1 //===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
9 // This pass implements the Bottom Up SLP vectorizer. It detects consecutive
10 // stores that can be put together into vector-stores. Next, it attempts to
11 // construct vectorizable tree using the use-def chains. If a profitable tree
12 // was found, the SLP vectorizer performs vectorization on the tree.
14 // The pass is inspired by the work described in the paper:
15 // "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
17 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Vectorize.h"
19 #include "llvm/ADT/MapVector.h"
20 #include "llvm/ADT/PostOrderIterator.h"
21 #include "llvm/ADT/SetVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/LoopInfo.h"
25 #include "llvm/Analysis/ScalarEvolution.h"
26 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
27 #include "llvm/Analysis/TargetTransformInfo.h"
28 #include "llvm/Analysis/ValueTracking.h"
29 #include "llvm/IR/DataLayout.h"
30 #include "llvm/IR/Dominators.h"
31 #include "llvm/IR/IRBuilder.h"
32 #include "llvm/IR/Instructions.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/Module.h"
35 #include "llvm/IR/NoFolder.h"
36 #include "llvm/IR/Type.h"
37 #include "llvm/IR/Value.h"
38 #include "llvm/IR/Verifier.h"
39 #include "llvm/Pass.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Transforms/Utils/VectorUtils.h"
50 #define SV_NAME "slp-vectorizer"
51 #define DEBUG_TYPE "SLP"
53 STATISTIC(NumVectorInstructions, "Number of vector instructions generated");
56 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
57 cl::desc("Only vectorize if you gain more than this "
61 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
62 cl::desc("Attempt to vectorize horizontal reductions"));
64 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
65 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
67 "Attempt to vectorize horizontal reductions feeding into a store"));
71 static const unsigned MinVecRegSize = 128;
73 static const unsigned RecursionMaxDepth = 12;
75 /// \returns the parent basic block if all of the instructions in \p VL
76 /// are in the same block or null otherwise.
77 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
78 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
81 BasicBlock *BB = I0->getParent();
82 for (int i = 1, e = VL.size(); i < e; i++) {
83 Instruction *I = dyn_cast<Instruction>(VL[i]);
87 if (BB != I->getParent())
93 /// \returns True if all of the values in \p VL are constants.
94 static bool allConstant(ArrayRef<Value *> VL) {
95 for (unsigned i = 0, e = VL.size(); i < e; ++i)
96 if (!isa<Constant>(VL[i]))
101 /// \returns True if all of the values in \p VL are identical.
102 static bool isSplat(ArrayRef<Value *> VL) {
103 for (unsigned i = 1, e = VL.size(); i < e; ++i)
109 ///\returns Opcode that can be clubbed with \p Op to create an alternate
110 /// sequence which can later be merged as a ShuffleVector instruction.
111 static unsigned getAltOpcode(unsigned Op) {
113 case Instruction::FAdd:
114 return Instruction::FSub;
115 case Instruction::FSub:
116 return Instruction::FAdd;
117 case Instruction::Add:
118 return Instruction::Sub;
119 case Instruction::Sub:
120 return Instruction::Add;
126 ///\returns bool representing if Opcode \p Op can be part
127 /// of an alternate sequence which can later be merged as
128 /// a ShuffleVector instruction.
129 static bool canCombineAsAltInst(unsigned Op) {
130 if (Op == Instruction::FAdd || Op == Instruction::FSub ||
131 Op == Instruction::Sub || Op == Instruction::Add)
136 /// \returns ShuffleVector instruction if intructions in \p VL have
137 /// alternate fadd,fsub / fsub,fadd/add,sub/sub,add sequence.
138 /// (i.e. e.g. opcodes of fadd,fsub,fadd,fsub...)
139 static unsigned isAltInst(ArrayRef<Value *> VL) {
140 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
141 unsigned Opcode = I0->getOpcode();
142 unsigned AltOpcode = getAltOpcode(Opcode);
143 for (int i = 1, e = VL.size(); i < e; i++) {
144 Instruction *I = dyn_cast<Instruction>(VL[i]);
145 if (!I || I->getOpcode() != ((i & 1) ? AltOpcode : Opcode))
148 return Instruction::ShuffleVector;
151 /// \returns The opcode if all of the Instructions in \p VL have the same
153 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
154 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
157 unsigned Opcode = I0->getOpcode();
158 for (int i = 1, e = VL.size(); i < e; i++) {
159 Instruction *I = dyn_cast<Instruction>(VL[i]);
160 if (!I || Opcode != I->getOpcode()) {
161 if (canCombineAsAltInst(Opcode) && i == 1)
162 return isAltInst(VL);
169 /// \returns \p I after propagating metadata from \p VL.
170 static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) {
171 Instruction *I0 = cast<Instruction>(VL[0]);
172 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
173 I0->getAllMetadataOtherThanDebugLoc(Metadata);
175 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) {
176 unsigned Kind = Metadata[i].first;
177 MDNode *MD = Metadata[i].second;
179 for (int i = 1, e = VL.size(); MD && i != e; i++) {
180 Instruction *I = cast<Instruction>(VL[i]);
181 MDNode *IMD = I->getMetadata(Kind);
185 MD = nullptr; // Remove unknown metadata
187 case LLVMContext::MD_tbaa:
188 MD = MDNode::getMostGenericTBAA(MD, IMD);
190 case LLVMContext::MD_alias_scope:
191 case LLVMContext::MD_noalias:
192 MD = MDNode::intersect(MD, IMD);
194 case LLVMContext::MD_fpmath:
195 MD = MDNode::getMostGenericFPMath(MD, IMD);
199 I->setMetadata(Kind, MD);
204 /// \returns The type that all of the values in \p VL have or null if there
205 /// are different types.
206 static Type* getSameType(ArrayRef<Value *> VL) {
207 Type *Ty = VL[0]->getType();
208 for (int i = 1, e = VL.size(); i < e; i++)
209 if (VL[i]->getType() != Ty)
215 /// \returns True if the ExtractElement instructions in VL can be vectorized
216 /// to use the original vector.
217 static bool CanReuseExtract(ArrayRef<Value *> VL) {
218 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
219 // Check if all of the extracts come from the same vector and from the
222 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
223 Value *Vec = E0->getOperand(0);
225 // We have to extract from the same vector type.
226 unsigned NElts = Vec->getType()->getVectorNumElements();
228 if (NElts != VL.size())
231 // Check that all of the indices extract from the correct offset.
232 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
233 if (!CI || CI->getZExtValue())
236 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
237 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
238 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
240 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
247 static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
248 SmallVectorImpl<Value *> &Left,
249 SmallVectorImpl<Value *> &Right) {
251 SmallVector<Value *, 16> OrigLeft, OrigRight;
253 bool AllSameOpcodeLeft = true;
254 bool AllSameOpcodeRight = true;
255 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
256 Instruction *I = cast<Instruction>(VL[i]);
257 Value *V0 = I->getOperand(0);
258 Value *V1 = I->getOperand(1);
260 OrigLeft.push_back(V0);
261 OrigRight.push_back(V1);
263 Instruction *I0 = dyn_cast<Instruction>(V0);
264 Instruction *I1 = dyn_cast<Instruction>(V1);
266 // Check whether all operands on one side have the same opcode. In this case
267 // we want to preserve the original order and not make things worse by
269 AllSameOpcodeLeft = I0;
270 AllSameOpcodeRight = I1;
272 if (i && AllSameOpcodeLeft) {
273 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) {
274 if(P0->getOpcode() != I0->getOpcode())
275 AllSameOpcodeLeft = false;
277 AllSameOpcodeLeft = false;
279 if (i && AllSameOpcodeRight) {
280 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) {
281 if(P1->getOpcode() != I1->getOpcode())
282 AllSameOpcodeRight = false;
284 AllSameOpcodeRight = false;
287 // Sort two opcodes. In the code below we try to preserve the ability to use
288 // broadcast of values instead of individual inserts.
295 // If we just sorted according to opcode we would leave the first line in
296 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load).
299 // Because vr2 and vr1 are from the same load we loose the opportunity of a
300 // broadcast for the packed right side in the backend: we have [vr1, vl2]
301 // instead of [vr1, vr2=vr1].
303 if(!i && I0->getOpcode() > I1->getOpcode()) {
306 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) {
307 // Try not to destroy a broad cast for no apparent benefit.
310 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) {
311 // Try preserve broadcasts.
314 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) {
315 // Try preserve broadcasts.
324 // One opcode, put the instruction on the right.
334 bool LeftBroadcast = isSplat(Left);
335 bool RightBroadcast = isSplat(Right);
337 // Don't reorder if the operands where good to begin with.
338 if (!(LeftBroadcast || RightBroadcast) &&
339 (AllSameOpcodeRight || AllSameOpcodeLeft)) {
345 /// Bottom Up SLP Vectorizer.
348 typedef SmallVector<Value *, 8> ValueList;
349 typedef SmallVector<Instruction *, 16> InstrList;
350 typedef SmallPtrSet<Value *, 16> ValueSet;
351 typedef SmallVector<StoreInst *, 8> StoreList;
353 BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl,
354 TargetTransformInfo *Tti, TargetLibraryInfo *TLi, AliasAnalysis *Aa,
355 LoopInfo *Li, DominatorTree *Dt)
356 : NumLoadsWantToKeepOrder(0), NumLoadsWantToChangeOrder(0),
357 F(Func), SE(Se), DL(Dl), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt),
358 Builder(Se->getContext()) {}
360 /// \brief Vectorize the tree that starts with the elements in \p VL.
361 /// Returns the vectorized root.
362 Value *vectorizeTree();
364 /// \returns the cost incurred by unwanted spills and fills, caused by
365 /// holding live values over call sites.
368 /// \returns the vectorization cost of the subtree that starts at \p VL.
369 /// A negative number means that this is profitable.
372 /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
373 /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
374 void buildTree(ArrayRef<Value *> Roots,
375 ArrayRef<Value *> UserIgnoreLst = None);
377 /// Clear the internal data structures that are created by 'buildTree'.
379 VectorizableTree.clear();
380 ScalarToTreeEntry.clear();
382 ExternalUses.clear();
383 NumLoadsWantToKeepOrder = 0;
384 NumLoadsWantToChangeOrder = 0;
385 for (auto &Iter : BlocksSchedules) {
386 BlockScheduling *BS = Iter.second.get();
391 /// \returns true if the memory operations A and B are consecutive.
392 bool isConsecutiveAccess(Value *A, Value *B);
394 /// \brief Perform LICM and CSE on the newly generated gather sequences.
395 void optimizeGatherSequence();
397 /// \returns true if it is benefitial to reverse the vector order.
398 bool shouldReorder() const {
399 return NumLoadsWantToChangeOrder > NumLoadsWantToKeepOrder;
405 /// \returns the cost of the vectorizable entry.
406 int getEntryCost(TreeEntry *E);
408 /// This is the recursive part of buildTree.
409 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
411 /// Vectorize a single entry in the tree.
412 Value *vectorizeTree(TreeEntry *E);
414 /// Vectorize a single entry in the tree, starting in \p VL.
415 Value *vectorizeTree(ArrayRef<Value *> VL);
417 /// \returns the pointer to the vectorized value if \p VL is already
418 /// vectorized, or NULL. They may happen in cycles.
419 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
421 /// \brief Take the pointer operand from the Load/Store instruction.
422 /// \returns NULL if this is not a valid Load/Store instruction.
423 static Value *getPointerOperand(Value *I);
425 /// \brief Take the address space operand from the Load/Store instruction.
426 /// \returns -1 if this is not a valid Load/Store instruction.
427 static unsigned getAddressSpaceOperand(Value *I);
429 /// \returns the scalarization cost for this type. Scalarization in this
430 /// context means the creation of vectors from a group of scalars.
431 int getGatherCost(Type *Ty);
433 /// \returns the scalarization cost for this list of values. Assuming that
434 /// this subtree gets vectorized, we may need to extract the values from the
435 /// roots. This method calculates the cost of extracting the values.
436 int getGatherCost(ArrayRef<Value *> VL);
438 /// \brief Set the Builder insert point to one after the last instruction in
440 void setInsertPointAfterBundle(ArrayRef<Value *> VL);
442 /// \returns a vector from a collection of scalars in \p VL.
443 Value *Gather(ArrayRef<Value *> VL, VectorType *Ty);
445 /// \returns whether the VectorizableTree is fully vectoriable and will
446 /// be beneficial even the tree height is tiny.
447 bool isFullyVectorizableTinyTree();
450 TreeEntry() : Scalars(), VectorizedValue(nullptr),
453 /// \returns true if the scalars in VL are equal to this entry.
