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/Analysis/AliasAnalysis.h"
23 #include "llvm/Analysis/LoopInfo.h"
24 #include "llvm/Analysis/ScalarEvolution.h"
25 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
26 #include "llvm/Analysis/TargetTransformInfo.h"
27 #include "llvm/Analysis/ValueTracking.h"
28 #include "llvm/IR/DataLayout.h"
29 #include "llvm/IR/Dominators.h"
30 #include "llvm/IR/IRBuilder.h"
31 #include "llvm/IR/Instructions.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/Module.h"
34 #include "llvm/IR/Type.h"
35 #include "llvm/IR/Value.h"
36 #include "llvm/IR/Verifier.h"
37 #include "llvm/Pass.h"
38 #include "llvm/Support/CommandLine.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Transforms/Utils/VectorUtils.h"
47 #define SV_NAME "slp-vectorizer"
48 #define DEBUG_TYPE "SLP"
51 SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
52 cl::desc("Only vectorize if you gain more than this "
56 ShouldVectorizeHor("slp-vectorize-hor", cl::init(false), cl::Hidden,
57 cl::desc("Attempt to vectorize horizontal reductions"));
59 static cl::opt<bool> ShouldStartVectorizeHorAtStore(
60 "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
62 "Attempt to vectorize horizontal reductions feeding into a store"));
66 static const unsigned MinVecRegSize = 128;
68 static const unsigned RecursionMaxDepth = 12;
70 /// A helper class for numbering instructions in multiple blocks.
71 /// Numbers start at zero for each basic block.
72 struct BlockNumbering {
74 BlockNumbering(BasicBlock *Bb) : BB(Bb), Valid(false) {}
76 BlockNumbering() : BB(nullptr), Valid(false) {}
78 void numberInstructions() {
82 // Number the instructions in the block.
83 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
85 InstrVec.push_back(it);
86 assert(InstrVec[InstrIdx[it]] == it && "Invalid allocation");
91 int getIndex(Instruction *I) {
92 assert(I->getParent() == BB && "Invalid instruction");
95 assert(InstrIdx.count(I) && "Unknown instruction");
99 Instruction *getInstruction(unsigned loc) {
101 numberInstructions();
102 assert(InstrVec.size() > loc && "Invalid Index");
103 return InstrVec[loc];
106 void forget() { Valid = false; }
109 /// The block we are numbering.
111 /// Is the block numbered.
113 /// Maps instructions to numbers and back.
114 SmallDenseMap<Instruction *, int> InstrIdx;
115 /// Maps integers to Instructions.
116 SmallVector<Instruction *, 32> InstrVec;
119 /// \returns the parent basic block if all of the instructions in \p VL
120 /// are in the same block or null otherwise.
121 static BasicBlock *getSameBlock(ArrayRef<Value *> VL) {
122 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
125 BasicBlock *BB = I0->getParent();
126 for (int i = 1, e = VL.size(); i < e; i++) {
127 Instruction *I = dyn_cast<Instruction>(VL[i]);
131 if (BB != I->getParent())
137 /// \returns True if all of the values in \p VL are constants.
138 static bool allConstant(ArrayRef<Value *> VL) {
139 for (unsigned i = 0, e = VL.size(); i < e; ++i)
140 if (!isa<Constant>(VL[i]))
145 /// \returns True if all of the values in \p VL are identical.
146 static bool isSplat(ArrayRef<Value *> VL) {
147 for (unsigned i = 1, e = VL.size(); i < e; ++i)
153 /// \returns The opcode if all of the Instructions in \p VL have the same
155 static unsigned getSameOpcode(ArrayRef<Value *> VL) {
156 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
159 unsigned Opcode = I0->getOpcode();
160 for (int i = 1, e = VL.size(); i < e; i++) {
161 Instruction *I = dyn_cast<Instruction>(VL[i]);
162 if (!I || Opcode != I->getOpcode())
168 /// \returns \p I after propagating metadata from \p VL.
169 static Instruction *propagateMetadata(Instruction *I, ArrayRef<Value *> VL) {
170 Instruction *I0 = cast<Instruction>(VL[0]);
171 SmallVector<std::pair<unsigned, MDNode *>, 4> Metadata;
172 I0->getAllMetadataOtherThanDebugLoc(Metadata);
174 for (unsigned i = 0, n = Metadata.size(); i != n; ++i) {
175 unsigned Kind = Metadata[i].first;
176 MDNode *MD = Metadata[i].second;
178 for (int i = 1, e = VL.size(); MD && i != e; i++) {
179 Instruction *I = cast<Instruction>(VL[i]);
180 MDNode *IMD = I->getMetadata(Kind);
184 MD = nullptr; // Remove unknown metadata
186 case LLVMContext::MD_tbaa:
187 MD = MDNode::getMostGenericTBAA(MD, IMD);
189 case LLVMContext::MD_fpmath:
190 MD = MDNode::getMostGenericFPMath(MD, IMD);
194 I->setMetadata(Kind, MD);
199 /// \returns The type that all of the values in \p VL have or null if there
200 /// are different types.
201 static Type* getSameType(ArrayRef<Value *> VL) {
202 Type *Ty = VL[0]->getType();
203 for (int i = 1, e = VL.size(); i < e; i++)
204 if (VL[i]->getType() != Ty)
210 /// \returns True if the ExtractElement instructions in VL can be vectorized
211 /// to use the original vector.
212 static bool CanReuseExtract(ArrayRef<Value *> VL) {
213 assert(Instruction::ExtractElement == getSameOpcode(VL) && "Invalid opcode");
214 // Check if all of the extracts come from the same vector and from the
217 ExtractElementInst *E0 = cast<ExtractElementInst>(VL0);
218 Value *Vec = E0->getOperand(0);
220 // We have to extract from the same vector type.
221 unsigned NElts = Vec->getType()->getVectorNumElements();
223 if (NElts != VL.size())
226 // Check that all of the indices extract from the correct offset.
227 ConstantInt *CI = dyn_cast<ConstantInt>(E0->getOperand(1));
228 if (!CI || CI->getZExtValue())
231 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
232 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
233 ConstantInt *CI = dyn_cast<ConstantInt>(E->getOperand(1));
235 if (!CI || CI->getZExtValue() != i || E->getOperand(0) != Vec)
242 static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
243 SmallVectorImpl<Value *> &Left,
244 SmallVectorImpl<Value *> &Right) {
246 SmallVector<Value *, 16> OrigLeft, OrigRight;
248 bool AllSameOpcodeLeft = true;
249 bool AllSameOpcodeRight = true;
250 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
251 Instruction *I = cast<Instruction>(VL[i]);
252 Value *V0 = I->getOperand(0);
253 Value *V1 = I->getOperand(1);
255 OrigLeft.push_back(V0);
256 OrigRight.push_back(V1);
258 Instruction *I0 = dyn_cast<Instruction>(V0);
259 Instruction *I1 = dyn_cast<Instruction>(V1);
261 // Check whether all operands on one side have the same opcode. In this case
262 // we want to preserve the original order and not make things worse by
264 AllSameOpcodeLeft = I0;
265 AllSameOpcodeRight = I1;
267 if (i && AllSameOpcodeLeft) {
268 if(Instruction *P0 = dyn_cast<Instruction>(OrigLeft[i-1])) {
269 if(P0->getOpcode() != I0->getOpcode())
270 AllSameOpcodeLeft = false;
272 AllSameOpcodeLeft = false;
274 if (i && AllSameOpcodeRight) {
275 if(Instruction *P1 = dyn_cast<Instruction>(OrigRight[i-1])) {
276 if(P1->getOpcode() != I1->getOpcode())
277 AllSameOpcodeRight = false;
279 AllSameOpcodeRight = false;
282 // Sort two opcodes. In the code below we try to preserve the ability to use
283 // broadcast of values instead of individual inserts.
290 // If we just sorted according to opcode we would leave the first line in
291 // tact but we would swap vl2 with vr2 because opcode(phi) > opcode(load).
294 // Because vr2 and vr1 are from the same load we loose the opportunity of a
295 // broadcast for the packed right side in the backend: we have [vr1, vl2]
296 // instead of [vr1, vr2=vr1].
298 if(!i && I0->getOpcode() > I1->getOpcode()) {
301 } else if (i && I0->getOpcode() > I1->getOpcode() && Right[i-1] != I1) {
302 // Try not to destroy a broad cast for no apparent benefit.
305 } else if (i && I0->getOpcode() == I1->getOpcode() && Right[i-1] == I0) {
306 // Try preserve broadcasts.
309 } else if (i && I0->getOpcode() == I1->getOpcode() && Left[i-1] == I1) {
310 // Try preserve broadcasts.
319 // One opcode, put the instruction on the right.
329 bool LeftBroadcast = isSplat(Left);
330 bool RightBroadcast = isSplat(Right);
332 // Don't reorder if the operands where good to begin with.
333 if (!(LeftBroadcast || RightBroadcast) &&
334 (AllSameOpcodeRight || AllSameOpcodeLeft)) {
340 /// Bottom Up SLP Vectorizer.
343 typedef SmallVector<Value *, 8> ValueList;
344 typedef SmallVector<Instruction *, 16> InstrList;
345 typedef SmallPtrSet<Value *, 16> ValueSet;
346 typedef SmallVector<StoreInst *, 8> StoreList;
348 BoUpSLP(Function *Func, ScalarEvolution *Se, const DataLayout *Dl,
349 TargetTransformInfo *Tti, TargetLibraryInfo *TLi, AliasAnalysis *Aa, LoopInfo *Li,
351 F(Func), SE(Se), DL(Dl), TTI(Tti), TLI(TLi), AA(Aa), LI(Li), DT(Dt),
352 Builder(Se->getContext()) {
353 // Setup the block numbering utility for all of the blocks in the
355 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
357 BlocksNumbers[BB] = BlockNumbering(BB);
361 /// \brief Vectorize the tree that starts with the elements in \p VL.
362 /// Returns the vectorized root.
363 Value *vectorizeTree();
365 /// \returns the vectorization cost of the subtree that starts at \p VL.
366 /// A negative number means that this is profitable.
369 /// Construct a vectorizable tree that starts at \p Roots and is possibly
370 /// used by a reduction of \p RdxOps.
371 void buildTree(ArrayRef<Value *> Roots, ValueSet *RdxOps = 0);
373 /// Clear the internal data structures that are created by 'buildTree'.
376 VectorizableTree.clear();
377 ScalarToTreeEntry.clear();
379 ExternalUses.clear();
380 MemBarrierIgnoreList.clear();
383 /// \returns true if the memory operations A and B are consecutive.
384 bool isConsecutiveAccess(Value *A, Value *B);
386 /// \brief Perform LICM and CSE on the newly generated gather sequences.
387 void optimizeGatherSequence();
391 /// \returns the cost of the vectorizable entry.
392 int getEntryCost(TreeEntry *E);
394 /// This is the recursive part of buildTree.
395 void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth);
397 /// Vectorize a single entry in the tree.
398 Value *vectorizeTree(TreeEntry *E);
400 /// Vectorize a single entry in the tree, starting in \p VL.
401 Value *vectorizeTree(ArrayRef<Value *> VL);
403 /// \returns the pointer to the vectorized value if \p VL is already
404 /// vectorized, or NULL. They may happen in cycles.
405 Value *alreadyVectorized(ArrayRef<Value *> VL) const;
407 /// \brief Take the pointer operand from the Load/Store instruction.
408 /// \returns NULL if this is not a valid Load/Store instruction.
409 static Value *getPointerOperand(Value *I);
411 /// \brief Take the address space operand from the Load/Store instruction.
412 /// \returns -1 if this is not a valid Load/Store instruction.
413 static unsigned getAddressSpaceOperand(Value *I);
415 /// \returns the scalarization cost for this type. Scalarization in this
416 /// context means the creation of vectors from a group of scalars.
417 int getGatherCost(Type *Ty);
419 /// \returns the scalarization cost for this list of values. Assuming that
420 /// this subtree gets vectorized, we may need to extract the values from the
421 /// roots. This method calculates the cost of extracting the values.
422 int getGatherCost(ArrayRef<Value *> VL);
424 /// \returns the AA location that is being access by the instruction.
425 AliasAnalysis::Location getLocation(Instruction *I);
427 /// \brief Checks if it is possible to sink an instruction from
428 /// \p Src to \p Dst.
429 /// \returns the pointer to the barrier instruction if we can't sink.
430 Value *getSinkBarrier(Instruction *Src, Instruction *Dst);
432 /// \returns the index of the last instruction in the BB from \p VL.
433 int getLastIndex(ArrayRef<Value *> VL);
435 /// \returns the Instruction in the bundle \p VL.
436 Instruction *getLastInstruction(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), LastScalarIndex(0),
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 /// The index in the basic block of the last scalar.
468 /// Do we need to gather this sequence ?
472 /// Create a new VectorizableTree entry.
473 TreeEntry *newTreeEntry(ArrayRef<Value *> VL, bool Vectorized) {
474 VectorizableTree.push_back(TreeEntry());
475 int idx = VectorizableTree.size() - 1;
476 TreeEntry *Last = &VectorizableTree[idx];
477 Last->Scalars.insert(Last->Scalars.begin(), VL.begin(), VL.end());
478 Last->NeedToGather = !Vectorized;
480 Last->LastScalarIndex = getLastIndex(VL);
481 for (int i = 0, e = VL.size(); i != e; ++i) {
482 assert(!ScalarToTreeEntry.count(VL[i]) && "Scalar already in tree!");
483 ScalarToTreeEntry[VL[i]] = idx;
486 Last->LastScalarIndex = 0;
487 MustGather.insert(VL.begin(), VL.end());
492 /// -- Vectorization State --
493 /// Holds all of the tree entries.
494 std::vector<TreeEntry> VectorizableTree;
496 /// Maps a specific scalar to its tree entry.
497 SmallDenseMap<Value*, int> ScalarToTreeEntry;
499 /// A list of scalars that we found that we need to keep as scalars.
502 /// This POD struct describes one external user in the vectorized tree.
503 struct ExternalUser {
504 ExternalUser (Value *S, llvm::User *U, int L) :
505 Scalar(S), User(U), Lane(L){};
506 // Which scalar in our function.
508 // Which user that uses the scalar.
510 // Which lane does the scalar belong to.
513 typedef SmallVector<ExternalUser, 16> UserList;
515 /// A list of values that need to extracted out of the tree.
516 /// This list holds pairs of (Internal Scalar : External User).
517 UserList ExternalUses;
519 /// A list of instructions to ignore while sinking
520 /// memory instructions. This map must be reset between runs of getCost.
521 ValueSet MemBarrierIgnoreList;
523 /// Holds all of the instructions that we gathered.
524 SetVector<Instruction *> GatherSeq;
525 /// A list of blocks that we are going to CSE.
526 SetVector<BasicBlock *> CSEBlocks;
528 /// Numbers instructions in different blocks.
529 DenseMap<BasicBlock *, BlockNumbering> BlocksNumbers;
531 /// Reduction operators.
534 // Analysis and block reference.
537 const DataLayout *DL;
538 TargetTransformInfo *TTI;
539 TargetLibraryInfo *TLI;
543 /// Instruction builder to construct the vectorized tree.
547 void BoUpSLP::buildTree(ArrayRef<Value *> Roots, ValueSet *Rdx) {
550 if (!getSameType(Roots))
552 buildTree_rec(Roots, 0);
554 // Collect the values that we need to extract from the tree.
555 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
556 TreeEntry *Entry = &VectorizableTree[EIdx];
559 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
560 Value *Scalar = Entry->Scalars[Lane];
562 // No need to handle users of gathered values.
563 if (Entry->NeedToGather)
566 for (User *U : Scalar->users()) {
567 DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
569 // Skip in-tree scalars that become vectors.
570 if (ScalarToTreeEntry.count(U)) {
571 DEBUG(dbgs() << "SLP: \tInternal user will be removed:" <<
573 int Idx = ScalarToTreeEntry[U]; (void) Idx;
574 assert(!VectorizableTree[Idx].NeedToGather && "Bad state");
577 Instruction *UserInst = dyn_cast<Instruction>(U);
581 // Ignore uses that are part of the reduction.
582 if (Rdx && std::find(Rdx->begin(), Rdx->end(), UserInst) != Rdx->end())
585 DEBUG(dbgs() << "SLP: Need to extract:" << *U << " from lane " <<
586 Lane << " from " << *Scalar << ".\n");
587 ExternalUses.push_back(ExternalUser(Scalar, U, Lane));
594 void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth) {
595 bool SameTy = getSameType(VL); (void)SameTy;
596 assert(SameTy && "Invalid types!");
598 if (Depth == RecursionMaxDepth) {
599 DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
600 newTreeEntry(VL, false);
604 // Don't handle vectors.
605 if (VL[0]->getType()->isVectorTy()) {
606 DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
607 newTreeEntry(VL, false);
611 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
612 if (SI->getValueOperand()->getType()->isVectorTy()) {
613 DEBUG(dbgs() << "SLP: Gathering due to store vector type.\n");
614 newTreeEntry(VL, false);
618 // If all of the operands are identical or constant we have a simple solution.
619 if (allConstant(VL) || isSplat(VL) || !getSameBlock(VL) ||
620 !getSameOpcode(VL)) {
621 DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O. \n");
622 newTreeEntry(VL, false);
626 // We now know that this is a vector of instructions of the same type from
629 // Check if this is a duplicate of another entry.
630 if (ScalarToTreeEntry.count(VL[0])) {
631 int Idx = ScalarToTreeEntry[VL[0]];
632 TreeEntry *E = &VectorizableTree[Idx];
633 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
634 DEBUG(dbgs() << "SLP: \tChecking bundle: " << *VL[i] << ".\n");
635 if (E->Scalars[i] != VL[i]) {
636 DEBUG(dbgs() << "SLP: Gathering due to partial overlap.\n");
637 newTreeEntry(VL, false);
641 DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *VL[0] << ".\n");
645 // Check that none of the instructions in the bundle are already in the tree.
646 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
647 if (ScalarToTreeEntry.count(VL[i])) {
648 DEBUG(dbgs() << "SLP: The instruction (" << *VL[i] <<
649 ") is already in tree.\n");
650 newTreeEntry(VL, false);
655 // If any of the scalars appears in the table OR it is marked as a value that
656 // needs to stat scalar then we need to gather the scalars.
657 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
658 if (ScalarToTreeEntry.count(VL[i]) || MustGather.count(VL[i])) {
659 DEBUG(dbgs() << "SLP: Gathering due to gathered scalar. \n");
660 newTreeEntry(VL, false);
665 // Check that all of the users of the scalars that we want to vectorize are
667 Instruction *VL0 = cast<Instruction>(VL[0]);
668 int MyLastIndex = getLastIndex(VL);
669 BasicBlock *BB = cast<Instruction>(VL0)->getParent();
671 for (unsigned i = 0, e = VL.size(); i != e; ++i) {
672 Instruction *Scalar = cast<Instruction>(VL[i]);
673 DEBUG(dbgs() << "SLP: Checking users of " << *Scalar << ". \n");
674 for (User *U : Scalar->users()) {
675 DEBUG(dbgs() << "SLP: \tUser " << *U << ". \n");
676 Instruction *UI = dyn_cast<Instruction>(U);
678 DEBUG(dbgs() << "SLP: Gathering due unknown user. \n");
679 newTreeEntry(VL, false);
683 // We don't care if the user is in a different basic block.
684 BasicBlock *UserBlock = UI->getParent();
685 if (UserBlock != BB) {
686 DEBUG(dbgs() << "SLP: User from a different basic block "
691 // If this is a PHINode within this basic block then we can place the
692 // extract wherever we want.
693 if (isa<PHINode>(*UI)) {
694 DEBUG(dbgs() << "SLP: \tWe can schedule PHIs:" << *UI << ". \n");
698 // Check if this is a safe in-tree user.
699 if (ScalarToTreeEntry.count(UI)) {
700 int Idx = ScalarToTreeEntry[UI];
701 int VecLocation = VectorizableTree[Idx].LastScalarIndex;
702 if (VecLocation <= MyLastIndex) {
703 DEBUG(dbgs() << "SLP: Gathering due to unschedulable vector. \n");
704 newTreeEntry(VL, false);
707 DEBUG(dbgs() << "SLP: In-tree user (" << *UI << ") at #" <<
708 VecLocation << " vector value (" << *Scalar << ") at #"
709 << MyLastIndex << ".\n");
713 // This user is part of the reduction.
714 if (RdxOps && RdxOps->count(UI))
717 // Make sure that we can schedule this unknown user.
718 BlockNumbering &BN = BlocksNumbers[BB];
719 int UserIndex = BN.getIndex(UI);
720 if (UserIndex < MyLastIndex) {
722 DEBUG(dbgs() << "SLP: Can't schedule extractelement for "
724 newTreeEntry(VL, false);
730 // Check that every instructions appears once in this bundle.
731 for (unsigned i = 0, e = VL.size(); i < e; ++i)
732 for (unsigned j = i+1; j < e; ++j)
733 if (VL[i] == VL[j]) {
734 DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
735 newTreeEntry(VL, false);
739 // Check that instructions in this bundle don't reference other instructions.
740 // The runtime of this check is O(N * N-1 * uses(N)) and a typical N is 4.
741 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
742 for (User *U : VL[i]->users()) {
743 for (unsigned j = 0; j < e; ++j) {
744 if (i != j && U == VL[j]) {
745 DEBUG(dbgs() << "SLP: Intra-bundle dependencies!" << *U << ". \n");
746 newTreeEntry(VL, false);
753 DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
755 unsigned Opcode = getSameOpcode(VL);
757 // Check if it is safe to sink the loads or the stores.
758 if (Opcode == Instruction::Load || Opcode == Instruction::Store) {
759 Instruction *Last = getLastInstruction(VL);
761 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
764 Value *Barrier = getSinkBarrier(cast<Instruction>(VL[i]), Last);
766 DEBUG(dbgs() << "SLP: Can't sink " << *VL[i] << "\n down to " << *Last
767 << "\n because of " << *Barrier << ". Gathering.\n");
768 newTreeEntry(VL, false);
775 case Instruction::PHI: {
776 PHINode *PH = dyn_cast<PHINode>(VL0);
778 // Check for terminator values (e.g. invoke).
779 for (unsigned j = 0; j < VL.size(); ++j)
780 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
781 TerminatorInst *Term = dyn_cast<TerminatorInst>(
782 cast<PHINode>(VL[j])->getIncomingValueForBlock(PH->getIncomingBlock(i)));
784 DEBUG(dbgs() << "SLP: Need to swizzle PHINodes (TerminatorInst use).\n");
785 newTreeEntry(VL, false);
790 newTreeEntry(VL, true);
791 DEBUG(dbgs() << "SLP: added a vector of PHINodes.\n");
793 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
795 // Prepare the operand vector.
796 for (unsigned j = 0; j < VL.size(); ++j)
797 Operands.push_back(cast<PHINode>(VL[j])->getIncomingValueForBlock(
798 PH->getIncomingBlock(i)));
800 buildTree_rec(Operands, Depth + 1);
804 case Instruction::ExtractElement: {
805 bool Reuse = CanReuseExtract(VL);
807 DEBUG(dbgs() << "SLP: Reusing extract sequence.\n");
809 newTreeEntry(VL, Reuse);
812 case Instruction::Load: {
813 // Check if the loads are consecutive or of we need to swizzle them.
814 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i) {
815 LoadInst *L = cast<LoadInst>(VL[i]);
816 if (!L->isSimple() || !isConsecutiveAccess(VL[i], VL[i + 1])) {
817 newTreeEntry(VL, false);
818 DEBUG(dbgs() << "SLP: Need to swizzle loads.\n");
822 newTreeEntry(VL, true);
823 DEBUG(dbgs() << "SLP: added a vector of loads.\n");
826 case Instruction::ZExt:
827 case Instruction::SExt:
828 case Instruction::FPToUI:
829 case Instruction::FPToSI:
830 case Instruction::FPExt:
831 case Instruction::PtrToInt:
832 case Instruction::IntToPtr:
833 case Instruction::SIToFP:
834 case Instruction::UIToFP:
835 case Instruction::Trunc:
836 case Instruction::FPTrunc:
837 case Instruction::BitCast: {
838 Type *SrcTy = VL0->getOperand(0)->getType();
839 for (unsigned i = 0; i < VL.size(); ++i) {
840 Type *Ty = cast<Instruction>(VL[i])->getOperand(0)->getType();
841 if (Ty != SrcTy || Ty->isAggregateType() || Ty->isVectorTy()) {
842 newTreeEntry(VL, false);
843 DEBUG(dbgs() << "SLP: Gathering casts with different src types.\n");
847 newTreeEntry(VL, true);
848 DEBUG(dbgs() << "SLP: added a vector of casts.\n");
850 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
852 // Prepare the operand vector.
853 for (unsigned j = 0; j < VL.size(); ++j)
854 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
856 buildTree_rec(Operands, Depth+1);
860 case Instruction::ICmp:
861 case Instruction::FCmp: {
862 // Check that all of the compares have the same predicate.
863 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
864 Type *ComparedTy = cast<Instruction>(VL[0])->getOperand(0)->getType();
865 for (unsigned i = 1, e = VL.size(); i < e; ++i) {
866 CmpInst *Cmp = cast<CmpInst>(VL[i]);
867 if (Cmp->getPredicate() != P0 ||
868 Cmp->getOperand(0)->getType() != ComparedTy) {
869 newTreeEntry(VL, false);
870 DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
875 newTreeEntry(VL, true);
876 DEBUG(dbgs() << "SLP: added a vector of compares.\n");
878 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
880 // Prepare the operand vector.
881 for (unsigned j = 0; j < VL.size(); ++j)
882 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
884 buildTree_rec(Operands, Depth+1);
888 case Instruction::Select:
889 case Instruction::Add:
890 case Instruction::FAdd:
891 case Instruction::Sub:
892 case Instruction::FSub:
893 case Instruction::Mul:
894 case Instruction::FMul:
895 case Instruction::UDiv:
896 case Instruction::SDiv:
897 case Instruction::FDiv:
898 case Instruction::URem:
899 case Instruction::SRem:
900 case Instruction::FRem:
901 case Instruction::Shl:
902 case Instruction::LShr:
903 case Instruction::AShr:
904 case Instruction::And:
905 case Instruction::Or:
906 case Instruction::Xor: {
907 newTreeEntry(VL, true);
908 DEBUG(dbgs() << "SLP: added a vector of bin op.\n");
910 // Sort operands of the instructions so that each side is more likely to
911 // have the same opcode.
912 if (isa<BinaryOperator>(VL0) && VL0->isCommutative()) {
913 ValueList Left, Right;
914 reorderInputsAccordingToOpcode(VL, Left, Right);
915 buildTree_rec(Left, Depth + 1);
916 buildTree_rec(Right, Depth + 1);
920 for (unsigned i = 0, e = VL0->getNumOperands(); i < e; ++i) {
922 // Prepare the operand vector.
923 for (unsigned j = 0; j < VL.size(); ++j)
924 Operands.push_back(cast<Instruction>(VL[j])->getOperand(i));
926 buildTree_rec(Operands, Depth+1);
930 case Instruction::Store: {
931 // Check if the stores are consecutive or of we need to swizzle them.
932 for (unsigned i = 0, e = VL.size() - 1; i < e; ++i)
933 if (!isConsecutiveAccess(VL[i], VL[i + 1])) {
934 newTreeEntry(VL, false);
935 DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
939 newTreeEntry(VL, true);
940 DEBUG(dbgs() << "SLP: added a vector of stores.\n");
943 for (unsigned j = 0; j < VL.size(); ++j)
944 Operands.push_back(cast<Instruction>(VL[j])->getOperand(0));
946 // We can ignore these values because we are sinking them down.
947 MemBarrierIgnoreList.insert(VL.begin(), VL.end());
948 buildTree_rec(Operands, Depth + 1);
951 case Instruction::Call: {
952 // Check if the calls are all to the same vectorizable intrinsic.
953 CallInst *CI = cast<CallInst>(VL[0]);
954 // Check if this is an Intrinsic call or something that can be
955 // represented by an intrinsic call
956 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
957 if (!isTriviallyVectorizable(ID)) {
958 newTreeEntry(VL, false);
959 DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
963 Function *Int = CI->getCalledFunction();
965 for (unsigned i = 1, e = VL.size(); i != e; ++i) {
966 CallInst *CI2 = dyn_cast<CallInst>(VL[i]);
967 if (!CI2 || CI2->getCalledFunction() != Int ||
968 getIntrinsicIDForCall(CI2, TLI) != ID) {
969 newTreeEntry(VL, false);
970 DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *VL[i]
976 newTreeEntry(VL, true);
977 for (unsigned i = 0, e = CI->getNumArgOperands(); i != e; ++i) {
979 // Prepare the operand vector.
980 for (unsigned j = 0; j < VL.size(); ++j) {
981 CallInst *CI2 = dyn_cast<CallInst>(VL[j]);
982 Operands.push_back(CI2->getArgOperand(i));
984 buildTree_rec(Operands, Depth + 1);
989 newTreeEntry(VL, false);
990 DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
995 int BoUpSLP::getEntryCost(TreeEntry *E) {
996 ArrayRef<Value*> VL = E->Scalars;
998 Type *ScalarTy = VL[0]->getType();
999 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1000 ScalarTy = SI->getValueOperand()->getType();
1001 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1003 if (E->NeedToGather) {
1004 if (allConstant(VL))
1007 return TTI->getShuffleCost(TargetTransformInfo::SK_Broadcast, VecTy, 0);
1009 return getGatherCost(E->Scalars);
1012 assert(getSameOpcode(VL) && getSameType(VL) && getSameBlock(VL) &&
1014 Instruction *VL0 = cast<Instruction>(VL[0]);
1015 unsigned Opcode = VL0->getOpcode();
1017 case Instruction::PHI: {
1020 case Instruction::ExtractElement: {
1021 if (CanReuseExtract(VL)) {
1023 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
1024 ExtractElementInst *E = cast<ExtractElementInst>(VL[i]);
1026 // Take credit for instruction that will become dead.
1028 TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, i);
1032 return getGatherCost(VecTy);
1034 case Instruction::ZExt:
1035 case Instruction::SExt:
1036 case Instruction::FPToUI:
1037 case Instruction::FPToSI:
1038 case Instruction::FPExt:
1039 case Instruction::PtrToInt:
1040 case Instruction::IntToPtr:
1041 case Instruction::SIToFP:
1042 case Instruction::UIToFP:
1043 case Instruction::Trunc:
1044 case Instruction::FPTrunc:
1045 case Instruction::BitCast: {
1046 Type *SrcTy = VL0->getOperand(0)->getType();
1048 // Calculate the cost of this instruction.
1049 int ScalarCost = VL.size() * TTI->getCastInstrCost(VL0->getOpcode(),
1050 VL0->getType(), SrcTy);
1052 VectorType *SrcVecTy = VectorType::get(SrcTy, VL.size());
1053 int VecCost = TTI->getCastInstrCost(VL0->getOpcode(), VecTy, SrcVecTy);
1054 return VecCost - ScalarCost;
1056 case Instruction::FCmp:
1057 case Instruction::ICmp:
1058 case Instruction::Select:
1059 case Instruction::Add:
1060 case Instruction::FAdd:
1061 case Instruction::Sub:
1062 case Instruction::FSub:
1063 case Instruction::Mul:
1064 case Instruction::FMul:
1065 case Instruction::UDiv:
1066 case Instruction::SDiv:
1067 case Instruction::FDiv:
1068 case Instruction::URem:
1069 case Instruction::SRem:
1070 case Instruction::FRem:
1071 case Instruction::Shl:
1072 case Instruction::LShr:
1073 case Instruction::AShr:
1074 case Instruction::And:
1075 case Instruction::Or:
1076 case Instruction::Xor: {
1077 // Calculate the cost of this instruction.
1080 if (Opcode == Instruction::FCmp || Opcode == Instruction::ICmp ||
1081 Opcode == Instruction::Select) {
1082 VectorType *MaskTy = VectorType::get(Builder.getInt1Ty(), VL.size());
1083 ScalarCost = VecTy->getNumElements() *
1084 TTI->getCmpSelInstrCost(Opcode, ScalarTy, Builder.getInt1Ty());
1085 VecCost = TTI->getCmpSelInstrCost(Opcode, VecTy, MaskTy);
1087 // Certain instructions can be cheaper to vectorize if they have a
1088 // constant second vector operand.
1089 TargetTransformInfo::OperandValueKind Op1VK =
1090 TargetTransformInfo::OK_AnyValue;
1091 TargetTransformInfo::OperandValueKind Op2VK =
1092 TargetTransformInfo::OK_UniformConstantValue;
1094 // If all operands are exactly the same ConstantInt then set the
1095 // operand kind to OK_UniformConstantValue.
1096 // If instead not all operands are constants, then set the operand kind
1097 // to OK_AnyValue. If all operands are constants but not the same,
1098 // then set the operand kind to OK_NonUniformConstantValue.
1099 ConstantInt *CInt = nullptr;
1100 for (unsigned i = 0; i < VL.size(); ++i) {
1101 const Instruction *I = cast<Instruction>(VL[i]);
1102 if (!isa<ConstantInt>(I->getOperand(1))) {
1103 Op2VK = TargetTransformInfo::OK_AnyValue;
1107 CInt = cast<ConstantInt>(I->getOperand(1));
1110 if (Op2VK == TargetTransformInfo::OK_UniformConstantValue &&
1111 CInt != cast<ConstantInt>(I->getOperand(1)))
1112 Op2VK = TargetTransformInfo::OK_NonUniformConstantValue;
1116 VecTy->getNumElements() *
1117 TTI->getArithmeticInstrCost(Opcode, ScalarTy, Op1VK, Op2VK);
1118 VecCost = TTI->getArithmeticInstrCost(Opcode, VecTy, Op1VK, Op2VK);
1120 return VecCost - ScalarCost;
1122 case Instruction::Load: {
1123 // Cost of wide load - cost of scalar loads.
1124 int ScalarLdCost = VecTy->getNumElements() *
1125 TTI->getMemoryOpCost(Instruction::Load, ScalarTy, 1, 0);
1126 int VecLdCost = TTI->getMemoryOpCost(Instruction::Load, VecTy, 1, 0);
1127 return VecLdCost - ScalarLdCost;
1129 case Instruction::Store: {
1130 // We know that we can merge the stores. Calculate the cost.
1131 int ScalarStCost = VecTy->getNumElements() *
1132 TTI->getMemoryOpCost(Instruction::Store, ScalarTy, 1, 0);
1133 int VecStCost = TTI->getMemoryOpCost(Instruction::Store, VecTy, 1, 0);
1134 return VecStCost - ScalarStCost;
1136 case Instruction::Call: {
1137 CallInst *CI = cast<CallInst>(VL0);
1138 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1140 // Calculate the cost of the scalar and vector calls.
1141 SmallVector<Type*, 4> ScalarTys, VecTys;
1142 for (unsigned op = 0, opc = CI->getNumArgOperands(); op!= opc; ++op) {
1143 ScalarTys.push_back(CI->getArgOperand(op)->getType());
1144 VecTys.push_back(VectorType::get(CI->getArgOperand(op)->getType(),
1145 VecTy->getNumElements()));
1148 int ScalarCallCost = VecTy->getNumElements() *
1149 TTI->getIntrinsicInstrCost(ID, ScalarTy, ScalarTys);
1151 int VecCallCost = TTI->getIntrinsicInstrCost(ID, VecTy, VecTys);
1153 DEBUG(dbgs() << "SLP: Call cost "<< VecCallCost - ScalarCallCost
1154 << " (" << VecCallCost << "-" << ScalarCallCost << ")"
1155 << " for " << *CI << "\n");
1157 return VecCallCost - ScalarCallCost;
1160 llvm_unreachable("Unknown instruction");
1164 bool BoUpSLP::isFullyVectorizableTinyTree() {
1165 DEBUG(dbgs() << "SLP: Check whether the tree with height " <<
1166 VectorizableTree.size() << " is fully vectorizable .\n");
1168 // We only handle trees of height 2.
1169 if (VectorizableTree.size() != 2)
1172 // Handle splat stores.
1173 if (!VectorizableTree[0].NeedToGather && isSplat(VectorizableTree[1].Scalars))
1176 // Gathering cost would be too much for tiny trees.
1177 if (VectorizableTree[0].NeedToGather || VectorizableTree[1].NeedToGather)
1183 int BoUpSLP::getTreeCost() {
1185 DEBUG(dbgs() << "SLP: Calculating cost for tree of size " <<
1186 VectorizableTree.size() << ".\n");
1188 // We only vectorize tiny trees if it is fully vectorizable.
1189 if (VectorizableTree.size() < 3 && !isFullyVectorizableTinyTree()) {
1190 if (!VectorizableTree.size()) {
1191 assert(!ExternalUses.size() && "We should not have any external users");
1196 unsigned BundleWidth = VectorizableTree[0].Scalars.size();
1198 for (unsigned i = 0, e = VectorizableTree.size(); i != e; ++i) {
1199 int C = getEntryCost(&VectorizableTree[i]);
1200 DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle that starts with "
1201 << *VectorizableTree[i].Scalars[0] << " .\n");
1205 SmallSet<Value *, 16> ExtractCostCalculated;
1206 int ExtractCost = 0;
1207 for (UserList::iterator I = ExternalUses.begin(), E = ExternalUses.end();
1209 // We only add extract cost once for the same scalar.
1210 if (!ExtractCostCalculated.insert(I->Scalar))
1213 VectorType *VecTy = VectorType::get(I->Scalar->getType(), BundleWidth);
1214 ExtractCost += TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy,
1218 DEBUG(dbgs() << "SLP: Total Cost " << Cost + ExtractCost<< ".\n");
1219 return Cost + ExtractCost;
1222 int BoUpSLP::getGatherCost(Type *Ty) {
1224 for (unsigned i = 0, e = cast<VectorType>(Ty)->getNumElements(); i < e; ++i)
1225 Cost += TTI->getVectorInstrCost(Instruction::InsertElement, Ty, i);
1229 int BoUpSLP::getGatherCost(ArrayRef<Value *> VL) {
1230 // Find the type of the operands in VL.
1231 Type *ScalarTy = VL[0]->getType();
1232 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1233 ScalarTy = SI->getValueOperand()->getType();
1234 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1235 // Find the cost of inserting/extracting values from the vector.
1236 return getGatherCost(VecTy);
1239 AliasAnalysis::Location BoUpSLP::getLocation(Instruction *I) {
1240 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1241 return AA->getLocation(SI);
1242 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1243 return AA->getLocation(LI);
1244 return AliasAnalysis::Location();
1247 Value *BoUpSLP::getPointerOperand(Value *I) {
1248 if (LoadInst *LI = dyn_cast<LoadInst>(I))
1249 return LI->getPointerOperand();
1250 if (StoreInst *SI = dyn_cast<StoreInst>(I))
1251 return SI->getPointerOperand();
1255 unsigned BoUpSLP::getAddressSpaceOperand(Value *I) {
1256 if (LoadInst *L = dyn_cast<LoadInst>(I))
1257 return L->getPointerAddressSpace();
1258 if (StoreInst *S = dyn_cast<StoreInst>(I))
1259 return S->getPointerAddressSpace();
1263 bool BoUpSLP::isConsecutiveAccess(Value *A, Value *B) {
1264 Value *PtrA = getPointerOperand(A);
1265 Value *PtrB = getPointerOperand(B);
1266 unsigned ASA = getAddressSpaceOperand(A);
1267 unsigned ASB = getAddressSpaceOperand(B);
1269 // Check that the address spaces match and that the pointers are valid.
1270 if (!PtrA || !PtrB || (ASA != ASB))
1273 // Make sure that A and B are different pointers of the same type.
1274 if (PtrA == PtrB || PtrA->getType() != PtrB->getType())
1277 unsigned PtrBitWidth = DL->getPointerSizeInBits(ASA);
1278 Type *Ty = cast<PointerType>(PtrA->getType())->getElementType();
1279 APInt Size(PtrBitWidth, DL->getTypeStoreSize(Ty));
1281 APInt OffsetA(PtrBitWidth, 0), OffsetB(PtrBitWidth, 0);
1282 PtrA = PtrA->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetA);
1283 PtrB = PtrB->stripAndAccumulateInBoundsConstantOffsets(*DL, OffsetB);
1285 APInt OffsetDelta = OffsetB - OffsetA;
1287 // Check if they are based on the same pointer. That makes the offsets
1290 return OffsetDelta == Size;
1292 // Compute the necessary base pointer delta to have the necessary final delta
1293 // equal to the size.
1294 APInt BaseDelta = Size - OffsetDelta;
1296 // Otherwise compute the distance with SCEV between the base pointers.
1297 const SCEV *PtrSCEVA = SE->getSCEV(PtrA);
1298 const SCEV *PtrSCEVB = SE->getSCEV(PtrB);
1299 const SCEV *C = SE->getConstant(BaseDelta);
1300 const SCEV *X = SE->getAddExpr(PtrSCEVA, C);
1301 return X == PtrSCEVB;
1304 Value *BoUpSLP::getSinkBarrier(Instruction *Src, Instruction *Dst) {
1305 assert(Src->getParent() == Dst->getParent() && "Not the same BB");
1306 BasicBlock::iterator I = Src, E = Dst;
1307 /// Scan all of the instruction from SRC to DST and check if
1308 /// the source may alias.
1309 for (++I; I != E; ++I) {
1310 // Ignore store instructions that are marked as 'ignore'.
1311 if (MemBarrierIgnoreList.count(I))
1313 if (Src->mayWriteToMemory()) /* Write */ {
1314 if (!I->mayReadOrWriteMemory())
1317 if (!I->mayWriteToMemory())
1320 AliasAnalysis::Location A = getLocation(&*I);
1321 AliasAnalysis::Location B = getLocation(Src);
1323 if (!A.Ptr || !B.Ptr || AA->alias(A, B))
1329 int BoUpSLP::getLastIndex(ArrayRef<Value *> VL) {
1330 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1331 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1332 BlockNumbering &BN = BlocksNumbers[BB];
1334 int MaxIdx = BN.getIndex(BB->getFirstNonPHI());
1335 for (unsigned i = 0, e = VL.size(); i < e; ++i)
1336 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1340 Instruction *BoUpSLP::getLastInstruction(ArrayRef<Value *> VL) {
1341 BasicBlock *BB = cast<Instruction>(VL[0])->getParent();
1342 assert(BB == getSameBlock(VL) && BlocksNumbers.count(BB) && "Invalid block");
1343 BlockNumbering &BN = BlocksNumbers[BB];
1345 int MaxIdx = BN.getIndex(cast<Instruction>(VL[0]));
1346 for (unsigned i = 1, e = VL.size(); i < e; ++i)
1347 MaxIdx = std::max(MaxIdx, BN.getIndex(cast<Instruction>(VL[i])));
1348 Instruction *I = BN.getInstruction(MaxIdx);
1349 assert(I && "bad location");
1353 void BoUpSLP::setInsertPointAfterBundle(ArrayRef<Value *> VL) {
1354 Instruction *VL0 = cast<Instruction>(VL[0]);
1355 Instruction *LastInst = getLastInstruction(VL);
1356 BasicBlock::iterator NextInst = LastInst;
1358 Builder.SetInsertPoint(VL0->getParent(), NextInst);
1359 Builder.SetCurrentDebugLocation(VL0->getDebugLoc());
1362 Value *BoUpSLP::Gather(ArrayRef<Value *> VL, VectorType *Ty) {
1363 Value *Vec = UndefValue::get(Ty);
1364 // Generate the 'InsertElement' instruction.
1365 for (unsigned i = 0; i < Ty->getNumElements(); ++i) {
1366 Vec = Builder.CreateInsertElement(Vec, VL[i], Builder.getInt32(i));
1367 if (Instruction *Insrt = dyn_cast<Instruction>(Vec)) {
1368 GatherSeq.insert(Insrt);
1369 CSEBlocks.insert(Insrt->getParent());
1371 // Add to our 'need-to-extract' list.
1372 if (ScalarToTreeEntry.count(VL[i])) {
1373 int Idx = ScalarToTreeEntry[VL[i]];
1374 TreeEntry *E = &VectorizableTree[Idx];
1375 // Find which lane we need to extract.
1377 for (unsigned Lane = 0, LE = VL.size(); Lane != LE; ++Lane) {
1378 // Is this the lane of the scalar that we are looking for ?
1379 if (E->Scalars[Lane] == VL[i]) {
1384 assert(FoundLane >= 0 && "Could not find the correct lane");
1385 ExternalUses.push_back(ExternalUser(VL[i], Insrt, FoundLane));
1393 Value *BoUpSLP::alreadyVectorized(ArrayRef<Value *> VL) const {
1394 SmallDenseMap<Value*, int>::const_iterator Entry
1395 = ScalarToTreeEntry.find(VL[0]);
1396 if (Entry != ScalarToTreeEntry.end()) {
1397 int Idx = Entry->second;
1398 const TreeEntry *En = &VectorizableTree[Idx];
1399 if (En->isSame(VL) && En->VectorizedValue)
1400 return En->VectorizedValue;
1405 Value *BoUpSLP::vectorizeTree(ArrayRef<Value *> VL) {
1406 if (ScalarToTreeEntry.count(VL[0])) {
1407 int Idx = ScalarToTreeEntry[VL[0]];
1408 TreeEntry *E = &VectorizableTree[Idx];
1410 return vectorizeTree(E);
1413 Type *ScalarTy = VL[0]->getType();
1414 if (StoreInst *SI = dyn_cast<StoreInst>(VL[0]))
1415 ScalarTy = SI->getValueOperand()->getType();
1416 VectorType *VecTy = VectorType::get(ScalarTy, VL.size());
1418 return Gather(VL, VecTy);
1421 Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
1422 IRBuilder<>::InsertPointGuard Guard(Builder);
1424 if (E->VectorizedValue) {
1425 DEBUG(dbgs() << "SLP: Diamond merged for " << *E->Scalars[0] << ".\n");
1426 return E->VectorizedValue;
1429 Instruction *VL0 = cast<Instruction>(E->Scalars[0]);
1430 Type *ScalarTy = VL0->getType();
1431 if (StoreInst *SI = dyn_cast<StoreInst>(VL0))
1432 ScalarTy = SI->getValueOperand()->getType();
1433 VectorType *VecTy = VectorType::get(ScalarTy, E->Scalars.size());
1435 if (E->NeedToGather) {
1436 setInsertPointAfterBundle(E->Scalars);
1437 return Gather(E->Scalars, VecTy);
1440 unsigned Opcode = VL0->getOpcode();
1441 assert(Opcode == getSameOpcode(E->Scalars) && "Invalid opcode");
1444 case Instruction::PHI: {
1445 PHINode *PH = dyn_cast<PHINode>(VL0);
1446 Builder.SetInsertPoint(PH->getParent()->getFirstNonPHI());
1447 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1448 PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
1449 E->VectorizedValue = NewPhi;
1451 // PHINodes may have multiple entries from the same block. We want to
1452 // visit every block once.
1453 SmallSet<BasicBlock*, 4> VisitedBBs;
1455 for (unsigned i = 0, e = PH->getNumIncomingValues(); i < e; ++i) {
1457 BasicBlock *IBB = PH->getIncomingBlock(i);
1459 if (!VisitedBBs.insert(IBB)) {
1460 NewPhi->addIncoming(NewPhi->getIncomingValueForBlock(IBB), IBB);
1464 // Prepare the operand vector.
1465 for (unsigned j = 0; j < E->Scalars.size(); ++j)
1466 Operands.push_back(cast<PHINode>(E->Scalars[j])->
1467 getIncomingValueForBlock(IBB));
1469 Builder.SetInsertPoint(IBB->getTerminator());
1470 Builder.SetCurrentDebugLocation(PH->getDebugLoc());
1471 Value *Vec = vectorizeTree(Operands);
1472 NewPhi->addIncoming(Vec, IBB);
1475 assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
1476 "Invalid number of incoming values");
1480 case Instruction::ExtractElement: {
1481 if (CanReuseExtract(E->Scalars)) {
1482 Value *V = VL0->getOperand(0);
1483 E->VectorizedValue = V;
1486 return Gather(E->Scalars, VecTy);
1488 case Instruction::ZExt:
1489 case Instruction::SExt:
1490 case Instruction::FPToUI:
1491 case Instruction::FPToSI:
1492 case Instruction::FPExt:
1493 case Instruction::PtrToInt:
1494 case Instruction::IntToPtr:
1495 case Instruction::SIToFP:
1496 case Instruction::UIToFP:
1497 case Instruction::Trunc:
1498 case Instruction::FPTrunc:
1499 case Instruction::BitCast: {
1501 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1502 INVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1504 setInsertPointAfterBundle(E->Scalars);
1506 Value *InVec = vectorizeTree(INVL);
1508 if (Value *V = alreadyVectorized(E->Scalars))
1511 CastInst *CI = dyn_cast<CastInst>(VL0);
1512 Value *V = Builder.CreateCast(CI->getOpcode(), InVec, VecTy);
1513 E->VectorizedValue = V;
1516 case Instruction::FCmp:
1517 case Instruction::ICmp: {
1518 ValueList LHSV, RHSV;
1519 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1520 LHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1521 RHSV.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1524 setInsertPointAfterBundle(E->Scalars);
1526 Value *L = vectorizeTree(LHSV);
1527 Value *R = vectorizeTree(RHSV);
1529 if (Value *V = alreadyVectorized(E->Scalars))
1532 CmpInst::Predicate P0 = dyn_cast<CmpInst>(VL0)->getPredicate();
1534 if (Opcode == Instruction::FCmp)
1535 V = Builder.CreateFCmp(P0, L, R);
1537 V = Builder.CreateICmp(P0, L, R);
1539 E->VectorizedValue = V;
1542 case Instruction::Select: {
1543 ValueList TrueVec, FalseVec, CondVec;
1544 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1545 CondVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1546 TrueVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1547 FalseVec.push_back(cast<Instruction>(E->Scalars[i])->getOperand(2));
1550 setInsertPointAfterBundle(E->Scalars);
1552 Value *Cond = vectorizeTree(CondVec);
1553 Value *True = vectorizeTree(TrueVec);
1554 Value *False = vectorizeTree(FalseVec);
1556 if (Value *V = alreadyVectorized(E->Scalars))
1559 Value *V = Builder.CreateSelect(Cond, True, False);
1560 E->VectorizedValue = V;
1563 case Instruction::Add:
1564 case Instruction::FAdd:
1565 case Instruction::Sub:
1566 case Instruction::FSub:
1567 case Instruction::Mul:
1568 case Instruction::FMul:
1569 case Instruction::UDiv:
1570 case Instruction::SDiv:
1571 case Instruction::FDiv:
1572 case Instruction::URem:
1573 case Instruction::SRem:
1574 case Instruction::FRem:
1575 case Instruction::Shl:
1576 case Instruction::LShr:
1577 case Instruction::AShr:
1578 case Instruction::And:
1579 case Instruction::Or:
1580 case Instruction::Xor: {
1581 ValueList LHSVL, RHSVL;
1582 if (isa<BinaryOperator>(VL0) && VL0->isCommutative())
1583 reorderInputsAccordingToOpcode(E->Scalars, LHSVL, RHSVL);
1585 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1586 LHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(0));
1587 RHSVL.push_back(cast<Instruction>(E->Scalars[i])->getOperand(1));
1590 setInsertPointAfterBundle(E->Scalars);
1592 Value *LHS = vectorizeTree(LHSVL);
1593 Value *RHS = vectorizeTree(RHSVL);
1595 if (LHS == RHS && isa<Instruction>(LHS)) {
1596 assert((VL0->getOperand(0) == VL0->getOperand(1)) && "Invalid order");
1599 if (Value *V = alreadyVectorized(E->Scalars))
1602 BinaryOperator *BinOp = cast<BinaryOperator>(VL0);
1603 Value *V = Builder.CreateBinOp(BinOp->getOpcode(), LHS, RHS);
1604 E->VectorizedValue = V;
1606 if (Instruction *I = dyn_cast<Instruction>(V))
1607 return propagateMetadata(I, E->Scalars);
1611 case Instruction::Load: {
1612 // Loads are inserted at the head of the tree because we don't want to
1613 // sink them all the way down past store instructions.
1614 setInsertPointAfterBundle(E->Scalars);
1616 LoadInst *LI = cast<LoadInst>(VL0);
1617 unsigned AS = LI->getPointerAddressSpace();
1619 Value *VecPtr = Builder.CreateBitCast(LI->getPointerOperand(),
1620 VecTy->getPointerTo(AS));
1621 unsigned Alignment = LI->getAlignment();
1622 LI = Builder.CreateLoad(VecPtr);
1623 LI->setAlignment(Alignment);
1624 E->VectorizedValue = LI;
1625 return propagateMetadata(LI, E->Scalars);
1627 case Instruction::Store: {
1628 StoreInst *SI = cast<StoreInst>(VL0);
1629 unsigned Alignment = SI->getAlignment();
1630 unsigned AS = SI->getPointerAddressSpace();
1633 for (int i = 0, e = E->Scalars.size(); i < e; ++i)
1634 ValueOp.push_back(cast<StoreInst>(E->Scalars[i])->getValueOperand());
1636 setInsertPointAfterBundle(E->Scalars);
1638 Value *VecValue = vectorizeTree(ValueOp);
1639 Value *VecPtr = Builder.CreateBitCast(SI->getPointerOperand(),
1640 VecTy->getPointerTo(AS));
1641 StoreInst *S = Builder.CreateStore(VecValue, VecPtr);
1642 S->setAlignment(Alignment);
1643 E->VectorizedValue = S;
1644 return propagateMetadata(S, E->Scalars);
1646 case Instruction::Call: {
1647 CallInst *CI = cast<CallInst>(VL0);
1648 setInsertPointAfterBundle(E->Scalars);
1649 std::vector<Value *> OpVecs;
1650 for (int j = 0, e = CI->getNumArgOperands(); j < e; ++j) {
1652 for (int i = 0, e = E->Scalars.size(); i < e; ++i) {
1653 CallInst *CEI = cast<CallInst>(E->Scalars[i]);
1654 OpVL.push_back(CEI->getArgOperand(j));
1657 Value *OpVec = vectorizeTree(OpVL);
1658 DEBUG(dbgs() << "SLP: OpVec[" << j << "]: " << *OpVec << "\n");
1659 OpVecs.push_back(OpVec);
1662 Module *M = F->getParent();
1663 Intrinsic::ID ID = getIntrinsicIDForCall(CI, TLI);
1664 Type *Tys[] = { VectorType::get(CI->getType(), E->Scalars.size()) };
1665 Function *CF = Intrinsic::getDeclaration(M, ID, Tys);
1666 Value *V = Builder.CreateCall(CF, OpVecs);
1667 E->VectorizedValue = V;
1671 llvm_unreachable("unknown inst");
1676 Value *BoUpSLP::vectorizeTree() {
1677 Builder.SetInsertPoint(F->getEntryBlock().begin());
1678 vectorizeTree(&VectorizableTree[0]);
1680 DEBUG(dbgs() << "SLP: Extracting " << ExternalUses.size() << " values .\n");
1682 // Extract all of the elements with the external uses.
1683 for (UserList::iterator it = ExternalUses.begin(), e = ExternalUses.end();
1685 Value *Scalar = it->Scalar;
1686 llvm::User *User = it->User;
1688 // Skip users that we already RAUW. This happens when one instruction
1689 // has multiple uses of the same value.
1690 if (std::find(Scalar->user_begin(), Scalar->user_end(), User) ==
1693 assert(ScalarToTreeEntry.count(Scalar) && "Invalid scalar");
1695 int Idx = ScalarToTreeEntry[Scalar];
1696 TreeEntry *E = &VectorizableTree[Idx];
1697 assert(!E->NeedToGather && "Extracting from a gather list");
1699 Value *Vec = E->VectorizedValue;
1700 assert(Vec && "Can't find vectorizable value");
1702 Value *Lane = Builder.getInt32(it->Lane);
1703 // Generate extracts for out-of-tree users.
1704 // Find the insertion point for the extractelement lane.
1705 if (isa<Instruction>(Vec)){
1706 if (PHINode *PH = dyn_cast<PHINode>(User)) {
1707 for (int i = 0, e = PH->getNumIncomingValues(); i != e; ++i) {
1708 if (PH->getIncomingValue(i) == Scalar) {
1709 Builder.SetInsertPoint(PH->getIncomingBlock(i)->getTerminator());
1710 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1711 CSEBlocks.insert(PH->getIncomingBlock(i));
1712 PH->setOperand(i, Ex);
1716 Builder.SetInsertPoint(cast<Instruction>(User));
1717 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1718 CSEBlocks.insert(cast<Instruction>(User)->getParent());
1719 User->replaceUsesOfWith(Scalar, Ex);
1722 Builder.SetInsertPoint(F->getEntryBlock().begin());
1723 Value *Ex = Builder.CreateExtractElement(Vec, Lane);
1724 CSEBlocks.insert(&F->getEntryBlock());
1725 User->replaceUsesOfWith(Scalar, Ex);
1728 DEBUG(dbgs() << "SLP: Replaced:" << *User << ".\n");
1731 // For each vectorized value:
1732 for (int EIdx = 0, EE = VectorizableTree.size(); EIdx < EE; ++EIdx) {
1733 TreeEntry *Entry = &VectorizableTree[EIdx];
1736 for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
1737 Value *Scalar = Entry->Scalars[Lane];
1739 // No need to handle users of gathered values.
1740 if (Entry->NeedToGather)
1743 assert(Entry->VectorizedValue && "Can't find vectorizable value");
1745 Type *Ty = Scalar->getType();
1746 if (!Ty->isVoidTy()) {
1748 for (User *U : Scalar->users()) {
1749 DEBUG(dbgs() << "SLP: \tvalidating user:" << *U << ".\n");
1751 assert((ScalarToTreeEntry.count(U) ||
1752 // It is legal to replace the reduction users by undef.
1753 (RdxOps && RdxOps->count(U))) &&
1754 "Replacing out-of-tree value with undef");
1757 Value *Undef = UndefValue::get(Ty);
1758 Scalar->replaceAllUsesWith(Undef);
1760 DEBUG(dbgs() << "SLP: \tErasing scalar:" << *Scalar << ".\n");
1761 cast<Instruction>(Scalar)->eraseFromParent();
1765 for (Function::iterator it = F->begin(), e = F->end(); it != e; ++it) {
1766 BlocksNumbers[it].forget();
1768 Builder.ClearInsertionPoint();
1770 return VectorizableTree[0].VectorizedValue;
1773 void BoUpSLP::optimizeGatherSequence() {
1774 DEBUG(dbgs() << "SLP: Optimizing " << GatherSeq.size()
1775 << " gather sequences instructions.\n");
1776 // LICM InsertElementInst sequences.
1777 for (SetVector<Instruction *>::iterator it = GatherSeq.begin(),
1778 e = GatherSeq.end(); it != e; ++it) {
1779 InsertElementInst *Insert = dyn_cast<InsertElementInst>(*it);
1784 // Check if this block is inside a loop.
1785 Loop *L = LI->getLoopFor(Insert->getParent());
1789 // Check if it has a preheader.
1790 BasicBlock *PreHeader = L->getLoopPreheader();
1794 // If the vector or the element that we insert into it are
1795 // instructions that are defined in this basic block then we can't
1796 // hoist this instruction.
1797 Instruction *CurrVec = dyn_cast<Instruction>(Insert->getOperand(0));
1798 Instruction *NewElem = dyn_cast<Instruction>(Insert->getOperand(1));
1799 if (CurrVec && L->contains(CurrVec))
1801 if (NewElem && L->contains(NewElem))
1804 // We can hoist this instruction. Move it to the pre-header.
1805 Insert->moveBefore(PreHeader->getTerminator());
1808 // Sort blocks by domination. This ensures we visit a block after all blocks
1809 // dominating it are visited.
1810 SmallVector<BasicBlock *, 8> CSEWorkList(CSEBlocks.begin(), CSEBlocks.end());
1811 std::stable_sort(CSEWorkList.begin(), CSEWorkList.end(),
1812 [this](const BasicBlock *A, const BasicBlock *B) {
1813 return DT->properlyDominates(A, B);
1816 // Perform O(N^2) search over the gather sequences and merge identical
1817 // instructions. TODO: We can further optimize this scan if we split the
1818 // instructions into different buckets based on the insert lane.
1819 SmallVector<Instruction *, 16> Visited;
1820 for (SmallVectorImpl<BasicBlock *>::iterator I = CSEWorkList.begin(),
1821 E = CSEWorkList.end();
1823 assert((I == CSEWorkList.begin() || !DT->dominates(*I, *std::prev(I))) &&
1824 "Worklist not sorted properly!");
1825 BasicBlock *BB = *I;
1826 // For all instructions in blocks containing gather sequences:
1827 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e;) {
1828 Instruction *In = it++;
1829 if (!isa<InsertElementInst>(In) && !isa<ExtractElementInst>(In))
1832 // Check if we can replace this instruction with any of the
1833 // visited instructions.
1834 for (SmallVectorImpl<Instruction *>::iterator v = Visited.begin(),
1837 if (In->isIdenticalTo(*v) &&
1838 DT->dominates((*v)->getParent(), In->getParent())) {
1839 In->replaceAllUsesWith(*v);
1840 In->eraseFromParent();
1846 assert(std::find(Visited.begin(), Visited.end(), In) == Visited.end());
1847 Visited.push_back(In);
1855 /// The SLPVectorizer Pass.
1856 struct SLPVectorizer : public FunctionPass {
1857 typedef SmallVector<StoreInst *, 8> StoreList;
1858 typedef MapVector<Value *, StoreList> StoreListMap;
1860 /// Pass identification, replacement for typeid
1863 explicit SLPVectorizer() : FunctionPass(ID) {
1864 initializeSLPVectorizerPass(*PassRegistry::getPassRegistry());
1867 ScalarEvolution *SE;
1868 const DataLayout *DL;
1869 TargetTransformInfo *TTI;
1870 TargetLibraryInfo *TLI;
1875 bool runOnFunction(Function &F) override {
1876 if (skipOptnoneFunction(F))
1879 SE = &getAnalysis<ScalarEvolution>();
1880 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
1881 DL = DLP ? &DLP->getDataLayout() : nullptr;
1882 TTI = &getAnalysis<TargetTransformInfo>();
1883 TLI = getAnalysisIfAvailable<TargetLibraryInfo>();
1884 AA = &getAnalysis<AliasAnalysis>();
1885 LI = &getAnalysis<LoopInfo>();
1886 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
1889 bool Changed = false;
1891 // If the target claims to have no vector registers don't attempt
1893 if (!TTI->getNumberOfRegisters(true))
1896 // Must have DataLayout. We can't require it because some tests run w/o
1901 // Don't vectorize when the attribute NoImplicitFloat is used.
1902 if (F.hasFnAttribute(Attribute::NoImplicitFloat))
1905 DEBUG(dbgs() << "SLP: Analyzing blocks in " << F.getName() << ".\n");
1907 // Use the bottom up slp vectorizer to construct chains that start with
1908 // he store instructions.
1909 BoUpSLP R(&F, SE, DL, TTI, TLI, AA, LI, DT);
1911 // Scan the blocks in the function in post order.
1912 for (po_iterator<BasicBlock*> it = po_begin(&F.getEntryBlock()),
1913 e = po_end(&F.getEntryBlock()); it != e; ++it) {
1914 BasicBlock *BB = *it;
1916 // Vectorize trees that end at stores.
1917 if (unsigned count = collectStores(BB, R)) {
1919 DEBUG(dbgs() << "SLP: Found " << count << " stores to vectorize.\n");
1920 Changed |= vectorizeStoreChains(R);
1923 // Vectorize trees that end at reductions.
1924 Changed |= vectorizeChainsInBlock(BB, R);
1928 R.optimizeGatherSequence();
1929 DEBUG(dbgs() << "SLP: vectorized \"" << F.getName() << "\"\n");
1930 DEBUG(verifyFunction(F));
1935 void getAnalysisUsage(AnalysisUsage &AU) const override {
1936 FunctionPass::getAnalysisUsage(AU);
1937 AU.addRequired<ScalarEvolution>();
1938 AU.addRequired<AliasAnalysis>();
1939 AU.addRequired<TargetTransformInfo>();
1940 AU.addRequired<LoopInfo>();
1941 AU.addRequired<DominatorTreeWrapperPass>();
1942 AU.addPreserved<LoopInfo>();
1943 AU.addPreserved<DominatorTreeWrapperPass>();
1944 AU.setPreservesCFG();
1949 /// \brief Collect memory references and sort them according to their base
1950 /// object. We sort the stores to their base objects to reduce the cost of the
1951 /// quadratic search on the stores. TODO: We can further reduce this cost
1952 /// if we flush the chain creation every time we run into a memory barrier.
1953 unsigned collectStores(BasicBlock *BB, BoUpSLP &R);
1955 /// \brief Try to vectorize a chain that starts at two arithmetic instrs.
1956 bool tryToVectorizePair(Value *A, Value *B, BoUpSLP &R);
1958 /// \brief Try to vectorize a list of operands.
1959 /// \returns true if a value was vectorized.
1960 bool tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R);
1962 /// \brief Try to vectorize a chain that may start at the operands of \V;
1963 bool tryToVectorize(BinaryOperator *V, BoUpSLP &R);
1965 /// \brief Vectorize the stores that were collected in StoreRefs.
1966 bool vectorizeStoreChains(BoUpSLP &R);
1968 /// \brief Scan the basic block and look for patterns that are likely to start
1969 /// a vectorization chain.
1970 bool vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R);
1972 bool vectorizeStoreChain(ArrayRef<Value *> Chain, int CostThreshold,
1975 bool vectorizeStores(ArrayRef<StoreInst *> Stores, int costThreshold,
1978 StoreListMap StoreRefs;
1981 /// \brief Check that the Values in the slice in VL array are still existent in
1982 /// the WeakVH array.
1983 /// Vectorization of part of the VL array may cause later values in the VL array
1984 /// to become invalid. We track when this has happened in the WeakVH array.
1985 static bool hasValueBeenRAUWed(ArrayRef<Value *> &VL,
1986 SmallVectorImpl<WeakVH> &VH,
1987 unsigned SliceBegin,
1988 unsigned SliceSize) {
1989 for (unsigned i = SliceBegin; i < SliceBegin + SliceSize; ++i)
1996 bool SLPVectorizer::vectorizeStoreChain(ArrayRef<Value *> Chain,
1997 int CostThreshold, BoUpSLP &R) {
1998 unsigned ChainLen = Chain.size();
1999 DEBUG(dbgs() << "SLP: Analyzing a store chain of length " << ChainLen
2001 Type *StoreTy = cast<StoreInst>(Chain[0])->getValueOperand()->getType();
2002 unsigned Sz = DL->getTypeSizeInBits(StoreTy);
2003 unsigned VF = MinVecRegSize / Sz;
2005 if (!isPowerOf2_32(Sz) || VF < 2)
2008 // Keep track of values that were deleted by vectorizing in the loop below.
2009 SmallVector<WeakVH, 8> TrackValues(Chain.begin(), Chain.end());
2011 bool Changed = false;
2012 // Look for profitable vectorizable trees at all offsets, starting at zero.
2013 for (unsigned i = 0, e = ChainLen; i < e; ++i) {
2017 // Check that a previous iteration of this loop did not delete the Value.
2018 if (hasValueBeenRAUWed(Chain, TrackValues, i, VF))
2021 DEBUG(dbgs() << "SLP: Analyzing " << VF << " stores at offset " << i
2023 ArrayRef<Value *> Operands = Chain.slice(i, VF);
2025 R.buildTree(Operands);
2027 int Cost = R.getTreeCost();
2029 DEBUG(dbgs() << "SLP: Found cost=" << Cost << " for VF=" << VF << "\n");
2030 if (Cost < CostThreshold) {
2031 DEBUG(dbgs() << "SLP: Decided to vectorize cost=" << Cost << "\n");
2034 // Move to the next bundle.
2043 bool SLPVectorizer::vectorizeStores(ArrayRef<StoreInst *> Stores,
2044 int costThreshold, BoUpSLP &R) {
2045 SetVector<Value *> Heads, Tails;
2046 SmallDenseMap<Value *, Value *> ConsecutiveChain;
2048 // We may run into multiple chains that merge into a single chain. We mark the
2049 // stores that we vectorized so that we don't visit the same store twice.
2050 BoUpSLP::ValueSet VectorizedStores;
2051 bool Changed = false;
2053 // Do a quadratic search on all of the given stores and find
2054 // all of the pairs of stores that follow each other.
2055 for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
2056 for (unsigned j = 0; j < e; ++j) {
2060 if (R.isConsecutiveAccess(Stores[i], Stores[j])) {
2061 Tails.insert(Stores[j]);
2062 Heads.insert(Stores[i]);
2063 ConsecutiveChain[Stores[i]] = Stores[j];
2068 // For stores that start but don't end a link in the chain:
2069 for (SetVector<Value *>::iterator it = Heads.begin(), e = Heads.end();
2071 if (Tails.count(*it))
2074 // We found a store instr that starts a chain. Now follow the chain and try
2076 BoUpSLP::ValueList Operands;
2078 // Collect the chain into a list.
2079 while (Tails.count(I) || Heads.count(I)) {
2080 if (VectorizedStores.count(I))
2082 Operands.push_back(I);
2083 // Move to the next value in the chain.
2084 I = ConsecutiveChain[I];
2087 bool Vectorized = vectorizeStoreChain(Operands, costThreshold, R);
2089 // Mark the vectorized stores so that we don't vectorize them again.
2091 VectorizedStores.insert(Operands.begin(), Operands.end());
2092 Changed |= Vectorized;
2099 unsigned SLPVectorizer::collectStores(BasicBlock *BB, BoUpSLP &R) {
2102 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
2103 StoreInst *SI = dyn_cast<StoreInst>(it);
2107 // Don't touch volatile stores.
2108 if (!SI->isSimple())
2111 // Check that the pointer points to scalars.
2112 Type *Ty = SI->getValueOperand()->getType();
2113 if (Ty->isAggregateType() || Ty->isVectorTy())
2116 // Find the base pointer.
2117 Value *Ptr = GetUnderlyingObject(SI->getPointerOperand(), DL);
2119 // Save the store locations.
2120 StoreRefs[Ptr].push_back(SI);
2126 bool SLPVectorizer::tryToVectorizePair(Value *A, Value *B, BoUpSLP &R) {
2129 Value *VL[] = { A, B };
2130 return tryToVectorizeList(VL, R);
2133 bool SLPVectorizer::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R) {
2137 DEBUG(dbgs() << "SLP: Vectorizing a list of length = " << VL.size() << ".\n");
2139 // Check that all of the parts are scalar instructions of the same type.
2140 Instruction *I0 = dyn_cast<Instruction>(VL[0]);
2144 unsigned Opcode0 = I0->getOpcode();
2146 Type *Ty0 = I0->getType();
2147 unsigned Sz = DL->getTypeSizeInBits(Ty0);
2148 unsigned VF = MinVecRegSize / Sz;
2150 for (int i = 0, e = VL.size(); i < e; ++i) {
2151 Type *Ty = VL[i]->getType();
2152 if (Ty->isAggregateType() || Ty->isVectorTy())
2154 Instruction *Inst = dyn_cast<Instruction>(VL[i]);
2155 if (!Inst || Inst->getOpcode() != Opcode0)
2159 bool Changed = false;
2161 // Keep track of values that were deleted by vectorizing in the loop below.
2162 SmallVector<WeakVH, 8> TrackValues(VL.begin(), VL.end());
2164 for (unsigned i = 0, e = VL.size(); i < e; ++i) {
2165 unsigned OpsWidth = 0;
2172 if (!isPowerOf2_32(OpsWidth) || OpsWidth < 2)
2175 // Check that a previous iteration of this loop did not delete the Value.
2176 if (hasValueBeenRAUWed(VL, TrackValues, i, OpsWidth))
2179 DEBUG(dbgs() << "SLP: Analyzing " << OpsWidth << " operations "
2181 ArrayRef<Value *> Ops = VL.slice(i, OpsWidth);
2184 int Cost = R.getTreeCost();
2186 if (Cost < -SLPCostThreshold) {
2187 DEBUG(dbgs() << "SLP: Vectorizing list at cost:" << Cost << ".\n");
2190 // Move to the next bundle.
2199 bool SLPVectorizer::tryToVectorize(BinaryOperator *V, BoUpSLP &R) {
2203 // Try to vectorize V.
2204 if (tryToVectorizePair(V->getOperand(0), V->getOperand(1), R))
2207 BinaryOperator *A = dyn_cast<BinaryOperator>(V->getOperand(0));
2208 BinaryOperator *B = dyn_cast<BinaryOperator>(V->getOperand(1));
2210 if (B && B->hasOneUse()) {
2211 BinaryOperator *B0 = dyn_cast<BinaryOperator>(B->getOperand(0));
2212 BinaryOperator *B1 = dyn_cast<BinaryOperator>(B->getOperand(1));
2213 if (tryToVectorizePair(A, B0, R)) {
2217 if (tryToVectorizePair(A, B1, R)) {
2224 if (A && A->hasOneUse()) {
2225 BinaryOperator *A0 = dyn_cast<BinaryOperator>(A->getOperand(0));
2226 BinaryOperator *A1 = dyn_cast<BinaryOperator>(A->getOperand(1));
2227 if (tryToVectorizePair(A0, B, R)) {
2231 if (tryToVectorizePair(A1, B, R)) {
2239 /// \brief Generate a shuffle mask to be used in a reduction tree.
2241 /// \param VecLen The length of the vector to be reduced.
2242 /// \param NumEltsToRdx The number of elements that should be reduced in the
2244 /// \param IsPairwise Whether the reduction is a pairwise or splitting
2245 /// reduction. A pairwise reduction will generate a mask of
2246 /// <0,2,...> or <1,3,..> while a splitting reduction will generate
2247 /// <2,3, undef,undef> for a vector of 4 and NumElts = 2.
2248 /// \param IsLeft True will generate a mask of even elements, odd otherwise.
2249 static Value *createRdxShuffleMask(unsigned VecLen, unsigned NumEltsToRdx,
2250 bool IsPairwise, bool IsLeft,
2251 IRBuilder<> &Builder) {
2252 assert((IsPairwise || !IsLeft) && "Don't support a <0,1,undef,...> mask");
2254 SmallVector<Constant *, 32> ShuffleMask(
2255 VecLen, UndefValue::get(Builder.getInt32Ty()));
2258 // Build a mask of 0, 2, ... (left) or 1, 3, ... (right).
2259 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2260 ShuffleMask[i] = Builder.getInt32(2 * i + !IsLeft);
2262 // Move the upper half of the vector to the lower half.
2263 for (unsigned i = 0; i != NumEltsToRdx; ++i)
2264 ShuffleMask[i] = Builder.getInt32(NumEltsToRdx + i);
2266 return ConstantVector::get(ShuffleMask);
2270 /// Model horizontal reductions.
2272 /// A horizontal reduction is a tree of reduction operations (currently add and
2273 /// fadd) that has operations that can be put into a vector as its leaf.
2274 /// For example, this tree:
2281 /// This tree has "mul" as its reduced values and "+" as its reduction
2282 /// operations. A reduction might be feeding into a store or a binary operation
2297 class HorizontalReduction {
2298 SmallPtrSet<Value *, 16> ReductionOps;
2299 SmallVector<Value *, 32> ReducedVals;
2301 BinaryOperator *ReductionRoot;
2302 PHINode *ReductionPHI;
2304 /// The opcode of the reduction.
2305 unsigned ReductionOpcode;
2306 /// The opcode of the values we perform a reduction on.
2307 unsigned ReducedValueOpcode;
2308 /// The width of one full horizontal reduction operation.
2309 unsigned ReduxWidth;
2310 /// Should we model this reduction as a pairwise reduction tree or a tree that
2311 /// splits the vector in halves and adds those halves.
2312 bool IsPairwiseReduction;
2315 HorizontalReduction()
2316 : ReductionRoot(nullptr), ReductionPHI(nullptr), ReductionOpcode(0),
2317 ReducedValueOpcode(0), ReduxWidth(0), IsPairwiseReduction(false) {}
2319 /// \brief Try to find a reduction tree.
2320 bool matchAssociativeReduction(PHINode *Phi, BinaryOperator *B,
2321 const DataLayout *DL) {
2323 std::find(Phi->op_begin(), Phi->op_end(), B) != Phi->op_end()) &&
2324 "Thi phi needs to use the binary operator");
2326 // We could have a initial reductions that is not an add.
2327 // r *= v1 + v2 + v3 + v4
2328 // In such a case start looking for a tree rooted in the first '+'.
2330 if (B->getOperand(0) == Phi) {
2332 B = dyn_cast<BinaryOperator>(B->getOperand(1));
2333 } else if (B->getOperand(1) == Phi) {
2335 B = dyn_cast<BinaryOperator>(B->getOperand(0));
2342 Type *Ty = B->getType();
2343 if (Ty->isVectorTy())
2346 ReductionOpcode = B->getOpcode();
2347 ReducedValueOpcode = 0;
2348 ReduxWidth = MinVecRegSize / DL->getTypeSizeInBits(Ty);
2355 // We currently only support adds.
2356 if (ReductionOpcode != Instruction::Add &&
2357 ReductionOpcode != Instruction::FAdd)
2360 // Post order traverse the reduction tree starting at B. We only handle true
2361 // trees containing only binary operators.
2362 SmallVector<std::pair<BinaryOperator *, unsigned>, 32> Stack;
2363 Stack.push_back(std::make_pair(B, 0));
2364 while (!Stack.empty()) {
2365 BinaryOperator *TreeN = Stack.back().first;
2366 unsigned EdgeToVist = Stack.back().second++;
2367 bool IsReducedValue = TreeN->getOpcode() != ReductionOpcode;
2369 // Only handle trees in the current basic block.
2370 if (TreeN->getParent() != B->getParent())
2373 // Each tree node needs to have one user except for the ultimate
2375 if (!TreeN->hasOneUse() && TreeN != B)
2379 if (EdgeToVist == 2 || IsReducedValue) {
2380 if (IsReducedValue) {
2381 // Make sure that the opcodes of the operations that we are going to
2383 if (!ReducedValueOpcode)
2384 ReducedValueOpcode = TreeN->getOpcode();
2385 else if (ReducedValueOpcode != TreeN->getOpcode())
2387 ReducedVals.push_back(TreeN);
2389 // We need to be able to reassociate the adds.
2390 if (!TreeN->isAssociative())
2392 ReductionOps.insert(TreeN);
2399 // Visit left or right.
2400 Value *NextV = TreeN->getOperand(EdgeToVist);
2401 BinaryOperator *Next = dyn_cast<BinaryOperator>(NextV);
2403 Stack.push_back(std::make_pair(Next, 0));
2404 else if (NextV != Phi)
2410 /// \brief Attempt to vectorize the tree found by
2411 /// matchAssociativeReduction.
2412 bool tryToReduce(BoUpSLP &V, TargetTransformInfo *TTI) {
2413 if (ReducedVals.empty())
2416 unsigned NumReducedVals = ReducedVals.size();
2417 if (NumReducedVals < ReduxWidth)
2420 Value *VectorizedTree = nullptr;
2421 IRBuilder<> Builder(ReductionRoot);
2422 FastMathFlags Unsafe;
2423 Unsafe.setUnsafeAlgebra();
2424 Builder.SetFastMathFlags(Unsafe);
2427 for (; i < NumReducedVals - ReduxWidth + 1; i += ReduxWidth) {
2428 ArrayRef<Value *> ValsToReduce(&ReducedVals[i], ReduxWidth);
2429 V.buildTree(ValsToReduce, &ReductionOps);
2432 int Cost = V.getTreeCost() + getReductionCost(TTI, ReducedVals[i]);
2433 if (Cost >= -SLPCostThreshold)
2436 DEBUG(dbgs() << "SLP: Vectorizing horizontal reduction at cost:" << Cost
2439 // Vectorize a tree.
2440 DebugLoc Loc = cast<Instruction>(ReducedVals[i])->getDebugLoc();
2441 Value *VectorizedRoot = V.vectorizeTree();
2443 // Emit a reduction.
2444 Value *ReducedSubTree = emitReduction(VectorizedRoot, Builder);
2445 if (VectorizedTree) {
2446 Builder.SetCurrentDebugLocation(Loc);
2447 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2448 ReducedSubTree, "bin.rdx");
2450 VectorizedTree = ReducedSubTree;
2453 if (VectorizedTree) {
2454 // Finish the reduction.
2455 for (; i < NumReducedVals; ++i) {
2456 Builder.SetCurrentDebugLocation(
2457 cast<Instruction>(ReducedVals[i])->getDebugLoc());
2458 VectorizedTree = createBinOp(Builder, ReductionOpcode, VectorizedTree,
2463 assert(ReductionRoot && "Need a reduction operation");
2464 ReductionRoot->setOperand(0, VectorizedTree);
2465 ReductionRoot->setOperand(1, ReductionPHI);
2467 ReductionRoot->replaceAllUsesWith(VectorizedTree);
2469 return VectorizedTree != nullptr;
2474 /// \brief Calcuate the cost of a reduction.
2475 int getReductionCost(TargetTransformInfo *TTI, Value *FirstReducedVal) {
2476 Type *ScalarTy = FirstReducedVal->getType();
2477 Type *VecTy = VectorType::get(ScalarTy, ReduxWidth);
2479 int PairwiseRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, true);
2480 int SplittingRdxCost = TTI->getReductionCost(ReductionOpcode, VecTy, false);
2482 IsPairwiseReduction = PairwiseRdxCost < SplittingRdxCost;
2483 int VecReduxCost = IsPairwiseReduction ? PairwiseRdxCost : SplittingRdxCost;
2485 int ScalarReduxCost =
2486 ReduxWidth * TTI->getArithmeticInstrCost(ReductionOpcode, VecTy);
2488 DEBUG(dbgs() << "SLP: Adding cost " << VecReduxCost - ScalarReduxCost
2489 << " for reduction that starts with " << *FirstReducedVal
2491 << (IsPairwiseReduction ? "pairwise" : "splitting")
2492 << " reduction)\n");
2494 return VecReduxCost - ScalarReduxCost;
2497 static Value *createBinOp(IRBuilder<> &Builder, unsigned Opcode, Value *L,
2498 Value *R, const Twine &Name = "") {
2499 if (Opcode == Instruction::FAdd)
2500 return Builder.CreateFAdd(L, R, Name);
2501 return Builder.CreateBinOp((Instruction::BinaryOps)Opcode, L, R, Name);
2504 /// \brief Emit a horizontal reduction of the vectorized value.
2505 Value *emitReduction(Value *VectorizedValue, IRBuilder<> &Builder) {
2506 assert(VectorizedValue && "Need to have a vectorized tree node");
2507 Instruction *ValToReduce = dyn_cast<Instruction>(VectorizedValue);
2508 assert(isPowerOf2_32(ReduxWidth) &&
2509 "We only handle power-of-two reductions for now");
2511 Value *TmpVec = ValToReduce;
2512 for (unsigned i = ReduxWidth / 2; i != 0; i >>= 1) {
2513 if (IsPairwiseReduction) {
2515 createRdxShuffleMask(ReduxWidth, i, true, true, Builder);
2517 createRdxShuffleMask(ReduxWidth, i, true, false, Builder);
2519 Value *LeftShuf = Builder.CreateShuffleVector(
2520 TmpVec, UndefValue::get(TmpVec->getType()), LeftMask, "rdx.shuf.l");
2521 Value *RightShuf = Builder.CreateShuffleVector(
2522 TmpVec, UndefValue::get(TmpVec->getType()), (RightMask),
2524 TmpVec = createBinOp(Builder, ReductionOpcode, LeftShuf, RightShuf,
2528 createRdxShuffleMask(ReduxWidth, i, false, false, Builder);
2529 Value *Shuf = Builder.CreateShuffleVector(
2530 TmpVec, UndefValue::get(TmpVec->getType()), UpperHalf, "rdx.shuf");
2531 TmpVec = createBinOp(Builder, ReductionOpcode, TmpVec, Shuf, "bin.rdx");
2535 // The result is in the first element of the vector.
2536 return Builder.CreateExtractElement(TmpVec, Builder.getInt32(0));
2540 /// \brief Recognize construction of vectors like
2541 /// %ra = insertelement <4 x float> undef, float %s0, i32 0
2542 /// %rb = insertelement <4 x float> %ra, float %s1, i32 1
2543 /// %rc = insertelement <4 x float> %rb, float %s2, i32 2
2544 /// %rd = insertelement <4 x float> %rc, float %s3, i32 3
2546 /// Returns true if it matches
2548 static bool findBuildVector(InsertElementInst *IE,
2549 SmallVectorImpl<Value *> &Ops) {
2550 if (!isa<UndefValue>(IE->getOperand(0)))
2554 Ops.push_back(IE->getOperand(1));
2556 if (IE->use_empty())
2559 InsertElementInst *NextUse = dyn_cast<InsertElementInst>(IE->user_back());
2563 // If this isn't the final use, make sure the next insertelement is the only
2564 // use. It's OK if the final constructed vector is used multiple times
2565 if (!IE->hasOneUse())
2574 static bool PhiTypeSorterFunc(Value *V, Value *V2) {
2575 return V->getType() < V2->getType();
2578 bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock *BB, BoUpSLP &R) {
2579 bool Changed = false;
2580 SmallVector<Value *, 4> Incoming;
2581 SmallSet<Value *, 16> VisitedInstrs;
2583 bool HaveVectorizedPhiNodes = true;
2584 while (HaveVectorizedPhiNodes) {
2585 HaveVectorizedPhiNodes = false;
2587 // Collect the incoming values from the PHIs.
2589 for (BasicBlock::iterator instr = BB->begin(), ie = BB->end(); instr != ie;
2591 PHINode *P = dyn_cast<PHINode>(instr);
2595 if (!VisitedInstrs.count(P))
2596 Incoming.push_back(P);
2600 std::stable_sort(Incoming.begin(), Incoming.end(), PhiTypeSorterFunc);
2602 // Try to vectorize elements base on their type.
2603 for (SmallVector<Value *, 4>::iterator IncIt = Incoming.begin(),
2607 // Look for the next elements with the same type.
2608 SmallVector<Value *, 4>::iterator SameTypeIt = IncIt;
2609 while (SameTypeIt != E &&
2610 (*SameTypeIt)->getType() == (*IncIt)->getType()) {
2611 VisitedInstrs.insert(*SameTypeIt);
2615 // Try to vectorize them.
2616 unsigned NumElts = (SameTypeIt - IncIt);
2617 DEBUG(errs() << "SLP: Trying to vectorize starting at PHIs (" << NumElts << ")\n");
2619 tryToVectorizeList(ArrayRef<Value *>(IncIt, NumElts), R)) {
2620 // Success start over because instructions might have been changed.
2621 HaveVectorizedPhiNodes = true;
2626 // Start over at the next instruction of a different type (or the end).
2631 VisitedInstrs.clear();
2633 for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; it++) {
2634 // We may go through BB multiple times so skip the one we have checked.
2635 if (!VisitedInstrs.insert(it))
2638 if (isa<DbgInfoIntrinsic>(it))
2641 // Try to vectorize reductions that use PHINodes.
2642 if (PHINode *P = dyn_cast<PHINode>(it)) {
2643 // Check that the PHI is a reduction PHI.
2644 if (P->getNumIncomingValues() != 2)
2647 (P->getIncomingBlock(0) == BB
2648 ? (P->getIncomingValue(0))
2649 : (P->getIncomingBlock(1) == BB ? P->getIncomingValue(1)
2651 // Check if this is a Binary Operator.
2652 BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
2656 // Try to match and vectorize a horizontal reduction.
2657 HorizontalReduction HorRdx;
2658 if (ShouldVectorizeHor &&
2659 HorRdx.matchAssociativeReduction(P, BI, DL) &&
2660 HorRdx.tryToReduce(R, TTI)) {
2667 Value *Inst = BI->getOperand(0);
2669 Inst = BI->getOperand(1);
2671 if (tryToVectorize(dyn_cast<BinaryOperator>(Inst), R)) {
2672 // We would like to start over since some instructions are deleted
2673 // and the iterator may become invalid value.
2683 // Try to vectorize horizontal reductions feeding into a store.
2684 if (ShouldStartVectorizeHorAtStore)
2685 if (StoreInst *SI = dyn_cast<StoreInst>(it))
2686 if (BinaryOperator *BinOp =
2687 dyn_cast<BinaryOperator>(SI->getValueOperand())) {
2688 HorizontalReduction HorRdx;
2689 if (((HorRdx.matchAssociativeReduction(nullptr, BinOp, DL) &&
2690 HorRdx.tryToReduce(R, TTI)) ||
2691 tryToVectorize(BinOp, R))) {
2699 // Try to vectorize trees that start at compare instructions.
2700 if (CmpInst *CI = dyn_cast<CmpInst>(it)) {
2701 if (tryToVectorizePair(CI->getOperand(0), CI->getOperand(1), R)) {
2703 // We would like to start over since some instructions are deleted
2704 // and the iterator may become invalid value.
2710 for (int i = 0; i < 2; ++i) {
2711 if (BinaryOperator *BI = dyn_cast<BinaryOperator>(CI->getOperand(i))) {
2712 if (tryToVectorizePair(BI->getOperand(0), BI->getOperand(1), R)) {
2714 // We would like to start over since some instructions are deleted
2715 // and the iterator may become invalid value.
2724 // Try to vectorize trees that start at insertelement instructions.
2725 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(it)) {
2726 SmallVector<Value *, 8> Ops;
2727 if (!findBuildVector(IE, Ops))
2730 if (tryToVectorizeList(Ops, R)) {
2743 bool SLPVectorizer::vectorizeStoreChains(BoUpSLP &R) {
2744 bool Changed = false;
2745 // Attempt to sort and vectorize each of the store-groups.
2746 for (StoreListMap::iterator it = StoreRefs.begin(), e = StoreRefs.end();
2748 if (it->second.size() < 2)
2751 DEBUG(dbgs() << "SLP: Analyzing a store chain of length "
2752 << it->second.size() << ".\n");
2754 // Process the stores in chunks of 16.
2755 for (unsigned CI = 0, CE = it->second.size(); CI < CE; CI+=16) {
2756 unsigned Len = std::min<unsigned>(CE - CI, 16);
2757 ArrayRef<StoreInst *> Chunk(&it->second[CI], Len);
2758 Changed |= vectorizeStores(Chunk, -SLPCostThreshold, R);
2764 } // end anonymous namespace
2766 char SLPVectorizer::ID = 0;
2767 static const char lv_name[] = "SLP Vectorizer";
2768 INITIALIZE_PASS_BEGIN(SLPVectorizer, SV_NAME, lv_name, false, false)
2769 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
2770 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
2771 INITIALIZE_PASS_DEPENDENCY(ScalarEvolution)
2772 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
2773 INITIALIZE_PASS_END(SLPVectorizer, SV_NAME, lv_name, false, false)
2776 Pass *createSLPVectorizerPass() { return new SLPVectorizer(); }