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 can't combine several GEPs into one vector if they operate on
1195 Type *Ty0 = cast<Instruction>(VL0)->getOperand(0)->getType();
1196 for (unsigned j = 0; j < VL.size(); ++j) {
1197 Type *CurTy = cast<Instruction>(VL[j])->getOperand(0)->getType();
1199 DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n");
1200 BS.cancelScheduling(VL);
1201 newTreeEntry(VL, false);
1206 // We don't combine GEPs with non-constant indexes.
1207 for (unsigned j = 0; j < VL.size(); ++j) {
1208 auto Op = cast<Instruction>(VL[j])->getOperand(1);
1209 if (!isa<ConstantInt>(Op)) {
1211 dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n");
1212 BS.cancelScheduling(VL);
1213 newTreeEntry(VL, false);
1218 newTreeEntry(VL, true);
1219 DEBUG(dbgs() << "SLP: added a vector of GEPs.\n");
1220 for (unsigned i = 0, e = 2; i < e; ++i) {
1222 // Prepare the operand vector.
1223 for (unsigned j = 0; j < VL.size(); ++j)
1224 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1226 buildTree_rec(Operands, Depth + 1);
1230 case Instruction::Store: {
1231 // Check if the stores are consecutive or of we need to swizzle them.
1232 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
1233 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
1234 BS.cancelScheduling(VL);
1235 newTreeEntry(VL, false);
1236 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
1240 newTreeEntry(VL, true);
1241 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
1244 for (unsigned j = 0; j < VL.size(); ++j)
1245 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
1247 buildTree_rec(Operands, Depth + 1);
1250 case Instruction::Call: {
1251 // Check if the calls are all to the same vectorizable intrinsic.
1252 CallInst *CI = cast<CallInst>(VL[0]);
1253 // Check if this is an Intrinsic call or something that can be
1254 // represented by an intrinsic call
1255 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1256 if (!isTriviallyVectorizable(ID)) {
1257 BS.cancelScheduling(VL);
1258 newTreeEntry(VL, false);
1259 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
1262 Function *Int = CI->getCalledFunction();
1263 Value *A1I = nullptr;
1264 if (hasVectorInstrinsicScalarOpd(ID, 1))
1265 A1I = CI->getArgOperand(1);
1266 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
1267 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
1268 if (!CI2 || CI2->getCalledFunction() != Int ||
1269 getIntrinsicIDForCall(CI2, TLI) != ID) {
1270 BS.cancelScheduling(VL);
1271 newTreeEntry(VL, false);
1272 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]
1276 // ctlz,cttz and powi are special intrinsics whose second argument
1277 // should be same in order for them to be vectorized.
1278 if (hasVectorInstrinsicScalarOpd(ID, 1)) {
1279 Value *A1J = CI2->getArgOperand(1);
1281 BS.cancelScheduling(VL);
1282 newTreeEntry(VL, false);
1283 DEBUG(dbgs() << "SLP: mismatched arguments in call:" << *CI
1284 << " argument "<< A1I<<"!=" << A1J
1291 newTreeEntry(VL, true);
1292 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
1294 // Prepare the operand vector.
1295 for (unsigned j = 0; j < VL.size(); ++j) {
1296 CallInst *CI2 = dyn_cast<CallInst>(VL[j]);
1297 Operands.push_back(CI2->getArgOperand(i));
1299 buildTree_rec(Operands, Depth + 1);
1303 case Instruction::ShuffleVector: {
1304 // If this is not an alternate sequence of opcode like add-sub
1305 // then do not vectorize this instruction.
1306 if (!isAltShuffle) {
1307 BS.cancelScheduling(VL);
1308 newTreeEntry(VL, false);
1309 DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n");
1312 newTreeEntry(VL, true);
1313 DEBUG(dbgs() << "SLP: added a ShuffleVector op.\n");
1314 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
1316 // Prepare the operand vector.
1317 for (unsigned j = 0; j < VL.size(); ++j)
1318 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
1320 buildTree_rec(Operands, Depth + 1);
1325 BS.cancelScheduling(VL);
1326 newTreeEntry(VL, false);
1327 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
1332 int BoUpSLP::getEntryCost(TreeEntry *E) {
1333 ArrayRef<Value*> VL = E->Scalars;
1335 Type *ScalarTy = VL[0]->getType();
1336 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1337 ScalarTy = SI->getValueOperand()->getType();
1338 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1340 if (E->NeedToGather) {
1341 if (allConstant(VL))
1344 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1346 return getGatherCost(E->Scalars);
1348 unsigned Opcode = getSameOpcode(VL);
1349 assert(Opcode && getSameType(VL) && getSameBlock(VL) && "Invalid VL");
1350 Instruction *VL0 = cast<Instruction>(VL[0]);
1352 case Instruction::PHI: {
1355 case Instruction::ExtractElement: {
1356 if (CanReuseExtract(VL)) {
1358 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1359 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1361 // Take credit for instruction that will become dead.
1363 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1367 return getGatherCost(VecTy);
1369 case Instruction::ZExt:
1370 case Instruction::SExt:
1371 case Instruction::FPToUI:
1372 case Instruction::FPToSI:
1373 case Instruction::FPExt:
1374 case Instruction::PtrToInt:
1375 case Instruction::IntToPtr:
1376 case Instruction::SIToFP:
1377 case Instruction::UIToFP:
1378 case Instruction::Trunc:
1379 case Instruction::FPTrunc:
1380 case Instruction::BitCast: {
1381 Type *SrcTy = VL0->getOperand(0)->getType();
1383 // Calculate the cost of this instruction.
1384 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1385 VL0->getType(), SrcTy);
1387 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1388 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1389 return VecCost - ScalarCost;
1391 case Instruction::FCmp:
1392 case Instruction::ICmp:
1393 case Instruction::Select:
1394 case Instruction::Add:
1395 case Instruction::FAdd:
1396 case Instruction::Sub:
1397 case Instruction::FSub:
1398 case Instruction::Mul:
1399 case Instruction::FMul:
1400 case Instruction::UDiv:
1401 case Instruction::SDiv:
1402 case Instruction::FDiv:
1403 case Instruction::URem:
1404 case Instruction::SRem:
1405 case Instruction::FRem:
1406 case Instruction::Shl:
1407 case Instruction::LShr:
1408 case Instruction::AShr:
1409 case Instruction::And:
1410 case Instruction::Or:
1411 case Instruction::Xor: {
1412 // Calculate the cost of this instruction.
1415 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1416 Opcode == Instruction::Select) {
1417 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1418 ScalarCost = VecTy->getNumElements() *
1419 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1420 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1422 // Certain instructions can be cheaper to vectorize if they have a
1423 // constant second vector operand.
1424 TargetTransformInfo::OperandValueKind Op1VK =
1425 TargetTransformInfo::OK_AnyValue;
1426 TargetTransformInfo::OperandValueKind Op2VK =
1427 TargetTransformInfo::OK_UniformConstantValue;
1429 // If all operands are exactly the same ConstantInt then set the
1430 // operand kind to OK_UniformConstantValue.
1431 // If instead not all operands are constants, then set the operand kind
1432 // to OK_AnyValue. If all operands are constants but not the same,
1433 // then set the operand kind to OK_NonUniformConstantValue.
1434 ConstantInt *CInt = nullptr;
1435 for (unsigned i = 0; i < VL.size(); ++i) {
1436 const Instruction *I = cast<Instruction>(VL[i]);
1437 if (!isa<ConstantInt>(I->getOperand(1))) {
1438 Op2VK = TargetTransformInfo::OK_AnyValue;
1442 CInt = cast<ConstantInt>(I->getOperand(1));
1445 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1446 CInt != cast<ConstantInt>(I->getOperand(1)))
1447 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1451 VecTy->getNumElements() *
1452 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1453 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1455 return VecCost - ScalarCost;
1457 case Instruction::GetElementPtr: {
1458 TargetTransformInfo::OperandValueKind Op1VK =
1459 TargetTransformInfo::OK_AnyValue;
1460 TargetTransformInfo::OperandValueKind Op2VK =
1461 TargetTransformInfo::OK_UniformConstantValue;
1464 VecTy->getNumElements() *
1465 TTI->getArithmeticInstrCost(Instruction::Add, ScalarTy, Op1VK, Op2VK);
1467 TTI->getArithmeticInstrCost(Instruction::Add, VecTy, Op1VK, Op2VK);
1469 return VecCost - ScalarCost;
1471 case Instruction::Load: {
1472 // Cost of wide load - cost of scalar loads.
1473 int ScalarLdCost = VecTy->getNumElements() *
1474 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1475 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1476 return VecLdCost - ScalarLdCost;
1478 case Instruction::Store: {
1479 // We know that we can merge the stores. Calculate the cost.
1480 int ScalarStCost = VecTy->getNumElements() *
1481 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1482 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1483 return VecStCost - ScalarStCost;
1485 case Instruction::Call: {
1486 CallInst *CI = cast<CallInst>(VL0);
1487 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1489 // Calculate the cost of the scalar and vector calls.
1490 SmallVector<Type*, 4> ScalarTys, VecTys;
1491 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1492 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1493 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1494 VecTy->getNumElements()));
1497 int ScalarCallCost = VecTy->getNumElements() *
1498 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1500 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1502 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1503 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1504 << " for " << *CI << "\n");
1506 return VecCallCost - ScalarCallCost;
1508 case Instruction::ShuffleVector: {
1509 TargetTransformInfo::OperandValueKind Op1VK =
1510 TargetTransformInfo::OK_AnyValue;
1511 TargetTransformInfo::OperandValueKind Op2VK =
1512 TargetTransformInfo::OK_AnyValue;
1515 for (unsigned i = 0; i < VL.size(); ++i) {
1516 Instruction *I = cast<Instruction>(VL[i]);
1520 TTI->getArithmeticInstrCost(I->getOpcode(), ScalarTy, Op1VK, Op2VK);
1522 // VecCost is equal to sum of the cost of creating 2 vectors
1523 // and the cost of creating shuffle.
1524 Instruction *I0 = cast<Instruction>(VL[0]);
1526 TTI->getArithmeticInstrCost(I0->getOpcode(), VecTy, Op1VK, Op2VK);
1527 Instruction *I1 = cast<Instruction>(VL[1]);
1529 TTI->getArithmeticInstrCost(I1->getOpcode(), VecTy, Op1VK, Op2VK);
1531 TTI->getShuffleCost(TargetTransformInfo::SK_Alternate, VecTy, 0);
1532 return VecCost - ScalarCost;
1535 llvm_unreachable("Unknown instruction");
1539 bool BoUpSLP::isFullyVectorizableTinyTree() {
1540 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1541 VectorizableTree.size() << " is fully vectorizable .\n");
1543 // We only handle trees of height 2.
1544 if (VectorizableTree.size() != 2)
1547 // Handle splat stores.
1548 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1551 // Gathering cost would be too much for tiny trees.
1552 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1558 int BoUpSLP::getSpillCost() {
1559 // Walk from the bottom of the tree to the top, tracking which values are
1560 // live. When we see a call instruction that is not part of our tree,
1561 // query TTI to see if there is a cost to keeping values live over it
1562 // (for example, if spills and fills are required).
1563 unsigned BundleWidth = VectorizableTree.front().Scalars.size();
1566 SmallPtrSet<Instruction*, 4> LiveValues;
1567 Instruction *PrevInst = nullptr;
1569 for (unsigned N = 0; N < VectorizableTree.size(); ++N) {
1570 Instruction *Inst = dyn_cast<Instruction>(VectorizableTree[N].Scalars[0]);
1580 dbgs() << "SLP: #LV: " << LiveValues.size();
1581 for (auto *X : LiveValues)
1582 dbgs() << " " << X->getName();
1583 dbgs() << ", Looking at ";
1587 // Update LiveValues.
1588 LiveValues.erase(PrevInst);
1589 for (auto &J : PrevInst->operands()) {
1590 if (isa<Instruction>(&*J) && ScalarToTreeEntry.count(&*J))
1591 LiveValues.insert(cast<Instruction>(&*J));
1594 // Now find the sequence of instructions between PrevInst and Inst.
1595 BasicBlock::reverse_iterator InstIt(Inst), PrevInstIt(PrevInst);
1597 while (InstIt != PrevInstIt) {
1598 if (PrevInstIt == PrevInst->getParent()->rend()) {
1599 PrevInstIt = Inst->getParent()->rbegin();
1603 if (isa<CallInst>(&*PrevInstIt) && &*PrevInstIt != PrevInst) {
1604 SmallVector<Type*, 4> V;
1605 for (auto *II : LiveValues)
1606 V.push_back(VectorType::get(II->getType(), BundleWidth));
1607 Cost += TTI->getCostOfKeepingLiveOverCall(V);
1616 DEBUG(dbgs() << "SLP: SpillCost=" << Cost << "\n");
1620 int BoUpSLP::getTreeCost() {
1622 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1623 VectorizableTree.size() << ".\n");
1625 // We only vectorize tiny trees if it is fully vectorizable.
1626 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1627 if (!VectorizableTree.size()) {
1628 assert(!ExternalUses.size() && "We should not have any external users");
1633 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1635 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1636 int C = getEntryCost(&VectorizableTree[i]);
1637 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1638 << *VectorizableTree[i].Scalars[0] << " .\n");
1642 SmallSet<Value *, 16> ExtractCostCalculated;
1643 int ExtractCost = 0;
1644 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1646 // We only add extract cost once for the same scalar.
1647 if (!ExtractCostCalculated.insert(I->Scalar))
1650 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1651 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1655 Cost += getSpillCost();
1657 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1658 return Cost + ExtractCost;
1661 int BoUpSLP::getGatherCost(Type *Ty) {
1663 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1664 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1668 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1669 // Find the type of the operands in VL.
1670 Type *ScalarTy = VL[0]->getType();
1671 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1672 ScalarTy = SI->getValueOperand()->getType();
1673 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1674 // Find the cost of inserting/extracting values from the vector.
1675 return getGatherCost(VecTy);
1678 Value *BoUpSLP::getPointerOperand(Value *I) {
1679 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1680 return LI->getPointerOperand();
1681 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1682 return SI->getPointerOperand();
1686 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1687 if (LoadInst *L = dyn_cast<LoadInst>(I))
1688 return L->getPointerAddressSpace();
1689 if (StoreInst *S = dyn_cast<StoreInst>(I))
1690 return S->getPointerAddressSpace();
1694 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1695 Value *PtrA = getPointerOperand(A);
1696 Value *PtrB = getPointerOperand(B);
1697 unsigned ASA = getAddressSpaceOperand(A);
1698 unsigned ASB = getAddressSpaceOperand(B);
1700 // Check that the address spaces match and that the pointers are valid.
1701 if (!PtrA || !PtrB || (ASA != ASB))
1704 // Make sure that A and B are different pointers of the same type.
1705 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1708 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1709 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1710 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1712 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1713 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1714 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1716 APInt OffsetDelta = OffsetB - OffsetA;
1718 // Check if they are based on the same pointer. That makes the offsets
1721 return OffsetDelta == Size;
1723 // Compute the necessary base pointer delta to have the necessary final delta
1724 // equal to the size.
1725 APInt BaseDelta = Size - OffsetDelta;
1727 // Otherwise compute the distance with SCEV between the base pointers.
1728 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1729 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1730 const SCEV *C = SE->getConstant(BaseDelta);
1731 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1732 return X == PtrSCEVB;
1735 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1736 Instruction *VL0 = cast<Instruction>(VL[0]);
1737 BasicBlock::iterator NextInst = VL0;
1739 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1740 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1743 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1744 Value *Vec = UndefValue::get(Ty);
1745 // Generate the 'InsertElement' instruction.
1746 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1747 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1748 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1749 GatherSeq.insert(Insrt);
1750 CSEBlocks.insert(Insrt->getParent());
1752 // Add to our 'need-to-extract' list.
1753 if (ScalarToTreeEntry.count(VL[i])) {
1754 int Idx = ScalarToTreeEntry[VL[i]];
1755 TreeEntry *E = &VectorizableTree[Idx];
1756 // Find which lane we need to extract.
1758 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1759 // Is this the lane of the scalar that we are looking for ?
1760 if (E->Scalars[Lane] == VL[i]) {
1765 assert(FoundLane >= 0 && "Could not find the correct lane");
1766 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1774 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1775 SmallDenseMap<Value*, int>::const_iterator Entry
1776 = ScalarToTreeEntry.find(VL[0]);
1777 if (Entry != ScalarToTreeEntry.end()) {
1778 int Idx = Entry->second;
1779 const TreeEntry *En = &VectorizableTree[Idx];
1780 if (En->isSame(VL) && En->VectorizedValue)
1781 return En->VectorizedValue;
1786 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1787 if (ScalarToTreeEntry.count(VL[0])) {
1788 int Idx = ScalarToTreeEntry[VL[0]];
1789 TreeEntry *E = &VectorizableTree[Idx];
1791 return vectorizeTree(E);
1794 Type *ScalarTy = VL[0]->getType();
1795 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1796 ScalarTy = SI->getValueOperand()->getType();
1797 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1799 return Gather(VL, VecTy);
1802 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1803 IRBuilder<>::InsertPointGuard Guard(Builder);
1805 if (E->VectorizedValue) {
1806 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1807 return E->VectorizedValue;
1810 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1811 Type *ScalarTy = VL0->getType();
1812 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1813 ScalarTy = SI->getValueOperand()->getType();
1814 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1816 if (E->NeedToGather) {
1817 setInsertPointAfterBundle(E->Scalars);
1818 return Gather(E->Scalars, VecTy);
1821 unsigned Opcode = getSameOpcode(E->Scalars);
1824 case Instruction::PHI: {
1825 PHINode *PH = dyn_cast<PHINode>(VL0);
1826 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1827 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1828 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1829 E->VectorizedValue = NewPhi;
1831 // PHINodes may have multiple entries from the same block. We want to
1832 // visit every block once.
1833 SmallSet<BasicBlock*, 4> VisitedBBs;
1835 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1837 BasicBlock *IBB = PH->getIncomingBlock(i);
1839 if (!VisitedBBs.insert(IBB)) {
1840 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1844 // Prepare the operand vector.
1845 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1846 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1847 getIncomingValueForBlock(IBB));
1849 Builder.SetInsertPoint(IBB->getTerminator());
1850 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1851 Value *Vec = vectorizeTree(Operands);
1852 NewPhi->addIncoming(Vec, IBB);
1855 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1856 "Invalid number of incoming values");
1860 case Instruction::ExtractElement: {
1861 if (CanReuseExtract(E->Scalars)) {
1862 Value *V = VL0->getOperand(0);
1863 E->VectorizedValue = V;
1866 return Gather(E->Scalars, VecTy);
1868 case Instruction::ZExt:
1869 case Instruction::SExt:
1870 case Instruction::FPToUI:
1871 case Instruction::FPToSI:
1872 case Instruction::FPExt:
1873 case Instruction::PtrToInt:
1874 case Instruction::IntToPtr:
1875 case Instruction::SIToFP:
1876 case Instruction::UIToFP:
1877 case Instruction::Trunc:
1878 case Instruction::FPTrunc:
1879 case Instruction::BitCast: {
1881 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1882 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1884 setInsertPointAfterBundle(E->Scalars);
1886 Value *InVec = vectorizeTree(INVL);
1888 if (Value *V = alreadyVectorized(E->Scalars))
1891 CastInst *CI = dyn_cast<CastInst>(VL0);
1892 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1893 E->VectorizedValue = V;
1894 ++NumVectorInstructions;
1897 case Instruction::FCmp:
1898 case Instruction::ICmp: {
1899 ValueList LHSV, RHSV;
1900 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1901 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1902 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1905 setInsertPointAfterBundle(E->Scalars);
1907 Value *L = vectorizeTree(LHSV);
1908 Value *R = vectorizeTree(RHSV);
1910 if (Value *V = alreadyVectorized(E->Scalars))
1913 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1915 if (Opcode == Instruction::FCmp)
1916 V = Builder.CreateFCmp(P0, L, R);
1918 V = Builder.CreateICmp(P0, L, R);
1920 E->VectorizedValue = V;
1921 ++NumVectorInstructions;
1924 case Instruction::Select: {
1925 ValueList TrueVec, FalseVec, CondVec;
1926 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1927 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1928 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1929 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1932 setInsertPointAfterBundle(E->Scalars);
1934 Value *Cond = vectorizeTree(CondVec);
1935 Value *True = vectorizeTree(TrueVec);
1936 Value *False = vectorizeTree(FalseVec);
1938 if (Value *V = alreadyVectorized(E->Scalars))
1941 Value *V = Builder.CreateSelect(Cond, True, False);
1942 E->VectorizedValue = V;
1943 ++NumVectorInstructions;
1946 case Instruction::Add:
1947 case Instruction::FAdd:
1948 case Instruction::Sub:
1949 case Instruction::FSub:
1950 case Instruction::Mul:
1951 case Instruction::FMul:
1952 case Instruction::UDiv:
1953 case Instruction::SDiv:
1954 case Instruction::FDiv:
1955 case Instruction::URem:
1956 case Instruction::SRem:
1957 case Instruction::FRem:
1958 case Instruction::Shl:
1959 case Instruction::LShr:
1960 case Instruction::AShr:
1961 case Instruction::And:
1962 case Instruction::Or:
1963 case Instruction::Xor: {
1964 ValueList LHSVL, RHSVL;
1965 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1966 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1968 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1969 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1970 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1973 setInsertPointAfterBundle(E->Scalars);
1975 Value *LHS = vectorizeTree(LHSVL);
1976 Value *RHS = vectorizeTree(RHSVL);
1978 if (LHS == RHS && isa<Instruction>(LHS)) {
1979 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1982 if (Value *V = alreadyVectorized(E->Scalars))
1985 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1986 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1987 E->VectorizedValue = V;
1988 ++NumVectorInstructions;
1990 if (Instruction *I = dyn_cast<Instruction>(V))
1991 return propagateMetadata(I, E->Scalars);
1995 case Instruction::Load: {
1996 // Loads are inserted at the head of the tree because we don't want to
1997 // sink them all the way down past store instructions.
1998 setInsertPointAfterBundle(E->Scalars);
2000 LoadInst *LI = cast<LoadInst>(VL0);
2001 Type *ScalarLoadTy = LI->getType();
2002 unsigned AS = LI->getPointerAddressSpace();
2004 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
2005 VecTy->getPointerTo(AS));
2006 unsigned Alignment = LI->getAlignment();
2007 LI = Builder.CreateLoad(VecPtr);
2009 Alignment = DL->getABITypeAlignment(ScalarLoadTy);
2010 LI->setAlignment(Alignment);
2011 E->VectorizedValue = LI;
2012 ++NumVectorInstructions;
2013 return propagateMetadata(LI, E->Scalars);
2015 case Instruction::Store: {
2016 StoreInst *SI = cast<StoreInst>(VL0);
2017 unsigned Alignment = SI->getAlignment();
2018 unsigned AS = SI->getPointerAddressSpace();
2021 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2022 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
2024 setInsertPointAfterBundle(E->Scalars);
2026 Value *VecValue = vectorizeTree(ValueOp);
2027 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
2028 VecTy->getPointerTo(AS));
2029 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
2031 Alignment = DL->getABITypeAlignment(SI->getValueOperand()->getType());
2032 S->setAlignment(Alignment);
2033 E->VectorizedValue = S;
2034 ++NumVectorInstructions;
2035 return propagateMetadata(S, E->Scalars);
2037 case Instruction::GetElementPtr: {
2038 setInsertPointAfterBundle(E->Scalars);
2041 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2042 Op0VL.push_back(cast<GetElementPtrInst>(E->Scalars[i])->getOperand(0));
2044 Value *Op0 = vectorizeTree(Op0VL);
2046 std::vector<Value *> OpVecs;
2047 for (int j = 1, e = cast<GetElementPtrInst>(VL0)->getNumOperands(); j < e;
2050 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
2051 OpVL.push_back(cast<GetElementPtrInst>(E->Scalars[i])->getOperand(j));
2053 Value *OpVec = vectorizeTree(OpVL);
2054 OpVecs.push_back(OpVec);
2057 Value *V = Builder.CreateGEP(Op0, OpVecs);
2058 E->VectorizedValue = V;
2059 ++NumVectorInstructions;
2061 if (Instruction *I = dyn_cast<Instruction>(V))
2062 return propagateMetadata(I, E->Scalars);
2066 case Instruction::Call: {
2067 CallInst *CI = cast<CallInst>(VL0);
2068 setInsertPointAfterBundle(E->Scalars);
2070 Intrinsic::ID IID = Intrinsic::not_intrinsic;
2071 if (CI && (FI = CI->getCalledFunction())) {
2072 IID = (Intrinsic::ID) FI->getIntrinsicID();
2074 std::vector<Value *> OpVecs;
2075 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
2077 // ctlz,cttz and powi are special intrinsics whose second argument is
2078 // a scalar. This argument should not be vectorized.
2079 if (hasVectorInstrinsicScalarOpd(IID, 1) && j == 1) {
2080 CallInst *CEI = cast<CallInst>(E->Scalars[0]);
2081 OpVecs.push_back(CEI->getArgOperand(j));
2084 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
2085 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
2086 OpVL.push_back(CEI->getArgOperand(j));
2089 Value *OpVec = vectorizeTree(OpVL);
2090 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
2091 OpVecs.push_back(OpVec);
2094 Module *M = F->getParent();
2095 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
2096 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
2097 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
2098 Value *V = Builder.CreateCall(CF, OpVecs);
2099 E->VectorizedValue = V;
2100 ++NumVectorInstructions;
2103 case Instruction::ShuffleVector: {
2104 ValueList LHSVL, RHSVL;
2105 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
2106 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
2107 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
2109 setInsertPointAfterBundle(E->Scalars);
2111 Value *LHS = vectorizeTree(LHSVL);
2112 Value *RHS = vectorizeTree(RHSVL);
2114 if (Value *V = alreadyVectorized(E->Scalars))
2117 // Create a vector of LHS op1 RHS
2118 BinaryOperator *BinOp0 = cast<BinaryOperator>(VL0);
2119 Value *V0 = Builder.CreateBinOp(BinOp0->getOpcode(), LHS, RHS);
2121 // Create a vector of LHS op2 RHS
2122 Instruction *VL1 = cast<Instruction>(E->Scalars[1]);
2123 BinaryOperator *BinOp1 = cast<BinaryOperator>(VL1);
2124 Value *V1 = Builder.CreateBinOp(BinOp1->getOpcode(), LHS, RHS);
2126 // Create appropriate shuffle to take alternative operations from
2128 std::vector<Constant *> Mask(E->Scalars.size());
2129 unsigned e = E->Scalars.size();
2130 for (unsigned i = 0; i < e; ++i) {
2132 Mask[i] = Builder.getInt32(e + i);
2134 Mask[i] = Builder.getInt32(i);
2137 Value *ShuffleMask = ConstantVector::get(Mask);
2139 Value *V = Builder.CreateShuffleVector(V0, V1, ShuffleMask);
2140 E->VectorizedValue = V;
2141 ++NumVectorInstructions;
2142 if (Instruction *I = dyn_cast<Instruction>(V))
2143 return propagateMetadata(I, E->Scalars);
2148 llvm_unreachable("unknown inst");
2153 Value *BoUpSLP::vectorizeTree() {
2155 // All blocks must be scheduled before any instructions are inserted.
2156 for (auto &BSIter : BlocksSchedules) {
2157 scheduleBlock(BSIter.second.get());
2160 Builder.SetInsertPoint(F->getEntryBlock().begin());
2161 vectorizeTree(&VectorizableTree[0]);
2163 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
2165 // Extract all of the elements with the external uses.
2166 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
2168 Value *Scalar = it->Scalar;
2169 llvm::User *User = it->User;
2171 // Skip users that we already RAUW. This happens when one instruction
2172 // has multiple uses of the same value.
2173 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
2176 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
2178 int Idx = ScalarToTreeEntry[Scalar];
2179 TreeEntry *E = &VectorizableTree[Idx];
2180 assert(!E->NeedToGather && "Extracting from a gather list");
2182 Value *Vec = E->VectorizedValue;
2183 assert(Vec && "Can't find vectorizable value");
2185 Value *Lane = Builder.getInt32(it->Lane);
2186 // Generate extracts for out-of-tree users.
2187 // Find the insertion point for the extractelement lane.
2188 if (isa<Instruction>(Vec)){
2189 if (PHINode *PH = dyn_cast<PHINode>(User)) {
2190 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
2191 if (PH->getIncomingValue(i) == Scalar) {
2192 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
2193 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2194 CSEBlocks.insert(PH->getIncomingBlock(i));
2195 PH->setOperand(i, Ex);
2199 Builder.SetInsertPoint(cast<Instruction>(User));
2200 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2201 CSEBlocks.insert(cast<Instruction>(User)->getParent());
2202 User->replaceUsesOfWith(Scalar, Ex);
2205 Builder.SetInsertPoint(F->getEntryBlock().begin());
2206 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
2207 CSEBlocks.insert(&F->getEntryBlock());
2208 User->replaceUsesOfWith(Scalar, Ex);
2211 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
2214 // For each vectorized value:
2215 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
2216 TreeEntry *Entry = &VectorizableTree[EIdx];
2219 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
2220 Value *Scalar = Entry->Scalars[Lane];
2221 // No need to handle users of gathered values.
2222 if (Entry->NeedToGather)
2225 assert(Entry->VectorizedValue && "Can't find vectorizable value");
2227 Type *Ty = Scalar->getType();
2228 if (!Ty->isVoidTy()) {
2230 for (User *U : Scalar->users()) {
2231 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
2233 assert((ScalarToTreeEntry.count(U) ||
2234 // It is legal to replace users in the ignorelist by undef.
2235 (std::find(UserIgnoreList.begin(), UserIgnoreList.end(), U) !=
2236 UserIgnoreList.end())) &&
2237 "Replacing out-of-tree value with undef");
2240 Value *Undef = UndefValue::get(Ty);
2241 Scalar->replaceAllUsesWith(Undef);
2243 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
2244 cast<Instruction>(Scalar)->eraseFromParent();
2248 Builder.ClearInsertionPoint();
2250 return VectorizableTree[0].VectorizedValue;
2253 void BoUpSLP::optimizeGatherSequence() {
2254 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
2255 << " gather sequences instructions.\n");
2256 // LICM InsertElementInst sequences.
2257 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
2258 e = GatherSeq.end(); it != e; ++it) {
2259 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
2264 // Check if this block is inside a loop.
2265 Loop *L = LI->getLoopFor(Insert->getParent());
2269 // Check if it has a preheader.
2270 BasicBlock *PreHeader = L->getLoopPreheader();
2274 // If the vector or the element that we insert into it are
2275 // instructions that are defined in this basic block then we can't
2276 // hoist this instruction.
2277 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
2278 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
2279 if (CurrVec && L->contains(CurrVec))
2281 if (NewElem && L->contains(NewElem))
2284 // We can hoist this instruction. Move it to the pre-header.
2285 Insert->moveBefore(PreHeader->getTerminator());
2288 // Make a list of all reachable blocks in our CSE queue.
2289 SmallVector<const DomTreeNode *, 8> CSEWorkList;
2290 CSEWorkList.reserve(CSEBlocks.size());
2291 for (BasicBlock *BB : CSEBlocks)
2292 if (DomTreeNode *N = DT->getNode(BB)) {
2293 assert(DT->isReachableFromEntry(N));
2294 CSEWorkList.push_back(N);
2297 // Sort blocks by domination. This ensures we visit a block after all blocks
2298 // dominating it are visited.
2299 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
2300 [this](const DomTreeNode *A, const DomTreeNode *B) {
2301 return DT->properlyDominates(A, B);
2304 // Perform O(N^2) search over the gather sequences and merge identical
2305 // instructions. TODO: We can further optimize this scan if we split the
2306 // instructions into different buckets based on the insert lane.
2307 SmallVector<Instruction *, 16> Visited;
2308 for (auto I = CSEWorkList.begin(), E = CSEWorkList.end(); I != E; ++I) {
2309 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
2310 "Worklist not sorted properly!");
2311 BasicBlock *BB = (*I)->getBlock();
2312 // For all instructions in blocks containing gather sequences:
2313 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
2314 Instruction *In = it++;
2315 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
2318 // Check if we can replace this instruction with any of the
2319 // visited instructions.
2320 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
2323 if (In->isIdenticalTo(*v) &&
2324 DT->dominates((*v)->getParent(), In->getParent())) {
2325 In->replaceAllUsesWith(*v);
2326 In->eraseFromParent();
2332 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
2333 Visited.push_back(In);
2341 // Groups the instructions to a bundle (which is then a single scheduling entity)
2342 // and schedules instructions until the bundle gets ready.
2343 bool BoUpSLP::BlockScheduling::tryScheduleBundle(ArrayRef<Value *> VL,
2344 AliasAnalysis *AA) {
2345 if (isa<PHINode>(VL[0]))
2348 // Initialize the instruction bundle.
2349 Instruction *OldScheduleEnd = ScheduleEnd;
2350 ScheduleData *PrevInBundle = nullptr;
2351 ScheduleData *Bundle = nullptr;
2352 bool ReSchedule = false;
2353 DEBUG(dbgs() << "SLP: bundle: " << *VL[0] << "\n");
2354 for (Value *V : VL) {
2355 extendSchedulingRegion(V);
2356 ScheduleData *BundleMember = getScheduleData(V);
2357 assert(BundleMember &&
2358 "no ScheduleData for bundle member (maybe not in same basic block)");
2359 if (BundleMember->IsScheduled) {
2360 // A bundle member was scheduled as single instruction before and now
2361 // needs to be scheduled as part of the bundle. We just get rid of the
2362 // existing schedule.
2363 DEBUG(dbgs() << "SLP: reset schedule because " << *BundleMember
2364 << " was already scheduled\n");
2367 assert(BundleMember->isSchedulingEntity() &&
2368 "bundle member already part of other bundle");
2370 PrevInBundle->NextInBundle = BundleMember;
2372 Bundle = BundleMember;
2374 BundleMember->UnscheduledDepsInBundle = 0;
2375 Bundle->UnscheduledDepsInBundle += BundleMember->UnscheduledDeps;
2377 // Group the instructions to a bundle.
2378 BundleMember->FirstInBundle = Bundle;
2379 PrevInBundle = BundleMember;
2381 if (ScheduleEnd != OldScheduleEnd) {
2382 // The scheduling region got new instructions at the lower end (or it is a
2383 // new region for the first bundle). This makes it necessary to
2384 // recalculate all dependencies.
2385 // It is seldom that this needs to be done a second time after adding the
2386 // initial bundle to the region.
2387 for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
2388 ScheduleData *SD = getScheduleData(I);
2389 SD->clearDependencies();
2395 initialFillReadyList(ReadyInsts);
2398 DEBUG(dbgs() << "SLP: try schedule bundle " << *Bundle << " in block "
2399 << BB->getName() << "\n");
2401 calculateDependencies(Bundle, true, AA);
2403 // Now try to schedule the new bundle. As soon as the bundle is "ready" it
2404 // means that there are no cyclic dependencies and we can schedule it.
2405 // Note that's important that we don't "schedule" the bundle yet (see
2406 // cancelScheduling).
2407 while (!Bundle->isReady() && !ReadyInsts.empty()) {
2409 ScheduleData *pickedSD = ReadyInsts.back();
2410 ReadyInsts.pop_back();
2412 if (pickedSD->isSchedulingEntity() && pickedSD->isReady()) {
2413 schedule(pickedSD, ReadyInsts);
2416 return Bundle->isReady();
2419 void BoUpSLP::BlockScheduling::cancelScheduling(ArrayRef<Value *> VL) {
2420 if (isa<PHINode>(VL[0]))
2423 ScheduleData *Bundle = getScheduleData(VL[0]);
2424 DEBUG(dbgs() << "SLP: cancel scheduling of " << *Bundle << "\n");
2425 assert(!Bundle->IsScheduled &&
2426 "Can't cancel bundle which is already scheduled");
2427 assert(Bundle->isSchedulingEntity() && Bundle->isPartOfBundle() &&
2428 "tried to unbundle something which is not a bundle");
2430 // Un-bundle: make single instructions out of the bundle.
2431 ScheduleData *BundleMember = Bundle;
2432 while (BundleMember) {
2433 assert(BundleMember->FirstInBundle == Bundle && "corrupt bundle links");
2434 BundleMember->FirstInBundle = BundleMember;
2435 ScheduleData *Next = BundleMember->NextInBundle;
2436 BundleMember->NextInBundle = nullptr;
2437 BundleMember->UnscheduledDepsInBundle = BundleMember->UnscheduledDeps;
2438 if (BundleMember->UnscheduledDepsInBundle == 0) {
2439 ReadyInsts.insert(BundleMember);
2441 BundleMember = Next;
2445 void BoUpSLP::BlockScheduling::extendSchedulingRegion(Value *V) {
2446 if (getScheduleData(V))
2448 Instruction *I = dyn_cast<Instruction>(V);
2449 assert(I && "bundle member must be an instruction");
2450 assert(!isa<PHINode>(I) && "phi nodes don't need to be scheduled");
2451 if (!ScheduleStart) {
2452 // It's the first instruction in the new region.
2453 initScheduleData(I, I->getNextNode(), nullptr, nullptr);
2455 ScheduleEnd = I->getNextNode();
2456 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?");
2457 DEBUG(dbgs() << "SLP: initialize schedule region to " << *I << "\n");
2460 // Search up and down at the same time, because we don't know if the new
2461 // instruction is above or below the existing scheduling region.
2462 BasicBlock::reverse_iterator UpIter(ScheduleStart);
2463 BasicBlock::reverse_iterator UpperEnd = BB->rend();
2464 BasicBlock::iterator DownIter(ScheduleEnd);
2465 BasicBlock::iterator LowerEnd = BB->end();
2467 if (UpIter != UpperEnd) {
2468 if (&*UpIter == I) {
2469 initScheduleData(I, ScheduleStart, nullptr, FirstLoadStoreInRegion);
2471 DEBUG(dbgs() << "SLP: extend schedule region start to " << *I << "\n");
2476 if (DownIter != LowerEnd) {
2477 if (&*DownIter == I) {
2478 initScheduleData(ScheduleEnd, I->getNextNode(), LastLoadStoreInRegion,
2480 ScheduleEnd = I->getNextNode();
2481 assert(ScheduleEnd && "tried to vectorize a TerminatorInst?");
2482 DEBUG(dbgs() << "SLP: extend schedule region end to " << *I << "\n");
2487 assert((UpIter != UpperEnd || DownIter != LowerEnd) &&
2488 "instruction not found in block");
2492 void BoUpSLP::BlockScheduling::initScheduleData(Instruction *FromI,
2494 ScheduleData *PrevLoadStore,
2495 ScheduleData *NextLoadStore) {
2496 ScheduleData *CurrentLoadStore = PrevLoadStore;
2497 for (Instruction *I = FromI; I != ToI; I = I->getNextNode()) {
2498 ScheduleData *SD = ScheduleDataMap[I];
2500 // Allocate a new ScheduleData for the instruction.
2501 if (ChunkPos >= ChunkSize) {
2502 ScheduleDataChunks.push_back(
2503 llvm::make_unique<ScheduleData[]>(ChunkSize));
2506 SD = &(ScheduleDataChunks.back()[ChunkPos++]);
2507 ScheduleDataMap[I] = SD;
2510 assert(!isInSchedulingRegion(SD) &&
2511 "new ScheduleData already in scheduling region");
2512 SD->init(SchedulingRegionID);
2514 if (I->mayReadOrWriteMemory()) {
2515 // Update the linked list of memory accessing instructions.
2516 if (CurrentLoadStore) {
2517 CurrentLoadStore->NextLoadStore = SD;
2519 FirstLoadStoreInRegion = SD;
2521 CurrentLoadStore = SD;
2524 if (NextLoadStore) {
2525 if (CurrentLoadStore)
2526 CurrentLoadStore->NextLoadStore = NextLoadStore;
2528 LastLoadStoreInRegion = CurrentLoadStore;
2532 /// \returns the AA location that is being access by the instruction.
2533 static AliasAnalysis::Location getLocation(Instruction *I, AliasAnalysis *AA) {
2534 if (StoreInst *SI = dyn_cast<StoreInst>(I))
2535 return AA->getLocation(SI);
2536 if (LoadInst *LI = dyn_cast<LoadInst>(I))
2537 return AA->getLocation(LI);
2538 return AliasAnalysis::Location();
2541 void BoUpSLP::BlockScheduling::calculateDependencies(ScheduleData *SD,
2542 bool InsertInReadyList,
2543 AliasAnalysis *AA) {
2544 assert(SD->isSchedulingEntity());
2546 SmallVector<ScheduleData *, 10> WorkList;
2547 WorkList.push_back(SD);
2549 while (!WorkList.empty()) {
2550 ScheduleData *SD = WorkList.back();
2551 WorkList.pop_back();
2553 ScheduleData *BundleMember = SD;
2554 while (BundleMember) {
2555 assert(isInSchedulingRegion(BundleMember));
2556 if (!BundleMember->hasValidDependencies()) {
2558 DEBUG(dbgs() << "SLP: update deps of " << *BundleMember << "\n");
2559 BundleMember->Dependencies = 0;
2560 BundleMember->resetUnscheduledDeps();
2562 // Handle def-use chain dependencies.
2563 for (User *U : BundleMember->Inst->users()) {
2564 if (isa<Instruction>(U)) {
2565 ScheduleData *UseSD = getScheduleData(U);
2566 if (UseSD && isInSchedulingRegion(UseSD->FirstInBundle)) {
2567 BundleMember->Dependencies++;
2568 ScheduleData *DestBundle = UseSD->FirstInBundle;
2569 if (!DestBundle->IsScheduled) {
2570 BundleMember->incrementUnscheduledDeps(1);
2572 if (!DestBundle->hasValidDependencies()) {
2573 WorkList.push_back(DestBundle);
2577 // I'm not sure if this can ever happen. But we need to be safe.
2578 // This lets the instruction/bundle never be scheduled and eventally
2579 // disable vectorization.
2580 BundleMember->Dependencies++;
2581 BundleMember->incrementUnscheduledDeps(1);
2585 // Handle the memory dependencies.
2586 ScheduleData *DepDest = BundleMember->NextLoadStore;
2588 AliasAnalysis::Location SrcLoc = getLocation(BundleMember->Inst, AA);
2589 bool SrcMayWrite = BundleMember->Inst->mayWriteToMemory();
2592 assert(isInSchedulingRegion(DepDest));
2593 if (SrcMayWrite || DepDest->Inst->mayWriteToMemory()) {
2594 AliasAnalysis::Location DstLoc = getLocation(DepDest->Inst, AA);
2595 if (!SrcLoc.Ptr || !DstLoc.Ptr || AA->alias(SrcLoc, DstLoc)) {
2596 DepDest->MemoryDependencies.push_back(BundleMember);
2597 BundleMember->Dependencies++;
2598 ScheduleData *DestBundle = DepDest->FirstInBundle;
2599 if (!DestBundle->IsScheduled) {
2600 BundleMember->incrementUnscheduledDeps(1);
2602 if (!DestBundle->hasValidDependencies()) {
2603 WorkList.push_back(DestBundle);
2607 DepDest = DepDest->NextLoadStore;
2611 BundleMember = BundleMember->NextInBundle;
2613 if (InsertInReadyList && SD->isReady()) {
2614 ReadyInsts.push_back(SD);
2615 DEBUG(dbgs() << "SLP: gets ready on update: " << *SD->Inst << "\n");
2620 void BoUpSLP::BlockScheduling::resetSchedule() {
2621 assert(ScheduleStart &&
2622 "tried to reset schedule on block which has not been scheduled");
2623 for (Instruction *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
2624 ScheduleData *SD = getScheduleData(I);
2625 assert(isInSchedulingRegion(SD));
2626 SD->IsScheduled = false;
2627 SD->resetUnscheduledDeps();
2632 void BoUpSLP::scheduleBlock(BlockScheduling *BS) {
2634 if (!BS->ScheduleStart)
2637 DEBUG(dbgs() << "SLP: schedule block " << BS->BB->getName() << "\n");
2639 BS->resetSchedule();
2641 // For the real scheduling we use a more sophisticated ready-list: it is
2642 // sorted by the original instruction location. This lets the final schedule
2643 // be as close as possible to the original instruction order.
2644 struct ScheduleDataCompare {
2645 bool operator()(ScheduleData *SD1, ScheduleData *SD2) {
2646 return SD2->SchedulingPriority < SD1->SchedulingPriority;
2649 std::set<ScheduleData *, ScheduleDataCompare> ReadyInsts;
2651 // Ensure that all depencency data is updated and fill the ready-list with
2652 // initial instructions.
2654 int NumToSchedule = 0;
2655 for (auto *I = BS->ScheduleStart; I != BS->ScheduleEnd;
2656 I = I->getNextNode()) {
2657 ScheduleData *SD = BS->getScheduleData(I);
2659 SD->isPartOfBundle() == (ScalarToTreeEntry.count(SD->Inst) != 0) &&
2660 "scheduler and vectorizer have different opinion on what is a bundle");
2661 SD->FirstInBundle->SchedulingPriority = Idx++;
2662 if (SD->isSchedulingEntity()) {
2663 BS->calculateDependencies(SD, false, AA);
2667 BS->initialFillReadyList(ReadyInsts);
2669 Instruction *LastScheduledInst = BS->ScheduleEnd;
2671 // Do the "real" scheduling.
2672 while (!ReadyInsts.empty()) {
2673 ScheduleData *picked = *ReadyInsts.begin();
2674 ReadyInsts.erase(ReadyInsts.begin());
2676 // Move the scheduled instruction(s) to their dedicated places, if not
2678 ScheduleData *BundleMember = picked;
2679 while (BundleMember) {
2680 Instruction *pickedInst = BundleMember->Inst;
2681 if (LastScheduledInst->getNextNode() != pickedInst) {
2682 BS->BB->getInstList().remove(pickedInst);
2683 BS->BB->getInstList().insert(LastScheduledInst, pickedInst);
2685 LastScheduledInst = pickedInst;
2686 BundleMember = BundleMember->NextInBundle;
2689 BS->schedule(picked, ReadyInsts);
2692 assert(NumToSchedule == 0 && "could not schedule all instructions");
2694 // Avoid duplicate scheduling of the block.
2695 BS->ScheduleStart = nullptr;
2698 /// The SLPVectorizer Pass.
2699 struct SLPVectorizer : public FunctionPass {
2700 typedef SmallVector<StoreInst *, 8> StoreList;
2701 typedef MapVector<Value *, StoreList> StoreListMap;
2703 /// Pass identification, replacement for typeid
2706 explicit SLPVectorizer() : FunctionPass(ID) {
2707 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
2710 ScalarEvolution *SE;
2711 const DataLayout *DL;
2712 TargetTransformInfo *TTI;
2713 TargetLibraryInfo *TLI;
2718 bool runOnFunction(Function &F) override {
2719 if (skipOptnoneFunction(F))
2722 SE = &getAnalysis<ScalarEvolution>();
2723 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2724 DL = DLP ? &DLP->getDataLayout() : nullptr;
2725 TTI = &getAnalysis<TargetTransformInfo>();
2726 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
2727 AA = &getAnalysis<AliasAnalysis>();
2728 LI = &getAnalysis<LoopInfo>();
2729 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2732 bool Changed = false;
2734 // If the target claims to have no vector registers don't attempt
2736 if (!TTI->getNumberOfRegisters(true))
2739 // Must have DataLayout. We can't require it because some tests run w/o
2744 // Don't vectorize when the attribute NoImplicitFloat is used.
2745 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
2748 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
2750 // Use the bottom up slp vectorizer to construct chains that start with
2751 // store instructions.
2752 BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT);
2754 // Scan the blocks in the function in post order.
2755 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
2756 e = po_end(&F.getEntryBlock()); it != e; ++it) {
2757 BasicBlock *BB = *it;
2758 // Vectorize trees that end at stores.
2759 if (unsigned count = collectStores(BB, R)) {
2761 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
2762 Changed |= vectorizeStoreChains(R);
2765 // Vectorize trees that end at reductions.
2766 Changed |= vectorizeChainsInBlock(BB, R);
2770 R.optimizeGatherSequence();
2771 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
2772 DEBUG(verifyFunction(F));
2777 void getAnalysisUsage(AnalysisUsage &AU) const override {
2778 FunctionPass::getAnalysisUsage(AU);
2779 AU.addRequired<ScalarEvolution>();
2780 AU.addRequired<AliasAnalysis>();
2781 AU.addRequired<TargetTransformInfo>();
2782 AU.addRequired<LoopInfo>();
2783 AU.addRequired<DominatorTreeWrapperPass>();
2784 AU.addPreserved<LoopInfo>();
2785 AU.addPreserved<DominatorTreeWrapperPass>();
2786 AU.setPreservesCFG();
2791 /// \brief Collect memory references and sort them according to their base
2792 /// object. We sort the stores to their base objects to reduce the cost of the
2793 /// quadratic search on the stores. TODO: We can further reduce this cost
2794 /// if we flush the chain creation every time we run into a memory barrier.
2795 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
2797 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
2798 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
2800 /// \brief Try to vectorize a list of operands.
2801 /// \@param BuildVector A list of users to ignore for the purpose of
2802 /// scheduling and that don't need extracting.
2803 /// \returns true if a value was vectorized.
2804 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2805 ArrayRef<Value *> BuildVector = None,
2806 bool allowReorder = false);
2808 /// \brief Try to vectorize a chain that may start at the operands of \V;
2809 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
2811 /// \brief Vectorize the stores that were collected in StoreRefs.
2812 bool vectorizeStoreChains(BoUpSLP &R);
2814 /// \brief Scan the basic block and look for patterns that are likely to start
2815 /// a vectorization chain.
2816 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
2818 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
2821 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
2824 StoreListMap StoreRefs;
2827 /// \brief Check that the Values in the slice in VL array are still existent in
2828 /// the WeakVH array.
2829 /// Vectorization of part of the VL array may cause later values in the VL array
2830 /// to become invalid. We track when this has happened in the WeakVH array.
2831 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
2832 SmallVectorImpl<WeakVH> &VH,
2833 unsigned SliceBegin,
2834 unsigned SliceSize) {
2835 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
2842 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
2843 int CostThreshold, BoUpSLP &R) {
2844 unsigned ChainLen = Chain.size();
2845 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2847 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2848 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2849 unsigned VF = MinVecRegSize / Sz;
2851 if (!isPowerOf2_32(Sz) || VF < 2)
2854 // Keep track of values that were deleted by vectorizing in the loop below.
2855 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2857 bool Changed = false;
2858 // Look for profitable vectorizable trees at all offsets, starting at zero.
2859 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2863 // Check that a previous iteration of this loop did not delete the Value.
2864 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2867 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2869 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2871 R.buildTree(Operands);
2873 int Cost = R.getTreeCost();
2875 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2876 if (Cost < CostThreshold) {
2877 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2880 // Move to the next bundle.
2889 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2890 int costThreshold, BoUpSLP &R) {
2891 SetVector<Value *> Heads, Tails;
2892 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2894 // We may run into multiple chains that merge into a single chain. We mark the
2895 // stores that we vectorized so that we don't visit the same store twice.
2896 BoUpSLP::ValueSet VectorizedStores;
2897 bool Changed = false;
2899 // Do a quadratic search on all of the given stores and find
2900 // all of the pairs of stores that follow each other.
2901 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2902 for (unsigned j = 0; j < e; ++j) {
2906 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2907 Tails.insert(Stores[j]);
2908 Heads.insert(Stores[i]);
2909 ConsecutiveChain[Stores[i]] = Stores[j];
2914 // For stores that start but don't end a link in the chain:
2915 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2917 if (Tails.count(*it))
2920 // We found a store instr that starts a chain. Now follow the chain and try
2922 BoUpSLP::ValueList Operands;
2924 // Collect the chain into a list.
2925 while (Tails.count(I) || Heads.count(I)) {
2926 if (VectorizedStores.count(I))
2928 Operands.push_back(I);
2929 // Move to the next value in the chain.
2930 I = ConsecutiveChain[I];
2933 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2935 // Mark the vectorized stores so that we don't vectorize them again.
2937 VectorizedStores.insert(Operands.begin(), Operands.end());
2938 Changed |= Vectorized;
2945 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2948 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2949 StoreInst *SI = dyn_cast<StoreInst>(it);
2953 // Don't touch volatile stores.
2954 if (!SI->isSimple())
2957 // Check that the pointer points to scalars.
2958 Type *Ty = SI->getValueOperand()->getType();
2959 if (Ty->isAggregateType() || Ty->isVectorTy())
2962 // Find the base pointer.
2963 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2965 // Save the store locations.
2966 StoreRefs[Ptr].push_back(SI);
2972 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2975 Value *VL[] = { A, B };
2976 return tryToVectorizeList(VL, R, None, true);
2979 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
2980 ArrayRef<Value *> BuildVector,
2981 bool allowReorder) {
2985 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2987 // Check that all of the parts are scalar instructions of the same type.
2988 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2992 unsigned Opcode0 = I0->getOpcode();
2994 Type *Ty0 = I0->getType();
2995 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2996 unsigned VF = MinVecRegSize / Sz;
2998 for (int i = 0, e = VL.size(); i < e; ++i) {
2999 Type *Ty = VL[i]->getType();
3000 if (Ty->isAggregateType() || Ty->isVectorTy())
3002 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
3003 if (!Inst || Inst->getOpcode() != Opcode0)
3007 bool Changed = false;
3009 // Keep track of values that were deleted by vectorizing in the loop below.
3010 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
3012 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
3013 unsigned OpsWidth = 0;
3020 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
3023 // Check that a previous iteration of this loop did not delete the Value.
3024 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
3027 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
3029 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
3031 ArrayRef<Value *> BuildVectorSlice;
3032 if (!BuildVector.empty())
3033 BuildVectorSlice = BuildVector.slice(i, OpsWidth);
3035 R.buildTree(Ops, BuildVectorSlice);
3036 // TODO: check if we can allow reordering also for other cases than
3037 // tryToVectorizePair()
3038 if (allowReorder && R.shouldReorder()) {
3039 assert(Ops.size() == 2);
3040 assert(BuildVectorSlice.empty());
3041 Value *ReorderedOps[] = { Ops[1], Ops[0] };
3042 R.buildTree(ReorderedOps, None);
3044 int Cost = R.getTreeCost();
3046 if (Cost < -SLPCostThreshold) {
3047 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
3048 Value *VectorizedRoot = R.vectorizeTree();
3050 // Reconstruct the build vector by extracting the vectorized root. This
3051 // way we handle the case where some elements of the vector are undefined.
3052 // (return (inserelt <4 xi32> (insertelt undef (opd0) 0) (opd1) 2))
3053 if (!BuildVectorSlice.empty()) {
3054 // The insert point is the last build vector instruction. The vectorized
3055 // root will precede it. This guarantees that we get an instruction. The
3056 // vectorized tree could have been constant folded.
3057 Instruction *InsertAfter = cast<Instruction>(BuildVectorSlice.back());
3058 unsigned VecIdx = 0;
3059 for (auto &V : BuildVectorSlice) {
3060 IRBuilder<true, NoFolder> Builder(
3061 ++BasicBlock::iterator(InsertAfter));
3062 InsertElementInst *IE = cast<InsertElementInst>(V);
3063 Instruction *Extract = cast<Instruction>(Builder.CreateExtractElement(
3064 VectorizedRoot, Builder.getInt32(VecIdx++)));
3065 IE->setOperand(1, Extract);
3066 IE->removeFromParent();
3067 IE->insertAfter(Extract);
3071 // Move to the next bundle.
3080 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
3084 // Try to vectorize V.
3085 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
3088 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
3089 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
3091 if (B && B->hasOneUse()) {
3092 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
3093 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
3094 if (tryToVectorizePair(A, B0, R)) {
3098 if (tryToVectorizePair(A, B1, R)) {
3105 if (A && A->hasOneUse()) {
3106 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
3107 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
3108 if (tryToVectorizePair(A0, B, R)) {
3112 if (tryToVectorizePair(A1, B, R)) {
3120 /// \brief Generate a shuffle mask to be used in a reduction tree.
3122 /// \param VecLen The length of the vector to be reduced.
3123 /// \param NumEltsToRdx The number of elements that should be reduced in the
3125 /// \param IsPairwise Whether the reduction is a pairwise or splitting
3126 /// reduction. A pairwise reduction will generate a mask of
3127 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
3128 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
3129 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
3130 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
3131 bool IsPairwise, bool IsLeft,
3132 IRBuilder<> &Builder) {
3133 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
3135 SmallVector<Constant *, 32> ShuffleMask(
3136 VecLen, UndefValue::get(Builder.getInt32Ty()));
3139 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
3140 for (unsigned i = 0; i != NumEltsToRdx; ++i)
3141 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
3143 // Move the upper half of the vector to the lower half.
3144 for (unsigned i = 0; i != NumEltsToRdx; ++i)
3145 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
3147 return ConstantVector::get(ShuffleMask);
3151 /// Model horizontal reductions.
3153 /// A horizontal reduction is a tree of reduction operations (currently add and
3154 /// fadd) that has operations that can be put into a vector as its leaf.
3155 /// For example, this tree:
3162 /// This tree has "mul" as its reduced values and "+" as its reduction
3163 /// operations. A reduction might be feeding into a store or a binary operation
3178 class HorizontalReduction {
3179 SmallVector<Value *, 16> ReductionOps;
3180 SmallVector<Value *, 32> ReducedVals;
3182 BinaryOperator *ReductionRoot;
3183 PHINode *ReductionPHI;
3185 /// The opcode of the reduction.
3186 unsigned ReductionOpcode;
3187 /// The opcode of the values we perform a reduction on.
3188 unsigned ReducedValueOpcode;
3189 /// The width of one full horizontal reduction operation.
3190 unsigned ReduxWidth;
3191 /// Should we model this reduction as a pairwise reduction tree or a tree that
3192 /// splits the vector in halves and adds those halves.
3193 bool IsPairwiseReduction;
3196 HorizontalReduction()
3197 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
3198 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
3200 /// \brief Try to find a reduction tree.
3201 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
3202 const DataLayout *DL) {
3204 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
3205 "Thi phi needs to use the binary operator");
3207 // We could have a initial reductions that is not an add.
3208 // r *= v1 + v2 + v3 + v4
3209 // In such a case start looking for a tree rooted in the first '+'.
3211 if (B->getOperand(0) == Phi) {
3213 B = dyn_cast<BinaryOperator>(B->getOperand(1));
3214 } else if (B->getOperand(1) == Phi) {
3216 B = dyn_cast<BinaryOperator>(B->getOperand(0));
3223 Type *Ty = B->getType();
3224 if (Ty->isVectorTy())
3227 ReductionOpcode = B->getOpcode();
3228 ReducedValueOpcode = 0;
3229 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
3236 // We currently only support adds.
3237 if (ReductionOpcode != Instruction::Add &&
3238 ReductionOpcode != Instruction::FAdd)
3241 // Post order traverse the reduction tree starting at B. We only handle true
3242 // trees containing only binary operators.
3243 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
3244 Stack.push_back(std::make_pair(B, 0));
3245 while (!Stack.empty()) {
3246 BinaryOperator *TreeN = Stack.back().first;
3247 unsigned EdgeToVist = Stack.back().second++;
3248 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
3250 // Only handle trees in the current basic block.
3251 if (TreeN->getParent() != B->getParent())
3254 // Each tree node needs to have one user except for the ultimate
3256 if (!TreeN->hasOneUse() && TreeN != B)
3260 if (EdgeToVist == 2 || IsReducedValue) {
3261 if (IsReducedValue) {
3262 // Make sure that the opcodes of the operations that we are going to
3264 if (!ReducedValueOpcode)
3265 ReducedValueOpcode = TreeN->getOpcode();
3266 else if (ReducedValueOpcode != TreeN->getOpcode())
3268 ReducedVals.push_back(TreeN);
3270 // We need to be able to reassociate the adds.
3271 if (!TreeN->isAssociative())
3273 ReductionOps.push_back(TreeN);
3280 // Visit left or right.
3281 Value *NextV = TreeN->getOperand(EdgeToVist);
3282 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
3284 Stack.push_back(std::make_pair(Next, 0));
3285 else if (NextV != Phi)
3291 /// \brief Attempt to vectorize the tree found by
3292 /// matchAssociativeReduction.
3293 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
3294 if (ReducedVals.empty())
3297 unsigned NumReducedVals = ReducedVals.size();
3298 if (NumReducedVals < ReduxWidth)
3301 Value *VectorizedTree = nullptr;
3302 IRBuilder<> Builder(ReductionRoot);
3303 FastMathFlags Unsafe;
3304 Unsafe.setUnsafeAlgebra();
3305 Builder.SetFastMathFlags(Unsafe);
3308 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
3309 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
3310 V.buildTree(ValsToReduce, ReductionOps);
3313 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
3314 if (Cost >= -SLPCostThreshold)
3317 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
3320 // Vectorize a tree.
3321 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
3322 Value *VectorizedRoot = V.vectorizeTree();
3324 // Emit a reduction.
3325 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
3326 if (VectorizedTree) {
3327 Builder.SetCurrentDebugLocation(Loc);
3328 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
3329 ReducedSubTree, "bin.rdx");
3331 VectorizedTree = ReducedSubTree;
3334 if (VectorizedTree) {
3335 // Finish the reduction.
3336 for (; i < NumReducedVals; ++i) {
3337 Builder.SetCurrentDebugLocation(
3338 cast<Instruction>(ReducedVals[i])->getDebugLoc());
3339 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
3344 assert(ReductionRoot && "Need a reduction operation");
3345 ReductionRoot->setOperand(0, VectorizedTree);
3346 ReductionRoot->setOperand(1, ReductionPHI);
3348 ReductionRoot->replaceAllUsesWith(VectorizedTree);
3350 return VectorizedTree != nullptr;
3355 /// \brief Calcuate the cost of a reduction.
3356 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
3357 Type *ScalarTy = FirstReducedVal->getType();
3358 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
3360 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
3361 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
3363 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
3364 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
3366 int ScalarReduxCost =
3367 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
3369 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
3370 << " for reduction that starts with " << *FirstReducedVal
3372 << (IsPairwiseReduction ? "pairwise" : "splitting")
3373 << " reduction)\n");
3375 return VecReduxCost - ScalarReduxCost;
3378 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
3379 Value *R, const Twine &Name = "") {
3380 if (Opcode == Instruction::FAdd)
3381 return Builder.CreateFAdd(L, R, Name);
3382 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
3385 /// \brief Emit a horizontal reduction of the vectorized value.
3386 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
3387 assert(VectorizedValue && "Need to have a vectorized tree node");
3388 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
3389 assert(isPowerOf2_32(ReduxWidth) &&
3390 "We only handle power-of-two reductions for now");
3392 Value *TmpVec = ValToReduce;
3393 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
3394 if (IsPairwiseReduction) {
3396 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
3398 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
3400 Value *LeftShuf = Builder.CreateShuffleVector(
3401 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
3402 Value *RightShuf = Builder.CreateShuffleVector(
3403 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
3405 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
3409 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
3410 Value *Shuf = Builder.CreateShuffleVector(
3411 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
3412 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
3416 // The result is in the first element of the vector.
3417 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
3421 /// \brief Recognize construction of vectors like
3422 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
3423 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
3424 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
3425 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
3427 /// Returns true if it matches
3429 static bool findBuildVector(InsertElementInst *FirstInsertElem,
3430 SmallVectorImpl<Value *> &BuildVector,
3431 SmallVectorImpl<Value *> &BuildVectorOpds) {
3432 if (!isa<UndefValue>(FirstInsertElem->getOperand(0)))
3435 InsertElementInst *IE = FirstInsertElem;
3437 BuildVector.push_back(IE);
3438 BuildVectorOpds.push_back(IE->getOperand(1));
3440 if (IE->use_empty())
3443 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
3447 // If this isn't the final use, make sure the next insertelement is the only
3448 // use. It's OK if the final constructed vector is used multiple times
3449 if (!IE->hasOneUse())
3458 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
3459 return V->getType() < V2->getType();
3462 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
3463 bool Changed = false;
3464 SmallVector<Value *, 4> Incoming;
3465 SmallSet<Value *, 16> VisitedInstrs;
3467 bool HaveVectorizedPhiNodes = true;
3468 while (HaveVectorizedPhiNodes) {
3469 HaveVectorizedPhiNodes = false;
3471 // Collect the incoming values from the PHIs.
3473 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
3475 PHINode *P = dyn_cast<PHINode>(instr);
3479 if (!VisitedInstrs.count(P))
3480 Incoming.push_back(P);
3484 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
3486 // Try to vectorize elements base on their type.
3487 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
3491 // Look for the next elements with the same type.
3492 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
3493 while (SameTypeIt != E &&
3494 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
3495 VisitedInstrs.insert(*SameTypeIt);
3499 // Try to vectorize them.
3500 unsigned NumElts = (SameTypeIt - IncIt);
3501 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
3503 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
3504 // Success start over because instructions might have been changed.
3505 HaveVectorizedPhiNodes = true;
3510 // Start over at the next instruction of a different type (or the end).
3515 VisitedInstrs.clear();
3517 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
3518 // We may go through BB multiple times so skip the one we have checked.
3519 if (!VisitedInstrs.insert(it))
3522 if (isa<DbgInfoIntrinsic>(it))
3525 // Try to vectorize reductions that use PHINodes.
3526 if (PHINode *P = dyn_cast<PHINode>(it)) {
3527 // Check that the PHI is a reduction PHI.
3528 if (P->getNumIncomingValues() != 2)
3531 (P->getIncomingBlock(0) == BB
3532 ? (P->getIncomingValue(0))
3533 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
3535 // Check if this is a Binary Operator.
3536 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
3540 // Try to match and vectorize a horizontal reduction.
3541 HorizontalReduction HorRdx;
3542 if (ShouldVectorizeHor &&
3543 HorRdx.matchAssociativeReduction(P, BI, DL) &&
3544 HorRdx.tryToReduce(R, TTI)) {
3551 Value *Inst = BI->getOperand(0);
3553 Inst = BI->getOperand(1);
3555 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
3556 // We would like to start over since some instructions are deleted
3557 // and the iterator may become invalid value.
3567 // Try to vectorize horizontal reductions feeding into a store.
3568 if (ShouldStartVectorizeHorAtStore)
3569 if (StoreInst *SI = dyn_cast<StoreInst>(it))
3570 if (BinaryOperator *BinOp =
3571 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
3572 HorizontalReduction HorRdx;
3573 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
3574 HorRdx.tryToReduce(R, TTI)) ||
3575 tryToVectorize(BinOp, R))) {
3583 // Try to vectorize trees that start at compare instructions.
3584 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
3585 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
3587 // We would like to start over since some instructions are deleted
3588 // and the iterator may become invalid value.
3594 for (int i = 0; i < 2; ++i) {
3595 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
3596 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
3598 // We would like to start over since some instructions are deleted
3599 // and the iterator may become invalid value.
3608 // Try to vectorize trees that start at insertelement instructions.
3609 if (InsertElementInst *FirstInsertElem = dyn_cast<InsertElementInst>(it)) {
3610 SmallVector<Value *, 16> BuildVector;
3611 SmallVector<Value *, 16> BuildVectorOpds;
3612 if (!findBuildVector(FirstInsertElem, BuildVector, BuildVectorOpds))
3615 // Vectorize starting with the build vector operands ignoring the
3616 // BuildVector instructions for the purpose of scheduling and user
3618 if (tryToVectorizeList(BuildVectorOpds, R, BuildVector)) {
3631 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
3632 bool Changed = false;
3633 // Attempt to sort and vectorize each of the store-groups.
3634 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
3636 if (it->second.size() < 2)
3639 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
3640 << it->second.size() << ".\n");
3642 // Process the stores in chunks of 16.
3643 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
3644 unsigned Len = std::min<unsigned>(CE - CI, 16);
3645 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
3646 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
3652 } // end anonymous namespace
3654 char SLPVectorizer::ID = 0;
3655 static const char lv_name[] = "SLP Vectorizer";
3656 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
3657 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
3658 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
3659 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
3660 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
3661 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
3664 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }