State) override;
1596 
1597 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1598   /// Print the recipe.
1599   void print(raw_ostream &O, const Twine &Indent,
1600              VPSlotTracker &SlotTracker) const override;
1601 #endif
1602 
1603   /// Returns true if the recipe only uses the first lane of operand \p Op.
1604   bool onlyFirstLaneUsed(const VPValue *Op) const override {
1605     assert(is_contained(operands(), Op) &&
1606            "Op must be an operand of the recipe");
1607     // Recursing through Blend recipes only, must terminate at header phi's the
1608     // latest.
1609     return all_of(users(),
1610                   [this](VPUser *U) { return U->onlyFirstLaneUsed(this); });
1611   }
1612 };
1613 
1614 /// VPInterleaveRecipe is a recipe for transforming an interleave group of load
1615 /// or stores into one wide load/store and shuffles. The first operand of a
1616 /// VPInterleave recipe is the address, followed by the stored values, followed
1617 /// by an optional mask.
1618 class VPInterleaveRecipe : public VPRecipeBase {
1619   const InterleaveGroup<Instruction> *IG;
1620 
1621   /// Indicates if the interleave group is in a conditional block and requires a
1622   /// mask.
1623   bool HasMask = false;
1624 
1625   /// Indicates if gaps between members of the group need to be masked out or if
1626   /// unusued gaps can be loaded speculatively.
1627   bool NeedsMaskForGaps = false;
1628 
1629 public:
1630   VPInterleaveRecipe(const InterleaveGroup<Instruction> *IG, VPValue *Addr,
1631                      ArrayRef<VPValue *> StoredValues, VPValue *Mask,
1632                      bool NeedsMaskForGaps)
1633       : VPRecipeBase(VPDef::VPInterleaveSC, {Addr}), IG(IG),
1634         NeedsMaskForGaps(NeedsMaskForGaps) {
1635     for (unsigned i = 0; i < IG->getFactor(); ++i)
1636       if (Instruction *I = IG->getMember(i)) {
1637         if (I->getType()->isVoidTy())
1638           continue;
1639         new VPValue(I, this);
1640       }
1641 
1642     for (auto *SV : StoredValues)
1643       addOperand(SV);
1644     if (Mask) {
1645       HasMask = true;
1646       addOperand(Mask);
1647     }
1648   }
1649   ~VPInterleaveRecipe() override = default;
1650 
1651   VP_CLASSOF_IMPL(VPDef::VPInterleaveSC)
1652 
1653   /// Return the address accessed by this recipe.
1654   VPValue *getAddr() const {
1655     return getOperand(0); // Address is the 1st, mandatory operand.
1656   }
1657 
1658   /// Return the mask used by this recipe. Note that a full mask is represented
1659   /// by a nullptr.
1660   VPValue *getMask() const {
1661     // Mask is optional and therefore the last, currently 2nd operand.
1662     return HasMask ? getOperand(getNumOperands() - 1) : nullptr;
1663   }
1664 
1665   /// Return the VPValues stored by this interleave group. If it is a load
1666   /// interleave group, return an empty ArrayRef.
1667   ArrayRef<VPValue *> getStoredValues() const {
1668     // The first operand is the address, followed by the stored values, followed
1669     // by an optional mask.
1670     return ArrayRef<VPValue *>(op_begin(), getNumOperands())
1671         .slice(1, getNumStoreOperands());
1672   }
1673 
1674   /// Generate the wide load or store, and shuffles.
1675   void execute(VPTransformState &State) override;
1676 
1677 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1678   /// Print the recipe.
1679   void print(raw_ostream &O, const Twine &Indent,
1680              VPSlotTracker &SlotTracker) const override;
1681 #endif
1682 
1683   const InterleaveGroup<Instruction> *getInterleaveGroup() { return IG; }
1684 
1685   /// Returns the number of stored operands of this interleave group. Returns 0
1686   /// for load interleave groups.
1687   unsigned getNumStoreOperands() const {
1688     return getNumOperands() - (HasMask ? 2 : 1);
1689   }
1690 
1691   /// The recipe only uses the first lane of the address.
1692   bool onlyFirstLaneUsed(const VPValue *Op) const override {
1693     assert(is_contained(operands(), Op) &&
1694            "Op must be an operand of the recipe");
1695     return Op == getAddr() && !llvm::is_contained(getStoredValues(), Op);
1696   }
1697 };
1698 
1699 /// A recipe to represent inloop reduction operations, performing a reduction on
1700 /// a vector operand into a scalar value, and adding the result to a chain.
1701 /// The Operands are {ChainOp, VecOp, [Condition]}.
1702 class VPReductionRecipe : public VPRecipeBase, public VPValue {
1703   /// The recurrence decriptor for the reduction in question.
1704   const RecurrenceDescriptor *RdxDesc;
1705   /// Pointer to the TTI, needed to create the target reduction
1706   const TargetTransformInfo *TTI;
1707 
1708 public:
1709   VPReductionRecipe(const RecurrenceDescriptor *R, Instruction *I,
1710                     VPValue *ChainOp, VPValue *VecOp, VPValue *CondOp,
1711                     const TargetTransformInfo *TTI)
1712       : VPRecipeBase(VPDef::VPReductionSC, {ChainOp, VecOp}), VPValue(this, I),
1713         RdxDesc(R), TTI(TTI) {
1714     if (CondOp)
1715       addOperand(CondOp);
1716   }
1717 
1718   ~VPReductionRecipe() override = default;
1719 
1720   VP_CLASSOF_IMPL(VPDef::VPReductionSC)
1721 
1722   /// Generate the reduction in the loop
1723   void execute(VPTransformState &State) override;
1724 
1725 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1726   /// Print the recipe.
1727   void print(raw_ostream &O, const Twine &Indent,
1728              VPSlotTracker &SlotTracker) const override;
1729 #endif
1730 
1731   /// The VPValue of the scalar Chain being accumulated.
1732   VPValue *getChainOp() const { return getOperand(0); }
1733   /// The VPValue of the vector value to be reduced.
1734   VPValue *getVecOp() const { return getOperand(1); }
1735   /// The VPValue of the condition for the block.
1736   VPValue *getCondOp() const {
1737     return getNumOperands() > 2 ? getOperand(2) : nullptr;
1738   }
1739 };
1740 
1741 /// VPReplicateRecipe replicates a given instruction producing multiple scalar
1742 /// copies of the original scalar type, one per lane, instead of producing a
1743 /// single copy of widened type for all lanes. If the instruction is known to be
1744 /// uniform only one copy, per lane zero, will be generated.
1745 class VPReplicateRecipe : public VPRecipeWithIRFlags, public VPValue {
1746   /// Indicator if only a single replica per lane is needed.
1747   bool IsUniform;
1748 
1749   /// Indicator if the replicas are also predicated.
1750   bool IsPredicated;
1751 
1752 public:
1753   template <typename IterT>
1754   VPReplicateRecipe(Instruction *I, iterator_range<IterT> Operands,
1755                     bool IsUniform, VPValue *Mask = nullptr)
1756       : VPRecipeWithIRFlags(VPDef::VPReplicateSC, Operands, *I),
1757         VPValue(this, I), IsUniform(IsUniform), IsPredicated(Mask) {
1758     if (Mask)
1759       addOperand(Mask);
1760   }
1761 
1762   ~VPReplicateRecipe() override = default;
1763 
1764   VP_CLASSOF_IMPL(VPDef::VPReplicateSC)
1765 
1766   /// Generate replicas of the desired Ingredient. Replicas will be generated
1767   /// for all parts and lanes unless a specific part and lane are specified in
1768   /// the \p State.
1769   void execute(VPTransformState &State) override;
1770 
1771 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1772   /// Print the recipe.
1773   void print(raw_ostream &O, const Twine &Indent,
1774              VPSlotTracker &SlotTracker) const override;
1775 #endif
1776 
1777   bool isUniform() const { return IsUniform; }
1778 
1779   bool isPredicated() const { return IsPredicated; }
1780 
1781   /// Returns true if the recipe only uses the first lane of operand \p Op.
1782   bool onlyFirstLaneUsed(const VPValue *Op) const override {
1783     assert(is_contained(operands(), Op) &&
1784            "Op must be an operand of the recipe");
1785     return isUniform();
1786   }
1787 
1788   /// Returns true if the recipe uses scalars of operand \p Op.
1789   bool usesScalars(const VPValue *Op) const override {
1790     assert(is_contained(operands(), Op) &&
1791            "Op must be an operand of the recipe");
1792     return true;
1793   }
1794 
1795   /// Returns true if the recipe is used by a widened recipe via an intervening
1796   /// VPPredInstPHIRecipe. In this case, the scalar values should also be packed
1797   /// in a vector.
1798   bool shouldPack() const;
1799 
1800   /// Return the mask of a predicated VPReplicateRecipe.
1801   VPValue *getMask() {
1802     assert(isPredicated() && "Trying to get the mask of a unpredicated recipe");
1803     return getOperand(getNumOperands() - 1);
1804   }
1805 };
1806 
1807 /// A recipe for generating conditional branches on the bits of a mask.
1808 class VPBranchOnMaskRecipe : public VPRecipeBase {
1809 public:
1810   VPBranchOnMaskRecipe(VPValue *BlockInMask)
1811       : VPRecipeBase(VPDef::VPBranchOnMaskSC, {}) {
1812     if (BlockInMask) // nullptr means all-one mask.
1813       addOperand(BlockInMask);
1814   }
1815 
1816   VP_CLASSOF_IMPL(VPDef::VPBranchOnMaskSC)
1817 
1818   /// Generate the extraction of the appropriate bit from the block mask and the
1819   /// conditional branch.
1820   void execute(VPTransformState &State) override;
1821 
1822 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1823   /// Print the recipe.
1824   void print(raw_ostream &O, const Twine &Indent,
1825              VPSlotTracker &SlotTracker) const override {
1826     O << Indent << "BRANCH-ON-MASK ";
1827     if (VPValue *Mask = getMask())
1828       Mask->printAsOperand(O, SlotTracker);
1829     else
1830       O << " All-One";
1831   }
1832 #endif
1833 
1834   /// Return the mask used by this recipe. Note that a full mask is represented
1835   /// by a nullptr.
1836   VPValue *getMask() const {
1837     assert(getNumOperands() <= 1 && "should have either 0 or 1 operands");
1838     // Mask is optional.
1839     return getNumOperands() == 1 ? getOperand(0) : nullptr;
1840   }
1841 
1842   /// Returns true if the recipe uses scalars of operand \p Op.
1843   bool usesScalars(const VPValue *Op) const override {
1844     assert(is_contained(operands(), Op) &&
1845            "Op must be an operand of the recipe");
1846     return true;
1847   }
1848 };
1849 
1850 /// VPPredInstPHIRecipe is a recipe for generating the phi nodes needed when
1851 /// control converges back from a Branch-on-Mask. The phi nodes are needed in
1852 /// order to merge values that are set under such a branch and feed their uses.
1853 /// The phi nodes can be scalar or vector depending on the users of the value.
1854 /// This recipe works in concert with VPBranchOnMaskRecipe.
1855 class VPPredInstPHIRecipe : public VPRecipeBase, public VPValue {
1856 public:
1857   /// Construct a VPPredInstPHIRecipe given \p PredInst whose value needs a phi
1858   /// nodes after merging back from a Branch-on-Mask.
1859   VPPredInstPHIRecipe(VPValue *PredV)
1860       : VPRecipeBase(VPDef::VPPredInstPHISC, PredV), VPValue(this) {}
1861   ~VPPredInstPHIRecipe() override = default;
1862 
1863   VP_CLASSOF_IMPL(VPDef::VPPredInstPHISC)
1864 
1865   /// Generates phi nodes for live-outs as needed to retain SSA form.
1866   void execute(VPTransformState &State) override;
1867 
1868 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1869   /// Print the recipe.
1870   void print(raw_ostream &O, const Twine &Indent,
1871              VPSlotTracker &SlotTracker) const override;
1872 #endif
1873 
1874   /// Returns true if the recipe uses scalars of operand \p Op.
1875   bool usesScalars(const VPValue *Op) const override {
1876     assert(is_contained(operands(), Op) &&
1877            "Op must be an operand of the recipe");
1878     return true;
1879   }
1880 };
1881 
1882 /// A Recipe for widening load/store operations.
1883 /// The recipe uses the following VPValues:
1884 /// - For load: Address, optional mask
1885 /// - For store: Address, stored value, optional mask
1886 /// TODO: We currently execute only per-part unless a specific instance is
1887 /// provided.
1888 class VPWidenMemoryInstructionRecipe : public VPRecipeBase {
1889   Instruction &Ingredient;
1890 
1891   // Whether the loaded-from / stored-to addresses are consecutive.
1892   bool Consecutive;
1893 
1894   // Whether the consecutive loaded/stored addresses are in reverse order.
1895   bool Reverse;
1896 
1897   void setMask(VPValue *Mask) {
1898     if (!Mask)
1899       return;
1900     addOperand(Mask);
1901   }
1902 
1903   bool isMasked() const {
1904     return isStore() ? getNumOperands() == 3 : getNumOperands() == 2;
1905   }
1906 
1907 public:
1908   VPWidenMemoryInstructionRecipe(LoadInst &Load, VPValue *Addr, VPValue *Mask,
1909                                  bool Consecutive, bool Reverse)
1910       : VPRecipeBase(VPDef::VPWidenMemoryInstructionSC, {Addr}),
1911         Ingredient(Load), Consecutive(Consecutive), Reverse(Reverse) {
1912     assert((Consecutive || !Reverse) && "Reverse implies consecutive");
1913     new VPValue(this, &Load);
1914     setMask(Mask);
1915   }
1916 
1917   VPWidenMemoryInstructionRecipe(StoreInst &Store, VPValue *Addr,
1918                                  VPValue *StoredValue, VPValue *Mask,
1919                                  bool Consecutive, bool Reverse)
1920       : VPRecipeBase(VPDef::VPWidenMemoryInstructionSC, {Addr, StoredValue}),
1921         Ingredient(Store), Consecutive(Consecutive), Reverse(Reverse) {
1922     assert((Consecutive || !Reverse) && "Reverse implies consecutive");
1923     setMask(Mask);
1924   }
1925 
1926   VP_CLASSOF_IMPL(VPDef::VPWidenMemoryInstructionSC)
1927 
1928   /// Return the address accessed by this recipe.
1929   VPValue *getAddr() const {
1930     return getOperand(0); // Address is the 1st, mandatory operand.
1931   }
1932 
1933   /// Return the mask used by this recipe. Note that a full mask is represented
1934   /// by a nullptr.
1935   VPValue *getMask() const {
1936     // Mask is optional and therefore the last operand.
1937     return isMasked() ? getOperand(getNumOperands() - 1) : nullptr;
1938   }
1939 
1940   /// Returns true if this recipe is a store.
1941   bool isStore() const { return isa<StoreInst>(Ingredient); }
1942 
1943   /// Return the address accessed by this recipe.
1944   VPValue *getStoredValue() const {
1945     assert(isStore() && "Stored value only available for store instructions");
1946     return getOperand(1); // Stored value is the 2nd, mandatory operand.
1947   }
1948 
1949   // Return whether the loaded-from / stored-to addresses are consecutive.
1950   bool isConsecutive() const { return Consecutive; }
1951 
1952   // Return whether the consecutive loaded/stored addresses are in reverse
1953   // order.
1954   bool isReverse() const { return Reverse; }
1955 
1956   /// Generate the wide load/store.
1957   void execute(VPTransformState &State) override;
1958 
1959 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1960   /// Print the recipe.
1961   void print(raw_ostream &O, const Twine &Indent,
1962              VPSlotTracker &SlotTracker) const override;
1963 #endif
1964 
1965   /// Returns true if the recipe only uses the first lane of operand \p Op.
1966   bool onlyFirstLaneUsed(const VPValue *Op) const override {
1967     assert(is_contained(operands(), Op) &&
1968            "Op must be an operand of the recipe");
1969 
1970     // Widened, consecutive memory operations only demand the first lane of
1971     // their address, unless the same operand is also stored. That latter can
1972     // happen with opaque pointers.
1973     return Op == getAddr() && isConsecutive() &&
1974            (!isStore() || Op != getStoredValue());
1975   }
1976 
1977   Instruction &getIngredient() const { return Ingredient; }
1978 };
1979 
1980 /// Recipe to expand a SCEV expression.
1981 class VPExpandSCEVRecipe : public VPRecipeBase, public VPValue {
1982   const SCEV *Expr;
1983   ScalarEvolution &SE;
1984 
1985 public:
1986   VPExpandSCEVRecipe(const SCEV *Expr, ScalarEvolution &SE)
1987       : VPRecipeBase(VPDef::VPExpandSCEVSC, {}), VPValue(this), Expr(Expr),
1988         SE(SE) {}
1989 
1990   ~VPExpandSCEVRecipe() override = default;
1991 
1992   VP_CLASSOF_IMPL(VPDef::VPExpandSCEVSC)
1993 
1994   /// Generate a canonical vector induction variable of the vector loop, with
1995   void execute(VPTransformState &State) override;
1996 
1997 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1998   /// Print the recipe.
1999   void print(raw_ostream &O, const Twine &Indent,
2000              VPSlotTracker &SlotTracker) const override;
2001 #endif
2002 
2003   const SCEV *getSCEV() const { return Expr; }
2004 };
2005 
2006 /// Canonical scalar induction phi of the vector loop. Starting at the specified
2007 /// start value (either 0 or the resume value when vectorizing the epilogue
2008 /// loop). VPWidenCanonicalIVRecipe represents the vector version of the
2009 /// canonical induction variable.
2010 class VPCanonicalIVPHIRecipe : public VPHeaderPHIRecipe {
2011   DebugLoc DL;
2012 
2013 public:
2014   VPCanonicalIVPHIRecipe(VPValue *StartV, DebugLoc DL)
2015       : VPHeaderPHIRecipe(VPDef::VPCanonicalIVPHISC, nullptr, StartV), DL(DL) {}
2016 
2017   ~VPCanonicalIVPHIRecipe() override = default;
2018 
2019   VP_CLASSOF_IMPL(VPDef::VPCanonicalIVPHISC)
2020 
2021   static inline bool classof(const VPHeaderPHIRecipe *D) {
2022     return D->getVPDefID() == VPDef::VPCanonicalIVPHISC;
2023   }
2024 
2025   /// Generate the canonical scalar induction phi of the vector loop.
2026   void execute(VPTransformState &State) override;
2027 
2028 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2029   /// Print the recipe.
2030   void print(raw_ostream &O, const Twine &Indent,
2031              VPSlotTracker &SlotTracker) const override;
2032 #endif
2033 
2034   /// Returns the scalar type of the induction.
2035   const Type *getScalarType() const {
2036     return getOperand(0)->getLiveInIRValue()->getType();
2037   }
2038 
2039   /// Returns true if the recipe only uses the first lane of operand \p Op.
2040   bool onlyFirstLaneUsed(const VPValue *Op) const override {
2041     assert(is_contained(operands(), Op) &&
2042            "Op must be an operand of the recipe");
2043     return true;
2044   }
2045 
2046   /// Check if the induction described by \p Kind, /p Start and \p Step is
2047   /// canonical, i.e.  has the same start, step (of 1), and type as the
2048   /// canonical IV.
2049   bool isCanonical(InductionDescriptor::InductionKind Kind, VPValue *Start,
2050                    VPValue *Step, Type *Ty) const;
2051 };
2052 
2053 /// A recipe for generating the active lane mask for the vector loop that is
2054 /// used to predicate the vector operations.
2055 /// TODO: It would be good to use the existing VPWidenPHIRecipe instead and
2056 /// remove VPActiveLaneMaskPHIRecipe.
2057 class VPActiveLaneMaskPHIRecipe : public VPHeaderPHIRecipe {
2058   DebugLoc DL;
2059 
2060 public:
2061   VPActiveLaneMaskPHIRecipe(VPValue *StartMask, DebugLoc DL)
2062       : VPHeaderPHIRecipe(VPDef::VPActiveLaneMaskPHISC, nullptr, StartMask),
2063         DL(DL) {}
2064 
2065   ~VPActiveLaneMaskPHIRecipe() override = default;
2066 
2067   VP_CLASSOF_IMPL(VPDef::VPActiveLaneMaskPHISC)
2068 
2069   static inline bool classof(const VPHeaderPHIRecipe *D) {
2070     return D->getVPDefID() == VPDef::VPActiveLaneMaskPHISC;
2071   }
2072 
2073   /// Generate the active lane mask phi of the vector loop.
2074   void execute(VPTransformState &State) override;
2075 
2076 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2077   /// Print the recipe.
2078   void print(raw_ostream &O, const Twine &Indent,
2079              VPSlotTracker &SlotTracker) const override;
2080 #endif
2081 };
2082 
2083 /// A Recipe for widening the canonical induction variable of the vector loop.
2084 class VPWidenCanonicalIVRecipe : public VPRecipeBase, public VPValue {
2085 public:
2086   VPWidenCanonicalIVRecipe(VPCanonicalIVPHIRecipe *CanonicalIV)
2087       : VPRecipeBase(VPDef::VPWidenCanonicalIVSC, {CanonicalIV}),
2088         VPValue(this) {}
2089 
2090   ~VPWidenCanonicalIVRecipe() override = default;
2091 
2092   VP_CLASSOF_IMPL(VPDef::VPWidenCanonicalIVSC)
2093 
2094   /// Generate a canonical vector induction variable of the vector loop, with
2095   /// start = {<Part*VF, Part*VF+1, ..., Part*VF+VF-1> for 0 <= Part < UF}, and
2096   /// step = <VF*UF, VF*UF, ..., VF*UF>.
2097   void execute(VPTransformState &State) override;
2098 
2099 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2100   /// Print the recipe.
2101   void print(raw_ostream &O, const Twine &Indent,
2102              VPSlotTracker &SlotTracker) const override;
2103 #endif
2104 
2105   /// Returns the scalar type of the induction.
2106   const Type *getScalarType() const {
2107     return cast<VPCanonicalIVPHIRecipe>(getOperand(0)->getDefiningRecipe())
2108         ->getScalarType();
2109   }
2110 };
2111 
2112 /// A recipe for converting the canonical IV value to the corresponding value of
2113 /// an IV with different start and step values, using Start + CanonicalIV *
2114 /// Step.
2115 class VPDerivedIVRecipe : public VPRecipeBase, public VPValue {
2116   /// The type of the result value. It may be smaller than the type of the
2117   /// induction and in this case it will get truncated to ResultTy.
2118   Type *ResultTy;
2119 
2120   /// Induction descriptor for the induction the canonical IV is transformed to.
2121   const InductionDescriptor &IndDesc;
2122 
2123 public:
2124   VPDerivedIVRecipe(const InductionDescriptor &IndDesc, VPValue *Start,
2125                     VPCanonicalIVPHIRecipe *CanonicalIV, VPValue *Step,
2126                     Type *ResultTy)
2127       : VPRecipeBase(VPDef::VPDerivedIVSC, {Start, CanonicalIV, Step}),
2128         VPValue(this), ResultTy(ResultTy), IndDesc(IndDesc) {}
2129 
2130   ~VPDerivedIVRecipe() override = default;
2131 
2132   VP_CLASSOF_IMPL(VPDef::VPDerivedIVSC)
2133 
2134   /// Generate the transformed value of the induction at offset StartValue (1.
2135   /// operand) + IV (2. operand) * StepValue (3, operand).
2136   void execute(VPTransformState &State) override;
2137 
2138 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2139   /// Print the recipe.
2140   void print(raw_ostream &O, const Twine &Indent,
2141              VPSlotTracker &SlotTracker) const override;
2142 #endif
2143 
2144   VPValue *getStartValue() const { return getOperand(0); }
2145   VPValue *getCanonicalIV() const { return getOperand(1); }
2146   VPValue *getStepValue() const { return getOperand(2); }
2147 
2148   /// Returns true if the recipe only uses the first lane of operand \p Op.
2149   bool onlyFirstLaneUsed(const VPValue *Op) const override {
2150     assert(is_contained(operands(), Op) &&
2151            "Op must be an operand of the recipe");
2152     return true;
2153   }
2154 };
2155 
2156 /// A recipe for handling phi nodes of integer and floating-point inductions,
2157 /// producing their scalar values.
2158 class VPScalarIVStepsRecipe : public VPRecipeBase, public VPValue {
2159   const InductionDescriptor &IndDesc;
2160 
2161 public:
2162   VPScalarIVStepsRecipe(const InductionDescriptor &IndDesc, VPValue *IV,
2163                         VPValue *Step)
2164       : VPRecipeBase(VPDef::VPScalarIVStepsSC, {IV, Step}), VPValue(this),
2165         IndDesc(IndDesc) {}
2166 
2167   ~VPScalarIVStepsRecipe() override = default;
2168 
2169   VP_CLASSOF_IMPL(VPDef::VPScalarIVStepsSC)
2170 
2171   /// Generate the scalarized versions of the phi node as needed by their users.
2172   void execute(VPTransformState &State) override;
2173 
2174 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2175   /// Print the recipe.
2176   void print(raw_ostream &O, const Twine &Indent,
2177              VPSlotTracker &SlotTracker) const override;
2178 #endif
2179 
2180   VPValue *getStepValue() const { return getOperand(1); }
2181 
2182   /// Returns true if the recipe only uses the first lane of operand \p Op.
2183   bool onlyFirstLaneUsed(const VPValue *Op) const override {
2184     assert(is_contained(operands(), Op) &&
2185            "Op must be an operand of the recipe");
2186     return true;
2187   }
2188 };
2189 
2190 /// VPBasicBlock serves as the leaf of the Hierarchical Control-Flow Graph. It
2191 /// holds a sequence of zero or more VPRecipe's each representing a sequence of
2192 /// output IR instructions. All PHI-like recipes must come before any non-PHI recipes.
2193 class VPBasicBlock : public VPBlockBase {
2194 public:
2195   using RecipeListTy = iplist<VPRecipeBase>;
2196 
2197 private:
2198   /// The VPRecipes held in the order of output instructions to generate.
2199   RecipeListTy Recipes;
2200 
2201 public:
2202   VPBasicBlock(const Twine &Name = "", VPRecipeBase *Recipe = nullptr)
2203       : VPBlockBase(VPBasicBlockSC, Name.str()) {
2204     if (Recipe)
2205       appendRecipe(Recipe);
2206   }
2207 
2208   ~VPBasicBlock() override {
2209     while (!Recipes.empty())
2210       Recipes.pop_back();
2211   }
2212 
2213   /// Instruction iterators...
2214   using iterator = RecipeListTy::iterator;
2215   using const_iterator = RecipeListTy::const_iterator;
2216   using reverse_iterator = RecipeListTy::reverse_iterator;
2217   using const_reverse_iterator = RecipeListTy::const_reverse_iterator;
2218 
2219   //===--------------------------------------------------------------------===//
2220   /// Recipe iterator methods
2221   ///
2222   inline iterator begin() { return Recipes.begin(); }
2223   inline const_iterator begin() const { return Recipes.begin(); }
2224   inline iterator end() { return Recipes.end(); }
2225   inline const_iterator end() const { return Recipes.end(); }
2226 
2227   inline reverse_iterator rbegin() { return Recipes.rbegin(); }
2228   inline const_reverse_iterator rbegin() const { return Recipes.rbegin(); }
2229   inline reverse_iterator rend() { return Recipes.rend(); }
2230   inline const_reverse_iterator rend() const { return Recipes.rend(); }
2231 
2232   inline size_t size() const { return Recipes.size(); }
2233   inline bool empty() const { return Recipes.empty(); }
2234   inline const VPRecipeBase &front() const { return Recipes.front(); }
2235   inline VPRecipeBase &front() { return Recipes.front(); }
2236   inline const VPRecipeBase &back() const { return Recipes.back(); }
2237   inline VPRecipeBase &back() { return Recipes.back(); }
2238 
2239   /// Returns a reference to the list of recipes.
2240   RecipeListTy &getRecipeList() { return Recipes; }
2241 
2242   /// Returns a pointer to a member of the recipe list.
2243   static RecipeListTy VPBasicBlock::*getSublistAccess(VPRecipeBase *) {
2244     return &VPBasicBlock::Recipes;
2245   }
2246 
2247   /// Method to support type inquiry through isa, cast, and dyn_cast.
2248   static inline bool classof(const VPBlockBase *V) {
2249     return V->getVPBlockID() == VPBlockBase::VPBasicBlockSC;
2250   }
2251 
2252   void insert(VPRecipeBase *Recipe, iterator InsertPt) {
2253     assert(Recipe && "No recipe to append.");
2254     assert(!Recipe->Parent && "Recipe already in VPlan");
2255     Recipe->Parent = this;
2256     Recipes.insert(InsertPt, Recipe);
2257   }
2258 
2259   /// Augment the existing recipes of a VPBasicBlock with an additional
2260   /// \p Recipe as the last recipe.
2261   void appendRecipe(VPRecipeBase *Recipe) { insert(Recipe, end()); }
2262 
2263   /// The method which generates the output IR instructions that correspond to
2264   /// this VPBasicBlock, thereby "executing" the VPlan.
2265   void execute(VPTransformState *State) override;
2266 
2267   /// Return the position of the first non-phi node recipe in the block.
2268   iterator getFirstNonPhi();
2269 
2270   /// Returns an iterator range over the PHI-like recipes in the block.
2271   iterator_range<iterator> phis() {
2272     return make_range(begin(), getFirstNonPhi());
2273   }
2274 
2275   void dropAllReferences(VPValue *NewValue) override;
2276 
2277   /// Split current block at \p SplitAt by inserting a new block between the
2278   /// current block and its successors and moving all recipes starting at
2279   /// SplitAt to the new block. Returns the new block.
2280   VPBasicBlock *splitAt(iterator SplitAt);
2281 
2282   VPRegionBlock *getEnclosingLoopRegion();
2283 
2284 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2285   /// Print this VPBsicBlock to \p O, prefixing all lines with \p Indent. \p
2286   /// SlotTracker is used to print unnamed VPValue's using consequtive numbers.
2287   ///
2288   /// Note that the numbering is applied to the whole VPlan, so printing
2289   /// individual blocks is consistent with the whole VPlan printing.
2290   void print(raw_ostream &O, const Twine &Indent,
2291              VPSlotTracker &SlotTracker) const override;
2292   using VPBlockBase::print; // Get the print(raw_stream &O) version.
2293 #endif
2294 
2295   /// If the block has multiple successors, return the branch recipe terminating
2296   /// the block. If there are no or only a single successor, return nullptr;
2297   VPRecipeBase *getTerminator();
2298   const VPRecipeBase *getTerminator() const;
2299 
2300   /// Returns true if the block is exiting it's parent region.
2301   bool isExiting() const;
2302 
2303 private:
2304   /// Create an IR BasicBlock to hold the output instructions generated by this
2305   /// VPBasicBlock, and return it. Update the CFGState accordingly.
2306   BasicBlock *createEmptyBasicBlock(VPTransformState::CFGState &CFG);
2307 };
2308 
2309 /// VPRegionBlock represents a collection of VPBasicBlocks and VPRegionBlocks
2310 /// which form a Single-Entry-Single-Exiting subgraph of the output IR CFG.
2311 /// A VPRegionBlock may indicate that its contents are to be replicated several
2312 /// times. This is designed to support predicated scalarization, in which a
2313 /// scalar if-then code structure needs to be generated VF * UF times. Having
2314 /// this replication indicator helps to keep a single model for multiple
2315 /// candidate VF's. The actual replication takes place only once the desired VF
2316 /// and UF have been determined.
2317 class VPRegionBlock : public VPBlockBase {
2318   /// Hold the Single Entry of the SESE region modelled by the VPRegionBlock.
2319   VPBlockBase *Entry;
2320 
2321   /// Hold the Single Exiting block of the SESE region modelled by the
2322   /// VPRegionBlock.
2323   VPBlockBase *Exiting;
2324 
2325   /// An indicator whether this region is to generate multiple replicated
2326   /// instances of output IR corresponding to its VPBlockBases.
2327   bool IsReplicator;
2328 
2329 public:
2330   VPRegionBlock(VPBlockBase *Entry, VPBlockBase *Exiting,
2331                 const std::string &Name = "", bool IsReplicator = false)
2332       : VPBlockBase(VPRegionBlockSC, Name), Entry(Entry), Exiting(Exiting),
2333         IsReplicator(IsReplicator) {
2334     assert(Entry->getPredecessors().empty() && "Entry block has predecessors.");
2335     assert(Exiting->getSuccessors().empty() && "Exit block has successors.");
2336     Entry->setParent(this);
2337     Exiting->setParent(this);
2338   }
2339   VPRegionBlock(const std::string &Name = "", bool IsReplicator = false)
2340       : VPBlockBase(VPRegionBlockSC, Name), Entry(nullptr), Exiting(nullptr),
2341         IsReplicator(IsReplicator) {}
2342 
2343   ~VPRegionBlock() override {
2344     if (Entry) {
2345       VPValue DummyValue;
2346       Entry->dropAllReferences(&DummyValue);
2347       deleteCFG(Entry);
2348     }
2349   }
2350 
2351   /// Method to support type inquiry through isa, cast, and dyn_cast.
2352   static inline bool classof(const VPBlockBase *V) {
2353     return V->getVPBlockID() == VPBlockBase::VPRegionBlockSC;
2354   }
2355 
2356   const VPBlockBase *getEntry() const { return Entry; }
2357   VPBlockBase *getEntry() { return Entry; }
2358 
2359   /// Set \p EntryBlock as the entry VPBlockBase of this VPRegionBlock. \p
2360   /// EntryBlock must have no predecessors.
2361   void setEntry(VPBlockBase *EntryBlock) {
2362     assert(EntryBlock->getPredecessors().empty() &&
2363            "Entry block cannot have predecessors.");
2364     Entry = EntryBlock;
2365     EntryBlock->setParent(this);
2366   }
2367 
2368   const VPBlockBase *getExiting() const { return Exiting; }
2369   VPBlockBase *getExiting() { return Exiting; }
2370 
2371   /// Set \p ExitingBlock as the exiting VPBlockBase of this VPRegionBlock. \p
2372   /// ExitingBlock must have no successors.
2373   void setExiting(VPBlockBase *ExitingBlock) {
2374     assert(ExitingBlock->getSuccessors().empty() &&
2375            "Exit block cannot have successors.");
2376     Exiting = ExitingBlock;
2377     ExitingBlock->setParent(this);
2378   }
2379 
2380   /// Returns the pre-header VPBasicBlock of the loop region.
2381   VPBasicBlock *getPreheaderVPBB() {
2382     assert(!isReplicator() && "should only get pre-header of loop regions");
2383     return getSinglePredecessor()->getExitingBasicBlock();
2384   }
2385 
2386   /// An indicator whether this region is to generate multiple replicated
2387   /// instances of output IR corresponding to its VPBlockBases.
2388   bool isReplicator() const { return IsReplicator; }
2389 
2390   /// The method which generates the output IR instructions that correspond to
2391   /// this VPRegionBlock, thereby "executing" the VPlan.
2392   void execute(VPTransformState *State) override;
2393 
2394   void dropAllReferences(VPValue *NewValue) override;
2395 
2396 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2397   /// Print this VPRegionBlock to \p O (recursively), prefixing all lines with
2398   /// \p Indent. \p SlotTracker is used to print unnamed VPValue's using
2399   /// consequtive numbers.
2400   ///
2401   /// Note that the numbering is applied to the whole VPlan, so printing
2402   /// individual regions is consistent with the whole VPlan printing.
2403   void print(raw_ostream &O, const Twine &Indent,
2404              VPSlotTracker &SlotTracker) const override;
2405   using VPBlockBase::print; // Get the print(raw_stream &O) version.
2406 #endif
2407 };
2408 
2409 /// VPlan models a candidate for vectorization, encoding various decisions take
2410 /// to produce efficient output IR, including which branches, basic-blocks and
2411 /// output IR instructions to generate, and their cost. VPlan holds a
2412 /// Hierarchical-CFG of VPBasicBlocks and VPRegionBlocks rooted at an Entry
2413 /// VPBasicBlock.
2414 class VPlan {
2415   friend class VPlanPrinter;
2416   friend class VPSlotTracker;
2417 
2418   /// Hold the single entry to the Hierarchical CFG of the VPlan, i.e. the
2419   /// preheader of the vector loop.
2420   VPBasicBlock *Entry;
2421 
2422   /// VPBasicBlock corresponding to the original preheader. Used to place
2423   /// VPExpandSCEV recipes for expressions used during skeleton creation and the
2424   /// rest of VPlan execution.
2425   VPBasicBlock *Preheader;
2426 
2427   /// Holds the VFs applicable to this VPlan.
2428   SmallSetVector<ElementCount, 2> VFs;
2429 
2430   /// Holds the UFs applicable to this VPlan. If empty, the VPlan is valid for
2431   /// any UF.
2432   SmallSetVector<unsigned, 2> UFs;
2433 
2434   /// Holds the name of the VPlan, for printing.
2435   std::string Name;
2436 
2437   /// Represents the trip count of the original loop, for folding
2438   /// the tail.
2439   VPValue *TripCount = nullptr;
2440 
2441   /// Represents the backedge taken count of the original loop, for folding
2442   /// the tail. It equals TripCount - 1.
2443   VPValue *BackedgeTakenCount = nullptr;
2444 
2445   /// Represents the vector trip count.
2446   VPValue VectorTripCount;
2447 
2448   /// Holds a mapping between Values and their corresponding VPValue inside
2449   /// VPlan.
2450   Value2VPValueTy Value2VPValue;
2451 
2452   /// Contains all the external definitions created for this VPlan. External
2453   /// definitions are VPValues that hold a pointer to their underlying IR.
2454   SmallVector<VPValue *, 16> VPLiveInsToFree;
2455 
2456   /// Indicates whether it is safe use the Value2VPValue mapping or if the
2457   /// mapping cannot be used any longer, because it is stale.
2458   bool Value2VPValueEnabled = true;
2459 
2460   /// Values used outside the plan.
2461   MapVector<PHINode *, VPLiveOut *> LiveOuts;
2462 
2463   /// Mapping from SCEVs to the VPValues representing their expansions.
2464   /// NOTE: This mapping is temporary and will be removed once all users have
2465   /// been modeled in VPlan directly.
2466   DenseMap<const SCEV *, VPValue *> SCEVToExpansion;
2467 
2468 public:
2469   /// Construct a VPlan with original preheader \p Preheader, trip count \p TC
2470   /// and \p Entry to the plan. At the moment, \p Preheader and \p Entry need to
2471   /// be disconnected, as the bypass blocks between them are not yet modeled in
2472   /// VPlan.
2473   VPlan(VPBasicBlock *Preheader, VPValue *TC, VPBasicBlock *Entry)
2474       : VPlan(Preheader, Entry) {
2475     TripCount = TC;
2476   }
2477 
2478   /// Construct a VPlan with original preheader \p Preheader and \p Entry to
2479   /// the plan. At the moment, \p Preheader and \p Entry need to be
2480   /// disconnected, as the bypass blocks between them are not yet modeled in
2481   /// VPlan.
2482   VPlan(VPBasicBlock *Preheader, VPBasicBlock *Entry)
2483       : Entry(Entry), Preheader(Preheader) {
2484     Entry->setPlan(this);
2485     Preheader->setPlan(this);
2486     assert(Preheader->getNumSuccessors() == 0 &&
2487            Preheader->getNumPredecessors() == 0 &&
2488            "preheader must be disconnected");
2489   }
2490 
2491   ~VPlan();
2492 
2493   /// Create an initial VPlan with preheader and entry blocks. Creates a
2494   /// VPExpandSCEVRecipe for \p TripCount and uses it as plan's trip count.
2495   static VPlanPtr createInitialVPlan(const SCEV *TripCount,
2496                                      ScalarEvolution &PSE);
2497 
2498   /// Prepare the plan for execution, setting up the required live-in values.
2499   void prepareToExecute(Value *TripCount, Value *VectorTripCount,
2500                         Value *CanonicalIVStartValue, VPTransformState &State,
2501                         bool IsEpilogueVectorization);
2502 
2503   /// Generate the IR code for this VPlan.
2504   void execute(VPTransformState *State);
2505 
2506   VPBasicBlock *getEntry() { return Entry; }
2507   const VPBasicBlock *getEntry() const { return Entry; }
2508 
2509   /// The trip count of the original loop.
2510   VPValue *getTripCount() const {
2511     assert(TripCount && "trip count needs to be set before accessing it");
2512     return TripCount;
2513   }
2514 
2515   /// The backedge taken count of the original loop.
2516   VPValue *getOrCreateBackedgeTakenCount() {
2517     if (!BackedgeTakenCount)
2518       BackedgeTakenCount = new VPValue();
2519     return BackedgeTakenCount;
2520   }
2521 
2522   /// The vector trip count.
2523   VPValue &getVectorTripCount() { return VectorTripCount; }
2524 
2525   /// Mark the plan to indicate that using Value2VPValue is not safe any
2526   /// longer, because it may be stale.
2527   void disableValue2VPValue() { Value2VPValueEnabled = false; }
2528 
2529   void addVF(ElementCount VF) { VFs.insert(VF); }
2530 
2531   void setVF(ElementCount VF) {
2532     assert(hasVF(VF) && "Cannot set VF not already in plan");
2533     VFs.clear();
2534     VFs.insert(VF);
2535   }
2536 
2537   bool hasVF(ElementCount VF) { return VFs.count(VF); }
2538 
2539   bool hasScalarVFOnly() const { return VFs.size() == 1 && VFs[0].isScalar(); }
2540 
2541   bool hasUF(unsigned UF) const { return UFs.empty() || UFs.contains(UF); }
2542 
2543   void setUF(unsigned UF) {
2544     assert(hasUF(UF) && "Cannot set the UF not already in plan");
2545     UFs.clear();
2546     UFs.insert(UF);
2547   }
2548 
2549   /// Return a string with the name of the plan and the applicable VFs and UFs.
2550   std::string getName() const;
2551 
2552   void setName(const Twine &newName) { Name = newName.str(); }
2553 
2554   void addVPValue(Value *V, VPValue *VPV) {
2555     assert((Value2VPValueEnabled || VPV->isLiveIn()) &&
2556            "Value2VPValue mapping may be out of date!");
2557     assert(V && "Trying to add a null Value to VPlan");
2558     assert(!Value2VPValue.count(V) && "Value already exists in VPlan");
2559     Value2VPValue[V] = VPV;
2560   }
2561 
2562   /// Returns the VPValue for \p V. \p OverrideAllowed can be used to disable
2563   ///   /// checking whether it is safe to query VPValues using IR Values.
2564   VPValue *getVPValue(Value *V, bool OverrideAllowed = false) {
2565     assert(V && "Trying to get the VPValue of a null Value");
2566     assert(Value2VPValue.count(V) && "Value does not exist in VPlan");
2567     assert((Value2VPValueEnabled || OverrideAllowed ||
2568             Value2VPValue[V]->isLiveIn()) &&
2569            "Value2VPValue mapping may be out of date!");
2570     return Value2VPValue[V];
2571   }
2572 
2573   /// Gets the VPValue for \p V or adds a new live-in (if none exists yet) for
2574   /// \p V.
2575   VPValue *getVPValueOrAddLiveIn(Value *V) {
2576     assert(V && "Trying to get or add the VPValue of a null Value");
2577     if (!Value2VPValue.count(V)) {
2578       VPValue *VPV = new VPValue(V);
2579       VPLiveInsToFree.push_back(VPV);
2580       addVPValue(V, VPV);
2581     }
2582 
2583     return getVPValue(V);
2584   }
2585 
2586   void removeVPValueFor(Value *V) {
2587     assert(Value2VPValueEnabled &&
2588            "IR value to VPValue mapping may be out of date!");
2589     Value2VPValue.erase(V);
2590   }
2591 
2592 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2593   /// Print this VPlan to \p O.
2594   void print(raw_ostream &O) const;
2595 
2596   /// Print this VPlan in DOT format to \p O.
2597   void printDOT(raw_ostream &O) const;
2598 
2599   /// Dump the plan to stderr (for debugging).
2600   LLVM_DUMP_METHOD void dump() const;
2601 #endif
2602 
2603   /// Returns a range mapping the values the range \p Operands to their
2604   /// corresponding VPValues.
2605   iterator_range<mapped_iterator<Use *, std::function<VPValue *(Value *)>>>
2606   mapToVPValues(User::op_range Operands) {
2607     std::function<VPValue *(Value *)> Fn = [this](Value *Op) {
2608       return getVPValueOrAddLiveIn(Op);
2609     };
2610     return map_range(Operands, Fn);
2611   }
2612 
2613   /// Returns the VPRegionBlock of the vector loop.
2614   VPRegionBlock *getVectorLoopRegion() {
2615     return cast<VPRegionBlock>(getEntry()->getSingleSuccessor());
2616   }
2617   const VPRegionBlock *getVectorLoopRegion() const {
2618     return cast<VPRegionBlock>(getEntry()->getSingleSuccessor());
2619   }
2620 
2621   /// Returns the canonical induction recipe of the vector loop.
2622   VPCanonicalIVPHIRecipe *getCanonicalIV() {
2623     VPBasicBlock *EntryVPBB = getVectorLoopRegion()->getEntryBasicBlock();
2624     if (EntryVPBB->empty()) {
2625       // VPlan native path.
2626       EntryVPBB = cast<VPBasicBlock>(EntryVPBB->getSingleSuccessor());
2627     }
2628     return cast<VPCanonicalIVPHIRecipe>(&*EntryVPBB->begin());
2629   }
2630 
2631   /// Find and return the VPActiveLaneMaskPHIRecipe from the header - there
2632   /// be only one at most. If there isn't one, then return nullptr.
2633   VPActiveLaneMaskPHIRecipe *getActiveLaneMaskPhi();
2634 
2635   void addLiveOut(PHINode *PN, VPValue *V);
2636 
2637   void removeLiveOut(PHINode *PN) {
2638     delete LiveOuts[PN];
2639     LiveOuts.erase(PN);
2640   }
2641 
2642   const MapVector<PHINode *, VPLiveOut *> &getLiveOuts() const {
2643     return LiveOuts;
2644   }
2645 
2646   VPValue *getSCEVExpansion(const SCEV *S) const {
2647     return SCEVToExpansion.lookup(S);
2648   }
2649 
2650   void addSCEVExpansion(const SCEV *S, VPValue *V) {
2651     assert(!SCEVToExpansion.contains(S) && "SCEV already expanded");
2652     SCEVToExpansion[S] = V;
2653   }
2654 
2655   /// \return The block corresponding to the original preheader.
2656   VPBasicBlock *getPreheader() { return Preheader; }
2657   const VPBasicBlock *getPreheader() const { return Preheader; }
2658 
2659 private:
2660   /// Add to the given dominator tree the header block and every new basic block
2661   /// that was created between it and the latch block, inclusive.
2662   static void updateDominatorTree(DominatorTree *DT, BasicBlock *LoopLatchBB,
2663                                   BasicBlock *LoopPreHeaderBB,
2664                                   BasicBlock *LoopExitBB);
2665 };
2666 
2667 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2668 /// VPlanPrinter prints a given VPlan to a given output stream. The printing is
2669 /// indented and follows the dot format.
2670 class VPlanPrinter {
2671   raw_ostream &OS;
2672   const VPlan &Plan;
2673   unsigned Depth = 0;
2674   unsigned TabWidth = 2;
2675   std::string Indent;
2676   unsigned BID = 0;
2677   SmallDenseMap<const VPBlockBase *, unsigned> BlockID;
2678 
2679   VPSlotTracker SlotTracker;
2680 
2681   /// Handle indentation.
2682   void bumpIndent(int b) { Indent = std::string((Depth += b) * TabWidth, ' '); }
2683 
2684   /// Print a given \p Block of the Plan.
2685   void dumpBlock(const VPBlockBase *Block);
2686 
2687   /// Print the information related to the CFG edges going out of a given
2688   /// \p Block, followed by printing the successor blocks themselves.
2689   void dumpEdges(const VPBlockBase *Block);
2690 
2691   /// Print a given \p BasicBlock, including its VPRecipes, followed by printing
2692   /// its successor blocks.
2693   void dumpBasicBlock(const VPBasicBlock *BasicBlock);
2694 
2695   /// Print a given \p Region of the Plan.
2696   void dumpRegion(const VPRegionBlock *Region);
2697 
2698   unsigned getOrCreateBID(const VPBlockBase *Block) {
2699     return BlockID.count(Block) ? BlockID[Block] : BlockID[Block] = BID++;
2700   }
2701 
2702   Twine getOrCreateName(const VPBlockBase *Block);
2703 
2704   Twine getUID(const VPBlockBase *Block);
2705 
2706   /// Print the information related to a CFG edge between two VPBlockBases.
2707   void drawEdge(const VPBlockBase *From, const VPBlockBase *To, bool Hidden,
2708                 const Twine &Label);
2709 
2710 public:
2711   VPlanPrinter(raw_ostream &O, const VPlan &P)
2712       : OS(O), Plan(P), SlotTracker(&P) {}
2713 
2714   LLVM_DUMP_METHOD void dump();
2715 };
2716 
2717 struct VPlanIngredient {
2718   const Value *V;
2719 
2720   VPlanIngredient(const Value *V) : V(V) {}
2721 
2722   void print(raw_ostream &O) const;
2723 };
2724 
2725 inline raw_ostream &operator<<(raw_ostream &OS, const VPlanIngredient &I) {
2726   I.print(OS);
2727   return OS;
2728 }
2729 
2730 inline raw_ostream &operator<<(raw_ostream &OS, const VPlan &Plan) {
2731   Plan.print(OS);
2732   return OS;
2733 }
2734 #endif
2735 
2736 //===----------------------------------------------------------------------===//
2737 // VPlan Utilities
2738 //===----------------------------------------------------------------------===//
2739 
2740 /// Class that provides utilities for VPBlockBases in VPlan.
2741 class VPBlockUtils {
2742 public:
2743   VPBlockUtils() = delete;
2744 
2745   /// Insert disconnected VPBlockBase \p NewBlock after \p BlockPtr. Add \p
2746   /// NewBlock as successor of \p BlockPtr and \p BlockPtr as predecessor of \p
2747   /// NewBlock, and propagate \p BlockPtr parent to \p NewBlock. \p BlockPtr's
2748   /// successors are moved from \p BlockPtr to \p NewBlock. \p NewBlock must
2749   /// have neither successors nor predecessors.
2750   static void insertBlockAfter(VPBlockBase *NewBlock, VPBlockBase *BlockPtr) {
2751     assert(NewBlock->getSuccessors().empty() &&
2752            NewBlock->getPredecessors().empty() &&
2753            "Can't insert new block with predecessors or successors.");
2754     NewBlock->setParent(BlockPtr->getParent());
2755     SmallVector<VPBlockBase *> Succs(BlockPtr->successors());
2756     for (VPBlockBase *Succ : Succs) {
2757       disconnectBlocks(BlockPtr, Succ);
2758       connectBlocks(NewBlock, Succ);
2759     }
2760     connectBlocks(BlockPtr, NewBlock);
2761   }
2762 
2763   /// Insert disconnected VPBlockBases \p IfTrue and \p IfFalse after \p
2764   /// BlockPtr. Add \p IfTrue and \p IfFalse as succesors of \p BlockPtr and \p
2765   /// BlockPtr as predecessor of \p IfTrue and \p IfFalse. Propagate \p BlockPtr
2766   /// parent to \p IfTrue and \p IfFalse. \p BlockPtr must have no successors
2767   /// and \p IfTrue and \p IfFalse must have neither successors nor
2768   /// predecessors.
2769   static void insertTwoBlocksAfter(VPBlockBase *IfTrue, VPBlockBase *IfFalse,
2770                                    VPBlockBase *BlockPtr) {
2771     assert(IfTrue->getSuccessors().empty() &&
2772            "Can't insert IfTrue with successors.");
2773     assert(IfFalse->getSuccessors().empty() &&
2774            "Can't insert IfFalse with successors.");
2775     BlockPtr->setTwoSuccessors(IfTrue, IfFalse);
2776     IfTrue->setPredecessors({BlockPtr});
2777     IfFalse->setPredecessors({BlockPtr});
2778     IfTrue->setParent(BlockPtr->getParent());
2779     IfFalse->setParent(BlockPtr->getParent());
2780   }
2781 
2782   /// Connect VPBlockBases \p From and \p To bi-directionally. Append \p To to
2783   /// the successors of \p From and \p From to the predecessors of \p To. Both
2784   /// VPBlockBases must have the same parent, which can be null. Both
2785   /// VPBlockBases can be already connected to other VPBlockBases.
2786   static void connectBlocks(VPBlockBase *From, VPBlockBase *To) {
2787     assert((From->getParent() == To->getParent()) &&
2788            "Can't connect two block with different parents");
2789     assert(From->getNumSuccessors() < 2 &&
2790            "Blocks can't have more than two successors.");
2791     From->appendSuccessor(To);
2792     To->appendPredecessor(From);
2793   }
2794 
2795   /// Disconnect VPBlockBases \p From and \p To bi-directionally. Remove \p To
2796   /// from the successors of \p From and \p From from the predecessors of \p To.
2797   static void disconnectBlocks(VPBlockBase *From, VPBlockBase *To) {
2798     assert(To && "Successor to disconnect is null.");
2799     From->removeSuccessor(To);
2800     To->removePredecessor(From);
2801   }
2802 
2803   /// Return an iterator range over \p Range which only includes \p BlockTy
2804   /// blocks. The accesses are casted to \p BlockTy.
2805   template <typename BlockTy, typename T>
2806   static auto blocksOnly(const T &Range) {
2807     // Create BaseTy with correct const-ness based on BlockTy.
2808     using BaseTy = std::conditional_t<std::is_const<BlockTy>::value,
2809                                       const VPBlockBase, VPBlockBase>;
2810 
2811     // We need to first create an iterator range over (const) BlocktTy & instead
2812     // of (const) BlockTy * for filter_range to work properly.
2813     auto Mapped =
2814         map_range(Range, [](BaseTy *Block) -> BaseTy & { return *Block; });
2815     auto Filter = make_filter_range(
2816         Mapped, [](BaseTy &Block) { return isa<BlockTy>(&Block); });
2817     return map_range(Filter, [](BaseTy &Block) -> BlockTy * {
2818       return cast<BlockTy>(&Block);
2819     });
2820   }
2821 };
2822 
2823 class VPInterleavedAccessInfo {
2824   DenseMap<VPInstruction *, InterleaveGroup<VPInstruction> *>
2825       InterleaveGroupMap;
2826 
2827   /// Type for mapping of instruction based interleave groups to VPInstruction
2828   /// interleave groups
2829   using Old2NewTy = DenseMap<InterleaveGroup<Instruction> *,
2830                              InterleaveGroup<VPInstruction> *>;
2831 
2832   /// Recursively \p Region and populate VPlan based interleave groups based on
2833   /// \p IAI.
2834   void visitRegion(VPRegionBlock *Region, Old2NewTy &Old2New,
2835                    InterleavedAccessInfo &IAI);
2836   /// Recursively traverse \p Block and populate VPlan based interleave groups
2837   /// based on \p IAI.
2838   void visitBlock(VPBlockBase *Block, Old2NewTy &Old2New,
2839                   InterleavedAccessInfo &IAI);
2840 
2841 public:
2842   VPInterleavedAccessInfo(VPlan &Plan, InterleavedAccessInfo &IAI);
2843 
2844   ~VPInterleavedAccessInfo() {
2845     SmallPtrSet<InterleaveGroup<VPInstruction> *, 4> DelSet;
2846     // Avoid releasing a pointer twice.
2847     for (auto &I : InterleaveGroupMap)
2848       DelSet.insert(I.second);
2849     for (auto *Ptr : DelSet)
2850       delete Ptr;
2851   }
2852 
2853   /// Get the interleave group that \p Instr belongs to.
2854   ///
2855   /// \returns nullptr if doesn't have such group.
2856   InterleaveGroup<VPInstruction> *
2857   getInterleaveGroup(VPInstruction *Instr) const {
2858     return InterleaveGroupMap.lookup(Instr);
2859   }
2860 };
2861 
2862 /// Class that maps (parts of) an existing VPlan to trees of combined
2863 /// VPInstructions.
2864 class VPlanSlp {
2865   enum class OpMode { Failed, Load, Opcode };
2866 
2867   /// A DenseMapInfo implementation for using SmallVector<VPValue *, 4> as
2868   /// DenseMap keys.
2869   struct BundleDenseMapInfo {
2870     static SmallVector<VPValue *, 4> getEmptyKey() {
2871       return {reinterpret_cast<VPValue *>(-1)};
2872     }
2873 
2874     static SmallVector<VPValue *, 4> getTombstoneKey() {
2875       return {reinterpret_cast<VPValue *>(-2)};
2876     }
2877 
2878     static unsigned getHashValue(const SmallVector<VPValue *, 4> &V) {
2879       return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
2880     }
2881 
2882     static bool isEqual(const SmallVector<VPValue *, 4> &LHS,
2883                         const SmallVector<VPValue *, 4> &RHS) {
2884       return LHS == RHS;
2885     }
2886   };
2887 
2888   /// Mapping of values in the original VPlan to a combined VPInstruction.
2889   DenseMap<SmallVector<VPValue *, 4>, VPInstruction *, BundleDenseMapInfo>
2890       BundleToCombined;
2891 
2892   VPInterleavedAccessInfo &IAI;
2893 
2894   /// Basic block to operate on. For now, only instructions in a single BB are
2895   /// considered.
2896   const VPBasicBlock &BB;
2897 
2898   /// Indicates whether we managed to combine all visited instructions or not.
2899   bool CompletelySLP = true;
2900 
2901   /// Width of the widest combined bundle in bits.
2902   unsigned WidestBundleBits = 0;
2903 
2904   using MultiNodeOpTy =
2905       typename std::pair<VPInstruction *, SmallVector<VPValue *, 4>>;
2906 
2907   // Input operand bundles for the current multi node. Each multi node operand
2908   // bundle contains values not matching the multi node's opcode. They will
2909   // be reordered in reorderMultiNodeOps, once we completed building a
2910   // multi node.
2911   SmallVector<MultiNodeOpTy, 4> MultiNodeOps;
2912 
2913   /// Indicates whether we are building a multi node currently.
2914   bool MultiNodeActive = false;
2915 
2916   /// Check if we can vectorize Operands together.
2917   bool areVectorizable(ArrayRef<VPValue *> Operands) const;
2918 
2919   /// Add combined instruction \p New for the bundle \p Operands.
2920   void addCombined(ArrayRef<VPValue *> Operands, VPInstruction *New);
2921 
2922   /// Indicate we hit a bundle we failed to combine. Returns nullptr for now.
2923   VPInstruction *markFailed();
2924 
2925   /// Reorder operands in the multi node to maximize sequential memory access
2926   /// and commutative operations.
2927   SmallVector<MultiNodeOpTy, 4> reorderMultiNodeOps();
2928 
2929   /// Choose the best candidate to use for the lane after \p Last. The set of
2930   /// candidates to choose from are values with an opcode matching \p Last's
2931   /// or loads consecutive to \p Last.
2932   std::pair<OpMode, VPValue *> getBest(OpMode Mode, VPValue *Last,
2933                                        SmallPtrSetImpl<VPValue *> &Candidates,
2934                                        VPInterleavedAccessInfo &IAI);
2935 
2936 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2937   /// Print bundle \p Values to dbgs().
2938   void dumpBundle(ArrayRef<VPValue *> Values);
2939 #endif
2940 
2941 public:
2942   VPlanSlp(VPInterleavedAccessInfo &IAI, VPBasicBlock &BB) : IAI(IAI), BB(BB) {}
2943 
2944   ~VPlanSlp() = default;
2945 
2946   /// Tries to build an SLP tree rooted at \p Operands and returns a
2947   /// VPInstruction combining \p Operands, if they can be combined.
2948   VPInstruction *buildGraph(ArrayRef<VPValue *> Operands);
2949 
2950   /// Return the width of the widest combined bundle in bits.
2951   unsigned getWidestBundleBits() const { return WidestBundleBits; }
2952 
2953   /// Return true if all visited instruction can be combined.
2954   bool isCompletelySLP() const { return CompletelySLP; }
2955 };
2956 
2957 namespace vputils {
2958 
2959 /// Returns true if only the first lane of \p Def is used.
2960 bool onlyFirstLaneUsed(VPValue *Def);
2961 
2962 /// Get or create a VPValue that corresponds to the expansion of \p Expr. If \p
2963 /// Expr is a SCEVConstant or SCEVUnknown, return a VPValue wrapping the live-in
2964 /// value. Otherwise return a VPExpandSCEVRecipe to expand \p Expr. If \p Plan's
2965 /// pre-header already contains a recipe expanding \p Expr, return it. If not,
2966 /// create a new one.
2967 VPValue *getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr,
2968                                        ScalarEvolution &SE);
2969 
2970 /// Returns true if \p VPV is uniform after vectorization.
2971 inline bool isUniformAfterVectorization(VPValue *VPV) {
2972   // A value defined outside the vector region must be uniform after
2973   // vectorization inside a vector region.
2974   if (VPV->isDefinedOutsideVectorRegions())
2975     return true;
2976   VPRecipeBase *Def = VPV->getDefiningRecipe();
2977   assert(Def && "Must have definition for value defined inside vector region");
2978   if (auto Rep = dyn_cast<VPReplicateRecipe>(Def))
2979     return Rep->isUniform();
2980   if (auto *GEP = dyn_cast<VPWidenGEPRecipe>(Def))
2981     return all_of(GEP->operands(), isUniformAfterVectorization);
2982   return false;
2983 }
2984 } // end namespace vputils
2985 
2986 } // end namespace llvm
2987 
2988 #endif // LLVM_TRANSFORMS_VECTORIZE_VPLAN_H
2989