454 bool isSame(ArrayRef<Value *> VL) const {
455 assert(VL.size() == Scalars.size() && "Invalid size");
456 return std::equal(VL.begin(), VL.end(), Scalars.begin());
459 /// A vector of scalars.
462 /// The Scalars are vectorized into this value. It is initialized to Null.
463 Value *VectorizedValue;
465 /// Do we need to gather this sequence ?
469 /// Create a new VectorizableTree entry.
470 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
471 VectorizableTree.push_back(TreeEntry());
472 int idx = VectorizableTree.size() - 1;
473 TreeEntry *Last = &VectorizableTree[idx];
474 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
475 Last->NeedToGather = !Vectorized;
477 for (int i = 0, e = VL.size(); i != e; ++i) {
478 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
479 ScalarToTreeEntry[VL[i]] = idx;
482 MustGather.insert(VL.begin(), VL.end());
487 /// -- Vectorization State --
488 /// Holds all of the tree entries.
489 std::vector<TreeEntry> VectorizableTree;
491 /// Maps a specific scalar to its tree entry.
492 SmallDenseMap<Value*, int> ScalarToTreeEntry;
494 /// A list of scalars that we found that we need to keep as scalars.
497 /// This POD struct describes one external user in the vectorized tree.
498 struct ExternalUser {
499 ExternalUser (Value *S, llvm::User *U, int L) :
500 Scalar(S), User(U), Lane(L){};
501 // Which scalar in our function.
503 // Which user that uses the scalar.
505 // Which lane does the scalar belong to.
508 typedef SmallVector<ExternalUser, 16> UserList;
510 /// A list of values that need to extracted out of the tree.
511 /// This list holds pairs of (Internal Scalar : External User).
512 UserList ExternalUses;
514 /// Holds all of the instructions that we gathered.
515 SetVector<Instruction *> GatherSeq;
516 /// A list of blocks that we are going to CSE.
517 SetVector<BasicBlock *> CSEBlocks;
519 /// Contains all scheduling relevant data for an instruction.
520 /// A ScheduleData either represents a single instruction or a member of an
521 /// instruction bundle (= a group of instructions which is combined into a
522 /// vector instruction).
523 struct ScheduleData {
525 // The initial value for the dependency counters. It means that the
526 // dependencies are not calculated yet.
527 enum { InvalidDeps = -1 };
530 : Inst(nullptr), FirstInBundle(nullptr), NextInBundle(nullptr),
531 NextLoadStore(nullptr), SchedulingRegionID(0), SchedulingPriority(0),
532 Dependencies(InvalidDeps), UnscheduledDeps(InvalidDeps),
533 UnscheduledDepsInBundle(InvalidDeps), IsScheduled(false) {}
535 void init(int BlockSchedulingRegionID) {
536 FirstInBundle = this;
537 NextInBundle = nullptr;
538 NextLoadStore = nullptr;
540 SchedulingRegionID = BlockSchedulingRegionID;
541 UnscheduledDepsInBundle = UnscheduledDeps;
545 /// Returns true if the dependency information has been calculated.
546 bool hasValidDependencies() const { return Dependencies != InvalidDeps; }
548 /// Returns true for single instructions and for bundle representatives
549 /// (= the head of a bundle).
550 bool isSchedulingEntity() const { return FirstInBundle == this; }
552 /// Returns true if it represents an instruction bundle and not only a
553 /// single instruction.
554 bool isPartOfBundle() const {
555 return NextInBundle != nullptr || FirstInBundle != this;
558 /// Returns true if it is ready for scheduling, i.e. it has no more
559 /// unscheduled depending instructions/bundles.
560 bool isReady() const {
561 assert(isSchedulingEntity() &&
562 "can't consider non-scheduling entity for ready list");
563 return UnscheduledDepsInBundle == 0 && !IsScheduled;
566 /// Modifies the number of unscheduled dependencies, also updating it for
567 /// the whole bundle.
568 int incrementUnscheduledDeps(int Incr) {
569 UnscheduledDeps += Incr;
570 return FirstInBundle->UnscheduledDepsInBundle += Incr;
573 /// Sets the number of unscheduled dependencies to the number of
575 void resetUnscheduledDeps() {
576 incrementUnscheduledDeps(Dependencies - UnscheduledDeps);
579 /// Clears all dependency information.
580 void clearDependencies() {
581 Dependencies = InvalidDeps;
582 resetUnscheduledDeps();
583 MemoryDependencies.clear();
586 void dump(raw_ostream &os) const {
587 if (!isSchedulingEntity()) {
589 } else if (NextInBundle) {
591 ScheduleData *SD = NextInBundle;
593 os << ';' << *SD->Inst;
594 SD = SD->NextInBundle;
604 /// Points to the head in an instruction bundle (and always to this for
605 /// single instructions).
606 ScheduleData *FirstInBundle;
608 /// Single linked list of all instructions in a bundle. Null if it is a
609 /// single instruction.
610 ScheduleData *NextInBundle;
612 /// Single linked list of all memory instructions (e.g. load, store, call)
613 /// in the block - until the end of the scheduling region.
614 ScheduleData *NextLoadStore;
616 /// The dependent memory instructions.
617 /// This list is derived on demand in calculateDependencies().
618 SmallVector<ScheduleData *, 4> MemoryDependencies;
620 /// This ScheduleData is in the current scheduling region if this matches
621 /// the current SchedulingRegionID of BlockScheduling.
622 int SchedulingRegionID;
624 /// Used for getting a "good" final ordering of instructions.
625 int SchedulingPriority;
627 /// The number of dependencies. Constitutes of the number of users of the
628 /// instruction plus the number of dependent memory instructions (if any).
629 /// This value is calculated on demand.
630 /// If InvalidDeps, the number of dependencies is not calculated yet.
634 /// The number of dependencies minus the number of dependencies of scheduled
635 /// instructions. As soon as this is zero, the instruction/bundle gets ready
637 /// Note that this is negative as long as Dependencies is not calculated.
640 /// The sum of UnscheduledDeps in a bundle. Equals to UnscheduledDeps for
641 /// single instructions.
642 int UnscheduledDepsInBundle;
644 /// True if this instruction is scheduled (or considered as scheduled in the
650 friend raw_ostream &operator<<(raw_ostream &os,
651 const BoUpSLP::ScheduleData &SD);
654 /// Contains all scheduling data for a basic block.
656 struct BlockScheduling {
658 BlockScheduling(BasicBlock *BB)
659 : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize),
660 ScheduleStart(nullptr), ScheduleEnd(nullptr),
661 FirstLoadStoreInRegion(nullptr), LastLoadStoreInRegion(nullptr),
662 // Make sure that the initial SchedulingRegionID is greater than the
663 // initial SchedulingRegionID in ScheduleData (which is 0).
664 SchedulingRegionID(1) {}
668 ScheduleStart = nullptr;
669 ScheduleEnd = nullptr;
670 FirstLoadStoreInRegion = nullptr;
671 LastLoadStoreInRegion = nullptr;
673 // Make a new scheduling region, i.e. all existing ScheduleData is not
674 // in the new region yet.
675 ++SchedulingRegionID;
678 ScheduleData *getScheduleData(Value *V) {
679 ScheduleData *SD = ScheduleDataMap[V];
680 if (SD && SD->SchedulingRegionID == SchedulingRegionID)
685 bool isInSchedulingRegion(ScheduleData *SD) {
686 return SD->SchedulingRegionID == SchedulingRegionID;
689 /// Marks an instruction as scheduled and puts all dependent ready
690 /// instructions into the ready-list.
691 template <typename ReadyListType>
692 void schedule(ScheduleData *SD, ReadyListType &ReadyList) {
693 SD->IsScheduled = true;
694 DEBUG(dbgs() << "SLP: schedule " << *SD << "\n");
696 ScheduleData *BundleMember = SD;
697 while (BundleMember) {
698 // Handle the def-use chain dependencies.
699 for (Use &U : BundleMember->Inst->operands()) {
700 ScheduleData *OpDef = getScheduleData(U.get());
701 if (OpDef && OpDef->hasValidDependencies() &&
702 OpDef->incrementUnscheduledDeps(-1) == 0) {
703 // There are no more unscheduled dependencies after decrementing,
704 // so we can put the dependent instruction into the ready list.
705 ScheduleData *DepBundle = OpDef->FirstInBundle;
706 assert(!DepBundle->IsScheduled &&
707 "already scheduled bundle gets ready");
708 ReadyList.insert(DepBundle);
709 DEBUG(dbgs() << "SLP: gets ready (def): " << *DepBundle << "\n");
712 // Handle the memory dependencies.
713 for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) {
714 if (MemoryDepSD->incrementUnscheduledDeps(-1) == 0) {
715 // There are no more unscheduled dependencies after decrementing,
716 // so we can put the dependent instruction into the ready list.
717 ScheduleData *DepBundle = MemoryDepSD->FirstInBundle;
718 assert(!DepBundle->IsScheduled &&
719 "already scheduled bundle gets ready");
720 ReadyList.insert(DepBundle);
721 DEBUG(dbgs() << "SLP: gets ready (mem): " << *DepBundle << "\n");
724 BundleMember = BundleMember->NextInBundle;
728 /// Put all instructions into the ReadyList which are ready for scheduling.
729 template <typename ReadyListType>
730 void initialFillReadyList(ReadyListType &ReadyList) {
731 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
732 ScheduleData *SD = getScheduleData(I);
733 if (SD->isSchedulingEntity() && SD->isReady()) {
734 ReadyList.insert(SD);
735 DEBUG(dbgs() << "SLP: initially in ready list: " << *I << "\n");
740 /// Checks if a bundle of instructions can be scheduled, i.e. has no
741 /// cyclic dependencies. This is only a dry-run, no instructions are
742 /// actually moved at this stage.
743 bool tryScheduleBundle(ArrayRef<Value *> VL, AliasAnalysis *AA);
745 /// Un-bundles a group of instructions.
746 void cancelScheduling(ArrayRef<Value *> VL);
748 /// Extends the scheduling region so that V is inside the region.
749 void extendSchedulingRegion(Value *V);
751 /// Initialize the ScheduleData structures for new instructions in the
752 /// scheduling region.
753 void initScheduleData(Instruction *FromI, Instruction *ToI,
754 ScheduleData *PrevLoadStore,
755 ScheduleData *NextLoadStore);
757 /// Updates the dependency information of a bundle and of all instructions/
758 /// bundles which depend on the original bundle.
759 void calculateDependencies(ScheduleData *SD, bool InsertInReadyList,
762 /// Sets all instruction in the scheduling region to un-scheduled.
763 void resetSchedule();
767 /// Simple memory allocation for ScheduleData.
768 std::vector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks;
770 /// The size of a ScheduleData array in ScheduleDataChunks.
773 /// The allocator position in the current chunk, which is the last entry
774 /// of ScheduleDataChunks.
777 /// Attaches ScheduleData to Instruction.
778 /// Note that the mapping survives during all vectorization iterations, i.e.
779 /// ScheduleData structures are recycled.
780 DenseMap<Value *, ScheduleData *> ScheduleDataMap;
782 struct ReadyList : SmallVector<ScheduleData *, 8> {
783 void insert(ScheduleData *SD) { push_back(SD); }
786 /// The ready-list for scheduling (only used for the dry-run).
787 ReadyList ReadyInsts;
789 /// The first instruction of the scheduling region.
790 Instruction *ScheduleStart;
792 /// The first instruction _after_ the scheduling region.
793 Instruction *ScheduleEnd;
795 /// The first memory accessing instruction in the scheduling region
797 ScheduleData *FirstLoadStoreInRegion;
799 /// The last memory accessing instruction in the scheduling region
801 ScheduleData *LastLoadStoreInRegion;
803 /// The ID of the scheduling region. For a new vectorization iteration this
804 /// is incremented which "removes" all ScheduleData from the region.
805 int SchedulingRegionID;
808 /// Attaches the BlockScheduling structures to basic blocks.
809 DenseMap<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules;
811 /// Performs the "real" scheduling. Done before vectorization is actually
812 /// performed in a basic block.
813 void scheduleBlock(BlockScheduling *BS);
815 /// List of users to ignore during scheduling and that don't need extracting.
816 ArrayRef<Value *> UserIgnoreList;
818 // Number of load-bundles, which contain consecutive loads.
819 int NumLoadsWantToKeepOrder;
821 // Number of load-bundles of size 2, which are consecutive loads if reversed.
822 int NumLoadsWantToChangeOrder;
824 // Analysis and block reference.
827 const DataLayout *DL;
828 TargetTransformInfo *TTI;
829 TargetLibraryInfo *TLI;
833 /// Instruction builder to construct the vectorized tree.
838 raw_ostream &operator<<(raw_ostream &os, const BoUpSLP::ScheduleData &SD) {
844 void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
845 ArrayRef<Value *> UserIgnoreLst) {
847 UserIgnoreList = UserIgnoreLst;
848 if (!getSameType(Roots))
850 buildTree_rec(Roots, 0);
852 // Collect the values that we need to extract from the tree.
853 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
854 TreeEntry *Entry = &VectorizableTree[EIdx];
857 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
858 Value *Scalar = Entry->Scalars[Lane];
860 // No need to handle users of gathered values.
861 if (Entry->NeedToGather)
864 for (User *U : Scalar->users()) {
865 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
867 // Skip in-tree scalars that become vectors.
868 if (ScalarToTreeEntry.count(U)) {
869 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
871 int Idx = ScalarToTreeEntry[U]; (void) Idx;
872 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
875 Instruction *UserInst = dyn_cast<Instruction>(U);
879 // Ignore users in the user ignore list.
880 if (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), UserInst) !=
881 UserIgnoreList.end())
884 DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<
885 Lane << " from " << *Scalar << ".\n");
886 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
893 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
894 bool SameTy = getSameType(VL); (void)SameTy;
895 bool isAltShuffle = false;
896 assert(SameTy && "Invalid types!");
898 if (Depth == RecursionMaxDepth) {
899 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
900 newTreeEntry(VL, false);
904 // Don't handle vectors.
905 if (VL[0]->getType()->isVectorTy()) {
906 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
907 newTreeEntry(VL, false);
911 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
912 if (SI->getValueOperand()->getType()->isVectorTy()) {
913 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
914 newTreeEntry(VL, false);
917 unsigned Opcode = getSameOpcode(VL);
919 // Check that this shuffle vector refers to the alternate
920 // sequence of opcodes.
921 if (Opcode == Instruction::ShuffleVector) {
922 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
923 unsigned Op = I0->getOpcode();
924 if (Op != Instruction::ShuffleVector)
928 // If all of the operands are identical or constant we have a simple solution.
929 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) || !Opcode) {
930 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
931 newTreeEntry(VL, false);
935 // We now know that this is a vector of instructions of the same type from
938 // Check if this is a duplicate of another entry.
939 if (ScalarToTreeEntry.count(VL[0])) {
940 int Idx = ScalarToTreeEntry[VL[0]];
941 TreeEntry *E = &VectorizableTree[Idx];
942 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
943 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
944 if (E->Scalars[i] != VL[i]) {
945 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
946 newTreeEntry(VL, false);
950 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
954 // Check that none of the instructions in the bundle are already in the tree.
955 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
956 if (ScalarToTreeEntry.count(VL[i])) {
957 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
958 ") is already in tree.\n");
959 newTreeEntry(VL, false);
964 // If any of the scalars appears in the table OR it is marked as a value that
965 // needs to stat scalar then we need to gather the scalars.
966 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
967 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
968 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
969 newTreeEntry(VL, false);
974 // Check that all of the users of the scalars that we want to vectorize are
976 Instruction *VL0 = cast<Instruction>(VL[0]);
977 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
979 if (!DT->isReachableFromEntry(BB)) {
980 // Don't go into unreachable blocks. They may contain instructions with
981 // dependency cycles which confuse the final scheduling.
982 DEBUG(dbgs() << "SLP: bundle in unreachable block.\n");
983 newTreeEntry(VL, false);
987 // Check that every instructions appears once in this bundle.
988 for (unsigned i = 0, e = VL.size(); i < e; ++i)
989 for (unsigned j = i+1; j < e; ++j)
990 if (VL[i] == VL[j]) {
991 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
992 newTreeEntry(VL, false);
996 auto &BSRef = BlocksSchedules[BB];
998 BSRef = llvm::make_unique<BlockScheduling>(BB);
1000 BlockScheduling &BS = *BSRef.get();
1002 if (!BS.tryScheduleBundle(VL, AA)) {
1003 DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n");
1004 BS.cancelScheduling(VL);
1005 newTreeEntry(VL, false);
1008 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
1011 case Instruction::PHI: {
1012 PHINode *PH = dyn_cast<PHINode>(VL0);
1014 // Check for terminator values (e.g. invoke).
1015 for (unsigned j = 0; j < VL.size(); ++j)
1016 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1017 TerminatorInst *Term = dyn_cast<TerminatorInst>(
1018 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
1020 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
1021 BS.cancelScheduling(VL);
1022 newTreeEntry(VL, false);
1027 newTreeEntry(VL, true);
1028 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
1030 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1032 // Prepare the operand vector.
1033 for (unsigned j = 0; j < VL.size(); ++j)
1034 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
1035 PH->getIncomingBlock(i)));
1037 buildTree_rec(Operands, Depth + 1);
1041 case Instruction::ExtractElement: {
1042 bool Reuse = CanReuseExtract(VL);
1044 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
1046 BS.cancelScheduling(VL);
1048 newTreeEntry(VL, Reuse);
1051 case Instruction::Load: {
1052 // Check if the loads are consecutive or of we need to swizzle them.
1053 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
1054 LoadInst *L = cast<LoadInst>(VL[i]);
1055 if (!L->isSimple()) {
1056 BS.cancelScheduling(VL);
1057 newTreeEntry(VL, false);
1058 DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n");
1061 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
1062 if (VL.size() == 2 && isConsecutiveAccess(VL[1], VL[0])) {
1063 ++NumLoadsWantToChangeOrder;
1065 BS.cancelScheduling(VL);
1066 newTreeEntry(VL, false);
1067 DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n");
1071 ++NumLoadsWantToKeepOrder;
1072 newTreeEntry(VL, true);
1073 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
1076 case Instruction::ZExt:
1077 case Instruction::SExt:
1078 case Instruction::FPToUI:
1079 case Instruction::FPToSI:
1080 case Instruction::FPExt:
1081 case Instruction::PtrToInt:
1082 case Instruction::IntToPtr:
1083 case Instruction::SIToFP:
1084 case Instruction::UIToFP:
1085 case Instruction::Trunc:
1086 case Instruction::FPTrunc:
1087 case Instruction::BitCast: {
1088 Type *SrcTy = VL0->getOperand(0)->getType();
1089 for (unsigned i = 0; i < VL.size(); ++i) {
1090 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
1091 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
1092 BS.cancelScheduling(VL);
1093 newTreeEntry(VL, false);
1094 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
1098 newTreeEntry(VL, true);
1099 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
1101 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1103 // Prepare the operand vector.
1104 for (unsigned j = 0; j < VL.size(); ++j)
1105 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1107 buildTree_rec(Operands, Depth+1);
1111 case Instruction::ICmp:
1112 case Instruction::FCmp: {
1113 // Check that all of the compares have the same predicate.
1114 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1115 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
1116 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
1117 CmpInst *Cmp = cast<CmpInst>(VL[i]);
1118 if (Cmp->getPredicate() != P0 ||
1119 Cmp->getOperand(0)->getType() != ComparedTy) {
1120 BS.cancelScheduling(VL);
1121 newTreeEntry(VL, false);
1122 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
1127 newTreeEntry(VL, true);
1128 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
1130 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1132 // Prepare the operand vector.
1133 for (unsigned j = 0; j < VL.size(); ++j)
1134 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1136 buildTree_rec(Operands, Depth+1);
1140 case Instruction::Select:
1141 case Instruction::Add:
1142 case Instruction::FAdd:
1143 case Instruction::Sub:
1144 case Instruction::FSub:
1145 case Instruction::Mul:
1146 case Instruction::FMul:
1147 case Instruction::UDiv:
1148 case Instruction::SDiv:
1149 case Instruction::FDiv:
1150 case Instruction::URem:
1151 case Instruction::SRem:
1152 case Instruction::FRem:
1153 case Instruction::Shl:
1154 case Instruction::LShr:
1155 case Instruction::AShr:
1156 case Instruction::And:
1157 case Instruction::Or:
1158 case Instruction::Xor: {
1159 newTreeEntry(VL, true);
1160 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
1162 // Sort operands of the instructions so that each side is more likely to
1163 // have the same opcode.
1164 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
1165 ValueList Left, Right;
1166 reorderInputsAccordingToOpcode(VL, Left, Right);
1167 buildTree_rec(Left, Depth + 1);
1168 buildTree_rec(Right, Depth + 1);
1172 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1174 // Prepare the operand vector.
1175 for (unsigned j = 0; j < VL.size(); ++j)
1176 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1178 buildTree_rec(Operands, Depth+1);
1182 case Instruction::GetElementPtr: {
1183 // We don't combine GEPs with complicated (nested) indexing.
1184 for (unsigned j = 0; j < VL.size(); ++j) {
1185 if (cast<Instruction>(VL[j])->getNumOperands() != 2) {
1186 DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n");
1187 BS.cancelScheduling(VL);
1188 newTreeEntry(VL, false);
1193 // We combine only GEPs with a single use.
1194 for (unsigned j = 0; j < VL.size(); ++j) {
1195 if (cast<Instruction>(VL[j])->getNumUses() > 1) {
1196 DEBUG(dbgs() << "SLP: not-vectorizable GEP (multiple uses).\n");
1197 BS.cancelScheduling(VL);
1198 newTreeEntry(VL, false);
1203 // We can't combine several GEPs into one vector if they operate on
1205 Type *Ty0 = cast<Instruction>(VL0)->getOperand(0)->getType();
1206 for (unsigned j = 0; j < VL.size(); ++j) {
1207 Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType();
1209 DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n");
1210 BS.cancelScheduling(VL);
1211 newTreeEntry(VL, false);
1216 // We don't combine GEPs with non-constant indexes.
1217 for (unsigned j = 0; j < VL.size(); ++j) {
1218 auto Op = cast<Instruction>(VL[j])->getOperand(1);
1219 if (!isa<ConstantInt>(Op)) {
1221 dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n");
1222 BS.cancelScheduling(VL);
1223 newTreeEntry(VL, false);
1228 newTreeEntry(VL, true);
1229 DEBUG(dbgs() << "SLP: added a vector of GEPs.\n");
1230 for (unsigned i = 0, e = 2; i < e; ++i) {
1232 // Prepare the operand vector.
1233 for (unsigned j = 0; j < VL.size(); ++j)
1234 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1236 buildTree_rec(Operands, Depth + 1);
1240 case Instruction::Store: {
1241 // Check if the stores are consecutive or of we need to swizzle them.
1242 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
1243 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
1244 BS.cancelScheduling(VL);
1245 newTreeEntry(VL, false);
1246 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
1250 newTreeEntry(VL, true);
1251 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
1254 for (unsigned j = 0; j < VL.size(); ++j)
1255 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
1257 buildTree_rec(Operands, Depth + 1);
1260 case Instruction::Call: {
1261 // Check if the calls are all to the same vectorizable intrinsic.
1262 CallInst *CI = cast<CallInst>(VL[0]);
1263 // Check if this is an Intrinsic call or something that can be
1264 // represented by an intrinsic call
1265 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1266 if (!isTriviallyVectorizable(ID)) {
1267 BS.cancelScheduling(VL);
1268 newTreeEntry(VL, false);
1269 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
1272 Function *Int = CI->getCalledFunction();
1273 Value *A1I = nullptr;
1274 if (hasVectorInstrinsicScalarOpd(ID, 1))
1275 A1I = CI->getArgOperand(1);
1276 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
1277 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
1278 if (!CI2 || CI2->getCalledFunction() != Int ||
1279 getIntrinsicIDForCall(CI2, TLI) != ID) {
1280 BS.cancelScheduling(VL);
1281 newTreeEntry(VL, false);
1282 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]
1286 // ctlz,cttz and powi are special intrinsics whose second argument
1287 // should be same in order for them to be vectorized.
1288 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
1289 Value *A1J = CI2->getArgOperand(1);
1291 BS.cancelScheduling(VL);
1292 newTreeEntry(VL, false);
1293 DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI
1294 << " argument "<< A1I<<"!=" << A1J
1301 newTreeEntry(VL, true);
1302 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
1304 // Prepare the operand vector.
1305 for (unsigned j = 0; j < VL.size(); ++j) {
1306 CallInst *CI2 = dyn_cast<CallInst>(VL[j]);
1307 Operands.push_back(CI2->getArgOperand(i));
1309 buildTree_rec(Operands, Depth + 1);
1313 case Instruction::ShuffleVector: {
1314 // If this is not an alternate sequence of opcode like add-sub
1315 // then do not vectorize this instruction.
1316 if (!isAltShuffle) {
1317 BS.cancelScheduling(VL);
1318 newTreeEntry(VL, false);
1319 DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n");
1322 newTreeEntry(VL, true);
1323 DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n");
1324 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1326 // Prepare the operand vector.
1327 for (unsigned j = 0; j < VL.size(); ++j)
1328 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1330 buildTree_rec(Operands, Depth + 1);
1335 BS.cancelScheduling(VL);
1336 newTreeEntry(VL, false);
1337 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
1342 int BoUpSLP::getEntryCost(TreeEntry *E) {
1343 ArrayRef<Value*> VL = E->Scalars;
1345 Type *ScalarTy = VL[0]->getType();
1346 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1347 ScalarTy = SI->getValueOperand()->getType();
1348 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1350 if (E->NeedToGather) {
1351 if (allConstant(VL))
1354 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1356 return getGatherCost(E->Scalars);
1358 unsigned Opcode = getSameOpcode(VL);
1359 assert(Opcode && getSameType(VL) && getSameBlock(VL) && "Invalid VL");
1360 Instruction *VL0 = cast<Instruction>(VL[0]);
1362 case Instruction::PHI: {
1365 case Instruction::ExtractElement: {
1366 if (CanReuseExtract(VL)) {
1368 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1369 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1371 // Take credit for instruction that will become dead.
1373 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1377 return getGatherCost(VecTy);
1379 case Instruction::ZExt:
1380 case Instruction::SExt:
1381 case Instruction::FPToUI:
1382 case Instruction::FPToSI:
1383 case Instruction::FPExt:
1384 case Instruction::PtrToInt:
1385 case Instruction::IntToPtr:
1386 case Instruction::SIToFP:
1387 case Instruction::UIToFP:
1388 case Instruction::Trunc:
1389 case Instruction::FPTrunc:
1390 case Instruction::BitCast: {
1391 Type *SrcTy = VL0->getOperand(0)->getType();
1393 // Calculate the cost of this instruction.
1394 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1395 VL0->getType(), SrcTy);
1397 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1398 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1399 return VecCost - ScalarCost;
1401 case Instruction::FCmp:
1402 case Instruction::ICmp:
1403 case Instruction::Select:
1404 case Instruction::Add:
1405 case Instruction::FAdd:
1406 case Instruction::Sub:
1407 case Instruction::FSub:
1408 case Instruction::Mul:
1409 case Instruction::FMul:
1410 case Instruction::UDiv:
1411 case Instruction::SDiv:
1412 case Instruction::FDiv:
1413 case Instruction::URem:
1414 case Instruction::SRem:
1415 case Instruction::FRem:
1416 case Instruction::Shl:
1417 case Instruction::LShr:
1418 case Instruction::AShr:
1419 case Instruction::And:
1420 case Instruction::Or:
1421 case Instruction::Xor: {
1422 // Calculate the cost of this instruction.
1425 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1426 Opcode == Instruction::Select) {
1427 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1428 ScalarCost = VecTy->getNumElements() *
1429 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1430 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1432 // Certain instructions can be cheaper to vectorize if they have a
1433 // constant second vector operand.
1434 TargetTransformInfo::OperandValueKind Op1VK =
1435 TargetTransformInfo::OK_AnyValue;
1436 TargetTransformInfo::OperandValueKind Op2VK =
1437 TargetTransformInfo::OK_UniformConstantValue;
1438 TargetTransformInfo::OperandValueProperties Op1VP =
1439 TargetTransformInfo::OP_None;
1440 TargetTransformInfo::OperandValueProperties Op2VP =
1441 TargetTransformInfo::OP_None;
1443 // If all operands are exactly the same ConstantInt then set the
1444 // operand kind to OK_UniformConstantValue.
1445 // If instead not all operands are constants, then set the operand kind
1446 // to OK_AnyValue. If all operands are constants but not the same,
1447 // then set the operand kind to OK_NonUniformConstantValue.
1448 ConstantInt *CInt = nullptr;
1449 for (unsigned i = 0; i < VL.size(); ++i) {
1450 const Instruction *I = cast<Instruction>(VL[i]);
1451 if (!isa<ConstantInt>(I->getOperand(1))) {
1452 Op2VK = TargetTransformInfo::OK_AnyValue;
1456 CInt = cast<ConstantInt>(I->getOperand(1));
1459 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1460 CInt != cast<ConstantInt>(I->getOperand(1)))
1461 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1463 // FIXME: Currently cost of model modification for division by
1464 // power of 2 is handled only for X86. Add support for other targets.
1465 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue && CInt &&
1466 CInt->getValue().isPowerOf2())
1467 Op2VP = TargetTransformInfo::OP_PowerOf2;
1469 ScalarCost = VecTy->getNumElements() *
1470 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK,
1472 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK,
1475 return VecCost - ScalarCost;
1477 case Instruction::GetElementPtr: {
1478 TargetTransformInfo::OperandValueKind Op1VK =
1479 TargetTransformInfo::OK_AnyValue;
1480 TargetTransformInfo::OperandValueKind Op2VK =
1481 TargetTransformInfo::OK_UniformConstantValue;
1484 VecTy->getNumElements() *
1485 TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK);
1487 TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK);
1489 return VecCost - ScalarCost;
1491 case Instruction::Load: {
1492 // Cost of wide load - cost of scalar loads.
1493 int ScalarLdCost = VecTy->getNumElements() *
1494 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1495 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1496 return VecLdCost - ScalarLdCost;
1498 case Instruction::Store: {
1499 // We know that we can merge the stores. Calculate the cost.
1500 int ScalarStCost = VecTy->getNumElements() *
1501 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1502 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1503 return VecStCost - ScalarStCost;
1505 case Instruction::Call: {
1506 CallInst *CI = cast<CallInst>(VL0);
1507 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1509 // Calculate the cost of the scalar and vector calls.
1510 SmallVector<Type*, 4> ScalarTys, VecTys;
1511 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1512 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1513 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1514 VecTy->getNumElements()));
1517 int ScalarCallCost = VecTy->getNumElements() *
1518 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1520 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1522 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1523 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1524 << " for " << *CI << "\n");
1526 return VecCallCost - ScalarCallCost;
1528 case Instruction::ShuffleVector: {
1529 TargetTransformInfo::OperandValueKind Op1VK =
1530 TargetTransformInfo::OK_AnyValue;
1531 TargetTransformInfo::OperandValueKind Op2VK =
1532 TargetTransformInfo::OK_AnyValue;
1535 for (unsigned i = 0; i < VL.size(); ++i) {
1536 Instruction *I = cast<Instruction>(VL[i]);
1540 TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK);
1542 // VecCost is equal to sum of the cost of creating 2 vectors
1543 // and the cost of creating shuffle.
1544 Instruction *I0 = cast<Instruction>(VL[0]);
1546 TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK);
1547 Instruction *I1 = cast<Instruction>(VL[1]);
1549 TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK);
1551 TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0);
1552 return VecCost - ScalarCost;
1555 llvm_unreachable("Unknown instruction");
1559 bool BoUpSLP::isFullyVectorizableTinyTree() {
1560 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1561 VectorizableTree.size() << " is fully vectorizable .\n");
1563 // We only handle trees of height 2.
1564 if (VectorizableTree.size() != 2)
1567 // Handle splat stores.
1568 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1571 // Gathering cost would be too much for tiny trees.
1572 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1578 int BoUpSLP::getSpillCost() {
1579 // Walk from the bottom of the tree to the top, tracking which values are
1580 // live. When we see a call instruction that is not part of our tree,
1581 // query TTI to see if there is a cost to keeping values live over it
1582 // (for example, if spills and fills are required).
1583 unsigned BundleWidth = VectorizableTree.front().Scalars.size();
1586 SmallPtrSet<Instruction*, 4> LiveValues;
1587 Instruction *PrevInst = nullptr;
1589 for (unsigned N = 0; N < VectorizableTree.size(); ++N) {
1590 Instruction *Inst = dyn_cast<Instruction>(VectorizableTree[N].Scalars[0]);
1600 dbgs() << "SLP: #LV: " << LiveValues.size();
1601 for (auto *X : LiveValues)
1602 dbgs() << " " << X->getName();
1603 dbgs() << ", Looking at ";
1607 // Update LiveValues.
1608 LiveValues.erase(PrevInst);
1609 for (auto &J : PrevInst->operands()) {
1610 if (isa<Instruction>(&*J) && ScalarToTreeEntry.count(&*J))
1611 LiveValues.insert(cast<Instruction>(&*J));
1614 // Now find the sequence of instructions between PrevInst and Inst.
1615 BasicBlock::reverse_iterator InstIt(Inst), PrevInstIt(PrevInst);
1617 while (InstIt != PrevInstIt) {
1618 if (PrevInstIt == PrevInst->getParent()->rend()) {
1619 PrevInstIt = Inst->getParent()->rbegin();
1623 if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) {
1624 SmallVector<Type*, 4> V;
1625 for (auto *II : LiveValues)
1626 V.push_back(VectorType::get(II->getType(), BundleWidth));
1627 Cost += TTI->getCostOfKeepingLiveOverCall(V);
1636 DEBUG(dbgs() << "SLP: SpillCost=" << Cost << "\n");
1640 int BoUpSLP::getTreeCost() {
1642 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1643 VectorizableTree.size() << ".\n");
1645 // We only vectorize tiny trees if it is fully vectorizable.
1646 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1647 if (!VectorizableTree.size()) {
1648 assert(!ExternalUses.size() && "We should not have any external users");
1653 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1655 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1656 int C = getEntryCost(&VectorizableTree[i]);
1657 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1658 << *VectorizableTree[i].Scalars[0] << " .\n");
1662 SmallSet<Value *, 16> ExtractCostCalculated;
1663 int ExtractCost = 0;
1664 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1666 // We only add extract cost once for the same scalar.
1667 if (!ExtractCostCalculated.insert(I->Scalar))
1670 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1671 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1675 Cost += getSpillCost();
1677 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1678 return Cost + ExtractCost;
1681 int BoUpSLP::getGatherCost(Type *Ty) {
1683 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1684 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1688 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1689 // Find the type of the operands in VL.
1690 Type *ScalarTy = VL[0]->getType();
1691 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1692 ScalarTy = SI->getValueOperand()->getType();
1693 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1694 // Find the cost of inserting/extracting values from the vector.
1695 return getGatherCost(VecTy);
1698 Value *BoUpSLP::getPointerOperand(Value *I) {
1699 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1700 return LI->getPointerOperand();
1701 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1702 return SI->getPointerOperand();
1706 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1707 if (LoadInst *L = dyn_cast<LoadInst>(I))
1708 return L->getPointerAddressSpace();
1709 if (StoreInst *S = dyn_cast<StoreInst>(I))
1710 return S->getPointerAddressSpace();
1714 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1715 Value *PtrA = getPointerOperand(A);
1716 Value *PtrB = getPointerOperand(B);
1717 unsigned ASA = getAddressSpaceOperand(A);
1718 unsigned ASB = getAddressSpaceOperand(B);
1720 // Check that the address spaces match and that the pointers are valid.
1721 if (!PtrA || !PtrB || (ASA != ASB))
1724 // Make sure that A and B are different pointers of the same type.
1725 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1728 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1729 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1730 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1732 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1733 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1734 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1736 APInt OffsetDelta = OffsetB - OffsetA;
1738 // Check if they are based on the same pointer. That makes the offsets
1741 return OffsetDelta == Size;
1743 // Compute the necessary base pointer delta to have the necessary final delta
1744 // equal to the size.
1745 APInt BaseDelta = Size - OffsetDelta;
1747 // Otherwise compute the distance with SCEV between the base pointers.
1748 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1749 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1750 const SCEV *C = SE->getConstant(BaseDelta);
1751 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1752 return X == PtrSCEVB;
1755 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1756 Instruction *VL0 = cast<Instruction>(VL[0]);
1757 BasicBlock::iterator NextInst = VL0;
1759 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1760 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1763 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1764 Value *Vec = UndefValue::get(Ty);
1765 // Generate the 'InsertElement' instruction.
1766 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1767 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1768 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1769 GatherSeq.insert(Insrt);
1770 CSEBlocks.insert(Insrt->getParent());
1772 // Add to our 'need-to-extract' list.
1773 if (ScalarToTreeEntry.count(VL[i])) {
1774 int Idx = ScalarToTreeEntry[VL[i]];
1775 TreeEntry *E = &VectorizableTree[Idx];
1776 // Find which lane we need to extract.
1778 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1779 // Is this the lane of the scalar that we are looking for ?
1780 if (E->Scalars[Lane] == VL[i]) {
1785 assert(FoundLane >= 0 && "Could not find the correct lane");
1786 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1794 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1795 SmallDenseMap<Value*, int>::const_iterator Entry
1796 = ScalarToTreeEntry.find(VL[0]);
1797 if (Entry != ScalarToTreeEntry.end()) {
1798 int Idx = Entry->second;
1799 const TreeEntry *En = &VectorizableTree[Idx];
1800 if (En->isSame(VL) && En->VectorizedValue)
1801 return En->VectorizedValue;
1806 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1807 if (ScalarToTreeEntry.count(VL[0])) {
1808 int Idx = ScalarToTreeEntry[VL[0]];
1809 TreeEntry *E = &VectorizableTree[Idx];
1811 return vectorizeTree(E);
1814 Type *ScalarTy = VL[0]->getType();
1815 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1816 ScalarTy = SI->getValueOperand()->getType();
1817 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1819 return Gather(VL, VecTy);
1822 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1823 IRBuilder<>::InsertPointGuard Guard(Builder);
1825 if (E->VectorizedValue) {
1826 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1827 return E->VectorizedValue;
1830 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1831 Type *ScalarTy = VL0->getType();
1832 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1833 ScalarTy = SI->getValueOperand()->getType();
1834 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1836 if (E->NeedToGather) {
1837 setInsertPointAfterBundle(E->Scalars);
1838 return Gather(E->Scalars, VecTy);
1841 unsigned Opcode = getSameOpcode(E->Scalars);
1844 case Instruction::PHI: {
1845 PHINode *PH = dyn_cast<PHINode>(VL0);
1846 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1847 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1848 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1849 E->VectorizedValue = NewPhi;
1851 // PHINodes may have multiple entries from the same block. We want to
1852 // visit every block once.
1853 SmallSet<BasicBlock*, 4> VisitedBBs;
1855 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1857 BasicBlock *IBB = PH->getIncomingBlock(i);
1859 if (!VisitedBBs.insert(IBB)) {
1860 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1864 // Prepare the operand vector.
1865 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1866 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1867 getIncomingValueForBlock(IBB));
1869 Builder.SetInsertPoint(IBB->getTerminator());
1870 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1871 Value *Vec = vectorizeTree(Operands);
1872 NewPhi->addIncoming(Vec, IBB);
1875 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1876 "Invalid number of incoming values");
1880 case Instruction::ExtractElement: {
1881 if (CanReuseExtract(E->Scalars)) {
1882 Value *V = VL0->getOperand(0);
1883 E->VectorizedValue = V;
1886 return Gather(E->Scalars, VecTy);
1888 case Instruction::ZExt:
1889 case Instruction::SExt:
1890 case Instruction::FPToUI:
1891 case Instruction::FPToSI:
1892 case Instruction::FPExt:
1893 case Instruction::PtrToInt:
1894 case Instruction::IntToPtr:
1895 case Instruction::SIToFP:
1896 case Instruction::UIToFP:
1897 case Instruction::Trunc:
1898 case Instruction::FPTrunc:
1899 case Instruction::BitCast: {
1901 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1902 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1904 setInsertPointAfterBundle(E->Scalars);
1906 Value *InVec = vectorizeTree(INVL);
1908 if (Value *V = alreadyVectorized(E->Scalars))
1911 CastInst *CI = dyn_cast<CastInst>(VL0);
1912 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1913 E->VectorizedValue = V;
1914 ++NumVectorInstructions;
1917 case Instruction::FCmp:
1918 case Instruction::ICmp: {
1919 ValueList LHSV, RHSV;
1920 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1921 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1922 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1925 setInsertPointAfterBundle(E->Scalars);
1927 Value *L = vectorizeTree(LHSV);
1928 Value *R = vectorizeTree(RHSV);
1930 if (Value *V = alreadyVectorized(E->Scalars))
1933 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1935 if (Opcode == Instruction::FCmp)
1936 V = Builder.CreateFCmp(P0, L, R);
1938 V = Builder.CreateICmp(P0, L, R);
1940 E->VectorizedValue = V;
1941 ++NumVectorInstructions;
1944 case Instruction::Select: {
1945 ValueList TrueVec, FalseVec, CondVec;
1946 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1947 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1948 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1949 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1952 setInsertPointAfterBundle(E->Scalars);
1954 Value *Cond = vectorizeTree(CondVec);
1955 Value *True = vectorizeTree(TrueVec);
1956 Value *False = vectorizeTree(FalseVec);
1958 if (Value *V = alreadyVectorized(E->Scalars))
1961 Value *V = Builder.CreateSelect(Cond, True, False);
1962 E->VectorizedValue = V;
1963 ++NumVectorInstructions;
1966 case Instruction::Add:
1967 case Instruction::FAdd:
1968 case Instruction::Sub:
1969 case Instruction::FSub:
1970 case Instruction::Mul:
1971 case Instruction::FMul:
1972 case Instruction::UDiv:
1973 case Instruction::SDiv:
1974 case Instruction::FDiv:
1975 case Instruction::URem:
1976 case Instruction::SRem:
1977 case Instruction::FRem:
1978 case Instruction::Shl:
1979 case Instruction::LShr:
1980 case Instruction::AShr:
1981 case Instruction::And:
1982 case Instruction::Or:
1983 case Instruction::Xor: {
1984 ValueList LHSVL, RHSVL;
1985 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1986 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1988 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1989 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1990 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1993 setInsertPointAfterBundle(E->Scalars);
1995 Value *LHS = vectorizeTree(LHSVL);
1996 Value *RHS = vectorizeTree(RHSVL);
1998 if (LHS == RHS && isa<Instruction>(LHS)) {
1999 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
2002 if (Value *V = alreadyVectorized(E->Scalars))
2005 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
2006 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
2007 E->VectorizedValue = V;
2008 ++NumVectorInstructions;
2010 if (Instruction *I = dyn_cast<Instruction>(V))
2011 return propagateMetadata(I, E->Scalars);
2015 case Instruction::Load: {
2016 // Loads are inserted at the head of the tree because we don't want to
2017 // sink them all the way down past store instructions.
2018 setInsertPointAfterBundle(E->Scalars);
2020 LoadInst *LI = cast<LoadInst>(VL0);
2021 Type *ScalarLoadTy = LI->getType();
2022 unsigned AS = LI->getPointerAddressSpace();
2024 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
2025 VecTy->getPointerTo(AS));
2026 unsigned Alignment = LI->getAlignment();
2027 LI = Builder.CreateLoad(VecPtr);
2029 Alignment = DL->getABITypeAlignment(ScalarLoadTy);
2030 LI->setAlignment(Alignment);
2031 E->VectorizedValue = LI;
2032 ++NumVectorInstructions;
2033 return propagateMetadata(LI, E->Scalars);
2035 case Instruction::Store: {
2036 StoreInst *SI = cast<StoreInst>(VL0);
2037 unsigned Alignment = SI->getAlignment();
2038 unsigned AS = SI->getPointerAddressSpace();
2041 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2042 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
2044 setInsertPointAfterBundle(E->Scalars);
2046 Value *VecValue = vectorizeTree(ValueOp);
2047 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
2048 VecTy->getPointerTo(AS));
2049 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
2051 Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType());
2052 S->setAlignment(Alignment);
2053 E->VectorizedValue = S;
2054 ++NumVectorInstructions;
2055 return propagateMetadata(S, E->Scalars);
2057 case Instruction::GetElementPtr: {
2058 setInsertPointAfterBundle(E->Scalars);
2061 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2062 Op0VL.push_back(cast<GetElementPtrInst>(E->Scalars[i])->getOperand(0));
2064 Value *Op0 = vectorizeTree(Op0VL);
2066 std::vector<Value *> OpVecs;
2067 for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e;
2070 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2071 OpVL.push_back(cast<GetElementPtrInst>(E->Scalars[i])->getOperand(j));
2073 Value *OpVec = vectorizeTree(OpVL);
2074 OpVecs.push_back(OpVec);
2077 Value *V = Builder.CreateGEP(Op0, OpVecs);
2078 E->VectorizedValue = V;
2079 ++NumVectorInstructions;
2081 if (Instruction *I = dyn_cast<Instruction>(V))
2082 return propagateMetadata(I, E->Scalars);
2086 case Instruction::Call: {
2087 CallInst *CI = cast<CallInst>(VL0);
2088 setInsertPointAfterBundle(E->Scalars);
2090 Intrinsic::ID IID = Intrinsic::not_intrinsic;
2091 if (CI && (FI = CI->getCalledFunction())) {
2092 IID = (Intrinsic::ID) FI->getIntrinsicID();
2094 std::vector<Value *> OpVecs;
2095 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
2097 // ctlz,cttz and powi are special intrinsics whose second argument is
2098 // a scalar. This argument should not be vectorized.
2099 if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) {
2100 CallInst *CEI = cast<CallInst>(E->Scalars[0]);
2101 OpVecs.push_back(CEI->getArgOperand(j));
2104 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
2105 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
2106 OpVL.push_back(CEI->getArgOperand(j));
2109 Value *OpVec = vectorizeTree(OpVL);
2110 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
2111 OpVecs.push_back(OpVec);
2114 Module *M = F->getParent();
2115 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2116 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
2117 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
2118 Value *V = Builder.CreateCall(CF, OpVecs);
2119 E->VectorizedValue = V;
2120 ++NumVectorInstructions;
2123 case Instruction::ShuffleVector: {
2124 ValueList LHSVL, RHSVL;
2125 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
2126 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
2127 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
2129 setInsertPointAfterBundle(E->Scalars);
2131 Value *LHS = vectorizeTree(LHSVL);
2132 Value *RHS = vectorizeTree(RHSVL);
2134 if (Value *V = alreadyVectorized(E->Scalars))
2137 // Create a vector of LHS op1 RHS
2138 BinaryOperator *BinOp0 = cast<BinaryOperator>(VL0);
2139 Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS);
2141 // Create a vector of LHS op2 RHS
2142 Instruction *VL1 = cast<Instruction>(E->Scalars[1]);
2143 BinaryOperator *BinOp1 = cast<BinaryOperator>(VL1);
2144 Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS);
2146 // Create appropriate shuffle to take alternative operations from
2148 std::vector<Constant *> Mask(E->Scalars.size());
2149 unsigned e = E->Scalars.size();
2150 for (unsigned i = 0; i < e; ++i) {
2152 Mask[i] = Builder.getInt32(e + i);
2154 Mask[i] = Builder.getInt32(i);
2157 Value *ShuffleMask = ConstantVector::get(Mask);
2159 Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2160 E->VectorizedValue = V;
2161 ++NumVectorInstructions;
2162 if (Instruction *I = dyn_cast<Instruction>(V))
2163 return propagateMetadata(I, E->Scalars);
2168 llvm_unreachable("unknown inst");
2173 Value *BoUpSLP::vectorizeTree() {
2175 // All blocks must be scheduled before any instructions are inserted.
2176 for (auto &BSIter : BlocksSchedules) {
2177 scheduleBlock(BSIter.second.get());
2180 Builder.SetInsertPoint(F->getEntryBlock().begin());
2181 vectorizeTree(&VectorizableTree[0]);
2183 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
2185 // Extract all of the elements with the external uses.
2186 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
2188 Value *Scalar = it->Scalar;
2189 llvm::User *User = it->User;
2191 // Skip users that we already RAUW. This happens when one instruction
2192 // has multiple uses of the same value.
2193 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
2196 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
2198 int Idx = ScalarToTreeEntry[Scalar];
2199 TreeEntry *E = &VectorizableTree[Idx];
2200 assert(!E->NeedToGather && "Extracting from a gather list");
2202 Value *Vec = E->VectorizedValue;
2203 assert(Vec && "Can't find vectorizable value");
2205 Value *Lane = Builder.getInt32(it->Lane);
2206 // Generate extracts for out-of-tree users.
2207 // Find the insertion point for the extractelement lane.
2208 if (isa<Instruction>(Vec)){
2209 if (PHINode *PH = dyn_cast<PHINode>(User)) {
2210 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
2211 if (PH->getIncomingValue(i) == Scalar) {
2212 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
2213 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2214 CSEBlocks.insert(PH->getIncomingBlock(i));
2215 PH->setOperand(i, Ex);
2219 Builder.SetInsertPoint(cast<Instruction>(User));
2220 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2221 CSEBlocks.insert(cast<Instruction>(User)->getParent());
2222 User->replaceUsesOfWith(Scalar, Ex);
2225 Builder.SetInsertPoint(F->getEntryBlock().begin());
2226 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2227 CSEBlocks.insert(&F->getEntryBlock());
2228 User->replaceUsesOfWith(Scalar, Ex);
2231 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
2234 // For each vectorized value:
2235 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
2236 TreeEntry *Entry = &VectorizableTree[EIdx];
2239 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
2240 Value *Scalar = Entry->Scalars[Lane];
2241 // No need to handle users of gathered values.
2242 if (Entry->NeedToGather)
2245 assert(Entry->VectorizedValue && "Can't find vectorizable value");
2247 Type *Ty = Scalar->getType();
2248 if (!Ty->isVoidTy()) {
2250 for (User *U : Scalar->users()) {
2251 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
2253 assert((ScalarToTreeEntry.count(U) ||
2254 // It is legal to replace users in the ignorelist by undef.
2255 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
2256 UserIgnoreList.end())) &&
2257 "Replacing out-of-tree value with undef");
2260 Value *Undef = UndefValue::get(Ty);
2261 Scalar->replaceAllUsesWith(Undef);
2263 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
2264 cast<Instruction>(Scalar)->eraseFromParent();
2268 Builder.ClearInsertionPoint();
2270 return VectorizableTree[0].VectorizedValue;
2273 void BoUpSLP::optimizeGatherSequence() {
2274 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
2275 << " gather sequences instructions.\n");
2276 // LICM InsertElementInst sequences.
2277 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
2278 e = GatherSeq.end(); it != e; ++it) {
2279 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
2284 // Check if this block is inside a loop.
2285 Loop *L = LI->getLoopFor(Insert->getParent());
2289 // Check if it has a preheader.
2290 BasicBlock *PreHeader = L->getLoopPreheader();
2294 // If the vector or the element that we insert into it are
2295 // instructions that are defined in this basic block then we can't
2296 // hoist this instruction.
2297 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
2298 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
2299 if (CurrVec && L->contains(CurrVec))
2301 if (NewElem && L->contains(NewElem))
2304 // We can hoist this instruction. Move it to the pre-header.
2305 Insert->moveBefore(PreHeader->getTerminator());
2308 // Make a list of all reachable blocks in our CSE queue.
2309 SmallVector<const DomTreeNode *, 8> CSEWorkList;
2310 CSEWorkList.reserve(CSEBlocks.size());
2311 for (BasicBlock *BB : CSEBlocks)
2312 if (DomTreeNode *N = DT->getNode(BB)) {
2313 assert(DT->isReachableFromEntry(N));
2314 CSEWorkList.push_back(N);
2317 // Sort blocks by domination. This ensures we visit a block after all blocks
2318 // dominating it are visited.
2319 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
2320 [this](const DomTreeNode *A, const DomTreeNode *B) {
2321 return DT->properlyDominates(A, B);
2324 // Perform O(N^2) search over the gather sequences and merge identical
2325 // instructions. TODO: We can further optimize this scan if we split the
2326 // instructions into different buckets based on the insert lane.
2327 SmallVector<Instruction *, 16> Visited;
2328 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
2329 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
2330 "Worklist not sorted properly!");
2331 BasicBlock *BB = (*I)->getBlock();
2332 // For all instructions in blocks containing gather sequences:
2333 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
2334 Instruction *In = it++;
2335 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
2338 // Check if we can replace this instruction with any of the
2339 // visited instructions.
2340 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
2343 if (In->isIdenticalTo(*v) &&
2344 DT->dominates((*v)->getParent(), In->getParent())) {
2345 In->replaceAllUsesWith(*v);
2346 In->eraseFromParent();
2352 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
2353 Visited.push_back(In);
2361 // Groups the instructions to a bundle (which is then a single scheduling entity)
2362 // and schedules instructions until the bundle gets ready.
2363 bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL,
2364 AliasAnalysis *AA) {
2365 if (isa<PHINode>(VL[0]))
2368 // Initialize the instruction bundle.
2369 Instruction *OldScheduleEnd = ScheduleEnd;
2370 ScheduleData *PrevInBundle = nullptr;
2371 ScheduleData *Bundle = nullptr;
2372 bool ReSchedule = false;
2373 DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n");
2374 for (Value *V : VL) {
2375 extendSchedulingRegion(V);
2376 ScheduleData *BundleMember = getScheduleData(V);
2377 assert(BundleMember &&
2378 "no ScheduleData for bundle member (maybe not in same basic block)");
2379 if (BundleMember->IsScheduled) {
2380 // A bundle member was scheduled as single instruction before and now
2381 // needs to be scheduled as part of the bundle. We just get rid of the
2382 // existing schedule.
2383 DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMember
2384 << " was already scheduled\n");
2387 assert(BundleMember->isSchedulingEntity() &&
2388 "bundle member already part of other bundle");
2390 PrevInBundle->NextInBundle = BundleMember;
2392 Bundle = BundleMember;
2394 BundleMember->UnscheduledDepsInBundle = 0;
2395 Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps;
2397 // Group the instructions to a bundle.
2398 BundleMember->FirstInBundle = Bundle;
2399 PrevInBundle = BundleMember;
2401 if (ScheduleEnd != OldScheduleEnd) {
2402 // The scheduling region got new instructions at the lower end (or it is a
2403 // new region for the first bundle). This makes it necessary to
2404 // recalculate all dependencies.
2405 // It is seldom that this needs to be done a second time after adding the
2406 // initial bundle to the region.
2407 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
2408 ScheduleData *SD = getScheduleData(I);
2409 SD->clearDependencies();
2415 initialFillReadyList(ReadyInsts);
2418 DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block "
2419 << BB->getName() << "\n");
2421 calculateDependencies(Bundle, true, AA);
2423 // Now try to schedule the new bundle. As soon as the bundle is "ready" it
2424 // means that there are no cyclic dependencies and we can schedule it.
2425 // Note that's important that we don't "schedule" the bundle yet (see
2426 // cancelScheduling).
2427 while (!Bundle->isReady() && !ReadyInsts.empty()) {
2429 ScheduleData *pickedSD = ReadyInsts.back();
2430 ReadyInsts.pop_back();
2432 if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) {
2433 schedule(pickedSD, ReadyInsts);
2436 return Bundle->isReady();
2439 void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL) {
2440 if (isa<PHINode>(VL[0]))
2443 ScheduleData *Bundle = getScheduleData(VL[0]);
2444 DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n");
2445 assert(!Bundle->IsScheduled &&
2446 "Can't cancel bundle which is already scheduled");
2447 assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&
2448 "tried to unbundle something which is not a bundle");
2450 // Un-bundle: make single instructions out of the bundle.
2451 ScheduleData *BundleMember = Bundle;
2452 while (BundleMember) {
2453 assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links");
2454 BundleMember->FirstInBundle = BundleMember;
2455 ScheduleData *Next = BundleMember->NextInBundle;
2456 BundleMember->NextInBundle = nullptr;
2457 BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps;
2458 if (BundleMember->UnscheduledDepsInBundle == 0) {
2459 ReadyInsts.insert(BundleMember);
2461 BundleMember = Next;
2465 void BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) {
2466 if (getScheduleData(V))
2468 Instruction *I = dyn_cast<Instruction>(V);
2469 assert(I && "bundle member must be an instruction");
2470 assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled");
2471 if (!ScheduleStart) {
2472 // It's the first instruction in the new region.
2473 initScheduleData(I, I->getNextNode(), nullptr, nullptr);
2475 ScheduleEnd = I->getNextNode();
2476 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?");
2477 DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n");
2480 // Search up and down at the same time, because we don't know if the new
2481 // instruction is above or below the existing scheduling region.
2482 BasicBlock::reverse_iterator UpIter(ScheduleStart);
2483 BasicBlock::reverse_iterator UpperEnd = BB->rend();
2484 BasicBlock::iterator DownIter(ScheduleEnd);
2485 BasicBlock::iterator LowerEnd = BB->end();
2487 if (UpIter != UpperEnd) {
2488 if (&*UpIter == I) {
2489 initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion);
2491 DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n");
2496 if (DownIter != LowerEnd) {
2497 if (&*DownIter == I) {
2498 initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion,
2500 ScheduleEnd = I->getNextNode();
2501 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?");
2502 DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n");
2507 assert((UpIter != UpperEnd || DownIter != LowerEnd) &&
2508 "instruction not found in block");
2512 void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI,
2514 ScheduleData *PrevLoadStore,
2515 ScheduleData *NextLoadStore) {
2516 ScheduleData *CurrentLoadStore = PrevLoadStore;
2517 for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) {
2518 ScheduleData *SD = ScheduleDataMap[I];
2520 // Allocate a new ScheduleData for the instruction.
2521 if (ChunkPos >= ChunkSize) {
2522 ScheduleDataChunks.push_back(
2523 llvm::make_unique<ScheduleData[]>(ChunkSize));
2526 SD = &(ScheduleDataChunks.back()[ChunkPos++]);
2527 ScheduleDataMap[I] = SD;
2530 assert(!isInSchedulingRegion(SD) &&
2531 "new ScheduleData already in scheduling region");
2532 SD->init(SchedulingRegionID);
2534 if (I->mayReadOrWriteMemory()) {
2535 // Update the linked list of memory accessing instructions.
2536 if (CurrentLoadStore) {
2537 CurrentLoadStore->NextLoadStore = SD;
2539 FirstLoadStoreInRegion = SD;
2541 CurrentLoadStore = SD;
2544 if (NextLoadStore) {
2545 if (CurrentLoadStore)
2546 CurrentLoadStore->NextLoadStore = NextLoadStore;
2548 LastLoadStoreInRegion = CurrentLoadStore;
2552 /// \returns the AA location that is being access by the instruction.
2553 static AliasAnalysis::Location getLocation(Instruction *I, AliasAnalysis *AA) {
2554 if (StoreInst *SI = dyn_cast<StoreInst>(I))
2555 return AA->getLocation(SI);
2556 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2557 return AA->getLocation(LI);
2558 return AliasAnalysis::Location();
2561 void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD,
2562 bool InsertInReadyList,
2563 AliasAnalysis *AA) {
2564 assert(SD->isSchedulingEntity());
2566 SmallVector<ScheduleData *, 10> WorkList;
2567 WorkList.push_back(SD);
2569 while (!WorkList.empty()) {
2570 ScheduleData *SD = WorkList.back();
2571 WorkList.pop_back();
2573 ScheduleData *BundleMember = SD;
2574 while (BundleMember) {
2575 assert(isInSchedulingRegion(BundleMember));
2576 if (!BundleMember->hasValidDependencies()) {
2578 DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n");
2579 BundleMember->Dependencies = 0;
2580 BundleMember->resetUnscheduledDeps();
2582 // Handle def-use chain dependencies.
2583 for (User *U : BundleMember->Inst->users()) {
2584 if (isa<Instruction>(U)) {
2585 ScheduleData *UseSD = getScheduleData(U);
2586 if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) {
2587 BundleMember->Dependencies++;
2588 ScheduleData *DestBundle = UseSD->FirstInBundle;
2589 if (!DestBundle->IsScheduled) {
2590 BundleMember->incrementUnscheduledDeps(1);
2592 if (!DestBundle->hasValidDependencies()) {
2593 WorkList.push_back(DestBundle);
2597 // I'm not sure if this can ever happen. But we need to be safe.
2598 // This lets the instruction/bundle never be scheduled and eventally
2599 // disable vectorization.
2600 BundleMember->Dependencies++;
2601 BundleMember->incrementUnscheduledDeps(1);
2605 // Handle the memory dependencies.
2606 ScheduleData *DepDest = BundleMember->NextLoadStore;
2608 AliasAnalysis::Location SrcLoc = getLocation(BundleMember->Inst, AA);
2609 bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory();
2612 assert(isInSchedulingRegion(DepDest));
2613 if (SrcMayWrite || DepDest->Inst->mayWriteToMemory()) {
2614 AliasAnalysis::Location DstLoc = getLocation(DepDest->Inst, AA);
2615 if (!SrcLoc.Ptr || !DstLoc.Ptr || AA->alias(SrcLoc, DstLoc)) {
2616 DepDest->MemoryDependencies.push_back(BundleMember);
2617 BundleMember->Dependencies++;
2618 ScheduleData *DestBundle = DepDest->FirstInBundle;
2619 if (!DestBundle->IsScheduled) {
2620 BundleMember->incrementUnscheduledDeps(1);
2622 if (!DestBundle->hasValidDependencies()) {
2623 WorkList.push_back(DestBundle);
2627 DepDest = DepDest->NextLoadStore;
2631 BundleMember = BundleMember->NextInBundle;
2633 if (InsertInReadyList && SD->isReady()) {
2634 ReadyInsts.push_back(SD);
2635 DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n");
2640 void BoUpSLP::BlockScheduling::resetSchedule() {
2641 assert(ScheduleStart &&
2642 "tried to reset schedule on block which has not been scheduled");
2643 for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
2644 ScheduleData *SD = getScheduleData(I);
2645 assert(isInSchedulingRegion(SD));
2646 SD->IsScheduled = false;
2647 SD->resetUnscheduledDeps();
2652 void BoUpSLP::scheduleBlock(BlockScheduling *BS) {
2654 if (!BS->ScheduleStart)
2657 DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n");
2659 BS->resetSchedule();
2661 // For the real scheduling we use a more sophisticated ready-list: it is
2662 // sorted by the original instruction location. This lets the final schedule
2663 // be as close as possible to the original instruction order.
2664 struct ScheduleDataCompare {
2665 bool operator()(ScheduleData *SD1, ScheduleData *SD2) {
2666 return SD2->SchedulingPriority < SD1->SchedulingPriority;
2669 std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts;
2671 // Ensure that all depencency data is updated and fill the ready-list with
2672 // initial instructions.
2674 int NumToSchedule = 0;
2675 for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd;
2676 I = I->getNextNode()) {
2677 ScheduleData *SD = BS->getScheduleData(I);
2679 SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) &&
2680 "scheduler and vectorizer have different opinion on what is a bundle");
2681 SD->FirstInBundle->SchedulingPriority = Idx++;
2682 if (SD->isSchedulingEntity()) {
2683 BS->calculateDependencies(SD, false, AA);
2687 BS->initialFillReadyList(ReadyInsts);
2689 Instruction *LastScheduledInst = BS->ScheduleEnd;
2691 // Do the "real" scheduling.
2692 while (!ReadyInsts.empty()) {
2693 ScheduleData *picked = *ReadyInsts.begin();
2694 ReadyInsts.erase(ReadyInsts.begin());
2696 // Move the scheduled instruction(s) to their dedicated places, if not
2698 ScheduleData *BundleMember = picked;
2699 while (BundleMember) {
2700 Instruction *pickedInst = BundleMember->Inst;
2701 if (LastScheduledInst->getNextNode() != pickedInst) {
2702 BS->BB->getInstList().remove(pickedInst);
2703 BS->BB->getInstList().insert(LastScheduledInst, pickedInst);
2705 LastScheduledInst = pickedInst;
2706 BundleMember = BundleMember->NextInBundle;
2709 BS->schedule(picked, ReadyInsts);
2712 assert(NumToSchedule == 0 && "could not schedule all instructions");
2714 // Avoid duplicate scheduling of the block.
2715 BS->ScheduleStart = nullptr;
2718 /// The SLPVectorizer Pass.
2719 struct SLPVectorizer : public FunctionPass {
2720 typedef SmallVector<StoreInst *, 8> StoreList;
2721 typedef MapVector<Value *, StoreList> StoreListMap;
2723 /// Pass identification, replacement for typeid
2726 explicit SLPVectorizer() : FunctionPass(ID) {
2727 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
2730 ScalarEvolution *SE;
2731 const DataLayout *DL;
2732 TargetTransformInfo *TTI;
2733 TargetLibraryInfo *TLI;
2738 bool runOnFunction(Function &F) override {
2739 if (skipOptnoneFunction(F))
2742 SE = &getAnalysis<ScalarEvolution>();
2743 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2744 DL = DLP ? &DLP->getDataLayout() : nullptr;
2745 TTI = &getAnalysis<TargetTransformInfo>();
2746 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
2747 AA = &getAnalysis<AliasAnalysis>();
2748 LI = &getAnalysis<LoopInfo>();
2749 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2752 bool Changed = false;
2754 // If the target claims to have no vector registers don't attempt
2756 if (!TTI->getNumberOfRegisters(true))
2759 // Must have DataLayout. We can't require it because some tests run w/o
2764 // Don't vectorize when the attribute NoImplicitFloat is used.
2765 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
2768 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
2770 // Use the bottom up slp vectorizer to construct chains that start with
2771 // store instructions.
2772 BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT);
2774 // Scan the blocks in the function in post order.
2775 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
2776 e = po_end(&F.getEntryBlock()); it != e; ++it) {
2777 BasicBlock *BB = *it;
2778 // Vectorize trees that end at stores.
2779 if (unsigned count = collectStores(BB, R)) {
2781 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
2782 Changed |= vectorizeStoreChains(R);
2785 // Vectorize trees that end at reductions.
2786 Changed |= vectorizeChainsInBlock(BB, R);
2790 R.optimizeGatherSequence();
2791 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
2792 DEBUG(verifyFunction(F));
2797 void getAnalysisUsage(AnalysisUsage &AU) const override {
2798 FunctionPass::getAnalysisUsage(AU);
2799 AU.addRequired<ScalarEvolution>();
2800 AU.addRequired<AliasAnalysis>();
2801 AU.addRequired<TargetTransformInfo>();
2802 AU.addRequired<LoopInfo>();
2803 AU.addRequired<DominatorTreeWrapperPass>();
2804 AU.addPreserved<LoopInfo>();
2805 AU.addPreserved<DominatorTreeWrapperPass>();
2806 AU.setPreservesCFG();
2811 /// \brief Collect memory references and sort them according to their base
2812 /// object. We sort the stores to their base objects to reduce the cost of the
2813 /// quadratic search on the stores. TODO: We can further reduce this cost
2814 /// if we flush the chain creation every time we run into a memory barrier.
2815 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
2817 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
2818 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
2820 /// \brief Try to vectorize a list of operands.
2821 /// \@param BuildVector A list of users to ignore for the purpose of
2822 /// scheduling and that don't need extracting.
2823 /// \returns true if a value was vectorized.
2824 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2825 ArrayRef<Value *> BuildVector = None,
2826 bool allowReorder = false);
2828 /// \brief Try to vectorize a chain that may start at the operands of \V;
2829 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
2831 /// \brief Vectorize the stores that were collected in StoreRefs.
2832 bool vectorizeStoreChains(BoUpSLP &R);
2834 /// \brief Scan the basic block and look for patterns that are likely to start
2835 /// a vectorization chain.
2836 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
2838 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
2841 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
2844 StoreListMap StoreRefs;
2847 /// \brief Check that the Values in the slice in VL array are still existent in
2848 /// the WeakVH array.
2849 /// Vectorization of part of the VL array may cause later values in the VL array
2850 /// to become invalid. We track when this has happened in the WeakVH array.
2851 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
2852 SmallVectorImpl<WeakVH> &VH,
2853 unsigned SliceBegin,
2854 unsigned SliceSize) {
2855 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
2862 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
2863 int CostThreshold, BoUpSLP &R) {
2864 unsigned ChainLen = Chain.size();
2865 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2867 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2868 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2869 unsigned VF = MinVecRegSize / Sz;
2871 if (!isPowerOf2_32(Sz) || VF < 2)
2874 // Keep track of values that were deleted by vectorizing in the loop below.
2875 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2877 bool Changed = false;
2878 // Look for profitable vectorizable trees at all offsets, starting at zero.
2879 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2883 // Check that a previous iteration of this loop did not delete the Value.
2884 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2887 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2889 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2891 R.buildTree(Operands);
2893 int Cost = R.getTreeCost();
2895 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2896 if (Cost < CostThreshold) {
2897 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2900 // Move to the next bundle.
2909 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2910 int costThreshold, BoUpSLP &R) {
2911 SetVector<Value *> Heads, Tails;
2912 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2914 // We may run into multiple chains that merge into a single chain. We mark the
2915 // stores that we vectorized so that we don't visit the same store twice.
2916 BoUpSLP::ValueSet VectorizedStores;
2917 bool Changed = false;
2919 // Do a quadratic search on all of the given stores and find
2920 // all of the pairs of stores that follow each other.
2921 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2922 for (unsigned j = 0; j < e; ++j) {
2926 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2927 Tails.insert(Stores[j]);
2928 Heads.insert(Stores[i]);
2929 ConsecutiveChain[Stores[i]] = Stores[j];
2934 // For stores that start but don't end a link in the chain:
2935 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2937 if (Tails.count(*it))
2940 // We found a store instr that starts a chain. Now follow the chain and try
2942 BoUpSLP::ValueList Operands;
2944 // Collect the chain into a list.
2945 while (Tails.count(I) || Heads.count(I)) {
2946 if (VectorizedStores.count(I))
2948 Operands.push_back(I);
2949 // Move to the next value in the chain.
2950 I = ConsecutiveChain[I];
2953 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2955 // Mark the vectorized stores so that we don't vectorize them again.
2957 VectorizedStores.insert(Operands.begin(), Operands.end());
2958 Changed |= Vectorized;
2965 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2968 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2969 StoreInst *SI = dyn_cast<StoreInst>(it);
2973 // Don't touch volatile stores.
2974 if (!SI->isSimple())
2977 // Check that the pointer points to scalars.
2978 Type *Ty = SI->getValueOperand()->getType();
2979 if (Ty->isAggregateType() || Ty->isVectorTy())
2982 // Find the base pointer.
2983 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2985 // Save the store locations.
2986 StoreRefs[Ptr].push_back(SI);
2992 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2995 Value *VL[] = { A, B };
2996 return tryToVectorizeList(VL, R, None, true);
2999 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
3000 ArrayRef<Value *> BuildVector,
3001 bool allowReorder) {
3005 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
3007 // Check that all of the parts are scalar instructions of the same type.
3008 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
3012 unsigned Opcode0 = I0->getOpcode();
3014 Type *Ty0 = I0->getType();
3015 unsigned Sz = DL->getTypeSizeInBits(Ty0);
3016 unsigned VF = MinVecRegSize / Sz;
3018 for (int i = 0, e = VL.size(); i < e; ++i) {
3019 Type *Ty = VL[i]->getType();
3020 if (Ty->isAggregateType() || Ty->isVectorTy())
3022 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
3023 if (!Inst || Inst->getOpcode() != Opcode0)
3027 bool Changed = false;
3029 // Keep track of values that were deleted by vectorizing in the loop below.
3030 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
3032 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
3033 unsigned OpsWidth = 0;
3040 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
3043 // Check that a previous iteration of this loop did not delete the Value.
3044 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
3047 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
3049 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
3051 ArrayRef<Value *> BuildVectorSlice;
3052 if (!BuildVector.empty())
3053 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
3055 R.buildTree(Ops, BuildVectorSlice);
3056 // TODO: check if we can allow reordering also for other cases than
3057 // tryToVectorizePair()
3058 if (allowReorder && R.shouldReorder()) {
3059 assert(Ops.size() == 2);
3060 assert(BuildVectorSlice.empty());
3061 Value *ReorderedOps[] = { Ops[1], Ops[0] };
3062 R.buildTree(ReorderedOps, None);
3064 int Cost = R.getTreeCost();
3066 if (Cost < -SLPCostThreshold) {
3067 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
3068 Value *VectorizedRoot = R.vectorizeTree();
3070 // Reconstruct the build vector by extracting the vectorized root. This
3071 // way we handle the case where some elements of the vector are undefined.
3072 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
3073 if (!BuildVectorSlice.empty()) {
3074 // The insert point is the last build vector instruction. The vectorized
3075 // root will precede it. This guarantees that we get an instruction. The
3076 // vectorized tree could have been constant folded.
3077 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
3078 unsigned VecIdx = 0;
3079 for (auto &V : BuildVectorSlice) {
3080 IRBuilder<true, NoFolder> Builder(
3081 ++BasicBlock::iterator(InsertAfter));
3082 InsertElementInst *IE = cast<InsertElementInst>(V);
3083 Instruction *Extract = cast<Instruction>(Builder.CreateExtractElement(
3084 VectorizedRoot, Builder.getInt32(VecIdx++)));
3085 IE->setOperand(1, Extract);
3086 IE->removeFromParent();
3087 IE->insertAfter(Extract);
3091 // Move to the next bundle.
3100 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
3104 // Try to vectorize V.
3105 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
3108 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
3109 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
3111 if (B && B->hasOneUse()) {
3112 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
3113 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
3114 if (tryToVectorizePair(A, B0, R)) {
3118 if (tryToVectorizePair(A, B1, R)) {
3125 if (A && A->hasOneUse()) {
3126 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
3127 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
3128 if (tryToVectorizePair(A0, B, R)) {
3132 if (tryToVectorizePair(A1, B, R)) {
3140 /// \brief Generate a shuffle mask to be used in a reduction tree.
3142 /// \param VecLen The length of the vector to be reduced.
3143 /// \param NumEltsToRdx The number of elements that should be reduced in the
3145 /// \param IsPairwise Whether the reduction is a pairwise or splitting
3146 /// reduction. A pairwise reduction will generate a mask of
3147 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
3148 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
3149 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
3150 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
3151 bool IsPairwise, bool IsLeft,
3152 IRBuilder<> &Builder) {
3153 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
3155 SmallVector<Constant *, 32> ShuffleMask(
3156 VecLen, UndefValue::get(Builder.getInt32Ty()));
3159 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
3160 for (unsigned i = 0; i != NumEltsToRdx; ++i)
3161 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
3163 // Move the upper half of the vector to the lower half.
3164 for (unsigned i = 0; i != NumEltsToRdx; ++i)
3165 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
3167 return ConstantVector::get(ShuffleMask);
3171 /// Model horizontal reductions.
3173 /// A horizontal reduction is a tree of reduction operations (currently add and
3174 /// fadd) that has operations that can be put into a vector as its leaf.
3175 /// For example, this tree:
3182 /// This tree has "mul" as its reduced values and "+" as its reduction
3183 /// operations. A reduction might be feeding into a store or a binary operation
3198 class HorizontalReduction {
3199 SmallVector<Value *, 16> ReductionOps;
3200 SmallVector<Value *, 32> ReducedVals;
3202 BinaryOperator *ReductionRoot;
3203 PHINode *ReductionPHI;
3205 /// The opcode of the reduction.
3206 unsigned ReductionOpcode;
3207 /// The opcode of the values we perform a reduction on.
3208 unsigned ReducedValueOpcode;
3209 /// The width of one full horizontal reduction operation.
3210 unsigned ReduxWidth;
3211 /// Should we model this reduction as a pairwise reduction tree or a tree that
3212 /// splits the vector in halves and adds those halves.
3213 bool IsPairwiseReduction;
3216 HorizontalReduction()
3217 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
3218 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
3220 /// \brief Try to find a reduction tree.
3221 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
3222 const DataLayout *DL) {
3224 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
3225 "Thi phi needs to use the binary operator");
3227 // We could have a initial reductions that is not an add.
3228 // r *= v1 + v2 + v3 + v4
3229 // In such a case start looking for a tree rooted in the first '+'.
3231 if (B->getOperand(0) == Phi) {
3233 B = dyn_cast<BinaryOperator>(B->getOperand(1));
3234 } else if (B->getOperand(1) == Phi) {
3236 B = dyn_cast<BinaryOperator>(B->getOperand(0));
3243 Type *Ty = B->getType();
3244 if (Ty->isVectorTy())
3247 ReductionOpcode = B->getOpcode();
3248 ReducedValueOpcode = 0;
3249 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
3256 // We currently only support adds.
3257 if (ReductionOpcode != Instruction::Add &&
3258 ReductionOpcode != Instruction::FAdd)
3261 // Post order traverse the reduction tree starting at B. We only handle true
3262 // trees containing only binary operators.
3263 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
3264 Stack.push_back(std::make_pair(B, 0));
3265 while (!Stack.empty()) {
3266 BinaryOperator *TreeN = Stack.back().first;
3267 unsigned EdgeToVist = Stack.back().second++;
3268 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
3270 // Only handle trees in the current basic block.
3271 if (TreeN->getParent() != B->getParent())
3274 // Each tree node needs to have one user except for the ultimate
3276 if (!TreeN->hasOneUse() && TreeN != B)
3280 if (EdgeToVist == 2 || IsReducedValue) {
3281 if (IsReducedValue) {
3282 // Make sure that the opcodes of the operations that we are going to
3284 if (!ReducedValueOpcode)
3285 ReducedValueOpcode = TreeN->getOpcode();
3286 else if (ReducedValueOpcode != TreeN->getOpcode())
3288 ReducedVals.push_back(TreeN);
3290 // We need to be able to reassociate the adds.
3291 if (!TreeN->isAssociative())
3293 ReductionOps.push_back(TreeN);
3300 // Visit left or right.
3301 Value *NextV = TreeN->getOperand(EdgeToVist);
3302 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
3304 Stack.push_back(std::make_pair(Next, 0));
3305 else if (NextV != Phi)
3311 /// \brief Attempt to vectorize the tree found by
3312 /// matchAssociativeReduction.
3313 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
3314 if (ReducedVals.empty())
3317 unsigned NumReducedVals = ReducedVals.size();
3318 if (NumReducedVals < ReduxWidth)
3321 Value *VectorizedTree = nullptr;
3322 IRBuilder<> Builder(ReductionRoot);
3323 FastMathFlags Unsafe;
3324 Unsafe.setUnsafeAlgebra();
3325 Builder.SetFastMathFlags(Unsafe);
3328 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
3329 V.buildTree(makeArrayRef(&ReducedVals[i], ReduxWidth), ReductionOps);
3332 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
3333 if (Cost >= -SLPCostThreshold)
3336 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
3339 // Vectorize a tree.
3340 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
3341 Value *VectorizedRoot = V.vectorizeTree();
3343 // Emit a reduction.
3344 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
3345 if (VectorizedTree) {
3346 Builder.SetCurrentDebugLocation(Loc);
3347 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
3348 ReducedSubTree, "bin.rdx");
3350 VectorizedTree = ReducedSubTree;
3353 if (VectorizedTree) {
3354 // Finish the reduction.
3355 for (; i < NumReducedVals; ++i) {
3356 Builder.SetCurrentDebugLocation(
3357 cast<Instruction>(ReducedVals[i])->getDebugLoc());
3358 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
3363 assert(ReductionRoot && "Need a reduction operation");
3364 ReductionRoot->setOperand(0, VectorizedTree);
3365 ReductionRoot->setOperand(1, ReductionPHI);
3367 ReductionRoot->replaceAllUsesWith(VectorizedTree);
3369 return VectorizedTree != nullptr;
3374 /// \brief Calcuate the cost of a reduction.
3375 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
3376 Type *ScalarTy = FirstReducedVal->getType();
3377 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
3379 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
3380 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
3382 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
3383 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
3385 int ScalarReduxCost =
3386 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
3388 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
3389 << " for reduction that starts with " << *FirstReducedVal
3391 << (IsPairwiseReduction ? "pairwise" : "splitting")
3392 << " reduction)\n");
3394 return VecReduxCost - ScalarReduxCost;
3397 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
3398 Value *R, const Twine &Name = "") {
3399 if (Opcode == Instruction::FAdd)
3400 return Builder.CreateFAdd(L, R, Name);
3401 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
3404 /// \brief Emit a horizontal reduction of the vectorized value.
3405 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
3406 assert(VectorizedValue && "Need to have a vectorized tree node");
3407 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
3408 assert(isPowerOf2_32(ReduxWidth) &&
3409 "We only handle power-of-two reductions for now");
3411 Value *TmpVec = ValToReduce;
3412 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
3413 if (IsPairwiseReduction) {
3415 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
3417 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
3419 Value *LeftShuf = Builder.CreateShuffleVector(
3420 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
3421 Value *RightShuf = Builder.CreateShuffleVector(
3422 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
3424 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
3428 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
3429 Value *Shuf = Builder.CreateShuffleVector(
3430 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
3431 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
3435 // The result is in the first element of the vector.
3436 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
3440 /// \brief Recognize construction of vectors like
3441 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
3442 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
3443 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
3444 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
3446 /// Returns true if it matches
3448 static bool findBuildVector(InsertElementInst *FirstInsertElem,
3449 SmallVectorImpl<Value *> &BuildVector,
3450 SmallVectorImpl<Value *> &BuildVectorOpds) {
3451 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
3454 InsertElementInst *IE = FirstInsertElem;
3456 BuildVector.push_back(IE);
3457 BuildVectorOpds.push_back(IE->getOperand(1));
3459 if (IE->use_empty())
3462 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
3466 // If this isn't the final use, make sure the next insertelement is the only
3467 // use. It's OK if the final constructed vector is used multiple times
3468 if (!IE->hasOneUse())
3477 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
3478 return V->getType() < V2->getType();
3481 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
3482 bool Changed = false;
3483 SmallVector<Value *, 4> Incoming;
3484 SmallSet<Value *, 16> VisitedInstrs;
3486 bool HaveVectorizedPhiNodes = true;
3487 while (HaveVectorizedPhiNodes) {
3488 HaveVectorizedPhiNodes = false;
3490 // Collect the incoming values from the PHIs.
3492 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
3494 PHINode *P = dyn_cast<PHINode>(instr);
3498 if (!VisitedInstrs.count(P))
3499 Incoming.push_back(P);
3503 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
3505 // Try to vectorize elements base on their type.
3506 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
3510 // Look for the next elements with the same type.
3511 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
3512 while (SameTypeIt != E &&
3513 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
3514 VisitedInstrs.insert(*SameTypeIt);
3518 // Try to vectorize them.
3519 unsigned NumElts = (SameTypeIt - IncIt);
3520 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
3521 if (NumElts > 1 && tryToVectorizeList(makeArrayRef(IncIt, NumElts), R)) {
3522 // Success start over because instructions might have been changed.
3523 HaveVectorizedPhiNodes = true;
3528 // Start over at the next instruction of a different type (or the end).
3533 VisitedInstrs.clear();
3535 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
3536 // We may go through BB multiple times so skip the one we have checked.
3537 if (!VisitedInstrs.insert(it))
3540 if (isa<DbgInfoIntrinsic>(it))
3543 // Try to vectorize reductions that use PHINodes.
3544 if (PHINode *P = dyn_cast<PHINode>(it)) {
3545 // Check that the PHI is a reduction PHI.
3546 if (P->getNumIncomingValues() != 2)
3549 (P->getIncomingBlock(0) == BB
3550 ? (P->getIncomingValue(0))
3551 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
3553 // Check if this is a Binary Operator.
3554 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
3558 // Try to match and vectorize a horizontal reduction.
3559 HorizontalReduction HorRdx;
3560 if (ShouldVectorizeHor &&
3561 HorRdx.matchAssociativeReduction(P, BI, DL) &&
3562 HorRdx.tryToReduce(R, TTI)) {
3569 Value *Inst = BI->getOperand(0);
3571 Inst = BI->getOperand(1);
3573 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
3574 // We would like to start over since some instructions are deleted
3575 // and the iterator may become invalid value.
3585 // Try to vectorize horizontal reductions feeding into a store.
3586 if (ShouldStartVectorizeHorAtStore)
3587 if (StoreInst *SI = dyn_cast<StoreInst>(it))
3588 if (BinaryOperator *BinOp =
3589 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
3590 HorizontalReduction HorRdx;
3591 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
3592 HorRdx.tryToReduce(R, TTI)) ||
3593 tryToVectorize(BinOp, R))) {
3601 // Try to vectorize trees that start at compare instructions.
3602 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
3603 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
3605 // We would like to start over since some instructions are deleted
3606 // and the iterator may become invalid value.
3612 for (int i = 0; i < 2; ++i) {
3613 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
3614 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
3616 // We would like to start over since some instructions are deleted
3617 // and the iterator may become invalid value.
3626 // Try to vectorize trees that start at insertelement instructions.
3627 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
3628 SmallVector<Value *, 16> BuildVector;
3629 SmallVector<Value *, 16> BuildVectorOpds;
3630 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
3633 // Vectorize starting with the build vector operands ignoring the
3634 // BuildVector instructions for the purpose of scheduling and user
3636 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
3649 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
3650 bool Changed = false;
3651 // Attempt to sort and vectorize each of the store-groups.
3652 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
3654 if (it->second.size() < 2)
3657 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
3658 << it->second.size() << ".\n");
3660 // Process the stores in chunks of 16.
3661 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
3662 unsigned Len = std::min<unsigned>(CE - CI, 16);
3663 Changed |= vectorizeStores(makeArrayRef(&it->second[CI], Len),
3664 -SLPCostThreshold, R);
3670 } // end anonymous namespace
3672 char SLPVectorizer::ID = 0;
3673 static const char lv_name[] = "SLP Vectorizer";
3674 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
3675 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3676 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3677 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3678 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3679 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
3682 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }